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<bool> ClTrackOrigins("msan-track-origins",
137 cl::desc("Track origins (allocation sites) of poisoned memory"),
138 cl::Hidden, cl::init(false));
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(bool TrackOrigins = false,
203 StringRef BlacklistFile = StringRef())
205 TrackOrigins(TrackOrigins || 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 MSan runtime replacements for memmove, memcpy and memset.
254 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
256 /// \brief Address mask used in application-to-shadow address calculation.
257 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
259 /// \brief Offset of the origin shadow from the "normal" shadow.
260 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
261 uint64_t OriginOffset;
262 /// \brief Branch weights for error reporting.
263 MDNode *ColdCallWeights;
264 /// \brief Branch weights for origin store.
265 MDNode *OriginStoreWeights;
266 /// \brief Path to blacklist file.
267 SmallString<64> BlacklistFile;
268 /// \brief The blacklist.
269 std::unique_ptr<SpecialCaseList> BL;
270 /// \brief An empty volatile inline asm that prevents callback merge.
273 bool WrapIndirectCalls;
274 /// \brief Run-time wrapper for indirect calls.
275 Value *IndirectCallWrapperFn;
276 // Argument and return type of IndirectCallWrapperFn: void (*f)(void).
277 Type *AnyFunctionPtrTy;
279 friend struct MemorySanitizerVisitor;
280 friend struct VarArgAMD64Helper;
284 char MemorySanitizer::ID = 0;
285 INITIALIZE_PASS(MemorySanitizer, "msan",
286 "MemorySanitizer: detects uninitialized reads.",
289 FunctionPass *llvm::createMemorySanitizerPass(bool TrackOrigins,
290 StringRef BlacklistFile) {
291 return new MemorySanitizer(TrackOrigins, BlacklistFile);
294 /// \brief Create a non-const global initialized with the given string.
296 /// Creates a writable global for Str so that we can pass it to the
297 /// run-time lib. Runtime uses first 4 bytes of the string to store the
298 /// frame ID, so the string needs to be mutable.
299 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
301 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
302 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
303 GlobalValue::PrivateLinkage, StrConst, "");
307 /// \brief Insert extern declaration of runtime-provided functions and globals.
308 void MemorySanitizer::initializeCallbacks(Module &M) {
309 // Only do this once.
314 // Create the callback.
315 // FIXME: this function should have "Cold" calling conv,
316 // which is not yet implemented.
317 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
318 : "__msan_warning_noreturn";
319 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
321 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
322 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
323 IRB.getInt8PtrTy(), IntptrTy, NULL);
324 MsanPoisonStackFn = M.getOrInsertFunction(
325 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
326 MemmoveFn = M.getOrInsertFunction(
327 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
328 IRB.getInt8PtrTy(), IntptrTy, NULL);
329 MemcpyFn = M.getOrInsertFunction(
330 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
332 MemsetFn = M.getOrInsertFunction(
333 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
337 RetvalTLS = new GlobalVariable(
338 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
339 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
340 GlobalVariable::InitialExecTLSModel);
341 RetvalOriginTLS = new GlobalVariable(
342 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
343 "__msan_retval_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
345 ParamTLS = new GlobalVariable(
346 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
347 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
348 GlobalVariable::InitialExecTLSModel);
349 ParamOriginTLS = new GlobalVariable(
350 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
351 0, "__msan_param_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
353 VAArgTLS = new GlobalVariable(
354 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
355 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
356 GlobalVariable::InitialExecTLSModel);
357 VAArgOverflowSizeTLS = new GlobalVariable(
358 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
359 "__msan_va_arg_overflow_size_tls", 0,
360 GlobalVariable::InitialExecTLSModel);
361 OriginTLS = new GlobalVariable(
362 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
363 "__msan_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
365 // We insert an empty inline asm after __msan_report* to avoid callback merge.
366 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
367 StringRef(""), StringRef(""),
368 /*hasSideEffects=*/true);
370 if (WrapIndirectCalls) {
372 PointerType::getUnqual(FunctionType::get(IRB.getVoidTy(), false));
373 IndirectCallWrapperFn = M.getOrInsertFunction(
374 ClWrapIndirectCalls, AnyFunctionPtrTy, AnyFunctionPtrTy, NULL);
377 if (ClWrapIndirectCallsFast) {
378 MsandrModuleStart = new GlobalVariable(
379 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
380 0, "__executable_start");
381 MsandrModuleStart->setVisibility(GlobalVariable::HiddenVisibility);
382 MsandrModuleEnd = new GlobalVariable(
383 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
385 MsandrModuleEnd->setVisibility(GlobalVariable::HiddenVisibility);
389 /// \brief Module-level initialization.
391 /// inserts a call to __msan_init to the module's constructor list.
392 bool MemorySanitizer::doInitialization(Module &M) {
393 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
396 DL = &DLP->getDataLayout();
398 BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
399 C = &(M.getContext());
400 unsigned PtrSize = DL->getPointerSizeInBits(/* AddressSpace */0);
403 ShadowMask = kShadowMask64;
404 OriginOffset = kOriginOffset64;
407 ShadowMask = kShadowMask32;
408 OriginOffset = kOriginOffset32;
411 report_fatal_error("unsupported pointer size");
416 IntptrTy = IRB.getIntPtrTy(DL);
417 OriginTy = IRB.getInt32Ty();
419 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
420 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
422 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
423 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
424 "__msan_init", IRB.getVoidTy(), NULL)), 0);
427 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
428 IRB.getInt32(TrackOrigins), "__msan_track_origins");
431 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
432 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
439 /// \brief A helper class that handles instrumentation of VarArg
440 /// functions on a particular platform.
442 /// Implementations are expected to insert the instrumentation
443 /// necessary to propagate argument shadow through VarArg function
444 /// calls. Visit* methods are called during an InstVisitor pass over
445 /// the function, and should avoid creating new basic blocks. A new
446 /// instance of this class is created for each instrumented function.
447 struct VarArgHelper {
448 /// \brief Visit a CallSite.
449 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
451 /// \brief Visit a va_start call.
452 virtual void visitVAStartInst(VAStartInst &I) = 0;
454 /// \brief Visit a va_copy call.
455 virtual void visitVACopyInst(VACopyInst &I) = 0;
457 /// \brief Finalize function instrumentation.
459 /// This method is called after visiting all interesting (see above)
460 /// instructions in a function.
461 virtual void finalizeInstrumentation() = 0;
463 virtual ~VarArgHelper() {}
466 struct MemorySanitizerVisitor;
469 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
470 MemorySanitizerVisitor &Visitor);
472 /// This class does all the work for a given function. Store and Load
473 /// instructions store and load corresponding shadow and origin
474 /// values. Most instructions propagate shadow from arguments to their
475 /// return values. Certain instructions (most importantly, BranchInst)
476 /// test their argument shadow and print reports (with a runtime call) if it's
478 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
481 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
482 ValueMap<Value*, Value*> ShadowMap, OriginMap;
483 std::unique_ptr<VarArgHelper> VAHelper;
485 // The following flags disable parts of MSan instrumentation based on
486 // blacklist contents and command-line options.
491 bool CheckReturnValue;
493 struct ShadowOriginAndInsertPoint {
496 Instruction *OrigIns;
497 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
498 : Shadow(S), Origin(O), OrigIns(I) { }
499 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
501 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
502 SmallVector<Instruction*, 16> StoreList;
503 SmallVector<CallSite, 16> IndirectCallList;
505 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
506 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
507 bool SanitizeFunction = !MS.BL->isIn(F) && F.getAttributes().hasAttribute(
508 AttributeSet::FunctionIndex,
509 Attribute::SanitizeMemory);
510 InsertChecks = SanitizeFunction;
511 LoadShadow = SanitizeFunction;
512 PoisonStack = SanitizeFunction && ClPoisonStack;
513 PoisonUndef = SanitizeFunction && ClPoisonUndef;
514 // FIXME: Consider using SpecialCaseList to specify a list of functions that
515 // must always return fully initialized values. For now, we hardcode "main".
