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 kOriginSize = 4;
124 static const unsigned kMinOriginAlignment = 4;
125 static const unsigned kShadowTLSAlignment = 8;
127 // These constants must be kept in sync with the ones in msan.h.
128 static const unsigned kParamTLSSize = 800;
129 static const unsigned kRetvalTLSSize = 800;
131 // Accesses sizes are powers of two: 1, 2, 4, 8.
132 static const size_t kNumberOfAccessSizes = 4;
134 /// \brief Track origins of uninitialized values.
136 /// Adds a section to MemorySanitizer report that points to the allocation
137 /// (stack or heap) the uninitialized bits came from originally.
138 static cl::opt<int> ClTrackOrigins("msan-track-origins",
139 cl::desc("Track origins (allocation sites) of poisoned memory"),
140 cl::Hidden, cl::init(0));
141 static cl::opt<bool> ClKeepGoing("msan-keep-going",
142 cl::desc("keep going after reporting a UMR"),
143 cl::Hidden, cl::init(false));
144 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
145 cl::desc("poison uninitialized stack variables"),
146 cl::Hidden, cl::init(true));
147 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
148 cl::desc("poison uninitialized stack variables with a call"),
149 cl::Hidden, cl::init(false));
150 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
151 cl::desc("poison uninitialized stack variables with the given patter"),
152 cl::Hidden, cl::init(0xff));
153 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
154 cl::desc("poison undef temps"),
155 cl::Hidden, cl::init(true));
157 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
158 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
159 cl::Hidden, cl::init(true));
161 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
162 cl::desc("exact handling of relational integer ICmp"),
163 cl::Hidden, cl::init(false));
165 // This flag controls whether we check the shadow of the address
166 // operand of load or store. Such bugs are very rare, since load from
167 // a garbage address typically results in SEGV, but still happen
168 // (e.g. only lower bits of address are garbage, or the access happens
169 // early at program startup where malloc-ed memory is more likely to
170 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
171 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
172 cl::desc("report accesses through a pointer which has poisoned shadow"),
173 cl::Hidden, cl::init(true));
175 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
176 cl::desc("print out instructions with default strict semantics"),
177 cl::Hidden, cl::init(false));
179 static cl::opt<int> ClInstrumentationWithCallThreshold(
180 "msan-instrumentation-with-call-threshold",
182 "If the function being instrumented requires more than "
183 "this number of checks and origin stores, use callbacks instead of "
184 "inline checks (-1 means never use callbacks)."),
185 cl::Hidden, cl::init(3500));
187 // This is an experiment to enable handling of cases where shadow is a non-zero
188 // compile-time constant. For some unexplainable reason they were silently
189 // ignored in the instrumentation.
190 static cl::opt<bool> ClCheckConstantShadow("msan-check-constant-shadow",
191 cl::desc("Insert checks for constant shadow values"),
192 cl::Hidden, cl::init(false));
196 // Memory map parameters used in application-to-shadow address calculation.
197 // Offset = (Addr & ~AndMask) ^ XorMask
198 // Shadow = ShadowBase + Offset
199 // Origin = OriginBase + Offset
200 struct MemoryMapParams {
207 struct PlatformMemoryMapParams {
208 const MemoryMapParams *bits32;
209 const MemoryMapParams *bits64;
213 static const MemoryMapParams Linux_I386_MemoryMapParams = {
214 0x000080000000, // AndMask
215 0, // XorMask (not used)
216 0, // ShadowBase (not used)
217 0x000040000000, // OriginBase
221 static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
222 0x400000000000, // AndMask
223 0, // XorMask (not used)
224 0, // ShadowBase (not used)
225 0x200000000000, // OriginBase
229 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
230 0x004000000000, // AndMask
231 0, // XorMask (not used)
232 0, // ShadowBase (not used)
233 0x002000000000, // OriginBase
237 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
238 0x000180000000, // AndMask
239 0x000040000000, // XorMask
240 0x000020000000, // ShadowBase
241 0x000700000000, // OriginBase
245 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
246 0xc00000000000, // AndMask
247 0x200000000000, // XorMask
248 0x100000000000, // ShadowBase
249 0x380000000000, // OriginBase
252 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
253 &Linux_I386_MemoryMapParams,
254 &Linux_X86_64_MemoryMapParams,
257 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
259 &Linux_MIPS64_MemoryMapParams,
262 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
263 &FreeBSD_I386_MemoryMapParams,
264 &FreeBSD_X86_64_MemoryMapParams,
267 /// \brief An instrumentation pass implementing detection of uninitialized
270 /// MemorySanitizer: instrument the code in module to find
271 /// uninitialized reads.
272 class MemorySanitizer : public FunctionPass {
274 MemorySanitizer(int TrackOrigins = 0)
276 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
278 WarningFn(nullptr) {}
279 const char *getPassName() const override { return "MemorySanitizer"; }
280 bool runOnFunction(Function &F) override;
281 bool doInitialization(Module &M) override;
282 static char ID; // Pass identification, replacement for typeid.
285 void initializeCallbacks(Module &M);
287 /// \brief Track origins (allocation points) of uninitialized values.
290 const DataLayout *DL;
294 /// \brief Thread-local shadow storage for function parameters.
295 GlobalVariable *ParamTLS;
296 /// \brief Thread-local origin storage for function parameters.
297 GlobalVariable *ParamOriginTLS;
298 /// \brief Thread-local shadow storage for function return value.
299 GlobalVariable *RetvalTLS;
300 /// \brief Thread-local origin storage for function return value.
301 GlobalVariable *RetvalOriginTLS;
302 /// \brief Thread-local shadow storage for in-register va_arg function
303 /// parameters (x86_64-specific).
304 GlobalVariable *VAArgTLS;
305 /// \brief Thread-local shadow storage for va_arg overflow area
306 /// (x86_64-specific).
307 GlobalVariable *VAArgOverflowSizeTLS;
308 /// \brief Thread-local space used to pass origin value to the UMR reporting
310 GlobalVariable *OriginTLS;
312 /// \brief The run-time callback to print a warning.
314 // These arrays are indexed by log2(AccessSize).
315 Value *MaybeWarningFn[kNumberOfAccessSizes];
316 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
318 /// \brief Run-time helper that generates a new origin value for a stack
320 Value *MsanSetAllocaOrigin4Fn;
321 /// \brief Run-time helper that poisons stack on function entry.
322 Value *MsanPoisonStackFn;
323 /// \brief Run-time helper that records a store (or any event) of an
324 /// uninitialized value and returns an updated origin id encoding this info.
325 Value *MsanChainOriginFn;
326 /// \brief MSan runtime replacements for memmove, memcpy and memset.
327 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
329 /// \brief Memory map parameters used in application-to-shadow calculation.
330 const MemoryMapParams *MapParams;
332 MDNode *ColdCallWeights;
333 /// \brief Branch weights for origin store.
334 MDNode *OriginStoreWeights;
335 /// \brief An empty volatile inline asm that prevents callback merge.
338 friend struct MemorySanitizerVisitor;
339 friend struct VarArgAMD64Helper;
340 friend struct VarArgMIPS64Helper;
344 char MemorySanitizer::ID = 0;
345 INITIALIZE_PASS(MemorySanitizer, "msan",
346 "MemorySanitizer: detects uninitialized reads.",
349 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
350 return new MemorySanitizer(TrackOrigins);
353 /// \brief Create a non-const global initialized with the given string.
355 /// Creates a writable global for Str so that we can pass it to the
356 /// run-time lib. Runtime uses first 4 bytes of the string to store the
357 /// frame ID, so the string needs to be mutable.
358 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
360 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
361 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
362 GlobalValue::PrivateLinkage, StrConst, "");
366 /// \brief Insert extern declaration of runtime-provided functions and globals.
367 void MemorySanitizer::initializeCallbacks(Module &M) {
368 // Only do this once.
373 // Create the callback.
374 // FIXME: this function should have "Cold" calling conv,
375 // which is not yet implemented.
376 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
377 : "__msan_warning_noreturn";
378 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), nullptr);
380 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
382 unsigned AccessSize = 1 << AccessSizeIndex;
383 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
384 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
385 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
386 IRB.getInt32Ty(), nullptr);
388 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
389 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
390 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
391 IRB.getInt8PtrTy(), IRB.getInt32Ty(), nullptr);
394 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
395 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
396 IRB.getInt8PtrTy(), IntptrTy, nullptr);
398 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
399 IRB.getInt8PtrTy(), IntptrTy, nullptr);
400 MsanChainOriginFn = M.getOrInsertFunction(
401 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), nullptr);
402 MemmoveFn = M.getOrInsertFunction(
403 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
404 IRB.getInt8PtrTy(), IntptrTy, nullptr);
405 MemcpyFn = M.getOrInsertFunction(
406 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
408 MemsetFn = M.getOrInsertFunction(
409 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
413 RetvalTLS = new GlobalVariable(
414 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
415 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
416 GlobalVariable::InitialExecTLSModel);
417 RetvalOriginTLS = new GlobalVariable(
418 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
419 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
421 ParamTLS = new GlobalVariable(
422 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
423 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
424 GlobalVariable::InitialExecTLSModel);
425 ParamOriginTLS = new GlobalVariable(
426 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
427 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
428 nullptr, GlobalVariable::InitialExecTLSModel);
430 VAArgTLS = new GlobalVariable(
431 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
432 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
433 GlobalVariable::InitialExecTLSModel);
434 VAArgOverflowSizeTLS = new GlobalVariable(
435 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
436 "__msan_va_arg_overflow_size_tls", nullptr,
437 GlobalVariable::InitialExecTLSModel);
438 OriginTLS = new GlobalVariable(
439 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
440 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
442 // We insert an empty inline asm after __msan_report* to avoid callback merge.
443 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
444 StringRef(""), StringRef(""),
445 /*hasSideEffects=*/true);
448 /// \brief Module-level initialization.
450 /// inserts a call to __msan_init to the module's constructor list.
