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));
194 static const char *const kMsanModuleCtorName = "msan.module_ctor";
195 static const char *const kMsanInitName = "__msan_init";
199 // Memory map parameters used in application-to-shadow address calculation.
200 // Offset = (Addr & ~AndMask) ^ XorMask
201 // Shadow = ShadowBase + Offset
202 // Origin = OriginBase + Offset
203 struct MemoryMapParams {
210 struct PlatformMemoryMapParams {
211 const MemoryMapParams *bits32;
212 const MemoryMapParams *bits64;
216 static const MemoryMapParams Linux_I386_MemoryMapParams = {
217 0x000080000000, // AndMask
218 0, // XorMask (not used)
219 0, // ShadowBase (not used)
220 0x000040000000, // OriginBase
224 static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
225 0x400000000000, // AndMask
226 0, // XorMask (not used)
227 0, // ShadowBase (not used)
228 0x200000000000, // OriginBase
232 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
233 0x004000000000, // AndMask
234 0, // XorMask (not used)
235 0, // ShadowBase (not used)
236 0x002000000000, // OriginBase
240 static const MemoryMapParams Linux_PowerPC64_MemoryMapParams = {
241 0x200000000000, // AndMask
242 0x100000000000, // XorMask
243 0x080000000000, // ShadowBase
244 0x1C0000000000, // OriginBase
248 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
249 0x000180000000, // AndMask
250 0x000040000000, // XorMask
251 0x000020000000, // ShadowBase
252 0x000700000000, // OriginBase
256 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
257 0xc00000000000, // AndMask
258 0x200000000000, // XorMask
259 0x100000000000, // ShadowBase
260 0x380000000000, // OriginBase
263 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
264 &Linux_I386_MemoryMapParams,
265 &Linux_X86_64_MemoryMapParams,
268 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
270 &Linux_MIPS64_MemoryMapParams,
273 static const PlatformMemoryMapParams Linux_PowerPC_MemoryMapParams = {
275 &Linux_PowerPC64_MemoryMapParams,
278 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
279 &FreeBSD_I386_MemoryMapParams,
280 &FreeBSD_X86_64_MemoryMapParams,
283 /// \brief An instrumentation pass implementing detection of uninitialized
286 /// MemorySanitizer: instrument the code in module to find
287 /// uninitialized reads.
288 class MemorySanitizer : public FunctionPass {
290 MemorySanitizer(int TrackOrigins = 0)
292 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
293 WarningFn(nullptr) {}
294 const char *getPassName() const override { return "MemorySanitizer"; }
295 bool runOnFunction(Function &F) override;
296 bool doInitialization(Module &M) override;
297 static char ID; // Pass identification, replacement for typeid.
300 void initializeCallbacks(Module &M);
302 /// \brief Track origins (allocation points) of uninitialized values.
308 /// \brief Thread-local shadow storage for function parameters.
309 GlobalVariable *ParamTLS;
310 /// \brief Thread-local origin storage for function parameters.
311 GlobalVariable *ParamOriginTLS;
312 /// \brief Thread-local shadow storage for function return value.
313 GlobalVariable *RetvalTLS;
314 /// \brief Thread-local origin storage for function return value.
315 GlobalVariable *RetvalOriginTLS;
316 /// \brief Thread-local shadow storage for in-register va_arg function
317 /// parameters (x86_64-specific).
318 GlobalVariable *VAArgTLS;
319 /// \brief Thread-local shadow storage for va_arg overflow area
320 /// (x86_64-specific).
321 GlobalVariable *VAArgOverflowSizeTLS;
322 /// \brief Thread-local space used to pass origin value to the UMR reporting
324 GlobalVariable *OriginTLS;
326 /// \brief The run-time callback to print a warning.
328 // These arrays are indexed by log2(AccessSize).
329 Value *MaybeWarningFn[kNumberOfAccessSizes];
330 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
332 /// \brief Run-time helper that generates a new origin value for a stack
334 Value *MsanSetAllocaOrigin4Fn;
335 /// \brief Run-time helper that poisons stack on function entry.
336 Value *MsanPoisonStackFn;
337 /// \brief Run-time helper that records a store (or any event) of an
338 /// uninitialized value and returns an updated origin id encoding this info.
339 Value *MsanChainOriginFn;
340 /// \brief MSan runtime replacements for memmove, memcpy and memset.
341 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
343 /// \brief Memory map parameters used in application-to-shadow calculation.
344 const MemoryMapParams *MapParams;
346 MDNode *ColdCallWeights;
347 /// \brief Branch weights for origin store.
348 MDNode *OriginStoreWeights;
349 /// \brief An empty volatile inline asm that prevents callback merge.
351 Function *MsanCtorFunction;
353 friend struct MemorySanitizerVisitor;
354 friend struct VarArgAMD64Helper;
355 friend struct VarArgMIPS64Helper;
359 char MemorySanitizer::ID = 0;
360 INITIALIZE_PASS(MemorySanitizer, "msan",
361 "MemorySanitizer: detects uninitialized reads.",
364 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
365 return new MemorySanitizer(TrackOrigins);
368 /// \brief Create a non-const global initialized with the given string.
370 /// Creates a writable global for Str so that we can pass it to the
371 /// run-time lib. Runtime uses first 4 bytes of the string to store the
372 /// frame ID, so the string needs to be mutable.
373 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
375 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
376 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
377 GlobalValue::PrivateLinkage, StrConst, "");
381 /// \brief Insert extern declaration of runtime-provided functions and globals.
382 void MemorySanitizer::initializeCallbacks(Module &M) {
383 // Only do this once.
388 // Create the callback.
389 // FIXME: this function should have "Cold" calling conv,
390 // which is not yet implemented.
391 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
392 : "__msan_warning_noreturn";
393 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), nullptr);
395 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
397 unsigned AccessSize = 1 << AccessSizeIndex;
398 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
399 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
400 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
401 IRB.getInt32Ty(), nullptr);
403 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
404 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
405 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
406 IRB.getInt8PtrTy(), IRB.getInt32Ty(), nullptr);
409 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
410 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
411 IRB.getInt8PtrTy(), IntptrTy, nullptr);
413 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
414 IRB.getInt8PtrTy(), IntptrTy, nullptr);
415 MsanChainOriginFn = M.getOrInsertFunction(
416 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), nullptr);
417 MemmoveFn = M.getOrInsertFunction(
418 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
419 IRB.getInt8PtrTy(), IntptrTy, nullptr);
420 MemcpyFn = M.getOrInsertFunction(
421 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
423 MemsetFn = M.getOrInsertFunction(
424 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
428 RetvalTLS = new GlobalVariable(
429 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
430 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
431 GlobalVariable::InitialExecTLSModel);
432 RetvalOriginTLS = new GlobalVariable(
433 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
434 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
436 ParamTLS = new GlobalVariable(
437 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
438 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
439 GlobalVariable::InitialExecTLSModel);
440 ParamOriginTLS = new GlobalVariable(
441 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
442 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
443 nullptr, GlobalVariable::InitialExecTLSModel);
445 VAArgTLS = new GlobalVariable(
446 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
447 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
448 GlobalVariable::InitialExecTLSModel);
449 VAArgOverflowSizeTLS = new GlobalVariable(
450 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
451 "__msan_va_arg_overflow_size_tls", nullptr,
452 GlobalVariable::InitialExecTLSModel);
453 OriginTLS = new GlobalVariable(
454 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
455 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
457 // We insert an empty inline asm after __msan_report* to avoid callback merge.
458 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
459 StringRef(""), StringRef(""),
460 /*hasSideEffects=*/true);
463 /// \brief Module-level initialization.
465 /// inserts a call to __msan_init to the module's constructor list.
466 bool MemorySanitizer::doInitialization(Module &M) {
467 auto &DL = M.getDataLayout();
469 Triple TargetTriple(M.getTargetTriple());
470 switch (TargetTriple.getOS()) {
471 case Triple::FreeBSD:
472 switch (TargetTriple.getArch()) {
474 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
477 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
480 report_fatal_error("unsupported architecture");
484 switch (TargetTriple.getArch()) {
486 MapParams = Linux_X86_MemoryMapParams.bits64;
489 MapParams = Linux_X86_MemoryMapParams.bits32;
492 case Triple::mips64el:
493 MapParams = Linux_MIPS_MemoryMapParams.bits64;
496 case Triple::ppc64le:
497 MapParams = Linux_PowerPC_MemoryMapParams.bits64;
500 report_fatal_error("unsupported architecture");
504 report_fatal_error("unsupported operating system");
507 C = &(M.getContext());
509 IntptrTy = IRB.getIntPtrTy(DL);
510 OriginTy = IRB.getInt32Ty();
512 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
513 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
515 std::tie(MsanCtorFunction, std::ignore) =
516 createSanitizerCtorAndInitFunctions(M, kMsanModuleCtorName, kMsanInitName,
520 appendToGlobalCtors(M, MsanCtorFunction, 0);
523 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
524 IRB.getInt32(TrackOrigins), "__msan_track_origins");
527 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
528 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
535 /// \brief A helper class that handles instrumentation of VarArg
536 /// functions on a particular platform.
538 /// Implementations are expected to insert the instrumentation
539 /// necessary to propagate argument shadow through VarArg function
540 /// calls. Visit* methods are called during an InstVisitor pass over
541 /// the function, and should avoid creating new basic blocks. A new
542 /// instance of this class is created for each instrumented function.
543 struct VarArgHelper {
544 /// \brief Visit a CallSite.
545 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
547 /// \brief Visit a va_start call.
548 virtual void visitVAStartInst(VAStartInst &I) = 0;
550 /// \brief Visit a va_copy call.
551 virtual void visitVACopyInst(VACopyInst &I) = 0;
553 /// \brief Finalize function instrumentation.
555 /// This method is called after visiting all interesting (see above)
556 /// instructions in a function.
