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 pattern"),
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 #ifdef MSAN_LINUX_X86_64_OLD_MAPPING
226 0x400000000000, // AndMask
227 0, // XorMask (not used)
228 0, // ShadowBase (not used)
229 0x200000000000, // OriginBase
231 0, // AndMask (not used)
232 0x500000000000, // XorMask
233 0, // ShadowBase (not used)
234 0x100000000000, // OriginBase
239 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
240 0x004000000000, // AndMask
241 0, // XorMask (not used)
242 0, // ShadowBase (not used)
243 0x002000000000, // OriginBase
247 static const MemoryMapParams Linux_PowerPC64_MemoryMapParams = {
248 0x200000000000, // AndMask
249 0x100000000000, // XorMask
250 0x080000000000, // ShadowBase
251 0x1C0000000000, // OriginBase
255 static const MemoryMapParams Linux_AArch64_MemoryMapParams = {
256 0, // AndMask (not used)
257 0x06000000000, // XorMask
258 0, // ShadowBase (not used)
259 0x01000000000, // OriginBase
263 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
264 0x000180000000, // AndMask
265 0x000040000000, // XorMask
266 0x000020000000, // ShadowBase
267 0x000700000000, // OriginBase
271 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
272 0xc00000000000, // AndMask
273 0x200000000000, // XorMask
274 0x100000000000, // ShadowBase
275 0x380000000000, // OriginBase
278 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
279 &Linux_I386_MemoryMapParams,
280 &Linux_X86_64_MemoryMapParams,
283 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
285 &Linux_MIPS64_MemoryMapParams,
288 static const PlatformMemoryMapParams Linux_PowerPC_MemoryMapParams = {
290 &Linux_PowerPC64_MemoryMapParams,
293 static const PlatformMemoryMapParams Linux_ARM_MemoryMapParams = {
295 &Linux_AArch64_MemoryMapParams,
298 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
299 &FreeBSD_I386_MemoryMapParams,
300 &FreeBSD_X86_64_MemoryMapParams,
303 /// \brief An instrumentation pass implementing detection of uninitialized
306 /// MemorySanitizer: instrument the code in module to find
307 /// uninitialized reads.
308 class MemorySanitizer : public FunctionPass {
310 MemorySanitizer(int TrackOrigins = 0)
312 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
313 WarningFn(nullptr) {}
314 const char *getPassName() const override { return "MemorySanitizer"; }
315 bool runOnFunction(Function &F) override;
316 bool doInitialization(Module &M) override;
317 static char ID; // Pass identification, replacement for typeid.
320 void initializeCallbacks(Module &M);
322 /// \brief Track origins (allocation points) of uninitialized values.
328 /// \brief Thread-local shadow storage for function parameters.
329 GlobalVariable *ParamTLS;
330 /// \brief Thread-local origin storage for function parameters.
331 GlobalVariable *ParamOriginTLS;
332 /// \brief Thread-local shadow storage for function return value.
333 GlobalVariable *RetvalTLS;
334 /// \brief Thread-local origin storage for function return value.
335 GlobalVariable *RetvalOriginTLS;
336 /// \brief Thread-local shadow storage for in-register va_arg function
337 /// parameters (x86_64-specific).
338 GlobalVariable *VAArgTLS;
339 /// \brief Thread-local shadow storage for va_arg overflow area
340 /// (x86_64-specific).
341 GlobalVariable *VAArgOverflowSizeTLS;
342 /// \brief Thread-local space used to pass origin value to the UMR reporting
344 GlobalVariable *OriginTLS;
346 /// \brief The run-time callback to print a warning.
348 // These arrays are indexed by log2(AccessSize).
349 Value *MaybeWarningFn[kNumberOfAccessSizes];
350 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
352 /// \brief Run-time helper that generates a new origin value for a stack
354 Value *MsanSetAllocaOrigin4Fn;
355 /// \brief Run-time helper that poisons stack on function entry.
356 Value *MsanPoisonStackFn;
357 /// \brief Run-time helper that records a store (or any event) of an
358 /// uninitialized value and returns an updated origin id encoding this info.
359 Value *MsanChainOriginFn;
360 /// \brief MSan runtime replacements for memmove, memcpy and memset.
361 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
363 /// \brief Memory map parameters used in application-to-shadow calculation.
364 const MemoryMapParams *MapParams;
366 MDNode *ColdCallWeights;
367 /// \brief Branch weights for origin store.
368 MDNode *OriginStoreWeights;
369 /// \brief An empty volatile inline asm that prevents callback merge.
371 Function *MsanCtorFunction;
373 friend struct MemorySanitizerVisitor;
374 friend struct VarArgAMD64Helper;
375 friend struct VarArgMIPS64Helper;
377 } // anonymous namespace
379 char MemorySanitizer::ID = 0;
380 INITIALIZE_PASS(MemorySanitizer, "msan",
381 "MemorySanitizer: detects uninitialized reads.",
384 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
385 return new MemorySanitizer(TrackOrigins);
388 /// \brief Create a non-const global initialized with the given string.
390 /// Creates a writable global for Str so that we can pass it to the
391 /// run-time lib. Runtime uses first 4 bytes of the string to store the
392 /// frame ID, so the string needs to be mutable.
393 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
395 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
396 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
397 GlobalValue::PrivateLinkage, StrConst, "");
400 /// \brief Insert extern declaration of runtime-provided functions and globals.
401 void MemorySanitizer::initializeCallbacks(Module &M) {
402 // Only do this once.
407 // Create the callback.
408 // FIXME: this function should have "Cold" calling conv,
409 // which is not yet implemented.
410 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
411 : "__msan_warning_noreturn";
412 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), nullptr);
414 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
416 unsigned AccessSize = 1 << AccessSizeIndex;
417 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
418 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
419 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
420 IRB.getInt32Ty(), nullptr);
422 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
423 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
424 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
425 IRB.getInt8PtrTy(), IRB.getInt32Ty(), nullptr);
428 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
429 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
430 IRB.getInt8PtrTy(), IntptrTy, nullptr);
432 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
433 IRB.getInt8PtrTy(), IntptrTy, nullptr);
434 MsanChainOriginFn = M.getOrInsertFunction(
435 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), nullptr);
436 MemmoveFn = M.getOrInsertFunction(
437 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
438 IRB.getInt8PtrTy(), IntptrTy, nullptr);
439 MemcpyFn = M.getOrInsertFunction(
440 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
442 MemsetFn = M.getOrInsertFunction(
443 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
447 RetvalTLS = new GlobalVariable(
448 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
449 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
450 GlobalVariable::InitialExecTLSModel);
451 RetvalOriginTLS = new GlobalVariable(
452 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
453 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
455 ParamTLS = new GlobalVariable(
456 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
457 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
458 GlobalVariable::InitialExecTLSModel);
459 ParamOriginTLS = new GlobalVariable(
460 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
461 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
462 nullptr, GlobalVariable::InitialExecTLSModel);
464 VAArgTLS = new GlobalVariable(
465 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
466 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
467 GlobalVariable::InitialExecTLSModel);
468 VAArgOverflowSizeTLS = new GlobalVariable(
469 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
470 "__msan_va_arg_overflow_size_tls", nullptr,
471 GlobalVariable::InitialExecTLSModel);
472 OriginTLS = new GlobalVariable(
473 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
474 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
476 // We insert an empty inline asm after __msan_report* to avoid callback merge.
477 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
478 StringRef(""), StringRef(""),
479 /*hasSideEffects=*/true);
482 /// \brief Module-level initialization.
484 /// inserts a call to __msan_init to the module's constructor list.
485 bool MemorySanitizer::doInitialization(Module &M) {
486 auto &DL = M.getDataLayout();
488 Triple TargetTriple(M.getTargetTriple());
489 switch (TargetTriple.getOS()) {
490 case Triple::FreeBSD:
491 switch (TargetTriple.getArch()) {
493 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
496 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
499 report_fatal_error("unsupported architecture");
503 switch (TargetTriple.getArch()) {
505 MapParams = Linux_X86_MemoryMapParams.bits64;
508 MapParams = Linux_X86_MemoryMapParams.bits32;
511 case Triple::mips64el:
512 MapParams = Linux_MIPS_MemoryMapParams.bits64;
515 case Triple::ppc64le:
516 MapParams = Linux_PowerPC_MemoryMapParams.bits64;
518 case Triple::aarch64:
519 case Triple::aarch64_be:
520 MapParams = Linux_ARM_MemoryMapParams.bits64;
523 report_fatal_error("unsupported architecture");
527 report_fatal_error("unsupported operating system");
530 C = &(M.getContext());
532 IntptrTy = IRB.getIntPtrTy(DL);
533 OriginTy = IRB.getInt32Ty();
535 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
536 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
538 std::tie(MsanCtorFunction, std::ignore) =
539 createSanitizerCtorAndInitFunctions(M, kMsanModuleCtorName, kMsanInitName,
543 appendToGlobalCtors(M, MsanCtorFunction, 0);
546 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
547 IRB.getInt32(TrackOrigins), "__msan_track_origins");
550 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
551 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
558 /// \brief A helper class that handles instrumentation of VarArg
559 /// functions on a particular platform.
561 /// Implementations are expected to insert the instrumentation
562 /// necessary to propagate argument shadow through VarArg function
563 /// calls. Visit* methods are called during an InstVisitor pass over
564 /// the function, and should avoid creating new basic blocks. A new
565 /// instance of this class is created for each instrumented function.
566 struct VarArgHelper {
567 /// \brief Visit a CallSite.
568 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
570 /// \brief Visit a va_start call.
571 virtual void visitVAStartInst(VAStartInst &I) = 0;
573 /// \brief Visit a va_copy call.
574 virtual void visitVACopyInst(VACopyInst &I) = 0;
576 /// \brief Finalize function instrumentation.
578 /// This method is called after visiting all interesting (see above)
579 /// instructions in a function.
580 virtual void finalizeInstrumentation() = 0;
582 virtual ~VarArgHelper() {}
585 struct MemorySanitizerVisitor;
588 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
589 MemorySanitizerVisitor &Visitor);
591 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
592 if (TypeSize <= 8) return 0;
593 return Log2_32_Ceil(TypeSize / 8);
596 /// This class does all the work for a given function. Store and Load
597 /// instructions store and load corresponding shadow and origin
598 /// values. Most instructions propagate shadow from arguments to their
599 /// return values. Certain instructions (most importantly, BranchInst)
600 /// test their argument shadow and print reports (with a runtime call) if it's
602 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
605 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
606 ValueMap<Value*, Value*> ShadowMap, OriginMap;
607 std::unique_ptr<VarArgHelper> VAHelper;
609 // The following flags disable parts of MSan instrumentation based on
610 // blacklist contents and command-line options.