451 bool MemorySanitizer::doInitialization(Module &M) {
452 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
454 report_fatal_error("data layout missing");
455 DL = &DLP->getDataLayout();
457 Triple TargetTriple(M.getTargetTriple());
458 switch (TargetTriple.getOS()) {
459 case Triple::FreeBSD:
460 switch (TargetTriple.getArch()) {
462 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
465 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
468 report_fatal_error("unsupported architecture");
472 switch (TargetTriple.getArch()) {
474 MapParams = Linux_X86_MemoryMapParams.bits64;
477 MapParams = Linux_X86_MemoryMapParams.bits32;
480 case Triple::mips64el:
481 MapParams = Linux_MIPS_MemoryMapParams.bits64;
484 report_fatal_error("unsupported architecture");
488 report_fatal_error("unsupported operating system");
491 C = &(M.getContext());
493 IntptrTy = IRB.getIntPtrTy(DL);
494 OriginTy = IRB.getInt32Ty();
496 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
497 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
499 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
500 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
501 "__msan_init", IRB.getVoidTy(), nullptr)), 0);
504 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
505 IRB.getInt32(TrackOrigins), "__msan_track_origins");
508 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
509 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
516 /// \brief A helper class that handles instrumentation of VarArg
517 /// functions on a particular platform.
519 /// Implementations are expected to insert the instrumentation
520 /// necessary to propagate argument shadow through VarArg function
521 /// calls. Visit* methods are called during an InstVisitor pass over
522 /// the function, and should avoid creating new basic blocks. A new
523 /// instance of this class is created for each instrumented function.
524 struct VarArgHelper {
525 /// \brief Visit a CallSite.
526 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
528 /// \brief Visit a va_start call.
529 virtual void visitVAStartInst(VAStartInst &I) = 0;
531 /// \brief Visit a va_copy call.
532 virtual void visitVACopyInst(VACopyInst &I) = 0;
534 /// \brief Finalize function instrumentation.
536 /// This method is called after visiting all interesting (see above)
537 /// instructions in a function.
538 virtual void finalizeInstrumentation() = 0;
540 virtual ~VarArgHelper() {}
543 struct MemorySanitizerVisitor;
546 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
547 MemorySanitizerVisitor &Visitor);
549 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
550 if (TypeSize <= 8) return 0;
551 return Log2_32_Ceil(TypeSize / 8);
554 /// This class does all the work for a given function. Store and Load
555 /// instructions store and load corresponding shadow and origin
556 /// values. Most instructions propagate shadow from arguments to their
557 /// return values. Certain instructions (most importantly, BranchInst)
558 /// test their argument shadow and print reports (with a runtime call) if it's
560 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
563 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
564 ValueMap<Value*, Value*> ShadowMap, OriginMap;
565 std::unique_ptr<VarArgHelper> VAHelper;
567 // The following flags disable parts of MSan instrumentation based on
568 // blacklist contents and command-line options.
570 bool PropagateShadow;
573 bool CheckReturnValue;
575 struct ShadowOriginAndInsertPoint {
578 Instruction *OrigIns;
579 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
580 : Shadow(S), Origin(O), OrigIns(I) { }
582 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
583 SmallVector<Instruction*, 16> StoreList;
585 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
586 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
587 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
588 InsertChecks = SanitizeFunction;
589 PropagateShadow = SanitizeFunction;
590 PoisonStack = SanitizeFunction && ClPoisonStack;
591 PoisonUndef = SanitizeFunction && ClPoisonUndef;
592 // FIXME: Consider using SpecialCaseList to specify a list of functions that
593 // must always return fully initialized values. For now, we hardcode "main".
594 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
596 DEBUG(if (!InsertChecks)
597 dbgs() << "MemorySanitizer is not inserting checks into '"
598 << F.getName() << "'\n");
601 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
602 if (MS.TrackOrigins <= 1) return V;
603 return IRB.CreateCall(MS.MsanChainOriginFn, V);
606 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
607 unsigned IntptrSize = MS.DL->getTypeStoreSize(MS.IntptrTy);
608 if (IntptrSize == kOriginSize) return Origin;
609 assert(IntptrSize == kOriginSize * 2);
610 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
611 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
614 /// \brief Fill memory range with the given origin value.
615 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
616 unsigned Size, unsigned Alignment) {
617 unsigned IntptrAlignment = MS.DL->getABITypeAlignment(MS.IntptrTy);
618 unsigned IntptrSize = MS.DL->getTypeStoreSize(MS.IntptrTy);
619 assert(IntptrAlignment >= kMinOriginAlignment);
620 assert(IntptrSize >= kOriginSize);
623 unsigned CurrentAlignment = Alignment;
624 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
625 Value *IntptrOrigin = originToIntptr(IRB, Origin);
626 Value *IntptrOriginPtr =
627 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
628 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
630 i ? IRB.CreateConstGEP1_32(IntptrOriginPtr, i) : IntptrOriginPtr;
631 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
632 Ofs += IntptrSize / kOriginSize;
633 CurrentAlignment = IntptrAlignment;
637 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
638 Value *GEP = i ? IRB.CreateConstGEP1_32(OriginPtr, i) : OriginPtr;
639 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
640 CurrentAlignment = kMinOriginAlignment;
644 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
645 unsigned Alignment, bool AsCall) {
646 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
647 unsigned StoreSize = MS.DL->getTypeStoreSize(Shadow->getType());
648 if (isa<StructType>(Shadow->getType())) {
649 paintOrigin(IRB, updateOrigin(Origin, IRB),
650 getOriginPtr(Addr, IRB, Alignment), StoreSize,
653 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
654 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
655 if (ConstantShadow) {
656 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
657 paintOrigin(IRB, updateOrigin(Origin, IRB),
658 getOriginPtr(Addr, IRB, Alignment), StoreSize,
663 unsigned TypeSizeInBits =
664 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
665 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
666 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
667 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
668 Value *ConvertedShadow2 = IRB.CreateZExt(
669 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
670 IRB.CreateCall3(Fn, ConvertedShadow2,
671 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
674 Value *Cmp = IRB.CreateICmpNE(
675 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
676 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
677 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
678 IRBuilder<> IRBNew(CheckTerm);
679 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
680 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
686 void materializeStores(bool InstrumentWithCalls) {
687 for (auto Inst : StoreList) {
688 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
690 IRBuilder<> IRB(&SI);
691 Value *Val = SI.getValueOperand();
692 Value *Addr = SI.getPointerOperand();
693 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
694 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
697 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
698 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
701 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
703 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
705 if (MS.TrackOrigins && !SI.isAtomic())
706 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
707 InstrumentWithCalls);
711 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
713 IRBuilder<> IRB(OrigIns);
714 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
715 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
716 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
718 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
719 if (ConstantShadow) {
720 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
721 if (MS.TrackOrigins) {
722 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
725 IRB.CreateCall(MS.WarningFn);
726 IRB.CreateCall(MS.EmptyAsm);
727 // FIXME: Insert UnreachableInst if !ClKeepGoing?
728 // This may invalidate some of the following checks and needs to be done
734 unsigned TypeSizeInBits =
735 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
736 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
737 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
738 Value *Fn = MS.MaybeWarningFn[SizeIndex];
739 Value *ConvertedShadow2 =
740 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
741 IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
743 : (Value *)IRB.getInt32(0));
745 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
746 getCleanShadow(ConvertedShadow), "_mscmp");
747 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
749 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
751 IRB.SetInsertPoint(CheckTerm);
752 if (MS.TrackOrigins) {
753 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
756 IRB.CreateCall(MS.WarningFn);
757 IRB.CreateCall(MS.EmptyAsm);
758 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
762 void materializeChecks(bool InstrumentWithCalls) {
763 for (const auto &ShadowData : InstrumentationList) {
764 Instruction *OrigIns = ShadowData.OrigIns;
765 Value *Shadow = ShadowData.Shadow;
766 Value *Origin = ShadowData.Origin;
767 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
769 DEBUG(dbgs() << "DONE:\n" << F);
772 /// \brief Add MemorySanitizer instrumentation to a function.
773 bool runOnFunction() {
774 MS.initializeCallbacks(*F.getParent());
775 if (!MS.DL) return false;
777 // In the presence of unreachable blocks, we may see Phi nodes with
778 // incoming nodes from such blocks. Since InstVisitor skips unreachable
779 // blocks, such nodes will not have any shadow value associated with them.
780 // It's easier to remove unreachable blocks than deal with missing shadow.
781 removeUnreachableBlocks(F);
783 // Iterate all BBs in depth-first order and create shadow instructions
784 // for all instructions (where applicable).
785 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
786 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
790 // Finalize PHI nodes.
791 for (PHINode *PN : ShadowPHINodes) {
792 PHINode *PNS = cast<PHINode>(getShadow(PN));
793 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
794 size_t NumValues = PN->getNumIncomingValues();
795 for (size_t v = 0; v < NumValues; v++) {
796 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
797 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
801 VAHelper->finalizeInstrumentation();
803 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
804 InstrumentationList.size() + StoreList.size() >
805 (unsigned)ClInstrumentationWithCallThreshold;
807 // Delayed instrumentation of StoreInst.
808 // This may add new checks to be inserted later.
809 materializeStores(InstrumentWithCalls);
811 // Insert shadow value checks.
812 materializeChecks(InstrumentWithCalls);
817 /// \brief Compute the shadow type that corresponds to a given Value.
818 Type *getShadowTy(Value *V) {
819 return getShadowTy(V->getType());
822 /// \brief Compute the shadow type that corresponds to a given Type.
823 Type *getShadowTy(Type *OrigTy) {
824 if (!OrigTy->isSized()) {
827 // For integer type, shadow is the same as the original type.
828 // This may return weird-sized types like i1.
829 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
831 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
832 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
833 return VectorType::get(IntegerType::get(*MS.C, EltSize),
834 VT->getNumElements());
836 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
837 return ArrayType::get(getShadowTy(AT->getElementType()),
838 AT->getNumElements());
840 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
841 SmallVector<Type*, 4> Elements;
842 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
843 Elements.push_back(getShadowTy(ST->getElementType(i)));
844 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
845 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
848 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
849 return IntegerType::get(*MS.C, TypeSize);
852 /// \brief Flatten a vector type.
853 Type *getShadowTyNoVec(Type *ty) {
854 if (VectorType *vt = dyn_cast<VectorType>(ty))
855 return IntegerType::get(*MS.C, vt->getBitWidth());
859 /// \brief Convert a shadow value to it's flattened variant.
