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 DL = &M.getDataLayout();
454 Triple TargetTriple(M.getTargetTriple());
455 switch (TargetTriple.getOS()) {
456 case Triple::FreeBSD:
457 switch (TargetTriple.getArch()) {
459 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
462 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
465 report_fatal_error("unsupported architecture");
469 switch (TargetTriple.getArch()) {
471 MapParams = Linux_X86_MemoryMapParams.bits64;
474 MapParams = Linux_X86_MemoryMapParams.bits32;
477 case Triple::mips64el:
478 MapParams = Linux_MIPS_MemoryMapParams.bits64;
481 report_fatal_error("unsupported architecture");
485 report_fatal_error("unsupported operating system");
488 C = &(M.getContext());
490 IntptrTy = IRB.getIntPtrTy(DL);
491 OriginTy = IRB.getInt32Ty();
493 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
494 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
496 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
497 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
498 "__msan_init", IRB.getVoidTy(), nullptr)), 0);
501 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
502 IRB.getInt32(TrackOrigins), "__msan_track_origins");
505 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
506 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
513 /// \brief A helper class that handles instrumentation of VarArg
514 /// functions on a particular platform.
516 /// Implementations are expected to insert the instrumentation
517 /// necessary to propagate argument shadow through VarArg function
518 /// calls. Visit* methods are called during an InstVisitor pass over
519 /// the function, and should avoid creating new basic blocks. A new
520 /// instance of this class is created for each instrumented function.
521 struct VarArgHelper {
522 /// \brief Visit a CallSite.
523 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
525 /// \brief Visit a va_start call.
526 virtual void visitVAStartInst(VAStartInst &I) = 0;
528 /// \brief Visit a va_copy call.
529 virtual void visitVACopyInst(VACopyInst &I) = 0;
531 /// \brief Finalize function instrumentation.
533 /// This method is called after visiting all interesting (see above)
534 /// instructions in a function.
535 virtual void finalizeInstrumentation() = 0;
537 virtual ~VarArgHelper() {}
540 struct MemorySanitizerVisitor;
543 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
544 MemorySanitizerVisitor &Visitor);
546 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
547 if (TypeSize <= 8) return 0;
548 return Log2_32_Ceil(TypeSize / 8);
551 /// This class does all the work for a given function. Store and Load
552 /// instructions store and load corresponding shadow and origin
553 /// values. Most instructions propagate shadow from arguments to their
554 /// return values. Certain instructions (most importantly, BranchInst)
555 /// test their argument shadow and print reports (with a runtime call) if it's
557 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
560 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
561 ValueMap<Value*, Value*> ShadowMap, OriginMap;
562 std::unique_ptr<VarArgHelper> VAHelper;
564 // The following flags disable parts of MSan instrumentation based on
565 // blacklist contents and command-line options.
567 bool PropagateShadow;
570 bool CheckReturnValue;
572 struct ShadowOriginAndInsertPoint {
575 Instruction *OrigIns;
576 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
577 : Shadow(S), Origin(O), OrigIns(I) { }
579 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
580 SmallVector<Instruction*, 16> StoreList;
582 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
583 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
584 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
585 InsertChecks = SanitizeFunction;
586 PropagateShadow = SanitizeFunction;
587 PoisonStack = SanitizeFunction && ClPoisonStack;
588 PoisonUndef = SanitizeFunction && ClPoisonUndef;
589 // FIXME: Consider using SpecialCaseList to specify a list of functions that
590 // must always return fully initialized values. For now, we hardcode "main".
591 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
593 DEBUG(if (!InsertChecks)
594 dbgs() << "MemorySanitizer is not inserting checks into '"
595 << F.getName() << "'\n");
598 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
599 if (MS.TrackOrigins <= 1) return V;
600 return IRB.CreateCall(MS.MsanChainOriginFn, V);
603 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
604 unsigned IntptrSize = MS.DL->getTypeStoreSize(MS.IntptrTy);
605 if (IntptrSize == kOriginSize) return Origin;
606 assert(IntptrSize == kOriginSize * 2);
607 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
608 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
611 /// \brief Fill memory range with the given origin value.
612 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
613 unsigned Size, unsigned Alignment) {
614 unsigned IntptrAlignment = MS.DL->getABITypeAlignment(MS.IntptrTy);
615 unsigned IntptrSize = MS.DL->getTypeStoreSize(MS.IntptrTy);
616 assert(IntptrAlignment >= kMinOriginAlignment);
617 assert(IntptrSize >= kOriginSize);
620 unsigned CurrentAlignment = Alignment;
621 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
622 Value *IntptrOrigin = originToIntptr(IRB, Origin);
623 Value *IntptrOriginPtr =
624 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
625 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
627 i ? IRB.CreateConstGEP1_32(IntptrOriginPtr, i) : IntptrOriginPtr;
628 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
629 Ofs += IntptrSize / kOriginSize;
630 CurrentAlignment = IntptrAlignment;
634 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
635 Value *GEP = i ? IRB.CreateConstGEP1_32(OriginPtr, i) : OriginPtr;
636 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
637 CurrentAlignment = kMinOriginAlignment;
641 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
642 unsigned Alignment, bool AsCall) {
643 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
644 unsigned StoreSize = MS.DL->getTypeStoreSize(Shadow->getType());
645 if (isa<StructType>(Shadow->getType())) {
646 paintOrigin(IRB, updateOrigin(Origin, IRB),
647 getOriginPtr(Addr, IRB, Alignment), StoreSize,
650 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
651 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
652 if (ConstantShadow) {
653 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
654 paintOrigin(IRB, updateOrigin(Origin, IRB),
655 getOriginPtr(Addr, IRB, Alignment), StoreSize,
660 unsigned TypeSizeInBits =
661 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
662 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
663 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
664 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
665 Value *ConvertedShadow2 = IRB.CreateZExt(
666 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
667 IRB.CreateCall3(Fn, ConvertedShadow2,
668 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
671 Value *Cmp = IRB.CreateICmpNE(
672 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
673 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
674 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
675 IRBuilder<> IRBNew(CheckTerm);
676 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
677 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
683 void materializeStores(bool InstrumentWithCalls) {
684 for (auto Inst : StoreList) {
685 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
687 IRBuilder<> IRB(&SI);
688 Value *Val = SI.getValueOperand();
689 Value *Addr = SI.getPointerOperand();
690 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
691 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
694 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
695 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
698 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
700 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
702 if (MS.TrackOrigins && !SI.isAtomic())
703 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
704 InstrumentWithCalls);
708 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
710 IRBuilder<> IRB(OrigIns);
711 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
712 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
713 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
715 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
716 if (ConstantShadow) {
717 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
718 if (MS.TrackOrigins) {
719 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
722 IRB.CreateCall(MS.WarningFn);
723 IRB.CreateCall(MS.EmptyAsm);
724 // FIXME: Insert UnreachableInst if !ClKeepGoing?
725 // This may invalidate some of the following checks and needs to be done
731 unsigned TypeSizeInBits =
732 MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
733 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
734 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
735 Value *Fn = MS.MaybeWarningFn[SizeIndex];
736 Value *ConvertedShadow2 =
737 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
738 IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
740 : (Value *)IRB.getInt32(0));
742 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
743 getCleanShadow(ConvertedShadow), "_mscmp");
744 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
746 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
748 IRB.SetInsertPoint(CheckTerm);
749 if (MS.TrackOrigins) {
750 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
753 IRB.CreateCall(MS.WarningFn);
754 IRB.CreateCall(MS.EmptyAsm);
755 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
759 void materializeChecks(bool InstrumentWithCalls) {
760 for (const auto &ShadowData : InstrumentationList) {
761 Instruction *OrigIns = ShadowData.OrigIns;
762 Value *Shadow = ShadowData.Shadow;
763 Value *Origin = ShadowData.Origin;
764 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
766 DEBUG(dbgs() << "DONE:\n" << F);
769 /// \brief Add MemorySanitizer instrumentation to a function.
770 bool runOnFunction() {
771 MS.initializeCallbacks(*F.getParent());
772 if (!MS.DL) return false;
774 // In the presence of unreachable blocks, we may see Phi nodes with
775 // incoming nodes from such blocks. Since InstVisitor skips unreachable
776 // blocks, such nodes will not have any shadow value associated with them.
777 // It's easier to remove unreachable blocks than deal with missing shadow.
778 removeUnreachableBlocks(F);
780 // Iterate all BBs in depth-first order and create shadow instructions
781 // for all instructions (where applicable).
782 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
783 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
787 // Finalize PHI nodes.
788 for (PHINode *PN : ShadowPHINodes) {
789 PHINode *PNS = cast<PHINode>(getShadow(PN));
790 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
791 size_t NumValues = PN->getNumIncomingValues();
792 for (size_t v = 0; v < NumValues; v++) {
793 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
794 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
798 VAHelper->finalizeInstrumentation();
800 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
801 InstrumentationList.size() + StoreList.size() >
802 (unsigned)ClInstrumentationWithCallThreshold;
804 // Delayed instrumentation of StoreInst.
