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 unsigned kMinOriginAlignment = 4;
124 static const unsigned kShadowTLSAlignment = 8;
126 // These constants must be kept in sync with the ones in msan.h.
127 static const unsigned kParamTLSSize = 800;
128 static const unsigned kRetvalTLSSize = 800;
130 // Accesses sizes are powers of two: 1, 2, 4, 8.
131 static const size_t kNumberOfAccessSizes = 4;
133 /// \brief Track origins of uninitialized values.
135 /// Adds a section to MemorySanitizer report that points to the allocation
136 /// (stack or heap) the uninitialized bits came from originally.
137 static cl::opt<int> ClTrackOrigins("msan-track-origins",
138 cl::desc("Track origins (allocation sites) of poisoned memory"),
139 cl::Hidden, cl::init(0));
140 static cl::opt<bool> ClKeepGoing("msan-keep-going",
141 cl::desc("keep going after reporting a UMR"),
142 cl::Hidden, cl::init(false));
143 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
144 cl::desc("poison uninitialized stack variables"),
145 cl::Hidden, cl::init(true));
146 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
147 cl::desc("poison uninitialized stack variables with a call"),
148 cl::Hidden, cl::init(false));
149 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
150 cl::desc("poison uninitialized stack variables with the given patter"),
151 cl::Hidden, cl::init(0xff));
152 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
153 cl::desc("poison undef temps"),
154 cl::Hidden, cl::init(true));
156 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
157 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
158 cl::Hidden, cl::init(true));
160 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
161 cl::desc("exact handling of relational integer ICmp"),
162 cl::Hidden, cl::init(false));
164 // This flag controls whether we check the shadow of the address
165 // operand of load or store. Such bugs are very rare, since load from
166 // a garbage address typically results in SEGV, but still happen
167 // (e.g. only lower bits of address are garbage, or the access happens
168 // early at program startup where malloc-ed memory is more likely to
169 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
170 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
171 cl::desc("report accesses through a pointer which has poisoned shadow"),
172 cl::Hidden, cl::init(true));
174 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
175 cl::desc("print out instructions with default strict semantics"),
176 cl::Hidden, cl::init(false));
178 static cl::opt<int> ClInstrumentationWithCallThreshold(
179 "msan-instrumentation-with-call-threshold",
181 "If the function being instrumented requires more than "
182 "this number of checks and origin stores, use callbacks instead of "
183 "inline checks (-1 means never use callbacks)."),
184 cl::Hidden, cl::init(3500));
186 // This is an experiment to enable handling of cases where shadow is a non-zero
187 // compile-time constant. For some unexplainable reason they were silently
188 // ignored in the instrumentation.
189 static cl::opt<bool> ClCheckConstantShadow("msan-check-constant-shadow",
190 cl::desc("Insert checks for constant shadow values"),
191 cl::Hidden, cl::init(false));
195 // Memory map parameters used in application-to-shadow address calculation.
196 // Offset = (Addr & ~AndMask) ^ XorMask
197 // Shadow = ShadowBase + Offset
198 // Origin = OriginBase + Offset
199 struct MemoryMapParams {
206 struct PlatformMemoryMapParams {
207 const MemoryMapParams *bits32;
208 const MemoryMapParams *bits64;
212 static const MemoryMapParams Linux_I386_MemoryMapParams = {
213 0x000080000000, // AndMask
214 0, // XorMask (not used)
215 0, // ShadowBase (not used)
216 0x000040000000, // OriginBase
220 static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
221 0x400000000000, // AndMask
222 0, // XorMask (not used)
223 0, // ShadowBase (not used)
224 0x200000000000, // OriginBase
228 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
229 0x004000000000, // AndMask
230 0, // XorMask (not used)
231 0, // ShadowBase (not used)
232 0x002000000000, // OriginBase
236 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
237 0x000180000000, // AndMask
238 0x000040000000, // XorMask
239 0x000020000000, // ShadowBase
240 0x000700000000, // OriginBase
244 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
245 0xc00000000000, // AndMask
246 0x200000000000, // XorMask
247 0x100000000000, // ShadowBase
248 0x380000000000, // OriginBase
251 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
252 &Linux_I386_MemoryMapParams,
253 &Linux_X86_64_MemoryMapParams,
256 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
258 &Linux_MIPS64_MemoryMapParams,
261 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
262 &FreeBSD_I386_MemoryMapParams,
263 &FreeBSD_X86_64_MemoryMapParams,
266 /// \brief An instrumentation pass implementing detection of uninitialized
269 /// MemorySanitizer: instrument the code in module to find
270 /// uninitialized reads.
271 class MemorySanitizer : public FunctionPass {
273 MemorySanitizer(int TrackOrigins = 0)
275 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
277 WarningFn(nullptr) {}
278 const char *getPassName() const override { return "MemorySanitizer"; }
279 bool runOnFunction(Function &F) override;
280 bool doInitialization(Module &M) override;
281 static char ID; // Pass identification, replacement for typeid.
284 void initializeCallbacks(Module &M);
286 /// \brief Track origins (allocation points) of uninitialized values.
289 const DataLayout *DL;
293 /// \brief Thread-local shadow storage for function parameters.
294 GlobalVariable *ParamTLS;
295 /// \brief Thread-local origin storage for function parameters.
296 GlobalVariable *ParamOriginTLS;
297 /// \brief Thread-local shadow storage for function return value.
298 GlobalVariable *RetvalTLS;
299 /// \brief Thread-local origin storage for function return value.
300 GlobalVariable *RetvalOriginTLS;
301 /// \brief Thread-local shadow storage for in-register va_arg function
302 /// parameters (x86_64-specific).
303 GlobalVariable *VAArgTLS;
304 /// \brief Thread-local shadow storage for va_arg overflow area
305 /// (x86_64-specific).
306 GlobalVariable *VAArgOverflowSizeTLS;
307 /// \brief Thread-local space used to pass origin value to the UMR reporting
309 GlobalVariable *OriginTLS;
311 /// \brief The run-time callback to print a warning.
313 // These arrays are indexed by log2(AccessSize).
314 Value *MaybeWarningFn[kNumberOfAccessSizes];
315 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
317 /// \brief Run-time helper that generates a new origin value for a stack
319 Value *MsanSetAllocaOrigin4Fn;
320 /// \brief Run-time helper that poisons stack on function entry.
321 Value *MsanPoisonStackFn;
322 /// \brief Run-time helper that records a store (or any event) of an
323 /// uninitialized value and returns an updated origin id encoding this info.
324 Value *MsanChainOriginFn;
325 /// \brief MSan runtime replacements for memmove, memcpy and memset.
326 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
328 /// \brief Memory map parameters used in application-to-shadow calculation.
329 const MemoryMapParams *MapParams;
331 MDNode *ColdCallWeights;
332 /// \brief Branch weights for origin store.
333 MDNode *OriginStoreWeights;
334 /// \brief An empty volatile inline asm that prevents callback merge.
337 friend struct MemorySanitizerVisitor;
338 friend struct VarArgAMD64Helper;
342 char MemorySanitizer::ID = 0;
343 INITIALIZE_PASS(MemorySanitizer, "msan",
344 "MemorySanitizer: detects uninitialized reads.",
347 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
348 return new MemorySanitizer(TrackOrigins);
351 /// \brief Create a non-const global initialized with the given string.
353 /// Creates a writable global for Str so that we can pass it to the
354 /// run-time lib. Runtime uses first 4 bytes of the string to store the
355 /// frame ID, so the string needs to be mutable.
356 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
358 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
359 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
360 GlobalValue::PrivateLinkage, StrConst, "");
364 /// \brief Insert extern declaration of runtime-provided functions and globals.
365 void MemorySanitizer::initializeCallbacks(Module &M) {
366 // Only do this once.
371 // Create the callback.
372 // FIXME: this function should have "Cold" calling conv,
373 // which is not yet implemented.
374 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
375 : "__msan_warning_noreturn";
376 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), nullptr);
378 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
380 unsigned AccessSize = 1 << AccessSizeIndex;
381 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
382 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
383 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
384 IRB.getInt32Ty(), nullptr);
386 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
387 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
388 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
389 IRB.getInt8PtrTy(), IRB.getInt32Ty(), nullptr);
392 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
393 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
394 IRB.getInt8PtrTy(), IntptrTy, nullptr);
396 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
397 IRB.getInt8PtrTy(), IntptrTy, nullptr);
398 MsanChainOriginFn = M.getOrInsertFunction(
399 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), nullptr);
400 MemmoveFn = M.getOrInsertFunction(
401 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
402 IRB.getInt8PtrTy(), IntptrTy, nullptr);
403 MemcpyFn = M.getOrInsertFunction(
404 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
406 MemsetFn = M.getOrInsertFunction(
407 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
411 RetvalTLS = new GlobalVariable(
412 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
413 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
414 GlobalVariable::InitialExecTLSModel);
415 RetvalOriginTLS = new GlobalVariable(
416 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
417 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
419 ParamTLS = new GlobalVariable(
420 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
421 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
422 GlobalVariable::InitialExecTLSModel);
423 ParamOriginTLS = new GlobalVariable(
424 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
425 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
426 nullptr, GlobalVariable::InitialExecTLSModel);
428 VAArgTLS = new GlobalVariable(
429 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
430 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
431 GlobalVariable::InitialExecTLSModel);
432 VAArgOverflowSizeTLS = new GlobalVariable(
433 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
434 "__msan_va_arg_overflow_size_tls", nullptr,
435 GlobalVariable::InitialExecTLSModel);
436 OriginTLS = new GlobalVariable(
437 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
438 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
440 // We insert an empty inline asm after __msan_report* to avoid callback merge.
441 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
442 StringRef(""), StringRef(""),
443 /*hasSideEffects=*/true);
446 /// \brief Module-level initialization.
448 /// inserts a call to __msan_init to the module's constructor list.