860 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
861 Type *Ty = V->getType();
862 Type *NoVecTy = getShadowTyNoVec(Ty);
863 if (Ty == NoVecTy) return V;
864 return IRB.CreateBitCast(V, NoVecTy);
867 /// \brief Compute the integer shadow offset that corresponds to a given
868 /// application address.
870 /// Offset = (Addr & ~AndMask) ^ XorMask
871 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
872 uint64_t AndMask = MS.MapParams->AndMask;
873 assert(AndMask != 0 && "AndMask shall be specified");
875 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
876 ConstantInt::get(MS.IntptrTy, ~AndMask));
878 uint64_t XorMask = MS.MapParams->XorMask;
880 OffsetLong = IRB.CreateXor(OffsetLong,
881 ConstantInt::get(MS.IntptrTy, XorMask));
885 /// \brief Compute the shadow address that corresponds to a given application
888 /// Shadow = ShadowBase + Offset
889 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
891 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
892 uint64_t ShadowBase = MS.MapParams->ShadowBase;
895 IRB.CreateAdd(ShadowLong,
896 ConstantInt::get(MS.IntptrTy, ShadowBase));
897 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
900 /// \brief Compute the origin address that corresponds to a given application
903 /// OriginAddr = (OriginBase + Offset) & ~3ULL
904 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
905 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
906 uint64_t OriginBase = MS.MapParams->OriginBase;
909 IRB.CreateAdd(OriginLong,
910 ConstantInt::get(MS.IntptrTy, OriginBase));
911 if (Alignment < kMinOriginAlignment) {
912 uint64_t Mask = kMinOriginAlignment - 1;
913 OriginLong = IRB.CreateAnd(OriginLong,
914 ConstantInt::get(MS.IntptrTy, ~Mask));
916 return IRB.CreateIntToPtr(OriginLong,
917 PointerType::get(IRB.getInt32Ty(), 0));
920 /// \brief Compute the shadow address for a given function argument.
922 /// Shadow = ParamTLS+ArgOffset.
923 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
925 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
926 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
927 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
931 /// \brief Compute the origin address for a given function argument.
932 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
934 if (!MS.TrackOrigins) return nullptr;
935 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
936 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
937 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
941 /// \brief Compute the shadow address for a retval.
942 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
943 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
944 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
948 /// \brief Compute the origin address for a retval.
949 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
950 // We keep a single origin for the entire retval. Might be too optimistic.
951 return MS.RetvalOriginTLS;
954 /// \brief Set SV to be the shadow value for V.
955 void setShadow(Value *V, Value *SV) {
956 assert(!ShadowMap.count(V) && "Values may only have one shadow");
957 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
960 /// \brief Set Origin to be the origin value for V.
961 void setOrigin(Value *V, Value *Origin) {
962 if (!MS.TrackOrigins) return;
963 assert(!OriginMap.count(V) && "Values may only have one origin");
964 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
965 OriginMap[V] = Origin;
968 /// \brief Create a clean shadow value for a given value.
970 /// Clean shadow (all zeroes) means all bits of the value are defined
972 Constant *getCleanShadow(Value *V) {
973 Type *ShadowTy = getShadowTy(V);
976 return Constant::getNullValue(ShadowTy);
979 /// \brief Create a dirty shadow of a given shadow type.
980 Constant *getPoisonedShadow(Type *ShadowTy) {
982 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
983 return Constant::getAllOnesValue(ShadowTy);
984 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
985 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
986 getPoisonedShadow(AT->getElementType()));
987 return ConstantArray::get(AT, Vals);
989 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
990 SmallVector<Constant *, 4> Vals;
991 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
992 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
993 return ConstantStruct::get(ST, Vals);
995 llvm_unreachable("Unexpected shadow type");
998 /// \brief Create a dirty shadow for a given value.
999 Constant *getPoisonedShadow(Value *V) {
1000 Type *ShadowTy = getShadowTy(V);
1003 return getPoisonedShadow(ShadowTy);
1006 /// \brief Create a clean (zero) origin.
1007 Value *getCleanOrigin() {
1008 return Constant::getNullValue(MS.OriginTy);
1011 /// \brief Get the shadow value for a given Value.
1013 /// This function either returns the value set earlier with setShadow,
1014 /// or extracts if from ParamTLS (for function arguments).
1015 Value *getShadow(Value *V) {
1016 if (!PropagateShadow) return getCleanShadow(V);
1017 if (Instruction *I = dyn_cast<Instruction>(V)) {
1018 // For instructions the shadow is already stored in the map.
1019 Value *Shadow = ShadowMap[V];
1021 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1023 assert(Shadow && "No shadow for a value");
1027 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1028 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1029 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1033 if (Argument *A = dyn_cast<Argument>(V)) {
1034 // For arguments we compute the shadow on demand and store it in the map.
1035 Value **ShadowPtr = &ShadowMap[V];
1038 Function *F = A->getParent();
1039 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1040 unsigned ArgOffset = 0;
1041 for (auto &FArg : F->args()) {
1042 if (!FArg.getType()->isSized()) {
1043 DEBUG(dbgs() << "Arg is not sized\n");
1046 unsigned Size = FArg.hasByValAttr()
1047 ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
1048 : MS.DL->getTypeAllocSize(FArg.getType());
1050 bool Overflow = ArgOffset + Size > kParamTLSSize;
1051 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1052 if (FArg.hasByValAttr()) {
1053 // ByVal pointer itself has clean shadow. We copy the actual
1054 // argument shadow to the underlying memory.
1055 // Figure out maximal valid memcpy alignment.
1056 unsigned ArgAlign = FArg.getParamAlignment();
1057 if (ArgAlign == 0) {
1058 Type *EltType = A->getType()->getPointerElementType();
1059 ArgAlign = MS.DL->getABITypeAlignment(EltType);
1062 // ParamTLS overflow.
1063 EntryIRB.CreateMemSet(
1064 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1065 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1067 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1068 Value *Cpy = EntryIRB.CreateMemCpy(
1069 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1071 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1074 *ShadowPtr = getCleanShadow(V);
1077 // ParamTLS overflow.
1078 *ShadowPtr = getCleanShadow(V);
1081 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1084 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1085 **ShadowPtr << "\n");
1086 if (MS.TrackOrigins && !Overflow) {
1088 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1089 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1091 setOrigin(A, getCleanOrigin());
1094 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1096 assert(*ShadowPtr && "Could not find shadow for an argument");
1099 // For everything else the shadow is zero.
1100 return getCleanShadow(V);
1103 /// \brief Get the shadow for i-th argument of the instruction I.
1104 Value *getShadow(Instruction *I, int i) {
1105 return getShadow(I->getOperand(i));
1108 /// \brief Get the origin for a value.
1109 Value *getOrigin(Value *V) {
1110 if (!MS.TrackOrigins) return nullptr;
1111 if (!PropagateShadow) return getCleanOrigin();
1112 if (isa<Constant>(V)) return getCleanOrigin();
1113 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1114 "Unexpected value type in getOrigin()");
1115 Value *Origin = OriginMap[V];
1116 assert(Origin && "Missing origin");
1120 /// \brief Get the origin for i-th argument of the instruction I.
1121 Value *getOrigin(Instruction *I, int i) {
1122 return getOrigin(I->getOperand(i));
1125 /// \brief Remember the place where a shadow check should be inserted.
1127 /// This location will be later instrumented with a check that will print a
1128 /// UMR warning in runtime if the shadow value is not 0.
1129 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1131 if (!InsertChecks) return;
1133 Type *ShadowTy = Shadow->getType();
1134 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1135 "Can only insert checks for integer and vector shadow types");
1137 InstrumentationList.push_back(
1138 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1141 /// \brief Remember the place where a shadow check should be inserted.
1143 /// This location will be later instrumented with a check that will print a
1144 /// UMR warning in runtime if the value is not fully defined.
1145 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1147 Value *Shadow, *Origin;
1148 if (ClCheckConstantShadow) {
1149 Shadow = getShadow(Val);
1150 if (!Shadow) return;
1151 Origin = getOrigin(Val);
1153 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1154 if (!Shadow) return;
1155 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1157 insertShadowCheck(Shadow, Origin, OrigIns);
1160 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1169 case AcquireRelease:
1170 return AcquireRelease;
1171 case SequentiallyConsistent:
1172 return SequentiallyConsistent;
1174 llvm_unreachable("Unknown ordering");
1177 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1186 case AcquireRelease:
1187 return AcquireRelease;
1188 case SequentiallyConsistent:
1189 return SequentiallyConsistent;
1191 llvm_unreachable("Unknown ordering");
1194 // ------------------- Visitors.
1196 /// \brief Instrument LoadInst
1198 /// Loads the corresponding shadow and (optionally) origin.
1199 /// Optionally, checks that the load address is fully defined.
1200 void visitLoadInst(LoadInst &I) {
1201 assert(I.getType()->isSized() && "Load type must have size");
1202 IRBuilder<> IRB(I.getNextNode());
1203 Type *ShadowTy = getShadowTy(&I);
1204 Value *Addr = I.getPointerOperand();
1205 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1206 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1208 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1210 setShadow(&I, getCleanShadow(&I));
1213 if (ClCheckAccessAddress)
1214 insertShadowCheck(I.getPointerOperand(), &I);
1217 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1219 if (MS.TrackOrigins) {
1220 if (PropagateShadow) {
1221 unsigned Alignment = I.getAlignment();
1222 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1223 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1226 setOrigin(&I, getCleanOrigin());
1231 /// \brief Instrument StoreInst
1233 /// Stores the corresponding shadow and (optionally) origin.
1234 /// Optionally, checks that the store address is fully defined.
1235 void visitStoreInst(StoreInst &I) {
1236 StoreList.push_back(&I);
1239 void handleCASOrRMW(Instruction &I) {
1240 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1242 IRBuilder<> IRB(&I);
1243 Value *Addr = I.getOperand(0);
1244 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1246 if (ClCheckAccessAddress)
1247 insertShadowCheck(Addr, &I);
1249 // Only test the conditional argument of cmpxchg instruction.
1250 // The other argument can potentially be uninitialized, but we can not
1251 // detect this situation reliably without possible false positives.
1252 if (isa<AtomicCmpXchgInst>(I))
1253 insertShadowCheck(I.getOperand(1), &I);
1255 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1257 setShadow(&I, getCleanShadow(&I));
1258 setOrigin(&I, getCleanOrigin());
1261 void visitAtomicRMWInst(AtomicRMWInst &I) {
1263 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1266 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1268 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1271 // Vector manipulation.
