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 (WrapIndirectCalls && 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 (auto Inst : StoreList) {
603 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
605 IRBuilder<> IRB(&SI);
606 Value *Val = SI.getValueOperand();
607 Value *Addr = SI.getPointerOperand();
608 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
609 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
612 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
613 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
616 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
618 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
620 if (MS.TrackOrigins) {
621 unsigned Alignment = std::max(kMinOriginAlignment, SI.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 (const auto &ShadowData : InstrumentationList) {
666 Instruction *OrigIns = ShadowData.OrigIns;
667 Value *Shadow = ShadowData.Shadow;
668 Value *Origin = ShadowData.Origin;
669 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
671 DEBUG(dbgs() << "DONE:\n" << F);
674 void materializeIndirectCalls() {
675 for (auto &CS : IndirectCallList) {
676 Instruction *I = CS.getInstruction();
677 BasicBlock *B = I->getParent();
679 Value *Fn0 = CS.getCalledValue();
680 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
682 if (ClWrapIndirectCallsFast) {
683 // Check that call target is inside this module limits.
685 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
686 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
688 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
689 IRB.CreateICmpUGE(Fn, End));
692 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
694 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
695 NotInThisModule, NewFnPhi,
696 /* Unreachable */ false, MS.ColdCallWeights);
698 IRB.SetInsertPoint(CheckTerm);
699 // Slow path: call wrapper function to possibly transform the call
701 Value *NewFn = IRB.CreateBitCast(
702 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
704 NewFnPhi->addIncoming(Fn0, B);
705 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
706 CS.setCalledFunction(NewFnPhi);
708 Value *NewFn = IRB.CreateBitCast(
709 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
710 CS.setCalledFunction(NewFn);
715 /// \brief Add MemorySanitizer instrumentation to a function.
716 bool runOnFunction() {
717 MS.initializeCallbacks(*F.getParent());
718 if (!MS.DL) return false;
720 // In the presence of unreachable blocks, we may see Phi nodes with
721 // incoming nodes from such blocks. Since InstVisitor skips unreachable
722 // blocks, such nodes will not have any shadow value associated with them.
723 // It's easier to remove unreachable blocks than deal with missing shadow.
724 removeUnreachableBlocks(F);
726 // Iterate all BBs in depth-first order and create shadow instructions
727 // for all instructions (where applicable).
728 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
729 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
733 // Finalize PHI nodes.
734 for (PHINode *PN : ShadowPHINodes) {
735 PHINode *PNS = cast<PHINode>(getShadow(PN));
736 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
737 size_t NumValues = PN->getNumIncomingValues();
738 for (size_t v = 0; v < NumValues; v++) {
739 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
741 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
745 VAHelper->finalizeInstrumentation();
747 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
748 InstrumentationList.size() + StoreList.size() >
749 (unsigned)ClInstrumentationWithCallThreshold;
751 // Delayed instrumentation of StoreInst.
752 // This may add new checks to be inserted later.
753 materializeStores(InstrumentWithCalls);
755 // Insert shadow value checks.
756 materializeChecks(InstrumentWithCalls);
758 // Wrap indirect calls.
759 materializeIndirectCalls();
764 /// \brief Compute the shadow type that corresponds to a given Value.
765 Type *getShadowTy(Value *V) {
766 return getShadowTy(V->getType());
769 /// \brief Compute the shadow type that corresponds to a given Type.
770 Type *getShadowTy(Type *OrigTy) {
771 if (!OrigTy->isSized()) {
774 // For integer type, shadow is the same as the original type.
775 // This may return weird-sized types like i1.
776 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
778 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
779 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
780 return VectorType::get(IntegerType::get(*MS.C, EltSize),
781 VT->getNumElements());
783 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
784 SmallVector<Type*, 4> Elements;
785 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
786 Elements.push_back(getShadowTy(ST->getElementType(i)));
787 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
788 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
791 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
792 return IntegerType::get(*MS.C, TypeSize);
795 /// \brief Flatten a vector type.
796 Type *getShadowTyNoVec(Type *ty) {
797 if (VectorType *vt = dyn_cast<VectorType>(ty))
798 return IntegerType::get(*MS.C, vt->getBitWidth());
802 /// \brief Convert a shadow value to it's flattened variant.
803 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
804 Type *Ty = V->getType();
805 Type *NoVecTy = getShadowTyNoVec(Ty);
806 if (Ty == NoVecTy) return V;
807 return IRB.CreateBitCast(V, NoVecTy);
810 /// \brief Compute the shadow address that corresponds to a given application
813 /// Shadow = Addr & ~ShadowMask.
814 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
817 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
818 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
819 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
822 /// \brief Compute the origin address that corresponds to a given application
825 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
826 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
828 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
829 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
831 IRB.CreateAdd(ShadowLong,
832 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
834 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
835 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
838 /// \brief Compute the shadow address for a given function argument.
840 /// Shadow = ParamTLS+ArgOffset.
841 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
843 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
844 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
845 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
849 /// \brief Compute the origin address for a given function argument.
850 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
852 if (!MS.TrackOrigins) return nullptr;
853 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
854 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
855 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
859 /// \brief Compute the shadow address for a retval.
860 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
861 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
862 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
866 /// \brief Compute the origin address for a retval.
867 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
868 // We keep a single origin for the entire retval. Might be too optimistic.
869 return MS.RetvalOriginTLS;
872 /// \brief Set SV to be the shadow value for V.
873 void setShadow(Value *V, Value *SV) {
874 assert(!ShadowMap.count(V) && "Values may only have one shadow");
878 /// \brief Set Origin to be the origin value for V.
879 void setOrigin(Value *V, Value *Origin) {
880 if (!MS.TrackOrigins) return;
881 assert(!OriginMap.count(V) && "Values may only have one origin");
882 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
883 OriginMap[V] = Origin;
886 /// \brief Create a clean shadow value for a given value.
888 /// Clean shadow (all zeroes) means all bits of the value are defined
890 Constant *getCleanShadow(Value *V) {
891 Type *ShadowTy = getShadowTy(V);
894 return Constant::getNullValue(ShadowTy);
897 /// \brief Create a dirty shadow of a given shadow type.
898 Constant *getPoisonedShadow(Type *ShadowTy) {
900 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
901 return Constant::getAllOnesValue(ShadowTy);
902 StructType *ST = cast<StructType>(ShadowTy);
903 SmallVector<Constant *, 4> Vals;
904 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
905 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
906 return ConstantStruct::get(ST, Vals);
909 /// \brief Create a dirty shadow for a given value.
910 Constant *getPoisonedShadow(Value *V) {
911 Type *ShadowTy = getShadowTy(V);
914 return getPoisonedShadow(ShadowTy);
917 /// \brief Create a clean (zero) origin.
918 Value *getCleanOrigin() {
919 return Constant::getNullValue(MS.OriginTy);
922 /// \brief Get the shadow value for a given Value.