516 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
518 DEBUG(if (!InsertChecks)
519 dbgs() << "MemorySanitizer is not inserting checks into '"
520 << F.getName() << "'\n");
523 void materializeStores() {
524 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
525 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
528 Value *Val = I.getValueOperand();
529 Value *Addr = I.getPointerOperand();
530 Value *Shadow = I.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
531 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
534 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
535 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
538 if (ClCheckAccessAddress)
539 insertShadowCheck(Addr, &I);
542 I.setOrdering(addReleaseOrdering(I.getOrdering()));
544 if (MS.TrackOrigins) {
545 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
546 if (isa<StructType>(Shadow->getType())) {
547 IRB.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRB),
550 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
552 // TODO(eugenis): handle non-zero constant shadow by inserting an
553 // unconditional check (can not simply fail compilation as this could
554 // be in the dead code).
555 if (isa<Constant>(ConvertedShadow))
558 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
559 getCleanShadow(ConvertedShadow), "_mscmp");
560 Instruction *CheckTerm =
561 SplitBlockAndInsertIfThen(Cmp, &I, false, MS.OriginStoreWeights);
562 IRBuilder<> IRBNew(CheckTerm);
563 IRBNew.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRBNew),
570 void materializeChecks() {
571 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
572 Value *Shadow = InstrumentationList[i].Shadow;
573 Instruction *OrigIns = InstrumentationList[i].OrigIns;
574 IRBuilder<> IRB(OrigIns);
575 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
576 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
577 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
578 // See the comment in materializeStores().
579 if (isa<Constant>(ConvertedShadow))
581 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
582 getCleanShadow(ConvertedShadow), "_mscmp");
583 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
585 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
587 IRB.SetInsertPoint(CheckTerm);
588 if (MS.TrackOrigins) {
589 Value *Origin = InstrumentationList[i].Origin;
590 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
593 CallInst *Call = IRB.CreateCall(MS.WarningFn);
594 Call->setDebugLoc(OrigIns->getDebugLoc());
595 IRB.CreateCall(MS.EmptyAsm);
596 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
598 DEBUG(dbgs() << "DONE:\n" << F);
601 void materializeIndirectCalls() {
602 for (size_t i = 0, n = IndirectCallList.size(); i < n; i++) {
603 CallSite CS = IndirectCallList[i];
604 Instruction *I = CS.getInstruction();
605 BasicBlock *B = I->getParent();
607 Value *Fn0 = CS.getCalledValue();
608 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
610 if (ClWrapIndirectCallsFast) {
611 // Check that call target is inside this module limits.
613 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
614 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
616 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
617 IRB.CreateICmpUGE(Fn, End));
620 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
622 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
623 NotInThisModule, NewFnPhi,
624 /* Unreachable */ false, MS.ColdCallWeights);
626 IRB.SetInsertPoint(CheckTerm);
627 // Slow path: call wrapper function to possibly transform the call
629 Value *NewFn = IRB.CreateBitCast(
630 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
632 NewFnPhi->addIncoming(Fn0, B);
633 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
634 CS.setCalledFunction(NewFnPhi);
636 Value *NewFn = IRB.CreateBitCast(
637 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
638 CS.setCalledFunction(NewFn);
643 /// \brief Add MemorySanitizer instrumentation to a function.
644 bool runOnFunction() {
645 MS.initializeCallbacks(*F.getParent());
646 if (!MS.DL) return false;
648 // In the presence of unreachable blocks, we may see Phi nodes with
649 // incoming nodes from such blocks. Since InstVisitor skips unreachable
650 // blocks, such nodes will not have any shadow value associated with them.
651 // It's easier to remove unreachable blocks than deal with missing shadow.
652 removeUnreachableBlocks(F);
654 // Iterate all BBs in depth-first order and create shadow instructions
655 // for all instructions (where applicable).
656 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
657 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
658 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
659 BasicBlock *BB = *DI;
663 // Finalize PHI nodes.
664 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
665 PHINode *PN = ShadowPHINodes[i];
666 PHINode *PNS = cast<PHINode>(getShadow(PN));
667 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
668 size_t NumValues = PN->getNumIncomingValues();
669 for (size_t v = 0; v < NumValues; v++) {
670 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
672 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
676 VAHelper->finalizeInstrumentation();
678 // Delayed instrumentation of StoreInst.
679 // This may add new checks to be inserted later.
682 // Insert shadow value checks.
685 // Wrap indirect calls.
686 materializeIndirectCalls();
691 /// \brief Compute the shadow type that corresponds to a given Value.
692 Type *getShadowTy(Value *V) {
693 return getShadowTy(V->getType());
696 /// \brief Compute the shadow type that corresponds to a given Type.
697 Type *getShadowTy(Type *OrigTy) {
698 if (!OrigTy->isSized()) {
701 // For integer type, shadow is the same as the original type.
702 // This may return weird-sized types like i1.
703 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
705 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
706 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
707 return VectorType::get(IntegerType::get(*MS.C, EltSize),
708 VT->getNumElements());
710 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
711 SmallVector<Type*, 4> Elements;
712 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
713 Elements.push_back(getShadowTy(ST->getElementType(i)));
714 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
715 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
718 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
719 return IntegerType::get(*MS.C, TypeSize);
722 /// \brief Flatten a vector type.
723 Type *getShadowTyNoVec(Type *ty) {
724 if (VectorType *vt = dyn_cast<VectorType>(ty))
725 return IntegerType::get(*MS.C, vt->getBitWidth());
729 /// \brief Convert a shadow value to it's flattened variant.
730 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
731 Type *Ty = V->getType();
732 Type *NoVecTy = getShadowTyNoVec(Ty);
733 if (Ty == NoVecTy) return V;
734 return IRB.CreateBitCast(V, NoVecTy);
737 /// \brief Compute the shadow address that corresponds to a given application
740 /// Shadow = Addr & ~ShadowMask.
741 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
744 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
745 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
746 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
749 /// \brief Compute the origin address that corresponds to a given application
752 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
753 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
755 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
756 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
758 IRB.CreateAdd(ShadowLong,
759 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
761 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
762 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
765 /// \brief Compute the shadow address for a given function argument.
767 /// Shadow = ParamTLS+ArgOffset.
768 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
770 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
771 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
772 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
776 /// \brief Compute the origin address for a given function argument.
777 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
779 if (!MS.TrackOrigins) return 0;
780 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
781 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
782 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
786 /// \brief Compute the shadow address for a retval.
787 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
788 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
789 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
793 /// \brief Compute the origin address for a retval.
794 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
795 // We keep a single origin for the entire retval. Might be too optimistic.
796 return MS.RetvalOriginTLS;
799 /// \brief Set SV to be the shadow value for V.
800 void setShadow(Value *V, Value *SV) {
801 assert(!ShadowMap.count(V) && "Values may only have one shadow");
805 /// \brief Set Origin to be the origin value for V.
806 void setOrigin(Value *V, Value *Origin) {
807 if (!MS.TrackOrigins) return;
808 assert(!OriginMap.count(V) && "Values may only have one origin");
809 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
810 OriginMap[V] = Origin;
813 /// \brief Create a clean shadow value for a given value.
815 /// Clean shadow (all zeroes) means all bits of the value are defined
817 Constant *getCleanShadow(Value *V) {
818 Type *ShadowTy = getShadowTy(V);
821 return Constant::getNullValue(ShadowTy);
824 /// \brief Create a dirty shadow of a given shadow type.
825 Constant *getPoisonedShadow(Type *ShadowTy) {
827 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
828 return Constant::getAllOnesValue(ShadowTy);
829 StructType *ST = cast<StructType>(ShadowTy);
830 SmallVector<Constant *, 4> Vals;
831 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
832 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
833 return ConstantStruct::get(ST, Vals);
836 /// \brief Create a dirty shadow for a given value.
837 Constant *getPoisonedShadow(Value *V) {
838 Type *ShadowTy = getShadowTy(V);
841 return getPoisonedShadow(ShadowTy);
844 /// \brief Create a clean (zero) origin.
845 Value *getCleanOrigin() {
846 return Constant::getNullValue(MS.OriginTy);
849 /// \brief Get the shadow value for a given Value.
