LLVM API Documentation

MemorySanitizer.cpp
Go to the documentation of this file.
00001 //===-- MemorySanitizer.cpp - detector of uninitialized reads -------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 /// \file
00010 /// This file is a part of MemorySanitizer, a detector of uninitialized
00011 /// reads.
00012 ///
00013 /// The algorithm of the tool is similar to Memcheck
00014 /// (http://goo.gl/QKbem). We associate a few shadow bits with every
00015 /// byte of the application memory, poison the shadow of the malloc-ed
00016 /// or alloca-ed memory, load the shadow bits on every memory read,
00017 /// propagate the shadow bits through some of the arithmetic
00018 /// instruction (including MOV), store the shadow bits on every memory
00019 /// write, report a bug on some other instructions (e.g. JMP) if the
00020 /// associated shadow is poisoned.
00021 ///
00022 /// But there are differences too. The first and the major one:
00023 /// compiler instrumentation instead of binary instrumentation. This
00024 /// gives us much better register allocation, possible compiler
00025 /// optimizations and a fast start-up. But this brings the major issue
00026 /// as well: msan needs to see all program events, including system
00027 /// calls and reads/writes in system libraries, so we either need to
00028 /// compile *everything* with msan or use a binary translation
00029 /// component (e.g. DynamoRIO) to instrument pre-built libraries.
00030 /// Another difference from Memcheck is that we use 8 shadow bits per
00031 /// byte of application memory and use a direct shadow mapping. This
00032 /// greatly simplifies the instrumentation code and avoids races on
00033 /// shadow updates (Memcheck is single-threaded so races are not a
00034 /// concern there. Memcheck uses 2 shadow bits per byte with a slow
00035 /// path storage that uses 8 bits per byte).
00036 ///
00037 /// The default value of shadow is 0, which means "clean" (not poisoned).
00038 ///
00039 /// Every module initializer should call __msan_init to ensure that the
00040 /// shadow memory is ready. On error, __msan_warning is called. Since
00041 /// parameters and return values may be passed via registers, we have a
00042 /// specialized thread-local shadow for return values
00043 /// (__msan_retval_tls) and parameters (__msan_param_tls).
00044 ///
00045 ///                           Origin tracking.
00046 ///
00047 /// MemorySanitizer can track origins (allocation points) of all uninitialized
00048 /// values. This behavior is controlled with a flag (msan-track-origins) and is
00049 /// disabled by default.
00050 ///
00051 /// Origins are 4-byte values created and interpreted by the runtime library.
00052 /// They are stored in a second shadow mapping, one 4-byte value for 4 bytes
00053 /// of application memory. Propagation of origins is basically a bunch of
00054 /// "select" instructions that pick the origin of a dirty argument, if an
00055 /// instruction has one.
00056 ///
00057 /// Every 4 aligned, consecutive bytes of application memory have one origin
00058 /// value associated with them. If these bytes contain uninitialized data
00059 /// coming from 2 different allocations, the last store wins. Because of this,
00060 /// MemorySanitizer reports can show unrelated origins, but this is unlikely in
00061 /// practice.
00062 ///
00063 /// Origins are meaningless for fully initialized values, so MemorySanitizer
00064 /// avoids storing origin to memory when a fully initialized value is stored.
00065 /// This way it avoids needless overwritting origin of the 4-byte region on
00066 /// a short (i.e. 1 byte) clean store, and it is also good for performance.
00067 ///
00068 ///                            Atomic handling.
00069 ///
00070 /// Ideally, every atomic store of application value should update the
00071 /// corresponding shadow location in an atomic way. Unfortunately, atomic store
00072 /// of two disjoint locations can not be done without severe slowdown.
00073 ///
00074 /// Therefore, we implement an approximation that may err on the safe side.
00075 /// In this implementation, every atomically accessed location in the program
00076 /// may only change from (partially) uninitialized to fully initialized, but
00077 /// not the other way around. We load the shadow _after_ the application load,
00078 /// and we store the shadow _before_ the app store. Also, we always store clean
00079 /// shadow (if the application store is atomic). This way, if the store-load
00080 /// pair constitutes a happens-before arc, shadow store and load are correctly
00081 /// ordered such that the load will get either the value that was stored, or
00082 /// some later value (which is always clean).