449 bool MemorySanitizer::doInitialization(Module &M) {
450 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
452 report_fatal_error("data layout missing");
453 DL = &DLP->getDataLayout();
455 Triple TargetTriple(M.getTargetTriple());
456 switch (TargetTriple.getOS()) {
457 case Triple::FreeBSD:
458 switch (TargetTriple.getArch()) {
460 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
463 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
466 report_fatal_error("unsupported architecture");
470 switch (TargetTriple.getArch()) {
472 MapParams = Linux_X86_MemoryMapParams.bits64;
475 MapParams = Linux_X86_MemoryMapParams.bits32;
478 case Triple::mips64el:
479 MapParams = Linux_MIPS_MemoryMapParams.bits64;
482 report_fatal_error("unsupported architecture");
486 report_fatal_error("unsupported operating system");
489 C = &(M.getContext());
491 IntptrTy = IRB.getIntPtrTy(DL);
492 OriginTy = IRB.getInt32Ty();
494 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
495 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
497 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
498 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
499 "__msan_init", IRB.getVoidTy(), nullptr)), 0);
502 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
503 IRB.getInt32(TrackOrigins), "__msan_track_origins");
506 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
507 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
514 /// \brief A helper class that handles instrumentation of VarArg
515 /// functions on a particular platform.
517 /// Implementations are expected to insert the instrumentation
518 /// necessary to propagate argument shadow through VarArg function
519 /// calls. Visit* methods are called during an InstVisitor pass over
520 /// the function, and should avoid creating new basic blocks. A new
521 /// instance of this class is created for each instrumented function.
522 struct VarArgHelper {
523 /// \brief Visit a CallSite.
524 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
526 /// \brief Visit a va_start call.
527 virtual void visitVAStartInst(VAStartInst &I) = 0;
529 /// \brief Visit a va_copy call.
530 virtual void visitVACopyInst(VACopyInst &I) = 0;
532 /// \brief Finalize function instrumentation.
534 /// This method is called after visiting all interesting (see above)
535 /// instructions in a function.
536 virtual void finalizeInstrumentation() = 0;
538 virtual ~VarArgHelper() {}
541 struct MemorySanitizerVisitor;
544 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
545 MemorySanitizerVisitor &Visitor);
547 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
548 if (TypeSize <= 8) return 0;
549 return Log2_32_Ceil(TypeSize / 8);
552 /// This class does all the work for a given function. Store and Load
553 /// instructions store and load corresponding shadow and origin
554 /// values. Most instructions propagate shadow from arguments to their
555 /// return values. Certain instructions (most importantly, BranchInst)
556 /// test their argument shadow and print reports (with a runtime call) if it's
558 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
561 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
562 ValueMap<Value*, Value*> ShadowMap, OriginMap;
563 std::unique_ptr<VarArgHelper> VAHelper;
565 // The following flags disable parts of MSan instrumentation based on
566 // blacklist contents and command-line options.
568 bool PropagateShadow;
571 bool CheckReturnValue;
573 struct ShadowOriginAndInsertPoint {
576 Instruction *OrigIns;
577 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
578 : Shadow(S), Origin(O), OrigIns(I) { }
580 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
581 SmallVector<Instruction*, 16> StoreList;
583 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
584 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
585 bool SanitizeFunction = F.getAttributes().hasAttribute(
586 AttributeSet::FunctionIndex, Attribute::SanitizeMemory);
587 InsertChecks = SanitizeFunction;
588 PropagateShadow = SanitizeFunction;
589 PoisonStack = SanitizeFunction && ClPoisonStack;
590 PoisonUndef = SanitizeFunction && ClPoisonUndef;
591 // FIXME: Consider using SpecialCaseList to specify a list of functions that
592 // must always return fully initialized values. For now, we hardcode "main".
593 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
595 DEBUG(if (!InsertChecks)
596 dbgs() << "MemorySanitizer is not inserting checks into '"
597 << F.getName() << "'\n");
600 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
601 if (MS.TrackOrigins <= 1) return V;
602 return IRB.CreateCall(MS.MsanChainOriginFn, V);
605 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
606 unsigned Alignment, bool AsCall) {
607 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
608 if (isa<StructType>(Shadow->getType())) {
609 IRB.CreateAlignedStore(updateOrigin(Origin, IRB),
610 getOriginPtr(Addr, IRB, Alignment),
613 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
614 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
615 if (ConstantShadow) {
616 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
617 IRB.CreateAlignedStore(updateOrigin(Origin, IRB),
618 getOriginPtr(Addr, IRB, Alignment),
623 unsigned TypeSizeInBits =
624 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
625 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
626 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
627 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
628 Value *ConvertedShadow2 = IRB.CreateZExt(
629 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
630 IRB.CreateCall3(Fn, ConvertedShadow2,
631 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
634 Value *Cmp = IRB.CreateICmpNE(
635 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
636 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
637 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
638 IRBuilder<> IRBNew(CheckTerm);
639 IRBNew.CreateAlignedStore(updateOrigin(Origin, IRBNew),
640 getOriginPtr(Addr, IRBNew, Alignment),
646 void materializeStores(bool InstrumentWithCalls) {
647 for (auto Inst : StoreList) {
648 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
650 IRBuilder<> IRB(&SI);
651 Value *Val = SI.getValueOperand();
652 Value *Addr = SI.getPointerOperand();
653 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
654 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
657 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
658 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
661 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
663 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
666 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
667 InstrumentWithCalls);
671 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
673 IRBuilder<> IRB(OrigIns);
674 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
675 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
676 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
678 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
679 if (ConstantShadow) {
680 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
681 if (MS.TrackOrigins) {
682 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
685 IRB.CreateCall(MS.WarningFn);
686 IRB.CreateCall(MS.EmptyAsm);
687 // FIXME: Insert UnreachableInst if !ClKeepGoing?
688 // This may invalidate some of the following checks and needs to be done
694 unsigned TypeSizeInBits =
695 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
696 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
697 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
698 Value *Fn = MS.MaybeWarningFn[SizeIndex];
699 Value *ConvertedShadow2 =
700 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
701 IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
703 : (Value *)IRB.getInt32(0));
705 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
706 getCleanShadow(ConvertedShadow), "_mscmp");
707 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
709 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
711 IRB.SetInsertPoint(CheckTerm);
712 if (MS.TrackOrigins) {
713 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
716 IRB.CreateCall(MS.WarningFn);
717 IRB.CreateCall(MS.EmptyAsm);
718 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
722 void materializeChecks(bool InstrumentWithCalls) {
723 for (const auto &ShadowData : InstrumentationList) {
724 Instruction *OrigIns = ShadowData.OrigIns;
725 Value *Shadow = ShadowData.Shadow;
726 Value *Origin = ShadowData.Origin;
727 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
729 DEBUG(dbgs() << "DONE:\n" << F);
732 /// \brief Add MemorySanitizer instrumentation to a function.
733 bool runOnFunction() {
734 MS.initializeCallbacks(*F.getParent());
735 if (!MS.DL) return false;
737 // In the presence of unreachable blocks, we may see Phi nodes with
738 // incoming nodes from such blocks. Since InstVisitor skips unreachable
739 // blocks, such nodes will not have any shadow value associated with them.
740 // It's easier to remove unreachable blocks than deal with missing shadow.
741 removeUnreachableBlocks(F);
743 // Iterate all BBs in depth-first order and create shadow instructions
744 // for all instructions (where applicable).
745 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
746 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
750 // Finalize PHI nodes.
751 for (PHINode *PN : ShadowPHINodes) {
752 PHINode *PNS = cast<PHINode>(getShadow(PN));
753 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
754 size_t NumValues = PN->getNumIncomingValues();
755 for (size_t v = 0; v < NumValues; v++) {
756 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
757 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
761 VAHelper->finalizeInstrumentation();
763 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
764 InstrumentationList.size() + StoreList.size() >
765 (unsigned)ClInstrumentationWithCallThreshold;
767 // Delayed instrumentation of StoreInst.
768 // This may add new checks to be inserted later.
769 materializeStores(InstrumentWithCalls);
771 // Insert shadow value checks.
772 materializeChecks(InstrumentWithCalls);
777 /// \brief Compute the shadow type that corresponds to a given Value.
778 Type *getShadowTy(Value *V) {
779 return getShadowTy(V->getType());
782 /// \brief Compute the shadow type that corresponds to a given Type.
783 Type *getShadowTy(Type *OrigTy) {
784 if (!OrigTy->isSized()) {
787 // For integer type, shadow is the same as the original type.
788 // This may return weird-sized types like i1.
789 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
791 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
792 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
793 return VectorType::get(IntegerType::get(*MS.C, EltSize),
794 VT->getNumElements());
796 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
797 return ArrayType::get(getShadowTy(AT->getElementType()),
798 AT->getNumElements());
800 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
801 SmallVector<Type*, 4> Elements;
802 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
803 Elements.push_back(getShadowTy(ST->getElementType(i)));
804 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
805 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
808 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
809 return IntegerType::get(*MS.C, TypeSize);
812 /// \brief Flatten a vector type.
813 Type *getShadowTyNoVec(Type *ty) {
814 if (VectorType *vt = dyn_cast<VectorType>(ty))
815 return IntegerType::get(*MS.C, vt->getBitWidth());
819 /// \brief Convert a shadow value to it's flattened variant.
820 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
821 Type *Ty = V->getType();
822 Type *NoVecTy = getShadowTyNoVec(Ty);
823 if (Ty == NoVecTy) return V;
824 return IRB.CreateBitCast(V, NoVecTy);
827 /// \brief Compute the integer shadow offset that corresponds to a given
828 /// application address.