1272 void visitExtractElementInst(ExtractElementInst &I) {
1273 insertShadowCheck(I.getOperand(1), &I);
1274 IRBuilder<> IRB(&I);
1275 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1277 setOrigin(&I, getOrigin(&I, 0));
1280 void visitInsertElementInst(InsertElementInst &I) {
1281 insertShadowCheck(I.getOperand(2), &I);
1282 IRBuilder<> IRB(&I);
1283 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1284 I.getOperand(2), "_msprop"));
1285 setOriginForNaryOp(I);
1288 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1289 insertShadowCheck(I.getOperand(2), &I);
1290 IRBuilder<> IRB(&I);
1291 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1292 I.getOperand(2), "_msprop"));
1293 setOriginForNaryOp(I);
1297 void visitSExtInst(SExtInst &I) {
1298 IRBuilder<> IRB(&I);
1299 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1300 setOrigin(&I, getOrigin(&I, 0));
1303 void visitZExtInst(ZExtInst &I) {
1304 IRBuilder<> IRB(&I);
1305 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1306 setOrigin(&I, getOrigin(&I, 0));
1309 void visitTruncInst(TruncInst &I) {
1310 IRBuilder<> IRB(&I);
1311 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1312 setOrigin(&I, getOrigin(&I, 0));
1315 void visitBitCastInst(BitCastInst &I) {
1316 IRBuilder<> IRB(&I);
1317 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1318 setOrigin(&I, getOrigin(&I, 0));
1321 void visitPtrToIntInst(PtrToIntInst &I) {
1322 IRBuilder<> IRB(&I);
1323 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1324 "_msprop_ptrtoint"));
1325 setOrigin(&I, getOrigin(&I, 0));
1328 void visitIntToPtrInst(IntToPtrInst &I) {
1329 IRBuilder<> IRB(&I);
1330 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1331 "_msprop_inttoptr"));
1332 setOrigin(&I, getOrigin(&I, 0));
1335 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1336 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1337 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1338 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1339 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1340 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1342 /// \brief Propagate shadow for bitwise AND.
1344 /// This code is exact, i.e. if, for example, a bit in the left argument
1345 /// is defined and 0, then neither the value not definedness of the
1346 /// corresponding bit in B don't affect the resulting shadow.
1347 void visitAnd(BinaryOperator &I) {
1348 IRBuilder<> IRB(&I);
1349 // "And" of 0 and a poisoned value results in unpoisoned value.
1350 // 1&1 => 1; 0&1 => 0; p&1 => p;
1351 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1352 // 1&p => p; 0&p => 0; p&p => p;
1353 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1354 Value *S1 = getShadow(&I, 0);
1355 Value *S2 = getShadow(&I, 1);
1356 Value *V1 = I.getOperand(0);
1357 Value *V2 = I.getOperand(1);
1358 if (V1->getType() != S1->getType()) {
1359 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1360 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1362 Value *S1S2 = IRB.CreateAnd(S1, S2);
1363 Value *V1S2 = IRB.CreateAnd(V1, S2);
1364 Value *S1V2 = IRB.CreateAnd(S1, V2);
1365 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1366 setOriginForNaryOp(I);
1369 void visitOr(BinaryOperator &I) {
1370 IRBuilder<> IRB(&I);
1371 // "Or" of 1 and a poisoned value results in unpoisoned value.
1372 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1373 // 1|0 => 1; 0|0 => 0; p|0 => p;
1374 // 1|p => 1; 0|p => p; p|p => p;
1375 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1376 Value *S1 = getShadow(&I, 0);
1377 Value *S2 = getShadow(&I, 1);
1378 Value *V1 = IRB.CreateNot(I.getOperand(0));
1379 Value *V2 = IRB.CreateNot(I.getOperand(1));
1380 if (V1->getType() != S1->getType()) {
1381 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1382 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1384 Value *S1S2 = IRB.CreateAnd(S1, S2);
1385 Value *V1S2 = IRB.CreateAnd(V1, S2);
1386 Value *S1V2 = IRB.CreateAnd(S1, V2);
1387 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1388 setOriginForNaryOp(I);
1391 /// \brief Default propagation of shadow and/or origin.
1393 /// This class implements the general case of shadow propagation, used in all
1394 /// cases where we don't know and/or don't care about what the operation
1395 /// actually does. It converts all input shadow values to a common type
1396 /// (extending or truncating as necessary), and bitwise OR's them.
1398 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1399 /// fully initialized), and less prone to false positives.
1401 /// This class also implements the general case of origin propagation. For a
1402 /// Nary operation, result origin is set to the origin of an argument that is
1403 /// not entirely initialized. If there is more than one such arguments, the
1404 /// rightmost of them is picked. It does not matter which one is picked if all
1405 /// arguments are initialized.
1406 template <bool CombineShadow>
1411 MemorySanitizerVisitor *MSV;
1414 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1415 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1417 /// \brief Add a pair of shadow and origin values to the mix.
1418 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1419 if (CombineShadow) {
1424 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1425 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1429 if (MSV->MS.TrackOrigins) {
1434 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1435 // No point in adding something that might result in 0 origin value.
1436 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1437 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1439 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1440 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1447 /// \brief Add an application value to the mix.
1448 Combiner &Add(Value *V) {
1449 Value *OpShadow = MSV->getShadow(V);
1450 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1451 return Add(OpShadow, OpOrigin);
1454 /// \brief Set the current combined values as the given instruction's shadow
1456 void Done(Instruction *I) {
1457 if (CombineShadow) {
1459 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1460 MSV->setShadow(I, Shadow);
1462 if (MSV->MS.TrackOrigins) {
1464 MSV->setOrigin(I, Origin);
1469 typedef Combiner<true> ShadowAndOriginCombiner;
1470 typedef Combiner<false> OriginCombiner;
1472 /// \brief Propagate origin for arbitrary operation.
1473 void setOriginForNaryOp(Instruction &I) {
1474 if (!MS.TrackOrigins) return;
1475 IRBuilder<> IRB(&I);
1476 OriginCombiner OC(this, IRB);
1477 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1482 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1483 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1484 "Vector of pointers is not a valid shadow type");
1485 return Ty->isVectorTy() ?
1486 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1487 Ty->getPrimitiveSizeInBits();
1490 /// \brief Cast between two shadow types, extending or truncating as
1492 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1493 bool Signed = false) {
1494 Type *srcTy = V->getType();
1495 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1496 return IRB.CreateIntCast(V, dstTy, Signed);
1497 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1498 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1499 return IRB.CreateIntCast(V, dstTy, Signed);
1500 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1501 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1502 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1504 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1505 return IRB.CreateBitCast(V2, dstTy);
1506 // TODO: handle struct types.
1509 /// \brief Cast an application value to the type of its own shadow.
1510 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1511 Type *ShadowTy = getShadowTy(V);
1512 if (V->getType() == ShadowTy)
1514 if (V->getType()->isPtrOrPtrVectorTy())
1515 return IRB.CreatePtrToInt(V, ShadowTy);
1517 return IRB.CreateBitCast(V, ShadowTy);
1520 /// \brief Propagate shadow for arbitrary operation.
1521 void handleShadowOr(Instruction &I) {
1522 IRBuilder<> IRB(&I);
1523 ShadowAndOriginCombiner SC(this, IRB);
1524 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1529 // \brief Handle multiplication by constant.
1531 // Handle a special case of multiplication by constant that may have one or
1532 // more zeros in the lower bits. This makes corresponding number of lower bits
1533 // of the result zero as well. We model it by shifting the other operand
1534 // shadow left by the required number of bits. Effectively, we transform
1535 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1536 // We use multiplication by 2**N instead of shift to cover the case of
1537 // multiplication by 0, which may occur in some elements of a vector operand.
1538 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1540 Constant *ShadowMul;
1541 Type *Ty = ConstArg->getType();
1542 if (Ty->isVectorTy()) {
1543 unsigned NumElements = Ty->getVectorNumElements();
1544 Type *EltTy = Ty->getSequentialElementType();
1545 SmallVector<Constant *, 16> Elements;
1546 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1548 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1549 APInt V = Elt->getValue();
1550 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1551 Elements.push_back(ConstantInt::get(EltTy, V2));
1553 ShadowMul = ConstantVector::get(Elements);
1555 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1556 APInt V = Elt->getValue();
1557 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1558 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1561 IRBuilder<> IRB(&I);
1563 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1564 setOrigin(&I, getOrigin(OtherArg));
1567 void visitMul(BinaryOperator &I) {
1568 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1569 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1570 if (constOp0 && !constOp1)
1571 handleMulByConstant(I, constOp0, I.getOperand(1));
1572 else if (constOp1 && !constOp0)
1573 handleMulByConstant(I, constOp1, I.getOperand(0));
1578 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1579 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1580 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1581 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1582 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1583 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1585 void handleDiv(Instruction &I) {
1586 IRBuilder<> IRB(&I);
1587 // Strict on the second argument.
1588 insertShadowCheck(I.getOperand(1), &I);
1589 setShadow(&I, getShadow(&I, 0));
1590 setOrigin(&I, getOrigin(&I, 0));
1593 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1594 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1595 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1596 void visitURem(BinaryOperator &I) { handleDiv(I); }
1597 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1598 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1600 /// \brief Instrument == and != comparisons.
1602 /// Sometimes the comparison result is known even if some of the bits of the
1603 /// arguments are not.
1604 void handleEqualityComparison(ICmpInst &I) {
1605 IRBuilder<> IRB(&I);
1606 Value *A = I.getOperand(0);
1607 Value *B = I.getOperand(1);
1608 Value *Sa = getShadow(A);
1609 Value *Sb = getShadow(B);
1611 // Get rid of pointers and vectors of pointers.
1612 // For ints (and vectors of ints), types of A and Sa match,
1613 // and this is a no-op.
