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 #include "llvm/Transforms/Instrumentation.h"
97 #include "llvm/ADT/DepthFirstIterator.h"
98 #include "llvm/ADT/SmallString.h"
99 #include "llvm/ADT/SmallVector.h"
100 #include "llvm/ADT/StringExtras.h"
101 #include "llvm/ADT/Triple.h"
102 #include "llvm/IR/DataLayout.h"
103 #include "llvm/IR/Function.h"
104 #include "llvm/IR/IRBuilder.h"
105 #include "llvm/IR/InlineAsm.h"
106 #include "llvm/IR/InstVisitor.h"
107 #include "llvm/IR/IntrinsicInst.h"
108 #include "llvm/IR/LLVMContext.h"
109 #include "llvm/IR/MDBuilder.h"
110 #include "llvm/IR/Module.h"
111 #include "llvm/IR/Type.h"
112 #include "llvm/IR/ValueMap.h"
113 #include "llvm/Support/CommandLine.h"
114 #include "llvm/Support/Compiler.h"
115 #include "llvm/Support/Debug.h"
116 #include "llvm/Support/raw_ostream.h"
117 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
118 #include "llvm/Transforms/Utils/Local.h"
119 #include "llvm/Transforms/Utils/ModuleUtils.h"
120 #include "llvm/Transforms/Utils/SpecialCaseList.h"
122 using namespace llvm;
124 #define DEBUG_TYPE "msan"
126 static const uint64_t kShadowMask32 = 1ULL << 31;
127 static const uint64_t kShadowMask64 = 1ULL << 46;
128 static const uint64_t kOriginOffset32 = 1ULL << 30;
129 static const uint64_t kOriginOffset64 = 1ULL << 45;
130 static const unsigned kMinOriginAlignment = 4;
131 static const unsigned kShadowTLSAlignment = 8;
133 // Accesses sizes are powers of two: 1, 2, 4, 8.
134 static const size_t kNumberOfAccessSizes = 4;
136 /// \brief Track origins of uninitialized values.
138 /// Adds a section to MemorySanitizer report that points to the allocation
139 /// (stack or heap) the uninitialized bits came from originally.
140 static cl::opt<int> ClTrackOrigins("msan-track-origins",
141 cl::desc("Track origins (allocation sites) of poisoned memory"),
142 cl::Hidden, cl::init(0));
143 static cl::opt<bool> ClKeepGoing("msan-keep-going",
144 cl::desc("keep going after reporting a UMR"),
145 cl::Hidden, cl::init(false));
146 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
147 cl::desc("poison uninitialized stack variables"),
148 cl::Hidden, cl::init(true));
149 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
150 cl::desc("poison uninitialized stack variables with a call"),
151 cl::Hidden, cl::init(false));
152 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
153 cl::desc("poison uninitialized stack variables with the given patter"),
154 cl::Hidden, cl::init(0xff));
155 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
156 cl::desc("poison undef temps"),
157 cl::Hidden, cl::init(true));
159 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
160 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
161 cl::Hidden, cl::init(true));
163 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
164 cl::desc("exact handling of relational integer ICmp"),
165 cl::Hidden, cl::init(false));
167 // This flag controls whether we check the shadow of the address
168 // operand of load or store. Such bugs are very rare, since load from
169 // a garbage address typically results in SEGV, but still happen
170 // (e.g. only lower bits of address are garbage, or the access happens
171 // early at program startup where malloc-ed memory is more likely to
172 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
173 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
174 cl::desc("report accesses through a pointer which has poisoned shadow"),
175 cl::Hidden, cl::init(true));
177 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
178 cl::desc("print out instructions with default strict semantics"),
179 cl::Hidden, cl::init(false));
181 static cl::opt<std::string> ClBlacklistFile("msan-blacklist",
182 cl::desc("File containing the list of functions where MemorySanitizer "
183 "should not report bugs"), cl::Hidden);
185 static cl::opt<int> ClInstrumentationWithCallThreshold(
186 "msan-instrumentation-with-call-threshold",
188 "If the function being instrumented requires more than "
189 "this number of checks and origin stores, use callbacks instead of "
190 "inline checks (-1 means never use callbacks)."),
191 cl::Hidden, cl::init(3500));
193 // Experimental. Wraps all indirect calls in the instrumented code with
194 // a call to the given function. This is needed to assist the dynamic
195 // helper tool (MSanDR) to regain control on transition between instrumented and
196 // non-instrumented code.
197 static cl::opt<std::string> ClWrapIndirectCalls("msan-wrap-indirect-calls",
198 cl::desc("Wrap indirect calls with a given function"),
201 static cl::opt<bool> ClWrapIndirectCallsFast("msan-wrap-indirect-calls-fast",
202 cl::desc("Do not wrap indirect calls with target in the same module"),
203 cl::Hidden, cl::init(true));
207 /// \brief An instrumentation pass implementing detection of uninitialized
210 /// MemorySanitizer: instrument the code in module to find
211 /// uninitialized reads.
212 class MemorySanitizer : public FunctionPass {
214 MemorySanitizer(int TrackOrigins = 0,
215 StringRef BlacklistFile = StringRef())
217 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
220 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile : BlacklistFile),
221 WrapIndirectCalls(!ClWrapIndirectCalls.empty()) {}
222 const char *getPassName() const override { return "MemorySanitizer"; }
223 bool runOnFunction(Function &F) override;
224 bool doInitialization(Module &M) override;
225 static char ID; // Pass identification, replacement for typeid.
228 void initializeCallbacks(Module &M);
230 /// \brief Track origins (allocation points) of uninitialized values.
233 const DataLayout *DL;
237 /// \brief Thread-local shadow storage for function parameters.
238 GlobalVariable *ParamTLS;
239 /// \brief Thread-local origin storage for function parameters.
240 GlobalVariable *ParamOriginTLS;
241 /// \brief Thread-local shadow storage for function return value.
242 GlobalVariable *RetvalTLS;
243 /// \brief Thread-local origin storage for function return value.
244 GlobalVariable *RetvalOriginTLS;
245 /// \brief Thread-local shadow storage for in-register va_arg function
246 /// parameters (x86_64-specific).
247 GlobalVariable *VAArgTLS;
248 /// \brief Thread-local shadow storage for va_arg overflow area
249 /// (x86_64-specific).
250 GlobalVariable *VAArgOverflowSizeTLS;
251 /// \brief Thread-local space used to pass origin value to the UMR reporting
253 GlobalVariable *OriginTLS;
255 GlobalVariable *MsandrModuleStart;
256 GlobalVariable *MsandrModuleEnd;
258 /// \brief The run-time callback to print a warning.
260 // These arrays are indexed by log2(AccessSize).
261 Value *MaybeWarningFn[kNumberOfAccessSizes];
262 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
264 /// \brief Run-time helper that generates a new origin value for a stack
266 Value *MsanSetAllocaOrigin4Fn;
267 /// \brief Run-time helper that poisons stack on function entry.
268 Value *MsanPoisonStackFn;
269 /// \brief Run-time helper that records a store (or any event) of an
270 /// uninitialized value and returns an updated origin id encoding this info.
271 Value *MsanChainOriginFn;
272 /// \brief MSan runtime replacements for memmove, memcpy and memset.
273 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
275 /// \brief Address mask used in application-to-shadow address calculation.
276 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
278 /// \brief Offset of the origin shadow from the "normal" shadow.
279 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
280 uint64_t OriginOffset;
281 /// \brief Branch weights for error reporting.
282 MDNode *ColdCallWeights;
283 /// \brief Branch weights for origin store.
284 MDNode *OriginStoreWeights;
285 /// \brief Path to blacklist file.
286 SmallString<64> BlacklistFile;
287 /// \brief The blacklist.
288 std::unique_ptr<SpecialCaseList> BL;
289 /// \brief An empty volatile inline asm that prevents callback merge.
292 bool WrapIndirectCalls;
293 /// \brief Run-time wrapper for indirect calls.
294 Value *IndirectCallWrapperFn;
295 // Argument and return type of IndirectCallWrapperFn: void (*f)(void).
296 Type *AnyFunctionPtrTy;
298 friend struct MemorySanitizerVisitor;
299 friend struct VarArgAMD64Helper;
303 char MemorySanitizer::ID = 0;
304 INITIALIZE_PASS(MemorySanitizer, "msan",
305 "MemorySanitizer: detects uninitialized reads.",
308 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins,
309 StringRef BlacklistFile) {
310 return new MemorySanitizer(TrackOrigins, BlacklistFile);
313 /// \brief Create a non-const global initialized with the given string.
315 /// Creates a writable global for Str so that we can pass it to the
316 /// run-time lib. Runtime uses first 4 bytes of the string to store the
317 /// frame ID, so the string needs to be mutable.
318 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
320 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
321 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
322 GlobalValue::PrivateLinkage, StrConst, "");
326 /// \brief Insert extern declaration of runtime-provided functions and globals.
327 void MemorySanitizer::initializeCallbacks(Module &M) {
328 // Only do this once.
333 // Create the callback.
334 // FIXME: this function should have "Cold" calling conv,
335 // which is not yet implemented.