851 /// This function either returns the value set earlier with setShadow,
852 /// or extracts if from ParamTLS (for function arguments).
853 Value *getShadow(Value *V) {
854 if (Instruction *I = dyn_cast<Instruction>(V)) {
855 // For instructions the shadow is already stored in the map.
856 Value *Shadow = ShadowMap[V];
858 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
860 assert(Shadow && "No shadow for a value");
864 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
865 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
866 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
870 if (Argument *A = dyn_cast<Argument>(V)) {
871 // For arguments we compute the shadow on demand and store it in the map.
872 Value **ShadowPtr = &ShadowMap[V];
875 Function *F = A->getParent();
876 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
877 unsigned ArgOffset = 0;
878 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
880 if (!AI->getType()->isSized()) {
881 DEBUG(dbgs() << "Arg is not sized\n");
884 unsigned Size = AI->hasByValAttr()
885 ? MS.DL->getTypeAllocSize(AI->getType()->getPointerElementType())
886 : MS.DL->getTypeAllocSize(AI->getType());
888 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
889 if (AI->hasByValAttr()) {
890 // ByVal pointer itself has clean shadow. We copy the actual
891 // argument shadow to the underlying memory.
892 // Figure out maximal valid memcpy alignment.
893 unsigned ArgAlign = AI->getParamAlignment();
895 Type *EltType = A->getType()->getPointerElementType();
896 ArgAlign = MS.DL->getABITypeAlignment(EltType);
898 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
899 Value *Cpy = EntryIRB.CreateMemCpy(
900 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
902 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
904 *ShadowPtr = getCleanShadow(V);
906 *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
908 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
909 **ShadowPtr << "\n");
910 if (MS.TrackOrigins) {
911 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
912 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
915 ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
917 assert(*ShadowPtr && "Could not find shadow for an argument");
920 // For everything else the shadow is zero.
921 return getCleanShadow(V);
924 /// \brief Get the shadow for i-th argument of the instruction I.
925 Value *getShadow(Instruction *I, int i) {
926 return getShadow(I->getOperand(i));
929 /// \brief Get the origin for a value.
930 Value *getOrigin(Value *V) {
931 if (!MS.TrackOrigins) return 0;
932 if (isa<Instruction>(V) || isa<Argument>(V)) {
933 Value *Origin = OriginMap[V];
935 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
936 Origin = getCleanOrigin();
940 return getCleanOrigin();
943 /// \brief Get the origin for i-th argument of the instruction I.
944 Value *getOrigin(Instruction *I, int i) {
945 return getOrigin(I->getOperand(i));
948 /// \brief Remember the place where a shadow check should be inserted.
950 /// This location will be later instrumented with a check that will print a
951 /// UMR warning in runtime if the shadow value is not 0.
952 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
954 if (!InsertChecks) return;
956 Type *ShadowTy = Shadow->getType();
957 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
958 "Can only insert checks for integer and vector shadow types");
960 InstrumentationList.push_back(
961 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
964 /// \brief Remember the place where a shadow check should be inserted.
966 /// This location will be later instrumented with a check that will print a
967 /// UMR warning in runtime if the value is not fully defined.
968 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
970 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
972 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
973 insertShadowCheck(Shadow, Origin, OrigIns);
976 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
986 return AcquireRelease;
987 case SequentiallyConsistent:
988 return SequentiallyConsistent;
990 llvm_unreachable("Unknown ordering");
993 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1002 case AcquireRelease:
1003 return AcquireRelease;
1004 case SequentiallyConsistent:
1005 return SequentiallyConsistent;
1007 llvm_unreachable("Unknown ordering");
1010 // ------------------- Visitors.
1012 /// \brief Instrument LoadInst
1014 /// Loads the corresponding shadow and (optionally) origin.
1015 /// Optionally, checks that the load address is fully defined.
1016 void visitLoadInst(LoadInst &I) {
1017 assert(I.getType()->isSized() && "Load type must have size");
1018 IRBuilder<> IRB(I.getNextNode());
1019 Type *ShadowTy = getShadowTy(&I);
1020 Value *Addr = I.getPointerOperand();
1022 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1024 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1026 setShadow(&I, getCleanShadow(&I));
1029 if (ClCheckAccessAddress)
1030 insertShadowCheck(I.getPointerOperand(), &I);
1033 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1035 if (MS.TrackOrigins) {
1037 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1039 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1041 setOrigin(&I, getCleanOrigin());
1046 /// \brief Instrument StoreInst
1048 /// Stores the corresponding shadow and (optionally) origin.
1049 /// Optionally, checks that the store address is fully defined.
1050 void visitStoreInst(StoreInst &I) {
1051 StoreList.push_back(&I);
1054 void handleCASOrRMW(Instruction &I) {
1055 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1057 IRBuilder<> IRB(&I);
1058 Value *Addr = I.getOperand(0);
1059 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1061 if (ClCheckAccessAddress)
1062 insertShadowCheck(Addr, &I);
1064 // Only test the conditional argument of cmpxchg instruction.
1065 // The other argument can potentially be uninitialized, but we can not
1066 // detect this situation reliably without possible false positives.
1067 if (isa<AtomicCmpXchgInst>(I))
1068 insertShadowCheck(I.getOperand(1), &I);
1070 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1072 setShadow(&I, getCleanShadow(&I));
1075 void visitAtomicRMWInst(AtomicRMWInst &I) {
1077 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1080 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1082 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1085 // Vector manipulation.
1086 void visitExtractElementInst(ExtractElementInst &I) {
1087 insertShadowCheck(I.getOperand(1), &I);
1088 IRBuilder<> IRB(&I);
1089 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1091 setOrigin(&I, getOrigin(&I, 0));
1094 void visitInsertElementInst(InsertElementInst &I) {
1095 insertShadowCheck(I.getOperand(2), &I);
1096 IRBuilder<> IRB(&I);
1097 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1098 I.getOperand(2), "_msprop"));
1099 setOriginForNaryOp(I);
1102 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1103 insertShadowCheck(I.getOperand(2), &I);
1104 IRBuilder<> IRB(&I);
1105 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1106 I.getOperand(2), "_msprop"));
1107 setOriginForNaryOp(I);
1111 void visitSExtInst(SExtInst &I) {
1112 IRBuilder<> IRB(&I);
1113 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1114 setOrigin(&I, getOrigin(&I, 0));
1117 void visitZExtInst(ZExtInst &I) {
1118 IRBuilder<> IRB(&I);
1119 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1120 setOrigin(&I, getOrigin(&I, 0));
1123 void visitTruncInst(TruncInst &I) {
1124 IRBuilder<> IRB(&I);
1125 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1126 setOrigin(&I, getOrigin(&I, 0));
1129 void visitBitCastInst(BitCastInst &I) {
1130 IRBuilder<> IRB(&I);
1131 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1132 setOrigin(&I, getOrigin(&I, 0));
1135 void visitPtrToIntInst(PtrToIntInst &I) {
1136 IRBuilder<> IRB(&I);
1137 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1138 "_msprop_ptrtoint"));
1139 setOrigin(&I, getOrigin(&I, 0));
1142 void visitIntToPtrInst(IntToPtrInst &I) {
1143 IRBuilder<> IRB(&I);
1144 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1145 "_msprop_inttoptr"));
1146 setOrigin(&I, getOrigin(&I, 0));
1149 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1150 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1151 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1152 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1153 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1154 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1156 /// \brief Propagate shadow for bitwise AND.
1158 /// This code is exact, i.e. if, for example, a bit in the left argument
1159 /// is defined and 0, then neither the value not definedness of the
1160 /// corresponding bit in B don't affect the resulting shadow.
1161 void visitAnd(BinaryOperator &I) {
1162 IRBuilder<> IRB(&I);
1163 // "And" of 0 and a poisoned value results in unpoisoned value.