830 /// Offset = (Addr & ~AndMask) ^ XorMask
831 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
832 uint64_t AndMask = MS.MapParams->AndMask;
833 assert(AndMask != 0 && "AndMask shall be specified");
835 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
836 ConstantInt::get(MS.IntptrTy, ~AndMask));
838 uint64_t XorMask = MS.MapParams->XorMask;
840 OffsetLong = IRB.CreateXor(OffsetLong,
841 ConstantInt::get(MS.IntptrTy, XorMask));
845 /// \brief Compute the shadow address that corresponds to a given application
848 /// Shadow = ShadowBase + Offset
849 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
851 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
852 uint64_t ShadowBase = MS.MapParams->ShadowBase;
855 IRB.CreateAdd(ShadowLong,
856 ConstantInt::get(MS.IntptrTy, ShadowBase));
857 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
860 /// \brief Compute the origin address that corresponds to a given application
863 /// OriginAddr = (OriginBase + Offset) & ~3ULL
864 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
865 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
866 uint64_t OriginBase = MS.MapParams->OriginBase;
869 IRB.CreateAdd(OriginLong,
870 ConstantInt::get(MS.IntptrTy, OriginBase));
871 if (Alignment < kMinOriginAlignment) {
872 uint64_t Mask = kMinOriginAlignment - 1;
873 OriginLong = IRB.CreateAnd(OriginLong,
874 ConstantInt::get(MS.IntptrTy, ~Mask));
876 return IRB.CreateIntToPtr(OriginLong,
877 PointerType::get(IRB.getInt32Ty(), 0));
880 /// \brief Compute the shadow address for a given function argument.
882 /// Shadow = ParamTLS+ArgOffset.
883 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
885 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
886 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
887 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
891 /// \brief Compute the origin address for a given function argument.
892 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
894 if (!MS.TrackOrigins) return nullptr;
895 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
896 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
897 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
901 /// \brief Compute the shadow address for a retval.
902 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
903 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
904 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
908 /// \brief Compute the origin address for a retval.
909 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
910 // We keep a single origin for the entire retval. Might be too optimistic.
911 return MS.RetvalOriginTLS;
914 /// \brief Set SV to be the shadow value for V.
915 void setShadow(Value *V, Value *SV) {
916 assert(!ShadowMap.count(V) && "Values may only have one shadow");
917 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
920 /// \brief Set Origin to be the origin value for V.
921 void setOrigin(Value *V, Value *Origin) {
922 if (!MS.TrackOrigins) return;
923 assert(!OriginMap.count(V) && "Values may only have one origin");
924 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
925 OriginMap[V] = Origin;
928 /// \brief Create a clean shadow value for a given value.
930 /// Clean shadow (all zeroes) means all bits of the value are defined
932 Constant *getCleanShadow(Value *V) {
933 Type *ShadowTy = getShadowTy(V);
936 return Constant::getNullValue(ShadowTy);
939 /// \brief Create a dirty shadow of a given shadow type.
940 Constant *getPoisonedShadow(Type *ShadowTy) {
942 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
943 return Constant::getAllOnesValue(ShadowTy);
944 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
945 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
946 getPoisonedShadow(AT->getElementType()));
947 return ConstantArray::get(AT, Vals);
949 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
950 SmallVector<Constant *, 4> Vals;
951 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
952 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
953 return ConstantStruct::get(ST, Vals);
955 llvm_unreachable("Unexpected shadow type");
958 /// \brief Create a dirty shadow for a given value.
959 Constant *getPoisonedShadow(Value *V) {
960 Type *ShadowTy = getShadowTy(V);
963 return getPoisonedShadow(ShadowTy);
966 /// \brief Create a clean (zero) origin.
967 Value *getCleanOrigin() {
968 return Constant::getNullValue(MS.OriginTy);
971 /// \brief Get the shadow value for a given Value.
973 /// This function either returns the value set earlier with setShadow,
974 /// or extracts if from ParamTLS (for function arguments).
975 Value *getShadow(Value *V) {
976 if (!PropagateShadow) return getCleanShadow(V);
977 if (Instruction *I = dyn_cast<Instruction>(V)) {
978 // For instructions the shadow is already stored in the map.
979 Value *Shadow = ShadowMap[V];
981 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
983 assert(Shadow && "No shadow for a value");
987 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
988 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
989 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
993 if (Argument *A = dyn_cast<Argument>(V)) {
994 // For arguments we compute the shadow on demand and store it in the map.
995 Value **ShadowPtr = &ShadowMap[V];
998 Function *F = A->getParent();
999 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1000 unsigned ArgOffset = 0;
1001 for (auto &FArg : F->args()) {
1002 if (!FArg.getType()->isSized()) {
1003 DEBUG(dbgs() << "Arg is not sized\n");
1006 unsigned Size = FArg.hasByValAttr()
1007 ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
1008 : MS.DL->getTypeAllocSize(FArg.getType());
1010 bool Overflow = ArgOffset + Size > kParamTLSSize;
1011 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1012 if (FArg.hasByValAttr()) {
1013 // ByVal pointer itself has clean shadow. We copy the actual
1014 // argument shadow to the underlying memory.
1015 // Figure out maximal valid memcpy alignment.
1016 unsigned ArgAlign = FArg.getParamAlignment();
1017 if (ArgAlign == 0) {
1018 Type *EltType = A->getType()->getPointerElementType();
1019 ArgAlign = MS.DL->getABITypeAlignment(EltType);
1022 // ParamTLS overflow.
1023 EntryIRB.CreateMemSet(
1024 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1025 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1027 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1028 Value *Cpy = EntryIRB.CreateMemCpy(
1029 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1031 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1034 *ShadowPtr = getCleanShadow(V);
1037 // ParamTLS overflow.
1038 *ShadowPtr = getCleanShadow(V);
1041 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1044 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1045 **ShadowPtr << "\n");
1046 if (MS.TrackOrigins && !Overflow) {
1048 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1049 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1051 setOrigin(A, getCleanOrigin());
1054 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1056 assert(*ShadowPtr && "Could not find shadow for an argument");
1059 // For everything else the shadow is zero.
1060 return getCleanShadow(V);
1063 /// \brief Get the shadow for i-th argument of the instruction I.
1064 Value *getShadow(Instruction *I, int i) {
1065 return getShadow(I->getOperand(i));
1068 /// \brief Get the origin for a value.
1069 Value *getOrigin(Value *V) {
1070 if (!MS.TrackOrigins) return nullptr;
1071 if (!PropagateShadow) return getCleanOrigin();
1072 if (isa<Constant>(V)) return getCleanOrigin();
1073 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1074 "Unexpected value type in getOrigin()");
1075 Value *Origin = OriginMap[V];
1076 assert(Origin && "Missing origin");
1080 /// \brief Get the origin for i-th argument of the instruction I.
1081 Value *getOrigin(Instruction *I, int i) {
1082 return getOrigin(I->getOperand(i));
1085 /// \brief Remember the place where a shadow check should be inserted.
1087 /// This location will be later instrumented with a check that will print a
1088 /// UMR warning in runtime if the shadow value is not 0.
1089 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1091 if (!InsertChecks) return;
1093 Type *ShadowTy = Shadow->getType();
1094 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1095 "Can only insert checks for integer and vector shadow types");
1097 InstrumentationList.push_back(
1098 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1101 /// \brief Remember the place where a shadow check should be inserted.
1103 /// This location will be later instrumented with a check that will print a
1104 /// UMR warning in runtime if the value is not fully defined.
1105 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1107 Value *Shadow, *Origin;
1108 if (ClCheckConstantShadow) {
1109 Shadow = getShadow(Val);
1110 if (!Shadow) return;
1111 Origin = getOrigin(Val);
1113 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1114 if (!Shadow) return;
1115 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1117 insertShadowCheck(Shadow, Origin, OrigIns);
1120 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1129 case AcquireRelease:
1130 return AcquireRelease;
1131 case SequentiallyConsistent:
1132 return SequentiallyConsistent;
1134 llvm_unreachable("Unknown ordering");
1137 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1146 case AcquireRelease:
1147 return AcquireRelease;
1148 case SequentiallyConsistent:
1149 return SequentiallyConsistent;
1151 llvm_unreachable("Unknown ordering");
1154 // ------------------- Visitors.
1156 /// \brief Instrument LoadInst
1158 /// Loads the corresponding shadow and (optionally) origin.
1159 /// Optionally, checks that the load address is fully defined.
1160 void visitLoadInst(LoadInst &I) {
1161 assert(I.getType()->isSized() && "Load type must have size");
1162 IRBuilder<> IRB(I.getNextNode());
1163 Type *ShadowTy = getShadowTy(&I);
1164 Value *Addr = I.getPointerOperand();
1165 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1166 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1168 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1170 setShadow(&I, getCleanShadow(&I));
1173 if (ClCheckAccessAddress)
1174 insertShadowCheck(I.getPointerOperand(), &I);
1177 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1179 if (MS.TrackOrigins) {
1180 if (PropagateShadow) {
1181 unsigned Alignment = I.getAlignment();
1182 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1183 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1186 setOrigin(&I, getCleanOrigin());
1191 /// \brief Instrument StoreInst
1193 /// Stores the corresponding shadow and (optionally) origin.
1194 /// Optionally, checks that the store address is fully defined.
1195 void visitStoreInst(StoreInst &I) {
1196 StoreList.push_back(&I);
1199 void handleCASOrRMW(Instruction &I) {
1200 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1202 IRBuilder<> IRB(&I);
1203 Value *Addr = I.getOperand(0);
1204 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1206 if (ClCheckAccessAddress)
1207 insertShadowCheck(Addr, &I);
1209 // Only test the conditional argument of cmpxchg instruction.
1210 // The other argument can potentially be uninitialized, but we can not
1211 // detect this situation reliably without possible false positives.
1212 if (isa<AtomicCmpXchgInst>(I))
1213 insertShadowCheck(I.getOperand(1), &I);
1215 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1217 setShadow(&I, getCleanShadow(&I));
1218 setOrigin(&I, getCleanOrigin());
1221 void visitAtomicRMWInst(AtomicRMWInst &I) {
1223 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1226 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1228 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1231 // Vector manipulation.