924 /// This function either returns the value set earlier with setShadow,
925 /// or extracts if from ParamTLS (for function arguments).
926 Value *getShadow(Value *V) {
927 if (Instruction *I = dyn_cast<Instruction>(V)) {
928 // For instructions the shadow is already stored in the map.
929 Value *Shadow = ShadowMap[V];
931 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
933 assert(Shadow && "No shadow for a value");
937 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
938 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
939 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
943 if (Argument *A = dyn_cast<Argument>(V)) {
944 // For arguments we compute the shadow on demand and store it in the map.
945 Value **ShadowPtr = &ShadowMap[V];
948 Function *F = A->getParent();
949 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
950 unsigned ArgOffset = 0;
951 for (auto &FArg : F->args()) {
952 if (!FArg.getType()->isSized()) {
953 DEBUG(dbgs() << "Arg is not sized\n");
956 unsigned Size = FArg.hasByValAttr()
957 ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
958 : MS.DL->getTypeAllocSize(FArg.getType());
960 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
961 if (FArg.hasByValAttr()) {
962 // ByVal pointer itself has clean shadow. We copy the actual
963 // argument shadow to the underlying memory.
964 // Figure out maximal valid memcpy alignment.
965 unsigned ArgAlign = FArg.getParamAlignment();
967 Type *EltType = A->getType()->getPointerElementType();
968 ArgAlign = MS.DL->getABITypeAlignment(EltType);
970 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
971 Value *Cpy = EntryIRB.CreateMemCpy(
972 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
974 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
976 *ShadowPtr = getCleanShadow(V);
978 *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
980 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
981 **ShadowPtr << "\n");
982 if (MS.TrackOrigins) {
984 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
985 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
988 ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
990 assert(*ShadowPtr && "Could not find shadow for an argument");
993 // For everything else the shadow is zero.
994 return getCleanShadow(V);
997 /// \brief Get the shadow for i-th argument of the instruction I.
998 Value *getShadow(Instruction *I, int i) {
999 return getShadow(I->getOperand(i));
1002 /// \brief Get the origin for a value.
1003 Value *getOrigin(Value *V) {
1004 if (!MS.TrackOrigins) return nullptr;
1005 if (isa<Instruction>(V) || isa<Argument>(V)) {
1006 Value *Origin = OriginMap[V];
1008 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
1009 Origin = getCleanOrigin();
1013 return getCleanOrigin();
1016 /// \brief Get the origin for i-th argument of the instruction I.
1017 Value *getOrigin(Instruction *I, int i) {
1018 return getOrigin(I->getOperand(i));
1021 /// \brief Remember the place where a shadow check should be inserted.
1023 /// This location will be later instrumented with a check that will print a
1024 /// UMR warning in runtime if the shadow value is not 0.
1025 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1027 if (!InsertChecks) return;
1029 Type *ShadowTy = Shadow->getType();
1030 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1031 "Can only insert checks for integer and vector shadow types");
1033 InstrumentationList.push_back(
1034 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1037 /// \brief Remember the place where a shadow check should be inserted.
1039 /// This location will be later instrumented with a check that will print a
1040 /// UMR warning in runtime if the value is not fully defined.
1041 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1043 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1044 if (!Shadow) return;
1045 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1046 insertShadowCheck(Shadow, Origin, OrigIns);
1049 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1058 case AcquireRelease:
1059 return AcquireRelease;
1060 case SequentiallyConsistent:
1061 return SequentiallyConsistent;
1063 llvm_unreachable("Unknown ordering");
1066 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1075 case AcquireRelease:
1076 return AcquireRelease;
1077 case SequentiallyConsistent:
1078 return SequentiallyConsistent;
1080 llvm_unreachable("Unknown ordering");
1083 // ------------------- Visitors.
1085 /// \brief Instrument LoadInst
1087 /// Loads the corresponding shadow and (optionally) origin.
1088 /// Optionally, checks that the load address is fully defined.
1089 void visitLoadInst(LoadInst &I) {
1090 assert(I.getType()->isSized() && "Load type must have size");
1091 IRBuilder<> IRB(I.getNextNode());
1092 Type *ShadowTy = getShadowTy(&I);
1093 Value *Addr = I.getPointerOperand();
1095 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1097 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1099 setShadow(&I, getCleanShadow(&I));
1102 if (ClCheckAccessAddress)
1103 insertShadowCheck(I.getPointerOperand(), &I);
1106 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1108 if (MS.TrackOrigins) {
1110 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1112 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1114 setOrigin(&I, getCleanOrigin());
1119 /// \brief Instrument StoreInst
1121 /// Stores the corresponding shadow and (optionally) origin.
1122 /// Optionally, checks that the store address is fully defined.
1123 void visitStoreInst(StoreInst &I) {
1124 StoreList.push_back(&I);
1127 void handleCASOrRMW(Instruction &I) {
1128 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1130 IRBuilder<> IRB(&I);
1131 Value *Addr = I.getOperand(0);
1132 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1134 if (ClCheckAccessAddress)
1135 insertShadowCheck(Addr, &I);
1137 // Only test the conditional argument of cmpxchg instruction.
1138 // The other argument can potentially be uninitialized, but we can not
1139 // detect this situation reliably without possible false positives.
1140 if (isa<AtomicCmpXchgInst>(I))
1141 insertShadowCheck(I.getOperand(1), &I);
1143 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1145 setShadow(&I, getCleanShadow(&I));
1148 void visitAtomicRMWInst(AtomicRMWInst &I) {
1150 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1153 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1155 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1158 // Vector manipulation.
1159 void visitExtractElementInst(ExtractElementInst &I) {
1160 insertShadowCheck(I.getOperand(1), &I);
1161 IRBuilder<> IRB(&I);
1162 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1164 setOrigin(&I, getOrigin(&I, 0));
1167 void visitInsertElementInst(InsertElementInst &I) {
1168 insertShadowCheck(I.getOperand(2), &I);
1169 IRBuilder<> IRB(&I);
1170 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1171 I.getOperand(2), "_msprop"));
1172 setOriginForNaryOp(I);
1175 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1176 insertShadowCheck(I.getOperand(2), &I);
1177 IRBuilder<> IRB(&I);
1178 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1179 I.getOperand(2), "_msprop"));
1180 setOriginForNaryOp(I);
1184 void visitSExtInst(SExtInst &I) {
1185 IRBuilder<> IRB(&I);
1186 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1187 setOrigin(&I, getOrigin(&I, 0));
1190 void visitZExtInst(ZExtInst &I) {
1191 IRBuilder<> IRB(&I);
1192 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1193 setOrigin(&I, getOrigin(&I, 0));
1196 void visitTruncInst(TruncInst &I) {
1197 IRBuilder<> IRB(&I);
1198 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1199 setOrigin(&I, getOrigin(&I, 0));
1202 void visitBitCastInst(BitCastInst &I) {
1203 IRBuilder<> IRB(&I);
1204 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1205 setOrigin(&I, getOrigin(&I, 0));
1208 void visitPtrToIntInst(PtrToIntInst &I) {
1209 IRBuilder<> IRB(&I);
1210 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1211 "_msprop_ptrtoint"));
1212 setOrigin(&I, getOrigin(&I, 0));
1215 void visitIntToPtrInst(IntToPtrInst &I) {
1216 IRBuilder<> IRB(&I);
1217 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1218 "_msprop_inttoptr"));
1219 setOrigin(&I, getOrigin(&I, 0));
1222 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1223 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1224 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1225 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1226 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1227 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1229 /// \brief Propagate shadow for bitwise AND.