805 // This may add new checks to be inserted later.
806 materializeStores(InstrumentWithCalls);
808 // Insert shadow value checks.
809 materializeChecks(InstrumentWithCalls);
814 /// \brief Compute the shadow type that corresponds to a given Value.
815 Type *getShadowTy(Value *V) {
816 return getShadowTy(V->getType());
819 /// \brief Compute the shadow type that corresponds to a given Type.
820 Type *getShadowTy(Type *OrigTy) {
821 if (!OrigTy->isSized()) {
824 // For integer type, shadow is the same as the original type.
825 // This may return weird-sized types like i1.
826 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
828 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
829 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
830 return VectorType::get(IntegerType::get(*MS.C, EltSize),
831 VT->getNumElements());
833 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
834 return ArrayType::get(getShadowTy(AT->getElementType()),
835 AT->getNumElements());
837 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
838 SmallVector<Type*, 4> Elements;
839 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
840 Elements.push_back(getShadowTy(ST->getElementType(i)));
841 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
842 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
845 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
846 return IntegerType::get(*MS.C, TypeSize);
849 /// \brief Flatten a vector type.
850 Type *getShadowTyNoVec(Type *ty) {
851 if (VectorType *vt = dyn_cast<VectorType>(ty))
852 return IntegerType::get(*MS.C, vt->getBitWidth());
856 /// \brief Convert a shadow value to it's flattened variant.
857 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
858 Type *Ty = V->getType();
859 Type *NoVecTy = getShadowTyNoVec(Ty);
860 if (Ty == NoVecTy) return V;
861 return IRB.CreateBitCast(V, NoVecTy);
864 /// \brief Compute the integer shadow offset that corresponds to a given
865 /// application address.
867 /// Offset = (Addr & ~AndMask) ^ XorMask
868 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
869 uint64_t AndMask = MS.MapParams->AndMask;
870 assert(AndMask != 0 && "AndMask shall be specified");
872 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
873 ConstantInt::get(MS.IntptrTy, ~AndMask));
875 uint64_t XorMask = MS.MapParams->XorMask;
877 OffsetLong = IRB.CreateXor(OffsetLong,
878 ConstantInt::get(MS.IntptrTy, XorMask));
882 /// \brief Compute the shadow address that corresponds to a given application
885 /// Shadow = ShadowBase + Offset
886 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
888 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
889 uint64_t ShadowBase = MS.MapParams->ShadowBase;
892 IRB.CreateAdd(ShadowLong,
893 ConstantInt::get(MS.IntptrTy, ShadowBase));
894 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
897 /// \brief Compute the origin address that corresponds to a given application
900 /// OriginAddr = (OriginBase + Offset) & ~3ULL
901 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
902 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
903 uint64_t OriginBase = MS.MapParams->OriginBase;
906 IRB.CreateAdd(OriginLong,
907 ConstantInt::get(MS.IntptrTy, OriginBase));
908 if (Alignment < kMinOriginAlignment) {
909 uint64_t Mask = kMinOriginAlignment - 1;
910 OriginLong = IRB.CreateAnd(OriginLong,
911 ConstantInt::get(MS.IntptrTy, ~Mask));
913 return IRB.CreateIntToPtr(OriginLong,
914 PointerType::get(IRB.getInt32Ty(), 0));
917 /// \brief Compute the shadow address for a given function argument.
919 /// Shadow = ParamTLS+ArgOffset.
920 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
922 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
923 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
924 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
928 /// \brief Compute the origin address for a given function argument.
929 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
931 if (!MS.TrackOrigins) return nullptr;
932 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
933 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
934 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
938 /// \brief Compute the shadow address for a retval.
939 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
940 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
941 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
945 /// \brief Compute the origin address for a retval.
946 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
947 // We keep a single origin for the entire retval. Might be too optimistic.
948 return MS.RetvalOriginTLS;
951 /// \brief Set SV to be the shadow value for V.
952 void setShadow(Value *V, Value *SV) {
953 assert(!ShadowMap.count(V) && "Values may only have one shadow");
954 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
957 /// \brief Set Origin to be the origin value for V.
958 void setOrigin(Value *V, Value *Origin) {
959 if (!MS.TrackOrigins) return;
960 assert(!OriginMap.count(V) && "Values may only have one origin");
961 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
962 OriginMap[V] = Origin;
965 /// \brief Create a clean shadow value for a given value.
967 /// Clean shadow (all zeroes) means all bits of the value are defined
969 Constant *getCleanShadow(Value *V) {
970 Type *ShadowTy = getShadowTy(V);
973 return Constant::getNullValue(ShadowTy);
976 /// \brief Create a dirty shadow of a given shadow type.
977 Constant *getPoisonedShadow(Type *ShadowTy) {
979 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
980 return Constant::getAllOnesValue(ShadowTy);
981 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
982 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
983 getPoisonedShadow(AT->getElementType()));
984 return ConstantArray::get(AT, Vals);
986 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
987 SmallVector<Constant *, 4> Vals;
988 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
989 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
990 return ConstantStruct::get(ST, Vals);
992 llvm_unreachable("Unexpected shadow type");
995 /// \brief Create a dirty shadow for a given value.
996 Constant *getPoisonedShadow(Value *V) {
997 Type *ShadowTy = getShadowTy(V);
1000 return getPoisonedShadow(ShadowTy);
1003 /// \brief Create a clean (zero) origin.
1004 Value *getCleanOrigin() {
1005 return Constant::getNullValue(MS.OriginTy);
1008 /// \brief Get the shadow value for a given Value.
1010 /// This function either returns the value set earlier with setShadow,
1011 /// or extracts if from ParamTLS (for function arguments).
1012 Value *getShadow(Value *V) {
1013 if (!PropagateShadow) return getCleanShadow(V);
1014 if (Instruction *I = dyn_cast<Instruction>(V)) {
1015 // For instructions the shadow is already stored in the map.
1016 Value *Shadow = ShadowMap[V];
1018 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1020 assert(Shadow && "No shadow for a value");
1024 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1025 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1026 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1030 if (Argument *A = dyn_cast<Argument>(V)) {
1031 // For arguments we compute the shadow on demand and store it in the map.
1032 Value **ShadowPtr = &ShadowMap[V];
1035 Function *F = A->getParent();
1036 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1037 unsigned ArgOffset = 0;
1038 for (auto &FArg : F->args()) {
1039 if (!FArg.getType()->isSized()) {
1040 DEBUG(dbgs() << "Arg is not sized\n");
1043 unsigned Size = FArg.hasByValAttr()
1044 ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
1045 : MS.DL->getTypeAllocSize(FArg.getType());
1047 bool Overflow = ArgOffset + Size > kParamTLSSize;
1048 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1049 if (FArg.hasByValAttr()) {
1050 // ByVal pointer itself has clean shadow. We copy the actual
1051 // argument shadow to the underlying memory.
1052 // Figure out maximal valid memcpy alignment.
1053 unsigned ArgAlign = FArg.getParamAlignment();
1054 if (ArgAlign == 0) {
1055 Type *EltType = A->getType()->getPointerElementType();
1056 ArgAlign = MS.DL->getABITypeAlignment(EltType);
1059 // ParamTLS overflow.
1060 EntryIRB.CreateMemSet(
1061 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1062 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1064 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1065 Value *Cpy = EntryIRB.CreateMemCpy(
1066 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1068 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1071 *ShadowPtr = getCleanShadow(V);
1074 // ParamTLS overflow.
1075 *ShadowPtr = getCleanShadow(V);
1078 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1081 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1082 **ShadowPtr << "\n");
1083 if (MS.TrackOrigins && !Overflow) {
1085 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1086 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1088 setOrigin(A, getCleanOrigin());
1091 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1093 assert(*ShadowPtr && "Could not find shadow for an argument");
1096 // For everything else the shadow is zero.
1097 return getCleanShadow(V);
1100 /// \brief Get the shadow for i-th argument of the instruction I.
1101 Value *getShadow(Instruction *I, int i) {
1102 return getShadow(I->getOperand(i));
1105 /// \brief Get the origin for a value.
1106 Value *getOrigin(Value *V) {
1107 if (!MS.TrackOrigins) return nullptr;
1108 if (!PropagateShadow) return getCleanOrigin();
1109 if (isa<Constant>(V)) return getCleanOrigin();
1110 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1111 "Unexpected value type in getOrigin()");
1112 Value *Origin = OriginMap[V];
1113 assert(Origin && "Missing origin");
1117 /// \brief Get the origin for i-th argument of the instruction I.
1118 Value *getOrigin(Instruction *I, int i) {
1119 return getOrigin(I->getOperand(i));
1122 /// \brief Remember the place where a shadow check should be inserted.
1124 /// This location will be later instrumented with a check that will print a
1125 /// UMR warning in runtime if the shadow value is not 0.
1126 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1128 if (!InsertChecks) return;
1130 Type *ShadowTy = Shadow->getType();
1131 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1132 "Can only insert checks for integer and vector shadow types");
1134 InstrumentationList.push_back(
1135 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1138 /// \brief Remember the place where a shadow check should be inserted.