336 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
337 : "__msan_warning_noreturn";
338 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
340 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
342 unsigned AccessSize = 1 << AccessSizeIndex;
343 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
344 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
345 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
346 IRB.getInt32Ty(), NULL);
348 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
349 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
350 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
351 IRB.getInt8PtrTy(), IRB.getInt32Ty(), NULL);
354 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
355 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
356 IRB.getInt8PtrTy(), IntptrTy, NULL);
357 MsanPoisonStackFn = M.getOrInsertFunction(
358 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
359 MsanChainOriginFn = M.getOrInsertFunction(
360 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), NULL);
361 MemmoveFn = M.getOrInsertFunction(
362 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
363 IRB.getInt8PtrTy(), IntptrTy, NULL);
364 MemcpyFn = M.getOrInsertFunction(
365 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
367 MemsetFn = M.getOrInsertFunction(
368 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
372 RetvalTLS = new GlobalVariable(
373 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
374 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
375 GlobalVariable::InitialExecTLSModel);
376 RetvalOriginTLS = new GlobalVariable(
377 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
378 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
380 ParamTLS = new GlobalVariable(
381 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
382 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
383 GlobalVariable::InitialExecTLSModel);
384 ParamOriginTLS = new GlobalVariable(
385 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
386 nullptr, "__msan_param_origin_tls", nullptr,
387 GlobalVariable::InitialExecTLSModel);
389 VAArgTLS = new GlobalVariable(
390 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
391 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
392 GlobalVariable::InitialExecTLSModel);
393 VAArgOverflowSizeTLS = new GlobalVariable(
394 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
395 "__msan_va_arg_overflow_size_tls", nullptr,
396 GlobalVariable::InitialExecTLSModel);
397 OriginTLS = new GlobalVariable(
398 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
399 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
401 // We insert an empty inline asm after __msan_report* to avoid callback merge.
402 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
403 StringRef(""), StringRef(""),
404 /*hasSideEffects=*/true);
406 if (WrapIndirectCalls) {
408 PointerType::getUnqual(FunctionType::get(IRB.getVoidTy(), false));
409 IndirectCallWrapperFn = M.getOrInsertFunction(
410 ClWrapIndirectCalls, AnyFunctionPtrTy, AnyFunctionPtrTy, NULL);
413 if (ClWrapIndirectCallsFast) {
414 MsandrModuleStart = new GlobalVariable(
415 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
416 nullptr, "__executable_start");
417 MsandrModuleStart->setVisibility(GlobalVariable::HiddenVisibility);
418 MsandrModuleEnd = new GlobalVariable(
419 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
421 MsandrModuleEnd->setVisibility(GlobalVariable::HiddenVisibility);
425 /// \brief Module-level initialization.
427 /// inserts a call to __msan_init to the module's constructor list.
428 bool MemorySanitizer::doInitialization(Module &M) {
429 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
431 report_fatal_error("data layout missing");
432 DL = &DLP->getDataLayout();
434 BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
435 C = &(M.getContext());
436 unsigned PtrSize = DL->getPointerSizeInBits(/* AddressSpace */0);
439 ShadowMask = kShadowMask64;
440 OriginOffset = kOriginOffset64;
443 ShadowMask = kShadowMask32;
444 OriginOffset = kOriginOffset32;
447 report_fatal_error("unsupported pointer size");
452 IntptrTy = IRB.getIntPtrTy(DL);
453 OriginTy = IRB.getInt32Ty();
455 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
456 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
458 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
459 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
460 "__msan_init", IRB.getVoidTy(), NULL)), 0);
463 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
464 IRB.getInt32(TrackOrigins), "__msan_track_origins");
467 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
468 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
475 /// \brief A helper class that handles instrumentation of VarArg
476 /// functions on a particular platform.
478 /// Implementations are expected to insert the instrumentation
479 /// necessary to propagate argument shadow through VarArg function
480 /// calls. Visit* methods are called during an InstVisitor pass over
481 /// the function, and should avoid creating new basic blocks. A new
482 /// instance of this class is created for each instrumented function.
483 struct VarArgHelper {
484 /// \brief Visit a CallSite.
485 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
487 /// \brief Visit a va_start call.
488 virtual void visitVAStartInst(VAStartInst &I) = 0;
490 /// \brief Visit a va_copy call.
491 virtual void visitVACopyInst(VACopyInst &I) = 0;
493 /// \brief Finalize function instrumentation.
495 /// This method is called after visiting all interesting (see above)
496 /// instructions in a function.
497 virtual void finalizeInstrumentation() = 0;
499 virtual ~VarArgHelper() {}
502 struct MemorySanitizerVisitor;
505 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
506 MemorySanitizerVisitor &Visitor);
508 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
509 if (TypeSize <= 8) return 0;
510 return Log2_32_Ceil(TypeSize / 8);
513 /// This class does all the work for a given function. Store and Load
514 /// instructions store and load corresponding shadow and origin
515 /// values. Most instructions propagate shadow from arguments to their
516 /// return values. Certain instructions (most importantly, BranchInst)
517 /// test their argument shadow and print reports (with a runtime call) if it's
519 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
522 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
523 ValueMap<Value*, Value*> ShadowMap, OriginMap;
524 std::unique_ptr<VarArgHelper> VAHelper;
526 // The following flags disable parts of MSan instrumentation based on
527 // blacklist contents and command-line options.
532 bool CheckReturnValue;
534 struct ShadowOriginAndInsertPoint {
537 Instruction *OrigIns;
538 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
539 : Shadow(S), Origin(O), OrigIns(I) { }
541 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
542 SmallVector<Instruction*, 16> StoreList;
543 SmallVector<CallSite, 16> IndirectCallList;
545 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
546 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
547 bool SanitizeFunction = !MS.BL->isIn(F) && F.getAttributes().hasAttribute(
548 AttributeSet::FunctionIndex,
549 Attribute::SanitizeMemory);
550 InsertChecks = SanitizeFunction;
551 LoadShadow = SanitizeFunction;
552 PoisonStack = SanitizeFunction && ClPoisonStack;
553 PoisonUndef = SanitizeFunction && ClPoisonUndef;
554 // FIXME: Consider using SpecialCaseList to specify a list of functions that
555 // must always return fully initialized values. For now, we hardcode "main".
556 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
558 DEBUG(if (!InsertChecks)
559 dbgs() << "MemorySanitizer is not inserting checks into '"
560 << F.getName() << "'\n");
563 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
564 if (MS.TrackOrigins <= 1) return V;
565 return IRB.CreateCall(MS.MsanChainOriginFn, V);
568 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
569 unsigned Alignment, bool AsCall) {
570 if (isa<StructType>(Shadow->getType())) {
571 IRB.CreateAlignedStore(updateOrigin(Origin, IRB), getOriginPtr(Addr, IRB),
574 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
575 // TODO(eugenis): handle non-zero constant shadow by inserting an
576 // unconditional check (can not simply fail compilation as this could
577 // be in the dead code).
578 if (isa<Constant>(ConvertedShadow)) return;
579 unsigned TypeSizeInBits =
580 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
581 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
582 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
583 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
584 Value *ConvertedShadow2 = IRB.CreateZExt(
585 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
586 IRB.CreateCall3(Fn, ConvertedShadow2,
587 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
588 updateOrigin(Origin, IRB));
590 Value *Cmp = IRB.CreateICmpNE(
591 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
592 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
593 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
594 IRBuilder<> IRBNew(CheckTerm);
595 IRBNew.CreateAlignedStore(updateOrigin(Origin, IRBNew),
596 getOriginPtr(Addr, IRBNew), Alignment);
601 void materializeStores(bool InstrumentWithCalls) {
602 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
603 StoreInst &I = *dyn_cast<StoreInst>(StoreList[i]);
606 Value *Val = I.getValueOperand();
607 Value *Addr = I.getPointerOperand();
608 Value *Shadow = I.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
609 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
612 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
613 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
616 if (ClCheckAccessAddress) insertShadowCheck(Addr, &I);
618 if (I.isAtomic()) I.setOrdering(addReleaseOrdering(I.getOrdering()));
620 if (MS.TrackOrigins) {
621 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
622 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), Alignment,
623 InstrumentWithCalls);
628 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
630 IRBuilder<> IRB(OrigIns);
631 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
632 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
633 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
634 // See the comment in materializeStores().
635 if (isa<Constant>(ConvertedShadow)) return;
636 unsigned TypeSizeInBits =
637 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
638 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
639 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
640 Value *Fn = MS.MaybeWarningFn[SizeIndex];
641 Value *ConvertedShadow2 =
642 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
643 IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
645 : (Value *)IRB.getInt32(0));
647 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
648 getCleanShadow(ConvertedShadow), "_mscmp");
649 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
651 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
653 IRB.SetInsertPoint(CheckTerm);
654 if (MS.TrackOrigins) {
655 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
658 IRB.CreateCall(MS.WarningFn);
659 IRB.CreateCall(MS.EmptyAsm);
660 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
664 void materializeChecks(bool InstrumentWithCalls) {
665 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
666 Instruction *OrigIns = InstrumentationList[i].OrigIns;
667 Value *Shadow = InstrumentationList[i].Shadow;
668 Value *Origin = InstrumentationList[i].Origin;
669 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
671 DEBUG(dbgs() << "DONE:\n" << F);
674 void materializeIndirectCalls() {
675 for (size_t i = 0, n = IndirectCallList.size(); i < n; i++) {
676 CallSite CS = IndirectCallList[i];
677 Instruction *I = CS.getInstruction();
678 BasicBlock *B = I->getParent();
680 Value *Fn0 = CS.getCalledValue();
681 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
683 if (ClWrapIndirectCallsFast) {
684 // Check that call target is inside this module limits.