00083 ///
00084 /// This does not work very well with Compare-And-Swap (CAS) and
00085 /// Read-Modify-Write (RMW) operations. To follow the above logic, CAS and RMW
00086 /// must store the new shadow before the app operation, and load the shadow
00087 /// after the app operation. Computers don't work this way. Current
00088 /// implementation ignores the load aspect of CAS/RMW, always returning a clean
00089 /// value. It implements the store part as a simple atomic store by storing a
00090 /// clean shadow.
00091 
00092 //===----------------------------------------------------------------------===//
00093 
00094 #include "llvm/Transforms/Instrumentation.h"
00095 #include "llvm/ADT/DepthFirstIterator.h"
00096 #include "llvm/ADT/SmallString.h"
00097 #include "llvm/ADT/SmallVector.h"
00098 #include "llvm/ADT/StringExtras.h"
00099 #include "llvm/ADT/Triple.h"
00100 #include "llvm/IR/DataLayout.h"
00101 #include "llvm/IR/Function.h"
00102 #include "llvm/IR/IRBuilder.h"
00103 #include "llvm/IR/InlineAsm.h"
00104 #include "llvm/IR/InstVisitor.h"
00105 #include "llvm/IR/IntrinsicInst.h"
00106 #include "llvm/IR/LLVMContext.h"
00107 #include "llvm/IR/MDBuilder.h"
00108 #include "llvm/IR/Module.h"
00109 #include "llvm/IR/Type.h"
00110 #include "llvm/IR/ValueMap.h"
00111 #include "llvm/Support/CommandLine.h"
00112 #include "llvm/Support/Compiler.h"
00113 #include "llvm/Support/Debug.h"
00114 #include "llvm/Support/raw_ostream.h"
00115 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
00116 #include "llvm/Transforms/Utils/Local.h"
00117 #include "llvm/Transforms/Utils/ModuleUtils.h"
00118 
00119 using namespace llvm;
00120 
00121 #define DEBUG_TYPE "msan"
00122 
00123 static const uint64_t kShadowMask32 = 1ULL << 31;
00124 static const uint64_t kShadowMask64 = 1ULL << 46;
00125 static const uint64_t kOriginOffset32 = 1ULL << 30;
00126 static const uint64_t kOriginOffset64 = 1ULL << 45;
00127 static const unsigned kMinOriginAlignment = 4;
00128 static const unsigned kShadowTLSAlignment = 8;
00129 
00130 // Accesses sizes are powers of two: 1, 2, 4, 8.
00131 static const size_t kNumberOfAccessSizes = 4;
00132 
00133 /// \brief Track origins of uninitialized values.
00134 ///
00135 /// Adds a section to MemorySanitizer report that points to the allocation
00136 /// (stack or heap) the uninitialized bits came from originally.
00137 static cl::opt<int> ClTrackOrigins("msan-track-origins",
00138        cl::desc("Track origins (allocation sites) of poisoned memory"),
00139        cl::Hidden, cl::init(0));
00140 static cl::opt<bool> ClKeepGoing("msan-keep-going",
00141        cl::desc("keep going after reporting a UMR"),
00142        cl::Hidden, cl::init(false));
00143 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
00144        cl::desc("poison uninitialized stack variables"),
00145        cl::Hidden, cl::init(true));
00146 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
00147        cl::desc("poison uninitialized stack variables with a call"),
00148        cl::Hidden, cl::init(false));
00149 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
00150        cl::desc("poison uninitialized stack variables with the given patter"),
00151        cl::Hidden, cl::init(0xff));
00152 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
00153        cl::desc("poison undef temps"),
00154        cl::Hidden, cl::init(true));
00155 
00156 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
00157        cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
00158        cl::Hidden, cl::init(true));
00159 
00160 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
00161        cl::desc("exact handling of relational integer ICmp"),
00162        cl::Hidden, cl::init(false));
00163 
00164 // This flag controls whether we check the shadow of the address
00165 // operand of load or store. Such bugs are very rare, since load from
00166 // a garbage address typically results in SEGV, but still happen
00167 // (e.g. only lower bits of address are garbage, or the access happens
00168 // early at program startup where malloc-ed memory is more likely to
00169 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
00170 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
00171        cl::desc("report accesses through a pointer which has poisoned shadow"),
00172        cl::Hidden, cl::init(true));
00173 
00174 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
00175        cl::desc("print out instructions with default strict semantics"),
00176        cl::Hidden, cl::init(false));
00177 
00178 static cl::opt<int> ClInstrumentationWithCallThreshold(
00179     "msan-instrumentation-with-call-threshold",
00180     cl::desc(
00181         "If the function being instrumented requires more than "
00182         "this number of checks and origin stores, use callbacks instead of "
00183         "inline checks (-1 means never use callbacks)."),
00184     cl::Hidden, cl::init(3500));
00185 
00186 // Experimental. Wraps all indirect calls in the instrumented code with
00187 // a call to the given function. This is needed to assist the dynamic
00188 // helper tool (MSanDR) to regain control on transition between instrumented and
00189 // non-instrumented code.