1231 /// This code is exact, i.e. if, for example, a bit in the left argument
1232 /// is defined and 0, then neither the value not definedness of the
1233 /// corresponding bit in B don't affect the resulting shadow.
1234 void visitAnd(BinaryOperator &I) {
1235 IRBuilder<> IRB(&I);
1236 // "And" of 0 and a poisoned value results in unpoisoned value.
1237 // 1&1 => 1; 0&1 => 0; p&1 => p;
1238 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1239 // 1&p => p; 0&p => 0; p&p => p;
1240 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1241 Value *S1 = getShadow(&I, 0);
1242 Value *S2 = getShadow(&I, 1);
1243 Value *V1 = I.getOperand(0);
1244 Value *V2 = I.getOperand(1);
1245 if (V1->getType() != S1->getType()) {
1246 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1247 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1249 Value *S1S2 = IRB.CreateAnd(S1, S2);
1250 Value *V1S2 = IRB.CreateAnd(V1, S2);
1251 Value *S1V2 = IRB.CreateAnd(S1, V2);
1252 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1253 setOriginForNaryOp(I);
1256 void visitOr(BinaryOperator &I) {
1257 IRBuilder<> IRB(&I);
1258 // "Or" of 1 and a poisoned value results in unpoisoned value.
1259 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1260 // 1|0 => 1; 0|0 => 0; p|0 => p;
1261 // 1|p => 1; 0|p => p; p|p => p;
1262 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1263 Value *S1 = getShadow(&I, 0);
1264 Value *S2 = getShadow(&I, 1);
1265 Value *V1 = IRB.CreateNot(I.getOperand(0));
1266 Value *V2 = IRB.CreateNot(I.getOperand(1));
1267 if (V1->getType() != S1->getType()) {
1268 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1269 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1271 Value *S1S2 = IRB.CreateAnd(S1, S2);
1272 Value *V1S2 = IRB.CreateAnd(V1, S2);
1273 Value *S1V2 = IRB.CreateAnd(S1, V2);
1274 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1275 setOriginForNaryOp(I);
1278 /// \brief Default propagation of shadow and/or origin.
1280 /// This class implements the general case of shadow propagation, used in all
1281 /// cases where we don't know and/or don't care about what the operation
1282 /// actually does. It converts all input shadow values to a common type
1283 /// (extending or truncating as necessary), and bitwise OR's them.
1285 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1286 /// fully initialized), and less prone to false positives.
1288 /// This class also implements the general case of origin propagation. For a
1289 /// Nary operation, result origin is set to the origin of an argument that is
1290 /// not entirely initialized. If there is more than one such arguments, the
1291 /// rightmost of them is picked. It does not matter which one is picked if all
1292 /// arguments are initialized.
1293 template <bool CombineShadow>
1298 MemorySanitizerVisitor *MSV;
1301 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1302 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1304 /// \brief Add a pair of shadow and origin values to the mix.
1305 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1306 if (CombineShadow) {
1311 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1312 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1316 if (MSV->MS.TrackOrigins) {
1321 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1322 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1323 MSV->getCleanShadow(FlatShadow));
1324 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1330 /// \brief Add an application value to the mix.
1331 Combiner &Add(Value *V) {
1332 Value *OpShadow = MSV->getShadow(V);
1333 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1334 return Add(OpShadow, OpOrigin);
1337 /// \brief Set the current combined values as the given instruction's shadow
1339 void Done(Instruction *I) {
1340 if (CombineShadow) {
1342 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1343 MSV->setShadow(I, Shadow);
1345 if (MSV->MS.TrackOrigins) {
1347 MSV->setOrigin(I, Origin);
1352 typedef Combiner<true> ShadowAndOriginCombiner;
1353 typedef Combiner<false> OriginCombiner;
1355 /// \brief Propagate origin for arbitrary operation.
1356 void setOriginForNaryOp(Instruction &I) {
1357 if (!MS.TrackOrigins) return;
1358 IRBuilder<> IRB(&I);
1359 OriginCombiner OC(this, IRB);
1360 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1365 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1366 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1367 "Vector of pointers is not a valid shadow type");
1368 return Ty->isVectorTy() ?
1369 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1370 Ty->getPrimitiveSizeInBits();
1373 /// \brief Cast between two shadow types, extending or truncating as
1375 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1376 bool Signed = false) {
1377 Type *srcTy = V->getType();
1378 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1379 return IRB.CreateIntCast(V, dstTy, Signed);
1380 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1381 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1382 return IRB.CreateIntCast(V, dstTy, Signed);
1383 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1384 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1385 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1387 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1388 return IRB.CreateBitCast(V2, dstTy);
1389 // TODO: handle struct types.
1392 /// \brief Cast an application value to the type of its own shadow.
1393 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1394 Type *ShadowTy = getShadowTy(V);
1395 if (V->getType() == ShadowTy)
1397 if (V->getType()->isPtrOrPtrVectorTy())
1398 return IRB.CreatePtrToInt(V, ShadowTy);
1400 return IRB.CreateBitCast(V, ShadowTy);
1403 /// \brief Propagate shadow for arbitrary operation.
1404 void handleShadowOr(Instruction &I) {
1405 IRBuilder<> IRB(&I);
1406 ShadowAndOriginCombiner SC(this, IRB);
1407 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1412 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1413 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1414 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1415 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1416 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1417 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1418 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1420 void handleDiv(Instruction &I) {
1421 IRBuilder<> IRB(&I);
1422 // Strict on the second argument.
1423 insertShadowCheck(I.getOperand(1), &I);
1424 setShadow(&I, getShadow(&I, 0));
1425 setOrigin(&I, getOrigin(&I, 0));
1428 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1429 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1430 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1431 void visitURem(BinaryOperator &I) { handleDiv(I); }
1432 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1433 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1435 /// \brief Instrument == and != comparisons.
1437 /// Sometimes the comparison result is known even if some of the bits of the
1438 /// arguments are not.
1439 void handleEqualityComparison(ICmpInst &I) {
1440 IRBuilder<> IRB(&I);
1441 Value *A = I.getOperand(0);
1442 Value *B = I.getOperand(1);
1443 Value *Sa = getShadow(A);
1444 Value *Sb = getShadow(B);
1446 // Get rid of pointers and vectors of pointers.
1447 // For ints (and vectors of ints), types of A and Sa match,
1448 // and this is a no-op.