686 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
687 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
689 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
690 IRB.CreateICmpUGE(Fn, End));
693 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
695 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
696 NotInThisModule, NewFnPhi,
697 /* Unreachable */ false, MS.ColdCallWeights);
699 IRB.SetInsertPoint(CheckTerm);
700 // Slow path: call wrapper function to possibly transform the call
702 Value *NewFn = IRB.CreateBitCast(
703 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
705 NewFnPhi->addIncoming(Fn0, B);
706 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
707 CS.setCalledFunction(NewFnPhi);
709 Value *NewFn = IRB.CreateBitCast(
710 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
711 CS.setCalledFunction(NewFn);
716 /// \brief Add MemorySanitizer instrumentation to a function.
717 bool runOnFunction() {
718 MS.initializeCallbacks(*F.getParent());
719 if (!MS.DL) return false;
721 // In the presence of unreachable blocks, we may see Phi nodes with
722 // incoming nodes from such blocks. Since InstVisitor skips unreachable
723 // blocks, such nodes will not have any shadow value associated with them.
724 // It's easier to remove unreachable blocks than deal with missing shadow.
725 removeUnreachableBlocks(F);
727 // Iterate all BBs in depth-first order and create shadow instructions
728 // for all instructions (where applicable).
729 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
730 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
734 // Finalize PHI nodes.
735 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
736 PHINode *PN = ShadowPHINodes[i];
737 PHINode *PNS = cast<PHINode>(getShadow(PN));
738 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
739 size_t NumValues = PN->getNumIncomingValues();
740 for (size_t v = 0; v < NumValues; v++) {
741 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
743 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
747 VAHelper->finalizeInstrumentation();
749 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
750 InstrumentationList.size() + StoreList.size() >
751 (unsigned)ClInstrumentationWithCallThreshold;
753 // Delayed instrumentation of StoreInst.
754 // This may add new checks to be inserted later.
755 materializeStores(InstrumentWithCalls);
757 // Insert shadow value checks.
758 materializeChecks(InstrumentWithCalls);
760 // Wrap indirect calls.
761 materializeIndirectCalls();
766 /// \brief Compute the shadow type that corresponds to a given Value.
767 Type *getShadowTy(Value *V) {
768 return getShadowTy(V->getType());
771 /// \brief Compute the shadow type that corresponds to a given Type.
772 Type *getShadowTy(Type *OrigTy) {
773 if (!OrigTy->isSized()) {
776 // For integer type, shadow is the same as the original type.
777 // This may return weird-sized types like i1.
778 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
780 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
781 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
782 return VectorType::get(IntegerType::get(*MS.C, EltSize),
783 VT->getNumElements());
785 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
786 SmallVector<Type*, 4> Elements;
787 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
788 Elements.push_back(getShadowTy(ST->getElementType(i)));
789 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
790 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
793 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
794 return IntegerType::get(*MS.C, TypeSize);
797 /// \brief Flatten a vector type.
798 Type *getShadowTyNoVec(Type *ty) {
799 if (VectorType *vt = dyn_cast<VectorType>(ty))
800 return IntegerType::get(*MS.C, vt->getBitWidth());
804 /// \brief Convert a shadow value to it's flattened variant.
805 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
806 Type *Ty = V->getType();
807 Type *NoVecTy = getShadowTyNoVec(Ty);
808 if (Ty == NoVecTy) return V;
809 return IRB.CreateBitCast(V, NoVecTy);
812 /// \brief Compute the shadow address that corresponds to a given application
815 /// Shadow = Addr & ~ShadowMask.
816 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
819 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
820 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
821 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
824 /// \brief Compute the origin address that corresponds to a given application
827 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
828 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
830 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
831 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
833 IRB.CreateAdd(ShadowLong,
834 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
836 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
837 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
840 /// \brief Compute the shadow address for a given function argument.
842 /// Shadow = ParamTLS+ArgOffset.
843 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
845 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
846 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
847 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
851 /// \brief Compute the origin address for a given function argument.
852 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
854 if (!MS.TrackOrigins) return nullptr;
855 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
856 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
857 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
861 /// \brief Compute the shadow address for a retval.
862 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
863 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
864 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
868 /// \brief Compute the origin address for a retval.
869 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
870 // We keep a single origin for the entire retval. Might be too optimistic.
871 return MS.RetvalOriginTLS;
874 /// \brief Set SV to be the shadow value for V.
875 void setShadow(Value *V, Value *SV) {
876 assert(!ShadowMap.count(V) && "Values may only have one shadow");
880 /// \brief Set Origin to be the origin value for V.
881 void setOrigin(Value *V, Value *Origin) {
882 if (!MS.TrackOrigins) return;
883 assert(!OriginMap.count(V) && "Values may only have one origin");
884 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
885 OriginMap[V] = Origin;
888 /// \brief Create a clean shadow value for a given value.
890 /// Clean shadow (all zeroes) means all bits of the value are defined
892 Constant *getCleanShadow(Value *V) {
893 Type *ShadowTy = getShadowTy(V);
896 return Constant::getNullValue(ShadowTy);
899 /// \brief Create a dirty shadow of a given shadow type.
900 Constant *getPoisonedShadow(Type *ShadowTy) {
902 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
903 return Constant::getAllOnesValue(ShadowTy);
904 StructType *ST = cast<StructType>(ShadowTy);
905 SmallVector<Constant *, 4> Vals;
906 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
907 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
908 return ConstantStruct::get(ST, Vals);
911 /// \brief Create a dirty shadow for a given value.
912 Constant *getPoisonedShadow(Value *V) {
913 Type *ShadowTy = getShadowTy(V);
916 return getPoisonedShadow(ShadowTy);
919 /// \brief Create a clean (zero) origin.
920 Value *getCleanOrigin() {
921 return Constant::getNullValue(MS.OriginTy);
924 /// \brief Get the shadow value for a given Value.
926 /// This function either returns the value set earlier with setShadow,
927 /// or extracts if from ParamTLS (for function arguments).
928 Value *getShadow(Value *V) {
929 if (Instruction *I = dyn_cast<Instruction>(V)) {
930 // For instructions the shadow is already stored in the map.
931 Value *Shadow = ShadowMap[V];
933 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
935 assert(Shadow && "No shadow for a value");
939 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
940 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
941 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
945 if (Argument *A = dyn_cast<Argument>(V)) {
946 // For arguments we compute the shadow on demand and store it in the map.
947 Value **ShadowPtr = &ShadowMap[V];
950 Function *F = A->getParent();
951 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
952 unsigned ArgOffset = 0;
953 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
955 if (!AI->getType()->isSized()) {
956 DEBUG(dbgs() << "Arg is not sized\n");
959 unsigned Size = AI->hasByValAttr()
960 ? MS.DL->getTypeAllocSize(AI->getType()->getPointerElementType())
961 : MS.DL->getTypeAllocSize(AI->getType());
963 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
964 if (AI->hasByValAttr()) {
965 // ByVal pointer itself has clean shadow. We copy the actual
966 // argument shadow to the underlying memory.
967 // Figure out maximal valid memcpy alignment.
968 unsigned ArgAlign = AI->getParamAlignment();
970 Type *EltType = A->getType()->getPointerElementType();
971 ArgAlign = MS.DL->getABITypeAlignment(EltType);
973 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
974 Value *Cpy = EntryIRB.CreateMemCpy(
975 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
977 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
979 *ShadowPtr = getCleanShadow(V);
981 *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
983 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
984 **ShadowPtr << "\n");
985 if (MS.TrackOrigins) {
986 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
987 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
990 ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
992 assert(*ShadowPtr && "Could not find shadow for an argument");
995 // For everything else the shadow is zero.
996 return getCleanShadow(V);
999 /// \brief Get the shadow for i-th argument of the instruction I.
1000 Value *getShadow(Instruction *I, int i) {
1001 return getShadow(I->getOperand(i));
1004 /// \brief Get the origin for a value.
1005 Value *getOrigin(Value *V) {
1006 if (!MS.TrackOrigins) return nullptr;
1007 if (isa<Instruction>(V) || isa<Argument>(V)) {
1008 Value *Origin = OriginMap[V];
1010 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
1011 Origin = getCleanOrigin();
1015 return getCleanOrigin();
1018 /// \brief Get the origin for i-th argument of the instruction I.
1019 Value *getOrigin(Instruction *I, int i) {
1020 return getOrigin(I->getOperand(i));
1023 /// \brief Remember the place where a shadow check should be inserted.
1025 /// This location will be later instrumented with a check that will print a
1026 /// UMR warning in runtime if the shadow value is not 0.