612 bool PropagateShadow;
615 bool CheckReturnValue;
617 struct ShadowOriginAndInsertPoint {
620 Instruction *OrigIns;
621 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
622 : Shadow(S), Origin(O), OrigIns(I) { }
624 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
625 SmallVector<Instruction*, 16> StoreList;
627 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
628 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
629 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
630 InsertChecks = SanitizeFunction;
631 PropagateShadow = SanitizeFunction;
632 PoisonStack = SanitizeFunction && ClPoisonStack;
633 PoisonUndef = SanitizeFunction && ClPoisonUndef;
634 // FIXME: Consider using SpecialCaseList to specify a list of functions that
635 // must always return fully initialized values. For now, we hardcode "main".
636 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
638 DEBUG(if (!InsertChecks)
639 dbgs() << "MemorySanitizer is not inserting checks into '"
640 << F.getName() << "'\n");
643 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
644 if (MS.TrackOrigins <= 1) return V;
645 return IRB.CreateCall(MS.MsanChainOriginFn, V);
648 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
649 const DataLayout &DL = F.getParent()->getDataLayout();
650 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
651 if (IntptrSize == kOriginSize) return Origin;
652 assert(IntptrSize == kOriginSize * 2);
653 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
654 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
657 /// \brief Fill memory range with the given origin value.
658 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
659 unsigned Size, unsigned Alignment) {
660 const DataLayout &DL = F.getParent()->getDataLayout();
661 unsigned IntptrAlignment = DL.getABITypeAlignment(MS.IntptrTy);
662 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
663 assert(IntptrAlignment >= kMinOriginAlignment);
664 assert(IntptrSize >= kOriginSize);
667 unsigned CurrentAlignment = Alignment;
668 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
669 Value *IntptrOrigin = originToIntptr(IRB, Origin);
670 Value *IntptrOriginPtr =
671 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
672 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
673 Value *Ptr = i ? IRB.CreateConstGEP1_32(MS.IntptrTy, IntptrOriginPtr, i)
675 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
676 Ofs += IntptrSize / kOriginSize;
677 CurrentAlignment = IntptrAlignment;
681 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
683 i ? IRB.CreateConstGEP1_32(nullptr, OriginPtr, i) : OriginPtr;
684 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
685 CurrentAlignment = kMinOriginAlignment;
689 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
690 unsigned Alignment, bool AsCall) {
691 const DataLayout &DL = F.getParent()->getDataLayout();
692 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
693 unsigned StoreSize = DL.getTypeStoreSize(Shadow->getType());
694 if (isa<StructType>(Shadow->getType())) {
695 paintOrigin(IRB, updateOrigin(Origin, IRB),
696 getOriginPtr(Addr, IRB, Alignment), StoreSize,
699 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
700 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
701 if (ConstantShadow) {
702 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
703 paintOrigin(IRB, updateOrigin(Origin, IRB),
704 getOriginPtr(Addr, IRB, Alignment), StoreSize,
709 unsigned TypeSizeInBits =
710 DL.getTypeSizeInBits(ConvertedShadow->getType());
711 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
712 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
713 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
714 Value *ConvertedShadow2 = IRB.CreateZExt(
715 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
716 IRB.CreateCall(Fn, {ConvertedShadow2,
717 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
720 Value *Cmp = IRB.CreateICmpNE(
721 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
722 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
723 Cmp, &*IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
724 IRBuilder<> IRBNew(CheckTerm);
725 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
726 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
732 void materializeStores(bool InstrumentWithCalls) {
733 for (auto Inst : StoreList) {
734 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
736 IRBuilder<> IRB(&SI);
737 Value *Val = SI.getValueOperand();
738 Value *Addr = SI.getPointerOperand();
739 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
740 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
743 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
744 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
747 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
749 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
751 if (MS.TrackOrigins && !SI.isAtomic())
752 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
753 InstrumentWithCalls);
757 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
759 IRBuilder<> IRB(OrigIns);
760 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
761 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
762 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
764 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
765 if (ConstantShadow) {
766 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
767 if (MS.TrackOrigins) {
768 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
771 IRB.CreateCall(MS.WarningFn, {});
772 IRB.CreateCall(MS.EmptyAsm, {});
773 // FIXME: Insert UnreachableInst if !ClKeepGoing?
774 // This may invalidate some of the following checks and needs to be done
780 const DataLayout &DL = OrigIns->getModule()->getDataLayout();
782 unsigned TypeSizeInBits = DL.getTypeSizeInBits(ConvertedShadow->getType());
783 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
784 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
785 Value *Fn = MS.MaybeWarningFn[SizeIndex];
786 Value *ConvertedShadow2 =
787 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
788 IRB.CreateCall(Fn, {ConvertedShadow2, MS.TrackOrigins && Origin
790 : (Value *)IRB.getInt32(0)});
792 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
793 getCleanShadow(ConvertedShadow), "_mscmp");
794 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
796 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
798 IRB.SetInsertPoint(CheckTerm);
799 if (MS.TrackOrigins) {
800 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
803 IRB.CreateCall(MS.WarningFn, {});
804 IRB.CreateCall(MS.EmptyAsm, {});
805 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
809 void materializeChecks(bool InstrumentWithCalls) {
810 for (const auto &ShadowData : InstrumentationList) {
811 Instruction *OrigIns = ShadowData.OrigIns;
812 Value *Shadow = ShadowData.Shadow;
813 Value *Origin = ShadowData.Origin;
814 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
816 DEBUG(dbgs() << "DONE:\n" << F);
819 /// \brief Add MemorySanitizer instrumentation to a function.
820 bool runOnFunction() {
821 MS.initializeCallbacks(*F.getParent());
823 // In the presence of unreachable blocks, we may see Phi nodes with
824 // incoming nodes from such blocks. Since InstVisitor skips unreachable
825 // blocks, such nodes will not have any shadow value associated with them.
826 // It's easier to remove unreachable blocks than deal with missing shadow.
827 removeUnreachableBlocks(F);
829 // Iterate all BBs in depth-first order and create shadow instructions
830 // for all instructions (where applicable).
831 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
832 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
836 // Finalize PHI nodes.
837 for (PHINode *PN : ShadowPHINodes) {
838 PHINode *PNS = cast<PHINode>(getShadow(PN));
839 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
840 size_t NumValues = PN->getNumIncomingValues();
841 for (size_t v = 0; v < NumValues; v++) {
842 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
843 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
847 VAHelper->finalizeInstrumentation();
849 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
850 InstrumentationList.size() + StoreList.size() >
851 (unsigned)ClInstrumentationWithCallThreshold;
853 // Delayed instrumentation of StoreInst.
854 // This may add new checks to be inserted later.
855 materializeStores(InstrumentWithCalls);
857 // Insert shadow value checks.
858 materializeChecks(InstrumentWithCalls);
863 /// \brief Compute the shadow type that corresponds to a given Value.
864 Type *getShadowTy(Value *V) {
865 return getShadowTy(V->getType());
868 /// \brief Compute the shadow type that corresponds to a given Type.
869 Type *getShadowTy(Type *OrigTy) {
870 if (!OrigTy->isSized()) {
873 // For integer type, shadow is the same as the original type.
874 // This may return weird-sized types like i1.
875 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
877 const DataLayout &DL = F.getParent()->getDataLayout();
878 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
879 uint32_t EltSize = DL.getTypeSizeInBits(VT->getElementType());
880 return VectorType::get(IntegerType::get(*MS.C, EltSize),
881 VT->getNumElements());
883 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
884 return ArrayType::get(getShadowTy(AT->getElementType()),
885 AT->getNumElements());
887 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
888 SmallVector<Type*, 4> Elements;
889 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
890 Elements.push_back(getShadowTy(ST->getElementType(i)));
891 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
892 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
895 uint32_t TypeSize = DL.getTypeSizeInBits(OrigTy);
896 return IntegerType::get(*MS.C, TypeSize);
899 /// \brief Flatten a vector type.
900 Type *getShadowTyNoVec(Type *ty) {
901 if (VectorType *vt = dyn_cast<VectorType>(ty))
902 return IntegerType::get(*MS.C, vt->getBitWidth());
906 /// \brief Convert a shadow value to it's flattened variant.
907 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
908 Type *Ty = V->getType();
909 Type *NoVecTy = getShadowTyNoVec(Ty);
910 if (Ty == NoVecTy) return V;
911 return IRB.CreateBitCast(V, NoVecTy);
914 /// \brief Compute the integer shadow offset that corresponds to a given
915 /// application address.
917 /// Offset = (Addr & ~AndMask) ^ XorMask
918 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
919 Value *OffsetLong = IRB.CreatePointerCast(Addr, MS.IntptrTy);
921 uint64_t AndMask = MS.MapParams->AndMask;
924 IRB.CreateAnd(OffsetLong, ConstantInt::get(MS.IntptrTy, ~AndMask));
926 uint64_t XorMask = MS.MapParams->XorMask;
929 IRB.CreateXor(OffsetLong, ConstantInt::get(MS.IntptrTy, XorMask));
933 /// \brief Compute the shadow address that corresponds to a given application
936 /// Shadow = ShadowBase + Offset
937 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
939 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
940 uint64_t ShadowBase = MS.MapParams->ShadowBase;
943 IRB.CreateAdd(ShadowLong,
944 ConstantInt::get(MS.IntptrTy, ShadowBase));
945 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
948 /// \brief Compute the origin address that corresponds to a given application
951 /// OriginAddr = (OriginBase + Offset) & ~3ULL
952 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
953 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
954 uint64_t OriginBase = MS.MapParams->OriginBase;
957 IRB.CreateAdd(OriginLong,
958 ConstantInt::get(MS.IntptrTy, OriginBase));
959 if (Alignment < kMinOriginAlignment) {
960 uint64_t Mask = kMinOriginAlignment - 1;
961 OriginLong = IRB.CreateAnd(OriginLong,
962 ConstantInt::get(MS.IntptrTy, ~Mask));
964 return IRB.CreateIntToPtr(OriginLong,
965 PointerType::get(IRB.getInt32Ty(), 0));
968 /// \brief Compute the shadow address for a given function argument.
970 /// Shadow = ParamTLS+ArgOffset.
971 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
973 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
974 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
975 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
979 /// \brief Compute the origin address for a given function argument.