1614 A = IRB.CreatePointerCast(A, Sa->getType());
1615 B = IRB.CreatePointerCast(B, Sb->getType());
1617 // A == B <==> (C = A^B) == 0
1618 // A != B <==> (C = A^B) != 0
1620 Value *C = IRB.CreateXor(A, B);
1621 Value *Sc = IRB.CreateOr(Sa, Sb);
1622 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1623 // Result is defined if one of the following is true
1624 // * there is a defined 1 bit in C
1625 // * C is fully defined
1626 // Si = !(C & ~Sc) && Sc
1627 Value *Zero = Constant::getNullValue(Sc->getType());
1628 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1630 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1632 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1633 Si->setName("_msprop_icmp");
1635 setOriginForNaryOp(I);
1638 /// \brief Build the lowest possible value of V, taking into account V's
1639 /// uninitialized bits.
1640 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1643 // Split shadow into sign bit and other bits.
1644 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1645 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1646 // Maximise the undefined shadow bit, minimize other undefined bits.
1648 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1650 // Minimize undefined bits.
1651 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1655 /// \brief Build the highest possible value of V, taking into account V's
1656 /// uninitialized bits.
1657 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1660 // Split shadow into sign bit and other bits.
1661 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1662 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1663 // Minimise the undefined shadow bit, maximise other undefined bits.
1665 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1667 // Maximize undefined bits.
1668 return IRB.CreateOr(A, Sa);
1672 /// \brief Instrument relational comparisons.
1674 /// This function does exact shadow propagation for all relational
1675 /// comparisons of integers, pointers and vectors of those.
1676 /// FIXME: output seems suboptimal when one of the operands is a constant
1677 void handleRelationalComparisonExact(ICmpInst &I) {
1678 IRBuilder<> IRB(&I);
1679 Value *A = I.getOperand(0);
1680 Value *B = I.getOperand(1);
1681 Value *Sa = getShadow(A);
1682 Value *Sb = getShadow(B);
1684 // Get rid of pointers and vectors of pointers.
1685 // For ints (and vectors of ints), types of A and Sa match,
1686 // and this is a no-op.
1687 A = IRB.CreatePointerCast(A, Sa->getType());
1688 B = IRB.CreatePointerCast(B, Sb->getType());
1690 // Let [a0, a1] be the interval of possible values of A, taking into account
1691 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1692 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1693 bool IsSigned = I.isSigned();
1694 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1695 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1696 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1697 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1698 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1699 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1700 Value *Si = IRB.CreateXor(S1, S2);
1702 setOriginForNaryOp(I);
1705 /// \brief Instrument signed relational comparisons.
1707 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1708 /// propagating the highest bit of the shadow. Everything else is delegated
1709 /// to handleShadowOr().
1710 void handleSignedRelationalComparison(ICmpInst &I) {
1711 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1712 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1713 Value* op = nullptr;
1714 CmpInst::Predicate pre = I.getPredicate();
1715 if (constOp0 && constOp0->isNullValue() &&
1716 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1717 op = I.getOperand(1);
1718 } else if (constOp1 && constOp1->isNullValue() &&
1719 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1720 op = I.getOperand(0);
1723 IRBuilder<> IRB(&I);
1725 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1726 setShadow(&I, Shadow);
1727 setOrigin(&I, getOrigin(op));
1733 void visitICmpInst(ICmpInst &I) {
1734 if (!ClHandleICmp) {
1738 if (I.isEquality()) {
1739 handleEqualityComparison(I);
1743 assert(I.isRelational());
1744 if (ClHandleICmpExact) {
1745 handleRelationalComparisonExact(I);
1749 handleSignedRelationalComparison(I);
1753 assert(I.isUnsigned());
1754 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1755 handleRelationalComparisonExact(I);
1762 void visitFCmpInst(FCmpInst &I) {
1766 void handleShift(BinaryOperator &I) {
1767 IRBuilder<> IRB(&I);
1768 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1769 // Otherwise perform the same shift on S1.
1770 Value *S1 = getShadow(&I, 0);
1771 Value *S2 = getShadow(&I, 1);
1772 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1774 Value *V2 = I.getOperand(1);
1775 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1776 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1777 setOriginForNaryOp(I);
1780 void visitShl(BinaryOperator &I) { handleShift(I); }
1781 void visitAShr(BinaryOperator &I) { handleShift(I); }
1782 void visitLShr(BinaryOperator &I) { handleShift(I); }
1784 /// \brief Instrument llvm.memmove
1786 /// At this point we don't know if llvm.memmove will be inlined or not.
1787 /// If we don't instrument it and it gets inlined,
1788 /// our interceptor will not kick in and we will lose the memmove.
1789 /// If we instrument the call here, but it does not get inlined,
1790 /// we will memove the shadow twice: which is bad in case
1791 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1793 /// Similar situation exists for memcpy and memset.
1794 void visitMemMoveInst(MemMoveInst &I) {
1795 IRBuilder<> IRB(&I);
1798 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1799 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1800 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1801 I.eraseFromParent();
1804 // Similar to memmove: avoid copying shadow twice.
1805 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1806 // FIXME: consider doing manual inline for small constant sizes and proper
1808 void visitMemCpyInst(MemCpyInst &I) {
1809 IRBuilder<> IRB(&I);
1812 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1813 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1814 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1815 I.eraseFromParent();
1819 void visitMemSetInst(MemSetInst &I) {
1820 IRBuilder<> IRB(&I);
1823 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1824 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1825 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1826 I.eraseFromParent();
1829 void visitVAStartInst(VAStartInst &I) {
1830 VAHelper->visitVAStartInst(I);
1833 void visitVACopyInst(VACopyInst &I) {
1834 VAHelper->visitVACopyInst(I);
1837 enum IntrinsicKind {
1838 IK_DoesNotAccessMemory,
1843 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1844 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1845 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1846 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1847 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1848 const int UnknownModRefBehavior = IK_WritesMemory;
1849 #define GET_INTRINSIC_MODREF_BEHAVIOR
1850 #define ModRefBehavior IntrinsicKind
1851 #include "llvm/IR/Intrinsics.gen"
1852 #undef ModRefBehavior
1853 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1856 /// \brief Handle vector store-like intrinsics.
1858 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1859 /// has 1 pointer argument and 1 vector argument, returns void.
1860 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1861 IRBuilder<> IRB(&I);
1862 Value* Addr = I.getArgOperand(0);
1863 Value *Shadow = getShadow(&I, 1);
1864 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1866 // We don't know the pointer alignment (could be unaligned SSE store!).
1867 // Have to assume to worst case.
1868 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1870 if (ClCheckAccessAddress)
1871 insertShadowCheck(Addr, &I);
1873 // FIXME: use ClStoreCleanOrigin
1874 // FIXME: factor out common code from materializeStores
1875 if (MS.TrackOrigins)
1876 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1880 /// \brief Handle vector load-like intrinsics.
1882 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1883 /// has 1 pointer argument, returns a vector.
1884 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1885 IRBuilder<> IRB(&I);
1886 Value *Addr = I.getArgOperand(0);
1888 Type *ShadowTy = getShadowTy(&I);
1889 if (PropagateShadow) {
1890 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1891 // We don't know the pointer alignment (could be unaligned SSE load!).
1892 // Have to assume to worst case.
1893 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1895 setShadow(&I, getCleanShadow(&I));
1898 if (ClCheckAccessAddress)
1899 insertShadowCheck(Addr, &I);
1901 if (MS.TrackOrigins) {
1902 if (PropagateShadow)
1903 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1905 setOrigin(&I, getCleanOrigin());
1910 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1912 /// Instrument intrinsics with any number of arguments of the same type,
1913 /// equal to the return type. The type should be simple (no aggregates or
1914 /// pointers; vectors are fine).
1915 /// Caller guarantees that this intrinsic does not access memory.
1916 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1917 Type *RetTy = I.getType();
1918 if (!(RetTy->isIntOrIntVectorTy() ||
1919 RetTy->isFPOrFPVectorTy() ||
1920 RetTy->isX86_MMXTy()))
1923 unsigned NumArgOperands = I.getNumArgOperands();
1925 for (unsigned i = 0; i < NumArgOperands; ++i) {
1926 Type *Ty = I.getArgOperand(i)->getType();
1931 IRBuilder<> IRB(&I);
1932 ShadowAndOriginCombiner SC(this, IRB);
1933 for (unsigned i = 0; i < NumArgOperands; ++i)
1934 SC.Add(I.getArgOperand(i));
1940 /// \brief Heuristically instrument unknown intrinsics.
1942 /// The main purpose of this code is to do something reasonable with all
1943 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1944 /// We recognize several classes of intrinsics by their argument types and
1945 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1946 /// sure that we know what the intrinsic does.
1948 /// We special-case intrinsics where this approach fails. See llvm.bswap
1949 /// handling as an example of that.
1950 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1951 unsigned NumArgOperands = I.getNumArgOperands();
1952 if (NumArgOperands == 0)
1955 Intrinsic::ID iid = I.getIntrinsicID();
1956 IntrinsicKind IK = getIntrinsicKind(iid);
1957 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1958 bool WritesMemory = IK == IK_WritesMemory;
1959 assert(!(OnlyReadsMemory && WritesMemory));
1961 if (NumArgOperands == 2 &&
1962 I.getArgOperand(0)->getType()->isPointerTy() &&
1963 I.getArgOperand(1)->getType()->isVectorTy() &&
1964 I.getType()->isVoidTy() &&
1966 // This looks like a vector store.
1967 return handleVectorStoreIntrinsic(I);
1970 if (NumArgOperands == 1 &&
1971 I.getArgOperand(0)->getType()->isPointerTy() &&
1972 I.getType()->isVectorTy() &&
1974 // This looks like a vector load.
1975 return handleVectorLoadIntrinsic(I);
1978 if (!OnlyReadsMemory && !WritesMemory)
1979 if (maybeHandleSimpleNomemIntrinsic(I))
1982 // FIXME: detect and handle SSE maskstore/maskload
1986 void handleBswap(IntrinsicInst &I) {
1987 IRBuilder<> IRB(&I);
1988 Value *Op = I.getArgOperand(0);
1989 Type *OpType = Op->getType();
1990 Function *BswapFunc = Intrinsic::getDeclaration(
1991 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
1992 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1993 setOrigin(&I, getOrigin(Op));
1996 // \brief Instrument vector convert instrinsic.
1998 // This function instruments intrinsics like cvtsi2ss:
1999 // %Out = int_xxx_cvtyyy(%ConvertOp)
2001 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2002 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2003 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2004 // elements from \p CopyOp.