1449 A = IRB.CreatePointerCast(A, Sa->getType());
1450 B = IRB.CreatePointerCast(B, Sb->getType());
1452 // A == B <==> (C = A^B) == 0
1453 // A != B <==> (C = A^B) != 0
1455 Value *C = IRB.CreateXor(A, B);
1456 Value *Sc = IRB.CreateOr(Sa, Sb);
1457 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1458 // Result is defined if one of the following is true
1459 // * there is a defined 1 bit in C
1460 // * C is fully defined
1461 // Si = !(C & ~Sc) && Sc
1462 Value *Zero = Constant::getNullValue(Sc->getType());
1463 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1465 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1467 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1468 Si->setName("_msprop_icmp");
1470 setOriginForNaryOp(I);
1473 /// \brief Build the lowest possible value of V, taking into account V's
1474 /// uninitialized bits.
1475 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1478 // Split shadow into sign bit and other bits.
1479 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1480 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1481 // Maximise the undefined shadow bit, minimize other undefined bits.
1483 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1485 // Minimize undefined bits.
1486 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1490 /// \brief Build the highest possible value of V, taking into account V's
1491 /// uninitialized bits.
1492 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1495 // Split shadow into sign bit and other bits.
1496 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1497 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1498 // Minimise the undefined shadow bit, maximise other undefined bits.
1500 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1502 // Maximize undefined bits.
1503 return IRB.CreateOr(A, Sa);
1507 /// \brief Instrument relational comparisons.
1509 /// This function does exact shadow propagation for all relational
1510 /// comparisons of integers, pointers and vectors of those.
1511 /// FIXME: output seems suboptimal when one of the operands is a constant
1512 void handleRelationalComparisonExact(ICmpInst &I) {
1513 IRBuilder<> IRB(&I);
1514 Value *A = I.getOperand(0);
1515 Value *B = I.getOperand(1);
1516 Value *Sa = getShadow(A);
1517 Value *Sb = getShadow(B);
1519 // Get rid of pointers and vectors of pointers.
1520 // For ints (and vectors of ints), types of A and Sa match,
1521 // and this is a no-op.
1522 A = IRB.CreatePointerCast(A, Sa->getType());
1523 B = IRB.CreatePointerCast(B, Sb->getType());
1525 // Let [a0, a1] be the interval of possible values of A, taking into account
1526 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1527 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1528 bool IsSigned = I.isSigned();
1529 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1530 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1531 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1532 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1533 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1534 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1535 Value *Si = IRB.CreateXor(S1, S2);
1537 setOriginForNaryOp(I);
1540 /// \brief Instrument signed relational comparisons.
1542 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1543 /// propagating the highest bit of the shadow. Everything else is delegated
1544 /// to handleShadowOr().
1545 void handleSignedRelationalComparison(ICmpInst &I) {
1546 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1547 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1548 Value* op = nullptr;
1549 CmpInst::Predicate pre = I.getPredicate();
1550 if (constOp0 && constOp0->isNullValue() &&
1551 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1552 op = I.getOperand(1);
1553 } else if (constOp1 && constOp1->isNullValue() &&
1554 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1555 op = I.getOperand(0);
1558 IRBuilder<> IRB(&I);
1560 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1561 setShadow(&I, Shadow);
1562 setOrigin(&I, getOrigin(op));
1568 void visitICmpInst(ICmpInst &I) {
1569 if (!ClHandleICmp) {
1573 if (I.isEquality()) {
1574 handleEqualityComparison(I);
1578 assert(I.isRelational());
1579 if (ClHandleICmpExact) {
1580 handleRelationalComparisonExact(I);
1584 handleSignedRelationalComparison(I);
1588 assert(I.isUnsigned());
1589 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1590 handleRelationalComparisonExact(I);
1597 void visitFCmpInst(FCmpInst &I) {
1601 void handleShift(BinaryOperator &I) {
1602 IRBuilder<> IRB(&I);
1603 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1604 // Otherwise perform the same shift on S1.
1605 Value *S1 = getShadow(&I, 0);
1606 Value *S2 = getShadow(&I, 1);
1607 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1609 Value *V2 = I.getOperand(1);
1610 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1611 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1612 setOriginForNaryOp(I);
1615 void visitShl(BinaryOperator &I) { handleShift(I); }
1616 void visitAShr(BinaryOperator &I) { handleShift(I); }
1617 void visitLShr(BinaryOperator &I) { handleShift(I); }
1619 /// \brief Instrument llvm.memmove
1621 /// At this point we don't know if llvm.memmove will be inlined or not.
1622 /// If we don't instrument it and it gets inlined,
1623 /// our interceptor will not kick in and we will lose the memmove.
1624 /// If we instrument the call here, but it does not get inlined,
1625 /// we will memove the shadow twice: which is bad in case
1626 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1628 /// Similar situation exists for memcpy and memset.
1629 void visitMemMoveInst(MemMoveInst &I) {
1630 IRBuilder<> IRB(&I);
1633 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1634 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1635 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1636 I.eraseFromParent();
1639 // Similar to memmove: avoid copying shadow twice.
1640 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1641 // FIXME: consider doing manual inline for small constant sizes and proper
1643 void visitMemCpyInst(MemCpyInst &I) {
1644 IRBuilder<> IRB(&I);
1647 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1648 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1649 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1650 I.eraseFromParent();
1654 void visitMemSetInst(MemSetInst &I) {
1655 IRBuilder<> IRB(&I);
1658 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1659 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1660 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1661 I.eraseFromParent();
1664 void visitVAStartInst(VAStartInst &I) {
1665 VAHelper->visitVAStartInst(I);
1668 void visitVACopyInst(VACopyInst &I) {
1669 VAHelper->visitVACopyInst(I);
1672 enum IntrinsicKind {
1673 IK_DoesNotAccessMemory,
1678 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1679 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1680 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1681 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1682 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1683 const int UnknownModRefBehavior = IK_WritesMemory;
1684 #define GET_INTRINSIC_MODREF_BEHAVIOR
1685 #define ModRefBehavior IntrinsicKind
1686 #include "llvm/IR/Intrinsics.gen"
1687 #undef ModRefBehavior
1688 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1691 /// \brief Handle vector store-like intrinsics.
1693 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1694 /// has 1 pointer argument and 1 vector argument, returns void.
1695 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1696 IRBuilder<> IRB(&I);
1697 Value* Addr = I.getArgOperand(0);
1698 Value *Shadow = getShadow(&I, 1);
1699 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1701 // We don't know the pointer alignment (could be unaligned SSE store!).
1702 // Have to assume to worst case.
1703 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1705 if (ClCheckAccessAddress)
1706 insertShadowCheck(Addr, &I);
1708 // FIXME: use ClStoreCleanOrigin
1709 // FIXME: factor out common code from materializeStores
1710 if (MS.TrackOrigins)
1711 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1715 /// \brief Handle vector load-like intrinsics.
1717 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1718 /// has 1 pointer argument, returns a vector.