980 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
982 if (!MS.TrackOrigins) return nullptr;
983 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
984 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
985 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
989 /// \brief Compute the shadow address for a retval.
990 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
991 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
992 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
996 /// \brief Compute the origin address for a retval.
997 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
998 // We keep a single origin for the entire retval. Might be too optimistic.
999 return MS.RetvalOriginTLS;
1002 /// \brief Set SV to be the shadow value for V.
1003 void setShadow(Value *V, Value *SV) {
1004 assert(!ShadowMap.count(V) && "Values may only have one shadow");
1005 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
1008 /// \brief Set Origin to be the origin value for V.
1009 void setOrigin(Value *V, Value *Origin) {
1010 if (!MS.TrackOrigins) return;
1011 assert(!OriginMap.count(V) && "Values may only have one origin");
1012 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
1013 OriginMap[V] = Origin;
1016 /// \brief Create a clean shadow value for a given value.
1018 /// Clean shadow (all zeroes) means all bits of the value are defined
1020 Constant *getCleanShadow(Value *V) {
1021 Type *ShadowTy = getShadowTy(V);
1024 return Constant::getNullValue(ShadowTy);
1027 /// \brief Create a dirty shadow of a given shadow type.
1028 Constant *getPoisonedShadow(Type *ShadowTy) {
1030 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
1031 return Constant::getAllOnesValue(ShadowTy);
1032 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
1033 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
1034 getPoisonedShadow(AT->getElementType()));
1035 return ConstantArray::get(AT, Vals);
1037 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
1038 SmallVector<Constant *, 4> Vals;
1039 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
1040 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
1041 return ConstantStruct::get(ST, Vals);
1043 llvm_unreachable("Unexpected shadow type");
1046 /// \brief Create a dirty shadow for a given value.
1047 Constant *getPoisonedShadow(Value *V) {
1048 Type *ShadowTy = getShadowTy(V);
1051 return getPoisonedShadow(ShadowTy);
1054 /// \brief Create a clean (zero) origin.
1055 Value *getCleanOrigin() {
1056 return Constant::getNullValue(MS.OriginTy);
1059 /// \brief Get the shadow value for a given Value.
1061 /// This function either returns the value set earlier with setShadow,
1062 /// or extracts if from ParamTLS (for function arguments).
1063 Value *getShadow(Value *V) {
1064 if (!PropagateShadow) return getCleanShadow(V);
1065 if (Instruction *I = dyn_cast<Instruction>(V)) {
1066 // For instructions the shadow is already stored in the map.
1067 Value *Shadow = ShadowMap[V];
1069 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1071 assert(Shadow && "No shadow for a value");
1075 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1076 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1077 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1081 if (Argument *A = dyn_cast<Argument>(V)) {
1082 // For arguments we compute the shadow on demand and store it in the map.
1083 Value **ShadowPtr = &ShadowMap[V];
1086 Function *F = A->getParent();
1087 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1088 unsigned ArgOffset = 0;
1089 const DataLayout &DL = F->getParent()->getDataLayout();
1090 for (auto &FArg : F->args()) {
1091 if (!FArg.getType()->isSized()) {
1092 DEBUG(dbgs() << "Arg is not sized\n");
1097 ? DL.getTypeAllocSize(FArg.getType()->getPointerElementType())
1098 : DL.getTypeAllocSize(FArg.getType());
1100 bool Overflow = ArgOffset + Size > kParamTLSSize;
1101 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1102 if (FArg.hasByValAttr()) {
1103 // ByVal pointer itself has clean shadow. We copy the actual
1104 // argument shadow to the underlying memory.
1105 // Figure out maximal valid memcpy alignment.
1106 unsigned ArgAlign = FArg.getParamAlignment();
1107 if (ArgAlign == 0) {
1108 Type *EltType = A->getType()->getPointerElementType();
1109 ArgAlign = DL.getABITypeAlignment(EltType);
1112 // ParamTLS overflow.
1113 EntryIRB.CreateMemSet(
1114 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1115 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1117 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1118 Value *Cpy = EntryIRB.CreateMemCpy(
1119 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1121 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1124 *ShadowPtr = getCleanShadow(V);
1127 // ParamTLS overflow.
1128 *ShadowPtr = getCleanShadow(V);
1131 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1134 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1135 **ShadowPtr << "\n");
1136 if (MS.TrackOrigins && !Overflow) {
1138 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1139 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1141 setOrigin(A, getCleanOrigin());
1144 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1146 assert(*ShadowPtr && "Could not find shadow for an argument");
1149 // For everything else the shadow is zero.
1150 return getCleanShadow(V);
1153 /// \brief Get the shadow for i-th argument of the instruction I.
1154 Value *getShadow(Instruction *I, int i) {
1155 return getShadow(I->getOperand(i));
1158 /// \brief Get the origin for a value.
1159 Value *getOrigin(Value *V) {
1160 if (!MS.TrackOrigins) return nullptr;
1161 if (!PropagateShadow) return getCleanOrigin();
1162 if (isa<Constant>(V)) return getCleanOrigin();
1163 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1164 "Unexpected value type in getOrigin()");
1165 Value *Origin = OriginMap[V];
1166 assert(Origin && "Missing origin");
1170 /// \brief Get the origin for i-th argument of the instruction I.
1171 Value *getOrigin(Instruction *I, int i) {
1172 return getOrigin(I->getOperand(i));
1175 /// \brief Remember the place where a shadow check should be inserted.
1177 /// This location will be later instrumented with a check that will print a
1178 /// UMR warning in runtime if the shadow value is not 0.
1179 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1181 if (!InsertChecks) return;
1183 Type *ShadowTy = Shadow->getType();
1184 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1185 "Can only insert checks for integer and vector shadow types");
1187 InstrumentationList.push_back(
1188 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1191 /// \brief Remember the place where a shadow check should be inserted.
1193 /// This location will be later instrumented with a check that will print a
1194 /// UMR warning in runtime if the value is not fully defined.
1195 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1197 Value *Shadow, *Origin;
1198 if (ClCheckConstantShadow) {
1199 Shadow = getShadow(Val);
1200 if (!Shadow) return;
1201 Origin = getOrigin(Val);
1203 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1204 if (!Shadow) return;
1205 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1207 insertShadowCheck(Shadow, Origin, OrigIns);
1210 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1219 case AcquireRelease:
1220 return AcquireRelease;
1221 case SequentiallyConsistent:
1222 return SequentiallyConsistent;
1224 llvm_unreachable("Unknown ordering");
1227 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1236 case AcquireRelease:
1237 return AcquireRelease;
1238 case SequentiallyConsistent:
1239 return SequentiallyConsistent;
1241 llvm_unreachable("Unknown ordering");
1244 // ------------------- Visitors.
1246 /// \brief Instrument LoadInst
1248 /// Loads the corresponding shadow and (optionally) origin.
1249 /// Optionally, checks that the load address is fully defined.
1250 void visitLoadInst(LoadInst &I) {
1251 assert(I.getType()->isSized() && "Load type must have size");
1252 IRBuilder<> IRB(I.getNextNode());
1253 Type *ShadowTy = getShadowTy(&I);
1254 Value *Addr = I.getPointerOperand();
1255 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1256 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1258 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1260 setShadow(&I, getCleanShadow(&I));
1263 if (ClCheckAccessAddress)
1264 insertShadowCheck(I.getPointerOperand(), &I);
1267 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1269 if (MS.TrackOrigins) {
1270 if (PropagateShadow) {
1271 unsigned Alignment = I.getAlignment();
1272 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1273 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1276 setOrigin(&I, getCleanOrigin());
1281 /// \brief Instrument StoreInst
1283 /// Stores the corresponding shadow and (optionally) origin.
1284 /// Optionally, checks that the store address is fully defined.
1285 void visitStoreInst(StoreInst &I) {
1286 StoreList.push_back(&I);
1289 void handleCASOrRMW(Instruction &I) {
1290 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1292 IRBuilder<> IRB(&I);
1293 Value *Addr = I.getOperand(0);
1294 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1296 if (ClCheckAccessAddress)
1297 insertShadowCheck(Addr, &I);
1299 // Only test the conditional argument of cmpxchg instruction.
1300 // The other argument can potentially be uninitialized, but we can not
1301 // detect this situation reliably without possible false positives.
1302 if (isa<AtomicCmpXchgInst>(I))
1303 insertShadowCheck(I.getOperand(1), &I);
1305 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1307 setShadow(&I, getCleanShadow(&I));
1308 setOrigin(&I, getCleanOrigin());
1311 void visitAtomicRMWInst(AtomicRMWInst &I) {
1313 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1316 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1318 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1321 // Vector manipulation.
1322 void visitExtractElementInst(ExtractElementInst &I) {
1323 insertShadowCheck(I.getOperand(1), &I);
1324 IRBuilder<> IRB(&I);
1325 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1327 setOrigin(&I, getOrigin(&I, 0));
1330 void visitInsertElementInst(InsertElementInst &I) {
1331 insertShadowCheck(I.getOperand(2), &I);
1332 IRBuilder<> IRB(&I);
1333 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1334 I.getOperand(2), "_msprop"));
1335 setOriginForNaryOp(I);
1338 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1339 insertShadowCheck(I.getOperand(2), &I);
1340 IRBuilder<> IRB(&I);
1341 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1342 I.getOperand(2), "_msprop"));
1343 setOriginForNaryOp(I);
1347 void visitSExtInst(SExtInst &I) {
1348 IRBuilder<> IRB(&I);
1349 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1350 setOrigin(&I, getOrigin(&I, 0));
1353 void visitZExtInst(ZExtInst &I) {
1354 IRBuilder<> IRB(&I);
1355 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1356 setOrigin(&I, getOrigin(&I, 0));
1359 void visitTruncInst(TruncInst &I) {
1360 IRBuilder<> IRB(&I);
1361 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1362 setOrigin(&I, getOrigin(&I, 0));
1365 void visitBitCastInst(BitCastInst &I) {
1366 // Special case: if this is the bitcast (there is exactly 1 allowed) between
1367 // a musttail call and a ret, don't instrument. New instructions are not
1368 // allowed after a musttail call.
1369 if (auto *CI = dyn_cast<CallInst>(I.getOperand(0)))
1370 if (CI->isMustTailCall())
1372 IRBuilder<> IRB(&I);
1373 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1374 setOrigin(&I, getOrigin(&I, 0));
1377 void visitPtrToIntInst(PtrToIntInst &I) {
1378 IRBuilder<> IRB(&I);
1379 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1380 "_msprop_ptrtoint"));
1381 setOrigin(&I, getOrigin(&I, 0));
1384 void visitIntToPtrInst(IntToPtrInst &I) {
1385 IRBuilder<> IRB(&I);
1386 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1387 "_msprop_inttoptr"));
1388 setOrigin(&I, getOrigin(&I, 0));
1391 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1392 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1393 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1394 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1395 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1396 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1398 /// \brief Propagate shadow for bitwise AND.