1140 /// This location will be later instrumented with a check that will print a
1141 /// UMR warning in runtime if the value is not fully defined.
1142 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1144 Value *Shadow, *Origin;
1145 if (ClCheckConstantShadow) {
1146 Shadow = getShadow(Val);
1147 if (!Shadow) return;
1148 Origin = getOrigin(Val);
1150 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1151 if (!Shadow) return;
1152 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1154 insertShadowCheck(Shadow, Origin, OrigIns);
1157 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1166 case AcquireRelease:
1167 return AcquireRelease;
1168 case SequentiallyConsistent:
1169 return SequentiallyConsistent;
1171 llvm_unreachable("Unknown ordering");
1174 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1183 case AcquireRelease:
1184 return AcquireRelease;
1185 case SequentiallyConsistent:
1186 return SequentiallyConsistent;
1188 llvm_unreachable("Unknown ordering");
1191 // ------------------- Visitors.
1193 /// \brief Instrument LoadInst
1195 /// Loads the corresponding shadow and (optionally) origin.
1196 /// Optionally, checks that the load address is fully defined.
1197 void visitLoadInst(LoadInst &I) {
1198 assert(I.getType()->isSized() && "Load type must have size");
1199 IRBuilder<> IRB(I.getNextNode());
1200 Type *ShadowTy = getShadowTy(&I);
1201 Value *Addr = I.getPointerOperand();
1202 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1203 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1205 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1207 setShadow(&I, getCleanShadow(&I));
1210 if (ClCheckAccessAddress)
1211 insertShadowCheck(I.getPointerOperand(), &I);
1214 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1216 if (MS.TrackOrigins) {
1217 if (PropagateShadow) {
1218 unsigned Alignment = I.getAlignment();
1219 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1220 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1223 setOrigin(&I, getCleanOrigin());
1228 /// \brief Instrument StoreInst
1230 /// Stores the corresponding shadow and (optionally) origin.
1231 /// Optionally, checks that the store address is fully defined.
1232 void visitStoreInst(StoreInst &I) {
1233 StoreList.push_back(&I);
1236 void handleCASOrRMW(Instruction &I) {
1237 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1239 IRBuilder<> IRB(&I);
1240 Value *Addr = I.getOperand(0);
1241 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1243 if (ClCheckAccessAddress)
1244 insertShadowCheck(Addr, &I);
1246 // Only test the conditional argument of cmpxchg instruction.
1247 // The other argument can potentially be uninitialized, but we can not
1248 // detect this situation reliably without possible false positives.
1249 if (isa<AtomicCmpXchgInst>(I))
1250 insertShadowCheck(I.getOperand(1), &I);
1252 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1254 setShadow(&I, getCleanShadow(&I));
1255 setOrigin(&I, getCleanOrigin());
1258 void visitAtomicRMWInst(AtomicRMWInst &I) {
1260 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1263 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1265 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1268 // Vector manipulation.
1269 void visitExtractElementInst(ExtractElementInst &I) {
1270 insertShadowCheck(I.getOperand(1), &I);
1271 IRBuilder<> IRB(&I);
1272 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1274 setOrigin(&I, getOrigin(&I, 0));
1277 void visitInsertElementInst(InsertElementInst &I) {
1278 insertShadowCheck(I.getOperand(2), &I);
1279 IRBuilder<> IRB(&I);
1280 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1281 I.getOperand(2), "_msprop"));
1282 setOriginForNaryOp(I);
1285 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1286 insertShadowCheck(I.getOperand(2), &I);
1287 IRBuilder<> IRB(&I);
1288 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1289 I.getOperand(2), "_msprop"));
1290 setOriginForNaryOp(I);
1294 void visitSExtInst(SExtInst &I) {
1295 IRBuilder<> IRB(&I);
1296 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1297 setOrigin(&I, getOrigin(&I, 0));
1300 void visitZExtInst(ZExtInst &I) {
1301 IRBuilder<> IRB(&I);
1302 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1303 setOrigin(&I, getOrigin(&I, 0));
1306 void visitTruncInst(TruncInst &I) {
1307 IRBuilder<> IRB(&I);
1308 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1309 setOrigin(&I, getOrigin(&I, 0));
1312 void visitBitCastInst(BitCastInst &I) {
1313 IRBuilder<> IRB(&I);
1314 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1315 setOrigin(&I, getOrigin(&I, 0));
1318 void visitPtrToIntInst(PtrToIntInst &I) {
1319 IRBuilder<> IRB(&I);
1320 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1321 "_msprop_ptrtoint"));
1322 setOrigin(&I, getOrigin(&I, 0));
1325 void visitIntToPtrInst(IntToPtrInst &I) {
1326 IRBuilder<> IRB(&I);
1327 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1328 "_msprop_inttoptr"));
1329 setOrigin(&I, getOrigin(&I, 0));
1332 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1333 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1334 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1335 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1336 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1337 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1339 /// \brief Propagate shadow for bitwise AND.
1341 /// This code is exact, i.e. if, for example, a bit in the left argument
1342 /// is defined and 0, then neither the value not definedness of the
1343 /// corresponding bit in B don't affect the resulting shadow.
1344 void visitAnd(BinaryOperator &I) {
1345 IRBuilder<> IRB(&I);
1346 // "And" of 0 and a poisoned value results in unpoisoned value.
1347 // 1&1 => 1; 0&1 => 0; p&1 => p;
1348 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1349 // 1&p => p; 0&p => 0; p&p => p;
1350 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1351 Value *S1 = getShadow(&I, 0);
1352 Value *S2 = getShadow(&I, 1);
1353 Value *V1 = I.getOperand(0);
1354 Value *V2 = I.getOperand(1);
1355 if (V1->getType() != S1->getType()) {
1356 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1357 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1359 Value *S1S2 = IRB.CreateAnd(S1, S2);
1360 Value *V1S2 = IRB.CreateAnd(V1, S2);
1361 Value *S1V2 = IRB.CreateAnd(S1, V2);
1362 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1363 setOriginForNaryOp(I);
1366 void visitOr(BinaryOperator &I) {
1367 IRBuilder<> IRB(&I);
1368 // "Or" of 1 and a poisoned value results in unpoisoned value.
1369 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1370 // 1|0 => 1; 0|0 => 0; p|0 => p;
1371 // 1|p => 1; 0|p => p; p|p => p;
1372 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1373 Value *S1 = getShadow(&I, 0);
1374 Value *S2 = getShadow(&I, 1);
1375 Value *V1 = IRB.CreateNot(I.getOperand(0));
1376 Value *V2 = IRB.CreateNot(I.getOperand(1));
1377 if (V1->getType() != S1->getType()) {
1378 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1379 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1381 Value *S1S2 = IRB.CreateAnd(S1, S2);
1382 Value *V1S2 = IRB.CreateAnd(V1, S2);
1383 Value *S1V2 = IRB.CreateAnd(S1, V2);
1384 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1385 setOriginForNaryOp(I);
1388 /// \brief Default propagation of shadow and/or origin.
1390 /// This class implements the general case of shadow propagation, used in all
1391 /// cases where we don't know and/or don't care about what the operation
1392 /// actually does. It converts all input shadow values to a common type
1393 /// (extending or truncating as necessary), and bitwise OR's them.
1395 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1396 /// fully initialized), and less prone to false positives.
1398 /// This class also implements the general case of origin propagation. For a
1399 /// Nary operation, result origin is set to the origin of an argument that is
1400 /// not entirely initialized. If there is more than one such arguments, the
1401 /// rightmost of them is picked. It does not matter which one is picked if all
1402 /// arguments are initialized.
1403 template <bool CombineShadow>
1408 MemorySanitizerVisitor *MSV;
1411 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1412 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1414 /// \brief Add a pair of shadow and origin values to the mix.
1415 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1416 if (CombineShadow) {
1421 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1422 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1426 if (MSV->MS.TrackOrigins) {
1431 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1432 // No point in adding something that might result in 0 origin value.
1433 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1434 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1436 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1437 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1444 /// \brief Add an application value to the mix.
1445 Combiner &Add(Value *V) {
1446 Value *OpShadow = MSV->getShadow(V);
1447 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1448 return Add(OpShadow, OpOrigin);
1451 /// \brief Set the current combined values as the given instruction's shadow
1453 void Done(Instruction *I) {
1454 if (CombineShadow) {
1456 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1457 MSV->setShadow(I, Shadow);
1459 if (MSV->MS.TrackOrigins) {
1461 MSV->setOrigin(I, Origin);
1466 typedef Combiner<true> ShadowAndOriginCombiner;
1467 typedef Combiner<false> OriginCombiner;
1469 /// \brief Propagate origin for arbitrary operation.
1470 void setOriginForNaryOp(Instruction &I) {
1471 if (!MS.TrackOrigins) return;
1472 IRBuilder<> IRB(&I);
1473 OriginCombiner OC(this, IRB);
1474 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1479 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1480 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1481 "Vector of pointers is not a valid shadow type");
1482 return Ty->isVectorTy() ?