557 virtual void finalizeInstrumentation() = 0;
559 virtual ~VarArgHelper() {}
562 struct MemorySanitizerVisitor;
565 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
566 MemorySanitizerVisitor &Visitor);
568 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
569 if (TypeSize <= 8) return 0;
570 return Log2_32_Ceil(TypeSize / 8);
573 /// This class does all the work for a given function. Store and Load
574 /// instructions store and load corresponding shadow and origin
575 /// values. Most instructions propagate shadow from arguments to their
576 /// return values. Certain instructions (most importantly, BranchInst)
577 /// test their argument shadow and print reports (with a runtime call) if it's
579 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
582 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
583 ValueMap<Value*, Value*> ShadowMap, OriginMap;
584 std::unique_ptr<VarArgHelper> VAHelper;
586 // The following flags disable parts of MSan instrumentation based on
587 // blacklist contents and command-line options.
589 bool PropagateShadow;
592 bool CheckReturnValue;
594 struct ShadowOriginAndInsertPoint {
597 Instruction *OrigIns;
598 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
599 : Shadow(S), Origin(O), OrigIns(I) { }
601 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
602 SmallVector<Instruction*, 16> StoreList;
604 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
605 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
606 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
607 InsertChecks = SanitizeFunction;
608 PropagateShadow = SanitizeFunction;
609 PoisonStack = SanitizeFunction && ClPoisonStack;
610 PoisonUndef = SanitizeFunction && ClPoisonUndef;
611 // FIXME: Consider using SpecialCaseList to specify a list of functions that
612 // must always return fully initialized values. For now, we hardcode "main".
613 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
615 DEBUG(if (!InsertChecks)
616 dbgs() << "MemorySanitizer is not inserting checks into '"
617 << F.getName() << "'\n");
620 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
621 if (MS.TrackOrigins <= 1) return V;
622 return IRB.CreateCall(MS.MsanChainOriginFn, V);
625 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
626 const DataLayout &DL = F.getParent()->getDataLayout();
627 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
628 if (IntptrSize == kOriginSize) return Origin;
629 assert(IntptrSize == kOriginSize * 2);
630 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
631 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
634 /// \brief Fill memory range with the given origin value.
635 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
636 unsigned Size, unsigned Alignment) {
637 const DataLayout &DL = F.getParent()->getDataLayout();
638 unsigned IntptrAlignment = DL.getABITypeAlignment(MS.IntptrTy);
639 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
640 assert(IntptrAlignment >= kMinOriginAlignment);
641 assert(IntptrSize >= kOriginSize);
644 unsigned CurrentAlignment = Alignment;
645 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
646 Value *IntptrOrigin = originToIntptr(IRB, Origin);
647 Value *IntptrOriginPtr =
648 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
649 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
650 Value *Ptr = i ? IRB.CreateConstGEP1_32(MS.IntptrTy, IntptrOriginPtr, i)
652 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
653 Ofs += IntptrSize / kOriginSize;
654 CurrentAlignment = IntptrAlignment;
658 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
660 i ? IRB.CreateConstGEP1_32(nullptr, OriginPtr, i) : OriginPtr;
661 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
662 CurrentAlignment = kMinOriginAlignment;
666 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
667 unsigned Alignment, bool AsCall) {
668 const DataLayout &DL = F.getParent()->getDataLayout();
669 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
670 unsigned StoreSize = DL.getTypeStoreSize(Shadow->getType());
671 if (isa<StructType>(Shadow->getType())) {
672 paintOrigin(IRB, updateOrigin(Origin, IRB),
673 getOriginPtr(Addr, IRB, Alignment), StoreSize,
676 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
677 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
678 if (ConstantShadow) {
679 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
680 paintOrigin(IRB, updateOrigin(Origin, IRB),
681 getOriginPtr(Addr, IRB, Alignment), StoreSize,
686 unsigned TypeSizeInBits =
687 DL.getTypeSizeInBits(ConvertedShadow->getType());
688 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
689 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
690 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
691 Value *ConvertedShadow2 = IRB.CreateZExt(
692 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
693 IRB.CreateCall(Fn, {ConvertedShadow2,
694 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
697 Value *Cmp = IRB.CreateICmpNE(
698 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
699 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
700 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
701 IRBuilder<> IRBNew(CheckTerm);
702 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
703 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
709 void materializeStores(bool InstrumentWithCalls) {
710 for (auto Inst : StoreList) {
711 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
713 IRBuilder<> IRB(&SI);
714 Value *Val = SI.getValueOperand();
715 Value *Addr = SI.getPointerOperand();
716 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
717 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
720 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
721 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
724 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
726 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
728 if (MS.TrackOrigins && !SI.isAtomic())
729 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
730 InstrumentWithCalls);
734 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
736 IRBuilder<> IRB(OrigIns);
737 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
738 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
739 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
741 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
742 if (ConstantShadow) {
743 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
744 if (MS.TrackOrigins) {
745 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
748 IRB.CreateCall(MS.WarningFn, {});
749 IRB.CreateCall(MS.EmptyAsm, {});
750 // FIXME: Insert UnreachableInst if !ClKeepGoing?
751 // This may invalidate some of the following checks and needs to be done
757 const DataLayout &DL = OrigIns->getModule()->getDataLayout();
759 unsigned TypeSizeInBits = DL.getTypeSizeInBits(ConvertedShadow->getType());
760 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
761 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
762 Value *Fn = MS.MaybeWarningFn[SizeIndex];
763 Value *ConvertedShadow2 =
764 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
765 IRB.CreateCall(Fn, {ConvertedShadow2, MS.TrackOrigins && Origin
767 : (Value *)IRB.getInt32(0)});
769 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
770 getCleanShadow(ConvertedShadow), "_mscmp");
771 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
773 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
775 IRB.SetInsertPoint(CheckTerm);
776 if (MS.TrackOrigins) {
777 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
780 IRB.CreateCall(MS.WarningFn, {});
781 IRB.CreateCall(MS.EmptyAsm, {});
782 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
786 void materializeChecks(bool InstrumentWithCalls) {
787 for (const auto &ShadowData : InstrumentationList) {
788 Instruction *OrigIns = ShadowData.OrigIns;
789 Value *Shadow = ShadowData.Shadow;
790 Value *Origin = ShadowData.Origin;
791 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
793 DEBUG(dbgs() << "DONE:\n" << F);
796 /// \brief Add MemorySanitizer instrumentation to a function.
797 bool runOnFunction() {
798 MS.initializeCallbacks(*F.getParent());
800 // In the presence of unreachable blocks, we may see Phi nodes with
801 // incoming nodes from such blocks. Since InstVisitor skips unreachable
802 // blocks, such nodes will not have any shadow value associated with them.
803 // It's easier to remove unreachable blocks than deal with missing shadow.
804 removeUnreachableBlocks(F);
806 // Iterate all BBs in depth-first order and create shadow instructions
807 // for all instructions (where applicable).
808 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
809 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
813 // Finalize PHI nodes.
814 for (PHINode *PN : ShadowPHINodes) {
815 PHINode *PNS = cast<PHINode>(getShadow(PN));
816 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
817 size_t NumValues = PN->getNumIncomingValues();
818 for (size_t v = 0; v < NumValues; v++) {
819 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
820 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
824 VAHelper->finalizeInstrumentation();
826 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
827 InstrumentationList.size() + StoreList.size() >
828 (unsigned)ClInstrumentationWithCallThreshold;
830 // Delayed instrumentation of StoreInst.
831 // This may add new checks to be inserted later.
832 materializeStores(InstrumentWithCalls);
834 // Insert shadow value checks.
835 materializeChecks(InstrumentWithCalls);
840 /// \brief Compute the shadow type that corresponds to a given Value.
841 Type *getShadowTy(Value *V) {
842 return getShadowTy(V->getType());
845 /// \brief Compute the shadow type that corresponds to a given Type.
846 Type *getShadowTy(Type *OrigTy) {
847 if (!OrigTy->isSized()) {
850 // For integer type, shadow is the same as the original type.
851 // This may return weird-sized types like i1.
852 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
854 const DataLayout &DL = F.getParent()->getDataLayout();
855 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
856 uint32_t EltSize = DL.getTypeSizeInBits(VT->getElementType());
857 return VectorType::get(IntegerType::get(*MS.C, EltSize),
858 VT->getNumElements());
860 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
861 return ArrayType::get(getShadowTy(AT->getElementType()),
862 AT->getNumElements());
864 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
865 SmallVector<Type*, 4> Elements;
866 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
867 Elements.push_back(getShadowTy(ST->getElementType(i)));
868 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
869 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
872 uint32_t TypeSize = DL.getTypeSizeInBits(OrigTy);
873 return IntegerType::get(*MS.C, TypeSize);
876 /// \brief Flatten a vector type.
877 Type *getShadowTyNoVec(Type *ty) {
878 if (VectorType *vt = dyn_cast<VectorType>(ty))
879 return IntegerType::get(*MS.C, vt->getBitWidth());
883 /// \brief Convert a shadow value to it's flattened variant.
884 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
885 Type *Ty = V->getType();
886 Type *NoVecTy = getShadowTyNoVec(Ty);
887 if (Ty == NoVecTy) return V;
888 return IRB.CreateBitCast(V, NoVecTy);
891 /// \brief Compute the integer shadow offset that corresponds to a given
892 /// application address.
894 /// Offset = (Addr & ~AndMask) ^ XorMask
895 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
896 uint64_t AndMask = MS.MapParams->AndMask;
897 assert(AndMask != 0 && "AndMask shall be specified");
899 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
900 ConstantInt::get(MS.IntptrTy, ~AndMask));
902 uint64_t XorMask = MS.MapParams->XorMask;
904 OffsetLong = IRB.CreateXor(OffsetLong,
905 ConstantInt::get(MS.IntptrTy, XorMask));
909 /// \brief Compute the shadow address that corresponds to a given application
912 /// Shadow = ShadowBase + Offset
913 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
915 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
916 uint64_t ShadowBase = MS.MapParams->ShadowBase;
919 IRB.CreateAdd(ShadowLong,
920 ConstantInt::get(MS.IntptrTy, ShadowBase));
921 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
924 /// \brief Compute the origin address that corresponds to a given application
927 /// OriginAddr = (OriginBase + Offset) & ~3ULL
928 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
929 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
930 uint64_t OriginBase = MS.MapParams->OriginBase;
933 IRB.CreateAdd(OriginLong,
934 ConstantInt::get(MS.IntptrTy, OriginBase));
935 if (Alignment < kMinOriginAlignment) {
936 uint64_t Mask = kMinOriginAlignment - 1;
937 OriginLong = IRB.CreateAnd(OriginLong,
938 ConstantInt::get(MS.IntptrTy, ~Mask));
940 return IRB.CreateIntToPtr(OriginLong,
941 PointerType::get(IRB.getInt32Ty(), 0));
944 /// \brief Compute the shadow address for a given function argument.