1400 /// This code is exact, i.e. if, for example, a bit in the left argument
1401 /// is defined and 0, then neither the value not definedness of the
1402 /// corresponding bit in B don't affect the resulting shadow.
1403 void visitAnd(BinaryOperator &I) {
1404 IRBuilder<> IRB(&I);
1405 // "And" of 0 and a poisoned value results in unpoisoned value.
1406 // 1&1 => 1; 0&1 => 0; p&1 => p;
1407 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1408 // 1&p => p; 0&p => 0; p&p => p;
1409 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1410 Value *S1 = getShadow(&I, 0);
1411 Value *S2 = getShadow(&I, 1);
1412 Value *V1 = I.getOperand(0);
1413 Value *V2 = I.getOperand(1);
1414 if (V1->getType() != S1->getType()) {
1415 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1416 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1418 Value *S1S2 = IRB.CreateAnd(S1, S2);
1419 Value *V1S2 = IRB.CreateAnd(V1, S2);
1420 Value *S1V2 = IRB.CreateAnd(S1, V2);
1421 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1422 setOriginForNaryOp(I);
1425 void visitOr(BinaryOperator &I) {
1426 IRBuilder<> IRB(&I);
1427 // "Or" of 1 and a poisoned value results in unpoisoned value.
1428 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1429 // 1|0 => 1; 0|0 => 0; p|0 => p;
1430 // 1|p => 1; 0|p => p; p|p => p;
1431 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1432 Value *S1 = getShadow(&I, 0);
1433 Value *S2 = getShadow(&I, 1);
1434 Value *V1 = IRB.CreateNot(I.getOperand(0));
1435 Value *V2 = IRB.CreateNot(I.getOperand(1));
1436 if (V1->getType() != S1->getType()) {
1437 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1438 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1440 Value *S1S2 = IRB.CreateAnd(S1, S2);
1441 Value *V1S2 = IRB.CreateAnd(V1, S2);
1442 Value *S1V2 = IRB.CreateAnd(S1, V2);
1443 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1444 setOriginForNaryOp(I);
1447 /// \brief Default propagation of shadow and/or origin.
1449 /// This class implements the general case of shadow propagation, used in all
1450 /// cases where we don't know and/or don't care about what the operation
1451 /// actually does. It converts all input shadow values to a common type
1452 /// (extending or truncating as necessary), and bitwise OR's them.
1454 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1455 /// fully initialized), and less prone to false positives.
1457 /// This class also implements the general case of origin propagation. For a
1458 /// Nary operation, result origin is set to the origin of an argument that is
1459 /// not entirely initialized. If there is more than one such arguments, the
1460 /// rightmost of them is picked. It does not matter which one is picked if all
1461 /// arguments are initialized.
1462 template <bool CombineShadow>
1467 MemorySanitizerVisitor *MSV;
1470 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1471 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1473 /// \brief Add a pair of shadow and origin values to the mix.
1474 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1475 if (CombineShadow) {
1480 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1481 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1485 if (MSV->MS.TrackOrigins) {
1490 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1491 // No point in adding something that might result in 0 origin value.
1492 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1493 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1495 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1496 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1503 /// \brief Add an application value to the mix.
1504 Combiner &Add(Value *V) {
1505 Value *OpShadow = MSV->getShadow(V);
1506 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1507 return Add(OpShadow, OpOrigin);
1510 /// \brief Set the current combined values as the given instruction's shadow
1512 void Done(Instruction *I) {
1513 if (CombineShadow) {
1515 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1516 MSV->setShadow(I, Shadow);
1518 if (MSV->MS.TrackOrigins) {
1520 MSV->setOrigin(I, Origin);
1525 typedef Combiner<true> ShadowAndOriginCombiner;
1526 typedef Combiner<false> OriginCombiner;
1528 /// \brief Propagate origin for arbitrary operation.
1529 void setOriginForNaryOp(Instruction &I) {
1530 if (!MS.TrackOrigins) return;
1531 IRBuilder<> IRB(&I);
1532 OriginCombiner OC(this, IRB);
1533 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1538 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1539 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1540 "Vector of pointers is not a valid shadow type");
1541 return Ty->isVectorTy() ?
1542 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1543 Ty->getPrimitiveSizeInBits();
1546 /// \brief Cast between two shadow types, extending or truncating as
1548 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1549 bool Signed = false) {
1550 Type *srcTy = V->getType();
1551 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1552 return IRB.CreateIntCast(V, dstTy, Signed);
1553 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1554 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1555 return IRB.CreateIntCast(V, dstTy, Signed);
1556 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1557 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1558 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1560 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1561 return IRB.CreateBitCast(V2, dstTy);
1562 // TODO: handle struct types.
1565 /// \brief Cast an application value to the type of its own shadow.
1566 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1567 Type *ShadowTy = getShadowTy(V);
1568 if (V->getType() == ShadowTy)
1570 if (V->getType()->isPtrOrPtrVectorTy())
1571 return IRB.CreatePtrToInt(V, ShadowTy);
1573 return IRB.CreateBitCast(V, ShadowTy);
1576 /// \brief Propagate shadow for arbitrary operation.
1577 void handleShadowOr(Instruction &I) {
1578 IRBuilder<> IRB(&I);
1579 ShadowAndOriginCombiner SC(this, IRB);
1580 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1585 // \brief Handle multiplication by constant.
1587 // Handle a special case of multiplication by constant that may have one or
1588 // more zeros in the lower bits. This makes corresponding number of lower bits
1589 // of the result zero as well. We model it by shifting the other operand
1590 // shadow left by the required number of bits. Effectively, we transform
1591 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1592 // We use multiplication by 2**N instead of shift to cover the case of
1593 // multiplication by 0, which may occur in some elements of a vector operand.
1594 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1596 Constant *ShadowMul;
1597 Type *Ty = ConstArg->getType();
1598 if (Ty->isVectorTy()) {
1599 unsigned NumElements = Ty->getVectorNumElements();
1600 Type *EltTy = Ty->getSequentialElementType();
1601 SmallVector<Constant *, 16> Elements;
1602 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1603 if (ConstantInt *Elt =
1604 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx))) {
1605 APInt V = Elt->getValue();
1606 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1607 Elements.push_back(ConstantInt::get(EltTy, V2));
1609 Elements.push_back(ConstantInt::get(EltTy, 1));
1612 ShadowMul = ConstantVector::get(Elements);
1614 if (ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg)) {
1615 APInt V = Elt->getValue();
1616 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1617 ShadowMul = ConstantInt::get(Ty, V2);
1619 ShadowMul = ConstantInt::get(Ty, 1);
1623 IRBuilder<> IRB(&I);
1625 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1626 setOrigin(&I, getOrigin(OtherArg));
1629 void visitMul(BinaryOperator &I) {
1630 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1631 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1632 if (constOp0 && !constOp1)
1633 handleMulByConstant(I, constOp0, I.getOperand(1));
1634 else if (constOp1 && !constOp0)
1635 handleMulByConstant(I, constOp1, I.getOperand(0));
1640 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1641 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1642 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1643 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1644 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1645 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1647 void handleDiv(Instruction &I) {
1648 IRBuilder<> IRB(&I);
1649 // Strict on the second argument.
1650 insertShadowCheck(I.getOperand(1), &I);
1651 setShadow(&I, getShadow(&I, 0));
1652 setOrigin(&I, getOrigin(&I, 0));
1655 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1656 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1657 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1658 void visitURem(BinaryOperator &I) { handleDiv(I); }
1659 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1660 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1662 /// \brief Instrument == and != comparisons.
1664 /// Sometimes the comparison result is known even if some of the bits of the
1665 /// arguments are not.
1666 void handleEqualityComparison(ICmpInst &I) {
1667 IRBuilder<> IRB(&I);
1668 Value *A = I.getOperand(0);
1669 Value *B = I.getOperand(1);
1670 Value *Sa = getShadow(A);
1671 Value *Sb = getShadow(B);
1673 // Get rid of pointers and vectors of pointers.
1674 // For ints (and vectors of ints), types of A and Sa match,
1675 // and this is a no-op.
1676 A = IRB.CreatePointerCast(A, Sa->getType());
1677 B = IRB.CreatePointerCast(B, Sb->getType());
1679 // A == B <==> (C = A^B) == 0
1680 // A != B <==> (C = A^B) != 0
1682 Value *C = IRB.CreateXor(A, B);
1683 Value *Sc = IRB.CreateOr(Sa, Sb);
1684 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1685 // Result is defined if one of the following is true
1686 // * there is a defined 1 bit in C
1687 // * C is fully defined
1688 // Si = !(C & ~Sc) && Sc
1689 Value *Zero = Constant::getNullValue(Sc->getType());
1690 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1692 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1694 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1695 Si->setName("_msprop_icmp");
1697 setOriginForNaryOp(I);
1700 /// \brief Build the lowest possible value of V, taking into account V's
1701 /// uninitialized bits.
1702 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1705 // Split shadow into sign bit and other bits.
1706 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1707 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1708 // Maximise the undefined shadow bit, minimize other undefined bits.
1710 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1712 // Minimize undefined bits.
1713 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1717 /// \brief Build the highest possible value of V, taking into account V's
1718 /// uninitialized bits.
1719 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1722 // Split shadow into sign bit and other bits.
1723 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1724 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1725 // Minimise the undefined shadow bit, maximise other undefined bits.
1727 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1729 // Maximize undefined bits.
1730 return IRB.CreateOr(A, Sa);
1734 /// \brief Instrument relational comparisons.
1736 /// This function does exact shadow propagation for all relational
1737 /// comparisons of integers, pointers and vectors of those.
1738 /// FIXME: output seems suboptimal when one of the operands is a constant
1739 void handleRelationalComparisonExact(ICmpInst &I) {
1740 IRBuilder<> IRB(&I);
1741 Value *A = I.getOperand(0);
1742 Value *B = I.getOperand(1);
1743 Value *Sa = getShadow(A);
1744 Value *Sb = getShadow(B);
1746 // Get rid of pointers and vectors of pointers.
1747 // For ints (and vectors of ints), types of A and Sa match,
1748 // and this is a no-op.