00190 static cl::opt<std::string> ClWrapIndirectCalls("msan-wrap-indirect-calls",
00191        cl::desc("Wrap indirect calls with a given function"),
00192        cl::Hidden);
00193 
00194 static cl::opt<bool> ClWrapIndirectCallsFast("msan-wrap-indirect-calls-fast",
00195        cl::desc("Do not wrap indirect calls with target in the same module"),
00196        cl::Hidden, cl::init(true));
00197 
00198 namespace {
00199 
00200 /// \brief An instrumentation pass implementing detection of uninitialized
00201 /// reads.
00202 ///
00203 /// MemorySanitizer: instrument the code in module to find
00204 /// uninitialized reads.
00205 class MemorySanitizer : public FunctionPass {
00206  public:
00207   MemorySanitizer(int TrackOrigins = 0)
00208       : FunctionPass(ID),
00209         TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
00210         DL(nullptr),
00211         WarningFn(nullptr),
00212         WrapIndirectCalls(!ClWrapIndirectCalls.empty()) {}
00213   const char *getPassName() const override { return "MemorySanitizer"; }
00214   bool runOnFunction(Function &F) override;
00215   bool doInitialization(Module &M) override;
00216   static char ID;  // Pass identification, replacement for typeid.
00217 
00218  private:
00219   void initializeCallbacks(Module &M);
00220 
00221   /// \brief Track origins (allocation points) of uninitialized values.
00222   int TrackOrigins;
00223 
00224   const DataLayout *DL;
00225   LLVMContext *C;
00226   Type *IntptrTy;
00227   Type *OriginTy;
00228   /// \brief Thread-local shadow storage for function parameters.
00229   GlobalVariable *ParamTLS;
00230   /// \brief Thread-local origin storage for function parameters.
00231   GlobalVariable *ParamOriginTLS;
00232   /// \brief Thread-local shadow storage for function return value.
00233   GlobalVariable *RetvalTLS;
00234   /// \brief Thread-local origin storage for function return value.
00235   GlobalVariable *RetvalOriginTLS;
00236   /// \brief Thread-local shadow storage for in-register va_arg function
00237   /// parameters (x86_64-specific).
00238   GlobalVariable *VAArgTLS;
00239   /// \brief Thread-local shadow storage for va_arg overflow area
00240   /// (x86_64-specific).
00241   GlobalVariable *VAArgOverflowSizeTLS;
00242   /// \brief Thread-local space used to pass origin value to the UMR reporting
00243   /// function.
00244   GlobalVariable *OriginTLS;
00245 
00246   GlobalVariable *MsandrModuleStart;
00247   GlobalVariable *MsandrModuleEnd;
00248 
00249   /// \brief The run-time callback to print a warning.
00250   Value *WarningFn;
00251   // These arrays are indexed by log2(AccessSize).
00252   Value *MaybeWarningFn[kNumberOfAccessSizes];
00253   Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
00254 
00255   /// \brief Run-time helper that generates a new origin value for a stack
00256   /// allocation.
00257   Value *MsanSetAllocaOrigin4Fn;
00258   /// \brief Run-time helper that poisons stack on function entry.