946 /// Shadow = ParamTLS+ArgOffset.
947 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
949 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
950 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
951 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
955 /// \brief Compute the origin address for a given function argument.
956 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
958 if (!MS.TrackOrigins) return nullptr;
959 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
960 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
961 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
965 /// \brief Compute the shadow address for a retval.
966 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
967 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
968 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
972 /// \brief Compute the origin address for a retval.
973 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
974 // We keep a single origin for the entire retval. Might be too optimistic.
975 return MS.RetvalOriginTLS;
978 /// \brief Set SV to be the shadow value for V.
979 void setShadow(Value *V, Value *SV) {
980 assert(!ShadowMap.count(V) && "Values may only have one shadow");
981 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
984 /// \brief Set Origin to be the origin value for V.
985 void setOrigin(Value *V, Value *Origin) {
986 if (!MS.TrackOrigins) return;
987 assert(!OriginMap.count(V) && "Values may only have one origin");
988 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
989 OriginMap[V] = Origin;
992 /// \brief Create a clean shadow value for a given value.
994 /// Clean shadow (all zeroes) means all bits of the value are defined
996 Constant *getCleanShadow(Value *V) {
997 Type *ShadowTy = getShadowTy(V);
1000 return Constant::getNullValue(ShadowTy);
1003 /// \brief Create a dirty shadow of a given shadow type.
1004 Constant *getPoisonedShadow(Type *ShadowTy) {
1006 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
1007 return Constant::getAllOnesValue(ShadowTy);
1008 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
1009 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
1010 getPoisonedShadow(AT->getElementType()));
1011 return ConstantArray::get(AT, Vals);
1013 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
1014 SmallVector<Constant *, 4> Vals;
1015 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
1016 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
1017 return ConstantStruct::get(ST, Vals);
1019 llvm_unreachable("Unexpected shadow type");
1022 /// \brief Create a dirty shadow for a given value.
1023 Constant *getPoisonedShadow(Value *V) {
1024 Type *ShadowTy = getShadowTy(V);
1027 return getPoisonedShadow(ShadowTy);
1030 /// \brief Create a clean (zero) origin.
1031 Value *getCleanOrigin() {
1032 return Constant::getNullValue(MS.OriginTy);
1035 /// \brief Get the shadow value for a given Value.
1037 /// This function either returns the value set earlier with setShadow,
1038 /// or extracts if from ParamTLS (for function arguments).
1039 Value *getShadow(Value *V) {
1040 if (!PropagateShadow) return getCleanShadow(V);
1041 if (Instruction *I = dyn_cast<Instruction>(V)) {
1042 // For instructions the shadow is already stored in the map.
1043 Value *Shadow = ShadowMap[V];
1045 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1047 assert(Shadow && "No shadow for a value");
1051 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1052 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1053 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1057 if (Argument *A = dyn_cast<Argument>(V)) {
1058 // For arguments we compute the shadow on demand and store it in the map.
1059 Value **ShadowPtr = &ShadowMap[V];
1062 Function *F = A->getParent();
1063 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1064 unsigned ArgOffset = 0;
1065 const DataLayout &DL = F->getParent()->getDataLayout();
1066 for (auto &FArg : F->args()) {
1067 if (!FArg.getType()->isSized()) {
1068 DEBUG(dbgs() << "Arg is not sized\n");
1073 ? DL.getTypeAllocSize(FArg.getType()->getPointerElementType())
1074 : DL.getTypeAllocSize(FArg.getType());
1076 bool Overflow = ArgOffset + Size > kParamTLSSize;
1077 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1078 if (FArg.hasByValAttr()) {
1079 // ByVal pointer itself has clean shadow. We copy the actual
1080 // argument shadow to the underlying memory.
1081 // Figure out maximal valid memcpy alignment.
1082 unsigned ArgAlign = FArg.getParamAlignment();
1083 if (ArgAlign == 0) {
1084 Type *EltType = A->getType()->getPointerElementType();
1085 ArgAlign = DL.getABITypeAlignment(EltType);
1088 // ParamTLS overflow.
1089 EntryIRB.CreateMemSet(
1090 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1091 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1093 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1094 Value *Cpy = EntryIRB.CreateMemCpy(
1095 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1097 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1100 *ShadowPtr = getCleanShadow(V);
1103 // ParamTLS overflow.
1104 *ShadowPtr = getCleanShadow(V);
1107 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1110 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1111 **ShadowPtr << "\n");
1112 if (MS.TrackOrigins && !Overflow) {
1114 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1115 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1117 setOrigin(A, getCleanOrigin());
1120 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1122 assert(*ShadowPtr && "Could not find shadow for an argument");
1125 // For everything else the shadow is zero.
1126 return getCleanShadow(V);
1129 /// \brief Get the shadow for i-th argument of the instruction I.
1130 Value *getShadow(Instruction *I, int i) {
1131 return getShadow(I->getOperand(i));
1134 /// \brief Get the origin for a value.
1135 Value *getOrigin(Value *V) {
1136 if (!MS.TrackOrigins) return nullptr;
1137 if (!PropagateShadow) return getCleanOrigin();
1138 if (isa<Constant>(V)) return getCleanOrigin();
1139 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1140 "Unexpected value type in getOrigin()");
1141 Value *Origin = OriginMap[V];
1142 assert(Origin && "Missing origin");
1146 /// \brief Get the origin for i-th argument of the instruction I.
1147 Value *getOrigin(Instruction *I, int i) {
1148 return getOrigin(I->getOperand(i));
1151 /// \brief Remember the place where a shadow check should be inserted.
1153 /// This location will be later instrumented with a check that will print a
1154 /// UMR warning in runtime if the shadow value is not 0.
1155 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1157 if (!InsertChecks) return;
1159 Type *ShadowTy = Shadow->getType();
1160 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1161 "Can only insert checks for integer and vector shadow types");
1163 InstrumentationList.push_back(
1164 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1167 /// \brief Remember the place where a shadow check should be inserted.
1169 /// This location will be later instrumented with a check that will print a
1170 /// UMR warning in runtime if the value is not fully defined.
1171 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1173 Value *Shadow, *Origin;
1174 if (ClCheckConstantShadow) {
1175 Shadow = getShadow(Val);
1176 if (!Shadow) return;
1177 Origin = getOrigin(Val);
1179 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1180 if (!Shadow) return;
1181 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1183 insertShadowCheck(Shadow, Origin, OrigIns);
1186 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1195 case AcquireRelease:
1196 return AcquireRelease;
1197 case SequentiallyConsistent:
1198 return SequentiallyConsistent;
1200 llvm_unreachable("Unknown ordering");
1203 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1212 case AcquireRelease:
1213 return AcquireRelease;
1214 case SequentiallyConsistent:
1215 return SequentiallyConsistent;
1217 llvm_unreachable("Unknown ordering");
1220 // ------------------- Visitors.
1222 /// \brief Instrument LoadInst
1224 /// Loads the corresponding shadow and (optionally) origin.
1225 /// Optionally, checks that the load address is fully defined.
1226 void visitLoadInst(LoadInst &I) {
1227 assert(I.getType()->isSized() && "Load type must have size");
1228 IRBuilder<> IRB(I.getNextNode());
1229 Type *ShadowTy = getShadowTy(&I);
1230 Value *Addr = I.getPointerOperand();
1231 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1232 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1234 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1236 setShadow(&I, getCleanShadow(&I));
1239 if (ClCheckAccessAddress)
1240 insertShadowCheck(I.getPointerOperand(), &I);
1243 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1245 if (MS.TrackOrigins) {
1246 if (PropagateShadow) {
1247 unsigned Alignment = I.getAlignment();
1248 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1249 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1252 setOrigin(&I, getCleanOrigin());
1257 /// \brief Instrument StoreInst
1259 /// Stores the corresponding shadow and (optionally) origin.
1260 /// Optionally, checks that the store address is fully defined.
1261 void visitStoreInst(StoreInst &I) {
1262 StoreList.push_back(&I);
1265 void handleCASOrRMW(Instruction &I) {
1266 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1268 IRBuilder<> IRB(&I);
1269 Value *Addr = I.getOperand(0);
1270 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1272 if (ClCheckAccessAddress)
1273 insertShadowCheck(Addr, &I);
1275 // Only test the conditional argument of cmpxchg instruction.
1276 // The other argument can potentially be uninitialized, but we can not
1277 // detect this situation reliably without possible false positives.
1278 if (isa<AtomicCmpXchgInst>(I))
1279 insertShadowCheck(I.getOperand(1), &I);
1281 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1283 setShadow(&I, getCleanShadow(&I));
1284 setOrigin(&I, getCleanOrigin());
1287 void visitAtomicRMWInst(AtomicRMWInst &I) {
1289 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1292 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1294 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1297 // Vector manipulation.
1298 void visitExtractElementInst(ExtractElementInst &I) {
1299 insertShadowCheck(I.getOperand(1), &I);
1300 IRBuilder<> IRB(&I);
1301 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1303 setOrigin(&I, getOrigin(&I, 0));
1306 void visitInsertElementInst(InsertElementInst &I) {
1307 insertShadowCheck(I.getOperand(2), &I);
1308 IRBuilder<> IRB(&I);
1309 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1310 I.getOperand(2), "_msprop"));
1311 setOriginForNaryOp(I);
1314 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1315 insertShadowCheck(I.getOperand(2), &I);
1316 IRBuilder<> IRB(&I);
1317 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1318 I.getOperand(2), "_msprop"));
1319 setOriginForNaryOp(I);
1323 void visitSExtInst(SExtInst &I) {
1324 IRBuilder<> IRB(&I);
1325 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1326 setOrigin(&I, getOrigin(&I, 0));
1329 void visitZExtInst(ZExtInst &I) {
1330 IRBuilder<> IRB(&I);
1331 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1332 setOrigin(&I, getOrigin(&I, 0));
1335 void visitTruncInst(TruncInst &I) {
1336 IRBuilder<> IRB(&I);
1337 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1338 setOrigin(&I, getOrigin(&I, 0));
1341 void visitBitCastInst(BitCastInst &I) {
1342 IRBuilder<> IRB(&I);
1343 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1344 setOrigin(&I, getOrigin(&I, 0));
1347 void visitPtrToIntInst(PtrToIntInst &I) {
1348 IRBuilder<> IRB(&I);
1349 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1350 "_msprop_ptrtoint"));
1351 setOrigin(&I, getOrigin(&I, 0));
1354 void visitIntToPtrInst(IntToPtrInst &I) {
1355 IRBuilder<> IRB(&I);
1356 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1357 "_msprop_inttoptr"));
1358 setOrigin(&I, getOrigin(&I, 0));
1361 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1362 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1363 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1364 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1365 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1366 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1368 /// \brief Propagate shadow for bitwise AND.