1027 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1029 if (!InsertChecks) return;
1031 Type *ShadowTy = Shadow->getType();
1032 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1033 "Can only insert checks for integer and vector shadow types");
1035 InstrumentationList.push_back(
1036 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1039 /// \brief Remember the place where a shadow check should be inserted.
1041 /// This location will be later instrumented with a check that will print a
1042 /// UMR warning in runtime if the value is not fully defined.
1043 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1045 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1046 if (!Shadow) return;
1047 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1048 insertShadowCheck(Shadow, Origin, OrigIns);
1051 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1060 case AcquireRelease:
1061 return AcquireRelease;
1062 case SequentiallyConsistent:
1063 return SequentiallyConsistent;
1065 llvm_unreachable("Unknown ordering");
1068 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1077 case AcquireRelease:
1078 return AcquireRelease;
1079 case SequentiallyConsistent:
1080 return SequentiallyConsistent;
1082 llvm_unreachable("Unknown ordering");
1085 // ------------------- Visitors.
1087 /// \brief Instrument LoadInst
1089 /// Loads the corresponding shadow and (optionally) origin.
1090 /// Optionally, checks that the load address is fully defined.
1091 void visitLoadInst(LoadInst &I) {
1092 assert(I.getType()->isSized() && "Load type must have size");
1093 IRBuilder<> IRB(I.getNextNode());
1094 Type *ShadowTy = getShadowTy(&I);
1095 Value *Addr = I.getPointerOperand();
1097 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1099 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1101 setShadow(&I, getCleanShadow(&I));
1104 if (ClCheckAccessAddress)
1105 insertShadowCheck(I.getPointerOperand(), &I);
1108 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1110 if (MS.TrackOrigins) {
1112 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1114 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1116 setOrigin(&I, getCleanOrigin());
1121 /// \brief Instrument StoreInst
1123 /// Stores the corresponding shadow and (optionally) origin.
1124 /// Optionally, checks that the store address is fully defined.
1125 void visitStoreInst(StoreInst &I) {
1126 StoreList.push_back(&I);
1129 void handleCASOrRMW(Instruction &I) {
1130 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1132 IRBuilder<> IRB(&I);
1133 Value *Addr = I.getOperand(0);
1134 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1136 if (ClCheckAccessAddress)
1137 insertShadowCheck(Addr, &I);
1139 // Only test the conditional argument of cmpxchg instruction.
1140 // The other argument can potentially be uninitialized, but we can not
1141 // detect this situation reliably without possible false positives.
1142 if (isa<AtomicCmpXchgInst>(I))
1143 insertShadowCheck(I.getOperand(1), &I);
1145 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1147 setShadow(&I, getCleanShadow(&I));
1150 void visitAtomicRMWInst(AtomicRMWInst &I) {
1152 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1155 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1157 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1160 // Vector manipulation.
1161 void visitExtractElementInst(ExtractElementInst &I) {
1162 insertShadowCheck(I.getOperand(1), &I);
1163 IRBuilder<> IRB(&I);
1164 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1166 setOrigin(&I, getOrigin(&I, 0));
1169 void visitInsertElementInst(InsertElementInst &I) {
1170 insertShadowCheck(I.getOperand(2), &I);
1171 IRBuilder<> IRB(&I);
1172 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1173 I.getOperand(2), "_msprop"));
1174 setOriginForNaryOp(I);
1177 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1178 insertShadowCheck(I.getOperand(2), &I);
1179 IRBuilder<> IRB(&I);
1180 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1181 I.getOperand(2), "_msprop"));
1182 setOriginForNaryOp(I);
1186 void visitSExtInst(SExtInst &I) {
1187 IRBuilder<> IRB(&I);
1188 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1189 setOrigin(&I, getOrigin(&I, 0));
1192 void visitZExtInst(ZExtInst &I) {
1193 IRBuilder<> IRB(&I);
1194 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1195 setOrigin(&I, getOrigin(&I, 0));
1198 void visitTruncInst(TruncInst &I) {
1199 IRBuilder<> IRB(&I);
1200 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1201 setOrigin(&I, getOrigin(&I, 0));
1204 void visitBitCastInst(BitCastInst &I) {
1205 IRBuilder<> IRB(&I);
1206 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1207 setOrigin(&I, getOrigin(&I, 0));
1210 void visitPtrToIntInst(PtrToIntInst &I) {
1211 IRBuilder<> IRB(&I);
1212 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1213 "_msprop_ptrtoint"));
1214 setOrigin(&I, getOrigin(&I, 0));
1217 void visitIntToPtrInst(IntToPtrInst &I) {
1218 IRBuilder<> IRB(&I);
1219 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1220 "_msprop_inttoptr"));
1221 setOrigin(&I, getOrigin(&I, 0));
1224 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1225 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1226 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1227 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1228 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1229 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1231 /// \brief Propagate shadow for bitwise AND.
1233 /// This code is exact, i.e. if, for example, a bit in the left argument
1234 /// is defined and 0, then neither the value not definedness of the
1235 /// corresponding bit in B don't affect the resulting shadow.
1236 void visitAnd(BinaryOperator &I) {
1237 IRBuilder<> IRB(&I);
1238 // "And" of 0 and a poisoned value results in unpoisoned value.
1239 // 1&1 => 1; 0&1 => 0; p&1 => p;
1240 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1241 // 1&p => p; 0&p => 0; p&p => p;
1242 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1243 Value *S1 = getShadow(&I, 0);
1244 Value *S2 = getShadow(&I, 1);
1245 Value *V1 = I.getOperand(0);
1246 Value *V2 = I.getOperand(1);
1247 if (V1->getType() != S1->getType()) {
1248 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1249 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1251 Value *S1S2 = IRB.CreateAnd(S1, S2);
1252 Value *V1S2 = IRB.CreateAnd(V1, S2);
1253 Value *S1V2 = IRB.CreateAnd(S1, V2);
1254 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1255 setOriginForNaryOp(I);
1258 void visitOr(BinaryOperator &I) {
1259 IRBuilder<> IRB(&I);
1260 // "Or" of 1 and a poisoned value results in unpoisoned value.
1261 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1262 // 1|0 => 1; 0|0 => 0; p|0 => p;
1263 // 1|p => 1; 0|p => p; p|p => p;
1264 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1265 Value *S1 = getShadow(&I, 0);
1266 Value *S2 = getShadow(&I, 1);
1267 Value *V1 = IRB.CreateNot(I.getOperand(0));
1268 Value *V2 = IRB.CreateNot(I.getOperand(1));
1269 if (V1->getType() != S1->getType()) {
1270 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1271 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1273 Value *S1S2 = IRB.CreateAnd(S1, S2);
1274 Value *V1S2 = IRB.CreateAnd(V1, S2);
1275 Value *S1V2 = IRB.CreateAnd(S1, V2);
1276 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1277 setOriginForNaryOp(I);
1280 /// \brief Default propagation of shadow and/or origin.
1282 /// This class implements the general case of shadow propagation, used in all
1283 /// cases where we don't know and/or don't care about what the operation
1284 /// actually does. It converts all input shadow values to a common type
1285 /// (extending or truncating as necessary), and bitwise OR's them.
1287 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1288 /// fully initialized), and less prone to false positives.
1290 /// This class also implements the general case of origin propagation. For a
1291 /// Nary operation, result origin is set to the origin of an argument that is
1292 /// not entirely initialized. If there is more than one such arguments, the
1293 /// rightmost of them is picked. It does not matter which one is picked if all
1294 /// arguments are initialized.
1295 template <bool CombineShadow>
1300 MemorySanitizerVisitor *MSV;
1303 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1304 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1306 /// \brief Add a pair of shadow and origin values to the mix.
1307 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1308 if (CombineShadow) {
1313 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1314 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1318 if (MSV->MS.TrackOrigins) {
1323 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1324 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1325 MSV->getCleanShadow(FlatShadow));
1326 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1332 /// \brief Add an application value to the mix.
1333 Combiner &Add(Value *V) {
1334 Value *OpShadow = MSV->getShadow(V);
1335 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1336 return Add(OpShadow, OpOrigin);
1339 /// \brief Set the current combined values as the given instruction's shadow
1341 void Done(Instruction *I) {
1342 if (CombineShadow) {
1344 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1345 MSV->setShadow(I, Shadow);
1347 if (MSV->MS.TrackOrigins) {
1349 MSV->setOrigin(I, Origin);
1354 typedef Combiner<true> ShadowAndOriginCombiner;
1355 typedef Combiner<false> OriginCombiner;
1357 /// \brief Propagate origin for arbitrary operation.
1358 void setOriginForNaryOp(Instruction &I) {
1359 if (!MS.TrackOrigins) return;
1360 IRBuilder<> IRB(&I);
1361 OriginCombiner OC(this, IRB);
1362 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1367 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1368 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1369 "Vector of pointers is not a valid shadow type");
1370 return Ty->isVectorTy() ?