2005 // In most cases conversion involves floating-point value which may trigger a
2006 // hardware exception when not fully initialized. For this reason we require
2007 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2008 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2009 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2010 // return a fully initialized value.
2011 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2012 IRBuilder<> IRB(&I);
2013 Value *CopyOp, *ConvertOp;
2015 switch (I.getNumArgOperands()) {
2017 CopyOp = I.getArgOperand(0);
2018 ConvertOp = I.getArgOperand(1);
2021 ConvertOp = I.getArgOperand(0);
2025 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2028 // The first *NumUsedElements* elements of ConvertOp are converted to the
2029 // same number of output elements. The rest of the output is copied from
2030 // CopyOp, or (if not available) filled with zeroes.
2031 // Combine shadow for elements of ConvertOp that are used in this operation,
2032 // and insert a check.
2033 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2034 // int->any conversion.
2035 Value *ConvertShadow = getShadow(ConvertOp);
2036 Value *AggShadow = nullptr;
2037 if (ConvertOp->getType()->isVectorTy()) {
2038 AggShadow = IRB.CreateExtractElement(
2039 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2040 for (int i = 1; i < NumUsedElements; ++i) {
2041 Value *MoreShadow = IRB.CreateExtractElement(
2042 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2043 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2046 AggShadow = ConvertShadow;
2048 assert(AggShadow->getType()->isIntegerTy());
2049 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2051 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2054 assert(CopyOp->getType() == I.getType());
2055 assert(CopyOp->getType()->isVectorTy());
2056 Value *ResultShadow = getShadow(CopyOp);
2057 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2058 for (int i = 0; i < NumUsedElements; ++i) {
2059 ResultShadow = IRB.CreateInsertElement(
2060 ResultShadow, ConstantInt::getNullValue(EltTy),
2061 ConstantInt::get(IRB.getInt32Ty(), i));
2063 setShadow(&I, ResultShadow);
2064 setOrigin(&I, getOrigin(CopyOp));
2066 setShadow(&I, getCleanShadow(&I));
2067 setOrigin(&I, getCleanOrigin());
2071 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2072 // zeroes if it is zero, and all ones otherwise.
2073 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2074 if (S->getType()->isVectorTy())
2075 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2076 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2077 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2078 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2081 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2082 Type *T = S->getType();
2083 assert(T->isVectorTy());
2084 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2085 return IRB.CreateSExt(S2, T);
2088 // \brief Instrument vector shift instrinsic.
2090 // This function instruments intrinsics like int_x86_avx2_psll_w.
2091 // Intrinsic shifts %In by %ShiftSize bits.
2092 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2093 // size, and the rest is ignored. Behavior is defined even if shift size is
2094 // greater than register (or field) width.
2095 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2096 assert(I.getNumArgOperands() == 2);
2097 IRBuilder<> IRB(&I);
2098 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2099 // Otherwise perform the same shift on S1.
2100 Value *S1 = getShadow(&I, 0);
2101 Value *S2 = getShadow(&I, 1);
2102 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2103 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2104 Value *V1 = I.getOperand(0);
2105 Value *V2 = I.getOperand(1);
2106 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
2107 IRB.CreateBitCast(S1, V1->getType()), V2);
2108 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2109 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2110 setOriginForNaryOp(I);
2113 // \brief Get an X86_MMX-sized vector type.
2114 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2115 const unsigned X86_MMXSizeInBits = 64;
2116 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2117 X86_MMXSizeInBits / EltSizeInBits);
2120 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2122 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2124 case llvm::Intrinsic::x86_sse2_packsswb_128:
2125 case llvm::Intrinsic::x86_sse2_packuswb_128:
2126 return llvm::Intrinsic::x86_sse2_packsswb_128;
2128 case llvm::Intrinsic::x86_sse2_packssdw_128:
2129 case llvm::Intrinsic::x86_sse41_packusdw:
2130 return llvm::Intrinsic::x86_sse2_packssdw_128;
2132 case llvm::Intrinsic::x86_avx2_packsswb:
2133 case llvm::Intrinsic::x86_avx2_packuswb:
2134 return llvm::Intrinsic::x86_avx2_packsswb;
2136 case llvm::Intrinsic::x86_avx2_packssdw:
2137 case llvm::Intrinsic::x86_avx2_packusdw:
2138 return llvm::Intrinsic::x86_avx2_packssdw;
2140 case llvm::Intrinsic::x86_mmx_packsswb:
2141 case llvm::Intrinsic::x86_mmx_packuswb:
2142 return llvm::Intrinsic::x86_mmx_packsswb;
2144 case llvm::Intrinsic::x86_mmx_packssdw:
2145 return llvm::Intrinsic::x86_mmx_packssdw;
2147 llvm_unreachable("unexpected intrinsic id");
2151 // \brief Instrument vector pack instrinsic.
2153 // This function instruments intrinsics like x86_mmx_packsswb, that
2154 // packs elements of 2 input vectors into half as many bits with saturation.
2155 // Shadow is propagated with the signed variant of the same intrinsic applied
2156 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2157 // EltSizeInBits is used only for x86mmx arguments.
2158 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2159 assert(I.getNumArgOperands() == 2);
2160 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2161 IRBuilder<> IRB(&I);
2162 Value *S1 = getShadow(&I, 0);
2163 Value *S2 = getShadow(&I, 1);
2164 assert(isX86_MMX || S1->getType()->isVectorTy());
2166 // SExt and ICmpNE below must apply to individual elements of input vectors.
2167 // In case of x86mmx arguments, cast them to appropriate vector types and
2169 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2171 S1 = IRB.CreateBitCast(S1, T);
2172 S2 = IRB.CreateBitCast(S2, T);
2174 Value *S1_ext = IRB.CreateSExt(
2175 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2176 Value *S2_ext = IRB.CreateSExt(
2177 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2179 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2180 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2181 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2184 Function *ShadowFn = Intrinsic::getDeclaration(
2185 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2187 Value *S = IRB.CreateCall2(ShadowFn, S1_ext, S2_ext, "_msprop_vector_pack");
2188 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2190 setOriginForNaryOp(I);
2193 // \brief Instrument sum-of-absolute-differencies intrinsic.
2194 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2195 const unsigned SignificantBitsPerResultElement = 16;
2196 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2197 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2198 unsigned ZeroBitsPerResultElement =
2199 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2201 IRBuilder<> IRB(&I);
2202 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2203 S = IRB.CreateBitCast(S, ResTy);
2204 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2206 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2207 S = IRB.CreateBitCast(S, getShadowTy(&I));
2209 setOriginForNaryOp(I);
2212 // \brief Instrument multiply-add intrinsic.
2213 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2214 unsigned EltSizeInBits = 0) {
2215 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2216 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2217 IRBuilder<> IRB(&I);
2218 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2219 S = IRB.CreateBitCast(S, ResTy);
2220 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2222 S = IRB.CreateBitCast(S, getShadowTy(&I));
2224 setOriginForNaryOp(I);
2227 void visitIntrinsicInst(IntrinsicInst &I) {
2228 switch (I.getIntrinsicID()) {
2229 case llvm::Intrinsic::bswap:
2232 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2233 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2234 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2235 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2236 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2237 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2238 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2239 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2240 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2241 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2242 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2243 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2244 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2245 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2246 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2247 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2248 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2249 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2250 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2251 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2252 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2253 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2254 case llvm::Intrinsic::x86_sse_cvtss2si64:
2255 case llvm::Intrinsic::x86_sse_cvtss2si:
2256 case llvm::Intrinsic::x86_sse_cvttss2si64:
2257 case llvm::Intrinsic::x86_sse_cvttss2si:
2258 handleVectorConvertIntrinsic(I, 1);
2260 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2261 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2262 case llvm::Intrinsic::x86_sse_cvtps2pi:
2263 case llvm::Intrinsic::x86_sse_cvttps2pi:
2264 handleVectorConvertIntrinsic(I, 2);
2266 case llvm::Intrinsic::x86_avx2_psll_w:
2267 case llvm::Intrinsic::x86_avx2_psll_d:
2268 case llvm::Intrinsic::x86_avx2_psll_q:
2269 case llvm::Intrinsic::x86_avx2_pslli_w:
2270 case llvm::Intrinsic::x86_avx2_pslli_d:
2271 case llvm::Intrinsic::x86_avx2_pslli_q:
2272 case llvm::Intrinsic::x86_avx2_psrl_w:
2273 case llvm::Intrinsic::x86_avx2_psrl_d:
2274 case llvm::Intrinsic::x86_avx2_psrl_q:
2275 case llvm::Intrinsic::x86_avx2_psra_w:
2276 case llvm::Intrinsic::x86_avx2_psra_d:
2277 case llvm::Intrinsic::x86_avx2_psrli_w:
2278 case llvm::Intrinsic::x86_avx2_psrli_d:
2279 case llvm::Intrinsic::x86_avx2_psrli_q:
2280 case llvm::Intrinsic::x86_avx2_psrai_w:
2281 case llvm::Intrinsic::x86_avx2_psrai_d:
2282 case llvm::Intrinsic::x86_sse2_psll_w:
2283 case llvm::Intrinsic::x86_sse2_psll_d:
2284 case llvm::Intrinsic::x86_sse2_psll_q:
2285 case llvm::Intrinsic::x86_sse2_pslli_w:
2286 case llvm::Intrinsic::x86_sse2_pslli_d:
2287 case llvm::Intrinsic::x86_sse2_pslli_q:
2288 case llvm::Intrinsic::x86_sse2_psrl_w:
2289 case llvm::Intrinsic::x86_sse2_psrl_d:
2290 case llvm::Intrinsic::x86_sse2_psrl_q:
2291 case llvm::Intrinsic::x86_sse2_psra_w:
2292 case llvm::Intrinsic::x86_sse2_psra_d:
2293 case llvm::Intrinsic::x86_sse2_psrli_w:
2294 case llvm::Intrinsic::x86_sse2_psrli_d:
2295 case llvm::Intrinsic::x86_sse2_psrli_q:
2296 case llvm::Intrinsic::x86_sse2_psrai_w:
2297 case llvm::Intrinsic::x86_sse2_psrai_d:
2298 case llvm::Intrinsic::x86_mmx_psll_w:
2299 case llvm::Intrinsic::x86_mmx_psll_d:
2300 case llvm::Intrinsic::x86_mmx_psll_q:
2301 case llvm::Intrinsic::x86_mmx_pslli_w:
2302 case llvm::Intrinsic::x86_mmx_pslli_d:
2303 case llvm::Intrinsic::x86_mmx_pslli_q:
2304 case llvm::Intrinsic::x86_mmx_psrl_w:
2305 case llvm::Intrinsic::x86_mmx_psrl_d:
2306 case llvm::Intrinsic::x86_mmx_psrl_q:
2307 case llvm::Intrinsic::x86_mmx_psra_w:
2308 case llvm::Intrinsic::x86_mmx_psra_d:
2309 case llvm::Intrinsic::x86_mmx_psrli_w:
2310 case llvm::Intrinsic::x86_mmx_psrli_d:
2311 case llvm::Intrinsic::x86_mmx_psrli_q:
2312 case llvm::Intrinsic::x86_mmx_psrai_w:
2313 case llvm::Intrinsic::x86_mmx_psrai_d:
2314 handleVectorShiftIntrinsic(I, /* Variable */ false);
2316 case llvm::Intrinsic::x86_avx2_psllv_d:
2317 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2318 case llvm::Intrinsic::x86_avx2_psllv_q:
2319 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2320 case llvm::Intrinsic::x86_avx2_psrlv_d:
2321 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2322 case llvm::Intrinsic::x86_avx2_psrlv_q:
2323 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2324 case llvm::Intrinsic::x86_avx2_psrav_d:
2325 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2326 handleVectorShiftIntrinsic(I, /* Variable */ true);
2329 case llvm::Intrinsic::x86_sse2_packsswb_128:
2330 case llvm::Intrinsic::x86_sse2_packssdw_128:
2331 case llvm::Intrinsic::x86_sse2_packuswb_128:
2332 case llvm::Intrinsic::x86_sse41_packusdw:
2333 case llvm::Intrinsic::x86_avx2_packsswb:
2334 case llvm::Intrinsic::x86_avx2_packssdw:
2335 case llvm::Intrinsic::x86_avx2_packuswb:
2336 case llvm::Intrinsic::x86_avx2_packusdw:
2337 handleVectorPackIntrinsic(I);
2340 case llvm::Intrinsic::x86_mmx_packsswb:
2341 case llvm::Intrinsic::x86_mmx_packuswb:
2342 handleVectorPackIntrinsic(I, 16);
2345 case llvm::Intrinsic::x86_mmx_packssdw:
2346 handleVectorPackIntrinsic(I, 32);
2349 case llvm::Intrinsic::x86_mmx_psad_bw:
2350 case llvm::Intrinsic::x86_sse2_psad_bw:
2351 case llvm::Intrinsic::x86_avx2_psad_bw:
2352 handleVectorSadIntrinsic(I);
2355 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2356 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2357 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2358 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2359 handleVectorPmaddIntrinsic(I);
2362 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2363 handleVectorPmaddIntrinsic(I, 8);
2366 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2367 handleVectorPmaddIntrinsic(I, 16);
2371 if (!handleUnknownIntrinsic(I))
2372 visitInstruction(I);
2377 void visitCallSite(CallSite CS) {
2378 Instruction &I = *CS.getInstruction();
2379 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2381 CallInst *Call = cast<CallInst>(&I);
2383 // For inline asm, do the usual thing: check argument shadow and mark all
2384 // outputs as clean. Note that any side effects of the inline asm that are
2385 // not immediately visible in its constraints are not handled.
2386 if (Call->isInlineAsm()) {
2387 visitInstruction(I);
2391 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2393 // We are going to insert code that relies on the fact that the callee
2394 // will become a non-readonly function after it is instrumented by us. To
2395 // prevent this code from being optimized out, mark that function
2396 // non-readonly in advance.
2397 if (Function *Func = Call->getCalledFunction()) {
2398 // Clear out readonly/readnone attributes.
2400 B.addAttribute(Attribute::ReadOnly)
2401 .addAttribute(Attribute::ReadNone);
2402 Func->removeAttributes(AttributeSet::FunctionIndex,
2403 AttributeSet::get(Func->getContext(),
2404 AttributeSet::FunctionIndex,
2408 IRBuilder<> IRB(&I);
2410 unsigned ArgOffset = 0;
2411 DEBUG(dbgs() << " CallSite: " << I << "\n");
2412 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2413 ArgIt != End; ++ArgIt) {
2415 unsigned i = ArgIt - CS.arg_begin();
2416 if (!A->getType()->isSized()) {
2417 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2421 Value *Store = nullptr;
2422 // Compute the Shadow for arg even if it is ByVal, because
2423 // in that case getShadow() will copy the actual arg shadow to
2424 // __msan_param_tls.
2425 Value *ArgShadow = getShadow(A);
2426 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2427 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2428 " Shadow: " << *ArgShadow << "\n");
2429 bool ArgIsInitialized = false;
2430 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2431 assert(A->getType()->isPointerTy() &&
2432 "ByVal argument is not a pointer!");
2433 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2434 if (ArgOffset + Size > kParamTLSSize) break;
2435 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2436 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2437 Store = IRB.CreateMemCpy(ArgShadowBase,
2438 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2441 Size = MS.DL->getTypeAllocSize(A->getType());
2442 if (ArgOffset + Size > kParamTLSSize) break;
2443 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2444 kShadowTLSAlignment);
2445 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2446 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2448 if (MS.TrackOrigins && !ArgIsInitialized)
2449 IRB.CreateStore(getOrigin(A),
2450 getOriginPtrForArgument(A, IRB, ArgOffset));
2452 assert(Size != 0 && Store != nullptr);
2453 DEBUG(dbgs() << " Param:" << *Store << "\n");
2454 ArgOffset += RoundUpToAlignment(Size, 8);
2456 DEBUG(dbgs() << " done with call args\n");
2459 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2460 if (FT->isVarArg()) {
2461 VAHelper->visitCallSite(CS, IRB);
2464 // Now, get the shadow for the RetVal.
2465 if (!I.getType()->isSized()) return;
2466 IRBuilder<> IRBBefore(&I);
2467 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2468 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2469 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2470 Instruction *NextInsn = nullptr;
2472 NextInsn = I.getNextNode();
2474 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2475 if (!NormalDest->getSinglePredecessor()) {
2476 // FIXME: this case is tricky, so we are just conservative here.
2477 // Perhaps we need to split the edge between this BB and NormalDest,
2478 // but a naive attempt to use SplitEdge leads to a crash.
2479 setShadow(&I, getCleanShadow(&I));
2480 setOrigin(&I, getCleanOrigin());
2483 NextInsn = NormalDest->getFirstInsertionPt();
2485 "Could not find insertion point for retval shadow load");
2487 IRBuilder<> IRBAfter(NextInsn);
2488 Value *RetvalShadow =
2489 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2490 kShadowTLSAlignment, "_msret");
2491 setShadow(&I, RetvalShadow);
2492 if (MS.TrackOrigins)
2493 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2496 void visitReturnInst(ReturnInst &I) {
2497 IRBuilder<> IRB(&I);
2498 Value *RetVal = I.getReturnValue();
2499 if (!RetVal) return;
2500 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2501 if (CheckReturnValue) {
2502 insertShadowCheck(RetVal, &I);
2503 Value *Shadow = getCleanShadow(RetVal);
2504 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2506 Value *Shadow = getShadow(RetVal);
2507 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2508 // FIXME: make it conditional if ClStoreCleanOrigin==0
2509 if (MS.TrackOrigins)
2510 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2514 void visitPHINode(PHINode &I) {
2515 IRBuilder<> IRB(&I);
2516 if (!PropagateShadow) {
2517 setShadow(&I, getCleanShadow(&I));
2518 setOrigin(&I, getCleanOrigin());
2522 ShadowPHINodes.push_back(&I);
2523 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2525 if (MS.TrackOrigins)
2526 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2530 void visitAllocaInst(AllocaInst &I) {
2531 setShadow(&I, getCleanShadow(&I));
2532 setOrigin(&I, getCleanOrigin());
2533 IRBuilder<> IRB(I.getNextNode());
2534 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2535 if (PoisonStack && ClPoisonStackWithCall) {
2536 IRB.CreateCall2(MS.MsanPoisonStackFn,
2537 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2538 ConstantInt::get(MS.IntptrTy, Size));
2540 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2541 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2542 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2545 if (PoisonStack && MS.TrackOrigins) {
2546 SmallString<2048> StackDescriptionStorage;
2547 raw_svector_ostream StackDescription(StackDescriptionStorage);
2548 // We create a string with a description of the stack allocation and
2549 // pass it into __msan_set_alloca_origin.
2550 // It will be printed by the run-time if stack-originated UMR is found.
2551 // The first 4 bytes of the string are set to '----' and will be replaced
2552 // by __msan_va_arg_overflow_size_tls at the first call.
2553 StackDescription << "----" << I.getName() << "@" << F.getName();
2555 createPrivateNonConstGlobalForString(*F.getParent(),
2556 StackDescription.str());
2558 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2559 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2560 ConstantInt::get(MS.IntptrTy, Size),
2561 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2562 IRB.CreatePointerCast(&F, MS.IntptrTy));
2566 void visitSelectInst(SelectInst& I) {
2567 IRBuilder<> IRB(&I);
2568 // a = select b, c, d
2569 Value *B = I.getCondition();
2570 Value *C = I.getTrueValue();
2571 Value *D = I.getFalseValue();
2572 Value *Sb = getShadow(B);
2573 Value *Sc = getShadow(C);
2574 Value *Sd = getShadow(D);
2576 // Result shadow if condition shadow is 0.
2577 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2579 if (I.getType()->isAggregateType()) {
2580 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2581 // an extra "select". This results in much more compact IR.
2582 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2583 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2585 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2586 // If Sb (condition is poisoned), look for bits in c and d that are equal
2587 // and both unpoisoned.
2588 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2590 // Cast arguments to shadow-compatible type.
2591 C = CreateAppToShadowCast(IRB, C);
2592 D = CreateAppToShadowCast(IRB, D);
2594 // Result shadow if condition shadow is 1.
2595 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2597 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2599 if (MS.TrackOrigins) {
2600 // Origins are always i32, so any vector conditions must be flattened.
2601 // FIXME: consider tracking vector origins for app vectors?