1370 /// This code is exact, i.e. if, for example, a bit in the left argument
1371 /// is defined and 0, then neither the value not definedness of the
1372 /// corresponding bit in B don't affect the resulting shadow.
1373 void visitAnd(BinaryOperator &I) {
1374 IRBuilder<> IRB(&I);
1375 // "And" of 0 and a poisoned value results in unpoisoned value.
1376 // 1&1 => 1; 0&1 => 0; p&1 => p;
1377 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1378 // 1&p => p; 0&p => 0; p&p => p;
1379 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1380 Value *S1 = getShadow(&I, 0);
1381 Value *S2 = getShadow(&I, 1);
1382 Value *V1 = I.getOperand(0);
1383 Value *V2 = I.getOperand(1);
1384 if (V1->getType() != S1->getType()) {
1385 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1386 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1388 Value *S1S2 = IRB.CreateAnd(S1, S2);
1389 Value *V1S2 = IRB.CreateAnd(V1, S2);
1390 Value *S1V2 = IRB.CreateAnd(S1, V2);
1391 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1392 setOriginForNaryOp(I);
1395 void visitOr(BinaryOperator &I) {
1396 IRBuilder<> IRB(&I);
1397 // "Or" of 1 and a poisoned value results in unpoisoned value.
1398 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1399 // 1|0 => 1; 0|0 => 0; p|0 => p;
1400 // 1|p => 1; 0|p => p; p|p => p;
1401 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1402 Value *S1 = getShadow(&I, 0);
1403 Value *S2 = getShadow(&I, 1);
1404 Value *V1 = IRB.CreateNot(I.getOperand(0));
1405 Value *V2 = IRB.CreateNot(I.getOperand(1));
1406 if (V1->getType() != S1->getType()) {
1407 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1408 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1410 Value *S1S2 = IRB.CreateAnd(S1, S2);
1411 Value *V1S2 = IRB.CreateAnd(V1, S2);
1412 Value *S1V2 = IRB.CreateAnd(S1, V2);
1413 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1414 setOriginForNaryOp(I);
1417 /// \brief Default propagation of shadow and/or origin.
1419 /// This class implements the general case of shadow propagation, used in all
1420 /// cases where we don't know and/or don't care about what the operation
1421 /// actually does. It converts all input shadow values to a common type
1422 /// (extending or truncating as necessary), and bitwise OR's them.
1424 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1425 /// fully initialized), and less prone to false positives.
1427 /// This class also implements the general case of origin propagation. For a
1428 /// Nary operation, result origin is set to the origin of an argument that is
1429 /// not entirely initialized. If there is more than one such arguments, the
1430 /// rightmost of them is picked. It does not matter which one is picked if all
1431 /// arguments are initialized.
1432 template <bool CombineShadow>
1437 MemorySanitizerVisitor *MSV;
1440 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1441 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1443 /// \brief Add a pair of shadow and origin values to the mix.
1444 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1445 if (CombineShadow) {
1450 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1451 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1455 if (MSV->MS.TrackOrigins) {
1460 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1461 // No point in adding something that might result in 0 origin value.
1462 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1463 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1465 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1466 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1473 /// \brief Add an application value to the mix.
1474 Combiner &Add(Value *V) {
1475 Value *OpShadow = MSV->getShadow(V);
1476 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1477 return Add(OpShadow, OpOrigin);
1480 /// \brief Set the current combined values as the given instruction's shadow
1482 void Done(Instruction *I) {
1483 if (CombineShadow) {
1485 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1486 MSV->setShadow(I, Shadow);
1488 if (MSV->MS.TrackOrigins) {
1490 MSV->setOrigin(I, Origin);
1495 typedef Combiner<true> ShadowAndOriginCombiner;
1496 typedef Combiner<false> OriginCombiner;
1498 /// \brief Propagate origin for arbitrary operation.
1499 void setOriginForNaryOp(Instruction &I) {
1500 if (!MS.TrackOrigins) return;
1501 IRBuilder<> IRB(&I);
1502 OriginCombiner OC(this, IRB);
1503 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1508 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1509 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1510 "Vector of pointers is not a valid shadow type");
1511 return Ty->isVectorTy() ?
1512 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1513 Ty->getPrimitiveSizeInBits();
1516 /// \brief Cast between two shadow types, extending or truncating as
1518 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1519 bool Signed = false) {
1520 Type *srcTy = V->getType();
1521 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1522 return IRB.CreateIntCast(V, dstTy, Signed);
1523 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1524 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1525 return IRB.CreateIntCast(V, dstTy, Signed);
1526 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1527 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1528 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1530 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1531 return IRB.CreateBitCast(V2, dstTy);
1532 // TODO: handle struct types.
1535 /// \brief Cast an application value to the type of its own shadow.
1536 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1537 Type *ShadowTy = getShadowTy(V);
1538 if (V->getType() == ShadowTy)
1540 if (V->getType()->isPtrOrPtrVectorTy())
1541 return IRB.CreatePtrToInt(V, ShadowTy);
1543 return IRB.CreateBitCast(V, ShadowTy);
1546 /// \brief Propagate shadow for arbitrary operation.
1547 void handleShadowOr(Instruction &I) {
1548 IRBuilder<> IRB(&I);
1549 ShadowAndOriginCombiner SC(this, IRB);
1550 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1555 // \brief Handle multiplication by constant.
1557 // Handle a special case of multiplication by constant that may have one or
1558 // more zeros in the lower bits. This makes corresponding number of lower bits
1559 // of the result zero as well. We model it by shifting the other operand
1560 // shadow left by the required number of bits. Effectively, we transform
1561 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1562 // We use multiplication by 2**N instead of shift to cover the case of
1563 // multiplication by 0, which may occur in some elements of a vector operand.
1564 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1566 Constant *ShadowMul;
1567 Type *Ty = ConstArg->getType();
1568 if (Ty->isVectorTy()) {
1569 unsigned NumElements = Ty->getVectorNumElements();
1570 Type *EltTy = Ty->getSequentialElementType();
1571 SmallVector<Constant *, 16> Elements;
1572 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1574 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1575 APInt V = Elt->getValue();
1576 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1577 Elements.push_back(ConstantInt::get(EltTy, V2));
1579 ShadowMul = ConstantVector::get(Elements);
1581 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1582 APInt V = Elt->getValue();
1583 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1584 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1587 IRBuilder<> IRB(&I);
1589 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1590 setOrigin(&I, getOrigin(OtherArg));
1593 void visitMul(BinaryOperator &I) {
1594 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1595 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1596 if (constOp0 && !constOp1)
1597 handleMulByConstant(I, constOp0, I.getOperand(1));
1598 else if (constOp1 && !constOp0)
1599 handleMulByConstant(I, constOp1, I.getOperand(0));
1604 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1605 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1606 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1607 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1608 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1609 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1611 void handleDiv(Instruction &I) {
1612 IRBuilder<> IRB(&I);
1613 // Strict on the second argument.
1614 insertShadowCheck(I.getOperand(1), &I);
1615 setShadow(&I, getShadow(&I, 0));
1616 setOrigin(&I, getOrigin(&I, 0));
1619 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1620 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1621 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1622 void visitURem(BinaryOperator &I) { handleDiv(I); }
1623 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1624 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1626 /// \brief Instrument == and != comparisons.
1628 /// Sometimes the comparison result is known even if some of the bits of the
1629 /// arguments are not.
1630 void handleEqualityComparison(ICmpInst &I) {
1631 IRBuilder<> IRB(&I);
1632 Value *A = I.getOperand(0);
1633 Value *B = I.getOperand(1);
1634 Value *Sa = getShadow(A);
1635 Value *Sb = getShadow(B);
1637 // Get rid of pointers and vectors of pointers.
1638 // For ints (and vectors of ints), types of A and Sa match,
1639 // and this is a no-op.
1640 A = IRB.CreatePointerCast(A, Sa->getType());
1641 B = IRB.CreatePointerCast(B, Sb->getType());
1643 // A == B <==> (C = A^B) == 0
1644 // A != B <==> (C = A^B) != 0
1646 Value *C = IRB.CreateXor(A, B);
1647 Value *Sc = IRB.CreateOr(Sa, Sb);
1648 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1649 // Result is defined if one of the following is true
1650 // * there is a defined 1 bit in C
1651 // * C is fully defined
1652 // Si = !(C & ~Sc) && Sc
1653 Value *Zero = Constant::getNullValue(Sc->getType());
1654 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1656 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1658 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1659 Si->setName("_msprop_icmp");
1661 setOriginForNaryOp(I);
1664 /// \brief Build the lowest possible value of V, taking into account V's
1665 /// uninitialized bits.
1666 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1669 // Split shadow into sign bit and other bits.
1670 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1671 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1672 // Maximise the undefined shadow bit, minimize other undefined bits.
1674 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1676 // Minimize undefined bits.
1677 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1681 /// \brief Build the highest possible value of V, taking into account V's
1682 /// uninitialized bits.
1683 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1686 // Split shadow into sign bit and other bits.
1687 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1688 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1689 // Minimise the undefined shadow bit, maximise other undefined bits.
1691 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1693 // Maximize undefined bits.
1694 return IRB.CreateOr(A, Sa);
1698 /// \brief Instrument relational comparisons.
1700 /// This function does exact shadow propagation for all relational
1701 /// comparisons of integers, pointers and vectors of those.
