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 /// The algorithm of the tool is similar to Memcheck
14 /// (http://goo.gl/QKbem). We associate a few shadow bits with every
15 /// byte of the application memory, poison the shadow of the malloc-ed
16 /// or alloca-ed memory, load the shadow bits on every memory read,
17 /// propagate the shadow bits through some of the arithmetic
18 /// instruction (including MOV), store the shadow bits on every memory
19 /// write, report a bug on some other instructions (e.g. JMP) if the
20 /// associated shadow is poisoned.
22 /// But there are differences too. The first and the major one:
23 /// compiler instrumentation instead of binary instrumentation. This
24 /// gives us much better register allocation, possible compiler
25 /// optimizations and a fast start-up. But this brings the major issue
26 /// as well: msan needs to see all program events, including system
27 /// calls and reads/writes in system libraries, so we either need to
28 /// compile *everything* with msan or use a binary translation
29 /// component (e.g. DynamoRIO) to instrument pre-built libraries.
30 /// Another difference from Memcheck is that we use 8 shadow bits per
31 /// byte of application memory and use a direct shadow mapping. This
32 /// greatly simplifies the instrumentation code and avoids races on
33 /// shadow updates (Memcheck is single-threaded so races are not a
34 /// concern there. Memcheck uses 2 shadow bits per byte with a slow
35 /// path storage that uses 8 bits per byte).
37 /// The default value of shadow is 0, which means "clean" (not poisoned).
39 /// Every module initializer should call __msan_init to ensure that the
40 /// shadow memory is ready. On error, __msan_warning is called. Since
41 /// parameters and return values may be passed via registers, we have a
42 /// specialized thread-local shadow for return values
43 /// (__msan_retval_tls) and parameters (__msan_param_tls).
47 /// MemorySanitizer can track origins (allocation points) of all uninitialized
48 /// values. This behavior is controlled with a flag (msan-track-origins) and is
49 /// disabled by default.
51 /// Origins are 4-byte values created and interpreted by the runtime library.
52 /// They are stored in a second shadow mapping, one 4-byte value for 4 bytes
53 /// of application memory. Propagation of origins is basically a bunch of
54 /// "select" instructions that pick the origin of a dirty argument, if an
55 /// instruction has one.
57 /// Every 4 aligned, consecutive bytes of application memory have one origin
58 /// value associated with them. If these bytes contain uninitialized data
59 /// coming from 2 different allocations, the last store wins. Because of this,
60 /// MemorySanitizer reports can show unrelated origins, but this is unlikely in
63 /// Origins are meaningless for fully initialized values, so MemorySanitizer
64 /// avoids storing origin to memory when a fully initialized value is stored.
65 /// This way it avoids needless overwritting origin of the 4-byte region on
66 /// a short (i.e. 1 byte) clean store, and it is also good for performance.
70 /// Ideally, every atomic store of application value should update the
71 /// corresponding shadow location in an atomic way. Unfortunately, atomic store
72 /// of two disjoint locations can not be done without severe slowdown.
74 /// Therefore, we implement an approximation that may err on the safe side.
75 /// In this implementation, every atomically accessed location in the program
76 /// may only change from (partially) uninitialized to fully initialized, but
77 /// not the other way around. We load the shadow _after_ the application load,
78 /// and we store the shadow _before_ the app store. Also, we always store clean
79 /// shadow (if the application store is atomic). This way, if the store-load
80 /// pair constitutes a happens-before arc, shadow store and load are correctly
81 /// ordered such that the load will get either the value that was stored, or
82 /// some later value (which is always clean).
84 /// This does not work very well with Compare-And-Swap (CAS) and
85 /// Read-Modify-Write (RMW) operations. To follow the above logic, CAS and RMW
86 /// must store the new shadow before the app operation, and load the shadow
87 /// after the app operation. Computers don't work this way. Current
88 /// implementation ignores the load aspect of CAS/RMW, always returning a clean
89 /// value. It implements the store part as a simple atomic store by storing a
92 //===----------------------------------------------------------------------===//
94 #include "llvm/Transforms/Instrumentation.h"
95 #include "llvm/ADT/DepthFirstIterator.h"
96 #include "llvm/ADT/SmallString.h"
97 #include "llvm/ADT/SmallVector.h"
98 #include "llvm/ADT/StringExtras.h"
99 #include "llvm/ADT/Triple.h"
100 #include "llvm/IR/DataLayout.h"
101 #include "llvm/IR/Function.h"
102 #include "llvm/IR/IRBuilder.h"
103 #include "llvm/IR/InlineAsm.h"
104 #include "llvm/IR/InstVisitor.h"
105 #include "llvm/IR/IntrinsicInst.h"
106 #include "llvm/IR/LLVMContext.h"
107 #include "llvm/IR/MDBuilder.h"
108 #include "llvm/IR/Module.h"
109 #include "llvm/IR/Type.h"
110 #include "llvm/IR/ValueMap.h"
111 #include "llvm/Support/CommandLine.h"
112 #include "llvm/Support/Compiler.h"
113 #include "llvm/Support/Debug.h"
114 #include "llvm/Support/raw_ostream.h"
115 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
116 #include "llvm/Transforms/Utils/Local.h"
117 #include "llvm/Transforms/Utils/ModuleUtils.h"
119 using namespace llvm;
121 #define DEBUG_TYPE "msan"
123 static const uint64_t kShadowMask32 = 1ULL << 31;
124 static const uint64_t kShadowMask64 = 1ULL << 46;
125 static const uint64_t kOriginOffset32 = 1ULL << 30;
126 static const uint64_t kOriginOffset64 = 1ULL << 45;
127 static const unsigned kMinOriginAlignment = 4;
128 static const unsigned kShadowTLSAlignment = 8;
130 // These constants must be kept in sync with the ones in msan.h.
131 static const unsigned kParamTLSSize = 800;
132 static const unsigned kRetvalTLSSize = 800;
134 // Accesses sizes are powers of two: 1, 2, 4, 8.
135 static const size_t kNumberOfAccessSizes = 4;
137 /// \brief Track origins of uninitialized values.
139 /// Adds a section to MemorySanitizer report that points to the allocation
140 /// (stack or heap) the uninitialized bits came from originally.
141 static cl::opt<int> ClTrackOrigins("msan-track-origins",
142 cl::desc("Track origins (allocation sites) of poisoned memory"),
143 cl::Hidden, cl::init(0));
144 static cl::opt<bool> ClKeepGoing("msan-keep-going",
145 cl::desc("keep going after reporting a UMR"),
146 cl::Hidden, cl::init(false));
147 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
148 cl::desc("poison uninitialized stack variables"),
149 cl::Hidden, cl::init(true));
150 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
151 cl::desc("poison uninitialized stack variables with a call"),
152 cl::Hidden, cl::init(false));
153 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
154 cl::desc("poison uninitialized stack variables with the given patter"),
155 cl::Hidden, cl::init(0xff));
156 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
157 cl::desc("poison undef temps"),
158 cl::Hidden, cl::init(true));
160 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
161 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
162 cl::Hidden, cl::init(true));
164 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
165 cl::desc("exact handling of relational integer ICmp"),
166 cl::Hidden, cl::init(false));
168 // This flag controls whether we check the shadow of the address
169 // operand of load or store. Such bugs are very rare, since load from
170 // a garbage address typically results in SEGV, but still happen
171 // (e.g. only lower bits of address are garbage, or the access happens
172 // early at program startup where malloc-ed memory is more likely to
173 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
174 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
175 cl::desc("report accesses through a pointer which has poisoned shadow"),
176 cl::Hidden, cl::init(true));
178 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
179 cl::desc("print out instructions with default strict semantics"),
180 cl::Hidden, cl::init(false));
182 static cl::opt<int> ClInstrumentationWithCallThreshold(
183 "msan-instrumentation-with-call-threshold",
185 "If the function being instrumented requires more than "
186 "this number of checks and origin stores, use callbacks instead of "
187 "inline checks (-1 means never use callbacks)."),
188 cl::Hidden, cl::init(3500));
190 // Experimental. Wraps all indirect calls in the instrumented code with
191 // a call to the given function. This is needed to assist the dynamic
192 // helper tool (MSanDR) to regain control on transition between instrumented and
193 // non-instrumented code.
194 static cl::opt<std::string> ClWrapIndirectCalls("msan-wrap-indirect-calls",
195 cl::desc("Wrap indirect calls with a given function"),
198 static cl::opt<bool> ClWrapIndirectCallsFast("msan-wrap-indirect-calls-fast",
199 cl::desc("Do not wrap indirect calls with target in the same module"),
200 cl::Hidden, cl::init(true));
202 // This is an experiment to enable handling of cases where shadow is a non-zero
203 // compile-time constant. For some unexplainable reason they were silently
204 // ignored in the instrumentation.
205 static cl::opt<bool> ClCheckConstantShadow("msan-check-constant-shadow",
206 cl::desc("Insert checks for constant shadow values"),
207 cl::Hidden, cl::init(false));
211 /// \brief An instrumentation pass implementing detection of uninitialized
214 /// MemorySanitizer: instrument the code in module to find
215 /// uninitialized reads.
216 class MemorySanitizer : public FunctionPass {
218 MemorySanitizer(int TrackOrigins = 0)
220 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
223 WrapIndirectCalls(!ClWrapIndirectCalls.empty()) {}
224 const char *getPassName() const override { return "MemorySanitizer"; }
225 bool runOnFunction(Function &F) override;
226 bool doInitialization(Module &M) override;
227 static char ID; // Pass identification, replacement for typeid.
230 void initializeCallbacks(Module &M);
232 /// \brief Track origins (allocation points) of uninitialized values.
235 const DataLayout *DL;
239 /// \brief Thread-local shadow storage for function parameters.
240 GlobalVariable *ParamTLS;
241 /// \brief Thread-local origin storage for function parameters.
242 GlobalVariable *ParamOriginTLS;
243 /// \brief Thread-local shadow storage for function return value.
244 GlobalVariable *RetvalTLS;
245 /// \brief Thread-local origin storage for function return value.
246 GlobalVariable *RetvalOriginTLS;
247 /// \brief Thread-local shadow storage for in-register va_arg function
248 /// parameters (x86_64-specific).