1164 // 1&1 => 1; 0&1 => 0; p&1 => p;
1165 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1166 // 1&p => p; 0&p => 0; p&p => p;
1167 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1168 Value *S1 = getShadow(&I, 0);
1169 Value *S2 = getShadow(&I, 1);
1170 Value *V1 = I.getOperand(0);
1171 Value *V2 = I.getOperand(1);
1172 if (V1->getType() != S1->getType()) {
1173 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1174 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1176 Value *S1S2 = IRB.CreateAnd(S1, S2);
1177 Value *V1S2 = IRB.CreateAnd(V1, S2);
1178 Value *S1V2 = IRB.CreateAnd(S1, V2);
1179 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1180 setOriginForNaryOp(I);
1183 void visitOr(BinaryOperator &I) {
1184 IRBuilder<> IRB(&I);
1185 // "Or" of 1 and a poisoned value results in unpoisoned value.
1186 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1187 // 1|0 => 1; 0|0 => 0; p|0 => p;
1188 // 1|p => 1; 0|p => p; p|p => p;
1189 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1190 Value *S1 = getShadow(&I, 0);
1191 Value *S2 = getShadow(&I, 1);
1192 Value *V1 = IRB.CreateNot(I.getOperand(0));
1193 Value *V2 = IRB.CreateNot(I.getOperand(1));
1194 if (V1->getType() != S1->getType()) {
1195 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1196 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1198 Value *S1S2 = IRB.CreateAnd(S1, S2);
1199 Value *V1S2 = IRB.CreateAnd(V1, S2);
1200 Value *S1V2 = IRB.CreateAnd(S1, V2);
1201 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1202 setOriginForNaryOp(I);
1205 /// \brief Default propagation of shadow and/or origin.
1207 /// This class implements the general case of shadow propagation, used in all
1208 /// cases where we don't know and/or don't care about what the operation
1209 /// actually does. It converts all input shadow values to a common type
1210 /// (extending or truncating as necessary), and bitwise OR's them.
1212 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1213 /// fully initialized), and less prone to false positives.
1215 /// This class also implements the general case of origin propagation. For a
1216 /// Nary operation, result origin is set to the origin of an argument that is
1217 /// not entirely initialized. If there is more than one such arguments, the
1218 /// rightmost of them is picked. It does not matter which one is picked if all
1219 /// arguments are initialized.
1220 template <bool CombineShadow>
1225 MemorySanitizerVisitor *MSV;
1228 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1229 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
1231 /// \brief Add a pair of shadow and origin values to the mix.
1232 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1233 if (CombineShadow) {
1238 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1239 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1243 if (MSV->MS.TrackOrigins) {
1248 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1249 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1250 MSV->getCleanShadow(FlatShadow));
1251 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1257 /// \brief Add an application value to the mix.
1258 Combiner &Add(Value *V) {
1259 Value *OpShadow = MSV->getShadow(V);
1260 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0;
1261 return Add(OpShadow, OpOrigin);
1264 /// \brief Set the current combined values as the given instruction's shadow
1266 void Done(Instruction *I) {
1267 if (CombineShadow) {
1269 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1270 MSV->setShadow(I, Shadow);
1272 if (MSV->MS.TrackOrigins) {
1274 MSV->setOrigin(I, Origin);
1279 typedef Combiner<true> ShadowAndOriginCombiner;
1280 typedef Combiner<false> OriginCombiner;
1282 /// \brief Propagate origin for arbitrary operation.
1283 void setOriginForNaryOp(Instruction &I) {
1284 if (!MS.TrackOrigins) return;
1285 IRBuilder<> IRB(&I);
1286 OriginCombiner OC(this, IRB);
1287 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1292 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1293 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1294 "Vector of pointers is not a valid shadow type");
1295 return Ty->isVectorTy() ?
1296 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1297 Ty->getPrimitiveSizeInBits();
1300 /// \brief Cast between two shadow types, extending or truncating as
1302 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1303 bool Signed = false) {
1304 Type *srcTy = V->getType();
1305 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1306 return IRB.CreateIntCast(V, dstTy, Signed);
1307 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1308 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1309 return IRB.CreateIntCast(V, dstTy, Signed);
1310 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1311 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1312 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1314 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1315 return IRB.CreateBitCast(V2, dstTy);
1316 // TODO: handle struct types.
1319 /// \brief Propagate shadow for arbitrary operation.
1320 void handleShadowOr(Instruction &I) {
1321 IRBuilder<> IRB(&I);
1322 ShadowAndOriginCombiner SC(this, IRB);
1323 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1328 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1329 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1330 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1331 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1332 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1333 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1334 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1336 void handleDiv(Instruction &I) {
1337 IRBuilder<> IRB(&I);
1338 // Strict on the second argument.
1339 insertShadowCheck(I.getOperand(1), &I);
1340 setShadow(&I, getShadow(&I, 0));
1341 setOrigin(&I, getOrigin(&I, 0));
1344 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1345 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1346 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1347 void visitURem(BinaryOperator &I) { handleDiv(I); }
1348 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1349 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1351 /// \brief Instrument == and != comparisons.
1353 /// Sometimes the comparison result is known even if some of the bits of the
1354 /// arguments are not.
1355 void handleEqualityComparison(ICmpInst &I) {
1356 IRBuilder<> IRB(&I);
1357 Value *A = I.getOperand(0);
1358 Value *B = I.getOperand(1);
1359 Value *Sa = getShadow(A);
1360 Value *Sb = getShadow(B);
1362 // Get rid of pointers and vectors of pointers.
1363 // For ints (and vectors of ints), types of A and Sa match,
1364 // and this is a no-op.
1365 A = IRB.CreatePointerCast(A, Sa->getType());
1366 B = IRB.CreatePointerCast(B, Sb->getType());
1368 // A == B <==> (C = A^B) == 0
1369 // A != B <==> (C = A^B) != 0
1371 Value *C = IRB.CreateXor(A, B);
1372 Value *Sc = IRB.CreateOr(Sa, Sb);
1373 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1374 // Result is defined if one of the following is true
1375 // * there is a defined 1 bit in C
1376 // * C is fully defined
1377 // Si = !(C & ~Sc) && Sc
1378 Value *Zero = Constant::getNullValue(Sc->getType());
1379 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1381 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1383 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1384 Si->setName("_msprop_icmp");
1386 setOriginForNaryOp(I);
1389 /// \brief Build the lowest possible value of V, taking into account V's
1390 /// uninitialized bits.
1391 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1394 // Split shadow into sign bit and other bits.
1395 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1396 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1397 // Maximise the undefined shadow bit, minimize other undefined bits.
1399 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1401 // Minimize undefined bits.
1402 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1406 /// \brief Build the highest possible value of V, taking into account V's
1407 /// uninitialized bits.
1408 Value *getHighestPossibleValue(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 // Minimise the undefined shadow bit, maximise other undefined bits.
1416 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1418 // Maximize undefined bits.
1419 return IRB.CreateOr(A, Sa);
1423 /// \brief Instrument relational comparisons.
1425 /// This function does exact shadow propagation for all relational
1426 /// comparisons of integers, pointers and vectors of those.
1427 /// FIXME: output seems suboptimal when one of the operands is a constant
1428 void handleRelationalComparisonExact(ICmpInst &I) {
1429 IRBuilder<> IRB(&I);
1430 Value *A = I.getOperand(0);
1431 Value *B = I.getOperand(1);
1432 Value *Sa = getShadow(A);
1433 Value *Sb = getShadow(B);
1435 // Get rid of pointers and vectors of pointers.
1436 // For ints (and vectors of ints), types of A and Sa match,
1437 // and this is a no-op.
1438 A = IRB.CreatePointerCast(A, Sa->getType());
1439 B = IRB.CreatePointerCast(B, Sb->getType());
1441 // Let [a0, a1] be the interval of possible values of A, taking into account
1442 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1443 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1444 bool IsSigned = I.isSigned();
1445 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1446 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1447 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1448 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1449 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1450 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1451 Value *Si = IRB.CreateXor(S1, S2);
1453 setOriginForNaryOp(I);
1456 /// \brief Instrument signed relational comparisons.
1458 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1459 /// propagating the highest bit of the shadow. Everything else is delegated
1460 /// to handleShadowOr().