1702 /// FIXME: output seems suboptimal when one of the operands is a constant
1703 void handleRelationalComparisonExact(ICmpInst &I) {
1704 IRBuilder<> IRB(&I);
1705 Value *A = I.getOperand(0);
1706 Value *B = I.getOperand(1);
1707 Value *Sa = getShadow(A);
1708 Value *Sb = getShadow(B);
1710 // Get rid of pointers and vectors of pointers.
1711 // For ints (and vectors of ints), types of A and Sa match,
1712 // and this is a no-op.
1713 A = IRB.CreatePointerCast(A, Sa->getType());
1714 B = IRB.CreatePointerCast(B, Sb->getType());
1716 // Let [a0, a1] be the interval of possible values of A, taking into account
1717 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1718 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1719 bool IsSigned = I.isSigned();
1720 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1721 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1722 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1723 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1724 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1725 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1726 Value *Si = IRB.CreateXor(S1, S2);
1728 setOriginForNaryOp(I);
1731 /// \brief Instrument signed relational comparisons.
1733 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1734 /// propagating the highest bit of the shadow. Everything else is delegated
1735 /// to handleShadowOr().
1736 void handleSignedRelationalComparison(ICmpInst &I) {
1737 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1738 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1739 Value* op = nullptr;
1740 CmpInst::Predicate pre = I.getPredicate();
1741 if (constOp0 && constOp0->isNullValue() &&
1742 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1743 op = I.getOperand(1);
1744 } else if (constOp1 && constOp1->isNullValue() &&
1745 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1746 op = I.getOperand(0);
1749 IRBuilder<> IRB(&I);
1751 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1752 setShadow(&I, Shadow);
1753 setOrigin(&I, getOrigin(op));
1759 void visitICmpInst(ICmpInst &I) {
1760 if (!ClHandleICmp) {
1764 if (I.isEquality()) {
1765 handleEqualityComparison(I);
1769 assert(I.isRelational());
1770 if (ClHandleICmpExact) {
1771 handleRelationalComparisonExact(I);
1775 handleSignedRelationalComparison(I);
1779 assert(I.isUnsigned());
1780 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1781 handleRelationalComparisonExact(I);
1788 void visitFCmpInst(FCmpInst &I) {
1792 void handleShift(BinaryOperator &I) {
1793 IRBuilder<> IRB(&I);
1794 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1795 // Otherwise perform the same shift on S1.
1796 Value *S1 = getShadow(&I, 0);
1797 Value *S2 = getShadow(&I, 1);
1798 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1800 Value *V2 = I.getOperand(1);
1801 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1802 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1803 setOriginForNaryOp(I);
1806 void visitShl(BinaryOperator &I) { handleShift(I); }
1807 void visitAShr(BinaryOperator &I) { handleShift(I); }
1808 void visitLShr(BinaryOperator &I) { handleShift(I); }
1810 /// \brief Instrument llvm.memmove
1812 /// At this point we don't know if llvm.memmove will be inlined or not.
1813 /// If we don't instrument it and it gets inlined,
1814 /// our interceptor will not kick in and we will lose the memmove.
1815 /// If we instrument the call here, but it does not get inlined,
1816 /// we will memove the shadow twice: which is bad in case
1817 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1819 /// Similar situation exists for memcpy and memset.
1820 void visitMemMoveInst(MemMoveInst &I) {
1821 IRBuilder<> IRB(&I);
1824 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1825 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1826 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1827 I.eraseFromParent();
1830 // Similar to memmove: avoid copying shadow twice.
1831 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1832 // FIXME: consider doing manual inline for small constant sizes and proper
1834 void visitMemCpyInst(MemCpyInst &I) {
1835 IRBuilder<> IRB(&I);
1838 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1839 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1840 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1841 I.eraseFromParent();
1845 void visitMemSetInst(MemSetInst &I) {
1846 IRBuilder<> IRB(&I);
1849 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1850 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1851 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1852 I.eraseFromParent();
1855 void visitVAStartInst(VAStartInst &I) {
1856 VAHelper->visitVAStartInst(I);
1859 void visitVACopyInst(VACopyInst &I) {
1860 VAHelper->visitVACopyInst(I);
1863 enum IntrinsicKind {
1864 IK_DoesNotAccessMemory,
1869 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1870 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1871 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1872 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1873 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1874 const int UnknownModRefBehavior = IK_WritesMemory;
1875 #define GET_INTRINSIC_MODREF_BEHAVIOR
1876 #define ModRefBehavior IntrinsicKind
1877 #include "llvm/IR/Intrinsics.gen"
1878 #undef ModRefBehavior
1879 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1882 /// \brief Handle vector store-like intrinsics.
1884 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1885 /// has 1 pointer argument and 1 vector argument, returns void.
1886 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1887 IRBuilder<> IRB(&I);
1888 Value* Addr = I.getArgOperand(0);
1889 Value *Shadow = getShadow(&I, 1);
1890 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1892 // We don't know the pointer alignment (could be unaligned SSE store!).
1893 // Have to assume to worst case.
1894 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1896 if (ClCheckAccessAddress)
1897 insertShadowCheck(Addr, &I);
1899 // FIXME: use ClStoreCleanOrigin
1900 // FIXME: factor out common code from materializeStores
1901 if (MS.TrackOrigins)
1902 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1906 /// \brief Handle vector load-like intrinsics.
1908 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1909 /// has 1 pointer argument, returns a vector.
1910 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1911 IRBuilder<> IRB(&I);
1912 Value *Addr = I.getArgOperand(0);
1914 Type *ShadowTy = getShadowTy(&I);
1915 if (PropagateShadow) {
1916 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1917 // We don't know the pointer alignment (could be unaligned SSE load!).
1918 // Have to assume to worst case.
1919 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1921 setShadow(&I, getCleanShadow(&I));
1924 if (ClCheckAccessAddress)
1925 insertShadowCheck(Addr, &I);
1927 if (MS.TrackOrigins) {
1928 if (PropagateShadow)
1929 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1931 setOrigin(&I, getCleanOrigin());
1936 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1938 /// Instrument intrinsics with any number of arguments of the same type,
1939 /// equal to the return type. The type should be simple (no aggregates or
1940 /// pointers; vectors are fine).
1941 /// Caller guarantees that this intrinsic does not access memory.
1942 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1943 Type *RetTy = I.getType();
1944 if (!(RetTy->isIntOrIntVectorTy() ||
1945 RetTy->isFPOrFPVectorTy() ||
1946 RetTy->isX86_MMXTy()))
1949 unsigned NumArgOperands = I.getNumArgOperands();
1951 for (unsigned i = 0; i < NumArgOperands; ++i) {
1952 Type *Ty = I.getArgOperand(i)->getType();
1957 IRBuilder<> IRB(&I);
1958 ShadowAndOriginCombiner SC(this, IRB);
1959 for (unsigned i = 0; i < NumArgOperands; ++i)
1960 SC.Add(I.getArgOperand(i));
1966 /// \brief Heuristically instrument unknown intrinsics.
1968 /// The main purpose of this code is to do something reasonable with all
1969 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1970 /// We recognize several classes of intrinsics by their argument types and
1971 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1972 /// sure that we know what the intrinsic does.
1974 /// We special-case intrinsics where this approach fails. See llvm.bswap
1975 /// handling as an example of that.
1976 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1977 unsigned NumArgOperands = I.getNumArgOperands();
1978 if (NumArgOperands == 0)
1981 Intrinsic::ID iid = I.getIntrinsicID();
1982 IntrinsicKind IK = getIntrinsicKind(iid);
1983 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1984 bool WritesMemory = IK == IK_WritesMemory;
1985 assert(!(OnlyReadsMemory && WritesMemory));
1987 if (NumArgOperands == 2 &&
1988 I.getArgOperand(0)->getType()->isPointerTy() &&
1989 I.getArgOperand(1)->getType()->isVectorTy() &&
1990 I.getType()->isVoidTy() &&
1992 // This looks like a vector store.
1993 return handleVectorStoreIntrinsic(I);
1996 if (NumArgOperands == 1 &&
1997 I.getArgOperand(0)->getType()->isPointerTy() &&
1998 I.getType()->isVectorTy() &&
2000 // This looks like a vector load.
2001 return handleVectorLoadIntrinsic(I);
2004 if (!OnlyReadsMemory && !WritesMemory)
2005 if (maybeHandleSimpleNomemIntrinsic(I))
2008 // FIXME: detect and handle SSE maskstore/maskload
2012 void handleBswap(IntrinsicInst &I) {
2013 IRBuilder<> IRB(&I);
2014 Value *Op = I.getArgOperand(0);
2015 Type *OpType = Op->getType();
2016 Function *BswapFunc = Intrinsic::getDeclaration(
2017 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
2018 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
2019 setOrigin(&I, getOrigin(Op));
2022 // \brief Instrument vector convert instrinsic.
2024 // This function instruments intrinsics like cvtsi2ss:
2025 // %Out = int_xxx_cvtyyy(%ConvertOp)
2027 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2028 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2029 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2030 // elements from \p CopyOp.
2031 // In most cases conversion involves floating-point value which may trigger a
2032 // hardware exception when not fully initialized. For this reason we require
2033 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2034 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2035 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2036 // return a fully initialized value.
2037 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2038 IRBuilder<> IRB(&I);
2039 Value *CopyOp, *ConvertOp;
2041 switch (I.getNumArgOperands()) {
2043 assert(isa<ConstantInt>(I.getArgOperand(2)) && "Invalid rounding mode");
2045 CopyOp = I.getArgOperand(0);
2046 ConvertOp = I.getArgOperand(1);
2049 ConvertOp = I.getArgOperand(0);
2053 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2056 // The first *NumUsedElements* elements of ConvertOp are converted to the
2057 // same number of output elements. The rest of the output is copied from
2058 // CopyOp, or (if not available) filled with zeroes.
2059 // Combine shadow for elements of ConvertOp that are used in this operation,
2060 // and insert a check.
2061 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2062 // int->any conversion.