00259   Value *MsanPoisonStackFn;
00260   /// \brief Run-time helper that records a store (or any event) of an
00261   /// uninitialized value and returns an updated origin id encoding this info.
00262   Value *MsanChainOriginFn;
00263   /// \brief MSan runtime replacements for memmove, memcpy and memset.
00264   Value *MemmoveFn, *MemcpyFn, *MemsetFn;
00265 
00266   /// \brief Address mask used in application-to-shadow address calculation.
00267   /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
00268   uint64_t ShadowMask;
00269   /// \brief Offset of the origin shadow from the "normal" shadow.
00270   /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
00271   uint64_t OriginOffset;
00272   /// \brief Branch weights for error reporting.
00273   MDNode *ColdCallWeights;
00274   /// \brief Branch weights for origin store.
00275   MDNode *OriginStoreWeights;
00276   /// \brief An empty volatile inline asm that prevents callback merge.
00277   InlineAsm *EmptyAsm;
00278 
00279   bool WrapIndirectCalls;
00280   /// \brief Run-time wrapper for indirect calls.
00281   Value *IndirectCallWrapperFn;
00282   // Argument and return type of IndirectCallWrapperFn: void (*f)(void).
00283   Type *AnyFunctionPtrTy;
00284 
00285   friend struct MemorySanitizerVisitor;
00286   friend struct VarArgAMD64Helper;
00287 };
00288 }  // namespace
00289 
00290 char MemorySanitizer::ID = 0;
00291 INITIALIZE_PASS(MemorySanitizer, "msan",
00292                 "MemorySanitizer: detects uninitialized reads.",
00293                 false, false)
00294 
00295 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
00296   return new MemorySanitizer(TrackOrigins);
00297 }
00298 
00299 /// \brief Create a non-const global initialized with the given string.
00300 ///
00301 /// Creates a writable global for Str so that we can pass it to the
00302 /// run-time lib. Runtime uses first 4 bytes of the string to store the
00303 /// frame ID, so the string needs to be mutable.
00304 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
00305                                                             StringRef Str) {
00306   Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
00307   return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
00308                             GlobalValue::PrivateLinkage, StrConst, "");
00309 }
00310 
00311 
00312 /// \brief Insert extern declaration of runtime-provided functions and globals.
00313 void MemorySanitizer::initializeCallbacks(Module &M) {
00314   // Only do this once.
00315   if (WarningFn)
00316     return;
00317 
00318   IRBuilder<> IRB(*C);
00319   // Create the callback.
00320   // FIXME: this function should have "Cold" calling conv,
00321   // which is not yet implemented.
00322   StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
00323                                         : "__msan_warning_noreturn";
00324   WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
00325 
00326   for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
00327        AccessSizeIndex++) {
00328     unsigned AccessSize = 1 << AccessSizeIndex;
00329     std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
00330     MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
00331         FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
00332         IRB.getInt32Ty(), NULL);
00333 
00334     FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
00335     MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
00336         FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
00337         IRB.getInt8PtrTy(), IRB.getInt32Ty(), NULL);
00338   }
00339 
00340   MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
00341     "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
00342     IRB.getInt8PtrTy(), IntptrTy, NULL);
00343   MsanPoisonStackFn = M.getOrInsertFunction(
00344     "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
00345   MsanChainOriginFn = M.getOrInsertFunction(
00346     "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), NULL);
00347   MemmoveFn = M.getOrInsertFunction(
00348     "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
00349     IRB.getInt8PtrTy(), IntptrTy, NULL);
00350   MemcpyFn = M.getOrInsertFunction(
00351     "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
00352     IntptrTy, NULL);
00353   MemsetFn = M.getOrInsertFunction(
00354     "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
00355     IntptrTy, NULL);
00356 
00357   // Create globals.