1461 void handleSignedRelationalComparison(ICmpInst &I) {
1462 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1463 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1465 CmpInst::Predicate pre = I.getPredicate();
1466 if (constOp0 && constOp0->isNullValue() &&
1467 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1468 op = I.getOperand(1);
1469 } else if (constOp1 && constOp1->isNullValue() &&
1470 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1471 op = I.getOperand(0);
1474 IRBuilder<> IRB(&I);
1476 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1477 setShadow(&I, Shadow);
1478 setOrigin(&I, getOrigin(op));
1484 void visitICmpInst(ICmpInst &I) {
1485 if (!ClHandleICmp) {
1489 if (I.isEquality()) {
1490 handleEqualityComparison(I);
1494 assert(I.isRelational());
1495 if (ClHandleICmpExact) {
1496 handleRelationalComparisonExact(I);
1500 handleSignedRelationalComparison(I);
1504 assert(I.isUnsigned());
1505 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1506 handleRelationalComparisonExact(I);
1513 void visitFCmpInst(FCmpInst &I) {
1517 void handleShift(BinaryOperator &I) {
1518 IRBuilder<> IRB(&I);
1519 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1520 // Otherwise perform the same shift on S1.
1521 Value *S1 = getShadow(&I, 0);
1522 Value *S2 = getShadow(&I, 1);
1523 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1525 Value *V2 = I.getOperand(1);
1526 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1527 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1528 setOriginForNaryOp(I);
1531 void visitShl(BinaryOperator &I) { handleShift(I); }
1532 void visitAShr(BinaryOperator &I) { handleShift(I); }
1533 void visitLShr(BinaryOperator &I) { handleShift(I); }
1535 /// \brief Instrument llvm.memmove
1537 /// At this point we don't know if llvm.memmove will be inlined or not.
1538 /// If we don't instrument it and it gets inlined,
1539 /// our interceptor will not kick in and we will lose the memmove.
1540 /// If we instrument the call here, but it does not get inlined,
1541 /// we will memove the shadow twice: which is bad in case
1542 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1544 /// Similar situation exists for memcpy and memset.
1545 void visitMemMoveInst(MemMoveInst &I) {
1546 IRBuilder<> IRB(&I);
1549 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1550 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1551 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1552 I.eraseFromParent();
1555 // Similar to memmove: avoid copying shadow twice.
1556 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1557 // FIXME: consider doing manual inline for small constant sizes and proper
1559 void visitMemCpyInst(MemCpyInst &I) {
1560 IRBuilder<> IRB(&I);
1563 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1564 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1565 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1566 I.eraseFromParent();
1570 void visitMemSetInst(MemSetInst &I) {
1571 IRBuilder<> IRB(&I);
1574 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1575 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1576 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1577 I.eraseFromParent();
1580 void visitVAStartInst(VAStartInst &I) {
1581 VAHelper->visitVAStartInst(I);
1584 void visitVACopyInst(VACopyInst &I) {
1585 VAHelper->visitVACopyInst(I);
1588 enum IntrinsicKind {
1589 IK_DoesNotAccessMemory,
1594 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1595 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1596 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1597 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1598 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1599 const int UnknownModRefBehavior = IK_WritesMemory;
1600 #define GET_INTRINSIC_MODREF_BEHAVIOR
1601 #define ModRefBehavior IntrinsicKind
1602 #include "llvm/IR/Intrinsics.gen"
1603 #undef ModRefBehavior
1604 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1607 /// \brief Handle vector store-like intrinsics.
1609 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1610 /// has 1 pointer argument and 1 vector argument, returns void.
1611 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1612 IRBuilder<> IRB(&I);
1613 Value* Addr = I.getArgOperand(0);
1614 Value *Shadow = getShadow(&I, 1);
1615 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1617 // We don't know the pointer alignment (could be unaligned SSE store!).
1618 // Have to assume to worst case.
1619 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1621 if (ClCheckAccessAddress)
1622 insertShadowCheck(Addr, &I);
1624 // FIXME: use ClStoreCleanOrigin
1625 // FIXME: factor out common code from materializeStores
1626 if (MS.TrackOrigins)
1627 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1631 /// \brief Handle vector load-like intrinsics.
1633 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1634 /// has 1 pointer argument, returns a vector.
1635 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1636 IRBuilder<> IRB(&I);
1637 Value *Addr = I.getArgOperand(0);
1639 Type *ShadowTy = getShadowTy(&I);
1641 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1642 // We don't know the pointer alignment (could be unaligned SSE load!).
1643 // Have to assume to worst case.
1644 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1646 setShadow(&I, getCleanShadow(&I));
1649 if (ClCheckAccessAddress)
1650 insertShadowCheck(Addr, &I);
1652 if (MS.TrackOrigins) {
1654 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1656 setOrigin(&I, getCleanOrigin());
1661 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1663 /// Instrument intrinsics with any number of arguments of the same type,
1664 /// equal to the return type. The type should be simple (no aggregates or
1665 /// pointers; vectors are fine).
1666 /// Caller guarantees that this intrinsic does not access memory.
1667 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1668 Type *RetTy = I.getType();
1669 if (!(RetTy->isIntOrIntVectorTy() ||
1670 RetTy->isFPOrFPVectorTy() ||
1671 RetTy->isX86_MMXTy()))
1674 unsigned NumArgOperands = I.getNumArgOperands();
1676 for (unsigned i = 0; i < NumArgOperands; ++i) {
1677 Type *Ty = I.getArgOperand(i)->getType();
1682 IRBuilder<> IRB(&I);
1683 ShadowAndOriginCombiner SC(this, IRB);
1684 for (unsigned i = 0; i < NumArgOperands; ++i)
1685 SC.Add(I.getArgOperand(i));
1691 /// \brief Heuristically instrument unknown intrinsics.
1693 /// The main purpose of this code is to do something reasonable with all
1694 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1695 /// We recognize several classes of intrinsics by their argument types and
1696 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1697 /// sure that we know what the intrinsic does.
1699 /// We special-case intrinsics where this approach fails. See llvm.bswap
1700 /// handling as an example of that.
1701 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1702 unsigned NumArgOperands = I.getNumArgOperands();
1703 if (NumArgOperands == 0)
1706 Intrinsic::ID iid = I.getIntrinsicID();
1707 IntrinsicKind IK = getIntrinsicKind(iid);
1708 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1709 bool WritesMemory = IK == IK_WritesMemory;
1710 assert(!(OnlyReadsMemory && WritesMemory));
1712 if (NumArgOperands == 2 &&
1713 I.getArgOperand(0)->getType()->isPointerTy() &&
1714 I.getArgOperand(1)->getType()->isVectorTy() &&
1715 I.getType()->isVoidTy() &&
1717 // This looks like a vector store.
1718 return handleVectorStoreIntrinsic(I);
1721 if (NumArgOperands == 1 &&
1722 I.getArgOperand(0)->getType()->isPointerTy() &&
1723 I.getType()->isVectorTy() &&
1725 // This looks like a vector load.
1726 return handleVectorLoadIntrinsic(I);
1729 if (!OnlyReadsMemory && !WritesMemory)
1730 if (maybeHandleSimpleNomemIntrinsic(I))
1733 // FIXME: detect and handle SSE maskstore/maskload
1737 void handleBswap(IntrinsicInst &I) {
1738 IRBuilder<> IRB(&I);
1739 Value *Op = I.getArgOperand(0);
1740 Type *OpType = Op->getType();
1741 Function *BswapFunc = Intrinsic::getDeclaration(
1742 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1743 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1744 setOrigin(&I, getOrigin(Op));
1747 // \brief Instrument vector convert instrinsic.
1749 // This function instruments intrinsics like cvtsi2ss:
1750 // %Out = int_xxx_cvtyyy(%ConvertOp)
1752 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1753 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1754 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1755 // elements from \p CopyOp.
1756 // In most cases conversion involves floating-point value which may trigger a
1757 // hardware exception when not fully initialized. For this reason we require
1758 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1759 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1760 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1761 // return a fully initialized value.