249 GlobalVariable *VAArgTLS;
250 /// \brief Thread-local shadow storage for va_arg overflow area
251 /// (x86_64-specific).
252 GlobalVariable *VAArgOverflowSizeTLS;
253 /// \brief Thread-local space used to pass origin value to the UMR reporting
255 GlobalVariable *OriginTLS;
257 GlobalVariable *MsandrModuleStart;
258 GlobalVariable *MsandrModuleEnd;
260 /// \brief The run-time callback to print a warning.
262 // These arrays are indexed by log2(AccessSize).
263 Value *MaybeWarningFn[kNumberOfAccessSizes];
264 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
266 /// \brief Run-time helper that generates a new origin value for a stack
268 Value *MsanSetAllocaOrigin4Fn;
269 /// \brief Run-time helper that poisons stack on function entry.
270 Value *MsanPoisonStackFn;
271 /// \brief Run-time helper that records a store (or any event) of an
272 /// uninitialized value and returns an updated origin id encoding this info.
273 Value *MsanChainOriginFn;
274 /// \brief MSan runtime replacements for memmove, memcpy and memset.
275 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
277 /// \brief Address mask used in application-to-shadow address calculation.
278 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
280 /// \brief Offset of the origin shadow from the "normal" shadow.
281 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
282 uint64_t OriginOffset;
283 /// \brief Branch weights for error reporting.
284 MDNode *ColdCallWeights;
285 /// \brief Branch weights for origin store.
286 MDNode *OriginStoreWeights;
287 /// \brief An empty volatile inline asm that prevents callback merge.
290 bool WrapIndirectCalls;
291 /// \brief Run-time wrapper for indirect calls.
292 Value *IndirectCallWrapperFn;
293 // Argument and return type of IndirectCallWrapperFn: void (*f)(void).
294 Type *AnyFunctionPtrTy;
296 friend struct MemorySanitizerVisitor;
297 friend struct VarArgAMD64Helper;
301 char MemorySanitizer::ID = 0;
302 INITIALIZE_PASS(MemorySanitizer, "msan",
303 "MemorySanitizer: detects uninitialized reads.",
306 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
307 return new MemorySanitizer(TrackOrigins);
310 /// \brief Create a non-const global initialized with the given string.
312 /// Creates a writable global for Str so that we can pass it to the
313 /// run-time lib. Runtime uses first 4 bytes of the string to store the
314 /// frame ID, so the string needs to be mutable.
315 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
317 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
318 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
319 GlobalValue::PrivateLinkage, StrConst, "");
323 /// \brief Insert extern declaration of runtime-provided functions and globals.
324 void MemorySanitizer::initializeCallbacks(Module &M) {
325 // Only do this once.
330 // Create the callback.
331 // FIXME: this function should have "Cold" calling conv,
332 // which is not yet implemented.
333 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
334 : "__msan_warning_noreturn";
335 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
337 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
339 unsigned AccessSize = 1 << AccessSizeIndex;
340 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
341 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
342 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
343 IRB.getInt32Ty(), NULL);
345 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
346 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
347 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
348 IRB.getInt8PtrTy(), IRB.getInt32Ty(), NULL);
351 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
352 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
353 IRB.getInt8PtrTy(), IntptrTy, NULL);
354 MsanPoisonStackFn = M.getOrInsertFunction(
355 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
356 MsanChainOriginFn = M.getOrInsertFunction(
357 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), NULL);
358 MemmoveFn = M.getOrInsertFunction(
359 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
360 IRB.getInt8PtrTy(), IntptrTy, NULL);
361 MemcpyFn = M.getOrInsertFunction(
362 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
364 MemsetFn = M.getOrInsertFunction(
365 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
369 RetvalTLS = new GlobalVariable(
370 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
371 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
372 GlobalVariable::InitialExecTLSModel);
373 RetvalOriginTLS = new GlobalVariable(
374 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
375 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
377 ParamTLS = new GlobalVariable(
378 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
379 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
380 GlobalVariable::InitialExecTLSModel);
381 ParamOriginTLS = new GlobalVariable(
382 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
383 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
384 nullptr, GlobalVariable::InitialExecTLSModel);
386 VAArgTLS = new GlobalVariable(
387 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
388 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
389 GlobalVariable::InitialExecTLSModel);
390 VAArgOverflowSizeTLS = new GlobalVariable(
391 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
392 "__msan_va_arg_overflow_size_tls", nullptr,
393 GlobalVariable::InitialExecTLSModel);
394 OriginTLS = new GlobalVariable(
395 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
396 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
398 // We insert an empty inline asm after __msan_report* to avoid callback merge.
399 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
400 StringRef(""), StringRef(""),
401 /*hasSideEffects=*/true);
403 if (WrapIndirectCalls) {
405 PointerType::getUnqual(FunctionType::get(IRB.getVoidTy(), false));
406 IndirectCallWrapperFn = M.getOrInsertFunction(
407 ClWrapIndirectCalls, AnyFunctionPtrTy, AnyFunctionPtrTy, NULL);
410 if (WrapIndirectCalls && ClWrapIndirectCallsFast) {
411 MsandrModuleStart = new GlobalVariable(
412 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
413 nullptr, "__executable_start");
414 MsandrModuleStart->setVisibility(GlobalVariable::HiddenVisibility);
415 MsandrModuleEnd = new GlobalVariable(
416 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
418 MsandrModuleEnd->setVisibility(GlobalVariable::HiddenVisibility);
422 /// \brief Module-level initialization.
424 /// inserts a call to __msan_init to the module's constructor list.
425 bool MemorySanitizer::doInitialization(Module &M) {
426 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
428 report_fatal_error("data layout missing");
429 DL = &DLP->getDataLayout();
431 C = &(M.getContext());
432 unsigned PtrSize = DL->getPointerSizeInBits(/* AddressSpace */0);
435 ShadowMask = kShadowMask64;
436 OriginOffset = kOriginOffset64;
439 ShadowMask = kShadowMask32;
440 OriginOffset = kOriginOffset32;
443 report_fatal_error("unsupported pointer size");
448 IntptrTy = IRB.getIntPtrTy(DL);
449 OriginTy = IRB.getInt32Ty();
451 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
452 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
454 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
455 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
456 "__msan_init", IRB.getVoidTy(), NULL)), 0);
459 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
460 IRB.getInt32(TrackOrigins), "__msan_track_origins");
463 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
464 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
471 /// \brief A helper class that handles instrumentation of VarArg
472 /// functions on a particular platform.
474 /// Implementations are expected to insert the instrumentation
475 /// necessary to propagate argument shadow through VarArg function
476 /// calls. Visit* methods are called during an InstVisitor pass over
477 /// the function, and should avoid creating new basic blocks. A new
478 /// instance of this class is created for each instrumented function.
479 struct VarArgHelper {
480 /// \brief Visit a CallSite.
481 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
483 /// \brief Visit a va_start call.
484 virtual void visitVAStartInst(VAStartInst &I) = 0;
486 /// \brief Visit a va_copy call.
487 virtual void visitVACopyInst(VACopyInst &I) = 0;
489 /// \brief Finalize function instrumentation.
491 /// This method is called after visiting all interesting (see above)
492 /// instructions in a function.
493 virtual void finalizeInstrumentation() = 0;
495 virtual ~VarArgHelper() {}
498 struct MemorySanitizerVisitor;
501 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
502 MemorySanitizerVisitor &Visitor);
504 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
505 if (TypeSize <= 8) return 0;
506 return Log2_32_Ceil(TypeSize / 8);
509 /// This class does all the work for a given function. Store and Load
510 /// instructions store and load corresponding shadow and origin
511 /// values. Most instructions propagate shadow from arguments to their
512 /// return values. Certain instructions (most importantly, BranchInst)
513 /// test their argument shadow and print reports (with a runtime call) if it's
515 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
518 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
519 ValueMap<Value*, Value*> ShadowMap, OriginMap;
520 std::unique_ptr<VarArgHelper> VAHelper;
522 // The following flags disable parts of MSan instrumentation based on
523 // blacklist contents and command-line options.
525 bool PropagateShadow;
528 bool CheckReturnValue;
530 struct ShadowOriginAndInsertPoint {
533 Instruction *OrigIns;
534 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
535 : Shadow(S), Origin(O), OrigIns(I) { }
537 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
538 SmallVector<Instruction*, 16> StoreList;
539 SmallVector<CallSite, 16> IndirectCallList;
541 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
542 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
543 bool SanitizeFunction = F.getAttributes().hasAttribute(
544 AttributeSet::FunctionIndex, Attribute::SanitizeMemory);
545 InsertChecks = SanitizeFunction;
546 PropagateShadow = SanitizeFunction;
547 PoisonStack = SanitizeFunction && ClPoisonStack;
548 PoisonUndef = SanitizeFunction && ClPoisonUndef;
549 // FIXME: Consider using SpecialCaseList to specify a list of functions that
550 // must always return fully initialized values. For now, we hardcode "main".
551 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
553 DEBUG(if (!InsertChecks)
554 dbgs() << "MemorySanitizer is not inserting checks into '"
555 << F.getName() << "'\n");
558 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
559 if (MS.TrackOrigins <= 1) return V;
560 return IRB.CreateCall(MS.MsanChainOriginFn, V);
563 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
564 unsigned Alignment, bool AsCall) {
565 if (isa<StructType>(Shadow->getType())) {
566 IRB.CreateAlignedStore(updateOrigin(Origin, IRB), getOriginPtr(Addr, IRB),
569 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
570 // TODO(eugenis): handle non-zero constant shadow by inserting an
571 // unconditional check (can not simply fail compilation as this could
572 // be in the dead code).