1483 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1484 Ty->getPrimitiveSizeInBits();
1487 /// \brief Cast between two shadow types, extending or truncating as
1489 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1490 bool Signed = false) {
1491 Type *srcTy = V->getType();
1492 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1493 return IRB.CreateIntCast(V, dstTy, Signed);
1494 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1495 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1496 return IRB.CreateIntCast(V, dstTy, Signed);
1497 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1498 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1499 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1501 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1502 return IRB.CreateBitCast(V2, dstTy);
1503 // TODO: handle struct types.
1506 /// \brief Cast an application value to the type of its own shadow.
1507 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1508 Type *ShadowTy = getShadowTy(V);
1509 if (V->getType() == ShadowTy)
1511 if (V->getType()->isPtrOrPtrVectorTy())
1512 return IRB.CreatePtrToInt(V, ShadowTy);
1514 return IRB.CreateBitCast(V, ShadowTy);
1517 /// \brief Propagate shadow for arbitrary operation.
1518 void handleShadowOr(Instruction &I) {
1519 IRBuilder<> IRB(&I);
1520 ShadowAndOriginCombiner SC(this, IRB);
1521 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1526 // \brief Handle multiplication by constant.
1528 // Handle a special case of multiplication by constant that may have one or
1529 // more zeros in the lower bits. This makes corresponding number of lower bits
1530 // of the result zero as well. We model it by shifting the other operand
1531 // shadow left by the required number of bits. Effectively, we transform
1532 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1533 // We use multiplication by 2**N instead of shift to cover the case of
1534 // multiplication by 0, which may occur in some elements of a vector operand.
1535 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1537 Constant *ShadowMul;
1538 Type *Ty = ConstArg->getType();
1539 if (Ty->isVectorTy()) {
1540 unsigned NumElements = Ty->getVectorNumElements();
1541 Type *EltTy = Ty->getSequentialElementType();
1542 SmallVector<Constant *, 16> Elements;
1543 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1545 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1546 APInt V = Elt->getValue();
1547 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1548 Elements.push_back(ConstantInt::get(EltTy, V2));
1550 ShadowMul = ConstantVector::get(Elements);
1552 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1553 APInt V = Elt->getValue();
1554 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1555 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1558 IRBuilder<> IRB(&I);
1560 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1561 setOrigin(&I, getOrigin(OtherArg));
1564 void visitMul(BinaryOperator &I) {
1565 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1566 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1567 if (constOp0 && !constOp1)
1568 handleMulByConstant(I, constOp0, I.getOperand(1));
1569 else if (constOp1 && !constOp0)
1570 handleMulByConstant(I, constOp1, I.getOperand(0));
1575 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1576 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1577 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1578 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1579 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1580 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1582 void handleDiv(Instruction &I) {
1583 IRBuilder<> IRB(&I);
1584 // Strict on the second argument.
1585 insertShadowCheck(I.getOperand(1), &I);
1586 setShadow(&I, getShadow(&I, 0));
1587 setOrigin(&I, getOrigin(&I, 0));
1590 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1591 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1592 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1593 void visitURem(BinaryOperator &I) { handleDiv(I); }
1594 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1595 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1597 /// \brief Instrument == and != comparisons.
1599 /// Sometimes the comparison result is known even if some of the bits of the
1600 /// arguments are not.
1601 void handleEqualityComparison(ICmpInst &I) {
1602 IRBuilder<> IRB(&I);
1603 Value *A = I.getOperand(0);
1604 Value *B = I.getOperand(1);
1605 Value *Sa = getShadow(A);
1606 Value *Sb = getShadow(B);
1608 // Get rid of pointers and vectors of pointers.
1609 // For ints (and vectors of ints), types of A and Sa match,
1610 // and this is a no-op.
1611 A = IRB.CreatePointerCast(A, Sa->getType());
1612 B = IRB.CreatePointerCast(B, Sb->getType());
1614 // A == B <==> (C = A^B) == 0
1615 // A != B <==> (C = A^B) != 0
1617 Value *C = IRB.CreateXor(A, B);
1618 Value *Sc = IRB.CreateOr(Sa, Sb);
1619 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1620 // Result is defined if one of the following is true
1621 // * there is a defined 1 bit in C
1622 // * C is fully defined
1623 // Si = !(C & ~Sc) && Sc
1624 Value *Zero = Constant::getNullValue(Sc->getType());
1625 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1627 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1629 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1630 Si->setName("_msprop_icmp");
1632 setOriginForNaryOp(I);
1635 /// \brief Build the lowest possible value of V, taking into account V's
1636 /// uninitialized bits.
1637 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1640 // Split shadow into sign bit and other bits.
1641 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1642 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1643 // Maximise the undefined shadow bit, minimize other undefined bits.
1645 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1647 // Minimize undefined bits.
1648 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1652 /// \brief Build the highest possible value of V, taking into account V's
1653 /// uninitialized bits.
1654 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1657 // Split shadow into sign bit and other bits.
1658 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1659 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1660 // Minimise the undefined shadow bit, maximise other undefined bits.
1662 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1664 // Maximize undefined bits.
1665 return IRB.CreateOr(A, Sa);
1669 /// \brief Instrument relational comparisons.
1671 /// This function does exact shadow propagation for all relational
1672 /// comparisons of integers, pointers and vectors of those.
1673 /// FIXME: output seems suboptimal when one of the operands is a constant
1674 void handleRelationalComparisonExact(ICmpInst &I) {
1675 IRBuilder<> IRB(&I);
1676 Value *A = I.getOperand(0);
1677 Value *B = I.getOperand(1);
1678 Value *Sa = getShadow(A);
1679 Value *Sb = getShadow(B);
1681 // Get rid of pointers and vectors of pointers.
1682 // For ints (and vectors of ints), types of A and Sa match,
1683 // and this is a no-op.
1684 A = IRB.CreatePointerCast(A, Sa->getType());
1685 B = IRB.CreatePointerCast(B, Sb->getType());
1687 // Let [a0, a1] be the interval of possible values of A, taking into account
1688 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1689 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1690 bool IsSigned = I.isSigned();
1691 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1692 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1693 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1694 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1695 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1696 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1697 Value *Si = IRB.CreateXor(S1, S2);
1699 setOriginForNaryOp(I);
1702 /// \brief Instrument signed relational comparisons.
1704 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1705 /// propagating the highest bit of the shadow. Everything else is delegated
1706 /// to handleShadowOr().
1707 void handleSignedRelationalComparison(ICmpInst &I) {
1708 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1709 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1710 Value* op = nullptr;
1711 CmpInst::Predicate pre = I.getPredicate();
1712 if (constOp0 && constOp0->isNullValue() &&
1713 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1714 op = I.getOperand(1);
1715 } else if (constOp1 && constOp1->isNullValue() &&
1716 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1717 op = I.getOperand(0);
1720 IRBuilder<> IRB(&I);
1722 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1723 setShadow(&I, Shadow);
1724 setOrigin(&I, getOrigin(op));
1730 void visitICmpInst(ICmpInst &I) {
1731 if (!ClHandleICmp) {
1735 if (I.isEquality()) {
1736 handleEqualityComparison(I);
1740 assert(I.isRelational());
1741 if (ClHandleICmpExact) {
1742 handleRelationalComparisonExact(I);
1746 handleSignedRelationalComparison(I);
1750 assert(I.isUnsigned());
1751 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1752 handleRelationalComparisonExact(I);
1759 void visitFCmpInst(FCmpInst &I) {
1763 void handleShift(BinaryOperator &I) {
1764 IRBuilder<> IRB(&I);
1765 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1766 // Otherwise perform the same shift on S1.
1767 Value *S1 = getShadow(&I, 0);
1768 Value *S2 = getShadow(&I, 1);
1769 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1771 Value *V2 = I.getOperand(1);
1772 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1773 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1774 setOriginForNaryOp(I);
1777 void visitShl(BinaryOperator &I) { handleShift(I); }
1778 void visitAShr(BinaryOperator &I) { handleShift(I); }
1779 void visitLShr(BinaryOperator &I) { handleShift(I); }
1781 /// \brief Instrument llvm.memmove
1783 /// At this point we don't know if llvm.memmove will be inlined or not.
1784 /// If we don't instrument it and it gets inlined,
1785 /// our interceptor will not kick in and we will lose the memmove.
1786 /// If we instrument the call here, but it does not get inlined,
1787 /// we will memove the shadow twice: which is bad in case
1788 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1790 /// Similar situation exists for memcpy and memset.
1791 void visitMemMoveInst(MemMoveInst &I) {
1792 IRBuilder<> IRB(&I);
1795 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1796 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1797 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1798 I.eraseFromParent();
1801 // Similar to memmove: avoid copying shadow twice.