1371 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1372 Ty->getPrimitiveSizeInBits();
1375 /// \brief Cast between two shadow types, extending or truncating as
1377 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1378 bool Signed = false) {
1379 Type *srcTy = V->getType();
1380 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1381 return IRB.CreateIntCast(V, dstTy, Signed);
1382 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1383 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1384 return IRB.CreateIntCast(V, dstTy, Signed);
1385 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1386 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1387 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1389 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1390 return IRB.CreateBitCast(V2, dstTy);
1391 // TODO: handle struct types.
1394 /// \brief Cast an application value to the type of its own shadow.
1395 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1396 Type *ShadowTy = getShadowTy(V);
1397 if (V->getType() == ShadowTy)
1399 if (V->getType()->isPtrOrPtrVectorTy())
1400 return IRB.CreatePtrToInt(V, ShadowTy);
1402 return IRB.CreateBitCast(V, ShadowTy);
1405 /// \brief Propagate shadow for arbitrary operation.
1406 void handleShadowOr(Instruction &I) {
1407 IRBuilder<> IRB(&I);
1408 ShadowAndOriginCombiner SC(this, IRB);
1409 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1414 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1415 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1416 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1417 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1418 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1419 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1420 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1422 void handleDiv(Instruction &I) {
1423 IRBuilder<> IRB(&I);
1424 // Strict on the second argument.
1425 insertShadowCheck(I.getOperand(1), &I);
1426 setShadow(&I, getShadow(&I, 0));
1427 setOrigin(&I, getOrigin(&I, 0));
1430 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1431 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1432 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1433 void visitURem(BinaryOperator &I) { handleDiv(I); }
1434 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1435 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1437 /// \brief Instrument == and != comparisons.
1439 /// Sometimes the comparison result is known even if some of the bits of the
1440 /// arguments are not.
1441 void handleEqualityComparison(ICmpInst &I) {
1442 IRBuilder<> IRB(&I);
1443 Value *A = I.getOperand(0);
1444 Value *B = I.getOperand(1);
1445 Value *Sa = getShadow(A);
1446 Value *Sb = getShadow(B);
1448 // Get rid of pointers and vectors of pointers.
1449 // For ints (and vectors of ints), types of A and Sa match,
1450 // and this is a no-op.
1451 A = IRB.CreatePointerCast(A, Sa->getType());
1452 B = IRB.CreatePointerCast(B, Sb->getType());
1454 // A == B <==> (C = A^B) == 0
1455 // A != B <==> (C = A^B) != 0
1457 Value *C = IRB.CreateXor(A, B);
1458 Value *Sc = IRB.CreateOr(Sa, Sb);
1459 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1460 // Result is defined if one of the following is true
1461 // * there is a defined 1 bit in C
1462 // * C is fully defined
1463 // Si = !(C & ~Sc) && Sc
1464 Value *Zero = Constant::getNullValue(Sc->getType());
1465 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1467 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1469 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1470 Si->setName("_msprop_icmp");
1472 setOriginForNaryOp(I);
1475 /// \brief Build the lowest possible value of V, taking into account V's
1476 /// uninitialized bits.
1477 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1480 // Split shadow into sign bit and other bits.
1481 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1482 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1483 // Maximise the undefined shadow bit, minimize other undefined bits.
1485 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1487 // Minimize undefined bits.
1488 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1492 /// \brief Build the highest possible value of V, taking into account V's
1493 /// uninitialized bits.
1494 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1497 // Split shadow into sign bit and other bits.
1498 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1499 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1500 // Minimise the undefined shadow bit, maximise other undefined bits.
1502 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1504 // Maximize undefined bits.
1505 return IRB.CreateOr(A, Sa);
1509 /// \brief Instrument relational comparisons.
1511 /// This function does exact shadow propagation for all relational
1512 /// comparisons of integers, pointers and vectors of those.
1513 /// FIXME: output seems suboptimal when one of the operands is a constant
1514 void handleRelationalComparisonExact(ICmpInst &I) {
1515 IRBuilder<> IRB(&I);
1516 Value *A = I.getOperand(0);
1517 Value *B = I.getOperand(1);
1518 Value *Sa = getShadow(A);
1519 Value *Sb = getShadow(B);
1521 // Get rid of pointers and vectors of pointers.
1522 // For ints (and vectors of ints), types of A and Sa match,
1523 // and this is a no-op.
1524 A = IRB.CreatePointerCast(A, Sa->getType());
1525 B = IRB.CreatePointerCast(B, Sb->getType());
1527 // Let [a0, a1] be the interval of possible values of A, taking into account
1528 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1529 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1530 bool IsSigned = I.isSigned();
1531 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1532 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1533 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1534 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1535 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1536 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1537 Value *Si = IRB.CreateXor(S1, S2);
1539 setOriginForNaryOp(I);
1542 /// \brief Instrument signed relational comparisons.
1544 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1545 /// propagating the highest bit of the shadow. Everything else is delegated
1546 /// to handleShadowOr().
1547 void handleSignedRelationalComparison(ICmpInst &I) {
1548 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1549 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1550 Value* op = nullptr;
1551 CmpInst::Predicate pre = I.getPredicate();
1552 if (constOp0 && constOp0->isNullValue() &&
1553 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1554 op = I.getOperand(1);
1555 } else if (constOp1 && constOp1->isNullValue() &&
1556 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1557 op = I.getOperand(0);
1560 IRBuilder<> IRB(&I);
1562 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1563 setShadow(&I, Shadow);
1564 setOrigin(&I, getOrigin(op));
1570 void visitICmpInst(ICmpInst &I) {
1571 if (!ClHandleICmp) {
1575 if (I.isEquality()) {
1576 handleEqualityComparison(I);
1580 assert(I.isRelational());
1581 if (ClHandleICmpExact) {
1582 handleRelationalComparisonExact(I);
1586 handleSignedRelationalComparison(I);
1590 assert(I.isUnsigned());
1591 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1592 handleRelationalComparisonExact(I);
1599 void visitFCmpInst(FCmpInst &I) {
1603 void handleShift(BinaryOperator &I) {
1604 IRBuilder<> IRB(&I);
1605 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1606 // Otherwise perform the same shift on S1.
1607 Value *S1 = getShadow(&I, 0);
1608 Value *S2 = getShadow(&I, 1);
1609 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1611 Value *V2 = I.getOperand(1);
1612 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1613 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1614 setOriginForNaryOp(I);
1617 void visitShl(BinaryOperator &I) { handleShift(I); }
1618 void visitAShr(BinaryOperator &I) { handleShift(I); }
1619 void visitLShr(BinaryOperator &I) { handleShift(I); }
1621 /// \brief Instrument llvm.memmove
1623 /// At this point we don't know if llvm.memmove will be inlined or not.
1624 /// If we don't instrument it and it gets inlined,
1625 /// our interceptor will not kick in and we will lose the memmove.
1626 /// If we instrument the call here, but it does not get inlined,
1627 /// we will memove the shadow twice: which is bad in case
1628 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1630 /// Similar situation exists for memcpy and memset.
1631 void visitMemMoveInst(MemMoveInst &I) {
1632 IRBuilder<> IRB(&I);
1635 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1636 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1637 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1638 I.eraseFromParent();
1641 // Similar to memmove: avoid copying shadow twice.
1642 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1643 // FIXME: consider doing manual inline for small constant sizes and proper
1645 void visitMemCpyInst(MemCpyInst &I) {
1646 IRBuilder<> IRB(&I);
1649 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1650 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1651 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1652 I.eraseFromParent();
1656 void visitMemSetInst(MemSetInst &I) {
1657 IRBuilder<> IRB(&I);
1660 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1661 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1662 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1663 I.eraseFromParent();
1666 void visitVAStartInst(VAStartInst &I) {
1667 VAHelper->visitVAStartInst(I);
1670 void visitVACopyInst(VACopyInst &I) {
1671 VAHelper->visitVACopyInst(I);
1674 enum IntrinsicKind {
1675 IK_DoesNotAccessMemory,
1680 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1681 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1682 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1683 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1684 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1685 const int UnknownModRefBehavior = IK_WritesMemory;
1686 #define GET_INTRINSIC_MODREF_BEHAVIOR
1687 #define ModRefBehavior IntrinsicKind
1688 #include "llvm/IR/Intrinsics.gen"
1689 #undef ModRefBehavior
1690 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1693 /// \brief Handle vector store-like intrinsics.
1695 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1696 /// has 1 pointer argument and 1 vector argument, returns void.
1697 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1698 IRBuilder<> IRB(&I);
1699 Value* Addr = I.getArgOperand(0);
1700 Value *Shadow = getShadow(&I, 1);
1701 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1703 // We don't know the pointer alignment (could be unaligned SSE store!).
1704 // Have to assume to worst case.
1705 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1707 if (ClCheckAccessAddress)
1708 insertShadowCheck(Addr, &I);
1710 // FIXME: use ClStoreCleanOrigin
1711 // FIXME: factor out common code from materializeStores
1712 if (MS.TrackOrigins)
1713 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1717 /// \brief Handle vector load-like intrinsics.
1719 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1720 /// has 1 pointer argument, returns a vector.
1721 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1722 IRBuilder<> IRB(&I);
1723 Value *Addr = I.getArgOperand(0);
1725 Type *ShadowTy = getShadowTy(&I);
1727 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1728 // We don't know the pointer alignment (could be unaligned SSE load!).