573 if (!ClCheckConstantShadow)
574 if (isa<Constant>(ConvertedShadow)) return;
575 unsigned TypeSizeInBits =
576 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
577 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
578 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
579 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
580 Value *ConvertedShadow2 = IRB.CreateZExt(
581 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
582 IRB.CreateCall3(Fn, ConvertedShadow2,
583 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
586 Value *Cmp = IRB.CreateICmpNE(
587 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
588 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
589 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
590 IRBuilder<> IRBNew(CheckTerm);
591 IRBNew.CreateAlignedStore(updateOrigin(Origin, IRBNew),
592 getOriginPtr(Addr, IRBNew), Alignment);
597 void materializeStores(bool InstrumentWithCalls) {
598 for (auto Inst : StoreList) {
599 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
601 IRBuilder<> IRB(&SI);
602 Value *Val = SI.getValueOperand();
603 Value *Addr = SI.getPointerOperand();
604 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
605 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
608 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
609 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
612 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
614 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
616 if (MS.TrackOrigins) {
617 unsigned Alignment = std::max(kMinOriginAlignment, SI.getAlignment());
618 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), Alignment,
619 InstrumentWithCalls);
624 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
626 IRBuilder<> IRB(OrigIns);
627 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
628 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
629 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
630 // See the comment in storeOrigin().
631 if (!ClCheckConstantShadow)
632 if (isa<Constant>(ConvertedShadow)) return;
633 unsigned TypeSizeInBits =
634 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
635 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
636 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
637 Value *Fn = MS.MaybeWarningFn[SizeIndex];
638 Value *ConvertedShadow2 =
639 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
640 IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
642 : (Value *)IRB.getInt32(0));
644 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
645 getCleanShadow(ConvertedShadow), "_mscmp");
646 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
648 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
650 IRB.SetInsertPoint(CheckTerm);
651 if (MS.TrackOrigins) {
652 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
655 IRB.CreateCall(MS.WarningFn);
656 IRB.CreateCall(MS.EmptyAsm);
657 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
661 void materializeChecks(bool InstrumentWithCalls) {
662 for (const auto &ShadowData : InstrumentationList) {
663 Instruction *OrigIns = ShadowData.OrigIns;
664 Value *Shadow = ShadowData.Shadow;
665 Value *Origin = ShadowData.Origin;
666 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
668 DEBUG(dbgs() << "DONE:\n" << F);
671 void materializeIndirectCalls() {
672 for (auto &CS : IndirectCallList) {
673 Instruction *I = CS.getInstruction();
674 BasicBlock *B = I->getParent();
676 Value *Fn0 = CS.getCalledValue();
677 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
679 if (ClWrapIndirectCallsFast) {
680 // Check that call target is inside this module limits.
682 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
683 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
685 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
686 IRB.CreateICmpUGE(Fn, End));
689 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
691 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
692 NotInThisModule, NewFnPhi,
693 /* Unreachable */ false, MS.ColdCallWeights);
695 IRB.SetInsertPoint(CheckTerm);
696 // Slow path: call wrapper function to possibly transform the call
698 Value *NewFn = IRB.CreateBitCast(
699 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
701 NewFnPhi->addIncoming(Fn0, B);
702 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
703 CS.setCalledFunction(NewFnPhi);
705 Value *NewFn = IRB.CreateBitCast(
706 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
707 CS.setCalledFunction(NewFn);
712 /// \brief Add MemorySanitizer instrumentation to a function.
713 bool runOnFunction() {
714 MS.initializeCallbacks(*F.getParent());
715 if (!MS.DL) return false;
717 // In the presence of unreachable blocks, we may see Phi nodes with
718 // incoming nodes from such blocks. Since InstVisitor skips unreachable
719 // blocks, such nodes will not have any shadow value associated with them.
720 // It's easier to remove unreachable blocks than deal with missing shadow.
721 removeUnreachableBlocks(F);
723 // Iterate all BBs in depth-first order and create shadow instructions
724 // for all instructions (where applicable).
725 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
726 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
730 // Finalize PHI nodes.
731 for (PHINode *PN : ShadowPHINodes) {
732 PHINode *PNS = cast<PHINode>(getShadow(PN));
733 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
734 size_t NumValues = PN->getNumIncomingValues();
735 for (size_t v = 0; v < NumValues; v++) {
736 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
737 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
741 VAHelper->finalizeInstrumentation();
743 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
744 InstrumentationList.size() + StoreList.size() >
745 (unsigned)ClInstrumentationWithCallThreshold;
747 // Delayed instrumentation of StoreInst.
748 // This may add new checks to be inserted later.
749 materializeStores(InstrumentWithCalls);
751 // Insert shadow value checks.
752 materializeChecks(InstrumentWithCalls);
754 // Wrap indirect calls.
755 materializeIndirectCalls();
760 /// \brief Compute the shadow type that corresponds to a given Value.
761 Type *getShadowTy(Value *V) {
762 return getShadowTy(V->getType());
765 /// \brief Compute the shadow type that corresponds to a given Type.
766 Type *getShadowTy(Type *OrigTy) {
767 if (!OrigTy->isSized()) {
770 // For integer type, shadow is the same as the original type.
771 // This may return weird-sized types like i1.
772 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
774 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
775 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
776 return VectorType::get(IntegerType::get(*MS.C, EltSize),
777 VT->getNumElements());
779 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
780 return ArrayType::get(getShadowTy(AT->getElementType()),
781 AT->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");
875 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
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 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
903 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
904 getPoisonedShadow(AT->getElementType()));
905 return ConstantArray::get(AT, Vals);
907 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
908 SmallVector<Constant *, 4> Vals;
909 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
910 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
911 return ConstantStruct::get(ST, Vals);
913 llvm_unreachable("Unexpected shadow type");
916 /// \brief Create a dirty shadow for a given value.
917 Constant *getPoisonedShadow(Value *V) {
918 Type *ShadowTy = getShadowTy(V);
921 return getPoisonedShadow(ShadowTy);
924 /// \brief Create a clean (zero) origin.
925 Value *getCleanOrigin() {
926 return Constant::getNullValue(MS.OriginTy);
929 /// \brief Get the shadow value for a given Value.
931 /// This function either returns the value set earlier with setShadow,
932 /// or extracts if from ParamTLS (for function arguments).
933 Value *getShadow(Value *V) {
934 if (!PropagateShadow) return getCleanShadow(V);
935 if (Instruction *I = dyn_cast<Instruction>(V)) {
936 // For instructions the shadow is already stored in the map.
937 Value *Shadow = ShadowMap[V];
939 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
941 assert(Shadow && "No shadow for a value");
945 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
946 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
947 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
951 if (Argument *A = dyn_cast<Argument>(V)) {
952 // For arguments we compute the shadow on demand and store it in the map.
953 Value **ShadowPtr = &ShadowMap[V];
956 Function *F = A->getParent();
957 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
958 unsigned ArgOffset = 0;
959 for (auto &FArg : F->args()) {
960 if (!FArg.getType()->isSized()) {
961 DEBUG(dbgs() << "Arg is not sized\n");
964 unsigned Size = FArg.hasByValAttr()
965 ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
966 : MS.DL->getTypeAllocSize(FArg.getType());
968 bool Overflow = ArgOffset + Size > kParamTLSSize;
969 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
970 if (FArg.hasByValAttr()) {
971 // ByVal pointer itself has clean shadow. We copy the actual
972 // argument shadow to the underlying memory.
973 // Figure out maximal valid memcpy alignment.
974 unsigned ArgAlign = FArg.getParamAlignment();
976 Type *EltType = A->getType()->getPointerElementType();
977 ArgAlign = MS.DL->getABITypeAlignment(EltType);
980 // ParamTLS overflow.
981 EntryIRB.CreateMemSet(
982 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
983 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
985 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
986 Value *Cpy = EntryIRB.CreateMemCpy(
987 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
989 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
992 *ShadowPtr = getCleanShadow(V);
995 // ParamTLS overflow.
996 *ShadowPtr = getCleanShadow(V);
999 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1002 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1003 **ShadowPtr << "\n");
1004 if (MS.TrackOrigins && !Overflow) {
1006 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1007 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1010 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1012 assert(*ShadowPtr && "Could not find shadow for an argument");
1015 // For everything else the shadow is zero.
1016 return getCleanShadow(V);
1019 /// \brief Get the shadow for i-th argument of the instruction I.
1020 Value *getShadow(Instruction *I, int i) {
1021 return getShadow(I->getOperand(i));
1024 /// \brief Get the origin for a value.
1025 Value *getOrigin(Value *V) {
1026 if (!MS.TrackOrigins) return nullptr;
1027 if (isa<Instruction>(V) || isa<Argument>(V)) {
1028 Value *Origin = OriginMap[V];
1030 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
1031 Origin = getCleanOrigin();
1035 return getCleanOrigin();
1038 /// \brief Get the origin for i-th argument of the instruction I.
1039 Value *getOrigin(Instruction *I, int i) {
1040 return getOrigin(I->getOperand(i));
1043 /// \brief Remember the place where a shadow check should be inserted.
1045 /// This location will be later instrumented with a check that will print a
1046 /// UMR warning in runtime if the shadow value is not 0.
1047 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1049 if (!InsertChecks) return;
1051 Type *ShadowTy = Shadow->getType();
1052 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1053 "Can only insert checks for integer and vector shadow types");
1055 InstrumentationList.push_back(
1056 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1059 /// \brief Remember the place where a shadow check should be inserted.
1061 /// This location will be later instrumented with a check that will print a
1062 /// UMR warning in runtime if the value is not fully defined.
1063 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1065 Value *Shadow, *Origin;
1066 if (ClCheckConstantShadow) {
1067 Shadow = getShadow(Val);
1068 if (!Shadow) return;
1069 Origin = getOrigin(Val);
1071 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1072 if (!Shadow) return;
1073 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1075 insertShadowCheck(Shadow, Origin, OrigIns);
1078 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1087 case AcquireRelease:
1088 return AcquireRelease;
1089 case SequentiallyConsistent:
1090 return SequentiallyConsistent;
1092 llvm_unreachable("Unknown ordering");
1095 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1104 case AcquireRelease:
1105 return AcquireRelease;
1106 case SequentiallyConsistent:
1107 return SequentiallyConsistent;
1109 llvm_unreachable("Unknown ordering");
1112 // ------------------- Visitors.
1114 /// \brief Instrument LoadInst
1116 /// Loads the corresponding shadow and (optionally) origin.
1117 /// Optionally, checks that the load address is fully defined.