1802 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1803 // FIXME: consider doing manual inline for small constant sizes and proper
1805 void visitMemCpyInst(MemCpyInst &I) {
1806 IRBuilder<> IRB(&I);
1809 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1810 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1811 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1812 I.eraseFromParent();
1816 void visitMemSetInst(MemSetInst &I) {
1817 IRBuilder<> IRB(&I);
1820 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1821 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1822 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1823 I.eraseFromParent();
1826 void visitVAStartInst(VAStartInst &I) {
1827 VAHelper->visitVAStartInst(I);
1830 void visitVACopyInst(VACopyInst &I) {
1831 VAHelper->visitVACopyInst(I);
1834 enum IntrinsicKind {
1835 IK_DoesNotAccessMemory,
1840 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1841 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1842 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1843 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1844 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1845 const int UnknownModRefBehavior = IK_WritesMemory;
1846 #define GET_INTRINSIC_MODREF_BEHAVIOR
1847 #define ModRefBehavior IntrinsicKind
1848 #include "llvm/IR/Intrinsics.gen"
1849 #undef ModRefBehavior
1850 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1853 /// \brief Handle vector store-like intrinsics.
1855 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1856 /// has 1 pointer argument and 1 vector argument, returns void.
1857 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1858 IRBuilder<> IRB(&I);
1859 Value* Addr = I.getArgOperand(0);
1860 Value *Shadow = getShadow(&I, 1);
1861 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1863 // We don't know the pointer alignment (could be unaligned SSE store!).
1864 // Have to assume to worst case.
1865 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1867 if (ClCheckAccessAddress)
1868 insertShadowCheck(Addr, &I);
1870 // FIXME: use ClStoreCleanOrigin
1871 // FIXME: factor out common code from materializeStores
1872 if (MS.TrackOrigins)
1873 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1877 /// \brief Handle vector load-like intrinsics.
1879 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1880 /// has 1 pointer argument, returns a vector.
1881 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1882 IRBuilder<> IRB(&I);
1883 Value *Addr = I.getArgOperand(0);
1885 Type *ShadowTy = getShadowTy(&I);
1886 if (PropagateShadow) {
1887 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1888 // We don't know the pointer alignment (could be unaligned SSE load!).
1889 // Have to assume to worst case.
1890 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1892 setShadow(&I, getCleanShadow(&I));
1895 if (ClCheckAccessAddress)
1896 insertShadowCheck(Addr, &I);
1898 if (MS.TrackOrigins) {
1899 if (PropagateShadow)
1900 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1902 setOrigin(&I, getCleanOrigin());
1907 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1909 /// Instrument intrinsics with any number of arguments of the same type,
1910 /// equal to the return type. The type should be simple (no aggregates or
1911 /// pointers; vectors are fine).
1912 /// Caller guarantees that this intrinsic does not access memory.
1913 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1914 Type *RetTy = I.getType();
1915 if (!(RetTy->isIntOrIntVectorTy() ||
1916 RetTy->isFPOrFPVectorTy() ||
1917 RetTy->isX86_MMXTy()))
1920 unsigned NumArgOperands = I.getNumArgOperands();
1922 for (unsigned i = 0; i < NumArgOperands; ++i) {
1923 Type *Ty = I.getArgOperand(i)->getType();
1928 IRBuilder<> IRB(&I);
1929 ShadowAndOriginCombiner SC(this, IRB);
1930 for (unsigned i = 0; i < NumArgOperands; ++i)
1931 SC.Add(I.getArgOperand(i));
1937 /// \brief Heuristically instrument unknown intrinsics.
1939 /// The main purpose of this code is to do something reasonable with all
1940 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1941 /// We recognize several classes of intrinsics by their argument types and
1942 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1943 /// sure that we know what the intrinsic does.
1945 /// We special-case intrinsics where this approach fails. See llvm.bswap
1946 /// handling as an example of that.
1947 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1948 unsigned NumArgOperands = I.getNumArgOperands();
1949 if (NumArgOperands == 0)
1952 Intrinsic::ID iid = I.getIntrinsicID();
1953 IntrinsicKind IK = getIntrinsicKind(iid);
1954 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1955 bool WritesMemory = IK == IK_WritesMemory;
1956 assert(!(OnlyReadsMemory && WritesMemory));
1958 if (NumArgOperands == 2 &&
1959 I.getArgOperand(0)->getType()->isPointerTy() &&
1960 I.getArgOperand(1)->getType()->isVectorTy() &&
1961 I.getType()->isVoidTy() &&
1963 // This looks like a vector store.
1964 return handleVectorStoreIntrinsic(I);
1967 if (NumArgOperands == 1 &&
1968 I.getArgOperand(0)->getType()->isPointerTy() &&
1969 I.getType()->isVectorTy() &&
1971 // This looks like a vector load.
1972 return handleVectorLoadIntrinsic(I);
1975 if (!OnlyReadsMemory && !WritesMemory)
1976 if (maybeHandleSimpleNomemIntrinsic(I))
1979 // FIXME: detect and handle SSE maskstore/maskload
1983 void handleBswap(IntrinsicInst &I) {
1984 IRBuilder<> IRB(&I);
1985 Value *Op = I.getArgOperand(0);
1986 Type *OpType = Op->getType();
1987 Function *BswapFunc = Intrinsic::getDeclaration(
1988 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
1989 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1990 setOrigin(&I, getOrigin(Op));
1993 // \brief Instrument vector convert instrinsic.
1995 // This function instruments intrinsics like cvtsi2ss:
1996 // %Out = int_xxx_cvtyyy(%ConvertOp)
1998 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1999 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2000 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2001 // elements from \p CopyOp.
2002 // In most cases conversion involves floating-point value which may trigger a
2003 // hardware exception when not fully initialized. For this reason we require
2004 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2005 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2006 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2007 // return a fully initialized value.
2008 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2009 IRBuilder<> IRB(&I);
2010 Value *CopyOp, *ConvertOp;
2012 switch (I.getNumArgOperands()) {
2014 CopyOp = I.getArgOperand(0);
2015 ConvertOp = I.getArgOperand(1);
2018 ConvertOp = I.getArgOperand(0);
2022 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2025 // The first *NumUsedElements* elements of ConvertOp are converted to the
2026 // same number of output elements. The rest of the output is copied from
2027 // CopyOp, or (if not available) filled with zeroes.
2028 // Combine shadow for elements of ConvertOp that are used in this operation,
2029 // and insert a check.
2030 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2031 // int->any conversion.
2032 Value *ConvertShadow = getShadow(ConvertOp);
2033 Value *AggShadow = nullptr;
2034 if (ConvertOp->getType()->isVectorTy()) {
2035 AggShadow = IRB.CreateExtractElement(
2036 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2037 for (int i = 1; i < NumUsedElements; ++i) {
2038 Value *MoreShadow = IRB.CreateExtractElement(
2039 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2040 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2043 AggShadow = ConvertShadow;
2045 assert(AggShadow->getType()->isIntegerTy());
2046 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2048 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2051 assert(CopyOp->getType() == I.getType());
2052 assert(CopyOp->getType()->isVectorTy());
2053 Value *ResultShadow = getShadow(CopyOp);
2054 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2055 for (int i = 0; i < NumUsedElements; ++i) {
2056 ResultShadow = IRB.CreateInsertElement(
2057 ResultShadow, ConstantInt::getNullValue(EltTy),
2058 ConstantInt::get(IRB.getInt32Ty(), i));
2060 setShadow(&I, ResultShadow);
2061 setOrigin(&I, getOrigin(CopyOp));
2063 setShadow(&I, getCleanShadow(&I));
2064 setOrigin(&I, getCleanOrigin());
2068 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2069 // zeroes if it is zero, and all ones otherwise.
2070 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2071 if (S->getType()->isVectorTy())
2072 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2073 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2074 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2075 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2078 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2079 Type *T = S->getType();
2080 assert(T->isVectorTy());
2081 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2082 return IRB.CreateSExt(S2, T);
2085 // \brief Instrument vector shift instrinsic.
2087 // This function instruments intrinsics like int_x86_avx2_psll_w.
2088 // Intrinsic shifts %In by %ShiftSize bits.
2089 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2090 // size, and the rest is ignored. Behavior is defined even if shift size is
2091 // greater than register (or field) width.
2092 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2093 assert(I.getNumArgOperands() == 2);
2094 IRBuilder<> IRB(&I);
2095 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2096 // Otherwise perform the same shift on S1.
2097 Value *S1 = getShadow(&I, 0);
2098 Value *S2 = getShadow(&I, 1);
2099 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2100 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2101 Value *V1 = I.getOperand(0);
2102 Value *V2 = I.getOperand(1);
2103 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
2104 IRB.CreateBitCast(S1, V1->getType()), V2);
2105 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2106 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2107 setOriginForNaryOp(I);
2110 // \brief Get an X86_MMX-sized vector type.