1719 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1720 IRBuilder<> IRB(&I);
1721 Value *Addr = I.getArgOperand(0);
1723 Type *ShadowTy = getShadowTy(&I);
1725 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1726 // We don't know the pointer alignment (could be unaligned SSE load!).
1727 // Have to assume to worst case.
1728 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1730 setShadow(&I, getCleanShadow(&I));
1733 if (ClCheckAccessAddress)
1734 insertShadowCheck(Addr, &I);
1736 if (MS.TrackOrigins) {
1738 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1740 setOrigin(&I, getCleanOrigin());
1745 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1747 /// Instrument intrinsics with any number of arguments of the same type,
1748 /// equal to the return type. The type should be simple (no aggregates or
1749 /// pointers; vectors are fine).
1750 /// Caller guarantees that this intrinsic does not access memory.
1751 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1752 Type *RetTy = I.getType();
1753 if (!(RetTy->isIntOrIntVectorTy() ||
1754 RetTy->isFPOrFPVectorTy() ||
1755 RetTy->isX86_MMXTy()))
1758 unsigned NumArgOperands = I.getNumArgOperands();
1760 for (unsigned i = 0; i < NumArgOperands; ++i) {
1761 Type *Ty = I.getArgOperand(i)->getType();
1766 IRBuilder<> IRB(&I);
1767 ShadowAndOriginCombiner SC(this, IRB);
1768 for (unsigned i = 0; i < NumArgOperands; ++i)
1769 SC.Add(I.getArgOperand(i));
1775 /// \brief Heuristically instrument unknown intrinsics.
1777 /// The main purpose of this code is to do something reasonable with all
1778 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1779 /// We recognize several classes of intrinsics by their argument types and
1780 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1781 /// sure that we know what the intrinsic does.
1783 /// We special-case intrinsics where this approach fails. See llvm.bswap
1784 /// handling as an example of that.
1785 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1786 unsigned NumArgOperands = I.getNumArgOperands();
1787 if (NumArgOperands == 0)
1790 Intrinsic::ID iid = I.getIntrinsicID();
1791 IntrinsicKind IK = getIntrinsicKind(iid);
1792 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1793 bool WritesMemory = IK == IK_WritesMemory;
1794 assert(!(OnlyReadsMemory && WritesMemory));
1796 if (NumArgOperands == 2 &&
1797 I.getArgOperand(0)->getType()->isPointerTy() &&
1798 I.getArgOperand(1)->getType()->isVectorTy() &&
1799 I.getType()->isVoidTy() &&
1801 // This looks like a vector store.
1802 return handleVectorStoreIntrinsic(I);
1805 if (NumArgOperands == 1 &&
1806 I.getArgOperand(0)->getType()->isPointerTy() &&
1807 I.getType()->isVectorTy() &&
1809 // This looks like a vector load.
1810 return handleVectorLoadIntrinsic(I);
1813 if (!OnlyReadsMemory && !WritesMemory)
1814 if (maybeHandleSimpleNomemIntrinsic(I))
1817 // FIXME: detect and handle SSE maskstore/maskload
1821 void handleBswap(IntrinsicInst &I) {
1822 IRBuilder<> IRB(&I);
1823 Value *Op = I.getArgOperand(0);
1824 Type *OpType = Op->getType();
1825 Function *BswapFunc = Intrinsic::getDeclaration(
1826 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1827 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1828 setOrigin(&I, getOrigin(Op));
1831 // \brief Instrument vector convert instrinsic.
1833 // This function instruments intrinsics like cvtsi2ss:
1834 // %Out = int_xxx_cvtyyy(%ConvertOp)
1836 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1837 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1838 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1839 // elements from \p CopyOp.
1840 // In most cases conversion involves floating-point value which may trigger a
1841 // hardware exception when not fully initialized. For this reason we require
1842 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1843 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1844 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1845 // return a fully initialized value.
1846 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1847 IRBuilder<> IRB(&I);
1848 Value *CopyOp, *ConvertOp;
1850 switch (I.getNumArgOperands()) {
1852 CopyOp = I.getArgOperand(0);
1853 ConvertOp = I.getArgOperand(1);
1856 ConvertOp = I.getArgOperand(0);
1860 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1863 // The first *NumUsedElements* elements of ConvertOp are converted to the
1864 // same number of output elements. The rest of the output is copied from
1865 // CopyOp, or (if not available) filled with zeroes.
1866 // Combine shadow for elements of ConvertOp that are used in this operation,
1867 // and insert a check.
1868 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1869 // int->any conversion.
1870 Value *ConvertShadow = getShadow(ConvertOp);
1871 Value *AggShadow = nullptr;
1872 if (ConvertOp->getType()->isVectorTy()) {
1873 AggShadow = IRB.CreateExtractElement(
1874 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1875 for (int i = 1; i < NumUsedElements; ++i) {
1876 Value *MoreShadow = IRB.CreateExtractElement(
1877 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1878 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1881 AggShadow = ConvertShadow;
1883 assert(AggShadow->getType()->isIntegerTy());
1884 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1886 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1889 assert(CopyOp->getType() == I.getType());
1890 assert(CopyOp->getType()->isVectorTy());
1891 Value *ResultShadow = getShadow(CopyOp);
1892 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1893 for (int i = 0; i < NumUsedElements; ++i) {
1894 ResultShadow = IRB.CreateInsertElement(
1895 ResultShadow, ConstantInt::getNullValue(EltTy),
1896 ConstantInt::get(IRB.getInt32Ty(), i));
1898 setShadow(&I, ResultShadow);
1899 setOrigin(&I, getOrigin(CopyOp));
1901 setShadow(&I, getCleanShadow(&I));
1905 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1906 // zeroes if it is zero, and all ones otherwise.
1907 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1908 if (S->getType()->isVectorTy())
1909 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1910 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1911 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1912 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1915 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1916 Type *T = S->getType();
1917 assert(T->isVectorTy());
1918 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1919 return IRB.CreateSExt(S2, T);
1922 // \brief Instrument vector shift instrinsic.
1924 // This function instruments intrinsics like int_x86_avx2_psll_w.
1925 // Intrinsic shifts %In by %ShiftSize bits.
1926 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
1927 // size, and the rest is ignored. Behavior is defined even if shift size is
1928 // greater than register (or field) width.
1929 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
1930 assert(I.getNumArgOperands() == 2);
1931 IRBuilder<> IRB(&I);
1932 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1933 // Otherwise perform the same shift on S1.