1118 void visitLoadInst(LoadInst &I) {
1119 assert(I.getType()->isSized() && "Load type must have size");
1120 IRBuilder<> IRB(I.getNextNode());
1121 Type *ShadowTy = getShadowTy(&I);
1122 Value *Addr = I.getPointerOperand();
1123 if (PropagateShadow) {
1124 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1126 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1128 setShadow(&I, getCleanShadow(&I));
1131 if (ClCheckAccessAddress)
1132 insertShadowCheck(I.getPointerOperand(), &I);
1135 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1137 if (MS.TrackOrigins) {
1138 if (PropagateShadow) {
1139 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1141 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1143 setOrigin(&I, getCleanOrigin());
1148 /// \brief Instrument StoreInst
1150 /// Stores the corresponding shadow and (optionally) origin.
1151 /// Optionally, checks that the store address is fully defined.
1152 void visitStoreInst(StoreInst &I) {
1153 StoreList.push_back(&I);
1156 void handleCASOrRMW(Instruction &I) {
1157 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1159 IRBuilder<> IRB(&I);
1160 Value *Addr = I.getOperand(0);
1161 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1163 if (ClCheckAccessAddress)
1164 insertShadowCheck(Addr, &I);
1166 // Only test the conditional argument of cmpxchg instruction.
1167 // The other argument can potentially be uninitialized, but we can not
1168 // detect this situation reliably without possible false positives.
1169 if (isa<AtomicCmpXchgInst>(I))
1170 insertShadowCheck(I.getOperand(1), &I);
1172 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1174 setShadow(&I, getCleanShadow(&I));
1177 void visitAtomicRMWInst(AtomicRMWInst &I) {
1179 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1182 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1184 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1187 // Vector manipulation.
1188 void visitExtractElementInst(ExtractElementInst &I) {
1189 insertShadowCheck(I.getOperand(1), &I);
1190 IRBuilder<> IRB(&I);
1191 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1193 setOrigin(&I, getOrigin(&I, 0));
1196 void visitInsertElementInst(InsertElementInst &I) {
1197 insertShadowCheck(I.getOperand(2), &I);
1198 IRBuilder<> IRB(&I);
1199 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1200 I.getOperand(2), "_msprop"));
1201 setOriginForNaryOp(I);
1204 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1205 insertShadowCheck(I.getOperand(2), &I);
1206 IRBuilder<> IRB(&I);
1207 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1208 I.getOperand(2), "_msprop"));
1209 setOriginForNaryOp(I);
1213 void visitSExtInst(SExtInst &I) {
1214 IRBuilder<> IRB(&I);
1215 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1216 setOrigin(&I, getOrigin(&I, 0));
1219 void visitZExtInst(ZExtInst &I) {
1220 IRBuilder<> IRB(&I);
1221 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1222 setOrigin(&I, getOrigin(&I, 0));
1225 void visitTruncInst(TruncInst &I) {
1226 IRBuilder<> IRB(&I);
1227 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1228 setOrigin(&I, getOrigin(&I, 0));
1231 void visitBitCastInst(BitCastInst &I) {
1232 IRBuilder<> IRB(&I);
1233 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1234 setOrigin(&I, getOrigin(&I, 0));
1237 void visitPtrToIntInst(PtrToIntInst &I) {
1238 IRBuilder<> IRB(&I);
1239 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1240 "_msprop_ptrtoint"));
1241 setOrigin(&I, getOrigin(&I, 0));
1244 void visitIntToPtrInst(IntToPtrInst &I) {
1245 IRBuilder<> IRB(&I);
1246 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1247 "_msprop_inttoptr"));
1248 setOrigin(&I, getOrigin(&I, 0));
1251 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1252 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1253 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1254 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1255 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1256 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1258 /// \brief Propagate shadow for bitwise AND.
1260 /// This code is exact, i.e. if, for example, a bit in the left argument
1261 /// is defined and 0, then neither the value not definedness of the
1262 /// corresponding bit in B don't affect the resulting shadow.
1263 void visitAnd(BinaryOperator &I) {
1264 IRBuilder<> IRB(&I);
1265 // "And" of 0 and a poisoned value results in unpoisoned value.
1266 // 1&1 => 1; 0&1 => 0; p&1 => p;
1267 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1268 // 1&p => p; 0&p => 0; p&p => p;
1269 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1270 Value *S1 = getShadow(&I, 0);
1271 Value *S2 = getShadow(&I, 1);
1272 Value *V1 = I.getOperand(0);
1273 Value *V2 = I.getOperand(1);
1274 if (V1->getType() != S1->getType()) {
1275 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1276 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1278 Value *S1S2 = IRB.CreateAnd(S1, S2);
1279 Value *V1S2 = IRB.CreateAnd(V1, S2);
1280 Value *S1V2 = IRB.CreateAnd(S1, V2);
1281 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1282 setOriginForNaryOp(I);
1285 void visitOr(BinaryOperator &I) {
1286 IRBuilder<> IRB(&I);
1287 // "Or" of 1 and a poisoned value results in unpoisoned value.
1288 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1289 // 1|0 => 1; 0|0 => 0; p|0 => p;
1290 // 1|p => 1; 0|p => p; p|p => p;
1291 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1292 Value *S1 = getShadow(&I, 0);
1293 Value *S2 = getShadow(&I, 1);
1294 Value *V1 = IRB.CreateNot(I.getOperand(0));
1295 Value *V2 = IRB.CreateNot(I.getOperand(1));
1296 if (V1->getType() != S1->getType()) {
1297 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1298 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1300 Value *S1S2 = IRB.CreateAnd(S1, S2);
1301 Value *V1S2 = IRB.CreateAnd(V1, S2);
1302 Value *S1V2 = IRB.CreateAnd(S1, V2);
1303 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1304 setOriginForNaryOp(I);
1307 /// \brief Default propagation of shadow and/or origin.
1309 /// This class implements the general case of shadow propagation, used in all
1310 /// cases where we don't know and/or don't care about what the operation
1311 /// actually does. It converts all input shadow values to a common type
1312 /// (extending or truncating as necessary), and bitwise OR's them.
1314 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1315 /// fully initialized), and less prone to false positives.
1317 /// This class also implements the general case of origin propagation. For a
1318 /// Nary operation, result origin is set to the origin of an argument that is
1319 /// not entirely initialized. If there is more than one such arguments, the
1320 /// rightmost of them is picked. It does not matter which one is picked if all
1321 /// arguments are initialized.
1322 template <bool CombineShadow>
1327 MemorySanitizerVisitor *MSV;
1330 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1331 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1333 /// \brief Add a pair of shadow and origin values to the mix.
1334 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1335 if (CombineShadow) {
1340 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1341 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1345 if (MSV->MS.TrackOrigins) {
1350 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1351 // No point in adding something that might result in 0 origin value.
1352 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1353 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1355 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1356 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1363 /// \brief Add an application value to the mix.
1364 Combiner &Add(Value *V) {
1365 Value *OpShadow = MSV->getShadow(V);
1366 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1367 return Add(OpShadow, OpOrigin);
1370 /// \brief Set the current combined values as the given instruction's shadow
1372 void Done(Instruction *I) {
1373 if (CombineShadow) {
1375 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1376 MSV->setShadow(I, Shadow);
1378 if (MSV->MS.TrackOrigins) {
1380 MSV->setOrigin(I, Origin);
1385 typedef Combiner<true> ShadowAndOriginCombiner;
1386 typedef Combiner<false> OriginCombiner;
1388 /// \brief Propagate origin for arbitrary operation.
1389 void setOriginForNaryOp(Instruction &I) {
1390 if (!MS.TrackOrigins) return;
1391 IRBuilder<> IRB(&I);
1392 OriginCombiner OC(this, IRB);
1393 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1398 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1399 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1400 "Vector of pointers is not a valid shadow type");
1401 return Ty->isVectorTy() ?
1402 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1403 Ty->getPrimitiveSizeInBits();
1406 /// \brief Cast between two shadow types, extending or truncating as
1408 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1409 bool Signed = false) {
1410 Type *srcTy = V->getType();
1411 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1412 return IRB.CreateIntCast(V, dstTy, Signed);
1413 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1414 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1415 return IRB.CreateIntCast(V, dstTy, Signed);
1416 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1417 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1418 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1420 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1421 return IRB.CreateBitCast(V2, dstTy);
1422 // TODO: handle struct types.
1425 /// \brief Cast an application value to the type of its own shadow.
1426 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1427 Type *ShadowTy = getShadowTy(V);
1428 if (V->getType() == ShadowTy)
1430 if (V->getType()->isPtrOrPtrVectorTy())
1431 return IRB.CreatePtrToInt(V, ShadowTy);
1433 return IRB.CreateBitCast(V, ShadowTy);
1436 /// \brief Propagate shadow for arbitrary operation.
1437 void handleShadowOr(Instruction &I) {
1438 IRBuilder<> IRB(&I);
1439 ShadowAndOriginCombiner SC(this, IRB);
1440 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1445 // \brief Handle multiplication by constant.
1447 // Handle a special case of multiplication by constant that may have one or
1448 // more zeros in the lower bits. This makes corresponding number of lower bits
1449 // of the result zero as well. We model it by shifting the other operand
1450 // shadow left by the required number of bits. Effectively, we transform
1451 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1452 // We use multiplication by 2**N instead of shift to cover the case of
1453 // multiplication by 0, which may occur in some elements of a vector operand.
1454 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1456 Constant *ShadowMul;
1457 Type *Ty = ConstArg->getType();
1458 if (Ty->isVectorTy()) {
1459 unsigned NumElements = Ty->getVectorNumElements();
1460 Type *EltTy = Ty->getSequentialElementType();
1461 SmallVector<Constant *, 16> Elements;
1462 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1464 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1465 APInt V = Elt->getValue();
1466 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1467 Elements.push_back(ConstantInt::get(EltTy, V2));
1469 ShadowMul = ConstantVector::get(Elements);
1471 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1472 APInt V = Elt->getValue();
1473 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1474 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1477 IRBuilder<> IRB(&I);
1479 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1480 setOrigin(&I, getOrigin(OtherArg));
1483 void visitMul(BinaryOperator &I) {
1484 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1485 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1486 if (constOp0 && !constOp1)
1487 handleMulByConstant(I, constOp0, I.getOperand(1));
1488 else if (constOp1 && !constOp0)
1489 handleMulByConstant(I, constOp1, I.getOperand(0));
1494 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1495 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1496 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1497 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1498 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1499 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1501 void handleDiv(Instruction &I) {
1502 IRBuilder<> IRB(&I);
1503 // Strict on the second argument.