1749 A = IRB.CreatePointerCast(A, Sa->getType());
1750 B = IRB.CreatePointerCast(B, Sb->getType());
1752 // Let [a0, a1] be the interval of possible values of A, taking into account
1753 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1754 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1755 bool IsSigned = I.isSigned();
1756 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1757 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1758 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1759 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1760 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1761 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1762 Value *Si = IRB.CreateXor(S1, S2);
1764 setOriginForNaryOp(I);
1767 /// \brief Instrument signed relational comparisons.
1769 /// Handle sign bit tests: x<0, x>=0, x<=-1, x>-1 by propagating the highest
1770 /// bit of the shadow. Everything else is delegated to handleShadowOr().
1771 void handleSignedRelationalComparison(ICmpInst &I) {
1773 Value *op = nullptr;
1774 CmpInst::Predicate pre;
1775 if ((constOp = dyn_cast<Constant>(I.getOperand(1)))) {
1776 op = I.getOperand(0);
1777 pre = I.getPredicate();
1778 } else if ((constOp = dyn_cast<Constant>(I.getOperand(0)))) {
1779 op = I.getOperand(1);
1780 pre = I.getSwappedPredicate();
1786 if ((constOp->isNullValue() &&
1787 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) ||
1788 (constOp->isAllOnesValue() &&
1789 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE))) {
1790 IRBuilder<> IRB(&I);
1791 Value *Shadow = IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op),
1793 setShadow(&I, Shadow);
1794 setOrigin(&I, getOrigin(op));
1800 void visitICmpInst(ICmpInst &I) {
1801 if (!ClHandleICmp) {
1805 if (I.isEquality()) {
1806 handleEqualityComparison(I);
1810 assert(I.isRelational());
1811 if (ClHandleICmpExact) {
1812 handleRelationalComparisonExact(I);
1816 handleSignedRelationalComparison(I);
1820 assert(I.isUnsigned());
1821 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1822 handleRelationalComparisonExact(I);
1829 void visitFCmpInst(FCmpInst &I) {
1833 void handleShift(BinaryOperator &I) {
1834 IRBuilder<> IRB(&I);
1835 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1836 // Otherwise perform the same shift on S1.
1837 Value *S1 = getShadow(&I, 0);
1838 Value *S2 = getShadow(&I, 1);
1839 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1841 Value *V2 = I.getOperand(1);
1842 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1843 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1844 setOriginForNaryOp(I);
1847 void visitShl(BinaryOperator &I) { handleShift(I); }
1848 void visitAShr(BinaryOperator &I) { handleShift(I); }
1849 void visitLShr(BinaryOperator &I) { handleShift(I); }
1851 /// \brief Instrument llvm.memmove
1853 /// At this point we don't know if llvm.memmove will be inlined or not.
1854 /// If we don't instrument it and it gets inlined,
1855 /// our interceptor will not kick in and we will lose the memmove.
1856 /// If we instrument the call here, but it does not get inlined,
1857 /// we will memove the shadow twice: which is bad in case
1858 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1860 /// Similar situation exists for memcpy and memset.
1861 void visitMemMoveInst(MemMoveInst &I) {
1862 IRBuilder<> IRB(&I);
1865 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1866 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1867 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1868 I.eraseFromParent();
1871 // Similar to memmove: avoid copying shadow twice.
1872 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1873 // FIXME: consider doing manual inline for small constant sizes and proper
1875 void visitMemCpyInst(MemCpyInst &I) {
1876 IRBuilder<> IRB(&I);
1879 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1880 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1881 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1882 I.eraseFromParent();
1886 void visitMemSetInst(MemSetInst &I) {
1887 IRBuilder<> IRB(&I);
1890 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1891 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1892 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1893 I.eraseFromParent();
1896 void visitVAStartInst(VAStartInst &I) {
1897 VAHelper->visitVAStartInst(I);
1900 void visitVACopyInst(VACopyInst &I) {
1901 VAHelper->visitVACopyInst(I);
1904 /// \brief Handle vector store-like intrinsics.
1906 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1907 /// has 1 pointer argument and 1 vector argument, returns void.
1908 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1909 IRBuilder<> IRB(&I);
1910 Value* Addr = I.getArgOperand(0);
1911 Value *Shadow = getShadow(&I, 1);
1912 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1914 // We don't know the pointer alignment (could be unaligned SSE store!).
1915 // Have to assume to worst case.
1916 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1918 if (ClCheckAccessAddress)
1919 insertShadowCheck(Addr, &I);
1921 // FIXME: use ClStoreCleanOrigin
1922 // FIXME: factor out common code from materializeStores
1923 if (MS.TrackOrigins)
1924 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1928 /// \brief Handle vector load-like intrinsics.
1930 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1931 /// has 1 pointer argument, returns a vector.
1932 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1933 IRBuilder<> IRB(&I);
1934 Value *Addr = I.getArgOperand(0);
1936 Type *ShadowTy = getShadowTy(&I);
1937 if (PropagateShadow) {
1938 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1939 // We don't know the pointer alignment (could be unaligned SSE load!).
1940 // Have to assume to worst case.
1941 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1943 setShadow(&I, getCleanShadow(&I));
1946 if (ClCheckAccessAddress)
1947 insertShadowCheck(Addr, &I);
1949 if (MS.TrackOrigins) {
1950 if (PropagateShadow)
1951 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1953 setOrigin(&I, getCleanOrigin());
1958 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1960 /// Instrument intrinsics with any number of arguments of the same type,
1961 /// equal to the return type. The type should be simple (no aggregates or
1962 /// pointers; vectors are fine).
1963 /// Caller guarantees that this intrinsic does not access memory.
1964 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1965 Type *RetTy = I.getType();
1966 if (!(RetTy->isIntOrIntVectorTy() ||
1967 RetTy->isFPOrFPVectorTy() ||
1968 RetTy->isX86_MMXTy()))
1971 unsigned NumArgOperands = I.getNumArgOperands();
1973 for (unsigned i = 0; i < NumArgOperands; ++i) {
1974 Type *Ty = I.getArgOperand(i)->getType();
1979 IRBuilder<> IRB(&I);
1980 ShadowAndOriginCombiner SC(this, IRB);
1981 for (unsigned i = 0; i < NumArgOperands; ++i)
1982 SC.Add(I.getArgOperand(i));
1988 /// \brief Heuristically instrument unknown intrinsics.
1990 /// The main purpose of this code is to do something reasonable with all
1991 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1992 /// We recognize several classes of intrinsics by their argument types and
1993 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1994 /// sure that we know what the intrinsic does.
1996 /// We special-case intrinsics where this approach fails. See llvm.bswap
1997 /// handling as an example of that.
1998 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1999 unsigned NumArgOperands = I.getNumArgOperands();
2000 if (NumArgOperands == 0)
2003 if (NumArgOperands == 2 &&
2004 I.getArgOperand(0)->getType()->isPointerTy() &&
2005 I.getArgOperand(1)->getType()->isVectorTy() &&
2006 I.getType()->isVoidTy() &&
2007 !I.onlyReadsMemory()) {
2008 // This looks like a vector store.
2009 return handleVectorStoreIntrinsic(I);
2012 if (NumArgOperands == 1 &&
2013 I.getArgOperand(0)->getType()->isPointerTy() &&
2014 I.getType()->isVectorTy() &&
2015 I.onlyReadsMemory()) {
2016 // This looks like a vector load.
2017 return handleVectorLoadIntrinsic(I);
2020 if (I.doesNotAccessMemory())
2021 if (maybeHandleSimpleNomemIntrinsic(I))
2024 // FIXME: detect and handle SSE maskstore/maskload
2028 void handleBswap(IntrinsicInst &I) {
2029 IRBuilder<> IRB(&I);
2030 Value *Op = I.getArgOperand(0);
2031 Type *OpType = Op->getType();
2032 Function *BswapFunc = Intrinsic::getDeclaration(
2033 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
2034 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
2035 setOrigin(&I, getOrigin(Op));
2038 // \brief Instrument vector convert instrinsic.
2040 // This function instruments intrinsics like cvtsi2ss:
2041 // %Out = int_xxx_cvtyyy(%ConvertOp)
2043 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2044 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2045 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2046 // elements from \p CopyOp.
2047 // In most cases conversion involves floating-point value which may trigger a
2048 // hardware exception when not fully initialized. For this reason we require
2049 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2050 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2051 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2052 // return a fully initialized value.
2053 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2054 IRBuilder<> IRB(&I);
2055 Value *CopyOp, *ConvertOp;
2057 switch (I.getNumArgOperands()) {
2059 assert(isa<ConstantInt>(I.getArgOperand(2)) && "Invalid rounding mode");
2061 CopyOp = I.getArgOperand(0);
2062 ConvertOp = I.getArgOperand(1);
2065 ConvertOp = I.getArgOperand(0);
2069 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2072 // The first *NumUsedElements* elements of ConvertOp are converted to the
2073 // same number of output elements. The rest of the output is copied from
2074 // CopyOp, or (if not available) filled with zeroes.
2075 // Combine shadow for elements of ConvertOp that are used in this operation,
2076 // and insert a check.
2077 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2078 // int->any conversion.
2079 Value *ConvertShadow = getShadow(ConvertOp);
2080 Value *AggShadow = nullptr;
2081 if (ConvertOp->getType()->isVectorTy()) {
2082 AggShadow = IRB.CreateExtractElement(
2083 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2084 for (int i = 1; i < NumUsedElements; ++i) {
2085 Value *MoreShadow = IRB.CreateExtractElement(
2086 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2087 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2090 AggShadow = ConvertShadow;
2092 assert(AggShadow->getType()->isIntegerTy());
2093 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2095 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2098 assert(CopyOp->getType() == I.getType());
2099 assert(CopyOp->getType()->isVectorTy());
2100 Value *ResultShadow = getShadow(CopyOp);
2101 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2102 for (int i = 0; i < NumUsedElements; ++i) {
2103 ResultShadow = IRB.CreateInsertElement(
2104 ResultShadow, ConstantInt::getNullValue(EltTy),
2105 ConstantInt::get(IRB.getInt32Ty(), i));
2107 setShadow(&I, ResultShadow);
2108 setOrigin(&I, getOrigin(CopyOp));
2110 setShadow(&I, getCleanShadow(&I));
2111 setOrigin(&I, getCleanOrigin());
2115 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2116 // zeroes if it is zero, and all ones otherwise.
2117 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2118 if (S->getType()->isVectorTy())
2119 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2120 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2121 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2122 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2125 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2126 Type *T = S->getType();
2127 assert(T->isVectorTy());
2128 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2129 return IRB.CreateSExt(S2, T);
2132 // \brief Instrument vector shift instrinsic.