1232 void visitExtractElementInst(ExtractElementInst &I) {
1233 insertShadowCheck(I.getOperand(1), &I);
1234 IRBuilder<> IRB(&I);
1235 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1237 setOrigin(&I, getOrigin(&I, 0));
1240 void visitInsertElementInst(InsertElementInst &I) {
1241 insertShadowCheck(I.getOperand(2), &I);
1242 IRBuilder<> IRB(&I);
1243 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1244 I.getOperand(2), "_msprop"));
1245 setOriginForNaryOp(I);
1248 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1249 insertShadowCheck(I.getOperand(2), &I);
1250 IRBuilder<> IRB(&I);
1251 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1252 I.getOperand(2), "_msprop"));
1253 setOriginForNaryOp(I);
1257 void visitSExtInst(SExtInst &I) {
1258 IRBuilder<> IRB(&I);
1259 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1260 setOrigin(&I, getOrigin(&I, 0));
1263 void visitZExtInst(ZExtInst &I) {
1264 IRBuilder<> IRB(&I);
1265 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1266 setOrigin(&I, getOrigin(&I, 0));
1269 void visitTruncInst(TruncInst &I) {
1270 IRBuilder<> IRB(&I);
1271 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1272 setOrigin(&I, getOrigin(&I, 0));
1275 void visitBitCastInst(BitCastInst &I) {
1276 IRBuilder<> IRB(&I);
1277 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1278 setOrigin(&I, getOrigin(&I, 0));
1281 void visitPtrToIntInst(PtrToIntInst &I) {
1282 IRBuilder<> IRB(&I);
1283 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1284 "_msprop_ptrtoint"));
1285 setOrigin(&I, getOrigin(&I, 0));
1288 void visitIntToPtrInst(IntToPtrInst &I) {
1289 IRBuilder<> IRB(&I);
1290 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1291 "_msprop_inttoptr"));
1292 setOrigin(&I, getOrigin(&I, 0));
1295 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1296 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1297 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1298 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1299 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1300 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1302 /// \brief Propagate shadow for bitwise AND.
1304 /// This code is exact, i.e. if, for example, a bit in the left argument
1305 /// is defined and 0, then neither the value not definedness of the
1306 /// corresponding bit in B don't affect the resulting shadow.
1307 void visitAnd(BinaryOperator &I) {
1308 IRBuilder<> IRB(&I);
1309 // "And" of 0 and a poisoned value results in unpoisoned value.
1310 // 1&1 => 1; 0&1 => 0; p&1 => p;
1311 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1312 // 1&p => p; 0&p => 0; p&p => p;
1313 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1314 Value *S1 = getShadow(&I, 0);
1315 Value *S2 = getShadow(&I, 1);
1316 Value *V1 = I.getOperand(0);
1317 Value *V2 = I.getOperand(1);
1318 if (V1->getType() != S1->getType()) {
1319 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1320 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1322 Value *S1S2 = IRB.CreateAnd(S1, S2);
1323 Value *V1S2 = IRB.CreateAnd(V1, S2);
1324 Value *S1V2 = IRB.CreateAnd(S1, V2);
1325 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1326 setOriginForNaryOp(I);
1329 void visitOr(BinaryOperator &I) {
1330 IRBuilder<> IRB(&I);
1331 // "Or" of 1 and a poisoned value results in unpoisoned value.
1332 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1333 // 1|0 => 1; 0|0 => 0; p|0 => p;
1334 // 1|p => 1; 0|p => p; p|p => p;
1335 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1336 Value *S1 = getShadow(&I, 0);
1337 Value *S2 = getShadow(&I, 1);
1338 Value *V1 = IRB.CreateNot(I.getOperand(0));
1339 Value *V2 = IRB.CreateNot(I.getOperand(1));
1340 if (V1->getType() != S1->getType()) {
1341 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1342 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1344 Value *S1S2 = IRB.CreateAnd(S1, S2);
1345 Value *V1S2 = IRB.CreateAnd(V1, S2);
1346 Value *S1V2 = IRB.CreateAnd(S1, V2);
1347 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1348 setOriginForNaryOp(I);
1351 /// \brief Default propagation of shadow and/or origin.
1353 /// This class implements the general case of shadow propagation, used in all
1354 /// cases where we don't know and/or don't care about what the operation
1355 /// actually does. It converts all input shadow values to a common type
1356 /// (extending or truncating as necessary), and bitwise OR's them.
1358 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1359 /// fully initialized), and less prone to false positives.
1361 /// This class also implements the general case of origin propagation. For a
1362 /// Nary operation, result origin is set to the origin of an argument that is
1363 /// not entirely initialized. If there is more than one such arguments, the
1364 /// rightmost of them is picked. It does not matter which one is picked if all
1365 /// arguments are initialized.
1366 template <bool CombineShadow>
1371 MemorySanitizerVisitor *MSV;
1374 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1375 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1377 /// \brief Add a pair of shadow and origin values to the mix.
1378 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1379 if (CombineShadow) {
1384 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1385 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1389 if (MSV->MS.TrackOrigins) {
1394 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1395 // No point in adding something that might result in 0 origin value.
1396 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1397 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1399 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1400 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1407 /// \brief Add an application value to the mix.
1408 Combiner &Add(Value *V) {
1409 Value *OpShadow = MSV->getShadow(V);
1410 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1411 return Add(OpShadow, OpOrigin);
1414 /// \brief Set the current combined values as the given instruction's shadow
1416 void Done(Instruction *I) {
1417 if (CombineShadow) {
1419 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1420 MSV->setShadow(I, Shadow);
1422 if (MSV->MS.TrackOrigins) {
1424 MSV->setOrigin(I, Origin);
1429 typedef Combiner<true> ShadowAndOriginCombiner;
1430 typedef Combiner<false> OriginCombiner;
1432 /// \brief Propagate origin for arbitrary operation.
1433 void setOriginForNaryOp(Instruction &I) {
1434 if (!MS.TrackOrigins) return;
1435 IRBuilder<> IRB(&I);
1436 OriginCombiner OC(this, IRB);
1437 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1442 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1443 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1444 "Vector of pointers is not a valid shadow type");
1445 return Ty->isVectorTy() ?
1446 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1447 Ty->getPrimitiveSizeInBits();
1450 /// \brief Cast between two shadow types, extending or truncating as
1452 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1453 bool Signed = false) {
1454 Type *srcTy = V->getType();
1455 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1456 return IRB.CreateIntCast(V, dstTy, Signed);
1457 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1458 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1459 return IRB.CreateIntCast(V, dstTy, Signed);
1460 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1461 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1462 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1464 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1465 return IRB.CreateBitCast(V2, dstTy);
1466 // TODO: handle struct types.
1469 /// \brief Cast an application value to the type of its own shadow.
1470 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1471 Type *ShadowTy = getShadowTy(V);
1472 if (V->getType() == ShadowTy)
1474 if (V->getType()->isPtrOrPtrVectorTy())
1475 return IRB.CreatePtrToInt(V, ShadowTy);
1477 return IRB.CreateBitCast(V, ShadowTy);
1480 /// \brief Propagate shadow for arbitrary operation.
1481 void handleShadowOr(Instruction &I) {
1482 IRBuilder<> IRB(&I);
1483 ShadowAndOriginCombiner SC(this, IRB);
1484 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1489 // \brief Handle multiplication by constant.
1491 // Handle a special case of multiplication by constant that may have one or
1492 // more zeros in the lower bits. This makes corresponding number of lower bits
1493 // of the result zero as well. We model it by shifting the other operand
1494 // shadow left by the required number of bits. Effectively, we transform
1495 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1496 // We use multiplication by 2**N instead of shift to cover the case of
1497 // multiplication by 0, which may occur in some elements of a vector operand.
1498 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1500 Constant *ShadowMul;
1501 Type *Ty = ConstArg->getType();
1502 if (Ty->isVectorTy()) {
1503 unsigned NumElements = Ty->getVectorNumElements();
1504 Type *EltTy = Ty->getSequentialElementType();
1505 SmallVector<Constant *, 16> Elements;
1506 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1508 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1509 APInt V = Elt->getValue();
1510 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1511 Elements.push_back(ConstantInt::get(EltTy, V2));
1513 ShadowMul = ConstantVector::get(Elements);
1515 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1516 APInt V = Elt->getValue();
1517 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1518 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1521 IRBuilder<> IRB(&I);
1523 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1524 setOrigin(&I, getOrigin(OtherArg));
1527 void visitMul(BinaryOperator &I) {
1528 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1529 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1530 if (constOp0 && !constOp1)
1531 handleMulByConstant(I, constOp0, I.getOperand(1));
1532 else if (constOp1 && !constOp0)
1533 handleMulByConstant(I, constOp1, I.getOperand(0));
1538 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1539 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1540 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1541 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1542 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1543 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1545 void handleDiv(Instruction &I) {
1546 IRBuilder<> IRB(&I);
1547 // Strict on the second argument.
1548 insertShadowCheck(I.getOperand(1), &I);
1549 setShadow(&I, getShadow(&I, 0));
1550 setOrigin(&I, getOrigin(&I, 0));
1553 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1554 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1555 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1556 void visitURem(BinaryOperator &I) { handleDiv(I); }
1557 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1558 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1560 /// \brief Instrument == and != comparisons.
1562 /// Sometimes the comparison result is known even if some of the bits of the
1563 /// arguments are not.
1564 void handleEqualityComparison(ICmpInst &I) {
1565 IRBuilder<> IRB(&I);
1566 Value *A = I.getOperand(0);
1567 Value *B = I.getOperand(1);
1568 Value *Sa = getShadow(A);
1569 Value *Sb = getShadow(B);
1571 // Get rid of pointers and vectors of pointers.