1934 Value *S1 = getShadow(&I, 0);
1935 Value *S2 = getShadow(&I, 1);
1936 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
1937 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
1938 Value *V1 = I.getOperand(0);
1939 Value *V2 = I.getOperand(1);
1940 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
1941 IRB.CreateBitCast(S1, V1->getType()), V2);
1942 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
1943 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1944 setOriginForNaryOp(I);
1947 void visitIntrinsicInst(IntrinsicInst &I) {
1948 switch (I.getIntrinsicID()) {
1949 case llvm::Intrinsic::bswap:
1952 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
1953 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
1954 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
1955 case llvm::Intrinsic::x86_avx512_cvtss2usi:
1956 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
1957 case llvm::Intrinsic::x86_avx512_cvttss2usi:
1958 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
1959 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
1960 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
1961 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
1962 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
1963 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
1964 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
1965 case llvm::Intrinsic::x86_sse2_cvtsd2si:
1966 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
1967 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
1968 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
1969 case llvm::Intrinsic::x86_sse2_cvtss2sd:
1970 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
1971 case llvm::Intrinsic::x86_sse2_cvttsd2si:
1972 case llvm::Intrinsic::x86_sse_cvtsi2ss:
1973 case llvm::Intrinsic::x86_sse_cvtsi642ss:
1974 case llvm::Intrinsic::x86_sse_cvtss2si64:
1975 case llvm::Intrinsic::x86_sse_cvtss2si:
1976 case llvm::Intrinsic::x86_sse_cvttss2si64:
1977 case llvm::Intrinsic::x86_sse_cvttss2si:
1978 handleVectorConvertIntrinsic(I, 1);
1980 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
1981 case llvm::Intrinsic::x86_sse2_cvtps2pd:
1982 case llvm::Intrinsic::x86_sse_cvtps2pi:
1983 case llvm::Intrinsic::x86_sse_cvttps2pi:
1984 handleVectorConvertIntrinsic(I, 2);
1986 case llvm::Intrinsic::x86_avx512_psll_dq:
1987 case llvm::Intrinsic::x86_avx512_psrl_dq:
1988 case llvm::Intrinsic::x86_avx2_psll_w:
1989 case llvm::Intrinsic::x86_avx2_psll_d:
1990 case llvm::Intrinsic::x86_avx2_psll_q:
1991 case llvm::Intrinsic::x86_avx2_pslli_w:
1992 case llvm::Intrinsic::x86_avx2_pslli_d:
1993 case llvm::Intrinsic::x86_avx2_pslli_q:
1994 case llvm::Intrinsic::x86_avx2_psll_dq:
1995 case llvm::Intrinsic::x86_avx2_psrl_w:
1996 case llvm::Intrinsic::x86_avx2_psrl_d:
1997 case llvm::Intrinsic::x86_avx2_psrl_q:
1998 case llvm::Intrinsic::x86_avx2_psra_w:
1999 case llvm::Intrinsic::x86_avx2_psra_d:
2000 case llvm::Intrinsic::x86_avx2_psrli_w:
2001 case llvm::Intrinsic::x86_avx2_psrli_d:
2002 case llvm::Intrinsic::x86_avx2_psrli_q:
2003 case llvm::Intrinsic::x86_avx2_psrai_w:
2004 case llvm::Intrinsic::x86_avx2_psrai_d:
2005 case llvm::Intrinsic::x86_avx2_psrl_dq:
2006 case llvm::Intrinsic::x86_sse2_psll_w:
2007 case llvm::Intrinsic::x86_sse2_psll_d:
2008 case llvm::Intrinsic::x86_sse2_psll_q:
2009 case llvm::Intrinsic::x86_sse2_pslli_w:
2010 case llvm::Intrinsic::x86_sse2_pslli_d:
2011 case llvm::Intrinsic::x86_sse2_pslli_q:
2012 case llvm::Intrinsic::x86_sse2_psll_dq:
2013 case llvm::Intrinsic::x86_sse2_psrl_w:
2014 case llvm::Intrinsic::x86_sse2_psrl_d:
2015 case llvm::Intrinsic::x86_sse2_psrl_q:
2016 case llvm::Intrinsic::x86_sse2_psra_w:
2017 case llvm::Intrinsic::x86_sse2_psra_d:
2018 case llvm::Intrinsic::x86_sse2_psrli_w:
2019 case llvm::Intrinsic::x86_sse2_psrli_d:
2020 case llvm::Intrinsic::x86_sse2_psrli_q:
2021 case llvm::Intrinsic::x86_sse2_psrai_w:
2022 case llvm::Intrinsic::x86_sse2_psrai_d:
2023 case llvm::Intrinsic::x86_sse2_psrl_dq:
2024 case llvm::Intrinsic::x86_mmx_psll_w:
2025 case llvm::Intrinsic::x86_mmx_psll_d:
2026 case llvm::Intrinsic::x86_mmx_psll_q:
2027 case llvm::Intrinsic::x86_mmx_pslli_w:
2028 case llvm::Intrinsic::x86_mmx_pslli_d:
2029 case llvm::Intrinsic::x86_mmx_pslli_q:
2030 case llvm::Intrinsic::x86_mmx_psrl_w:
2031 case llvm::Intrinsic::x86_mmx_psrl_d:
2032 case llvm::Intrinsic::x86_mmx_psrl_q:
2033 case llvm::Intrinsic::x86_mmx_psra_w:
2034 case llvm::Intrinsic::x86_mmx_psra_d:
2035 case llvm::Intrinsic::x86_mmx_psrli_w:
2036 case llvm::Intrinsic::x86_mmx_psrli_d:
2037 case llvm::Intrinsic::x86_mmx_psrli_q:
2038 case llvm::Intrinsic::x86_mmx_psrai_w:
2039 case llvm::Intrinsic::x86_mmx_psrai_d:
2040 handleVectorShiftIntrinsic(I, /* Variable */ false);
2042 case llvm::Intrinsic::x86_avx2_psllv_d:
2043 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2044 case llvm::Intrinsic::x86_avx2_psllv_q:
2045 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2046 case llvm::Intrinsic::x86_avx2_psrlv_d:
2047 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2048 case llvm::Intrinsic::x86_avx2_psrlv_q:
2049 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2050 case llvm::Intrinsic::x86_avx2_psrav_d:
2051 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2052 handleVectorShiftIntrinsic(I, /* Variable */ true);
2055 // Byte shifts are not implemented.
2056 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
2057 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
2058 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
2059 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
2060 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
2061 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
2064 if (!handleUnknownIntrinsic(I))
2065 visitInstruction(I);
2070 void visitCallSite(CallSite CS) {
2071 Instruction &I = *CS.getInstruction();
2072 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2074 CallInst *Call = cast<CallInst>(&I);
2076 // For inline asm, do the usual thing: check argument shadow and mark all
2077 // outputs as clean. Note that any side effects of the inline asm that are
2078 // not immediately visible in its constraints are not handled.
2079 if (Call->isInlineAsm()) {
2080 visitInstruction(I);
2084 // Allow only tail calls with the same types, otherwise
2085 // we may have a false positive: shadow for a non-void RetVal
2086 // will get propagated to a void RetVal.
2087 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
2088 Call->setTailCall(false);
2090 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2092 // We are going to insert code that relies on the fact that the callee
2093 // will become a non-readonly function after it is instrumented by us. To
2094 // prevent this code from being optimized out, mark that function
2095 // non-readonly in advance.