1504 insertShadowCheck(I.getOperand(1), &I);
1505 setShadow(&I, getShadow(&I, 0));
1506 setOrigin(&I, getOrigin(&I, 0));
1509 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1510 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1511 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1512 void visitURem(BinaryOperator &I) { handleDiv(I); }
1513 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1514 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1516 /// \brief Instrument == and != comparisons.
1518 /// Sometimes the comparison result is known even if some of the bits of the
1519 /// arguments are not.
1520 void handleEqualityComparison(ICmpInst &I) {
1521 IRBuilder<> IRB(&I);
1522 Value *A = I.getOperand(0);
1523 Value *B = I.getOperand(1);
1524 Value *Sa = getShadow(A);
1525 Value *Sb = getShadow(B);
1527 // Get rid of pointers and vectors of pointers.
1528 // For ints (and vectors of ints), types of A and Sa match,
1529 // and this is a no-op.
1530 A = IRB.CreatePointerCast(A, Sa->getType());
1531 B = IRB.CreatePointerCast(B, Sb->getType());
1533 // A == B <==> (C = A^B) == 0
1534 // A != B <==> (C = A^B) != 0
1536 Value *C = IRB.CreateXor(A, B);
1537 Value *Sc = IRB.CreateOr(Sa, Sb);
1538 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1539 // Result is defined if one of the following is true
1540 // * there is a defined 1 bit in C
1541 // * C is fully defined
1542 // Si = !(C & ~Sc) && Sc
1543 Value *Zero = Constant::getNullValue(Sc->getType());
1544 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1546 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1548 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1549 Si->setName("_msprop_icmp");
1551 setOriginForNaryOp(I);
1554 /// \brief Build the lowest possible value of V, taking into account V's
1555 /// uninitialized bits.
1556 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1559 // Split shadow into sign bit and other bits.
1560 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1561 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1562 // Maximise the undefined shadow bit, minimize other undefined bits.
1564 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1566 // Minimize undefined bits.
1567 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1571 /// \brief Build the highest possible value of V, taking into account V's
1572 /// uninitialized bits.
1573 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1576 // Split shadow into sign bit and other bits.
1577 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1578 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1579 // Minimise the undefined shadow bit, maximise other undefined bits.
1581 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1583 // Maximize undefined bits.
1584 return IRB.CreateOr(A, Sa);
1588 /// \brief Instrument relational comparisons.
1590 /// This function does exact shadow propagation for all relational
1591 /// comparisons of integers, pointers and vectors of those.
1592 /// FIXME: output seems suboptimal when one of the operands is a constant
1593 void handleRelationalComparisonExact(ICmpInst &I) {
1594 IRBuilder<> IRB(&I);
1595 Value *A = I.getOperand(0);
1596 Value *B = I.getOperand(1);
1597 Value *Sa = getShadow(A);
1598 Value *Sb = getShadow(B);
1600 // Get rid of pointers and vectors of pointers.
1601 // For ints (and vectors of ints), types of A and Sa match,
1602 // and this is a no-op.
1603 A = IRB.CreatePointerCast(A, Sa->getType());
1604 B = IRB.CreatePointerCast(B, Sb->getType());
1606 // Let [a0, a1] be the interval of possible values of A, taking into account
1607 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1608 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1609 bool IsSigned = I.isSigned();
1610 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1611 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1612 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1613 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1614 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1615 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1616 Value *Si = IRB.CreateXor(S1, S2);
1618 setOriginForNaryOp(I);
1621 /// \brief Instrument signed relational comparisons.
1623 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1624 /// propagating the highest bit of the shadow. Everything else is delegated
1625 /// to handleShadowOr().
1626 void handleSignedRelationalComparison(ICmpInst &I) {
1627 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1628 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1629 Value* op = nullptr;
1630 CmpInst::Predicate pre = I.getPredicate();
1631 if (constOp0 && constOp0->isNullValue() &&
1632 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1633 op = I.getOperand(1);
1634 } else if (constOp1 && constOp1->isNullValue() &&
1635 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1636 op = I.getOperand(0);
1639 IRBuilder<> IRB(&I);
1641 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1642 setShadow(&I, Shadow);
1643 setOrigin(&I, getOrigin(op));
1649 void visitICmpInst(ICmpInst &I) {
1650 if (!ClHandleICmp) {
1654 if (I.isEquality()) {
1655 handleEqualityComparison(I);
1659 assert(I.isRelational());
1660 if (ClHandleICmpExact) {
1661 handleRelationalComparisonExact(I);
1665 handleSignedRelationalComparison(I);
1669 assert(I.isUnsigned());
1670 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1671 handleRelationalComparisonExact(I);
1678 void visitFCmpInst(FCmpInst &I) {
1682 void handleShift(BinaryOperator &I) {
1683 IRBuilder<> IRB(&I);
1684 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1685 // Otherwise perform the same shift on S1.
1686 Value *S1 = getShadow(&I, 0);
1687 Value *S2 = getShadow(&I, 1);
1688 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1690 Value *V2 = I.getOperand(1);
1691 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1692 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1693 setOriginForNaryOp(I);
1696 void visitShl(BinaryOperator &I) { handleShift(I); }
1697 void visitAShr(BinaryOperator &I) { handleShift(I); }
1698 void visitLShr(BinaryOperator &I) { handleShift(I); }
1700 /// \brief Instrument llvm.memmove
1702 /// At this point we don't know if llvm.memmove will be inlined or not.
1703 /// If we don't instrument it and it gets inlined,
1704 /// our interceptor will not kick in and we will lose the memmove.
1705 /// If we instrument the call here, but it does not get inlined,
1706 /// we will memove the shadow twice: which is bad in case
1707 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1709 /// Similar situation exists for memcpy and memset.
1710 void visitMemMoveInst(MemMoveInst &I) {
1711 IRBuilder<> IRB(&I);
1714 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1715 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1716 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1717 I.eraseFromParent();
1720 // Similar to memmove: avoid copying shadow twice.
1721 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1722 // FIXME: consider doing manual inline for small constant sizes and proper
1724 void visitMemCpyInst(MemCpyInst &I) {
1725 IRBuilder<> IRB(&I);
1728 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1729 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1730 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1731 I.eraseFromParent();
1735 void visitMemSetInst(MemSetInst &I) {
1736 IRBuilder<> IRB(&I);
1739 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1740 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1741 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1742 I.eraseFromParent();
1745 void visitVAStartInst(VAStartInst &I) {
1746 VAHelper->visitVAStartInst(I);
1749 void visitVACopyInst(VACopyInst &I) {
1750 VAHelper->visitVACopyInst(I);
1753 enum IntrinsicKind {
1754 IK_DoesNotAccessMemory,
1759 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1760 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1761 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1762 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1763 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1764 const int UnknownModRefBehavior = IK_WritesMemory;
1765 #define GET_INTRINSIC_MODREF_BEHAVIOR
1766 #define ModRefBehavior IntrinsicKind
1767 #include "llvm/IR/Intrinsics.gen"
1768 #undef ModRefBehavior
1769 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1772 /// \brief Handle vector store-like intrinsics.
1774 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1775 /// has 1 pointer argument and 1 vector argument, returns void.
1776 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1777 IRBuilder<> IRB(&I);
1778 Value* Addr = I.getArgOperand(0);
1779 Value *Shadow = getShadow(&I, 1);
1780 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1782 // We don't know the pointer alignment (could be unaligned SSE store!).
1783 // Have to assume to worst case.
1784 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1786 if (ClCheckAccessAddress)
1787 insertShadowCheck(Addr, &I);
1789 // FIXME: use ClStoreCleanOrigin
1790 // FIXME: factor out common code from materializeStores
1791 if (MS.TrackOrigins)
1792 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1796 /// \brief Handle vector load-like intrinsics.
1798 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1799 /// has 1 pointer argument, returns a vector.
1800 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1801 IRBuilder<> IRB(&I);
1802 Value *Addr = I.getArgOperand(0);
1804 Type *ShadowTy = getShadowTy(&I);
1805 if (PropagateShadow) {
1806 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1807 // We don't know the pointer alignment (could be unaligned SSE load!).
1808 // Have to assume to worst case.
1809 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1811 setShadow(&I, getCleanShadow(&I));
1814 if (ClCheckAccessAddress)
1815 insertShadowCheck(Addr, &I);
1817 if (MS.TrackOrigins) {
1818 if (PropagateShadow)
1819 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1821 setOrigin(&I, getCleanOrigin());
1826 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1828 /// Instrument intrinsics with any number of arguments of the same type,
1829 /// equal to the return type. The type should be simple (no aggregates or
1830 /// pointers; vectors are fine).
1831 /// Caller guarantees that this intrinsic does not access memory.
1832 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1833 Type *RetTy = I.getType();
1834 if (!(RetTy->isIntOrIntVectorTy() ||
1835 RetTy->isFPOrFPVectorTy() ||
1836 RetTy->isX86_MMXTy()))
1839 unsigned NumArgOperands = I.getNumArgOperands();
1841 for (unsigned i = 0; i < NumArgOperands; ++i) {
1842 Type *Ty = I.getArgOperand(i)->getType();
1847 IRBuilder<> IRB(&I);
1848 ShadowAndOriginCombiner SC(this, IRB);
1849 for (unsigned i = 0; i < NumArgOperands; ++i)
1850 SC.Add(I.getArgOperand(i));
1856 /// \brief Heuristically instrument unknown intrinsics.
1858 /// The main purpose of this code is to do something reasonable with all
1859 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1860 /// We recognize several classes of intrinsics by their argument types and
1861 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1862 /// sure that we know what the intrinsic does.
1864 /// We special-case intrinsics where this approach fails. See llvm.bswap
1865 /// handling as an example of that.