1762 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1763 IRBuilder<> IRB(&I);
1764 Value *CopyOp, *ConvertOp;
1766 switch (I.getNumArgOperands()) {
1768 CopyOp = I.getArgOperand(0);
1769 ConvertOp = I.getArgOperand(1);
1772 ConvertOp = I.getArgOperand(0);
1776 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1779 // The first *NumUsedElements* elements of ConvertOp are converted to the
1780 // same number of output elements. The rest of the output is copied from
1781 // CopyOp, or (if not available) filled with zeroes.
1782 // Combine shadow for elements of ConvertOp that are used in this operation,
1783 // and insert a check.
1784 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1785 // int->any conversion.
1786 Value *ConvertShadow = getShadow(ConvertOp);
1787 Value *AggShadow = 0;
1788 if (ConvertOp->getType()->isVectorTy()) {
1789 AggShadow = IRB.CreateExtractElement(
1790 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1791 for (int i = 1; i < NumUsedElements; ++i) {
1792 Value *MoreShadow = IRB.CreateExtractElement(
1793 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1794 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1797 AggShadow = ConvertShadow;
1799 assert(AggShadow->getType()->isIntegerTy());
1800 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1802 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1805 assert(CopyOp->getType() == I.getType());
1806 assert(CopyOp->getType()->isVectorTy());
1807 Value *ResultShadow = getShadow(CopyOp);
1808 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1809 for (int i = 0; i < NumUsedElements; ++i) {
1810 ResultShadow = IRB.CreateInsertElement(
1811 ResultShadow, ConstantInt::getNullValue(EltTy),
1812 ConstantInt::get(IRB.getInt32Ty(), i));
1814 setShadow(&I, ResultShadow);
1815 setOrigin(&I, getOrigin(CopyOp));
1817 setShadow(&I, getCleanShadow(&I));
1821 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1822 // zeroes if it is zero, and all ones otherwise.
1823 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1824 if (S->getType()->isVectorTy())
1825 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1826 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1827 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1828 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1831 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1832 Type *T = S->getType();
1833 assert(T->isVectorTy());
1834 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1835 return IRB.CreateSExt(S2, T);
1838 // \brief Instrument vector shift instrinsic.
1840 // This function instruments intrinsics like int_x86_avx2_psll_w.
1841 // Intrinsic shifts %In by %ShiftSize bits.
1842 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
1843 // size, and the rest is ignored. Behavior is defined even if shift size is
1844 // greater than register (or field) width.
1845 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
1846 assert(I.getNumArgOperands() == 2);
1847 IRBuilder<> IRB(&I);
1848 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1849 // Otherwise perform the same shift on S1.
1850 Value *S1 = getShadow(&I, 0);
1851 Value *S2 = getShadow(&I, 1);
1852 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
1853 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
1854 Value *V1 = I.getOperand(0);
1855 Value *V2 = I.getOperand(1);
1856 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
1857 IRB.CreateBitCast(S1, V1->getType()), V2);
1858 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
1859 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1860 setOriginForNaryOp(I);
1863 void visitIntrinsicInst(IntrinsicInst &I) {
1864 switch (I.getIntrinsicID()) {
1865 case llvm::Intrinsic::bswap:
1868 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
1869 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
1870 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
1871 case llvm::Intrinsic::x86_avx512_cvtss2usi:
1872 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
1873 case llvm::Intrinsic::x86_avx512_cvttss2usi:
1874 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
1875 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
1876 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
1877 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
1878 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
1879 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
1880 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
1881 case llvm::Intrinsic::x86_sse2_cvtsd2si:
1882 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
1883 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
1884 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
1885 case llvm::Intrinsic::x86_sse2_cvtss2sd:
1886 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
1887 case llvm::Intrinsic::x86_sse2_cvttsd2si:
1888 case llvm::Intrinsic::x86_sse_cvtsi2ss:
1889 case llvm::Intrinsic::x86_sse_cvtsi642ss:
1890 case llvm::Intrinsic::x86_sse_cvtss2si64:
1891 case llvm::Intrinsic::x86_sse_cvtss2si:
1892 case llvm::Intrinsic::x86_sse_cvttss2si64:
1893 case llvm::Intrinsic::x86_sse_cvttss2si:
1894 handleVectorConvertIntrinsic(I, 1);
1896 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
1897 case llvm::Intrinsic::x86_sse2_cvtps2pd:
1898 case llvm::Intrinsic::x86_sse_cvtps2pi:
1899 case llvm::Intrinsic::x86_sse_cvttps2pi:
1900 handleVectorConvertIntrinsic(I, 2);
1902 case llvm::Intrinsic::x86_avx512_psll_dq:
1903 case llvm::Intrinsic::x86_avx512_psrl_dq:
1904 case llvm::Intrinsic::x86_avx2_psll_w:
1905 case llvm::Intrinsic::x86_avx2_psll_d:
1906 case llvm::Intrinsic::x86_avx2_psll_q:
1907 case llvm::Intrinsic::x86_avx2_pslli_w:
1908 case llvm::Intrinsic::x86_avx2_pslli_d:
1909 case llvm::Intrinsic::x86_avx2_pslli_q:
1910 case llvm::Intrinsic::x86_avx2_psll_dq:
1911 case llvm::Intrinsic::x86_avx2_psrl_w:
1912 case llvm::Intrinsic::x86_avx2_psrl_d:
1913 case llvm::Intrinsic::x86_avx2_psrl_q:
1914 case llvm::Intrinsic::x86_avx2_psra_w:
1915 case llvm::Intrinsic::x86_avx2_psra_d:
1916 case llvm::Intrinsic::x86_avx2_psrli_w:
1917 case llvm::Intrinsic::x86_avx2_psrli_d:
1918 case llvm::Intrinsic::x86_avx2_psrli_q:
1919 case llvm::Intrinsic::x86_avx2_psrai_w:
1920 case llvm::Intrinsic::x86_avx2_psrai_d:
1921 case llvm::Intrinsic::x86_avx2_psrl_dq:
1922 case llvm::Intrinsic::x86_sse2_psll_w:
1923 case llvm::Intrinsic::x86_sse2_psll_d:
1924 case llvm::Intrinsic::x86_sse2_psll_q:
1925 case llvm::Intrinsic::x86_sse2_pslli_w:
1926 case llvm::Intrinsic::x86_sse2_pslli_d:
1927 case llvm::Intrinsic::x86_sse2_pslli_q:
1928 case llvm::Intrinsic::x86_sse2_psll_dq:
1929 case llvm::Intrinsic::x86_sse2_psrl_w:
1930 case llvm::Intrinsic::x86_sse2_psrl_d:
1931 case llvm::Intrinsic::x86_sse2_psrl_q:
1932 case llvm::Intrinsic::x86_sse2_psra_w:
1933 case llvm::Intrinsic::x86_sse2_psra_d:
1934 case llvm::Intrinsic::x86_sse2_psrli_w:
1935 case llvm::Intrinsic::x86_sse2_psrli_d:
1936 case llvm::Intrinsic::x86_sse2_psrli_q:
1937 case llvm::Intrinsic::x86_sse2_psrai_w:
1938 case llvm::Intrinsic::x86_sse2_psrai_d:
1939 case llvm::Intrinsic::x86_sse2_psrl_dq:
1940 case llvm::Intrinsic::x86_mmx_psll_w:
1941 case llvm::Intrinsic::x86_mmx_psll_d:
1942 case llvm::Intrinsic::x86_mmx_psll_q:
1943 case llvm::Intrinsic::x86_mmx_pslli_w:
1944 case llvm::Intrinsic::x86_mmx_pslli_d:
1945 case llvm::Intrinsic::x86_mmx_pslli_q:
1946 case llvm::Intrinsic::x86_mmx_psrl_w:
1947 case llvm::Intrinsic::x86_mmx_psrl_d:
1948 case llvm::Intrinsic::x86_mmx_psrl_q:
1949 case llvm::Intrinsic::x86_mmx_psra_w:
1950 case llvm::Intrinsic::x86_mmx_psra_d:
1951 case llvm::Intrinsic::x86_mmx_psrli_w:
1952 case llvm::Intrinsic::x86_mmx_psrli_d:
1953 case llvm::Intrinsic::x86_mmx_psrli_q:
1954 case llvm::Intrinsic::x86_mmx_psrai_w:
1955 case llvm::Intrinsic::x86_mmx_psrai_d:
1956 handleVectorShiftIntrinsic(I, /* Variable */ false);
1958 case llvm::Intrinsic::x86_avx2_psllv_d:
1959 case llvm::Intrinsic::x86_avx2_psllv_d_256:
1960 case llvm::Intrinsic::x86_avx2_psllv_q:
1961 case llvm::Intrinsic::x86_avx2_psllv_q_256:
1962 case llvm::Intrinsic::x86_avx2_psrlv_d:
1963 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
1964 case llvm::Intrinsic::x86_avx2_psrlv_q:
1965 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
1966 case llvm::Intrinsic::x86_avx2_psrav_d:
1967 case llvm::Intrinsic::x86_avx2_psrav_d_256:
1968 handleVectorShiftIntrinsic(I, /* Variable */ true);
1971 // Byte shifts are not implemented.