2111 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2112 const unsigned X86_MMXSizeInBits = 64;
2113 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2114 X86_MMXSizeInBits / EltSizeInBits);
2117 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2119 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2121 case llvm::Intrinsic::x86_sse2_packsswb_128:
2122 case llvm::Intrinsic::x86_sse2_packuswb_128:
2123 return llvm::Intrinsic::x86_sse2_packsswb_128;
2125 case llvm::Intrinsic::x86_sse2_packssdw_128:
2126 case llvm::Intrinsic::x86_sse41_packusdw:
2127 return llvm::Intrinsic::x86_sse2_packssdw_128;
2129 case llvm::Intrinsic::x86_avx2_packsswb:
2130 case llvm::Intrinsic::x86_avx2_packuswb:
2131 return llvm::Intrinsic::x86_avx2_packsswb;
2133 case llvm::Intrinsic::x86_avx2_packssdw:
2134 case llvm::Intrinsic::x86_avx2_packusdw:
2135 return llvm::Intrinsic::x86_avx2_packssdw;
2137 case llvm::Intrinsic::x86_mmx_packsswb:
2138 case llvm::Intrinsic::x86_mmx_packuswb:
2139 return llvm::Intrinsic::x86_mmx_packsswb;
2141 case llvm::Intrinsic::x86_mmx_packssdw:
2142 return llvm::Intrinsic::x86_mmx_packssdw;
2144 llvm_unreachable("unexpected intrinsic id");
2148 // \brief Instrument vector pack instrinsic.
2150 // This function instruments intrinsics like x86_mmx_packsswb, that
2151 // packs elements of 2 input vectors into half as many bits with saturation.
2152 // Shadow is propagated with the signed variant of the same intrinsic applied
2153 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2154 // EltSizeInBits is used only for x86mmx arguments.
2155 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2156 assert(I.getNumArgOperands() == 2);
2157 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2158 IRBuilder<> IRB(&I);
2159 Value *S1 = getShadow(&I, 0);
2160 Value *S2 = getShadow(&I, 1);
2161 assert(isX86_MMX || S1->getType()->isVectorTy());
2163 // SExt and ICmpNE below must apply to individual elements of input vectors.
2164 // In case of x86mmx arguments, cast them to appropriate vector types and
2166 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2168 S1 = IRB.CreateBitCast(S1, T);
2169 S2 = IRB.CreateBitCast(S2, T);
2171 Value *S1_ext = IRB.CreateSExt(
2172 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2173 Value *S2_ext = IRB.CreateSExt(
2174 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2176 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2177 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2178 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2181 Function *ShadowFn = Intrinsic::getDeclaration(
2182 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2184 Value *S = IRB.CreateCall2(ShadowFn, S1_ext, S2_ext, "_msprop_vector_pack");
2185 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2187 setOriginForNaryOp(I);
2190 // \brief Instrument sum-of-absolute-differencies intrinsic.
2191 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2192 const unsigned SignificantBitsPerResultElement = 16;
2193 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2194 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2195 unsigned ZeroBitsPerResultElement =
2196 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2198 IRBuilder<> IRB(&I);
2199 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2200 S = IRB.CreateBitCast(S, ResTy);
2201 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2203 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2204 S = IRB.CreateBitCast(S, getShadowTy(&I));
2206 setOriginForNaryOp(I);
2209 // \brief Instrument multiply-add intrinsic.
2210 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2211 unsigned EltSizeInBits = 0) {
2212 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2213 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2214 IRBuilder<> IRB(&I);
2215 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2216 S = IRB.CreateBitCast(S, ResTy);
2217 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2219 S = IRB.CreateBitCast(S, getShadowTy(&I));
2221 setOriginForNaryOp(I);
2224 void visitIntrinsicInst(IntrinsicInst &I) {
2225 switch (I.getIntrinsicID()) {
2226 case llvm::Intrinsic::bswap:
2229 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2230 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2231 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2232 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2233 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2234 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2235 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2236 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2237 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2238 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2239 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2240 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2241 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2242 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2243 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2244 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2245 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2246 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2247 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2248 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2249 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2250 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2251 case llvm::Intrinsic::x86_sse_cvtss2si64:
2252 case llvm::Intrinsic::x86_sse_cvtss2si:
2253 case llvm::Intrinsic::x86_sse_cvttss2si64:
2254 case llvm::Intrinsic::x86_sse_cvttss2si:
2255 handleVectorConvertIntrinsic(I, 1);
2257 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2258 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2259 case llvm::Intrinsic::x86_sse_cvtps2pi:
2260 case llvm::Intrinsic::x86_sse_cvttps2pi:
2261 handleVectorConvertIntrinsic(I, 2);
2263 case llvm::Intrinsic::x86_avx2_psll_w:
2264 case llvm::Intrinsic::x86_avx2_psll_d:
2265 case llvm::Intrinsic::x86_avx2_psll_q:
2266 case llvm::Intrinsic::x86_avx2_pslli_w:
2267 case llvm::Intrinsic::x86_avx2_pslli_d:
2268 case llvm::Intrinsic::x86_avx2_pslli_q:
2269 case llvm::Intrinsic::x86_avx2_psrl_w:
2270 case llvm::Intrinsic::x86_avx2_psrl_d:
2271 case llvm::Intrinsic::x86_avx2_psrl_q:
2272 case llvm::Intrinsic::x86_avx2_psra_w:
2273 case llvm::Intrinsic::x86_avx2_psra_d:
2274 case llvm::Intrinsic::x86_avx2_psrli_w:
2275 case llvm::Intrinsic::x86_avx2_psrli_d:
2276 case llvm::Intrinsic::x86_avx2_psrli_q:
2277 case llvm::Intrinsic::x86_avx2_psrai_w:
2278 case llvm::Intrinsic::x86_avx2_psrai_d:
2279 case llvm::Intrinsic::x86_sse2_psll_w:
2280 case llvm::Intrinsic::x86_sse2_psll_d:
2281 case llvm::Intrinsic::x86_sse2_psll_q:
2282 case llvm::Intrinsic::x86_sse2_pslli_w:
2283 case llvm::Intrinsic::x86_sse2_pslli_d:
2284 case llvm::Intrinsic::x86_sse2_pslli_q:
2285 case llvm::Intrinsic::x86_sse2_psrl_w:
2286 case llvm::Intrinsic::x86_sse2_psrl_d:
2287 case llvm::Intrinsic::x86_sse2_psrl_q:
2288 case llvm::Intrinsic::x86_sse2_psra_w:
2289 case llvm::Intrinsic::x86_sse2_psra_d:
2290 case llvm::Intrinsic::x86_sse2_psrli_w:
2291 case llvm::Intrinsic::x86_sse2_psrli_d:
2292 case llvm::Intrinsic::x86_sse2_psrli_q:
2293 case llvm::Intrinsic::x86_sse2_psrai_w:
2294 case llvm::Intrinsic::x86_sse2_psrai_d:
2295 case llvm::Intrinsic::x86_mmx_psll_w:
2296 case llvm::Intrinsic::x86_mmx_psll_d:
2297 case llvm::Intrinsic::x86_mmx_psll_q:
2298 case llvm::Intrinsic::x86_mmx_pslli_w:
2299 case llvm::Intrinsic::x86_mmx_pslli_d:
2300 case llvm::Intrinsic::x86_mmx_pslli_q:
2301 case llvm::Intrinsic::x86_mmx_psrl_w:
2302 case llvm::Intrinsic::x86_mmx_psrl_d:
2303 case llvm::Intrinsic::x86_mmx_psrl_q:
2304 case llvm::Intrinsic::x86_mmx_psra_w:
2305 case llvm::Intrinsic::x86_mmx_psra_d:
2306 case llvm::Intrinsic::x86_mmx_psrli_w:
2307 case llvm::Intrinsic::x86_mmx_psrli_d:
2308 case llvm::Intrinsic::x86_mmx_psrli_q:
2309 case llvm::Intrinsic::x86_mmx_psrai_w:
2310 case llvm::Intrinsic::x86_mmx_psrai_d:
2311 handleVectorShiftIntrinsic(I, /* Variable */ false);
2313 case llvm::Intrinsic::x86_avx2_psllv_d:
2314 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2315 case llvm::Intrinsic::x86_avx2_psllv_q:
2316 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2317 case llvm::Intrinsic::x86_avx2_psrlv_d:
2318 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2319 case llvm::Intrinsic::x86_avx2_psrlv_q:
2320 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2321 case llvm::Intrinsic::x86_avx2_psrav_d:
2322 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2323 handleVectorShiftIntrinsic(I, /* Variable */ true);
2326 case llvm::Intrinsic::x86_sse2_packsswb_128:
2327 case llvm::Intrinsic::x86_sse2_packssdw_128:
2328 case llvm::Intrinsic::x86_sse2_packuswb_128:
2329 case llvm::Intrinsic::x86_sse41_packusdw:
2330 case llvm::Intrinsic::x86_avx2_packsswb:
2331 case llvm::Intrinsic::x86_avx2_packssdw:
2332 case llvm::Intrinsic::x86_avx2_packuswb:
2333 case llvm::Intrinsic::x86_avx2_packusdw:
2334 handleVectorPackIntrinsic(I);
2337 case llvm::Intrinsic::x86_mmx_packsswb:
2338 case llvm::Intrinsic::x86_mmx_packuswb:
2339 handleVectorPackIntrinsic(I, 16);
2342 case llvm::Intrinsic::x86_mmx_packssdw:
2343 handleVectorPackIntrinsic(I, 32);
2346 case llvm::Intrinsic::x86_mmx_psad_bw:
2347 case llvm::Intrinsic::x86_sse2_psad_bw:
2348 case llvm::Intrinsic::x86_avx2_psad_bw:
2349 handleVectorSadIntrinsic(I);
2352 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2353 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2354 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2355 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2356 handleVectorPmaddIntrinsic(I);
2359 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2360 handleVectorPmaddIntrinsic(I, 8);
2363 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2364 handleVectorPmaddIntrinsic(I, 16);
2368 if (!handleUnknownIntrinsic(I))
2369 visitInstruction(I);
2374 void visitCallSite(CallSite CS) {
2375 Instruction &I = *CS.getInstruction();
2376 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2378 CallInst *Call = cast<CallInst>(&I);
2380 // For inline asm, do the usual thing: check argument shadow and mark all
2381 // outputs as clean. Note that any side effects of the inline asm that are
2382 // not immediately visible in its constraints are not handled.