1866 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1867 unsigned NumArgOperands = I.getNumArgOperands();
1868 if (NumArgOperands == 0)
1871 Intrinsic::ID iid = I.getIntrinsicID();
1872 IntrinsicKind IK = getIntrinsicKind(iid);
1873 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1874 bool WritesMemory = IK == IK_WritesMemory;
1875 assert(!(OnlyReadsMemory && WritesMemory));
1877 if (NumArgOperands == 2 &&
1878 I.getArgOperand(0)->getType()->isPointerTy() &&
1879 I.getArgOperand(1)->getType()->isVectorTy() &&
1880 I.getType()->isVoidTy() &&
1882 // This looks like a vector store.
1883 return handleVectorStoreIntrinsic(I);
1886 if (NumArgOperands == 1 &&
1887 I.getArgOperand(0)->getType()->isPointerTy() &&
1888 I.getType()->isVectorTy() &&
1890 // This looks like a vector load.
1891 return handleVectorLoadIntrinsic(I);
1894 if (!OnlyReadsMemory && !WritesMemory)
1895 if (maybeHandleSimpleNomemIntrinsic(I))
1898 // FIXME: detect and handle SSE maskstore/maskload
1902 void handleBswap(IntrinsicInst &I) {
1903 IRBuilder<> IRB(&I);
1904 Value *Op = I.getArgOperand(0);
1905 Type *OpType = Op->getType();
1906 Function *BswapFunc = Intrinsic::getDeclaration(
1907 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
1908 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1909 setOrigin(&I, getOrigin(Op));
1912 // \brief Instrument vector convert instrinsic.
1914 // This function instruments intrinsics like cvtsi2ss:
1915 // %Out = int_xxx_cvtyyy(%ConvertOp)
1917 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1918 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1919 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1920 // elements from \p CopyOp.
1921 // In most cases conversion involves floating-point value which may trigger a
1922 // hardware exception when not fully initialized. For this reason we require
1923 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1924 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1925 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1926 // return a fully initialized value.
1927 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1928 IRBuilder<> IRB(&I);
1929 Value *CopyOp, *ConvertOp;
1931 switch (I.getNumArgOperands()) {
1933 CopyOp = I.getArgOperand(0);
1934 ConvertOp = I.getArgOperand(1);
1937 ConvertOp = I.getArgOperand(0);
1941 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1944 // The first *NumUsedElements* elements of ConvertOp are converted to the
1945 // same number of output elements. The rest of the output is copied from
1946 // CopyOp, or (if not available) filled with zeroes.
1947 // Combine shadow for elements of ConvertOp that are used in this operation,
1948 // and insert a check.
1949 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1950 // int->any conversion.
1951 Value *ConvertShadow = getShadow(ConvertOp);
1952 Value *AggShadow = nullptr;
1953 if (ConvertOp->getType()->isVectorTy()) {
1954 AggShadow = IRB.CreateExtractElement(
1955 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1956 for (int i = 1; i < NumUsedElements; ++i) {
1957 Value *MoreShadow = IRB.CreateExtractElement(
1958 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1959 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1962 AggShadow = ConvertShadow;
1964 assert(AggShadow->getType()->isIntegerTy());
1965 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1967 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1970 assert(CopyOp->getType() == I.getType());
1971 assert(CopyOp->getType()->isVectorTy());
1972 Value *ResultShadow = getShadow(CopyOp);
1973 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1974 for (int i = 0; i < NumUsedElements; ++i) {
1975 ResultShadow = IRB.CreateInsertElement(
1976 ResultShadow, ConstantInt::getNullValue(EltTy),
1977 ConstantInt::get(IRB.getInt32Ty(), i));
1979 setShadow(&I, ResultShadow);
1980 setOrigin(&I, getOrigin(CopyOp));
1982 setShadow(&I, getCleanShadow(&I));
1986 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1987 // zeroes if it is zero, and all ones otherwise.
1988 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1989 if (S->getType()->isVectorTy())
1990 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1991 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1992 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1993 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1996 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1997 Type *T = S->getType();
1998 assert(T->isVectorTy());
1999 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2000 return IRB.CreateSExt(S2, T);
2003 // \brief Instrument vector shift instrinsic.
2005 // This function instruments intrinsics like int_x86_avx2_psll_w.
2006 // Intrinsic shifts %In by %ShiftSize bits.
2007 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2008 // size, and the rest is ignored. Behavior is defined even if shift size is
2009 // greater than register (or field) width.
2010 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2011 assert(I.getNumArgOperands() == 2);
2012 IRBuilder<> IRB(&I);
2013 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2014 // Otherwise perform the same shift on S1.
2015 Value *S1 = getShadow(&I, 0);
2016 Value *S2 = getShadow(&I, 1);
2017 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2018 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2019 Value *V1 = I.getOperand(0);
2020 Value *V2 = I.getOperand(1);
2021 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
2022 IRB.CreateBitCast(S1, V1->getType()), V2);
2023 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2024 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2025 setOriginForNaryOp(I);
2028 // \brief Get an X86_MMX-sized vector type.
2029 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2030 const unsigned X86_MMXSizeInBits = 64;
2031 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2032 X86_MMXSizeInBits / EltSizeInBits);
2035 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2037 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2039 case llvm::Intrinsic::x86_sse2_packsswb_128:
2040 case llvm::Intrinsic::x86_sse2_packuswb_128:
2041 return llvm::Intrinsic::x86_sse2_packsswb_128;
2043 case llvm::Intrinsic::x86_sse2_packssdw_128:
2044 case llvm::Intrinsic::x86_sse41_packusdw:
2045 return llvm::Intrinsic::x86_sse2_packssdw_128;
2047 case llvm::Intrinsic::x86_avx2_packsswb:
2048 case llvm::Intrinsic::x86_avx2_packuswb:
2049 return llvm::Intrinsic::x86_avx2_packsswb;
2051 case llvm::Intrinsic::x86_avx2_packssdw:
2052 case llvm::Intrinsic::x86_avx2_packusdw:
2053 return llvm::Intrinsic::x86_avx2_packssdw;
2055 case llvm::Intrinsic::x86_mmx_packsswb:
2056 case llvm::Intrinsic::x86_mmx_packuswb:
2057 return llvm::Intrinsic::x86_mmx_packsswb;
2059 case llvm::Intrinsic::x86_mmx_packssdw:
2060 return llvm::Intrinsic::x86_mmx_packssdw;
2062 llvm_unreachable("unexpected intrinsic id");
2066 // \brief Instrument vector pack instrinsic.
2068 // This function instruments intrinsics like x86_mmx_packsswb, that
2069 // packs elements of 2 input vectors into half as many bits with saturation.
2070 // Shadow is propagated with the signed variant of the same intrinsic applied
2071 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2072 // EltSizeInBits is used only for x86mmx arguments.
2073 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2074 assert(I.getNumArgOperands() == 2);
2075 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2076 IRBuilder<> IRB(&I);
2077 Value *S1 = getShadow(&I, 0);
2078 Value *S2 = getShadow(&I, 1);
2079 assert(isX86_MMX || S1->getType()->isVectorTy());
2081 // SExt and ICmpNE below must apply to individual elements of input vectors.
2082 // In case of x86mmx arguments, cast them to appropriate vector types and
2084 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2086 S1 = IRB.CreateBitCast(S1, T);
2087 S2 = IRB.CreateBitCast(S2, T);
2089 Value *S1_ext = IRB.CreateSExt(
2090 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2091 Value *S2_ext = IRB.CreateSExt(
2092 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2094 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2095 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2096 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2099 Function *ShadowFn = Intrinsic::getDeclaration(
2100 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2102 Value *S = IRB.CreateCall2(ShadowFn, S1_ext, S2_ext, "_msprop_vector_pack");
2103 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2105 setOriginForNaryOp(I);
2108 // \brief Instrument sum-of-absolute-differencies intrinsic.
2109 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2110 const unsigned SignificantBitsPerResultElement = 16;
2111 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2112 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2113 unsigned ZeroBitsPerResultElement =
2114 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2116 IRBuilder<> IRB(&I);
2117 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2118 S = IRB.CreateBitCast(S, ResTy);
2119 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2121 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2122 S = IRB.CreateBitCast(S, getShadowTy(&I));
2124 setOriginForNaryOp(I);
2127 // \brief Instrument multiply-add intrinsic.