1972 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
1973 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
1974 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
1975 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
1976 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
1977 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
1980 if (!handleUnknownIntrinsic(I))
1981 visitInstruction(I);
1986 void visitCallSite(CallSite CS) {
1987 Instruction &I = *CS.getInstruction();
1988 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1990 CallInst *Call = cast<CallInst>(&I);
1992 // For inline asm, do the usual thing: check argument shadow and mark all
1993 // outputs as clean. Note that any side effects of the inline asm that are
1994 // not immediately visible in its constraints are not handled.
1995 if (Call->isInlineAsm()) {
1996 visitInstruction(I);
2000 // Allow only tail calls with the same types, otherwise
2001 // we may have a false positive: shadow for a non-void RetVal
2002 // will get propagated to a void RetVal.
2003 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
2004 Call->setTailCall(false);
2006 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2008 // We are going to insert code that relies on the fact that the callee
2009 // will become a non-readonly function after it is instrumented by us. To
2010 // prevent this code from being optimized out, mark that function
2011 // non-readonly in advance.
2012 if (Function *Func = Call->getCalledFunction()) {
2013 // Clear out readonly/readnone attributes.
2015 B.addAttribute(Attribute::ReadOnly)
2016 .addAttribute(Attribute::ReadNone);
2017 Func->removeAttributes(AttributeSet::FunctionIndex,
2018 AttributeSet::get(Func->getContext(),
2019 AttributeSet::FunctionIndex,
2023 IRBuilder<> IRB(&I);
2025 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2026 IndirectCallList.push_back(CS);
2028 unsigned ArgOffset = 0;
2029 DEBUG(dbgs() << " CallSite: " << I << "\n");
2030 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2031 ArgIt != End; ++ArgIt) {
2033 unsigned i = ArgIt - CS.arg_begin();
2034 if (!A->getType()->isSized()) {
2035 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2040 // Compute the Shadow for arg even if it is ByVal, because
2041 // in that case getShadow() will copy the actual arg shadow to
2042 // __msan_param_tls.
2043 Value *ArgShadow = getShadow(A);
2044 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2045 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2046 " Shadow: " << *ArgShadow << "\n");
2047 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2048 assert(A->getType()->isPointerTy() &&
2049 "ByVal argument is not a pointer!");
2050 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2051 unsigned Alignment = CS.getParamAlignment(i + 1);
2052 Store = IRB.CreateMemCpy(ArgShadowBase,
2053 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2056 Size = MS.DL->getTypeAllocSize(A->getType());
2057 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2058 kShadowTLSAlignment);
2060 if (MS.TrackOrigins)
2061 IRB.CreateStore(getOrigin(A),
2062 getOriginPtrForArgument(A, IRB, ArgOffset));
2064 assert(Size != 0 && Store != 0);
2065 DEBUG(dbgs() << " Param:" << *Store << "\n");
2066 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
2068 DEBUG(dbgs() << " done with call args\n");
2071 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2072 if (FT->isVarArg()) {
2073 VAHelper->visitCallSite(CS, IRB);
2076 // Now, get the shadow for the RetVal.
2077 if (!I.getType()->isSized()) return;
2078 IRBuilder<> IRBBefore(&I);
2079 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2080 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2081 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2082 Instruction *NextInsn = 0;
2084 NextInsn = I.getNextNode();
2086 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2087 if (!NormalDest->getSinglePredecessor()) {
2088 // FIXME: this case is tricky, so we are just conservative here.
2089 // Perhaps we need to split the edge between this BB and NormalDest,
2090 // but a naive attempt to use SplitEdge leads to a crash.
2091 setShadow(&I, getCleanShadow(&I));
2092 setOrigin(&I, getCleanOrigin());
2095 NextInsn = NormalDest->getFirstInsertionPt();
2097 "Could not find insertion point for retval shadow load");
2099 IRBuilder<> IRBAfter(NextInsn);
2100 Value *RetvalShadow =
2101 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2102 kShadowTLSAlignment, "_msret");
2103 setShadow(&I, RetvalShadow);
2104 if (MS.TrackOrigins)
2105 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2108 void visitReturnInst(ReturnInst &I) {
2109 IRBuilder<> IRB(&I);
2110 Value *RetVal = I.getReturnValue();
2111 if (!RetVal) return;
2112 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2113 if (CheckReturnValue) {
2114 insertShadowCheck(RetVal, &I);
2115 Value *Shadow = getCleanShadow(RetVal);
2116 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2118 Value *Shadow = getShadow(RetVal);
2119 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2120 // FIXME: make it conditional if ClStoreCleanOrigin==0
2121 if (MS.TrackOrigins)
2122 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2126 void visitPHINode(PHINode &I) {
2127 IRBuilder<> IRB(&I);
2128 ShadowPHINodes.push_back(&I);
2129 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2131 if (MS.TrackOrigins)
2132 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2136 void visitAllocaInst(AllocaInst &I) {
2137 setShadow(&I, getCleanShadow(&I));
2138 IRBuilder<> IRB(I.getNextNode());
2139 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2140 if (PoisonStack && ClPoisonStackWithCall) {
2141 IRB.CreateCall2(MS.MsanPoisonStackFn,
2142 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2143 ConstantInt::get(MS.IntptrTy, Size));
2145 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2146 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2147 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2150 if (PoisonStack && MS.TrackOrigins) {
2151 setOrigin(&I, getCleanOrigin());
2152 SmallString<2048> StackDescriptionStorage;
2153 raw_svector_ostream StackDescription(StackDescriptionStorage);
2154 // We create a string with a description of the stack allocation and
2155 // pass it into __msan_set_alloca_origin.
2156 // It will be printed by the run-time if stack-originated UMR is found.
2157 // The first 4 bytes of the string are set to '----' and will be replaced
2158 // by __msan_va_arg_overflow_size_tls at the first call.
2159 StackDescription << "----" << I.getName() << "@" << F.getName();
2161 createPrivateNonConstGlobalForString(*F.getParent(),
2162 StackDescription.str());
2164 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2165 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2166 ConstantInt::get(MS.IntptrTy, Size),
2167 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2168 IRB.CreatePointerCast(&F, MS.IntptrTy));
2172 void visitSelectInst(SelectInst& I) {
2173 IRBuilder<> IRB(&I);
2174 // a = select b, c, d
2175 Value *S = IRB.CreateSelect(I.getCondition(), getShadow(I.getTrueValue()),
2176 getShadow(I.getFalseValue()));
2177 if (I.getType()->isAggregateType()) {
2178 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2179 // an extra "select". This results in much more compact IR.
2180 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2181 S = IRB.CreateSelect(getShadow(I.getCondition()),
2182 getPoisonedShadow(getShadowTy(I.getType())), S,
2183 "_msprop_select_agg");
2185 // Sa = (sext Sb) | (select b, Sc, Sd)
2186 S = IRB.CreateOr(S, CreateShadowCast(IRB, getShadow(I.getCondition()),
2187 S->getType(), true),
2191 if (MS.TrackOrigins) {
2192 // Origins are always i32, so any vector conditions must be flattened.