2383 if (Call->isInlineAsm()) {
2384 visitInstruction(I);
2388 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2390 // We are going to insert code that relies on the fact that the callee
2391 // will become a non-readonly function after it is instrumented by us. To
2392 // prevent this code from being optimized out, mark that function
2393 // non-readonly in advance.
2394 if (Function *Func = Call->getCalledFunction()) {
2395 // Clear out readonly/readnone attributes.
2397 B.addAttribute(Attribute::ReadOnly)
2398 .addAttribute(Attribute::ReadNone);
2399 Func->removeAttributes(AttributeSet::FunctionIndex,
2400 AttributeSet::get(Func->getContext(),
2401 AttributeSet::FunctionIndex,
2405 IRBuilder<> IRB(&I);
2407 unsigned ArgOffset = 0;
2408 DEBUG(dbgs() << " CallSite: " << I << "\n");
2409 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2410 ArgIt != End; ++ArgIt) {
2412 unsigned i = ArgIt - CS.arg_begin();
2413 if (!A->getType()->isSized()) {
2414 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2418 Value *Store = nullptr;
2419 // Compute the Shadow for arg even if it is ByVal, because
2420 // in that case getShadow() will copy the actual arg shadow to
2421 // __msan_param_tls.
2422 Value *ArgShadow = getShadow(A);
2423 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2424 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2425 " Shadow: " << *ArgShadow << "\n");
2426 bool ArgIsInitialized = false;
2427 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2428 assert(A->getType()->isPointerTy() &&
2429 "ByVal argument is not a pointer!");
2430 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2431 if (ArgOffset + Size > kParamTLSSize) break;
2432 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2433 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2434 Store = IRB.CreateMemCpy(ArgShadowBase,
2435 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2438 Size = MS.DL->getTypeAllocSize(A->getType());
2439 if (ArgOffset + Size > kParamTLSSize) break;
2440 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2441 kShadowTLSAlignment);
2442 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2443 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2445 if (MS.TrackOrigins && !ArgIsInitialized)
2446 IRB.CreateStore(getOrigin(A),
2447 getOriginPtrForArgument(A, IRB, ArgOffset));
2449 assert(Size != 0 && Store != nullptr);
2450 DEBUG(dbgs() << " Param:" << *Store << "\n");
2451 ArgOffset += RoundUpToAlignment(Size, 8);
2453 DEBUG(dbgs() << " done with call args\n");
2456 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2457 if (FT->isVarArg()) {
2458 VAHelper->visitCallSite(CS, IRB);
2461 // Now, get the shadow for the RetVal.
2462 if (!I.getType()->isSized()) return;
2463 IRBuilder<> IRBBefore(&I);
2464 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2465 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2466 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2467 Instruction *NextInsn = nullptr;
2469 NextInsn = I.getNextNode();
2471 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2472 if (!NormalDest->getSinglePredecessor()) {
2473 // FIXME: this case is tricky, so we are just conservative here.
2474 // Perhaps we need to split the edge between this BB and NormalDest,
2475 // but a naive attempt to use SplitEdge leads to a crash.
2476 setShadow(&I, getCleanShadow(&I));
2477 setOrigin(&I, getCleanOrigin());
2480 NextInsn = NormalDest->getFirstInsertionPt();
2482 "Could not find insertion point for retval shadow load");
2484 IRBuilder<> IRBAfter(NextInsn);
2485 Value *RetvalShadow =
2486 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2487 kShadowTLSAlignment, "_msret");
2488 setShadow(&I, RetvalShadow);
2489 if (MS.TrackOrigins)
2490 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2493 void visitReturnInst(ReturnInst &I) {
2494 IRBuilder<> IRB(&I);
2495 Value *RetVal = I.getReturnValue();
2496 if (!RetVal) return;
2497 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2498 if (CheckReturnValue) {
2499 insertShadowCheck(RetVal, &I);
2500 Value *Shadow = getCleanShadow(RetVal);
2501 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2503 Value *Shadow = getShadow(RetVal);
2504 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2505 // FIXME: make it conditional if ClStoreCleanOrigin==0
2506 if (MS.TrackOrigins)
2507 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2511 void visitPHINode(PHINode &I) {
2512 IRBuilder<> IRB(&I);
2513 if (!PropagateShadow) {
2514 setShadow(&I, getCleanShadow(&I));
2515 setOrigin(&I, getCleanOrigin());
2519 ShadowPHINodes.push_back(&I);
2520 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2522 if (MS.TrackOrigins)
2523 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2527 void visitAllocaInst(AllocaInst &I) {
2528 setShadow(&I, getCleanShadow(&I));
2529 setOrigin(&I, getCleanOrigin());
2530 IRBuilder<> IRB(I.getNextNode());
2531 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2532 if (PoisonStack && ClPoisonStackWithCall) {
2533 IRB.CreateCall2(MS.MsanPoisonStackFn,
2534 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2535 ConstantInt::get(MS.IntptrTy, Size));
2537 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2538 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2539 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2542 if (PoisonStack && MS.TrackOrigins) {
2543 SmallString<2048> StackDescriptionStorage;
2544 raw_svector_ostream StackDescription(StackDescriptionStorage);
2545 // We create a string with a description of the stack allocation and
2546 // pass it into __msan_set_alloca_origin.
2547 // It will be printed by the run-time if stack-originated UMR is found.
2548 // The first 4 bytes of the string are set to '----' and will be replaced
2549 // by __msan_va_arg_overflow_size_tls at the first call.
2550 StackDescription << "----" << I.getName() << "@" << F.getName();
2552 createPrivateNonConstGlobalForString(*F.getParent(),
2553 StackDescription.str());
2555 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2556 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2557 ConstantInt::get(MS.IntptrTy, Size),
2558 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2559 IRB.CreatePointerCast(&F, MS.IntptrTy));
2563 void visitSelectInst(SelectInst& I) {
2564 IRBuilder<> IRB(&I);
2565 // a = select b, c, d
2566 Value *B = I.getCondition();
2567 Value *C = I.getTrueValue();
2568 Value *D = I.getFalseValue();
2569 Value *Sb = getShadow(B);
2570 Value *Sc = getShadow(C);
2571 Value *Sd = getShadow(D);
2573 // Result shadow if condition shadow is 0.
2574 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2576 if (I.getType()->isAggregateType()) {
2577 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2578 // an extra "select". This results in much more compact IR.
2579 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2580 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2582 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2583 // If Sb (condition is poisoned), look for bits in c and d that are equal
2584 // and both unpoisoned.
2585 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2587 // Cast arguments to shadow-compatible type.
2588 C = CreateAppToShadowCast(IRB, C);
2589 D = CreateAppToShadowCast(IRB, D);
2591 // Result shadow if condition shadow is 1.
2592 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2594 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2596 if (MS.TrackOrigins) {
2597 // Origins are always i32, so any vector conditions must be flattened.
2598 // FIXME: consider tracking vector origins for app vectors?
2599 if (B->getType()->isVectorTy()) {
2600 Type *FlatTy = getShadowTyNoVec(B->getType());
2601 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2602 ConstantInt::getNullValue(FlatTy));
2603 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2604 ConstantInt::getNullValue(FlatTy));
2606 // a = select b, c, d
2607 // Oa = Sb ? Ob : (b ? Oc : Od)
2609 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2610 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2611 getOrigin(I.getFalseValue()))));
2615 void visitLandingPadInst(LandingPadInst &I) {
2617 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2618 setShadow(&I, getCleanShadow(&I));
2619 setOrigin(&I, getCleanOrigin());
2622 void visitGetElementPtrInst(GetElementPtrInst &I) {
2626 void visitExtractValueInst(ExtractValueInst &I) {
2627 IRBuilder<> IRB(&I);
2628 Value *Agg = I.getAggregateOperand();
2629 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2630 Value *AggShadow = getShadow(Agg);
2631 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2632 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2633 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2634 setShadow(&I, ResShadow);
2635 setOriginForNaryOp(I);
2638 void visitInsertValueInst(InsertValueInst &I) {
2639 IRBuilder<> IRB(&I);
2640 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2641 Value *AggShadow = getShadow(I.getAggregateOperand());
2642 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2643 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2644 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2645 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2646 DEBUG(dbgs() << " Res: " << *Res << "\n");
2648 setOriginForNaryOp(I);
2651 void dumpInst(Instruction &I) {
2652 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2653 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2655 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2657 errs() << "QQQ " << I << "\n";
2660 void visitResumeInst(ResumeInst &I) {
2661 DEBUG(dbgs() << "Resume: " << I << "\n");
2662 // Nothing to do here.