2063 Value *ConvertShadow = getShadow(ConvertOp);
2064 Value *AggShadow = nullptr;
2065 if (ConvertOp->getType()->isVectorTy()) {
2066 AggShadow = IRB.CreateExtractElement(
2067 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2068 for (int i = 1; i < NumUsedElements; ++i) {
2069 Value *MoreShadow = IRB.CreateExtractElement(
2070 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2071 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2074 AggShadow = ConvertShadow;
2076 assert(AggShadow->getType()->isIntegerTy());
2077 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2079 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2082 assert(CopyOp->getType() == I.getType());
2083 assert(CopyOp->getType()->isVectorTy());
2084 Value *ResultShadow = getShadow(CopyOp);
2085 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2086 for (int i = 0; i < NumUsedElements; ++i) {
2087 ResultShadow = IRB.CreateInsertElement(
2088 ResultShadow, ConstantInt::getNullValue(EltTy),
2089 ConstantInt::get(IRB.getInt32Ty(), i));
2091 setShadow(&I, ResultShadow);
2092 setOrigin(&I, getOrigin(CopyOp));
2094 setShadow(&I, getCleanShadow(&I));
2095 setOrigin(&I, getCleanOrigin());
2099 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2100 // zeroes if it is zero, and all ones otherwise.
2101 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2102 if (S->getType()->isVectorTy())
2103 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2104 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2105 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2106 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2109 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2110 Type *T = S->getType();
2111 assert(T->isVectorTy());
2112 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2113 return IRB.CreateSExt(S2, T);
2116 // \brief Instrument vector shift instrinsic.
2118 // This function instruments intrinsics like int_x86_avx2_psll_w.
2119 // Intrinsic shifts %In by %ShiftSize bits.
2120 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2121 // size, and the rest is ignored. Behavior is defined even if shift size is
2122 // greater than register (or field) width.
2123 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2124 assert(I.getNumArgOperands() == 2);
2125 IRBuilder<> IRB(&I);
2126 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2127 // Otherwise perform the same shift on S1.
2128 Value *S1 = getShadow(&I, 0);
2129 Value *S2 = getShadow(&I, 1);
2130 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2131 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2132 Value *V1 = I.getOperand(0);
2133 Value *V2 = I.getOperand(1);
2134 Value *Shift = IRB.CreateCall(I.getCalledValue(),
2135 {IRB.CreateBitCast(S1, V1->getType()), V2});
2136 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2137 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2138 setOriginForNaryOp(I);
2141 // \brief Get an X86_MMX-sized vector type.
2142 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2143 const unsigned X86_MMXSizeInBits = 64;
2144 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2145 X86_MMXSizeInBits / EltSizeInBits);
2148 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2150 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2152 case llvm::Intrinsic::x86_sse2_packsswb_128:
2153 case llvm::Intrinsic::x86_sse2_packuswb_128:
2154 return llvm::Intrinsic::x86_sse2_packsswb_128;
2156 case llvm::Intrinsic::x86_sse2_packssdw_128:
2157 case llvm::Intrinsic::x86_sse41_packusdw:
2158 return llvm::Intrinsic::x86_sse2_packssdw_128;
2160 case llvm::Intrinsic::x86_avx2_packsswb:
2161 case llvm::Intrinsic::x86_avx2_packuswb:
2162 return llvm::Intrinsic::x86_avx2_packsswb;
2164 case llvm::Intrinsic::x86_avx2_packssdw:
2165 case llvm::Intrinsic::x86_avx2_packusdw:
2166 return llvm::Intrinsic::x86_avx2_packssdw;
2168 case llvm::Intrinsic::x86_mmx_packsswb:
2169 case llvm::Intrinsic::x86_mmx_packuswb:
2170 return llvm::Intrinsic::x86_mmx_packsswb;
2172 case llvm::Intrinsic::x86_mmx_packssdw:
2173 return llvm::Intrinsic::x86_mmx_packssdw;
2175 llvm_unreachable("unexpected intrinsic id");
2179 // \brief Instrument vector pack instrinsic.
2181 // This function instruments intrinsics like x86_mmx_packsswb, that
2182 // packs elements of 2 input vectors into half as many bits with saturation.
2183 // Shadow is propagated with the signed variant of the same intrinsic applied
2184 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2185 // EltSizeInBits is used only for x86mmx arguments.
2186 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2187 assert(I.getNumArgOperands() == 2);
2188 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2189 IRBuilder<> IRB(&I);
2190 Value *S1 = getShadow(&I, 0);
2191 Value *S2 = getShadow(&I, 1);
2192 assert(isX86_MMX || S1->getType()->isVectorTy());
2194 // SExt and ICmpNE below must apply to individual elements of input vectors.
2195 // In case of x86mmx arguments, cast them to appropriate vector types and
2197 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2199 S1 = IRB.CreateBitCast(S1, T);
2200 S2 = IRB.CreateBitCast(S2, T);
2202 Value *S1_ext = IRB.CreateSExt(
2203 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2204 Value *S2_ext = IRB.CreateSExt(
2205 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2207 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2208 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2209 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2212 Function *ShadowFn = Intrinsic::getDeclaration(
2213 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2216 IRB.CreateCall(ShadowFn, {S1_ext, S2_ext}, "_msprop_vector_pack");
2217 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2219 setOriginForNaryOp(I);
2222 // \brief Instrument sum-of-absolute-differencies intrinsic.
2223 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2224 const unsigned SignificantBitsPerResultElement = 16;
2225 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2226 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2227 unsigned ZeroBitsPerResultElement =
2228 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2230 IRBuilder<> IRB(&I);
2231 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2232 S = IRB.CreateBitCast(S, ResTy);
2233 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2235 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2236 S = IRB.CreateBitCast(S, getShadowTy(&I));
2238 setOriginForNaryOp(I);
2241 // \brief Instrument multiply-add intrinsic.
2242 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2243 unsigned EltSizeInBits = 0) {
2244 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2245 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2246 IRBuilder<> IRB(&I);
2247 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2248 S = IRB.CreateBitCast(S, ResTy);
2249 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2251 S = IRB.CreateBitCast(S, getShadowTy(&I));
2253 setOriginForNaryOp(I);
2256 void visitIntrinsicInst(IntrinsicInst &I) {
2257 switch (I.getIntrinsicID()) {
2258 case llvm::Intrinsic::bswap:
2261 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2262 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2263 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2264 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2265 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2266 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2267 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2268 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2269 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2270 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2271 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2272 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2273 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2274 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2275 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2276 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2277 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2278 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2279 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2280 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2281 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2282 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2283 case llvm::Intrinsic::x86_sse_cvtss2si64:
2284 case llvm::Intrinsic::x86_sse_cvtss2si:
2285 case llvm::Intrinsic::x86_sse_cvttss2si64:
2286 case llvm::Intrinsic::x86_sse_cvttss2si:
2287 handleVectorConvertIntrinsic(I, 1);
2289 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2290 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2291 case llvm::Intrinsic::x86_sse_cvtps2pi:
2292 case llvm::Intrinsic::x86_sse_cvttps2pi:
2293 handleVectorConvertIntrinsic(I, 2);
2295 case llvm::Intrinsic::x86_avx2_psll_w:
2296 case llvm::Intrinsic::x86_avx2_psll_d:
2297 case llvm::Intrinsic::x86_avx2_psll_q:
2298 case llvm::Intrinsic::x86_avx2_pslli_w:
2299 case llvm::Intrinsic::x86_avx2_pslli_d:
2300 case llvm::Intrinsic::x86_avx2_pslli_q:
2301 case llvm::Intrinsic::x86_avx2_psrl_w:
2302 case llvm::Intrinsic::x86_avx2_psrl_d:
2303 case llvm::Intrinsic::x86_avx2_psrl_q:
2304 case llvm::Intrinsic::x86_avx2_psra_w:
2305 case llvm::Intrinsic::x86_avx2_psra_d:
2306 case llvm::Intrinsic::x86_avx2_psrli_w:
2307 case llvm::Intrinsic::x86_avx2_psrli_d:
2308 case llvm::Intrinsic::x86_avx2_psrli_q:
2309 case llvm::Intrinsic::x86_avx2_psrai_w:
2310 case llvm::Intrinsic::x86_avx2_psrai_d:
2311 case llvm::Intrinsic::x86_sse2_psll_w:
2312 case llvm::Intrinsic::x86_sse2_psll_d:
2313 case llvm::Intrinsic::x86_sse2_psll_q:
2314 case llvm::Intrinsic::x86_sse2_pslli_w:
2315 case llvm::Intrinsic::x86_sse2_pslli_d:
2316 case llvm::Intrinsic::x86_sse2_pslli_q:
2317 case llvm::Intrinsic::x86_sse2_psrl_w:
2318 case llvm::Intrinsic::x86_sse2_psrl_d:
2319 case llvm::Intrinsic::x86_sse2_psrl_q:
2320 case llvm::Intrinsic::x86_sse2_psra_w:
2321 case llvm::Intrinsic::x86_sse2_psra_d:
2322 case llvm::Intrinsic::x86_sse2_psrli_w:
2323 case llvm::Intrinsic::x86_sse2_psrli_d:
2324 case llvm::Intrinsic::x86_sse2_psrli_q:
2325 case llvm::Intrinsic::x86_sse2_psrai_w:
2326 case llvm::Intrinsic::x86_sse2_psrai_d:
2327 case llvm::Intrinsic::x86_mmx_psll_w:
2328 case llvm::Intrinsic::x86_mmx_psll_d:
2329 case llvm::Intrinsic::x86_mmx_psll_q:
2330 case llvm::Intrinsic::x86_mmx_pslli_w:
2331 case llvm::Intrinsic::x86_mmx_pslli_d:
2332 case llvm::Intrinsic::x86_mmx_pslli_q:
2333 case llvm::Intrinsic::x86_mmx_psrl_w:
2334 case llvm::Intrinsic::x86_mmx_psrl_d:
2335 case llvm::Intrinsic::x86_mmx_psrl_q:
2336 case llvm::Intrinsic::x86_mmx_psra_w:
2337 case llvm::Intrinsic::x86_mmx_psra_d:
2338 case llvm::Intrinsic::x86_mmx_psrli_w:
2339 case llvm::Intrinsic::x86_mmx_psrli_d:
2340 case llvm::Intrinsic::x86_mmx_psrli_q:
2341 case llvm::Intrinsic::x86_mmx_psrai_w:
2342 case llvm::Intrinsic::x86_mmx_psrai_d:
2343 handleVectorShiftIntrinsic(I, /* Variable */ false);
2345 case llvm::Intrinsic::x86_avx2_psllv_d:
2346 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2347 case llvm::Intrinsic::x86_avx2_psllv_q:
2348 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2349 case llvm::Intrinsic::x86_avx2_psrlv_d:
2350 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2351 case llvm::Intrinsic::x86_avx2_psrlv_q:
2352 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2353 case llvm::Intrinsic::x86_avx2_psrav_d:
2354 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2355 handleVectorShiftIntrinsic(I, /* Variable */ true);
2358 case llvm::Intrinsic::x86_sse2_packsswb_128:
2359 case llvm::Intrinsic::x86_sse2_packssdw_128:
2360 case llvm::Intrinsic::x86_sse2_packuswb_128:
2361 case llvm::Intrinsic::x86_sse41_packusdw:
2362 case llvm::Intrinsic::x86_avx2_packsswb:
2363 case llvm::Intrinsic::x86_avx2_packssdw:
2364 case llvm::Intrinsic::x86_avx2_packuswb:
2365 case llvm::Intrinsic::x86_avx2_packusdw:
2366 handleVectorPackIntrinsic(I);
2369 case llvm::Intrinsic::x86_mmx_packsswb:
2370 case llvm::Intrinsic::x86_mmx_packuswb:
2371 handleVectorPackIntrinsic(I, 16);
2374 case llvm::Intrinsic::x86_mmx_packssdw:
2375 handleVectorPackIntrinsic(I, 32);
2378 case llvm::Intrinsic::x86_mmx_psad_bw:
2379 case llvm::Intrinsic::x86_sse2_psad_bw:
2380 case llvm::Intrinsic::x86_avx2_psad_bw:
2381 handleVectorSadIntrinsic(I);
2384 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2385 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2386 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2387 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2388 handleVectorPmaddIntrinsic(I);
2391 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2392 handleVectorPmaddIntrinsic(I, 8);
2395 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2396 handleVectorPmaddIntrinsic(I, 16);
2400 if (!handleUnknownIntrinsic(I))
2401 visitInstruction(I);
2406 void visitCallSite(CallSite CS) {
2407 Instruction &I = *CS.getInstruction();
2408 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2410 CallInst *Call = cast<CallInst>(&I);
2412 // For inline asm, do the usual thing: check argument shadow and mark all
2413 // outputs as clean. Note that any side effects of the inline asm that are
2414 // not immediately visible in its constraints are not handled.