00358   RetvalTLS = new GlobalVariable(
00359     M, ArrayType::get(IRB.getInt64Ty(), 8), false,
00360     GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
00361     GlobalVariable::InitialExecTLSModel);
00362   RetvalOriginTLS = new GlobalVariable(
00363     M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
00364     "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
00365 
00366   ParamTLS = new GlobalVariable(
00367     M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
00368     GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
00369     GlobalVariable::InitialExecTLSModel);
00370   ParamOriginTLS = new GlobalVariable(
00371     M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
00372     nullptr, "__msan_param_origin_tls", nullptr,
00373     GlobalVariable::InitialExecTLSModel);
00374 
00375   VAArgTLS = new GlobalVariable(
00376     M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
00377     GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
00378     GlobalVariable::InitialExecTLSModel);
00379   VAArgOverflowSizeTLS = new GlobalVariable(
00380     M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
00381     "__msan_va_arg_overflow_size_tls", nullptr,
00382     GlobalVariable::InitialExecTLSModel);
00383   OriginTLS = new GlobalVariable(
00384     M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
00385     "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
00386 
00387   // We insert an empty inline asm after __msan_report* to avoid callback merge.
00388   EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
00389                             StringRef(""), StringRef(""),
00390                             /*hasSideEffects=*/true);
00391 
00392   if (WrapIndirectCalls) {
00393     AnyFunctionPtrTy =
00394         PointerType::getUnqual(FunctionType::get(IRB.getVoidTy(), false));
00395     IndirectCallWrapperFn = M.getOrInsertFunction(
00396         ClWrapIndirectCalls, AnyFunctionPtrTy, AnyFunctionPtrTy, NULL);
00397   }
00398 
00399   if (WrapIndirectCalls && ClWrapIndirectCallsFast) {
00400     MsandrModuleStart = new GlobalVariable(
00401         M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
00402         nullptr, "__executable_start");
00403     MsandrModuleStart->setVisibility(GlobalVariable::HiddenVisibility);
00404     MsandrModuleEnd = new GlobalVariable(
00405         M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
00406         nullptr, "_end");
00407     MsandrModuleEnd->setVisibility(GlobalVariable::HiddenVisibility);
00408   }
00409 }
00410 
00411 /// \brief Module-level initialization.
00412 ///
00413 /// inserts a call to __msan_init to the module's constructor list.
00414 bool MemorySanitizer::doInitialization(Module &M) {
00415   DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
00416   if (!DLP)
00417     report_fatal_error("data layout missing");
00418   DL = &DLP->getDataLayout();
00419 
00420   C = &(M.getContext());
00421   unsigned PtrSize = DL->getPointerSizeInBits(/* AddressSpace */0);
00422   switch (PtrSize) {
00423     case 64:
00424       ShadowMask = kShadowMask64;
00425       OriginOffset = kOriginOffset64;
00426       break;
00427     case 32:
00428       ShadowMask = kShadowMask32;
00429       OriginOffset = kOriginOffset32;
00430       break;
00431     default:
00432       report_fatal_error("unsupported pointer size");
00433       break;
00434   }
00435 
00436   IRBuilder<> IRB(*C);
00437   IntptrTy = IRB.getIntPtrTy(DL);
00438   OriginTy = IRB.getInt32Ty();
00439 
00440   ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
00441   OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
00442 
00443   // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
00444   appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
00445                       "__msan_init", IRB.getVoidTy(), NULL)), 0);
00446 
00447   if (TrackOrigins)
00448     new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
00449                        IRB.getInt32(TrackOrigins), "__msan_track_origins");
00450 
00451   if (ClKeepGoing)
00452     new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
00453                        IRB.getInt32(ClKeepGoing), "__msan_keep_going");
00454 
00455   return true;
00456 }
00457 
00458 namespace {
00459 
00460 /// \brief A helper class that handles instrumentation of VarArg
00461 /// functions on a particular platform.
00462 ///
00463 /// Implementations are expected to insert the instrumentation
00464 /// necessary to propagate argument shadow through VarArg function
00465 /// calls. Visit* methods are called during an InstVisitor pass over
00466 /// the function, and should avoid creating new basic blocks. A new
00467 /// instance of this class is created for each instrumented function.
00468 struct VarArgHelper {
00469   /// \brief Visit a CallSite.
00470   virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
00471 
00472   /// \brief Visit a va_start call.
00473   virtual void visitVAStartInst(VAStartInst &I) = 0;
00474 
00475   /// \brief Visit a va_copy call.
00476   virtual void visitVACopyInst(VACopyInst &I) = 0;
00477 
00478   /// \brief Finalize function instrumentation.