2665 void visitInstruction(Instruction &I) {
2666 // Everything else: stop propagating and check for poisoned shadow.
2667 if (ClDumpStrictInstructions)
2669 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2670 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2671 insertShadowCheck(I.getOperand(i), &I);
2672 setShadow(&I, getCleanShadow(&I));
2673 setOrigin(&I, getCleanOrigin());
2677 /// \brief AMD64-specific implementation of VarArgHelper.
2678 struct VarArgAMD64Helper : public VarArgHelper {
2679 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2680 // See a comment in visitCallSite for more details.
2681 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2682 static const unsigned AMD64FpEndOffset = 176;
2685 MemorySanitizer &MS;
2686 MemorySanitizerVisitor &MSV;
2687 Value *VAArgTLSCopy;
2688 Value *VAArgOverflowSize;
2690 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2692 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2693 MemorySanitizerVisitor &MSV)
2694 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2695 VAArgOverflowSize(nullptr) {}
2697 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2699 ArgKind classifyArgument(Value* arg) {
2700 // A very rough approximation of X86_64 argument classification rules.
2701 Type *T = arg->getType();
2702 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2703 return AK_FloatingPoint;
2704 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2705 return AK_GeneralPurpose;
2706 if (T->isPointerTy())
2707 return AK_GeneralPurpose;
2711 // For VarArg functions, store the argument shadow in an ABI-specific format
2712 // that corresponds to va_list layout.
2713 // We do this because Clang lowers va_arg in the frontend, and this pass
2714 // only sees the low level code that deals with va_list internals.
2715 // A much easier alternative (provided that Clang emits va_arg instructions)
2716 // would have been to associate each live instance of va_list with a copy of
2717 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2719 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2720 unsigned GpOffset = 0;
2721 unsigned FpOffset = AMD64GpEndOffset;
2722 unsigned OverflowOffset = AMD64FpEndOffset;
2723 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2724 ArgIt != End; ++ArgIt) {
2726 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2727 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2729 // ByVal arguments always go to the overflow area.
2730 assert(A->getType()->isPointerTy());
2731 Type *RealTy = A->getType()->getPointerElementType();
2732 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2733 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2734 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2735 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2736 ArgSize, kShadowTLSAlignment);
2738 ArgKind AK = classifyArgument(A);
2739 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2741 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2745 case AK_GeneralPurpose:
2746 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2749 case AK_FloatingPoint:
2750 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2754 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2755 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2756 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2758 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2761 Constant *OverflowSize =
2762 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2763 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2766 /// \brief Compute the shadow address for a given va_arg.
2767 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2769 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2770 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2771 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2775 void visitVAStartInst(VAStartInst &I) override {
2776 IRBuilder<> IRB(&I);
2777 VAStartInstrumentationList.push_back(&I);
2778 Value *VAListTag = I.getArgOperand(0);
2779 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2781 // Unpoison the whole __va_list_tag.
2782 // FIXME: magic ABI constants.
2783 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2784 /* size */24, /* alignment */8, false);
2787 void visitVACopyInst(VACopyInst &I) override {
2788 IRBuilder<> IRB(&I);
2789 Value *VAListTag = I.getArgOperand(0);
2790 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2792 // Unpoison the whole __va_list_tag.
2793 // FIXME: magic ABI constants.
2794 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2795 /* size */24, /* alignment */8, false);
2798 void finalizeInstrumentation() override {
2799 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2800 "finalizeInstrumentation called twice");
2801 if (!VAStartInstrumentationList.empty()) {
2802 // If there is a va_start in this function, make a backup copy of
2803 // va_arg_tls somewhere in the function entry block.
2804 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2805 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2807 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2809 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2810 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2813 // Instrument va_start.
2814 // Copy va_list shadow from the backup copy of the TLS contents.
2815 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2816 CallInst *OrigInst = VAStartInstrumentationList[i];
2817 IRBuilder<> IRB(OrigInst->getNextNode());
2818 Value *VAListTag = OrigInst->getArgOperand(0);
2820 Value *RegSaveAreaPtrPtr =
2822 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2823 ConstantInt::get(MS.IntptrTy, 16)),
2824 Type::getInt64PtrTy(*MS.C));
2825 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2826 Value *RegSaveAreaShadowPtr =
2827 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2828 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2829 AMD64FpEndOffset, 16);
2831 Value *OverflowArgAreaPtrPtr =
2833 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2834 ConstantInt::get(MS.IntptrTy, 8)),
2835 Type::getInt64PtrTy(*MS.C));
2836 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2837 Value *OverflowArgAreaShadowPtr =
2838 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2839 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2840 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2845 /// \brief MIPS64-specific implementation of VarArgHelper.
2846 struct VarArgMIPS64Helper : public VarArgHelper {
2848 MemorySanitizer &MS;
2849 MemorySanitizerVisitor &MSV;
2850 Value *VAArgTLSCopy;
2853 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2855 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
2856 MemorySanitizerVisitor &MSV)
2857 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2858 VAArgSize(nullptr) {}
2860 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2861 unsigned VAArgOffset = 0;
2862 for (CallSite::arg_iterator ArgIt = CS.arg_begin() + 1, End = CS.arg_end();
2863 ArgIt != End; ++ArgIt) {
2866 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2867 #if defined(__MIPSEB__) || defined(MIPSEB)
2868 // Adjusting the shadow for argument with size < 8 to match the placement
2869 // of bits in big endian system
2871 VAArgOffset += (8 - ArgSize);
2873 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
2874 VAArgOffset += ArgSize;
2875 VAArgOffset = RoundUpToAlignment(VAArgOffset, 8);
2876 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2879 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
2880 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
2881 // a new class member i.e. it is the total size of all VarArgs.
2882 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
2885 /// \brief Compute the shadow address for a given va_arg.
2886 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2888 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2889 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2890 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2894 void visitVAStartInst(VAStartInst &I) override {
2895 IRBuilder<> IRB(&I);
2896 VAStartInstrumentationList.push_back(&I);
2897 Value *VAListTag = I.getArgOperand(0);
2898 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2899 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2900 /* size */8, /* alignment */8, false);
2903 void visitVACopyInst(VACopyInst &I) override {
2904 IRBuilder<> IRB(&I);
2905 Value *VAListTag = I.getArgOperand(0);
2906 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2907 // Unpoison the whole __va_list_tag.
2908 // FIXME: magic ABI constants.
2909 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2910 /* size */8, /* alignment */8, false);
2913 void finalizeInstrumentation() override {
2914 assert(!VAArgSize && !VAArgTLSCopy &&
2915 "finalizeInstrumentation called twice");
2916 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2917 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2918 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
2921 if (!VAStartInstrumentationList.empty()) {
2922 // If there is a va_start in this function, make a backup copy of
2923 // va_arg_tls somewhere in the function entry block.
2924 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2925 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2928 // Instrument va_start.
2929 // Copy va_list shadow from the backup copy of the TLS contents.
2930 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2931 CallInst *OrigInst = VAStartInstrumentationList[i];
2932 IRBuilder<> IRB(OrigInst->getNextNode());
2933 Value *VAListTag = OrigInst->getArgOperand(0);
2934 Value *RegSaveAreaPtrPtr =
2935 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2936 Type::getInt64PtrTy(*MS.C));
2937 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2938 Value *RegSaveAreaShadowPtr =
2939 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2940 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
2945 /// \brief A no-op implementation of VarArgHelper.
2946 struct VarArgNoOpHelper : public VarArgHelper {
2947 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2948 MemorySanitizerVisitor &MSV) {}
2950 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2952 void visitVAStartInst(VAStartInst &I) override {}
2954 void visitVACopyInst(VACopyInst &I) override {}
2956 void finalizeInstrumentation() override {}
2959 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2960 MemorySanitizerVisitor &Visitor) {
2961 // VarArg handling is only implemented on AMD64. False positives are possible
2962 // on other platforms.
2963 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2964 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2965 return new VarArgAMD64Helper(Func, Msan, Visitor);
2966 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
2967 TargetTriple.getArch() == llvm::Triple::mips64el)
2968 return new VarArgMIPS64Helper(Func, Msan, Visitor);
2970 return new VarArgNoOpHelper(Func, Msan, Visitor);
2975 bool MemorySanitizer::runOnFunction(Function &F) {
2976 MemorySanitizerVisitor Visitor(F, *this);
2978 // Clear out readonly/readnone attributes.
2980 B.addAttribute(Attribute::ReadOnly)
2981 .addAttribute(Attribute::ReadNone);
2982 F.removeAttributes(AttributeSet::FunctionIndex,
2983 AttributeSet::get(F.getContext(),
2984 AttributeSet::FunctionIndex, B));
2986 return Visitor.runOnFunction();