2134 // This function instruments intrinsics like int_x86_avx2_psll_w.
2135 // Intrinsic shifts %In by %ShiftSize bits.
2136 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2137 // size, and the rest is ignored. Behavior is defined even if shift size is
2138 // greater than register (or field) width.
2139 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2140 assert(I.getNumArgOperands() == 2);
2141 IRBuilder<> IRB(&I);
2142 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2143 // Otherwise perform the same shift on S1.
2144 Value *S1 = getShadow(&I, 0);
2145 Value *S2 = getShadow(&I, 1);
2146 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2147 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2148 Value *V1 = I.getOperand(0);
2149 Value *V2 = I.getOperand(1);
2150 Value *Shift = IRB.CreateCall(I.getCalledValue(),
2151 {IRB.CreateBitCast(S1, V1->getType()), V2});
2152 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2153 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2154 setOriginForNaryOp(I);
2157 // \brief Get an X86_MMX-sized vector type.
2158 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2159 const unsigned X86_MMXSizeInBits = 64;
2160 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2161 X86_MMXSizeInBits / EltSizeInBits);
2164 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2166 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2168 case llvm::Intrinsic::x86_sse2_packsswb_128:
2169 case llvm::Intrinsic::x86_sse2_packuswb_128:
2170 return llvm::Intrinsic::x86_sse2_packsswb_128;
2172 case llvm::Intrinsic::x86_sse2_packssdw_128:
2173 case llvm::Intrinsic::x86_sse41_packusdw:
2174 return llvm::Intrinsic::x86_sse2_packssdw_128;
2176 case llvm::Intrinsic::x86_avx2_packsswb:
2177 case llvm::Intrinsic::x86_avx2_packuswb:
2178 return llvm::Intrinsic::x86_avx2_packsswb;
2180 case llvm::Intrinsic::x86_avx2_packssdw:
2181 case llvm::Intrinsic::x86_avx2_packusdw:
2182 return llvm::Intrinsic::x86_avx2_packssdw;
2184 case llvm::Intrinsic::x86_mmx_packsswb:
2185 case llvm::Intrinsic::x86_mmx_packuswb:
2186 return llvm::Intrinsic::x86_mmx_packsswb;
2188 case llvm::Intrinsic::x86_mmx_packssdw:
2189 return llvm::Intrinsic::x86_mmx_packssdw;
2191 llvm_unreachable("unexpected intrinsic id");
2195 // \brief Instrument vector pack instrinsic.
2197 // This function instruments intrinsics like x86_mmx_packsswb, that
2198 // packs elements of 2 input vectors into half as many bits with saturation.
2199 // Shadow is propagated with the signed variant of the same intrinsic applied
2200 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2201 // EltSizeInBits is used only for x86mmx arguments.
2202 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2203 assert(I.getNumArgOperands() == 2);
2204 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2205 IRBuilder<> IRB(&I);
2206 Value *S1 = getShadow(&I, 0);
2207 Value *S2 = getShadow(&I, 1);
2208 assert(isX86_MMX || S1->getType()->isVectorTy());
2210 // SExt and ICmpNE below must apply to individual elements of input vectors.
2211 // In case of x86mmx arguments, cast them to appropriate vector types and
2213 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2215 S1 = IRB.CreateBitCast(S1, T);
2216 S2 = IRB.CreateBitCast(S2, T);
2218 Value *S1_ext = IRB.CreateSExt(
2219 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2220 Value *S2_ext = IRB.CreateSExt(
2221 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2223 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2224 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2225 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2228 Function *ShadowFn = Intrinsic::getDeclaration(
2229 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2232 IRB.CreateCall(ShadowFn, {S1_ext, S2_ext}, "_msprop_vector_pack");
2233 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2235 setOriginForNaryOp(I);
2238 // \brief Instrument sum-of-absolute-differencies intrinsic.
2239 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2240 const unsigned SignificantBitsPerResultElement = 16;
2241 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2242 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2243 unsigned ZeroBitsPerResultElement =
2244 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
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.CreateLShr(S, ZeroBitsPerResultElement);
2252 S = IRB.CreateBitCast(S, getShadowTy(&I));
2254 setOriginForNaryOp(I);
2257 // \brief Instrument multiply-add intrinsic.
2258 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2259 unsigned EltSizeInBits = 0) {
2260 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2261 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2262 IRBuilder<> IRB(&I);
2263 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2264 S = IRB.CreateBitCast(S, ResTy);
2265 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2267 S = IRB.CreateBitCast(S, getShadowTy(&I));
2269 setOriginForNaryOp(I);
2272 void visitIntrinsicInst(IntrinsicInst &I) {
2273 switch (I.getIntrinsicID()) {
2274 case llvm::Intrinsic::bswap:
2277 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2278 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2279 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2280 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2281 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2282 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2283 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2284 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2285 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2286 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2287 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2288 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2289 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2290 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2291 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2292 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2293 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2294 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2295 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2296 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2297 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2298 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2299 case llvm::Intrinsic::x86_sse_cvtss2si64:
2300 case llvm::Intrinsic::x86_sse_cvtss2si:
2301 case llvm::Intrinsic::x86_sse_cvttss2si64:
2302 case llvm::Intrinsic::x86_sse_cvttss2si:
2303 handleVectorConvertIntrinsic(I, 1);
2305 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2306 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2307 case llvm::Intrinsic::x86_sse_cvtps2pi:
2308 case llvm::Intrinsic::x86_sse_cvttps2pi:
2309 handleVectorConvertIntrinsic(I, 2);
2311 case llvm::Intrinsic::x86_avx2_psll_w:
2312 case llvm::Intrinsic::x86_avx2_psll_d:
2313 case llvm::Intrinsic::x86_avx2_psll_q:
2314 case llvm::Intrinsic::x86_avx2_pslli_w:
2315 case llvm::Intrinsic::x86_avx2_pslli_d:
2316 case llvm::Intrinsic::x86_avx2_pslli_q:
2317 case llvm::Intrinsic::x86_avx2_psrl_w:
2318 case llvm::Intrinsic::x86_avx2_psrl_d:
2319 case llvm::Intrinsic::x86_avx2_psrl_q:
2320 case llvm::Intrinsic::x86_avx2_psra_w:
2321 case llvm::Intrinsic::x86_avx2_psra_d:
2322 case llvm::Intrinsic::x86_avx2_psrli_w:
2323 case llvm::Intrinsic::x86_avx2_psrli_d:
2324 case llvm::Intrinsic::x86_avx2_psrli_q:
2325 case llvm::Intrinsic::x86_avx2_psrai_w:
2326 case llvm::Intrinsic::x86_avx2_psrai_d:
2327 case llvm::Intrinsic::x86_sse2_psll_w:
2328 case llvm::Intrinsic::x86_sse2_psll_d:
2329 case llvm::Intrinsic::x86_sse2_psll_q:
2330 case llvm::Intrinsic::x86_sse2_pslli_w:
2331 case llvm::Intrinsic::x86_sse2_pslli_d:
2332 case llvm::Intrinsic::x86_sse2_pslli_q:
2333 case llvm::Intrinsic::x86_sse2_psrl_w:
2334 case llvm::Intrinsic::x86_sse2_psrl_d:
2335 case llvm::Intrinsic::x86_sse2_psrl_q:
2336 case llvm::Intrinsic::x86_sse2_psra_w:
2337 case llvm::Intrinsic::x86_sse2_psra_d:
2338 case llvm::Intrinsic::x86_sse2_psrli_w:
2339 case llvm::Intrinsic::x86_sse2_psrli_d:
2340 case llvm::Intrinsic::x86_sse2_psrli_q:
2341 case llvm::Intrinsic::x86_sse2_psrai_w:
2342 case llvm::Intrinsic::x86_sse2_psrai_d:
2343 case llvm::Intrinsic::x86_mmx_psll_w:
2344 case llvm::Intrinsic::x86_mmx_psll_d:
2345 case llvm::Intrinsic::x86_mmx_psll_q:
2346 case llvm::Intrinsic::x86_mmx_pslli_w:
2347 case llvm::Intrinsic::x86_mmx_pslli_d:
2348 case llvm::Intrinsic::x86_mmx_pslli_q:
2349 case llvm::Intrinsic::x86_mmx_psrl_w:
2350 case llvm::Intrinsic::x86_mmx_psrl_d:
2351 case llvm::Intrinsic::x86_mmx_psrl_q:
2352 case llvm::Intrinsic::x86_mmx_psra_w:
2353 case llvm::Intrinsic::x86_mmx_psra_d:
2354 case llvm::Intrinsic::x86_mmx_psrli_w:
2355 case llvm::Intrinsic::x86_mmx_psrli_d:
2356 case llvm::Intrinsic::x86_mmx_psrli_q:
2357 case llvm::Intrinsic::x86_mmx_psrai_w:
2358 case llvm::Intrinsic::x86_mmx_psrai_d:
2359 handleVectorShiftIntrinsic(I, /* Variable */ false);
2361 case llvm::Intrinsic::x86_avx2_psllv_d:
2362 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2363 case llvm::Intrinsic::x86_avx2_psllv_q:
2364 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2365 case llvm::Intrinsic::x86_avx2_psrlv_d:
2366 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2367 case llvm::Intrinsic::x86_avx2_psrlv_q:
2368 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2369 case llvm::Intrinsic::x86_avx2_psrav_d:
2370 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2371 handleVectorShiftIntrinsic(I, /* Variable */ true);
2374 case llvm::Intrinsic::x86_sse2_packsswb_128:
2375 case llvm::Intrinsic::x86_sse2_packssdw_128:
2376 case llvm::Intrinsic::x86_sse2_packuswb_128:
2377 case llvm::Intrinsic::x86_sse41_packusdw:
2378 case llvm::Intrinsic::x86_avx2_packsswb:
2379 case llvm::Intrinsic::x86_avx2_packssdw:
2380 case llvm::Intrinsic::x86_avx2_packuswb:
2381 case llvm::Intrinsic::x86_avx2_packusdw:
2382 handleVectorPackIntrinsic(I);
2385 case llvm::Intrinsic::x86_mmx_packsswb:
2386 case llvm::Intrinsic::x86_mmx_packuswb:
2387 handleVectorPackIntrinsic(I, 16);
2390 case llvm::Intrinsic::x86_mmx_packssdw:
2391 handleVectorPackIntrinsic(I, 32);
2394 case llvm::Intrinsic::x86_mmx_psad_bw:
2395 case llvm::Intrinsic::x86_sse2_psad_bw:
2396 case llvm::Intrinsic::x86_avx2_psad_bw:
2397 handleVectorSadIntrinsic(I);
2400 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2401 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2402 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2403 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2404 handleVectorPmaddIntrinsic(I);
2407 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2408 handleVectorPmaddIntrinsic(I, 8);
2411 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2412 handleVectorPmaddIntrinsic(I, 16);
2416 if (!handleUnknownIntrinsic(I))
2417 visitInstruction(I);
2422 void visitCallSite(CallSite CS) {
2423 Instruction &I = *CS.getInstruction();
2424 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2426 CallInst *Call = cast<CallInst>(&I);
2428 // For inline asm, do the usual thing: check argument shadow and mark all
2429 // outputs as clean. Note that any side effects of the inline asm that are
2430 // not immediately visible in its constraints are not handled.