00479   ///
00480   /// This method is called after visiting all interesting (see above)
00481   /// instructions in a function.
00482   virtual void finalizeInstrumentation() = 0;
00483 
00484   virtual ~VarArgHelper() {}
00485 };
00486 
00487 struct MemorySanitizerVisitor;
00488 
00489 VarArgHelper*
00490 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
00491                    MemorySanitizerVisitor &Visitor);
00492 
00493 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
00494   if (TypeSize <= 8) return 0;
00495   return Log2_32_Ceil(TypeSize / 8);
00496 }
00497 
00498 /// This class does all the work for a given function. Store and Load
00499 /// instructions store and load corresponding shadow and origin
00500 /// values. Most instructions propagate shadow from arguments to their
00501 /// return values. Certain instructions (most importantly, BranchInst)
00502 /// test their argument shadow and print reports (with a runtime call) if it's
00503 /// non-zero.
00504 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
00505   Function &F;
00506   MemorySanitizer &MS;
00507   SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
00508   ValueMap<Value*, Value*> ShadowMap, OriginMap;
00509   std::unique_ptr<VarArgHelper> VAHelper;
00510 
00511   // The following flags disable parts of MSan instrumentation based on
00512   // blacklist contents and command-line options.
00513   bool InsertChecks;
00514   bool PropagateShadow;
00515   bool PoisonStack;
00516   bool PoisonUndef;
00517   bool CheckReturnValue;
00518 
00519   struct ShadowOriginAndInsertPoint {
00520     Value *Shadow;
00521     Value *Origin;
00522     Instruction *OrigIns;
00523     ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
00524       : Shadow(S), Origin(O), OrigIns(I) { }
00525   };
00526   SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
00527   SmallVector<Instruction*, 16> StoreList;
00528   SmallVector<CallSite, 16> IndirectCallList;
00529 
00530   MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
00531       : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
00532     bool SanitizeFunction = F.getAttributes().hasAttribute(
00533         AttributeSet::FunctionIndex, Attribute::SanitizeMemory);
00534     InsertChecks = SanitizeFunction;
00535     PropagateShadow = SanitizeFunction;
00536     PoisonStack = SanitizeFunction && ClPoisonStack;
00537     PoisonUndef = SanitizeFunction && ClPoisonUndef;
00538     // FIXME: Consider using SpecialCaseList to specify a list of functions that
00539     // must always return fully initialized values. For now, we hardcode "main".
00540     CheckReturnValue = SanitizeFunction && (F.getName() == "main");
00541 
00542     DEBUG(if (!InsertChecks)
00543           dbgs() << "MemorySanitizer is not inserting checks into '"
00544                  << F.getName() << "'\n");
00545   }
00546 
00547   Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
00548     if (MS.TrackOrigins <= 1) return V;
00549     return IRB.CreateCall(MS.MsanChainOriginFn, V);
00550   }
00551 
00552   void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
00553                    unsigned Alignment, bool AsCall) {
00554     if (isa<StructType>(Shadow->getType())) {
00555       IRB.CreateAlignedStore(updateOrigin(Origin, IRB), getOriginPtr(Addr, IRB),
00556                              Alignment);
00557     } else {
00558       Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
00559       // TODO(eugenis): handle non-zero constant shadow by inserting an
00560       // unconditional check (can not simply fail compilation as this could
00561       // be in the dead code).