2096 if (Function *Func = Call->getCalledFunction()) {
2097 // Clear out readonly/readnone attributes.
2099 B.addAttribute(Attribute::ReadOnly)
2100 .addAttribute(Attribute::ReadNone);
2101 Func->removeAttributes(AttributeSet::FunctionIndex,
2102 AttributeSet::get(Func->getContext(),
2103 AttributeSet::FunctionIndex,
2107 IRBuilder<> IRB(&I);
2109 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2110 IndirectCallList.push_back(CS);
2112 unsigned ArgOffset = 0;
2113 DEBUG(dbgs() << " CallSite: " << I << "\n");
2114 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2115 ArgIt != End; ++ArgIt) {
2117 unsigned i = ArgIt - CS.arg_begin();
2118 if (!A->getType()->isSized()) {
2119 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2123 Value *Store = nullptr;
2124 // Compute the Shadow for arg even if it is ByVal, because
2125 // in that case getShadow() will copy the actual arg shadow to
2126 // __msan_param_tls.
2127 Value *ArgShadow = getShadow(A);
2128 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2129 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2130 " Shadow: " << *ArgShadow << "\n");
2131 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2132 assert(A->getType()->isPointerTy() &&
2133 "ByVal argument is not a pointer!");
2134 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2135 unsigned Alignment = CS.getParamAlignment(i + 1);
2136 Store = IRB.CreateMemCpy(ArgShadowBase,
2137 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2140 Size = MS.DL->getTypeAllocSize(A->getType());
2141 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2142 kShadowTLSAlignment);
2144 if (MS.TrackOrigins)
2145 IRB.CreateStore(getOrigin(A),
2146 getOriginPtrForArgument(A, IRB, ArgOffset));
2148 assert(Size != 0 && Store != nullptr);
2149 DEBUG(dbgs() << " Param:" << *Store << "\n");
2150 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
2152 DEBUG(dbgs() << " done with call args\n");
2155 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2156 if (FT->isVarArg()) {
2157 VAHelper->visitCallSite(CS, IRB);
2160 // Now, get the shadow for the RetVal.
2161 if (!I.getType()->isSized()) return;
2162 IRBuilder<> IRBBefore(&I);
2163 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2164 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2165 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2166 Instruction *NextInsn = nullptr;
2168 NextInsn = I.getNextNode();
2170 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2171 if (!NormalDest->getSinglePredecessor()) {
2172 // FIXME: this case is tricky, so we are just conservative here.
2173 // Perhaps we need to split the edge between this BB and NormalDest,
2174 // but a naive attempt to use SplitEdge leads to a crash.
2175 setShadow(&I, getCleanShadow(&I));
2176 setOrigin(&I, getCleanOrigin());
2179 NextInsn = NormalDest->getFirstInsertionPt();
2181 "Could not find insertion point for retval shadow load");
2183 IRBuilder<> IRBAfter(NextInsn);
2184 Value *RetvalShadow =
2185 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2186 kShadowTLSAlignment, "_msret");
2187 setShadow(&I, RetvalShadow);
2188 if (MS.TrackOrigins)
2189 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2192 void visitReturnInst(ReturnInst &I) {
2193 IRBuilder<> IRB(&I);
2194 Value *RetVal = I.getReturnValue();
2195 if (!RetVal) return;
2196 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2197 if (CheckReturnValue) {
2198 insertShadowCheck(RetVal, &I);
2199 Value *Shadow = getCleanShadow(RetVal);
2200 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2202 Value *Shadow = getShadow(RetVal);
2203 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2204 // FIXME: make it conditional if ClStoreCleanOrigin==0
2205 if (MS.TrackOrigins)
2206 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2210 void visitPHINode(PHINode &I) {
2211 IRBuilder<> IRB(&I);
2212 ShadowPHINodes.push_back(&I);
2213 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2215 if (MS.TrackOrigins)
2216 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2220 void visitAllocaInst(AllocaInst &I) {
2221 setShadow(&I, getCleanShadow(&I));
2222 IRBuilder<> IRB(I.getNextNode());
2223 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2224 if (PoisonStack && ClPoisonStackWithCall) {
2225 IRB.CreateCall2(MS.MsanPoisonStackFn,
2226 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2227 ConstantInt::get(MS.IntptrTy, Size));
2229 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2230 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2231 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2234 if (PoisonStack && MS.TrackOrigins) {
2235 setOrigin(&I, getCleanOrigin());
2236 SmallString<2048> StackDescriptionStorage;
2237 raw_svector_ostream StackDescription(StackDescriptionStorage);
2238 // We create a string with a description of the stack allocation and
2239 // pass it into __msan_set_alloca_origin.
2240 // It will be printed by the run-time if stack-originated UMR is found.
2241 // The first 4 bytes of the string are set to '----' and will be replaced
2242 // by __msan_va_arg_overflow_size_tls at the first call.
2243 StackDescription << "----" << I.getName() << "@" << F.getName();
2245 createPrivateNonConstGlobalForString(*F.getParent(),
2246 StackDescription.str());
2248 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2249 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2250 ConstantInt::get(MS.IntptrTy, Size),
2251 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2252 IRB.CreatePointerCast(&F, MS.IntptrTy));
2256 void visitSelectInst(SelectInst& I) {
2257 IRBuilder<> IRB(&I);
2258 // a = select b, c, d
2259 Value *B = I.getCondition();
2260 Value *C = I.getTrueValue();
2261 Value *D = I.getFalseValue();
2262 Value *Sb = getShadow(B);
2263 Value *Sc = getShadow(C);
2264 Value *Sd = getShadow(D);
2266 // Result shadow if condition shadow is 0.
2267 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2269 if (I.getType()->isAggregateType()) {
2270 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2271 // an extra "select". This results in much more compact IR.
2272 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2273 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2275 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2276 // If Sb (condition is poisoned), look for bits in c and d that are equal
2277 // and both unpoisoned.
2278 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2280 // Cast arguments to shadow-compatible type.
2281 C = CreateAppToShadowCast(IRB, C);
2282 D = CreateAppToShadowCast(IRB, D);
2284 // Result shadow if condition shadow is 1.
2285 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2287 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2289 if (MS.TrackOrigins) {
2290 // Origins are always i32, so any vector conditions must be flattened.
2291 // FIXME: consider tracking vector origins for app vectors?