2602 if (B->getType()->isVectorTy()) {
2603 Type *FlatTy = getShadowTyNoVec(B->getType());
2604 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2605 ConstantInt::getNullValue(FlatTy));
2606 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2607 ConstantInt::getNullValue(FlatTy));
2609 // a = select b, c, d
2610 // Oa = Sb ? Ob : (b ? Oc : Od)
2612 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2613 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2614 getOrigin(I.getFalseValue()))));
2618 void visitLandingPadInst(LandingPadInst &I) {
2620 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2621 setShadow(&I, getCleanShadow(&I));
2622 setOrigin(&I, getCleanOrigin());
2625 void visitGetElementPtrInst(GetElementPtrInst &I) {
2629 void visitExtractValueInst(ExtractValueInst &I) {
2630 IRBuilder<> IRB(&I);
2631 Value *Agg = I.getAggregateOperand();
2632 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2633 Value *AggShadow = getShadow(Agg);
2634 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2635 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2636 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2637 setShadow(&I, ResShadow);
2638 setOriginForNaryOp(I);
2641 void visitInsertValueInst(InsertValueInst &I) {
2642 IRBuilder<> IRB(&I);
2643 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2644 Value *AggShadow = getShadow(I.getAggregateOperand());
2645 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2646 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2647 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2648 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2649 DEBUG(dbgs() << " Res: " << *Res << "\n");
2651 setOriginForNaryOp(I);
2654 void dumpInst(Instruction &I) {
2655 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2656 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2658 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2660 errs() << "QQQ " << I << "\n";
2663 void visitResumeInst(ResumeInst &I) {
2664 DEBUG(dbgs() << "Resume: " << I << "\n");
2665 // Nothing to do here.
2668 void visitInstruction(Instruction &I) {
2669 // Everything else: stop propagating and check for poisoned shadow.
2670 if (ClDumpStrictInstructions)
2672 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2673 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2674 insertShadowCheck(I.getOperand(i), &I);
2675 setShadow(&I, getCleanShadow(&I));
2676 setOrigin(&I, getCleanOrigin());
2680 /// \brief AMD64-specific implementation of VarArgHelper.
2681 struct VarArgAMD64Helper : public VarArgHelper {
2682 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2683 // See a comment in visitCallSite for more details.
2684 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2685 static const unsigned AMD64FpEndOffset = 176;
2688 MemorySanitizer &MS;
2689 MemorySanitizerVisitor &MSV;
2690 Value *VAArgTLSCopy;
2691 Value *VAArgOverflowSize;
2693 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2695 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2696 MemorySanitizerVisitor &MSV)
2697 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2698 VAArgOverflowSize(nullptr) {}
2700 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2702 ArgKind classifyArgument(Value* arg) {
2703 // A very rough approximation of X86_64 argument classification rules.
2704 Type *T = arg->getType();
2705 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2706 return AK_FloatingPoint;
2707 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2708 return AK_GeneralPurpose;
2709 if (T->isPointerTy())
2710 return AK_GeneralPurpose;
2714 // For VarArg functions, store the argument shadow in an ABI-specific format
2715 // that corresponds to va_list layout.
2716 // We do this because Clang lowers va_arg in the frontend, and this pass
2717 // only sees the low level code that deals with va_list internals.
2718 // A much easier alternative (provided that Clang emits va_arg instructions)
2719 // would have been to associate each live instance of va_list with a copy of
2720 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2722 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2723 unsigned GpOffset = 0;
2724 unsigned FpOffset = AMD64GpEndOffset;
2725 unsigned OverflowOffset = AMD64FpEndOffset;
2726 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2727 ArgIt != End; ++ArgIt) {
2729 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2730 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2732 // ByVal arguments always go to the overflow area.
2733 assert(A->getType()->isPointerTy());
2734 Type *RealTy = A->getType()->getPointerElementType();
2735 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2736 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2737 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2738 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2739 ArgSize, kShadowTLSAlignment);
2741 ArgKind AK = classifyArgument(A);
2742 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2744 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2748 case AK_GeneralPurpose:
2749 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2752 case AK_FloatingPoint:
2753 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2757 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2758 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2759 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2761 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2764 Constant *OverflowSize =
2765 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2766 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2769 /// \brief Compute the shadow address for a given va_arg.
2770 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2772 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2773 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2774 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2778 void visitVAStartInst(VAStartInst &I) override {
2779 IRBuilder<> IRB(&I);
2780 VAStartInstrumentationList.push_back(&I);
2781 Value *VAListTag = I.getArgOperand(0);
2782 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2784 // Unpoison the whole __va_list_tag.
2785 // FIXME: magic ABI constants.
2786 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2787 /* size */24, /* alignment */8, false);
2790 void visitVACopyInst(VACopyInst &I) override {
2791 IRBuilder<> IRB(&I);
2792 Value *VAListTag = I.getArgOperand(0);
2793 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2795 // Unpoison the whole __va_list_tag.
2796 // FIXME: magic ABI constants.
2797 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2798 /* size */24, /* alignment */8, false);
2801 void finalizeInstrumentation() override {
2802 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2803 "finalizeInstrumentation called twice");
2804 if (!VAStartInstrumentationList.empty()) {
2805 // If there is a va_start in this function, make a backup copy of
2806 // va_arg_tls somewhere in the function entry block.
2807 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2808 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2810 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2812 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2813 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2816 // Instrument va_start.
2817 // Copy va_list shadow from the backup copy of the TLS contents.
2818 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2819 CallInst *OrigInst = VAStartInstrumentationList[i];
2820 IRBuilder<> IRB(OrigInst->getNextNode());
2821 Value *VAListTag = OrigInst->getArgOperand(0);
2823 Value *RegSaveAreaPtrPtr =
2825 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2826 ConstantInt::get(MS.IntptrTy, 16)),
2827 Type::getInt64PtrTy(*MS.C));
2828 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2829 Value *RegSaveAreaShadowPtr =
2830 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2831 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2832 AMD64FpEndOffset, 16);
2834 Value *OverflowArgAreaPtrPtr =
2836 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2837 ConstantInt::get(MS.IntptrTy, 8)),
2838 Type::getInt64PtrTy(*MS.C));
2839 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2840 Value *OverflowArgAreaShadowPtr =
2841 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2842 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2843 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2848 /// \brief MIPS64-specific implementation of VarArgHelper.
2849 struct VarArgMIPS64Helper : public VarArgHelper {
2851 MemorySanitizer &MS;
2852 MemorySanitizerVisitor &MSV;
2853 Value *VAArgTLSCopy;
2856 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2858 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
2859 MemorySanitizerVisitor &MSV)
2860 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2861 VAArgSize(nullptr) {}
2863 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2864 unsigned VAArgOffset = 0;
2865 for (CallSite::arg_iterator ArgIt = CS.arg_begin() + 1, End = CS.arg_end();
2866 ArgIt != End; ++ArgIt) {
2869 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2870 #if defined(__MIPSEB__) || defined(MIPSEB)
2871 // Adjusting the shadow for argument with size < 8 to match the placement
2872 // of bits in big endian system
2874 VAArgOffset += (8 - ArgSize);
2876 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
2877 VAArgOffset += ArgSize;
2878 VAArgOffset = RoundUpToAlignment(VAArgOffset, 8);
2879 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2882 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
2883 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
2884 // a new class member i.e. it is the total size of all VarArgs.
2885 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
2888 /// \brief Compute the shadow address for a given va_arg.
2889 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2891 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2892 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2893 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2897 void visitVAStartInst(VAStartInst &I) override {
2898 IRBuilder<> IRB(&I);
2899 VAStartInstrumentationList.push_back(&I);
2900 Value *VAListTag = I.getArgOperand(0);
2901 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2902 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2903 /* size */8, /* alignment */8, false);
2906 void visitVACopyInst(VACopyInst &I) override {
2907 IRBuilder<> IRB(&I);
2908 Value *VAListTag = I.getArgOperand(0);
2909 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2910 // Unpoison the whole __va_list_tag.
2911 // FIXME: magic ABI constants.
2912 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2913 /* size */8, /* alignment */8, false);
2916 void finalizeInstrumentation() override {
2917 assert(!VAArgSize && !VAArgTLSCopy &&
2918 "finalizeInstrumentation called twice");
2919 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2920 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2921 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
2924 if (!VAStartInstrumentationList.empty()) {
2925 // If there is a va_start in this function, make a backup copy of
2926 // va_arg_tls somewhere in the function entry block.
2927 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2928 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2931 // Instrument va_start.
2932 // Copy va_list shadow from the backup copy of the TLS contents.
2933 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2934 CallInst *OrigInst = VAStartInstrumentationList[i];
2935 IRBuilder<> IRB(OrigInst->getNextNode());
2936 Value *VAListTag = OrigInst->getArgOperand(0);
2937 Value *RegSaveAreaPtrPtr =
2938 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2939 Type::getInt64PtrTy(*MS.C));
2940 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2941 Value *RegSaveAreaShadowPtr =
2942 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2943 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
2948 /// \brief A no-op implementation of VarArgHelper.
2949 struct VarArgNoOpHelper : public VarArgHelper {
2950 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2951 MemorySanitizerVisitor &MSV) {}
2953 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2955 void visitVAStartInst(VAStartInst &I) override {}
2957 void visitVACopyInst(VACopyInst &I) override {}
2959 void finalizeInstrumentation() override {}
2962 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2963 MemorySanitizerVisitor &Visitor) {
2964 // VarArg handling is only implemented on AMD64. False positives are possible
2965 // on other platforms.
2966 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2967 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2968 return new VarArgAMD64Helper(Func, Msan, Visitor);
2969 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
2970 TargetTriple.getArch() == llvm::Triple::mips64el)
2971 return new VarArgMIPS64Helper(Func, Msan, Visitor);
2973 return new VarArgNoOpHelper(Func, Msan, Visitor);
2978 bool MemorySanitizer::runOnFunction(Function &F) {
2979 MemorySanitizerVisitor Visitor(F, *this);
2981 // Clear out readonly/readnone attributes.
2983 B.addAttribute(Attribute::ReadOnly)
2984 .addAttribute(Attribute::ReadNone);
2985 F.removeAttributes(AttributeSet::FunctionIndex,
2986 AttributeSet::get(F.getContext(),
2987 AttributeSet::FunctionIndex, B));
2989 return Visitor.runOnFunction();