2415 if (Call->isInlineAsm()) {
2416 visitInstruction(I);
2420 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2422 // We are going to insert code that relies on the fact that the callee
2423 // will become a non-readonly function after it is instrumented by us. To
2424 // prevent this code from being optimized out, mark that function
2425 // non-readonly in advance.
2426 if (Function *Func = Call->getCalledFunction()) {
2427 // Clear out readonly/readnone attributes.
2429 B.addAttribute(Attribute::ReadOnly)
2430 .addAttribute(Attribute::ReadNone);
2431 Func->removeAttributes(AttributeSet::FunctionIndex,
2432 AttributeSet::get(Func->getContext(),
2433 AttributeSet::FunctionIndex,
2437 IRBuilder<> IRB(&I);
2439 unsigned ArgOffset = 0;
2440 DEBUG(dbgs() << " CallSite: " << I << "\n");
2441 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2442 ArgIt != End; ++ArgIt) {
2444 unsigned i = ArgIt - CS.arg_begin();
2445 if (!A->getType()->isSized()) {
2446 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2450 Value *Store = nullptr;
2451 // Compute the Shadow for arg even if it is ByVal, because
2452 // in that case getShadow() will copy the actual arg shadow to
2453 // __msan_param_tls.
2454 Value *ArgShadow = getShadow(A);
2455 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2456 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2457 " Shadow: " << *ArgShadow << "\n");
2458 bool ArgIsInitialized = false;
2459 const DataLayout &DL = F.getParent()->getDataLayout();
2460 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2461 assert(A->getType()->isPointerTy() &&
2462 "ByVal argument is not a pointer!");
2463 Size = DL.getTypeAllocSize(A->getType()->getPointerElementType());
2464 if (ArgOffset + Size > kParamTLSSize) break;
2465 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2466 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2467 Store = IRB.CreateMemCpy(ArgShadowBase,
2468 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2471 Size = DL.getTypeAllocSize(A->getType());
2472 if (ArgOffset + Size > kParamTLSSize) break;
2473 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2474 kShadowTLSAlignment);
2475 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2476 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2478 if (MS.TrackOrigins && !ArgIsInitialized)
2479 IRB.CreateStore(getOrigin(A),
2480 getOriginPtrForArgument(A, IRB, ArgOffset));
2482 assert(Size != 0 && Store != nullptr);
2483 DEBUG(dbgs() << " Param:" << *Store << "\n");
2484 ArgOffset += RoundUpToAlignment(Size, 8);
2486 DEBUG(dbgs() << " done with call args\n");
2489 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2490 if (FT->isVarArg()) {
2491 VAHelper->visitCallSite(CS, IRB);
2494 // Now, get the shadow for the RetVal.
2495 if (!I.getType()->isSized()) return;
2496 IRBuilder<> IRBBefore(&I);
2497 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2498 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2499 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2500 Instruction *NextInsn = nullptr;
2502 NextInsn = I.getNextNode();
2504 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2505 if (!NormalDest->getSinglePredecessor()) {
2506 // FIXME: this case is tricky, so we are just conservative here.
2507 // Perhaps we need to split the edge between this BB and NormalDest,
2508 // but a naive attempt to use SplitEdge leads to a crash.
2509 setShadow(&I, getCleanShadow(&I));
2510 setOrigin(&I, getCleanOrigin());
2513 NextInsn = NormalDest->getFirstInsertionPt();
2515 "Could not find insertion point for retval shadow load");
2517 IRBuilder<> IRBAfter(NextInsn);
2518 Value *RetvalShadow =
2519 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2520 kShadowTLSAlignment, "_msret");
2521 setShadow(&I, RetvalShadow);
2522 if (MS.TrackOrigins)
2523 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2526 void visitReturnInst(ReturnInst &I) {
2527 IRBuilder<> IRB(&I);
2528 Value *RetVal = I.getReturnValue();
2529 if (!RetVal) return;
2530 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2531 if (CheckReturnValue) {
2532 insertShadowCheck(RetVal, &I);
2533 Value *Shadow = getCleanShadow(RetVal);
2534 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2536 Value *Shadow = getShadow(RetVal);
2537 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2538 // FIXME: make it conditional if ClStoreCleanOrigin==0
2539 if (MS.TrackOrigins)
2540 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2544 void visitPHINode(PHINode &I) {
2545 IRBuilder<> IRB(&I);
2546 if (!PropagateShadow) {
2547 setShadow(&I, getCleanShadow(&I));
2548 setOrigin(&I, getCleanOrigin());
2552 ShadowPHINodes.push_back(&I);
2553 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2555 if (MS.TrackOrigins)
2556 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2560 void visitAllocaInst(AllocaInst &I) {
2561 setShadow(&I, getCleanShadow(&I));
2562 setOrigin(&I, getCleanOrigin());
2563 IRBuilder<> IRB(I.getNextNode());
2564 const DataLayout &DL = F.getParent()->getDataLayout();
2565 uint64_t Size = DL.getTypeAllocSize(I.getAllocatedType());
2566 if (PoisonStack && ClPoisonStackWithCall) {
2567 IRB.CreateCall(MS.MsanPoisonStackFn,
2568 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2569 ConstantInt::get(MS.IntptrTy, Size)});
2571 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2572 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2573 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2576 if (PoisonStack && MS.TrackOrigins) {
2577 SmallString<2048> StackDescriptionStorage;
2578 raw_svector_ostream StackDescription(StackDescriptionStorage);
2579 // We create a string with a description of the stack allocation and
2580 // pass it into __msan_set_alloca_origin.
2581 // It will be printed by the run-time if stack-originated UMR is found.
2582 // The first 4 bytes of the string are set to '----' and will be replaced
2583 // by __msan_va_arg_overflow_size_tls at the first call.
2584 StackDescription << "----" << I.getName() << "@" << F.getName();
2586 createPrivateNonConstGlobalForString(*F.getParent(),
2587 StackDescription.str());
2589 IRB.CreateCall(MS.MsanSetAllocaOrigin4Fn,
2590 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2591 ConstantInt::get(MS.IntptrTy, Size),
2592 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2593 IRB.CreatePointerCast(&F, MS.IntptrTy)});
2597 void visitSelectInst(SelectInst& I) {
2598 IRBuilder<> IRB(&I);
2599 // a = select b, c, d
2600 Value *B = I.getCondition();
2601 Value *C = I.getTrueValue();
2602 Value *D = I.getFalseValue();
2603 Value *Sb = getShadow(B);
2604 Value *Sc = getShadow(C);
2605 Value *Sd = getShadow(D);
2607 // Result shadow if condition shadow is 0.
2608 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2610 if (I.getType()->isAggregateType()) {
2611 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2612 // an extra "select". This results in much more compact IR.
2613 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2614 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2616 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2617 // If Sb (condition is poisoned), look for bits in c and d that are equal
2618 // and both unpoisoned.
2619 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2621 // Cast arguments to shadow-compatible type.
2622 C = CreateAppToShadowCast(IRB, C);
2623 D = CreateAppToShadowCast(IRB, D);
2625 // Result shadow if condition shadow is 1.
2626 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2628 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2630 if (MS.TrackOrigins) {
2631 // Origins are always i32, so any vector conditions must be flattened.
2632 // FIXME: consider tracking vector origins for app vectors?