1572 // For ints (and vectors of ints), types of A and Sa match,
1573 // and this is a no-op.
1574 A = IRB.CreatePointerCast(A, Sa->getType());
1575 B = IRB.CreatePointerCast(B, Sb->getType());
1577 // A == B <==> (C = A^B) == 0
1578 // A != B <==> (C = A^B) != 0
1580 Value *C = IRB.CreateXor(A, B);
1581 Value *Sc = IRB.CreateOr(Sa, Sb);
1582 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1583 // Result is defined if one of the following is true
1584 // * there is a defined 1 bit in C
1585 // * C is fully defined
1586 // Si = !(C & ~Sc) && Sc
1587 Value *Zero = Constant::getNullValue(Sc->getType());
1588 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1590 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1592 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1593 Si->setName("_msprop_icmp");
1595 setOriginForNaryOp(I);
1598 /// \brief Build the lowest possible value of V, taking into account V's
1599 /// uninitialized bits.
1600 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1603 // Split shadow into sign bit and other bits.
1604 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1605 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1606 // Maximise the undefined shadow bit, minimize other undefined bits.
1608 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1610 // Minimize undefined bits.
1611 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1615 /// \brief Build the highest possible value of V, taking into account V's
1616 /// uninitialized bits.
1617 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1620 // Split shadow into sign bit and other bits.
1621 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1622 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1623 // Minimise the undefined shadow bit, maximise other undefined bits.
1625 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1627 // Maximize undefined bits.
1628 return IRB.CreateOr(A, Sa);
1632 /// \brief Instrument relational comparisons.
1634 /// This function does exact shadow propagation for all relational
1635 /// comparisons of integers, pointers and vectors of those.
1636 /// FIXME: output seems suboptimal when one of the operands is a constant
1637 void handleRelationalComparisonExact(ICmpInst &I) {
1638 IRBuilder<> IRB(&I);
1639 Value *A = I.getOperand(0);
1640 Value *B = I.getOperand(1);
1641 Value *Sa = getShadow(A);
1642 Value *Sb = getShadow(B);
1644 // Get rid of pointers and vectors of pointers.
1645 // For ints (and vectors of ints), types of A and Sa match,
1646 // and this is a no-op.
1647 A = IRB.CreatePointerCast(A, Sa->getType());
1648 B = IRB.CreatePointerCast(B, Sb->getType());
1650 // Let [a0, a1] be the interval of possible values of A, taking into account
1651 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1652 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1653 bool IsSigned = I.isSigned();
1654 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1655 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1656 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1657 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1658 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1659 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1660 Value *Si = IRB.CreateXor(S1, S2);
1662 setOriginForNaryOp(I);
1665 /// \brief Instrument signed relational comparisons.
1667 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1668 /// propagating the highest bit of the shadow. Everything else is delegated
1669 /// to handleShadowOr().
1670 void handleSignedRelationalComparison(ICmpInst &I) {
1671 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1672 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1673 Value* op = nullptr;
1674 CmpInst::Predicate pre = I.getPredicate();
1675 if (constOp0 && constOp0->isNullValue() &&
1676 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1677 op = I.getOperand(1);
1678 } else if (constOp1 && constOp1->isNullValue() &&
1679 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1680 op = I.getOperand(0);
1683 IRBuilder<> IRB(&I);
1685 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1686 setShadow(&I, Shadow);
1687 setOrigin(&I, getOrigin(op));
1693 void visitICmpInst(ICmpInst &I) {
1694 if (!ClHandleICmp) {
1698 if (I.isEquality()) {
1699 handleEqualityComparison(I);
1703 assert(I.isRelational());
1704 if (ClHandleICmpExact) {
1705 handleRelationalComparisonExact(I);
1709 handleSignedRelationalComparison(I);
1713 assert(I.isUnsigned());
1714 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1715 handleRelationalComparisonExact(I);
1722 void visitFCmpInst(FCmpInst &I) {
1726 void handleShift(BinaryOperator &I) {
1727 IRBuilder<> IRB(&I);
1728 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1729 // Otherwise perform the same shift on S1.
1730 Value *S1 = getShadow(&I, 0);
1731 Value *S2 = getShadow(&I, 1);
1732 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1734 Value *V2 = I.getOperand(1);
1735 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1736 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1737 setOriginForNaryOp(I);
1740 void visitShl(BinaryOperator &I) { handleShift(I); }
1741 void visitAShr(BinaryOperator &I) { handleShift(I); }
1742 void visitLShr(BinaryOperator &I) { handleShift(I); }
1744 /// \brief Instrument llvm.memmove
1746 /// At this point we don't know if llvm.memmove will be inlined or not.
1747 /// If we don't instrument it and it gets inlined,
1748 /// our interceptor will not kick in and we will lose the memmove.
1749 /// If we instrument the call here, but it does not get inlined,
1750 /// we will memove the shadow twice: which is bad in case
1751 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1753 /// Similar situation exists for memcpy and memset.
1754 void visitMemMoveInst(MemMoveInst &I) {
1755 IRBuilder<> IRB(&I);
1758 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1759 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1760 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1761 I.eraseFromParent();
1764 // Similar to memmove: avoid copying shadow twice.
1765 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1766 // FIXME: consider doing manual inline for small constant sizes and proper
1768 void visitMemCpyInst(MemCpyInst &I) {
1769 IRBuilder<> IRB(&I);
1772 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1773 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1774 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1775 I.eraseFromParent();
1779 void visitMemSetInst(MemSetInst &I) {
1780 IRBuilder<> IRB(&I);
1783 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1784 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1785 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1786 I.eraseFromParent();
1789 void visitVAStartInst(VAStartInst &I) {
1790 VAHelper->visitVAStartInst(I);
1793 void visitVACopyInst(VACopyInst &I) {
1794 VAHelper->visitVACopyInst(I);
1797 enum IntrinsicKind {
1798 IK_DoesNotAccessMemory,
1803 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1804 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1805 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1806 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1807 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1808 const int UnknownModRefBehavior = IK_WritesMemory;
1809 #define GET_INTRINSIC_MODREF_BEHAVIOR
1810 #define ModRefBehavior IntrinsicKind
1811 #include "llvm/IR/Intrinsics.gen"
1812 #undef ModRefBehavior
1813 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1816 /// \brief Handle vector store-like intrinsics.
1818 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1819 /// has 1 pointer argument and 1 vector argument, returns void.
1820 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1821 IRBuilder<> IRB(&I);
1822 Value* Addr = I.getArgOperand(0);
1823 Value *Shadow = getShadow(&I, 1);
1824 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1826 // We don't know the pointer alignment (could be unaligned SSE store!).
1827 // Have to assume to worst case.
1828 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1830 if (ClCheckAccessAddress)
1831 insertShadowCheck(Addr, &I);
1833 // FIXME: use ClStoreCleanOrigin
1834 // FIXME: factor out common code from materializeStores
1835 if (MS.TrackOrigins)
1836 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1840 /// \brief Handle vector load-like intrinsics.
1842 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1843 /// has 1 pointer argument, returns a vector.
1844 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1845 IRBuilder<> IRB(&I);
1846 Value *Addr = I.getArgOperand(0);
1848 Type *ShadowTy = getShadowTy(&I);
1849 if (PropagateShadow) {
1850 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1851 // We don't know the pointer alignment (could be unaligned SSE load!).
1852 // Have to assume to worst case.
1853 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1855 setShadow(&I, getCleanShadow(&I));
1858 if (ClCheckAccessAddress)
1859 insertShadowCheck(Addr, &I);
1861 if (MS.TrackOrigins) {
1862 if (PropagateShadow)
1863 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1865 setOrigin(&I, getCleanOrigin());
1870 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1872 /// Instrument intrinsics with any number of arguments of the same type,
1873 /// equal to the return type. The type should be simple (no aggregates or
1874 /// pointers; vectors are fine).
1875 /// Caller guarantees that this intrinsic does not access memory.
1876 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1877 Type *RetTy = I.getType();
1878 if (!(RetTy->isIntOrIntVectorTy() ||
1879 RetTy->isFPOrFPVectorTy() ||
1880 RetTy->isX86_MMXTy()))
1883 unsigned NumArgOperands = I.getNumArgOperands();
1885 for (unsigned i = 0; i < NumArgOperands; ++i) {
1886 Type *Ty = I.getArgOperand(i)->getType();
1891 IRBuilder<> IRB(&I);
1892 ShadowAndOriginCombiner SC(this, IRB);
1893 for (unsigned i = 0; i < NumArgOperands; ++i)
1894 SC.Add(I.getArgOperand(i));
1900 /// \brief Heuristically instrument unknown intrinsics.
1902 /// The main purpose of this code is to do something reasonable with all
1903 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1904 /// We recognize several classes of intrinsics by their argument types and
1905 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1906 /// sure that we know what the intrinsic does.
1908 /// We special-case intrinsics where this approach fails. See llvm.bswap
1909 /// handling as an example of that.
1910 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1911 unsigned NumArgOperands = I.getNumArgOperands();
1912 if (NumArgOperands == 0)
1915 Intrinsic::ID iid = I.getIntrinsicID();
1916 IntrinsicKind IK = getIntrinsicKind(iid);
1917 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1918 bool WritesMemory = IK == IK_WritesMemory;
1919 assert(!(OnlyReadsMemory && WritesMemory));
1921 if (NumArgOperands == 2 &&
1922 I.getArgOperand(0)->getType()->isPointerTy() &&
1923 I.getArgOperand(1)->getType()->isVectorTy() &&
1924 I.getType()->isVoidTy() &&
1926 // This looks like a vector store.
1927 return handleVectorStoreIntrinsic(I);
1930 if (NumArgOperands == 1 &&
1931 I.getArgOperand(0)->getType()->isPointerTy() &&
1932 I.getType()->isVectorTy() &&
1934 // This looks like a vector load.