2128 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2129 unsigned EltSizeInBits = 0) {
2130 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2131 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2132 IRBuilder<> IRB(&I);
2133 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2134 S = IRB.CreateBitCast(S, ResTy);
2135 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2137 S = IRB.CreateBitCast(S, getShadowTy(&I));
2139 setOriginForNaryOp(I);
2142 void visitIntrinsicInst(IntrinsicInst &I) {
2143 switch (I.getIntrinsicID()) {
2144 case llvm::Intrinsic::bswap:
2147 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2148 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2149 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2150 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2151 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2152 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2153 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2154 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2155 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2156 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2157 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2158 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2159 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2160 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2161 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2162 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2163 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2164 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2165 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2166 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2167 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2168 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2169 case llvm::Intrinsic::x86_sse_cvtss2si64:
2170 case llvm::Intrinsic::x86_sse_cvtss2si:
2171 case llvm::Intrinsic::x86_sse_cvttss2si64:
2172 case llvm::Intrinsic::x86_sse_cvttss2si:
2173 handleVectorConvertIntrinsic(I, 1);
2175 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2176 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2177 case llvm::Intrinsic::x86_sse_cvtps2pi:
2178 case llvm::Intrinsic::x86_sse_cvttps2pi:
2179 handleVectorConvertIntrinsic(I, 2);
2181 case llvm::Intrinsic::x86_avx512_psll_dq:
2182 case llvm::Intrinsic::x86_avx512_psrl_dq:
2183 case llvm::Intrinsic::x86_avx2_psll_w:
2184 case llvm::Intrinsic::x86_avx2_psll_d:
2185 case llvm::Intrinsic::x86_avx2_psll_q:
2186 case llvm::Intrinsic::x86_avx2_pslli_w:
2187 case llvm::Intrinsic::x86_avx2_pslli_d:
2188 case llvm::Intrinsic::x86_avx2_pslli_q:
2189 case llvm::Intrinsic::x86_avx2_psll_dq:
2190 case llvm::Intrinsic::x86_avx2_psrl_w:
2191 case llvm::Intrinsic::x86_avx2_psrl_d:
2192 case llvm::Intrinsic::x86_avx2_psrl_q:
2193 case llvm::Intrinsic::x86_avx2_psra_w:
2194 case llvm::Intrinsic::x86_avx2_psra_d:
2195 case llvm::Intrinsic::x86_avx2_psrli_w:
2196 case llvm::Intrinsic::x86_avx2_psrli_d:
2197 case llvm::Intrinsic::x86_avx2_psrli_q:
2198 case llvm::Intrinsic::x86_avx2_psrai_w:
2199 case llvm::Intrinsic::x86_avx2_psrai_d:
2200 case llvm::Intrinsic::x86_avx2_psrl_dq:
2201 case llvm::Intrinsic::x86_sse2_psll_w:
2202 case llvm::Intrinsic::x86_sse2_psll_d:
2203 case llvm::Intrinsic::x86_sse2_psll_q:
2204 case llvm::Intrinsic::x86_sse2_pslli_w:
2205 case llvm::Intrinsic::x86_sse2_pslli_d:
2206 case llvm::Intrinsic::x86_sse2_pslli_q:
2207 case llvm::Intrinsic::x86_sse2_psll_dq:
2208 case llvm::Intrinsic::x86_sse2_psrl_w:
2209 case llvm::Intrinsic::x86_sse2_psrl_d:
2210 case llvm::Intrinsic::x86_sse2_psrl_q:
2211 case llvm::Intrinsic::x86_sse2_psra_w:
2212 case llvm::Intrinsic::x86_sse2_psra_d:
2213 case llvm::Intrinsic::x86_sse2_psrli_w:
2214 case llvm::Intrinsic::x86_sse2_psrli_d:
2215 case llvm::Intrinsic::x86_sse2_psrli_q:
2216 case llvm::Intrinsic::x86_sse2_psrai_w:
2217 case llvm::Intrinsic::x86_sse2_psrai_d:
2218 case llvm::Intrinsic::x86_sse2_psrl_dq:
2219 case llvm::Intrinsic::x86_mmx_psll_w:
2220 case llvm::Intrinsic::x86_mmx_psll_d:
2221 case llvm::Intrinsic::x86_mmx_psll_q:
2222 case llvm::Intrinsic::x86_mmx_pslli_w:
2223 case llvm::Intrinsic::x86_mmx_pslli_d:
2224 case llvm::Intrinsic::x86_mmx_pslli_q:
2225 case llvm::Intrinsic::x86_mmx_psrl_w:
2226 case llvm::Intrinsic::x86_mmx_psrl_d:
2227 case llvm::Intrinsic::x86_mmx_psrl_q:
2228 case llvm::Intrinsic::x86_mmx_psra_w:
2229 case llvm::Intrinsic::x86_mmx_psra_d:
2230 case llvm::Intrinsic::x86_mmx_psrli_w:
2231 case llvm::Intrinsic::x86_mmx_psrli_d:
2232 case llvm::Intrinsic::x86_mmx_psrli_q:
2233 case llvm::Intrinsic::x86_mmx_psrai_w:
2234 case llvm::Intrinsic::x86_mmx_psrai_d:
2235 handleVectorShiftIntrinsic(I, /* Variable */ false);
2237 case llvm::Intrinsic::x86_avx2_psllv_d:
2238 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2239 case llvm::Intrinsic::x86_avx2_psllv_q:
2240 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2241 case llvm::Intrinsic::x86_avx2_psrlv_d:
2242 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2243 case llvm::Intrinsic::x86_avx2_psrlv_q:
2244 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2245 case llvm::Intrinsic::x86_avx2_psrav_d:
2246 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2247 handleVectorShiftIntrinsic(I, /* Variable */ true);
2250 // Byte shifts are not implemented.
2251 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
2252 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
2253 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
2254 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
2255 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
2256 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
2258 case llvm::Intrinsic::x86_sse2_packsswb_128:
2259 case llvm::Intrinsic::x86_sse2_packssdw_128:
2260 case llvm::Intrinsic::x86_sse2_packuswb_128:
2261 case llvm::Intrinsic::x86_sse41_packusdw:
2262 case llvm::Intrinsic::x86_avx2_packsswb:
2263 case llvm::Intrinsic::x86_avx2_packssdw:
2264 case llvm::Intrinsic::x86_avx2_packuswb:
2265 case llvm::Intrinsic::x86_avx2_packusdw:
2266 handleVectorPackIntrinsic(I);
2269 case llvm::Intrinsic::x86_mmx_packsswb:
2270 case llvm::Intrinsic::x86_mmx_packuswb:
2271 handleVectorPackIntrinsic(I, 16);
2274 case llvm::Intrinsic::x86_mmx_packssdw:
2275 handleVectorPackIntrinsic(I, 32);
2278 case llvm::Intrinsic::x86_mmx_psad_bw:
2279 case llvm::Intrinsic::x86_sse2_psad_bw:
2280 case llvm::Intrinsic::x86_avx2_psad_bw:
2281 handleVectorSadIntrinsic(I);
2284 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2285 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2286 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2287 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2288 handleVectorPmaddIntrinsic(I);
2291 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2292 handleVectorPmaddIntrinsic(I, 8);
2295 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2296 handleVectorPmaddIntrinsic(I, 16);
2300 if (!handleUnknownIntrinsic(I))
2301 visitInstruction(I);
2306 void visitCallSite(CallSite CS) {
2307 Instruction &I = *CS.getInstruction();
2308 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2310 CallInst *Call = cast<CallInst>(&I);
2312 // For inline asm, do the usual thing: check argument shadow and mark all
2313 // outputs as clean. Note that any side effects of the inline asm that are
2314 // not immediately visible in its constraints are not handled.
2315 if (Call->isInlineAsm()) {
2316 visitInstruction(I);
2320 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2322 // We are going to insert code that relies on the fact that the callee
2323 // will become a non-readonly function after it is instrumented by us. To
2324 // prevent this code from being optimized out, mark that function
2325 // non-readonly in advance.
2326 if (Function *Func = Call->getCalledFunction()) {
2327 // Clear out readonly/readnone attributes.
2329 B.addAttribute(Attribute::ReadOnly)
2330 .addAttribute(Attribute::ReadNone);
2331 Func->removeAttributes(AttributeSet::FunctionIndex,
2332 AttributeSet::get(Func->getContext(),
2333 AttributeSet::FunctionIndex,
2337 IRBuilder<> IRB(&I);
2339 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2340 IndirectCallList.push_back(CS);
2342 unsigned ArgOffset = 0;
2343 DEBUG(dbgs() << " CallSite: " << I << "\n");
2344 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2345 ArgIt != End; ++ArgIt) {
2347 unsigned i = ArgIt - CS.arg_begin();
2348 if (!A->getType()->isSized()) {
2349 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2353 Value *Store = nullptr;
2354 // Compute the Shadow for arg even if it is ByVal, because
2355 // in that case getShadow() will copy the actual arg shadow to
2356 // __msan_param_tls.
2357 Value *ArgShadow = getShadow(A);
2358 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2359 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2360 " Shadow: " << *ArgShadow << "\n");
2361 bool ArgIsInitialized = false;
2362 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2363 assert(A->getType()->isPointerTy() &&
2364 "ByVal argument is not a pointer!");
2365 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2366 if (ArgOffset + Size > kParamTLSSize) break;
2367 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2368 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2369 Store = IRB.CreateMemCpy(ArgShadowBase,
2370 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2373 Size = MS.DL->getTypeAllocSize(A->getType());
2374 if (ArgOffset + Size > kParamTLSSize) break;
2375 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2376 kShadowTLSAlignment);
2377 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2378 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2380 if (MS.TrackOrigins && !ArgIsInitialized)
2381 IRB.CreateStore(getOrigin(A),
2382 getOriginPtrForArgument(A, IRB, ArgOffset));
2384 assert(Size != 0 && Store != nullptr);
2385 DEBUG(dbgs() << " Param:" << *Store << "\n");
2386 ArgOffset += RoundUpToAlignment(Size, 8);
2388 DEBUG(dbgs() << " done with call args\n");
2391 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2392 if (FT->isVarArg()) {
2393 VAHelper->visitCallSite(CS, IRB);
2396 // Now, get the shadow for the RetVal.
2397 if (!I.getType()->isSized()) return;
2398 IRBuilder<> IRBBefore(&I);
2399 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2400 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2401 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2402 Instruction *NextInsn = nullptr;
2404 NextInsn = I.getNextNode();
2406 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2407 if (!NormalDest->getSinglePredecessor()) {
2408 // FIXME: this case is tricky, so we are just conservative here.
2409 // Perhaps we need to split the edge between this BB and NormalDest,
2410 // but a naive attempt to use SplitEdge leads to a crash.
2411 setShadow(&I, getCleanShadow(&I));
2412 setOrigin(&I, getCleanOrigin());
2415 NextInsn = NormalDest->getFirstInsertionPt();
2417 "Could not find insertion point for retval shadow load");
2419 IRBuilder<> IRBAfter(NextInsn);
2420 Value *RetvalShadow =
2421 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2422 kShadowTLSAlignment, "_msret");
2423 setShadow(&I, RetvalShadow);
2424 if (MS.TrackOrigins)
2425 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2428 void visitReturnInst(ReturnInst &I) {
2429 IRBuilder<> IRB(&I);
2430 Value *RetVal = I.getReturnValue();
2431 if (!RetVal) return;
2432 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2433 if (CheckReturnValue) {
2434 insertShadowCheck(RetVal, &I);
2435 Value *Shadow = getCleanShadow(RetVal);
2436 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2438 Value *Shadow = getShadow(RetVal);
2439 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2440 // FIXME: make it conditional if ClStoreCleanOrigin==0
2441 if (MS.TrackOrigins)
2442 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2446 void visitPHINode(PHINode &I) {
2447 IRBuilder<> IRB(&I);
2448 if (!PropagateShadow) {
2449 setShadow(&I, getCleanShadow(&I));
2453 ShadowPHINodes.push_back(&I);
2454 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2456 if (MS.TrackOrigins)
2457 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2461 void visitAllocaInst(AllocaInst &I) {
2462 setShadow(&I, getCleanShadow(&I));
2463 IRBuilder<> IRB(I.getNextNode());
2464 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2465 if (PoisonStack && ClPoisonStackWithCall) {
2466 IRB.CreateCall2(MS.MsanPoisonStackFn,
2467 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2468 ConstantInt::get(MS.IntptrTy, Size));
2470 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2471 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2472 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2475 if (PoisonStack && MS.TrackOrigins) {
2476 setOrigin(&I, getCleanOrigin());
2477 SmallString<2048> StackDescriptionStorage;
2478 raw_svector_ostream StackDescription(StackDescriptionStorage);
2479 // We create a string with a description of the stack allocation and
2480 // pass it into __msan_set_alloca_origin.