00562       if (isa<Constant>(ConvertedShadow)) return;
00563       unsigned TypeSizeInBits =
00564           MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
00565       unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
00566       if (AsCall && SizeIndex < kNumberOfAccessSizes) {
00567         Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
00568         Value *ConvertedShadow2 = IRB.CreateZExt(
00569             ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
00570         IRB.CreateCall3(Fn, ConvertedShadow2,
00571                         IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
00572                         Origin);
00573       } else {
00574         Value *Cmp = IRB.CreateICmpNE(
00575             ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
00576         Instruction *CheckTerm = SplitBlockAndInsertIfThen(
00577             Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
00578         IRBuilder<> IRBNew(CheckTerm);
00579         IRBNew.CreateAlignedStore(updateOrigin(Origin, IRBNew),
00580                                   getOriginPtr(Addr, IRBNew), Alignment);
00581       }
00582     }
00583   }
00584 
00585   void materializeStores(bool InstrumentWithCalls) {
00586     for (auto Inst : StoreList) {
00587       StoreInst &SI = *dyn_cast<StoreInst>(Inst);
00588 
00589       IRBuilder<> IRB(&SI);
00590       Value *Val = SI.getValueOperand();
00591       Value *Addr = SI.getPointerOperand();
00592       Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
00593       Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
00594 
00595       StoreInst *NewSI =
00596           IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
00597       DEBUG(dbgs() << "  STORE: " << *NewSI << "\n");
00598       (void)NewSI;
00599 
00600       if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
00601 
00602       if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
00603 
00604       if (MS.TrackOrigins) {
00605         unsigned Alignment = std::max(kMinOriginAlignment, SI.getAlignment());
00606         storeOrigin(IRB, Addr, Shadow, getOrigin(Val), Alignment,
00607                     InstrumentWithCalls);
00608       }
00609     }
00610   }
00611 
00612   void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
00613                            bool AsCall) {
00614     IRBuilder<> IRB(OrigIns);
00615     DEBUG(dbgs() << "  SHAD0 : " << *Shadow << "\n");
00616     Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
00617     DEBUG(dbgs() << "  SHAD1 : " << *ConvertedShadow << "\n");
00618     // See the comment in materializeStores().
00619     if (isa<Constant>(ConvertedShadow)) return;
00620     unsigned TypeSizeInBits =
00621         MS.DL->getTypeSizeInBits(ConvertedShadow->getType());
00622     unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
00623     if (AsCall && SizeIndex < kNumberOfAccessSizes) {
00624       Value *Fn = MS.MaybeWarningFn[SizeIndex];
00625       Value *ConvertedShadow2 =
00626           IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
00627       IRB.CreateCall2(Fn, ConvertedShadow2, MS.TrackOrigins && Origin
00628                                                 ? Origin
00629                                                 : (Value *)IRB.getInt32(0));
00630     } else {
00631       Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
00632                                     getCleanShadow(ConvertedShadow), "_mscmp");
00633       Instruction *CheckTerm = SplitBlockAndInsertIfThen(
00634           Cmp, OrigIns,
00635           /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
00636 
00637       IRB.SetInsertPoint(CheckTerm);
00638       if (MS.TrackOrigins) {
00639         IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
00640                         MS.OriginTLS);
00641       }
00642       IRB.CreateCall(MS.WarningFn);
00643       IRB.CreateCall(MS.EmptyAsm);
00644       DEBUG(dbgs() << "  CHECK: " << *Cmp << "\n");
00645     }
00646   }
00647 
00648   void materializeChecks(bool InstrumentWithCalls) {
00649     for (const auto &ShadowData : InstrumentationList) {
00650       Instruction *OrigIns = ShadowData.OrigIns;
00651       Value *Shadow = ShadowData.Shadow;
00652       Value *Origin = ShadowData.Origin;
00653       materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
00654     }
00655     DEBUG(dbgs() << "DONE:\n" << F);
00656   }
00657 
00658   void materializeIndirectCalls() {
00659     for (auto &CS : IndirectCallList) {
00660       Instruction *I = CS.getInstruction();
00661       BasicBlock *B = I->getParent();
00662       IRBuilder<> IRB(I);
00663       Value *Fn0 = CS.getCalledValue();
00664       Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
00665 
00666       if (ClWrapIndirectCallsFast) {
00667         // Check that call target is inside this module limits.
00668         Value *Start =
00669             IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
00670         Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
00671 
00672         Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
00673                                               IRB.CreateICmpUGE(Fn, End));
00674 
00675         PHINode *NewFnPhi =
00676             IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
00677 
00678         Instruction *CheckTerm = SplitBlockAndInsertIfThen(
00679             NotInThisModule, NewFnPhi,
00680             /* Unreachable */ false, MS.ColdCallWeights);
00681 
00682         IRB.SetInsertPoint(CheckTerm);
00683         // Slow path: call wrapper function to possibly transform the call
00684         // target.