1729 // Have to assume to worst case.
1730 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1732 setShadow(&I, getCleanShadow(&I));
1735 if (ClCheckAccessAddress)
1736 insertShadowCheck(Addr, &I);
1738 if (MS.TrackOrigins) {
1740 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1742 setOrigin(&I, getCleanOrigin());
1747 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1749 /// Instrument intrinsics with any number of arguments of the same type,
1750 /// equal to the return type. The type should be simple (no aggregates or
1751 /// pointers; vectors are fine).
1752 /// Caller guarantees that this intrinsic does not access memory.
1753 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1754 Type *RetTy = I.getType();
1755 if (!(RetTy->isIntOrIntVectorTy() ||
1756 RetTy->isFPOrFPVectorTy() ||
1757 RetTy->isX86_MMXTy()))
1760 unsigned NumArgOperands = I.getNumArgOperands();
1762 for (unsigned i = 0; i < NumArgOperands; ++i) {
1763 Type *Ty = I.getArgOperand(i)->getType();
1768 IRBuilder<> IRB(&I);
1769 ShadowAndOriginCombiner SC(this, IRB);
1770 for (unsigned i = 0; i < NumArgOperands; ++i)
1771 SC.Add(I.getArgOperand(i));
1777 /// \brief Heuristically instrument unknown intrinsics.
1779 /// The main purpose of this code is to do something reasonable with all
1780 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1781 /// We recognize several classes of intrinsics by their argument types and
1782 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1783 /// sure that we know what the intrinsic does.
1785 /// We special-case intrinsics where this approach fails. See llvm.bswap
1786 /// handling as an example of that.
1787 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1788 unsigned NumArgOperands = I.getNumArgOperands();
1789 if (NumArgOperands == 0)
1792 Intrinsic::ID iid = I.getIntrinsicID();
1793 IntrinsicKind IK = getIntrinsicKind(iid);
1794 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1795 bool WritesMemory = IK == IK_WritesMemory;
1796 assert(!(OnlyReadsMemory && WritesMemory));
1798 if (NumArgOperands == 2 &&
1799 I.getArgOperand(0)->getType()->isPointerTy() &&
1800 I.getArgOperand(1)->getType()->isVectorTy() &&
1801 I.getType()->isVoidTy() &&
1803 // This looks like a vector store.
1804 return handleVectorStoreIntrinsic(I);
1807 if (NumArgOperands == 1 &&
1808 I.getArgOperand(0)->getType()->isPointerTy() &&
1809 I.getType()->isVectorTy() &&
1811 // This looks like a vector load.
1812 return handleVectorLoadIntrinsic(I);
1815 if (!OnlyReadsMemory && !WritesMemory)
1816 if (maybeHandleSimpleNomemIntrinsic(I))
1819 // FIXME: detect and handle SSE maskstore/maskload
1823 void handleBswap(IntrinsicInst &I) {
1824 IRBuilder<> IRB(&I);
1825 Value *Op = I.getArgOperand(0);
1826 Type *OpType = Op->getType();
1827 Function *BswapFunc = Intrinsic::getDeclaration(
1828 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1829 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1830 setOrigin(&I, getOrigin(Op));
1833 // \brief Instrument vector convert instrinsic.
1835 // This function instruments intrinsics like cvtsi2ss:
1836 // %Out = int_xxx_cvtyyy(%ConvertOp)
1838 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1839 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1840 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1841 // elements from \p CopyOp.
1842 // In most cases conversion involves floating-point value which may trigger a
1843 // hardware exception when not fully initialized. For this reason we require
1844 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1845 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1846 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1847 // return a fully initialized value.
1848 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1849 IRBuilder<> IRB(&I);
1850 Value *CopyOp, *ConvertOp;
1852 switch (I.getNumArgOperands()) {
1854 CopyOp = I.getArgOperand(0);
1855 ConvertOp = I.getArgOperand(1);
1858 ConvertOp = I.getArgOperand(0);
1862 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1865 // The first *NumUsedElements* elements of ConvertOp are converted to the
1866 // same number of output elements. The rest of the output is copied from
1867 // CopyOp, or (if not available) filled with zeroes.
1868 // Combine shadow for elements of ConvertOp that are used in this operation,
1869 // and insert a check.
1870 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1871 // int->any conversion.
1872 Value *ConvertShadow = getShadow(ConvertOp);
1873 Value *AggShadow = nullptr;
1874 if (ConvertOp->getType()->isVectorTy()) {
1875 AggShadow = IRB.CreateExtractElement(
1876 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1877 for (int i = 1; i < NumUsedElements; ++i) {
1878 Value *MoreShadow = IRB.CreateExtractElement(
1879 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1880 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1883 AggShadow = ConvertShadow;
1885 assert(AggShadow->getType()->isIntegerTy());
1886 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1888 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1891 assert(CopyOp->getType() == I.getType());
1892 assert(CopyOp->getType()->isVectorTy());
1893 Value *ResultShadow = getShadow(CopyOp);
1894 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1895 for (int i = 0; i < NumUsedElements; ++i) {
1896 ResultShadow = IRB.CreateInsertElement(
1897 ResultShadow, ConstantInt::getNullValue(EltTy),
1898 ConstantInt::get(IRB.getInt32Ty(), i));
1900 setShadow(&I, ResultShadow);
1901 setOrigin(&I, getOrigin(CopyOp));
1903 setShadow(&I, getCleanShadow(&I));
1907 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1908 // zeroes if it is zero, and all ones otherwise.
1909 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1910 if (S->getType()->isVectorTy())
1911 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1912 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1913 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1914 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1917 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1918 Type *T = S->getType();
1919 assert(T->isVectorTy());
1920 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1921 return IRB.CreateSExt(S2, T);
1924 // \brief Instrument vector shift instrinsic.
1926 // This function instruments intrinsics like int_x86_avx2_psll_w.
1927 // Intrinsic shifts %In by %ShiftSize bits.
1928 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
1929 // size, and the rest is ignored. Behavior is defined even if shift size is
1930 // greater than register (or field) width.
1931 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
1932 assert(I.getNumArgOperands() == 2);
1933 IRBuilder<> IRB(&I);
1934 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1935 // Otherwise perform the same shift on S1.
1936 Value *S1 = getShadow(&I, 0);
1937 Value *S2 = getShadow(&I, 1);
1938 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
1939 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
1940 Value *V1 = I.getOperand(0);
1941 Value *V2 = I.getOperand(1);
1942 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
1943 IRB.CreateBitCast(S1, V1->getType()), V2);
1944 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
1945 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1946 setOriginForNaryOp(I);
1949 void visitIntrinsicInst(IntrinsicInst &I) {
1950 switch (I.getIntrinsicID()) {
1951 case llvm::Intrinsic::bswap:
1954 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
1955 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
1956 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
1957 case llvm::Intrinsic::x86_avx512_cvtss2usi:
1958 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
1959 case llvm::Intrinsic::x86_avx512_cvttss2usi:
1960 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
1961 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
1962 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
1963 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
1964 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
1965 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
1966 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
1967 case llvm::Intrinsic::x86_sse2_cvtsd2si:
1968 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
1969 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
1970 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
1971 case llvm::Intrinsic::x86_sse2_cvtss2sd:
1972 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
1973 case llvm::Intrinsic::x86_sse2_cvttsd2si:
1974 case llvm::Intrinsic::x86_sse_cvtsi2ss:
1975 case llvm::Intrinsic::x86_sse_cvtsi642ss:
1976 case llvm::Intrinsic::x86_sse_cvtss2si64:
1977 case llvm::Intrinsic::x86_sse_cvtss2si:
1978 case llvm::Intrinsic::x86_sse_cvttss2si64:
1979 case llvm::Intrinsic::x86_sse_cvttss2si:
1980 handleVectorConvertIntrinsic(I, 1);
1982 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
1983 case llvm::Intrinsic::x86_sse2_cvtps2pd:
1984 case llvm::Intrinsic::x86_sse_cvtps2pi:
1985 case llvm::Intrinsic::x86_sse_cvttps2pi:
1986 handleVectorConvertIntrinsic(I, 2);
1988 case llvm::Intrinsic::x86_avx512_psll_dq:
1989 case llvm::Intrinsic::x86_avx512_psrl_dq:
1990 case llvm::Intrinsic::x86_avx2_psll_w:
1991 case llvm::Intrinsic::x86_avx2_psll_d:
1992 case llvm::Intrinsic::x86_avx2_psll_q:
1993 case llvm::Intrinsic::x86_avx2_pslli_w:
1994 case llvm::Intrinsic::x86_avx2_pslli_d:
1995 case llvm::Intrinsic::x86_avx2_pslli_q:
1996 case llvm::Intrinsic::x86_avx2_psll_dq:
1997 case llvm::Intrinsic::x86_avx2_psrl_w:
1998 case llvm::Intrinsic::x86_avx2_psrl_d:
1999 case llvm::Intrinsic::x86_avx2_psrl_q:
2000 case llvm::Intrinsic::x86_avx2_psra_w:
2001 case llvm::Intrinsic::x86_avx2_psra_d:
2002 case llvm::Intrinsic::x86_avx2_psrli_w:
2003 case llvm::Intrinsic::x86_avx2_psrli_d:
2004 case llvm::Intrinsic::x86_avx2_psrli_q:
2005 case llvm::Intrinsic::x86_avx2_psrai_w:
2006 case llvm::Intrinsic::x86_avx2_psrai_d:
2007 case llvm::Intrinsic::x86_avx2_psrl_dq:
2008 case llvm::Intrinsic::x86_sse2_psll_w:
2009 case llvm::Intrinsic::x86_sse2_psll_d:
2010 case llvm::Intrinsic::x86_sse2_psll_q:
2011 case llvm::Intrinsic::x86_sse2_pslli_w:
2012 case llvm::Intrinsic::x86_sse2_pslli_d:
2013 case llvm::Intrinsic::x86_sse2_pslli_q:
2014 case llvm::Intrinsic::x86_sse2_psll_dq:
2015 case llvm::Intrinsic::x86_sse2_psrl_w:
2016 case llvm::Intrinsic::x86_sse2_psrl_d:
2017 case llvm::Intrinsic::x86_sse2_psrl_q:
2018 case llvm::Intrinsic::x86_sse2_psra_w:
2019 case llvm::Intrinsic::x86_sse2_psra_d:
2020 case llvm::Intrinsic::x86_sse2_psrli_w:
2021 case llvm::Intrinsic::x86_sse2_psrli_d:
2022 case llvm::Intrinsic::x86_sse2_psrli_q:
2023 case llvm::Intrinsic::x86_sse2_psrai_w:
2024 case llvm::Intrinsic::x86_sse2_psrai_d:
2025 case llvm::Intrinsic::x86_sse2_psrl_dq:
2026 case llvm::Intrinsic::x86_mmx_psll_w:
2027 case llvm::Intrinsic::x86_mmx_psll_d:
2028 case llvm::Intrinsic::x86_mmx_psll_q:
2029 case llvm::Intrinsic::x86_mmx_pslli_w:
2030 case llvm::Intrinsic::x86_mmx_pslli_d:
2031 case llvm::Intrinsic::x86_mmx_pslli_q:
2032 case llvm::Intrinsic::x86_mmx_psrl_w:
2033 case llvm::Intrinsic::x86_mmx_psrl_d:
2034 case llvm::Intrinsic::x86_mmx_psrl_q:
2035 case llvm::Intrinsic::x86_mmx_psra_w:
2036 case llvm::Intrinsic::x86_mmx_psra_d:
2037 case llvm::Intrinsic::x86_mmx_psrli_w:
2038 case llvm::Intrinsic::x86_mmx_psrli_d:
2039 case llvm::Intrinsic::x86_mmx_psrli_q:
2040 case llvm::Intrinsic::x86_mmx_psrai_w:
2041 case llvm::Intrinsic::x86_mmx_psrai_d:
2042 handleVectorShiftIntrinsic(I, /* Variable */ false);
2044 case llvm::Intrinsic::x86_avx2_psllv_d:
2045 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2046 case llvm::Intrinsic::x86_avx2_psllv_q:
2047 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2048 case llvm::Intrinsic::x86_avx2_psrlv_d:
2049 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2050 case llvm::Intrinsic::x86_avx2_psrlv_q:
2051 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2052 case llvm::Intrinsic::x86_avx2_psrav_d:
2053 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2054 handleVectorShiftIntrinsic(I, /* Variable */ true);
2057 // Byte shifts are not implemented.
2058 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
2059 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
2060 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
2061 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
2062 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
2063 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
2066 if (!handleUnknownIntrinsic(I))
2067 visitInstruction(I);
2072 void visitCallSite(CallSite CS) {
2073 Instruction &I = *CS.getInstruction();
2074 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2076 CallInst *Call = cast<CallInst>(&I);
2078 // For inline asm, do the usual thing: check argument shadow and mark all
2079 // outputs as clean. Note that any side effects of the inline asm that are
2080 // not immediately visible in its constraints are not handled.
2081 if (Call->isInlineAsm()) {
2082 visitInstruction(I);
2086 // Allow only tail calls with the same types, otherwise
2087 // we may have a false positive: shadow for a non-void RetVal
2088 // will get propagated to a void RetVal.
2089 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
2090 Call->setTailCall(false);
2092 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2094 // We are going to insert code that relies on the fact that the callee
2095 // will become a non-readonly function after it is instrumented by us. To
2096 // prevent this code from being optimized out, mark that function
2097 // non-readonly in advance.
2098 if (Function *Func = Call->getCalledFunction()) {
2099 // Clear out readonly/readnone attributes.
2101 B.addAttribute(Attribute::ReadOnly)
2102 .addAttribute(Attribute::ReadNone);
2103 Func->removeAttributes(AttributeSet::FunctionIndex,
2104 AttributeSet::get(Func->getContext(),
2105 AttributeSet::FunctionIndex,
2109 IRBuilder<> IRB(&I);
2111 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2112 IndirectCallList.push_back(CS);
2114 unsigned ArgOffset = 0;
2115 DEBUG(dbgs() << " CallSite: " << I << "\n");
2116 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2117 ArgIt != End; ++ArgIt) {
2119 unsigned i = ArgIt - CS.arg_begin();
2120 if (!A->getType()->isSized()) {
2121 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2125 Value *Store = nullptr;
2126 // Compute the Shadow for arg even if it is ByVal, because
2127 // in that case getShadow() will copy the actual arg shadow to
2128 // __msan_param_tls.
2129 Value *ArgShadow = getShadow(A);
2130 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2131 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2132 " Shadow: " << *ArgShadow << "\n");
2133 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2134 assert(A->getType()->isPointerTy() &&
2135 "ByVal argument is not a pointer!");
2136 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2137 unsigned Alignment = CS.getParamAlignment(i + 1);
2138 Store = IRB.CreateMemCpy(ArgShadowBase,
2139 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2142 Size = MS.DL->getTypeAllocSize(A->getType());
2143 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2144 kShadowTLSAlignment);
2146 if (MS.TrackOrigins)
2147 IRB.CreateStore(getOrigin(A),
2148 getOriginPtrForArgument(A, IRB, ArgOffset));
2150 assert(Size != 0 && Store != 0);
2151 DEBUG(dbgs() << " Param:" << *Store << "\n");
2152 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
2154 DEBUG(dbgs() << " done with call args\n");
2157 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2158 if (FT->isVarArg()) {
2159 VAHelper->visitCallSite(CS, IRB);
2162 // Now, get the shadow for the RetVal.
2163 if (!I.getType()->isSized()) return;
2164 IRBuilder<> IRBBefore(&I);
2165 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2166 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2167 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2168 Instruction *NextInsn = nullptr;
2170 NextInsn = I.getNextNode();
2172 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2173 if (!NormalDest->getSinglePredecessor()) {
2174 // FIXME: this case is tricky, so we are just conservative here.
2175 // Perhaps we need to split the edge between this BB and NormalDest,
2176 // but a naive attempt to use SplitEdge leads to a crash.
2177 setShadow(&I, getCleanShadow(&I));
2178 setOrigin(&I, getCleanOrigin());
2181 NextInsn = NormalDest->getFirstInsertionPt();
2183 "Could not find insertion point for retval shadow load");
2185 IRBuilder<> IRBAfter(NextInsn);
2186 Value *RetvalShadow =
2187 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2188 kShadowTLSAlignment, "_msret");
2189 setShadow(&I, RetvalShadow);
2190 if (MS.TrackOrigins)
2191 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2194 void visitReturnInst(ReturnInst &I) {
2195 IRBuilder<> IRB(&I);
2196 Value *RetVal = I.getReturnValue();
2197 if (!RetVal) return;
2198 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2199 if (CheckReturnValue) {
2200 insertShadowCheck(RetVal, &I);
2201 Value *Shadow = getCleanShadow(RetVal);
2202 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2204 Value *Shadow = getShadow(RetVal);
2205 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2206 // FIXME: make it conditional if ClStoreCleanOrigin==0
2207 if (MS.TrackOrigins)
2208 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2212 void visitPHINode(PHINode &I) {
2213 IRBuilder<> IRB(&I);
2214 ShadowPHINodes.push_back(&I);
2215 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2217 if (MS.TrackOrigins)
2218 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2222 void visitAllocaInst(AllocaInst &I) {
2223 setShadow(&I, getCleanShadow(&I));
2224 IRBuilder<> IRB(I.getNextNode());
2225 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2226 if (PoisonStack && ClPoisonStackWithCall) {
2227 IRB.CreateCall2(MS.MsanPoisonStackFn,
2228 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2229 ConstantInt::get(MS.IntptrTy, Size));
2231 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2232 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2233 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2236 if (PoisonStack && MS.TrackOrigins) {
2237 setOrigin(&I, getCleanOrigin());
2238 SmallString<2048> StackDescriptionStorage;
2239 raw_svector_ostream StackDescription(StackDescriptionStorage);
2240 // We create a string with a description of the stack allocation and
2241 // pass it into __msan_set_alloca_origin.
2242 // It will be printed by the run-time if stack-originated UMR is found.