2633 if (B->getType()->isVectorTy()) {
2634 Type *FlatTy = getShadowTyNoVec(B->getType());
2635 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2636 ConstantInt::getNullValue(FlatTy));
2637 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2638 ConstantInt::getNullValue(FlatTy));
2640 // a = select b, c, d
2641 // Oa = Sb ? Ob : (b ? Oc : Od)
2643 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2644 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2645 getOrigin(I.getFalseValue()))));
2649 void visitLandingPadInst(LandingPadInst &I) {
2651 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2652 setShadow(&I, getCleanShadow(&I));
2653 setOrigin(&I, getCleanOrigin());
2656 void visitGetElementPtrInst(GetElementPtrInst &I) {
2660 void visitExtractValueInst(ExtractValueInst &I) {
2661 IRBuilder<> IRB(&I);
2662 Value *Agg = I.getAggregateOperand();
2663 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2664 Value *AggShadow = getShadow(Agg);
2665 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2666 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2667 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2668 setShadow(&I, ResShadow);
2669 setOriginForNaryOp(I);
2672 void visitInsertValueInst(InsertValueInst &I) {
2673 IRBuilder<> IRB(&I);
2674 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2675 Value *AggShadow = getShadow(I.getAggregateOperand());
2676 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2677 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2678 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2679 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2680 DEBUG(dbgs() << " Res: " << *Res << "\n");
2682 setOriginForNaryOp(I);
2685 void dumpInst(Instruction &I) {
2686 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2687 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2689 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2691 errs() << "QQQ " << I << "\n";
2694 void visitResumeInst(ResumeInst &I) {
2695 DEBUG(dbgs() << "Resume: " << I << "\n");
2696 // Nothing to do here.
2699 void visitInstruction(Instruction &I) {
2700 // Everything else: stop propagating and check for poisoned shadow.
2701 if (ClDumpStrictInstructions)
2703 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2704 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2705 insertShadowCheck(I.getOperand(i), &I);
2706 setShadow(&I, getCleanShadow(&I));
2707 setOrigin(&I, getCleanOrigin());
2711 /// \brief AMD64-specific implementation of VarArgHelper.
2712 struct VarArgAMD64Helper : public VarArgHelper {
2713 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2714 // See a comment in visitCallSite for more details.
2715 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2716 static const unsigned AMD64FpEndOffset = 176;
2719 MemorySanitizer &MS;
2720 MemorySanitizerVisitor &MSV;
2721 Value *VAArgTLSCopy;
2722 Value *VAArgOverflowSize;
2724 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2726 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2727 MemorySanitizerVisitor &MSV)
2728 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2729 VAArgOverflowSize(nullptr) {}
2731 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2733 ArgKind classifyArgument(Value* arg) {
2734 // A very rough approximation of X86_64 argument classification rules.
2735 Type *T = arg->getType();
2736 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2737 return AK_FloatingPoint;
2738 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2739 return AK_GeneralPurpose;
2740 if (T->isPointerTy())
2741 return AK_GeneralPurpose;
2745 // For VarArg functions, store the argument shadow in an ABI-specific format
2746 // that corresponds to va_list layout.
2747 // We do this because Clang lowers va_arg in the frontend, and this pass
2748 // only sees the low level code that deals with va_list internals.
2749 // A much easier alternative (provided that Clang emits va_arg instructions)
2750 // would have been to associate each live instance of va_list with a copy of
2751 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2753 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2754 unsigned GpOffset = 0;
2755 unsigned FpOffset = AMD64GpEndOffset;
2756 unsigned OverflowOffset = AMD64FpEndOffset;
2757 const DataLayout &DL = F.getParent()->getDataLayout();
2758 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2759 ArgIt != End; ++ArgIt) {
2761 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2762 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2764 // ByVal arguments always go to the overflow area.
2765 assert(A->getType()->isPointerTy());
2766 Type *RealTy = A->getType()->getPointerElementType();
2767 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
2768 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2769 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2770 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2771 ArgSize, kShadowTLSAlignment);
2773 ArgKind AK = classifyArgument(A);
2774 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2776 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2780 case AK_GeneralPurpose:
2781 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2784 case AK_FloatingPoint:
2785 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2789 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
2790 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2791 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2793 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2796 Constant *OverflowSize =
2797 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2798 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2801 /// \brief Compute the shadow address for a given va_arg.
2802 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2804 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2805 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2806 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2810 void visitVAStartInst(VAStartInst &I) override {
2811 IRBuilder<> IRB(&I);
2812 VAStartInstrumentationList.push_back(&I);
2813 Value *VAListTag = I.getArgOperand(0);
2814 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2816 // Unpoison the whole __va_list_tag.
2817 // FIXME: magic ABI constants.
2818 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2819 /* size */24, /* alignment */8, false);
2822 void visitVACopyInst(VACopyInst &I) override {
2823 IRBuilder<> IRB(&I);
2824 Value *VAListTag = I.getArgOperand(0);
2825 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2827 // Unpoison the whole __va_list_tag.
2828 // FIXME: magic ABI constants.
2829 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2830 /* size */24, /* alignment */8, false);
2833 void finalizeInstrumentation() override {
2834 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2835 "finalizeInstrumentation called twice");
2836 if (!VAStartInstrumentationList.empty()) {
2837 // If there is a va_start in this function, make a backup copy of
2838 // va_arg_tls somewhere in the function entry block.
2839 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2840 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2842 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2844 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2845 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2848 // Instrument va_start.
2849 // Copy va_list shadow from the backup copy of the TLS contents.
2850 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2851 CallInst *OrigInst = VAStartInstrumentationList[i];
2852 IRBuilder<> IRB(OrigInst->getNextNode());
2853 Value *VAListTag = OrigInst->getArgOperand(0);
2855 Value *RegSaveAreaPtrPtr =
2857 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2858 ConstantInt::get(MS.IntptrTy, 16)),
2859 Type::getInt64PtrTy(*MS.C));
2860 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2861 Value *RegSaveAreaShadowPtr =
2862 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2863 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2864 AMD64FpEndOffset, 16);
2866 Value *OverflowArgAreaPtrPtr =
2868 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2869 ConstantInt::get(MS.IntptrTy, 8)),
2870 Type::getInt64PtrTy(*MS.C));
2871 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2872 Value *OverflowArgAreaShadowPtr =
2873 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2874 Value *SrcPtr = IRB.CreateConstGEP1_32(IRB.getInt8Ty(), VAArgTLSCopy,
2876 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2881 /// \brief MIPS64-specific implementation of VarArgHelper.
2882 struct VarArgMIPS64Helper : public VarArgHelper {
2884 MemorySanitizer &MS;
2885 MemorySanitizerVisitor &MSV;
2886 Value *VAArgTLSCopy;
2889 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2891 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
2892 MemorySanitizerVisitor &MSV)
2893 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2894 VAArgSize(nullptr) {}
2896 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2897 unsigned VAArgOffset = 0;
2898 const DataLayout &DL = F.getParent()->getDataLayout();
2899 for (CallSite::arg_iterator ArgIt = CS.arg_begin() + 1, End = CS.arg_end();
2900 ArgIt != End; ++ArgIt) {
2903 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
2904 #if defined(__MIPSEB__) || defined(MIPSEB)
2905 // Adjusting the shadow for argument with size < 8 to match the placement
2906 // of bits in big endian system
2908 VAArgOffset += (8 - ArgSize);
2910 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
2911 VAArgOffset += ArgSize;
2912 VAArgOffset = RoundUpToAlignment(VAArgOffset, 8);
2913 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2916 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
2917 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
2918 // a new class member i.e. it is the total size of all VarArgs.
2919 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
2922 /// \brief Compute the shadow address for a given va_arg.
2923 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2925 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2926 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2927 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2931 void visitVAStartInst(VAStartInst &I) override {
2932 IRBuilder<> IRB(&I);
2933 VAStartInstrumentationList.push_back(&I);
2934 Value *VAListTag = I.getArgOperand(0);
2935 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2936 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2937 /* size */8, /* alignment */8, false);
2940 void visitVACopyInst(VACopyInst &I) override {
2941 IRBuilder<> IRB(&I);
2942 Value *VAListTag = I.getArgOperand(0);
2943 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2944 // Unpoison the whole __va_list_tag.
2945 // FIXME: magic ABI constants.
2946 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2947 /* size */8, /* alignment */8, false);
2950 void finalizeInstrumentation() override {
2951 assert(!VAArgSize && !VAArgTLSCopy &&
2952 "finalizeInstrumentation called twice");
2953 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2954 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2955 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
2958 if (!VAStartInstrumentationList.empty()) {
2959 // If there is a va_start in this function, make a backup copy of
2960 // va_arg_tls somewhere in the function entry block.
2961 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2962 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2965 // Instrument va_start.
2966 // Copy va_list shadow from the backup copy of the TLS contents.
2967 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2968 CallInst *OrigInst = VAStartInstrumentationList[i];
2969 IRBuilder<> IRB(OrigInst->getNextNode());
2970 Value *VAListTag = OrigInst->getArgOperand(0);
2971 Value *RegSaveAreaPtrPtr =
2972 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2973 Type::getInt64PtrTy(*MS.C));
2974 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2975 Value *RegSaveAreaShadowPtr =
2976 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2977 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
2982 /// \brief A no-op implementation of VarArgHelper.
2983 struct VarArgNoOpHelper : public VarArgHelper {
2984 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2985 MemorySanitizerVisitor &MSV) {}
2987 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2989 void visitVAStartInst(VAStartInst &I) override {}
2991 void visitVACopyInst(VACopyInst &I) override {}
2993 void finalizeInstrumentation() override {}
2996 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2997 MemorySanitizerVisitor &Visitor) {
2998 // VarArg handling is only implemented on AMD64. False positives are possible
2999 // on other platforms.
3000 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
3001 if (TargetTriple.getArch() == llvm::Triple::x86_64)
3002 return new VarArgAMD64Helper(Func, Msan, Visitor);
3003 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
3004 TargetTriple.getArch() == llvm::Triple::mips64el)
3005 return new VarArgMIPS64Helper(Func, Msan, Visitor);
3007 return new VarArgNoOpHelper(Func, Msan, Visitor);
3012 bool MemorySanitizer::runOnFunction(Function &F) {
3013 if (&F == MsanCtorFunction)
3015 MemorySanitizerVisitor Visitor(F, *this);
3017 // Clear out readonly/readnone attributes.
3019 B.addAttribute(Attribute::ReadOnly)
3020 .addAttribute(Attribute::ReadNone);
3021 F.removeAttributes(AttributeSet::FunctionIndex,
3022 AttributeSet::get(F.getContext(),
3023 AttributeSet::FunctionIndex, B));
3025 return Visitor.runOnFunction();