1935 return handleVectorLoadIntrinsic(I);
1938 if (!OnlyReadsMemory && !WritesMemory)
1939 if (maybeHandleSimpleNomemIntrinsic(I))
1942 // FIXME: detect and handle SSE maskstore/maskload
1946 void handleBswap(IntrinsicInst &I) {
1947 IRBuilder<> IRB(&I);
1948 Value *Op = I.getArgOperand(0);
1949 Type *OpType = Op->getType();
1950 Function *BswapFunc = Intrinsic::getDeclaration(
1951 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
1952 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1953 setOrigin(&I, getOrigin(Op));
1956 // \brief Instrument vector convert instrinsic.
1958 // This function instruments intrinsics like cvtsi2ss:
1959 // %Out = int_xxx_cvtyyy(%ConvertOp)
1961 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1962 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1963 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1964 // elements from \p CopyOp.
1965 // In most cases conversion involves floating-point value which may trigger a
1966 // hardware exception when not fully initialized. For this reason we require
1967 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1968 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1969 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1970 // return a fully initialized value.
1971 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1972 IRBuilder<> IRB(&I);
1973 Value *CopyOp, *ConvertOp;
1975 switch (I.getNumArgOperands()) {
1977 CopyOp = I.getArgOperand(0);
1978 ConvertOp = I.getArgOperand(1);
1981 ConvertOp = I.getArgOperand(0);
1985 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1988 // The first *NumUsedElements* elements of ConvertOp are converted to the
1989 // same number of output elements. The rest of the output is copied from
1990 // CopyOp, or (if not available) filled with zeroes.
1991 // Combine shadow for elements of ConvertOp that are used in this operation,
1992 // and insert a check.
1993 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1994 // int->any conversion.
1995 Value *ConvertShadow = getShadow(ConvertOp);
1996 Value *AggShadow = nullptr;
1997 if (ConvertOp->getType()->isVectorTy()) {
1998 AggShadow = IRB.CreateExtractElement(
1999 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2000 for (int i = 1; i < NumUsedElements; ++i) {
2001 Value *MoreShadow = IRB.CreateExtractElement(
2002 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2003 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2006 AggShadow = ConvertShadow;
2008 assert(AggShadow->getType()->isIntegerTy());
2009 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2011 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2014 assert(CopyOp->getType() == I.getType());
2015 assert(CopyOp->getType()->isVectorTy());
2016 Value *ResultShadow = getShadow(CopyOp);
2017 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2018 for (int i = 0; i < NumUsedElements; ++i) {
2019 ResultShadow = IRB.CreateInsertElement(
2020 ResultShadow, ConstantInt::getNullValue(EltTy),
2021 ConstantInt::get(IRB.getInt32Ty(), i));
2023 setShadow(&I, ResultShadow);
2024 setOrigin(&I, getOrigin(CopyOp));
2026 setShadow(&I, getCleanShadow(&I));
2027 setOrigin(&I, getCleanOrigin());
2031 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2032 // zeroes if it is zero, and all ones otherwise.
2033 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2034 if (S->getType()->isVectorTy())
2035 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2036 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2037 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2038 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2041 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2042 Type *T = S->getType();
2043 assert(T->isVectorTy());
2044 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2045 return IRB.CreateSExt(S2, T);
2048 // \brief Instrument vector shift instrinsic.
2050 // This function instruments intrinsics like int_x86_avx2_psll_w.
2051 // Intrinsic shifts %In by %ShiftSize bits.
2052 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2053 // size, and the rest is ignored. Behavior is defined even if shift size is
2054 // greater than register (or field) width.
2055 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2056 assert(I.getNumArgOperands() == 2);
2057 IRBuilder<> IRB(&I);
2058 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2059 // Otherwise perform the same shift on S1.
2060 Value *S1 = getShadow(&I, 0);
2061 Value *S2 = getShadow(&I, 1);
2062 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2063 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2064 Value *V1 = I.getOperand(0);
2065 Value *V2 = I.getOperand(1);
2066 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
2067 IRB.CreateBitCast(S1, V1->getType()), V2);
2068 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2069 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2070 setOriginForNaryOp(I);
2073 // \brief Get an X86_MMX-sized vector type.
2074 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2075 const unsigned X86_MMXSizeInBits = 64;
2076 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2077 X86_MMXSizeInBits / EltSizeInBits);
2080 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2082 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2084 case llvm::Intrinsic::x86_sse2_packsswb_128:
2085 case llvm::Intrinsic::x86_sse2_packuswb_128:
2086 return llvm::Intrinsic::x86_sse2_packsswb_128;
2088 case llvm::Intrinsic::x86_sse2_packssdw_128:
2089 case llvm::Intrinsic::x86_sse41_packusdw:
2090 return llvm::Intrinsic::x86_sse2_packssdw_128;
2092 case llvm::Intrinsic::x86_avx2_packsswb:
2093 case llvm::Intrinsic::x86_avx2_packuswb:
2094 return llvm::Intrinsic::x86_avx2_packsswb;
2096 case llvm::Intrinsic::x86_avx2_packssdw:
2097 case llvm::Intrinsic::x86_avx2_packusdw:
2098 return llvm::Intrinsic::x86_avx2_packssdw;
2100 case llvm::Intrinsic::x86_mmx_packsswb:
2101 case llvm::Intrinsic::x86_mmx_packuswb:
2102 return llvm::Intrinsic::x86_mmx_packsswb;
2104 case llvm::Intrinsic::x86_mmx_packssdw:
2105 return llvm::Intrinsic::x86_mmx_packssdw;
2107 llvm_unreachable("unexpected intrinsic id");
2111 // \brief Instrument vector pack instrinsic.
2113 // This function instruments intrinsics like x86_mmx_packsswb, that
2114 // packs elements of 2 input vectors into half as many bits with saturation.
2115 // Shadow is propagated with the signed variant of the same intrinsic applied
2116 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2117 // EltSizeInBits is used only for x86mmx arguments.
2118 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2119 assert(I.getNumArgOperands() == 2);
2120 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2121 IRBuilder<> IRB(&I);
2122 Value *S1 = getShadow(&I, 0);
2123 Value *S2 = getShadow(&I, 1);
2124 assert(isX86_MMX || S1->getType()->isVectorTy());
2126 // SExt and ICmpNE below must apply to individual elements of input vectors.
2127 // In case of x86mmx arguments, cast them to appropriate vector types and
2129 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2131 S1 = IRB.CreateBitCast(S1, T);
2132 S2 = IRB.CreateBitCast(S2, T);
2134 Value *S1_ext = IRB.CreateSExt(
2135 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2136 Value *S2_ext = IRB.CreateSExt(
2137 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2139 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2140 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2141 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2144 Function *ShadowFn = Intrinsic::getDeclaration(
2145 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2147 Value *S = IRB.CreateCall2(ShadowFn, S1_ext, S2_ext, "_msprop_vector_pack");
2148 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2150 setOriginForNaryOp(I);
2153 // \brief Instrument sum-of-absolute-differencies intrinsic.
2154 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2155 const unsigned SignificantBitsPerResultElement = 16;
2156 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2157 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2158 unsigned ZeroBitsPerResultElement =
2159 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2161 IRBuilder<> IRB(&I);
2162 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2163 S = IRB.CreateBitCast(S, ResTy);
2164 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2166 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2167 S = IRB.CreateBitCast(S, getShadowTy(&I));
2169 setOriginForNaryOp(I);
2172 // \brief Instrument multiply-add intrinsic.