2431 if (Call->isInlineAsm()) {
2432 visitInstruction(I);
2436 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2438 // We are going to insert code that relies on the fact that the callee
2439 // will become a non-readonly function after it is instrumented by us. To
2440 // prevent this code from being optimized out, mark that function
2441 // non-readonly in advance.
2442 if (Function *Func = Call->getCalledFunction()) {
2443 // Clear out readonly/readnone attributes.
2445 B.addAttribute(Attribute::ReadOnly)
2446 .addAttribute(Attribute::ReadNone);
2447 Func->removeAttributes(AttributeSet::FunctionIndex,
2448 AttributeSet::get(Func->getContext(),
2449 AttributeSet::FunctionIndex,
2453 IRBuilder<> IRB(&I);
2455 unsigned ArgOffset = 0;
2456 DEBUG(dbgs() << " CallSite: " << I << "\n");
2457 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2458 ArgIt != End; ++ArgIt) {
2460 unsigned i = ArgIt - CS.arg_begin();
2461 if (!A->getType()->isSized()) {
2462 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2466 Value *Store = nullptr;
2467 // Compute the Shadow for arg even if it is ByVal, because
2468 // in that case getShadow() will copy the actual arg shadow to
2469 // __msan_param_tls.
2470 Value *ArgShadow = getShadow(A);
2471 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2472 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2473 " Shadow: " << *ArgShadow << "\n");
2474 bool ArgIsInitialized = false;
2475 const DataLayout &DL = F.getParent()->getDataLayout();
2476 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2477 assert(A->getType()->isPointerTy() &&
2478 "ByVal argument is not a pointer!");
2479 Size = DL.getTypeAllocSize(A->getType()->getPointerElementType());
2480 if (ArgOffset + Size > kParamTLSSize) break;
2481 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2482 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2483 Store = IRB.CreateMemCpy(ArgShadowBase,
2484 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2487 Size = DL.getTypeAllocSize(A->getType());
2488 if (ArgOffset + Size > kParamTLSSize) break;
2489 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2490 kShadowTLSAlignment);
2491 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2492 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2494 if (MS.TrackOrigins && !ArgIsInitialized)
2495 IRB.CreateStore(getOrigin(A),
2496 getOriginPtrForArgument(A, IRB, ArgOffset));
2498 assert(Size != 0 && Store != nullptr);
2499 DEBUG(dbgs() << " Param:" << *Store << "\n");
2500 ArgOffset += RoundUpToAlignment(Size, 8);
2502 DEBUG(dbgs() << " done with call args\n");
2505 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2506 if (FT->isVarArg()) {
2507 VAHelper->visitCallSite(CS, IRB);
2510 // Now, get the shadow for the RetVal.
2511 if (!I.getType()->isSized()) return;
2512 // Don't emit the epilogue for musttail call returns.
2513 if (CS.isCall() && cast<CallInst>(&I)->isMustTailCall()) return;
2514 IRBuilder<> IRBBefore(&I);
2515 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2516 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2517 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2518 BasicBlock::iterator NextInsn;
2520 NextInsn = ++I.getIterator();
2521 assert(NextInsn != I.getParent()->end());
2523 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2524 if (!NormalDest->getSinglePredecessor()) {
2525 // FIXME: this case is tricky, so we are just conservative here.
2526 // Perhaps we need to split the edge between this BB and NormalDest,
2527 // but a naive attempt to use SplitEdge leads to a crash.
2528 setShadow(&I, getCleanShadow(&I));
2529 setOrigin(&I, getCleanOrigin());
2532 NextInsn = NormalDest->getFirstInsertionPt();
2533 assert(NextInsn != NormalDest->end() &&
2534 "Could not find insertion point for retval shadow load");
2536 IRBuilder<> IRBAfter(&*NextInsn);
2537 Value *RetvalShadow =
2538 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2539 kShadowTLSAlignment, "_msret");
2540 setShadow(&I, RetvalShadow);
2541 if (MS.TrackOrigins)
2542 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2545 bool isAMustTailRetVal(Value *RetVal) {
2546 if (auto *I = dyn_cast<BitCastInst>(RetVal)) {
2547 RetVal = I->getOperand(0);
2549 if (auto *I = dyn_cast<CallInst>(RetVal)) {
2550 return I->isMustTailCall();
2555 void visitReturnInst(ReturnInst &I) {
2556 IRBuilder<> IRB(&I);
2557 Value *RetVal = I.getReturnValue();
2558 if (!RetVal) return;
2559 // Don't emit the epilogue for musttail call returns.
2560 if (isAMustTailRetVal(RetVal)) return;
2561 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2562 if (CheckReturnValue) {
2563 insertShadowCheck(RetVal, &I);
2564 Value *Shadow = getCleanShadow(RetVal);
2565 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2567 Value *Shadow = getShadow(RetVal);
2568 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2569 // FIXME: make it conditional if ClStoreCleanOrigin==0
2570 if (MS.TrackOrigins)
2571 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2575 void visitPHINode(PHINode &I) {
2576 IRBuilder<> IRB(&I);
2577 if (!PropagateShadow) {
2578 setShadow(&I, getCleanShadow(&I));
2579 setOrigin(&I, getCleanOrigin());
2583 ShadowPHINodes.push_back(&I);
2584 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2586 if (MS.TrackOrigins)
2587 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2591 void visitAllocaInst(AllocaInst &I) {
2592 setShadow(&I, getCleanShadow(&I));
2593 setOrigin(&I, getCleanOrigin());
2594 IRBuilder<> IRB(I.getNextNode());
2595 const DataLayout &DL = F.getParent()->getDataLayout();
2596 uint64_t Size = DL.getTypeAllocSize(I.getAllocatedType());
2597 if (PoisonStack && ClPoisonStackWithCall) {
2598 IRB.CreateCall(MS.MsanPoisonStackFn,
2599 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2600 ConstantInt::get(MS.IntptrTy, Size)});
2602 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2603 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2604 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2607 if (PoisonStack && MS.TrackOrigins) {
2608 SmallString<2048> StackDescriptionStorage;
2609 raw_svector_ostream StackDescription(StackDescriptionStorage);
2610 // We create a string with a description of the stack allocation and
2611 // pass it into __msan_set_alloca_origin.
2612 // It will be printed by the run-time if stack-originated UMR is found.
2613 // The first 4 bytes of the string are set to '----' and will be replaced
2614 // by __msan_va_arg_overflow_size_tls at the first call.
2615 StackDescription << "----" << I.getName() << "@" << F.getName();
2617 createPrivateNonConstGlobalForString(*F.getParent(),
2618 StackDescription.str());
2620 IRB.CreateCall(MS.MsanSetAllocaOrigin4Fn,
2621 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2622 ConstantInt::get(MS.IntptrTy, Size),
2623 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2624 IRB.CreatePointerCast(&F, MS.IntptrTy)});
2628 void visitSelectInst(SelectInst& I) {
2629 IRBuilder<> IRB(&I);
2630 // a = select b, c, d
2631 Value *B = I.getCondition();
2632 Value *C = I.getTrueValue();
2633 Value *D = I.getFalseValue();
2634 Value *Sb = getShadow(B);
2635 Value *Sc = getShadow(C);
2636 Value *Sd = getShadow(D);
2638 // Result shadow if condition shadow is 0.
2639 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2641 if (I.getType()->isAggregateType()) {
2642 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2643 // an extra "select". This results in much more compact IR.
2644 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2645 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2647 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2648 // If Sb (condition is poisoned), look for bits in c and d that are equal
2649 // and both unpoisoned.
2650 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2652 // Cast arguments to shadow-compatible type.
2653 C = CreateAppToShadowCast(IRB, C);
2654 D = CreateAppToShadowCast(IRB, D);
2656 // Result shadow if condition shadow is 1.
2657 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2659 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2661 if (MS.TrackOrigins) {
2662 // Origins are always i32, so any vector conditions must be flattened.
2663 // FIXME: consider tracking vector origins for app vectors?
2664 if (B->getType()->isVectorTy()) {
2665 Type *FlatTy = getShadowTyNoVec(B->getType());
2666 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2667 ConstantInt::getNullValue(FlatTy));
2668 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2669 ConstantInt::getNullValue(FlatTy));
2671 // a = select b, c, d
2672 // Oa = Sb ? Ob : (b ? Oc : Od)
2674 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2675 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2676 getOrigin(I.getFalseValue()))));
2680 void visitLandingPadInst(LandingPadInst &I) {
2682 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2683 setShadow(&I, getCleanShadow(&I));
2684 setOrigin(&I, getCleanOrigin());
2687 void visitCleanupPadInst(CleanupPadInst &I) {
2688 setShadow(&I, getCleanShadow(&I));
2689 setOrigin(&I, getCleanOrigin());
2692 void visitCatchPad(CatchPadInst &I) {
2693 setShadow(&I, getCleanShadow(&I));
2694 setOrigin(&I, getCleanOrigin());
2697 void visitTerminatePad(TerminatePadInst &I) {
2698 DEBUG(dbgs() << "TerminatePad: " << I << "\n");
2699 // Nothing to do here.
2702 void visitCatchEndPadInst(CatchEndPadInst &I) {
2703 DEBUG(dbgs() << "CatchEndPad: " << I << "\n");
2704 // Nothing to do here.
2707 void visitCleanupEndPadInst(CleanupEndPadInst &I) {
2708 DEBUG(dbgs() << "CleanupEndPad: " << I << "\n");
2709 // Nothing to do here.