2481 // It will be printed by the run-time if stack-originated UMR is found.
2482 // The first 4 bytes of the string are set to '----' and will be replaced
2483 // by __msan_va_arg_overflow_size_tls at the first call.
2484 StackDescription << "----" << I.getName() << "@" << F.getName();
2486 createPrivateNonConstGlobalForString(*F.getParent(),
2487 StackDescription.str());
2489 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2490 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2491 ConstantInt::get(MS.IntptrTy, Size),
2492 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2493 IRB.CreatePointerCast(&F, MS.IntptrTy));
2497 void visitSelectInst(SelectInst& I) {
2498 IRBuilder<> IRB(&I);
2499 // a = select b, c, d
2500 Value *B = I.getCondition();
2501 Value *C = I.getTrueValue();
2502 Value *D = I.getFalseValue();
2503 Value *Sb = getShadow(B);
2504 Value *Sc = getShadow(C);
2505 Value *Sd = getShadow(D);
2507 // Result shadow if condition shadow is 0.
2508 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2510 if (I.getType()->isAggregateType()) {
2511 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2512 // an extra "select". This results in much more compact IR.
2513 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2514 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2516 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2517 // If Sb (condition is poisoned), look for bits in c and d that are equal
2518 // and both unpoisoned.
2519 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2521 // Cast arguments to shadow-compatible type.
2522 C = CreateAppToShadowCast(IRB, C);
2523 D = CreateAppToShadowCast(IRB, D);
2525 // Result shadow if condition shadow is 1.
2526 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2528 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2530 if (MS.TrackOrigins) {
2531 // Origins are always i32, so any vector conditions must be flattened.
2532 // FIXME: consider tracking vector origins for app vectors?
2533 if (B->getType()->isVectorTy()) {
2534 Type *FlatTy = getShadowTyNoVec(B->getType());
2535 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2536 ConstantInt::getNullValue(FlatTy));
2537 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2538 ConstantInt::getNullValue(FlatTy));
2540 // a = select b, c, d
2541 // Oa = Sb ? Ob : (b ? Oc : Od)
2542 setOrigin(&I, IRB.CreateSelect(
2543 Sb, getOrigin(I.getCondition()),
2544 IRB.CreateSelect(B, getOrigin(C), getOrigin(D))));
2548 void visitLandingPadInst(LandingPadInst &I) {
2550 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2551 setShadow(&I, getCleanShadow(&I));
2552 setOrigin(&I, getCleanOrigin());
2555 void visitGetElementPtrInst(GetElementPtrInst &I) {
2559 void visitExtractValueInst(ExtractValueInst &I) {
2560 IRBuilder<> IRB(&I);
2561 Value *Agg = I.getAggregateOperand();
2562 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2563 Value *AggShadow = getShadow(Agg);
2564 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2565 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2566 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2567 setShadow(&I, ResShadow);
2568 setOriginForNaryOp(I);
2571 void visitInsertValueInst(InsertValueInst &I) {
2572 IRBuilder<> IRB(&I);
2573 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2574 Value *AggShadow = getShadow(I.getAggregateOperand());
2575 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2576 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2577 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2578 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2579 DEBUG(dbgs() << " Res: " << *Res << "\n");
2581 setOriginForNaryOp(I);
2584 void dumpInst(Instruction &I) {
2585 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2586 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2588 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2590 errs() << "QQQ " << I << "\n";
2593 void visitResumeInst(ResumeInst &I) {
2594 DEBUG(dbgs() << "Resume: " << I << "\n");
2595 // Nothing to do here.
2598 void visitInstruction(Instruction &I) {
2599 // Everything else: stop propagating and check for poisoned shadow.
2600 if (ClDumpStrictInstructions)
2602 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2603 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2604 insertShadowCheck(I.getOperand(i), &I);
2605 setShadow(&I, getCleanShadow(&I));
2606 setOrigin(&I, getCleanOrigin());
2610 /// \brief AMD64-specific implementation of VarArgHelper.
2611 struct VarArgAMD64Helper : public VarArgHelper {
2612 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2613 // See a comment in visitCallSite for more details.
2614 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2615 static const unsigned AMD64FpEndOffset = 176;
2618 MemorySanitizer &MS;
2619 MemorySanitizerVisitor &MSV;
2620 Value *VAArgTLSCopy;
2621 Value *VAArgOverflowSize;
2623 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2625 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2626 MemorySanitizerVisitor &MSV)
2627 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2628 VAArgOverflowSize(nullptr) {}
2630 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2632 ArgKind classifyArgument(Value* arg) {
2633 // A very rough approximation of X86_64 argument classification rules.
2634 Type *T = arg->getType();
2635 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2636 return AK_FloatingPoint;
2637 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2638 return AK_GeneralPurpose;
2639 if (T->isPointerTy())
2640 return AK_GeneralPurpose;
2644 // For VarArg functions, store the argument shadow in an ABI-specific format
2645 // that corresponds to va_list layout.
2646 // We do this because Clang lowers va_arg in the frontend, and this pass
2647 // only sees the low level code that deals with va_list internals.
2648 // A much easier alternative (provided that Clang emits va_arg instructions)
2649 // would have been to associate each live instance of va_list with a copy of
2650 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2652 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2653 unsigned GpOffset = 0;
2654 unsigned FpOffset = AMD64GpEndOffset;
2655 unsigned OverflowOffset = AMD64FpEndOffset;
2656 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2657 ArgIt != End; ++ArgIt) {
2659 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2660 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2662 // ByVal arguments always go to the overflow area.
2663 assert(A->getType()->isPointerTy());
2664 Type *RealTy = A->getType()->getPointerElementType();
2665 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2666 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2667 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2668 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2669 ArgSize, kShadowTLSAlignment);
2671 ArgKind AK = classifyArgument(A);
2672 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2674 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2678 case AK_GeneralPurpose:
2679 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2682 case AK_FloatingPoint:
2683 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2687 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2688 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2689 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2691 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2694 Constant *OverflowSize =
2695 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2696 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2699 /// \brief Compute the shadow address for a given va_arg.
2700 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2702 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2703 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2704 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2708 void visitVAStartInst(VAStartInst &I) override {
2709 IRBuilder<> IRB(&I);
2710 VAStartInstrumentationList.push_back(&I);
2711 Value *VAListTag = I.getArgOperand(0);
2712 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2714 // Unpoison the whole __va_list_tag.
2715 // FIXME: magic ABI constants.
2716 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2717 /* size */24, /* alignment */8, false);
2720 void visitVACopyInst(VACopyInst &I) override {
2721 IRBuilder<> IRB(&I);
2722 Value *VAListTag = I.getArgOperand(0);
2723 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2725 // Unpoison the whole __va_list_tag.
2726 // FIXME: magic ABI constants.
2727 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2728 /* size */24, /* alignment */8, false);
2731 void finalizeInstrumentation() override {
2732 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2733 "finalizeInstrumentation called twice");
2734 if (!VAStartInstrumentationList.empty()) {
2735 // If there is a va_start in this function, make a backup copy of
2736 // va_arg_tls somewhere in the function entry block.
2737 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2738 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2740 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2742 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2743 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2746 // Instrument va_start.
2747 // Copy va_list shadow from the backup copy of the TLS contents.
2748 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2749 CallInst *OrigInst = VAStartInstrumentationList[i];
2750 IRBuilder<> IRB(OrigInst->getNextNode());
2751 Value *VAListTag = OrigInst->getArgOperand(0);
2753 Value *RegSaveAreaPtrPtr =
2755 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2756 ConstantInt::get(MS.IntptrTy, 16)),
2757 Type::getInt64PtrTy(*MS.C));
2758 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2759 Value *RegSaveAreaShadowPtr =
2760 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2761 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2762 AMD64FpEndOffset, 16);
2764 Value *OverflowArgAreaPtrPtr =
2766 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2767 ConstantInt::get(MS.IntptrTy, 8)),
2768 Type::getInt64PtrTy(*MS.C));
2769 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2770 Value *OverflowArgAreaShadowPtr =
2771 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2772 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2773 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2778 /// \brief A no-op implementation of VarArgHelper.
2779 struct VarArgNoOpHelper : public VarArgHelper {
2780 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2781 MemorySanitizerVisitor &MSV) {}
2783 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2785 void visitVAStartInst(VAStartInst &I) override {}
2787 void visitVACopyInst(VACopyInst &I) override {}
2789 void finalizeInstrumentation() override {}
2792 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2793 MemorySanitizerVisitor &Visitor) {
2794 // VarArg handling is only implemented on AMD64. False positives are possible
2795 // on other platforms.
2796 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2797 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2798 return new VarArgAMD64Helper(Func, Msan, Visitor);
2800 return new VarArgNoOpHelper(Func, Msan, Visitor);
2805 bool MemorySanitizer::runOnFunction(Function &F) {
2806 MemorySanitizerVisitor Visitor(F, *this);
2808 // Clear out readonly/readnone attributes.
2810 B.addAttribute(Attribute::ReadOnly)
2811 .addAttribute(Attribute::ReadNone);
2812 F.removeAttributes(AttributeSet::FunctionIndex,
2813 AttributeSet::get(F.getContext(),
2814 AttributeSet::FunctionIndex, B));
2816 return Visitor.runOnFunction();