2243 // The first 4 bytes of the string are set to '----' and will be replaced
2244 // by __msan_va_arg_overflow_size_tls at the first call.
2245 StackDescription << "----" << I.getName() << "@" << F.getName();
2247 createPrivateNonConstGlobalForString(*F.getParent(),
2248 StackDescription.str());
2250 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2251 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2252 ConstantInt::get(MS.IntptrTy, Size),
2253 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2254 IRB.CreatePointerCast(&F, MS.IntptrTy));
2258 void visitSelectInst(SelectInst& I) {
2259 IRBuilder<> IRB(&I);
2260 // a = select b, c, d
2261 Value *B = I.getCondition();
2262 Value *C = I.getTrueValue();
2263 Value *D = I.getFalseValue();
2264 Value *Sb = getShadow(B);
2265 Value *Sc = getShadow(C);
2266 Value *Sd = getShadow(D);
2268 // Result shadow if condition shadow is 0.
2269 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2271 if (I.getType()->isAggregateType()) {
2272 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2273 // an extra "select". This results in much more compact IR.
2274 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2275 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2277 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2278 // If Sb (condition is poisoned), look for bits in c and d that are equal
2279 // and both unpoisoned.
2280 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2282 // Cast arguments to shadow-compatible type.
2283 C = CreateAppToShadowCast(IRB, C);
2284 D = CreateAppToShadowCast(IRB, D);
2286 // Result shadow if condition shadow is 1.
2287 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2289 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2291 if (MS.TrackOrigins) {
2292 // Origins are always i32, so any vector conditions must be flattened.
2293 // FIXME: consider tracking vector origins for app vectors?
2294 if (B->getType()->isVectorTy()) {
2295 Type *FlatTy = getShadowTyNoVec(B->getType());
2296 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2297 ConstantInt::getNullValue(FlatTy));
2298 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2299 ConstantInt::getNullValue(FlatTy));
2301 // a = select b, c, d
2302 // Oa = Sb ? Ob : (b ? Oc : Od)
2303 setOrigin(&I, IRB.CreateSelect(
2304 Sb, getOrigin(I.getCondition()),
2305 IRB.CreateSelect(B, getOrigin(C), getOrigin(D))));
2309 void visitLandingPadInst(LandingPadInst &I) {
2311 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2312 setShadow(&I, getCleanShadow(&I));
2313 setOrigin(&I, getCleanOrigin());
2316 void visitGetElementPtrInst(GetElementPtrInst &I) {
2320 void visitExtractValueInst(ExtractValueInst &I) {
2321 IRBuilder<> IRB(&I);
2322 Value *Agg = I.getAggregateOperand();
2323 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2324 Value *AggShadow = getShadow(Agg);
2325 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2326 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2327 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2328 setShadow(&I, ResShadow);
2329 setOriginForNaryOp(I);
2332 void visitInsertValueInst(InsertValueInst &I) {
2333 IRBuilder<> IRB(&I);
2334 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2335 Value *AggShadow = getShadow(I.getAggregateOperand());
2336 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2337 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2338 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2339 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2340 DEBUG(dbgs() << " Res: " << *Res << "\n");
2342 setOriginForNaryOp(I);
2345 void dumpInst(Instruction &I) {
2346 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2347 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2349 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2351 errs() << "QQQ " << I << "\n";
2354 void visitResumeInst(ResumeInst &I) {
2355 DEBUG(dbgs() << "Resume: " << I << "\n");
2356 // Nothing to do here.
2359 void visitInstruction(Instruction &I) {
2360 // Everything else: stop propagating and check for poisoned shadow.
2361 if (ClDumpStrictInstructions)
2363 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2364 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2365 insertShadowCheck(I.getOperand(i), &I);
2366 setShadow(&I, getCleanShadow(&I));
2367 setOrigin(&I, getCleanOrigin());
2371 /// \brief AMD64-specific implementation of VarArgHelper.
2372 struct VarArgAMD64Helper : public VarArgHelper {
2373 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2374 // See a comment in visitCallSite for more details.
2375 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2376 static const unsigned AMD64FpEndOffset = 176;
2379 MemorySanitizer &MS;
2380 MemorySanitizerVisitor &MSV;
2381 Value *VAArgTLSCopy;
2382 Value *VAArgOverflowSize;
2384 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2386 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2387 MemorySanitizerVisitor &MSV)
2388 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2389 VAArgOverflowSize(nullptr) {}
2391 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2393 ArgKind classifyArgument(Value* arg) {
2394 // A very rough approximation of X86_64 argument classification rules.
2395 Type *T = arg->getType();
2396 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2397 return AK_FloatingPoint;
2398 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2399 return AK_GeneralPurpose;
2400 if (T->isPointerTy())
2401 return AK_GeneralPurpose;
2405 // For VarArg functions, store the argument shadow in an ABI-specific format
2406 // that corresponds to va_list layout.
2407 // We do this because Clang lowers va_arg in the frontend, and this pass
2408 // only sees the low level code that deals with va_list internals.
2409 // A much easier alternative (provided that Clang emits va_arg instructions)
2410 // would have been to associate each live instance of va_list with a copy of
2411 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2413 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2414 unsigned GpOffset = 0;
2415 unsigned FpOffset = AMD64GpEndOffset;
2416 unsigned OverflowOffset = AMD64FpEndOffset;
2417 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2418 ArgIt != End; ++ArgIt) {
2420 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2421 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2423 // ByVal arguments always go to the overflow area.
2424 assert(A->getType()->isPointerTy());
2425 Type *RealTy = A->getType()->getPointerElementType();
2426 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2427 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2428 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2429 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2430 ArgSize, kShadowTLSAlignment);
2432 ArgKind AK = classifyArgument(A);
2433 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2435 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2439 case AK_GeneralPurpose:
2440 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2443 case AK_FloatingPoint:
2444 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2448 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2449 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2450 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2452 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2455 Constant *OverflowSize =
2456 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2457 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2460 /// \brief Compute the shadow address for a given va_arg.
2461 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2463 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2464 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2465 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2469 void visitVAStartInst(VAStartInst &I) override {
2470 IRBuilder<> IRB(&I);
2471 VAStartInstrumentationList.push_back(&I);
2472 Value *VAListTag = I.getArgOperand(0);
2473 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2475 // Unpoison the whole __va_list_tag.
2476 // FIXME: magic ABI constants.
2477 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2478 /* size */24, /* alignment */8, false);
2481 void visitVACopyInst(VACopyInst &I) override {
2482 IRBuilder<> IRB(&I);
2483 Value *VAListTag = I.getArgOperand(0);
2484 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2486 // Unpoison the whole __va_list_tag.
2487 // FIXME: magic ABI constants.
2488 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2489 /* size */24, /* alignment */8, false);
2492 void finalizeInstrumentation() override {
2493 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2494 "finalizeInstrumentation called twice");
2495 if (!VAStartInstrumentationList.empty()) {
2496 // If there is a va_start in this function, make a backup copy of
2497 // va_arg_tls somewhere in the function entry block.
2498 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2499 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2501 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2503 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2504 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2507 // Instrument va_start.
2508 // Copy va_list shadow from the backup copy of the TLS contents.
2509 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2510 CallInst *OrigInst = VAStartInstrumentationList[i];
2511 IRBuilder<> IRB(OrigInst->getNextNode());
2512 Value *VAListTag = OrigInst->getArgOperand(0);
2514 Value *RegSaveAreaPtrPtr =
2516 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2517 ConstantInt::get(MS.IntptrTy, 16)),
2518 Type::getInt64PtrTy(*MS.C));
2519 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2520 Value *RegSaveAreaShadowPtr =
2521 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2522 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2523 AMD64FpEndOffset, 16);
2525 Value *OverflowArgAreaPtrPtr =
2527 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2528 ConstantInt::get(MS.IntptrTy, 8)),
2529 Type::getInt64PtrTy(*MS.C));
2530 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2531 Value *OverflowArgAreaShadowPtr =
2532 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2533 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2534 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2539 /// \brief A no-op implementation of VarArgHelper.
2540 struct VarArgNoOpHelper : public VarArgHelper {
2541 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2542 MemorySanitizerVisitor &MSV) {}
2544 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2546 void visitVAStartInst(VAStartInst &I) override {}
2548 void visitVACopyInst(VACopyInst &I) override {}
2550 void finalizeInstrumentation() override {}
2553 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2554 MemorySanitizerVisitor &Visitor) {
2555 // VarArg handling is only implemented on AMD64. False positives are possible
2556 // on other platforms.
2557 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2558 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2559 return new VarArgAMD64Helper(Func, Msan, Visitor);
2561 return new VarArgNoOpHelper(Func, Msan, Visitor);
2566 bool MemorySanitizer::runOnFunction(Function &F) {
2567 MemorySanitizerVisitor Visitor(F, *this);
2569 // Clear out readonly/readnone attributes.
2571 B.addAttribute(Attribute::ReadOnly)
2572 .addAttribute(Attribute::ReadNone);
2573 F.removeAttributes(AttributeSet::FunctionIndex,
2574 AttributeSet::get(F.getContext(),
2575 AttributeSet::FunctionIndex, B));
2577 return Visitor.runOnFunction();