2193 // FIXME: consider tracking vector origins for app vectors?
2194 Value *Cond = I.getCondition();
2195 Value *CondShadow = getShadow(Cond);
2196 if (Cond->getType()->isVectorTy()) {
2197 Type *FlatTy = getShadowTyNoVec(Cond->getType());
2198 Cond = IRB.CreateICmpNE(IRB.CreateBitCast(Cond, FlatTy),
2199 ConstantInt::getNullValue(FlatTy));
2200 CondShadow = IRB.CreateICmpNE(IRB.CreateBitCast(CondShadow, FlatTy),
2201 ConstantInt::getNullValue(FlatTy));
2203 // a = select b, c, d
2204 // Oa = Sb ? Ob : (b ? Oc : Od)
2205 setOrigin(&I, IRB.CreateSelect(
2206 CondShadow, getOrigin(I.getCondition()),
2207 IRB.CreateSelect(Cond, getOrigin(I.getTrueValue()),
2208 getOrigin(I.getFalseValue()))));
2212 void visitLandingPadInst(LandingPadInst &I) {
2214 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2215 setShadow(&I, getCleanShadow(&I));
2216 setOrigin(&I, getCleanOrigin());
2219 void visitGetElementPtrInst(GetElementPtrInst &I) {
2223 void visitExtractValueInst(ExtractValueInst &I) {
2224 IRBuilder<> IRB(&I);
2225 Value *Agg = I.getAggregateOperand();
2226 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2227 Value *AggShadow = getShadow(Agg);
2228 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2229 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2230 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2231 setShadow(&I, ResShadow);
2232 setOriginForNaryOp(I);
2235 void visitInsertValueInst(InsertValueInst &I) {
2236 IRBuilder<> IRB(&I);
2237 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2238 Value *AggShadow = getShadow(I.getAggregateOperand());
2239 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2240 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2241 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2242 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2243 DEBUG(dbgs() << " Res: " << *Res << "\n");
2245 setOriginForNaryOp(I);
2248 void dumpInst(Instruction &I) {
2249 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2250 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2252 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2254 errs() << "QQQ " << I << "\n";
2257 void visitResumeInst(ResumeInst &I) {
2258 DEBUG(dbgs() << "Resume: " << I << "\n");
2259 // Nothing to do here.
2262 void visitInstruction(Instruction &I) {
2263 // Everything else: stop propagating and check for poisoned shadow.
2264 if (ClDumpStrictInstructions)
2266 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2267 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2268 insertShadowCheck(I.getOperand(i), &I);
2269 setShadow(&I, getCleanShadow(&I));
2270 setOrigin(&I, getCleanOrigin());
2274 /// \brief AMD64-specific implementation of VarArgHelper.
2275 struct VarArgAMD64Helper : public VarArgHelper {
2276 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2277 // See a comment in visitCallSite for more details.
2278 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2279 static const unsigned AMD64FpEndOffset = 176;
2282 MemorySanitizer &MS;
2283 MemorySanitizerVisitor &MSV;
2284 Value *VAArgTLSCopy;
2285 Value *VAArgOverflowSize;
2287 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2289 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2290 MemorySanitizerVisitor &MSV)
2291 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
2293 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2295 ArgKind classifyArgument(Value* arg) {
2296 // A very rough approximation of X86_64 argument classification rules.
2297 Type *T = arg->getType();
2298 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2299 return AK_FloatingPoint;
2300 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2301 return AK_GeneralPurpose;
2302 if (T->isPointerTy())
2303 return AK_GeneralPurpose;
2307 // For VarArg functions, store the argument shadow in an ABI-specific format
2308 // that corresponds to va_list layout.
2309 // We do this because Clang lowers va_arg in the frontend, and this pass
2310 // only sees the low level code that deals with va_list internals.
2311 // A much easier alternative (provided that Clang emits va_arg instructions)
2312 // would have been to associate each live instance of va_list with a copy of
2313 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2315 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2316 unsigned GpOffset = 0;
2317 unsigned FpOffset = AMD64GpEndOffset;
2318 unsigned OverflowOffset = AMD64FpEndOffset;
2319 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2320 ArgIt != End; ++ArgIt) {
2322 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2323 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2325 // ByVal arguments always go to the overflow area.
2326 assert(A->getType()->isPointerTy());
2327 Type *RealTy = A->getType()->getPointerElementType();
2328 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2329 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2330 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2331 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2332 ArgSize, kShadowTLSAlignment);
2334 ArgKind AK = classifyArgument(A);
2335 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2337 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2341 case AK_GeneralPurpose:
2342 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2345 case AK_FloatingPoint:
2346 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2350 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2351 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2352 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2354 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2357 Constant *OverflowSize =
2358 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2359 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2362 /// \brief Compute the shadow address for a given va_arg.
2363 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2365 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2366 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2367 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2371 void visitVAStartInst(VAStartInst &I) override {
2372 IRBuilder<> IRB(&I);
2373 VAStartInstrumentationList.push_back(&I);
2374 Value *VAListTag = I.getArgOperand(0);
2375 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2377 // Unpoison the whole __va_list_tag.
2378 // FIXME: magic ABI constants.
2379 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2380 /* size */24, /* alignment */8, false);
2383 void visitVACopyInst(VACopyInst &I) override {
2384 IRBuilder<> IRB(&I);
2385 Value *VAListTag = I.getArgOperand(0);
2386 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2388 // Unpoison the whole __va_list_tag.
2389 // FIXME: magic ABI constants.
2390 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2391 /* size */24, /* alignment */8, false);
2394 void finalizeInstrumentation() override {
2395 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2396 "finalizeInstrumentation called twice");
2397 if (!VAStartInstrumentationList.empty()) {
2398 // If there is a va_start in this function, make a backup copy of
2399 // va_arg_tls somewhere in the function entry block.
2400 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2401 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2403 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2405 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2406 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2409 // Instrument va_start.
2410 // Copy va_list shadow from the backup copy of the TLS contents.
2411 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2412 CallInst *OrigInst = VAStartInstrumentationList[i];
2413 IRBuilder<> IRB(OrigInst->getNextNode());
2414 Value *VAListTag = OrigInst->getArgOperand(0);
2416 Value *RegSaveAreaPtrPtr =
2418 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2419 ConstantInt::get(MS.IntptrTy, 16)),
2420 Type::getInt64PtrTy(*MS.C));
2421 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2422 Value *RegSaveAreaShadowPtr =
2423 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2424 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2425 AMD64FpEndOffset, 16);
2427 Value *OverflowArgAreaPtrPtr =
2429 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2430 ConstantInt::get(MS.IntptrTy, 8)),
2431 Type::getInt64PtrTy(*MS.C));
2432 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2433 Value *OverflowArgAreaShadowPtr =
2434 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2435 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2436 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2441 /// \brief A no-op implementation of VarArgHelper.
2442 struct VarArgNoOpHelper : public VarArgHelper {
2443 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2444 MemorySanitizerVisitor &MSV) {}
2446 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2448 void visitVAStartInst(VAStartInst &I) override {}
2450 void visitVACopyInst(VACopyInst &I) override {}
2452 void finalizeInstrumentation() override {}
2455 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2456 MemorySanitizerVisitor &Visitor) {
2457 // VarArg handling is only implemented on AMD64. False positives are possible
2458 // on other platforms.
2459 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2460 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2461 return new VarArgAMD64Helper(Func, Msan, Visitor);
2463 return new VarArgNoOpHelper(Func, Msan, Visitor);
2468 bool MemorySanitizer::runOnFunction(Function &F) {
2469 MemorySanitizerVisitor Visitor(F, *this);
2471 // Clear out readonly/readnone attributes.
2473 B.addAttribute(Attribute::ReadOnly)
2474 .addAttribute(Attribute::ReadNone);
2475 F.removeAttributes(AttributeSet::FunctionIndex,
2476 AttributeSet::get(F.getContext(),
2477 AttributeSet::FunctionIndex, B));
2479 return Visitor.runOnFunction();