2173 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2174 unsigned EltSizeInBits = 0) {
2175 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2176 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2177 IRBuilder<> IRB(&I);
2178 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2179 S = IRB.CreateBitCast(S, ResTy);
2180 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2182 S = IRB.CreateBitCast(S, getShadowTy(&I));
2184 setOriginForNaryOp(I);
2187 void visitIntrinsicInst(IntrinsicInst &I) {
2188 switch (I.getIntrinsicID()) {
2189 case llvm::Intrinsic::bswap:
2192 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2193 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2194 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2195 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2196 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2197 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2198 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2199 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2200 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2201 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2202 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2203 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2204 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2205 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2206 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2207 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2208 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2209 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2210 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2211 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2212 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2213 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2214 case llvm::Intrinsic::x86_sse_cvtss2si64:
2215 case llvm::Intrinsic::x86_sse_cvtss2si:
2216 case llvm::Intrinsic::x86_sse_cvttss2si64:
2217 case llvm::Intrinsic::x86_sse_cvttss2si:
2218 handleVectorConvertIntrinsic(I, 1);
2220 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2221 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2222 case llvm::Intrinsic::x86_sse_cvtps2pi:
2223 case llvm::Intrinsic::x86_sse_cvttps2pi:
2224 handleVectorConvertIntrinsic(I, 2);
2226 case llvm::Intrinsic::x86_avx512_psll_dq:
2227 case llvm::Intrinsic::x86_avx512_psrl_dq:
2228 case llvm::Intrinsic::x86_avx2_psll_w:
2229 case llvm::Intrinsic::x86_avx2_psll_d:
2230 case llvm::Intrinsic::x86_avx2_psll_q:
2231 case llvm::Intrinsic::x86_avx2_pslli_w:
2232 case llvm::Intrinsic::x86_avx2_pslli_d:
2233 case llvm::Intrinsic::x86_avx2_pslli_q:
2234 case llvm::Intrinsic::x86_avx2_psll_dq:
2235 case llvm::Intrinsic::x86_avx2_psrl_w:
2236 case llvm::Intrinsic::x86_avx2_psrl_d:
2237 case llvm::Intrinsic::x86_avx2_psrl_q:
2238 case llvm::Intrinsic::x86_avx2_psra_w:
2239 case llvm::Intrinsic::x86_avx2_psra_d:
2240 case llvm::Intrinsic::x86_avx2_psrli_w:
2241 case llvm::Intrinsic::x86_avx2_psrli_d:
2242 case llvm::Intrinsic::x86_avx2_psrli_q:
2243 case llvm::Intrinsic::x86_avx2_psrai_w:
2244 case llvm::Intrinsic::x86_avx2_psrai_d:
2245 case llvm::Intrinsic::x86_avx2_psrl_dq:
2246 case llvm::Intrinsic::x86_sse2_psll_w:
2247 case llvm::Intrinsic::x86_sse2_psll_d:
2248 case llvm::Intrinsic::x86_sse2_psll_q:
2249 case llvm::Intrinsic::x86_sse2_pslli_w:
2250 case llvm::Intrinsic::x86_sse2_pslli_d:
2251 case llvm::Intrinsic::x86_sse2_pslli_q:
2252 case llvm::Intrinsic::x86_sse2_psll_dq:
2253 case llvm::Intrinsic::x86_sse2_psrl_w:
2254 case llvm::Intrinsic::x86_sse2_psrl_d:
2255 case llvm::Intrinsic::x86_sse2_psrl_q:
2256 case llvm::Intrinsic::x86_sse2_psra_w:
2257 case llvm::Intrinsic::x86_sse2_psra_d:
2258 case llvm::Intrinsic::x86_sse2_psrli_w:
2259 case llvm::Intrinsic::x86_sse2_psrli_d:
2260 case llvm::Intrinsic::x86_sse2_psrli_q:
2261 case llvm::Intrinsic::x86_sse2_psrai_w:
2262 case llvm::Intrinsic::x86_sse2_psrai_d:
2263 case llvm::Intrinsic::x86_sse2_psrl_dq:
2264 case llvm::Intrinsic::x86_mmx_psll_w:
2265 case llvm::Intrinsic::x86_mmx_psll_d:
2266 case llvm::Intrinsic::x86_mmx_psll_q:
2267 case llvm::Intrinsic::x86_mmx_pslli_w:
2268 case llvm::Intrinsic::x86_mmx_pslli_d:
2269 case llvm::Intrinsic::x86_mmx_pslli_q:
2270 case llvm::Intrinsic::x86_mmx_psrl_w:
2271 case llvm::Intrinsic::x86_mmx_psrl_d:
2272 case llvm::Intrinsic::x86_mmx_psrl_q:
2273 case llvm::Intrinsic::x86_mmx_psra_w:
2274 case llvm::Intrinsic::x86_mmx_psra_d:
2275 case llvm::Intrinsic::x86_mmx_psrli_w:
2276 case llvm::Intrinsic::x86_mmx_psrli_d:
2277 case llvm::Intrinsic::x86_mmx_psrli_q:
2278 case llvm::Intrinsic::x86_mmx_psrai_w:
2279 case llvm::Intrinsic::x86_mmx_psrai_d:
2280 handleVectorShiftIntrinsic(I, /* Variable */ false);
2282 case llvm::Intrinsic::x86_avx2_psllv_d:
2283 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2284 case llvm::Intrinsic::x86_avx2_psllv_q:
2285 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2286 case llvm::Intrinsic::x86_avx2_psrlv_d:
2287 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2288 case llvm::Intrinsic::x86_avx2_psrlv_q:
2289 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2290 case llvm::Intrinsic::x86_avx2_psrav_d:
2291 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2292 handleVectorShiftIntrinsic(I, /* Variable */ true);
2295 // Byte shifts are not implemented.
2296 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
2297 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
2298 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
2299 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
2300 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
2301 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
2303 case llvm::Intrinsic::x86_sse2_packsswb_128:
2304 case llvm::Intrinsic::x86_sse2_packssdw_128:
2305 case llvm::Intrinsic::x86_sse2_packuswb_128:
2306 case llvm::Intrinsic::x86_sse41_packusdw:
2307 case llvm::Intrinsic::x86_avx2_packsswb:
2308 case llvm::Intrinsic::x86_avx2_packssdw:
2309 case llvm::Intrinsic::x86_avx2_packuswb:
2310 case llvm::Intrinsic::x86_avx2_packusdw:
2311 handleVectorPackIntrinsic(I);
2314 case llvm::Intrinsic::x86_mmx_packsswb:
2315 case llvm::Intrinsic::x86_mmx_packuswb:
2316 handleVectorPackIntrinsic(I, 16);
2319 case llvm::Intrinsic::x86_mmx_packssdw:
2320 handleVectorPackIntrinsic(I, 32);
2323 case llvm::Intrinsic::x86_mmx_psad_bw:
2324 case llvm::Intrinsic::x86_sse2_psad_bw:
2325 case llvm::Intrinsic::x86_avx2_psad_bw:
2326 handleVectorSadIntrinsic(I);
2329 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2330 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2331 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2332 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2333 handleVectorPmaddIntrinsic(I);
2336 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2337 handleVectorPmaddIntrinsic(I, 8);
2340 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2341 handleVectorPmaddIntrinsic(I, 16);
2345 if (!handleUnknownIntrinsic(I))
2346 visitInstruction(I);
2351 void visitCallSite(CallSite CS) {
2352 Instruction &I = *CS.getInstruction();
2353 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2355 CallInst *Call = cast<CallInst>(&I);
2357 // For inline asm, do the usual thing: check argument shadow and mark all
2358 // outputs as clean. Note that any side effects of the inline asm that are
2359 // not immediately visible in its constraints are not handled.
2360 if (Call->isInlineAsm()) {
2361 visitInstruction(I);
2365 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2367 // We are going to insert code that relies on the fact that the callee
2368 // will become a non-readonly function after it is instrumented by us. To
2369 // prevent this code from being optimized out, mark that function
2370 // non-readonly in advance.
2371 if (Function *Func = Call->getCalledFunction()) {
2372 // Clear out readonly/readnone attributes.
2374 B.addAttribute(Attribute::ReadOnly)
2375 .addAttribute(Attribute::ReadNone);
2376 Func->removeAttributes(AttributeSet::FunctionIndex,
2377 AttributeSet::get(Func->getContext(),
2378 AttributeSet::FunctionIndex,
2382 IRBuilder<> IRB(&I);
2384 unsigned ArgOffset = 0;
2385 DEBUG(dbgs() << " CallSite: " << I << "\n");
2386 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2387 ArgIt != End; ++ArgIt) {
2389 unsigned i = ArgIt - CS.arg_begin();
2390 if (!A->getType()->isSized()) {
2391 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2395 Value *Store = nullptr;
2396 // Compute the Shadow for arg even if it is ByVal, because
2397 // in that case getShadow() will copy the actual arg shadow to
2398 // __msan_param_tls.
2399 Value *ArgShadow = getShadow(A);
2400 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2401 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2402 " Shadow: " << *ArgShadow << "\n");
2403 bool ArgIsInitialized = false;
2404 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2405 assert(A->getType()->isPointerTy() &&
2406 "ByVal argument is not a pointer!");
2407 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2408 if (ArgOffset + Size > kParamTLSSize) break;
2409 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2410 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2411 Store = IRB.CreateMemCpy(ArgShadowBase,
2412 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2415 Size = MS.DL->getTypeAllocSize(A->getType());
2416 if (ArgOffset + Size > kParamTLSSize) break;
2417 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2418 kShadowTLSAlignment);
2419 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2420 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2422 if (MS.TrackOrigins && !ArgIsInitialized)
2423 IRB.CreateStore(getOrigin(A),
2424 getOriginPtrForArgument(A, IRB, ArgOffset));
2426 assert(Size != 0 && Store != nullptr);
2427 DEBUG(dbgs() << " Param:" << *Store << "\n");
2428 ArgOffset += RoundUpToAlignment(Size, 8);
2430 DEBUG(dbgs() << " done with call args\n");
2433 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2434 if (FT->isVarArg()) {
2435 VAHelper->visitCallSite(CS, IRB);
2438 // Now, get the shadow for the RetVal.
2439 if (!I.getType()->isSized()) return;
2440 IRBuilder<> IRBBefore(&I);
2441 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2442 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2443 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2444 Instruction *NextInsn = nullptr;
2446 NextInsn = I.getNextNode();
2448 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2449 if (!NormalDest->getSinglePredecessor()) {
2450 // FIXME: this case is tricky, so we are just conservative here.
2451 // Perhaps we need to split the edge between this BB and NormalDest,
2452 // but a naive attempt to use SplitEdge leads to a crash.
2453 setShadow(&I, getCleanShadow(&I));
2454 setOrigin(&I, getCleanOrigin());
2457 NextInsn = NormalDest->getFirstInsertionPt();
2459 "Could not find insertion point for retval shadow load");
2461 IRBuilder<> IRBAfter(NextInsn);
2462 Value *RetvalShadow =
2463 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2464 kShadowTLSAlignment, "_msret");
2465 setShadow(&I, RetvalShadow);
2466 if (MS.TrackOrigins)
2467 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2470 void visitReturnInst(ReturnInst &I) {
2471 IRBuilder<> IRB(&I);
2472 Value *RetVal = I.getReturnValue();
2473 if (!RetVal) return;
2474 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2475 if (CheckReturnValue) {
2476 insertShadowCheck(RetVal, &I);
2477 Value *Shadow = getCleanShadow(RetVal);
2478 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2480 Value *Shadow = getShadow(RetVal);
2481 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2482 // FIXME: make it conditional if ClStoreCleanOrigin==0
2483 if (MS.TrackOrigins)
2484 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2488 void visitPHINode(PHINode &I) {
2489 IRBuilder<> IRB(&I);
2490 if (!PropagateShadow) {
2491 setShadow(&I, getCleanShadow(&I));
2492 setOrigin(&I, getCleanOrigin());
2496 ShadowPHINodes.push_back(&I);
2497 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2499 if (MS.TrackOrigins)
2500 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2504 void visitAllocaInst(AllocaInst &I) {
2505 setShadow(&I, getCleanShadow(&I));
2506 setOrigin(&I, getCleanOrigin());
2507 IRBuilder<> IRB(I.getNextNode());
2508 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2509 if (PoisonStack && ClPoisonStackWithCall) {
2510 IRB.CreateCall2(MS.MsanPoisonStackFn,
2511 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2512 ConstantInt::get(MS.IntptrTy, Size));
2514 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2515 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2516 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2519 if (PoisonStack && MS.TrackOrigins) {
2520 SmallString<2048> StackDescriptionStorage;
2521 raw_svector_ostream StackDescription(StackDescriptionStorage);
2522 // We create a string with a description of the stack allocation and
2523 // pass it into __msan_set_alloca_origin.