2292 if (B->getType()->isVectorTy()) {
2293 Type *FlatTy = getShadowTyNoVec(B->getType());
2294 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2295 ConstantInt::getNullValue(FlatTy));
2296 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2297 ConstantInt::getNullValue(FlatTy));
2299 // a = select b, c, d
2300 // Oa = Sb ? Ob : (b ? Oc : Od)
2301 setOrigin(&I, IRB.CreateSelect(
2302 Sb, getOrigin(I.getCondition()),
2303 IRB.CreateSelect(B, getOrigin(C), getOrigin(D))));
2307 void visitLandingPadInst(LandingPadInst &I) {
2309 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2310 setShadow(&I, getCleanShadow(&I));
2311 setOrigin(&I, getCleanOrigin());
2314 void visitGetElementPtrInst(GetElementPtrInst &I) {
2318 void visitExtractValueInst(ExtractValueInst &I) {
2319 IRBuilder<> IRB(&I);
2320 Value *Agg = I.getAggregateOperand();
2321 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2322 Value *AggShadow = getShadow(Agg);
2323 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2324 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2325 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2326 setShadow(&I, ResShadow);
2327 setOriginForNaryOp(I);
2330 void visitInsertValueInst(InsertValueInst &I) {
2331 IRBuilder<> IRB(&I);
2332 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2333 Value *AggShadow = getShadow(I.getAggregateOperand());
2334 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2335 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2336 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2337 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2338 DEBUG(dbgs() << " Res: " << *Res << "\n");
2340 setOriginForNaryOp(I);
2343 void dumpInst(Instruction &I) {
2344 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2345 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2347 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2349 errs() << "QQQ " << I << "\n";
2352 void visitResumeInst(ResumeInst &I) {
2353 DEBUG(dbgs() << "Resume: " << I << "\n");
2354 // Nothing to do here.
2357 void visitInstruction(Instruction &I) {
2358 // Everything else: stop propagating and check for poisoned shadow.
2359 if (ClDumpStrictInstructions)
2361 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2362 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2363 insertShadowCheck(I.getOperand(i), &I);
2364 setShadow(&I, getCleanShadow(&I));
2365 setOrigin(&I, getCleanOrigin());
2369 /// \brief AMD64-specific implementation of VarArgHelper.
2370 struct VarArgAMD64Helper : public VarArgHelper {
2371 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2372 // See a comment in visitCallSite for more details.
2373 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2374 static const unsigned AMD64FpEndOffset = 176;
2377 MemorySanitizer &MS;
2378 MemorySanitizerVisitor &MSV;
2379 Value *VAArgTLSCopy;
2380 Value *VAArgOverflowSize;
2382 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2384 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2385 MemorySanitizerVisitor &MSV)
2386 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2387 VAArgOverflowSize(nullptr) {}
2389 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2391 ArgKind classifyArgument(Value* arg) {
2392 // A very rough approximation of X86_64 argument classification rules.
2393 Type *T = arg->getType();
2394 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2395 return AK_FloatingPoint;
2396 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2397 return AK_GeneralPurpose;
2398 if (T->isPointerTy())
2399 return AK_GeneralPurpose;
2403 // For VarArg functions, store the argument shadow in an ABI-specific format
2404 // that corresponds to va_list layout.
2405 // We do this because Clang lowers va_arg in the frontend, and this pass
2406 // only sees the low level code that deals with va_list internals.
2407 // A much easier alternative (provided that Clang emits va_arg instructions)
2408 // would have been to associate each live instance of va_list with a copy of
2409 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2411 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2412 unsigned GpOffset = 0;
2413 unsigned FpOffset = AMD64GpEndOffset;
2414 unsigned OverflowOffset = AMD64FpEndOffset;
2415 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2416 ArgIt != End; ++ArgIt) {
2418 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2419 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2421 // ByVal arguments always go to the overflow area.
2422 assert(A->getType()->isPointerTy());
2423 Type *RealTy = A->getType()->getPointerElementType();
2424 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2425 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2426 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2427 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2428 ArgSize, kShadowTLSAlignment);
2430 ArgKind AK = classifyArgument(A);
2431 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2433 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2437 case AK_GeneralPurpose:
2438 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2441 case AK_FloatingPoint:
2442 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2446 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2447 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2448 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2450 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2453 Constant *OverflowSize =
2454 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2455 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2458 /// \brief Compute the shadow address for a given va_arg.
2459 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2461 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2462 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2463 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2467 void visitVAStartInst(VAStartInst &I) override {
2468 IRBuilder<> IRB(&I);
2469 VAStartInstrumentationList.push_back(&I);
2470 Value *VAListTag = I.getArgOperand(0);
2471 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2473 // Unpoison the whole __va_list_tag.
2474 // FIXME: magic ABI constants.
2475 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2476 /* size */24, /* alignment */8, false);
2479 void visitVACopyInst(VACopyInst &I) override {
2480 IRBuilder<> IRB(&I);
2481 Value *VAListTag = I.getArgOperand(0);
2482 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2484 // Unpoison the whole __va_list_tag.
2485 // FIXME: magic ABI constants.
2486 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2487 /* size */24, /* alignment */8, false);
2490 void finalizeInstrumentation() override {
2491 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2492 "finalizeInstrumentation called twice");
2493 if (!VAStartInstrumentationList.empty()) {
2494 // If there is a va_start in this function, make a backup copy of
2495 // va_arg_tls somewhere in the function entry block.
2496 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2497 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2499 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2501 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2502 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2505 // Instrument va_start.
2506 // Copy va_list shadow from the backup copy of the TLS contents.
2507 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2508 CallInst *OrigInst = VAStartInstrumentationList[i];
2509 IRBuilder<> IRB(OrigInst->getNextNode());
2510 Value *VAListTag = OrigInst->getArgOperand(0);
2512 Value *RegSaveAreaPtrPtr =
2514 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2515 ConstantInt::get(MS.IntptrTy, 16)),
2516 Type::getInt64PtrTy(*MS.C));
2517 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2518 Value *RegSaveAreaShadowPtr =
2519 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2520 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2521 AMD64FpEndOffset, 16);
2523 Value *OverflowArgAreaPtrPtr =
2525 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2526 ConstantInt::get(MS.IntptrTy, 8)),
2527 Type::getInt64PtrTy(*MS.C));
2528 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2529 Value *OverflowArgAreaShadowPtr =
2530 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2531 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2532 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2537 /// \brief A no-op implementation of VarArgHelper.
2538 struct VarArgNoOpHelper : public VarArgHelper {
2539 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2540 MemorySanitizerVisitor &MSV) {}
2542 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2544 void visitVAStartInst(VAStartInst &I) override {}
2546 void visitVACopyInst(VACopyInst &I) override {}
2548 void finalizeInstrumentation() override {}
2551 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2552 MemorySanitizerVisitor &Visitor) {
2553 // VarArg handling is only implemented on AMD64. False positives are possible
2554 // on other platforms.
2555 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2556 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2557 return new VarArgAMD64Helper(Func, Msan, Visitor);
2559 return new VarArgNoOpHelper(Func, Msan, Visitor);
2564 bool MemorySanitizer::runOnFunction(Function &F) {
2565 MemorySanitizerVisitor Visitor(F, *this);
2567 // Clear out readonly/readnone attributes.
2569 B.addAttribute(Attribute::ReadOnly)
2570 .addAttribute(Attribute::ReadNone);
2571 F.removeAttributes(AttributeSet::FunctionIndex,
2572 AttributeSet::get(F.getContext(),
2573 AttributeSet::FunctionIndex, B));
2575 return Visitor.runOnFunction();