2712 void visitGetElementPtrInst(GetElementPtrInst &I) {
2716 void visitExtractValueInst(ExtractValueInst &I) {
2717 IRBuilder<> IRB(&I);
2718 Value *Agg = I.getAggregateOperand();
2719 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2720 Value *AggShadow = getShadow(Agg);
2721 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2722 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2723 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2724 setShadow(&I, ResShadow);
2725 setOriginForNaryOp(I);
2728 void visitInsertValueInst(InsertValueInst &I) {
2729 IRBuilder<> IRB(&I);
2730 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2731 Value *AggShadow = getShadow(I.getAggregateOperand());
2732 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2733 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2734 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2735 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2736 DEBUG(dbgs() << " Res: " << *Res << "\n");
2738 setOriginForNaryOp(I);
2741 void dumpInst(Instruction &I) {
2742 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2743 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2745 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2747 errs() << "QQQ " << I << "\n";
2750 void visitResumeInst(ResumeInst &I) {
2751 DEBUG(dbgs() << "Resume: " << I << "\n");
2752 // Nothing to do here.
2755 void visitCleanupReturnInst(CleanupReturnInst &CRI) {
2756 DEBUG(dbgs() << "CleanupReturn: " << CRI << "\n");
2757 // Nothing to do here.
2760 void visitCatchReturnInst(CatchReturnInst &CRI) {
2761 DEBUG(dbgs() << "CatchReturn: " << CRI << "\n");
2762 // Nothing to do here.
2765 void visitInstruction(Instruction &I) {
2766 // Everything else: stop propagating and check for poisoned shadow.
2767 if (ClDumpStrictInstructions)
2769 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2770 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2771 insertShadowCheck(I.getOperand(i), &I);
2772 setShadow(&I, getCleanShadow(&I));
2773 setOrigin(&I, getCleanOrigin());
2777 /// \brief AMD64-specific implementation of VarArgHelper.
2778 struct VarArgAMD64Helper : public VarArgHelper {
2779 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2780 // See a comment in visitCallSite for more details.
2781 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2782 static const unsigned AMD64FpEndOffset = 176;
2785 MemorySanitizer &MS;
2786 MemorySanitizerVisitor &MSV;
2787 Value *VAArgTLSCopy;
2788 Value *VAArgOverflowSize;
2790 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2792 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2793 MemorySanitizerVisitor &MSV)
2794 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2795 VAArgOverflowSize(nullptr) {}
2797 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2799 ArgKind classifyArgument(Value* arg) {
2800 // A very rough approximation of X86_64 argument classification rules.
2801 Type *T = arg->getType();
2802 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2803 return AK_FloatingPoint;
2804 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2805 return AK_GeneralPurpose;
2806 if (T->isPointerTy())
2807 return AK_GeneralPurpose;
2811 // For VarArg functions, store the argument shadow in an ABI-specific format
2812 // that corresponds to va_list layout.
2813 // We do this because Clang lowers va_arg in the frontend, and this pass
2814 // only sees the low level code that deals with va_list internals.
2815 // A much easier alternative (provided that Clang emits va_arg instructions)
2816 // would have been to associate each live instance of va_list with a copy of
2817 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2819 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2820 unsigned GpOffset = 0;
2821 unsigned FpOffset = AMD64GpEndOffset;
2822 unsigned OverflowOffset = AMD64FpEndOffset;
2823 const DataLayout &DL = F.getParent()->getDataLayout();
2824 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2825 ArgIt != End; ++ArgIt) {
2827 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2828 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2830 // ByVal arguments always go to the overflow area.
2831 assert(A->getType()->isPointerTy());
2832 Type *RealTy = A->getType()->getPointerElementType();
2833 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
2834 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2835 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2836 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2837 ArgSize, kShadowTLSAlignment);
2839 ArgKind AK = classifyArgument(A);
2840 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2842 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2846 case AK_GeneralPurpose:
2847 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2850 case AK_FloatingPoint:
2851 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2855 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
2856 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2857 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2859 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2862 Constant *OverflowSize =
2863 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2864 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2867 /// \brief Compute the shadow address for a given va_arg.
2868 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2870 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2871 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2872 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2876 void visitVAStartInst(VAStartInst &I) override {
2877 if (F.getCallingConv() == CallingConv::X86_64_Win64)
2879 IRBuilder<> IRB(&I);
2880 VAStartInstrumentationList.push_back(&I);
2881 Value *VAListTag = I.getArgOperand(0);
2882 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2884 // Unpoison the whole __va_list_tag.
2885 // FIXME: magic ABI constants.
2886 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2887 /* size */24, /* alignment */8, false);
2890 void visitVACopyInst(VACopyInst &I) override {
2891 if (F.getCallingConv() == CallingConv::X86_64_Win64)
2893 IRBuilder<> IRB(&I);
2894 Value *VAListTag = I.getArgOperand(0);
2895 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2897 // Unpoison the whole __va_list_tag.
2898 // FIXME: magic ABI constants.
2899 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2900 /* size */24, /* alignment */8, false);
2903 void finalizeInstrumentation() override {
2904 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2905 "finalizeInstrumentation called twice");
2906 if (!VAStartInstrumentationList.empty()) {
2907 // If there is a va_start in this function, make a backup copy of
2908 // va_arg_tls somewhere in the function entry block.
2909 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2910 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2912 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2914 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2915 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2918 // Instrument va_start.
2919 // Copy va_list shadow from the backup copy of the TLS contents.
2920 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2921 CallInst *OrigInst = VAStartInstrumentationList[i];
2922 IRBuilder<> IRB(OrigInst->getNextNode());
2923 Value *VAListTag = OrigInst->getArgOperand(0);
2925 Value *RegSaveAreaPtrPtr =
2927 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2928 ConstantInt::get(MS.IntptrTy, 16)),
2929 Type::getInt64PtrTy(*MS.C));
2930 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2931 Value *RegSaveAreaShadowPtr =
2932 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2933 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2934 AMD64FpEndOffset, 16);
2936 Value *OverflowArgAreaPtrPtr =
2938 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2939 ConstantInt::get(MS.IntptrTy, 8)),
2940 Type::getInt64PtrTy(*MS.C));
2941 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2942 Value *OverflowArgAreaShadowPtr =
2943 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2944 Value *SrcPtr = IRB.CreateConstGEP1_32(IRB.getInt8Ty(), VAArgTLSCopy,
2946 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2951 /// \brief MIPS64-specific implementation of VarArgHelper.
2952 struct VarArgMIPS64Helper : public VarArgHelper {
2954 MemorySanitizer &MS;
2955 MemorySanitizerVisitor &MSV;
2956 Value *VAArgTLSCopy;
2959 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2961 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
2962 MemorySanitizerVisitor &MSV)
2963 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2964 VAArgSize(nullptr) {}
2966 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2967 unsigned VAArgOffset = 0;
2968 const DataLayout &DL = F.getParent()->getDataLayout();
2969 for (CallSite::arg_iterator ArgIt = CS.arg_begin() + 1, End = CS.arg_end();
2970 ArgIt != End; ++ArgIt) {
2973 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
2974 #if defined(__MIPSEB__) || defined(MIPSEB)
2975 // Adjusting the shadow for argument with size < 8 to match the placement
2976 // of bits in big endian system
2978 VAArgOffset += (8 - ArgSize);
2980 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
2981 VAArgOffset += ArgSize;
2982 VAArgOffset = RoundUpToAlignment(VAArgOffset, 8);
2983 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2986 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
2987 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
2988 // a new class member i.e. it is the total size of all VarArgs.
2989 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
2992 /// \brief Compute the shadow address for a given va_arg.
2993 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2995 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2996 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2997 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3001 void visitVAStartInst(VAStartInst &I) override {
3002 IRBuilder<> IRB(&I);
3003 VAStartInstrumentationList.push_back(&I);
3004 Value *VAListTag = I.getArgOperand(0);
3005 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3006 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3007 /* size */8, /* alignment */8, false);
3010 void visitVACopyInst(VACopyInst &I) override {
3011 IRBuilder<> IRB(&I);
3012 Value *VAListTag = I.getArgOperand(0);
3013 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3014 // Unpoison the whole __va_list_tag.
3015 // FIXME: magic ABI constants.
3016 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3017 /* size */8, /* alignment */8, false);
3020 void finalizeInstrumentation() override {
3021 assert(!VAArgSize && !VAArgTLSCopy &&
3022 "finalizeInstrumentation called twice");
3023 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3024 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3025 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3028 if (!VAStartInstrumentationList.empty()) {
3029 // If there is a va_start in this function, make a backup copy of
3030 // va_arg_tls somewhere in the function entry block.
3031 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3032 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3035 // Instrument va_start.
3036 // Copy va_list shadow from the backup copy of the TLS contents.
3037 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3038 CallInst *OrigInst = VAStartInstrumentationList[i];
3039 IRBuilder<> IRB(OrigInst->getNextNode());
3040 Value *VAListTag = OrigInst->getArgOperand(0);
3041 Value *RegSaveAreaPtrPtr =
3042 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3043 Type::getInt64PtrTy(*MS.C));
3044 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3045 Value *RegSaveAreaShadowPtr =
3046 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3047 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3052 /// \brief A no-op implementation of VarArgHelper.
3053 struct VarArgNoOpHelper : public VarArgHelper {
3054 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
3055 MemorySanitizerVisitor &MSV) {}
3057 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
3059 void visitVAStartInst(VAStartInst &I) override {}
3061 void visitVACopyInst(VACopyInst &I) override {}
3063 void finalizeInstrumentation() override {}
3066 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
3067 MemorySanitizerVisitor &Visitor) {
3068 // VarArg handling is only implemented on AMD64. False positives are possible
3069 // on other platforms.
3070 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
3071 if (TargetTriple.getArch() == llvm::Triple::x86_64)
3072 return new VarArgAMD64Helper(Func, Msan, Visitor);
3073 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
3074 TargetTriple.getArch() == llvm::Triple::mips64el)
3075 return new VarArgMIPS64Helper(Func, Msan, Visitor);
3077 return new VarArgNoOpHelper(Func, Msan, Visitor);
3080 } // anonymous namespace
3082 bool MemorySanitizer::runOnFunction(Function &F) {
3083 if (&F == MsanCtorFunction)
3085 MemorySanitizerVisitor Visitor(F, *this);
3087 // Clear out readonly/readnone attributes.
3089 B.addAttribute(Attribute::ReadOnly)
3090 .addAttribute(Attribute::ReadNone);
3091 F.removeAttributes(AttributeSet::FunctionIndex,
3092 AttributeSet::get(F.getContext(),
3093 AttributeSet::FunctionIndex, B));
3095 return Visitor.runOnFunction();