2524 // It will be printed by the run-time if stack-originated UMR is found.
2525 // The first 4 bytes of the string are set to '----' and will be replaced
2526 // by __msan_va_arg_overflow_size_tls at the first call.
2527 StackDescription << "----" << I.getName() << "@" << F.getName();
2529 createPrivateNonConstGlobalForString(*F.getParent(),
2530 StackDescription.str());
2532 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2533 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2534 ConstantInt::get(MS.IntptrTy, Size),
2535 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2536 IRB.CreatePointerCast(&F, MS.IntptrTy));
2540 void visitSelectInst(SelectInst& I) {
2541 IRBuilder<> IRB(&I);
2542 // a = select b, c, d
2543 Value *B = I.getCondition();
2544 Value *C = I.getTrueValue();
2545 Value *D = I.getFalseValue();
2546 Value *Sb = getShadow(B);
2547 Value *Sc = getShadow(C);
2548 Value *Sd = getShadow(D);
2550 // Result shadow if condition shadow is 0.
2551 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2553 if (I.getType()->isAggregateType()) {
2554 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2555 // an extra "select". This results in much more compact IR.
2556 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2557 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2559 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2560 // If Sb (condition is poisoned), look for bits in c and d that are equal
2561 // and both unpoisoned.
2562 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2564 // Cast arguments to shadow-compatible type.
2565 C = CreateAppToShadowCast(IRB, C);
2566 D = CreateAppToShadowCast(IRB, D);
2568 // Result shadow if condition shadow is 1.
2569 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2571 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2573 if (MS.TrackOrigins) {
2574 // Origins are always i32, so any vector conditions must be flattened.
2575 // FIXME: consider tracking vector origins for app vectors?
2576 if (B->getType()->isVectorTy()) {
2577 Type *FlatTy = getShadowTyNoVec(B->getType());
2578 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2579 ConstantInt::getNullValue(FlatTy));
2580 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2581 ConstantInt::getNullValue(FlatTy));
2583 // a = select b, c, d
2584 // Oa = Sb ? Ob : (b ? Oc : Od)
2586 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2587 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2588 getOrigin(I.getFalseValue()))));
2592 void visitLandingPadInst(LandingPadInst &I) {
2594 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2595 setShadow(&I, getCleanShadow(&I));
2596 setOrigin(&I, getCleanOrigin());
2599 void visitGetElementPtrInst(GetElementPtrInst &I) {
2603 void visitExtractValueInst(ExtractValueInst &I) {
2604 IRBuilder<> IRB(&I);
2605 Value *Agg = I.getAggregateOperand();
2606 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2607 Value *AggShadow = getShadow(Agg);
2608 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2609 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2610 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2611 setShadow(&I, ResShadow);
2612 setOriginForNaryOp(I);
2615 void visitInsertValueInst(InsertValueInst &I) {
2616 IRBuilder<> IRB(&I);
2617 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2618 Value *AggShadow = getShadow(I.getAggregateOperand());
2619 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2620 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2621 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2622 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2623 DEBUG(dbgs() << " Res: " << *Res << "\n");
2625 setOriginForNaryOp(I);
2628 void dumpInst(Instruction &I) {
2629 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2630 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2632 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2634 errs() << "QQQ " << I << "\n";
2637 void visitResumeInst(ResumeInst &I) {
2638 DEBUG(dbgs() << "Resume: " << I << "\n");
2639 // Nothing to do here.
2642 void visitInstruction(Instruction &I) {
2643 // Everything else: stop propagating and check for poisoned shadow.
2644 if (ClDumpStrictInstructions)
2646 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2647 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2648 insertShadowCheck(I.getOperand(i), &I);
2649 setShadow(&I, getCleanShadow(&I));
2650 setOrigin(&I, getCleanOrigin());
2654 /// \brief AMD64-specific implementation of VarArgHelper.
2655 struct VarArgAMD64Helper : public VarArgHelper {
2656 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2657 // See a comment in visitCallSite for more details.
2658 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2659 static const unsigned AMD64FpEndOffset = 176;
2662 MemorySanitizer &MS;
2663 MemorySanitizerVisitor &MSV;
2664 Value *VAArgTLSCopy;
2665 Value *VAArgOverflowSize;
2667 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2669 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2670 MemorySanitizerVisitor &MSV)
2671 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2672 VAArgOverflowSize(nullptr) {}
2674 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2676 ArgKind classifyArgument(Value* arg) {
2677 // A very rough approximation of X86_64 argument classification rules.
2678 Type *T = arg->getType();
2679 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2680 return AK_FloatingPoint;
2681 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2682 return AK_GeneralPurpose;
2683 if (T->isPointerTy())
2684 return AK_GeneralPurpose;
2688 // For VarArg functions, store the argument shadow in an ABI-specific format
2689 // that corresponds to va_list layout.
2690 // We do this because Clang lowers va_arg in the frontend, and this pass
2691 // only sees the low level code that deals with va_list internals.
2692 // A much easier alternative (provided that Clang emits va_arg instructions)
2693 // would have been to associate each live instance of va_list with a copy of
2694 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2696 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2697 unsigned GpOffset = 0;
2698 unsigned FpOffset = AMD64GpEndOffset;
2699 unsigned OverflowOffset = AMD64FpEndOffset;
2700 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2701 ArgIt != End; ++ArgIt) {
2703 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2704 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2706 // ByVal arguments always go to the overflow area.
2707 assert(A->getType()->isPointerTy());
2708 Type *RealTy = A->getType()->getPointerElementType();
2709 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2710 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2711 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2712 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2713 ArgSize, kShadowTLSAlignment);
2715 ArgKind AK = classifyArgument(A);
2716 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2718 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2722 case AK_GeneralPurpose:
2723 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2726 case AK_FloatingPoint:
2727 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2731 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2732 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2733 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2735 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2738 Constant *OverflowSize =
2739 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2740 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2743 /// \brief Compute the shadow address for a given va_arg.
2744 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2746 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2747 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2748 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2752 void visitVAStartInst(VAStartInst &I) override {
2753 IRBuilder<> IRB(&I);
2754 VAStartInstrumentationList.push_back(&I);
2755 Value *VAListTag = I.getArgOperand(0);
2756 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2758 // Unpoison the whole __va_list_tag.
2759 // FIXME: magic ABI constants.
2760 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2761 /* size */24, /* alignment */8, false);
2764 void visitVACopyInst(VACopyInst &I) override {
2765 IRBuilder<> IRB(&I);
2766 Value *VAListTag = I.getArgOperand(0);
2767 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2769 // Unpoison the whole __va_list_tag.
2770 // FIXME: magic ABI constants.
2771 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2772 /* size */24, /* alignment */8, false);
2775 void finalizeInstrumentation() override {
2776 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2777 "finalizeInstrumentation called twice");
2778 if (!VAStartInstrumentationList.empty()) {
2779 // If there is a va_start in this function, make a backup copy of
2780 // va_arg_tls somewhere in the function entry block.
2781 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2782 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2784 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2786 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2787 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2790 // Instrument va_start.
2791 // Copy va_list shadow from the backup copy of the TLS contents.
2792 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2793 CallInst *OrigInst = VAStartInstrumentationList[i];
2794 IRBuilder<> IRB(OrigInst->getNextNode());
2795 Value *VAListTag = OrigInst->getArgOperand(0);
2797 Value *RegSaveAreaPtrPtr =
2799 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2800 ConstantInt::get(MS.IntptrTy, 16)),
2801 Type::getInt64PtrTy(*MS.C));
2802 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2803 Value *RegSaveAreaShadowPtr =
2804 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2805 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2806 AMD64FpEndOffset, 16);
2808 Value *OverflowArgAreaPtrPtr =
2810 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2811 ConstantInt::get(MS.IntptrTy, 8)),
2812 Type::getInt64PtrTy(*MS.C));
2813 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2814 Value *OverflowArgAreaShadowPtr =
2815 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2816 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2817 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2822 /// \brief A no-op implementation of VarArgHelper.
2823 struct VarArgNoOpHelper : public VarArgHelper {
2824 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2825 MemorySanitizerVisitor &MSV) {}
2827 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2829 void visitVAStartInst(VAStartInst &I) override {}
2831 void visitVACopyInst(VACopyInst &I) override {}
2833 void finalizeInstrumentation() override {}
2836 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2837 MemorySanitizerVisitor &Visitor) {
2838 // VarArg handling is only implemented on AMD64. False positives are possible
2839 // on other platforms.
2840 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2841 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2842 return new VarArgAMD64Helper(Func, Msan, Visitor);
2844 return new VarArgNoOpHelper(Func, Msan, Visitor);
2849 bool MemorySanitizer::runOnFunction(Function &F) {
2850 MemorySanitizerVisitor Visitor(F, *this);
2852 // Clear out readonly/readnone attributes.
2854 B.addAttribute(Attribute::ReadOnly)
2855 .addAttribute(Attribute::ReadNone);
2856 F.removeAttributes(AttributeSet::FunctionIndex,
2857 AttributeSet::get(F.getContext(),
2858 AttributeSet::FunctionIndex, B));
2860 return Visitor.runOnFunction();