00685         Value *NewFn = IRB.CreateBitCast(
00686             IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
00687 
00688         NewFnPhi->addIncoming(Fn0, B);
00689         NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
00690         CS.setCalledFunction(NewFnPhi);
00691       } else {
00692         Value *NewFn = IRB.CreateBitCast(
00693             IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
00694         CS.setCalledFunction(NewFn);
00695       }
00696     }
00697   }
00698 
00699   /// \brief Add MemorySanitizer instrumentation to a function.
00700   bool runOnFunction() {
00701     MS.initializeCallbacks(*F.getParent());
00702     if (!MS.DL) return false;
00703 
00704     // In the presence of unreachable blocks, we may see Phi nodes with
00705     // incoming nodes from such blocks. Since InstVisitor skips unreachable
00706     // blocks, such nodes will not have any shadow value associated with them.
00707     // It's easier to remove unreachable blocks than deal with missing shadow.
00708     removeUnreachableBlocks(F);
00709 
00710     // Iterate all BBs in depth-first order and create shadow instructions
00711     // for all instructions (where applicable).
00712     // For PHI nodes we create dummy shadow PHIs which will be finalized later.
00713     for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
00714       visit(*BB);
00715 
00716 
00717     // Finalize PHI nodes.
00718     for (PHINode *PN : ShadowPHINodes) {
00719       PHINode *PNS = cast<PHINode>(getShadow(PN));
00720       PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
00721       size_t NumValues = PN->getNumIncomingValues();
00722       for (size_t v = 0; v < NumValues; v++) {
00723         PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
00724         if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
00725       }
00726     }
00727 
00728     VAHelper->finalizeInstrumentation();
00729 
00730     bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
00731                                InstrumentationList.size() + StoreList.size() >
00732                                    (unsigned)ClInstrumentationWithCallThreshold;
00733 
00734     // Delayed instrumentation of StoreInst.
00735     // This may add new checks to be inserted later.
00736     materializeStores(InstrumentWithCalls);
00737 
00738     // Insert shadow value checks.
00739     materializeChecks(InstrumentWithCalls);
00740 
00741     // Wrap indirect calls.
00742     materializeIndirectCalls();
00743 
00744     return true;
00745   }
00746 
00747   /// \brief Compute the shadow type that corresponds to a given Value.
00748   Type *getShadowTy(Value *V) {
00749     return getShadowTy(V->getType());
00750   }
00751 
00752   /// \brief Compute the shadow type that corresponds to a given Type.
00753   Type *getShadowTy(Type *OrigTy) {
00754     if (!OrigTy->isSized()) {
00755       return nullptr;
00756     }
00757     // For integer type, shadow is the same as the original type.
00758     // This may return weird-sized types like i1.
00759     if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
00760       return IT;
00761     if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
00762       uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
00763       return VectorType::get(IntegerType::get(*MS.C, EltSize),
00764                              VT->getNumElements());
00765     }
00766     if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
00767       return ArrayType::get(getShadowTy(AT->getElementType()),
00768                             AT->getNumElements());
00769     }
00770     if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
00771       SmallVector<Type*, 4> Elements;
00772       for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
00773         Elements.push_back(getShadowTy(ST->getElementType(i)));
00774       StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
00775       DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
00776       return Res;
00777     }
00778     uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
00779     return IntegerType::get(*MS.C, TypeSize);
00780   }
00781 
00782   /// \brief Flatten a vector type.
00783   Type *getShadowTyNoVec(Type *ty) {
00784     if (VectorType *vt = dyn_cast<VectorType>(ty))
00785       return IntegerType::get(*MS.C, vt->getBitWidth());
00786     return ty;
00787   }
00788 
00789   /// \brief Convert a shadow value to it's flattened variant.
00790   Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
00791     Type *Ty = V->getType();
00792     Type *NoVecTy = getShadowTyNoVec(Ty);
00793     if (Ty == NoVecTy) return V;
00794     return IRB.CreateBitCast(V, NoVecTy);
00795   }
00796 
00797   /// \brief Compute the shadow address that corresponds to a given application
00798   /// address.
00799   ///
00800   /// Shadow = Addr & ~ShadowMask.
00801   Value *getShadowPtr(Value *Addr, Type *ShadowTy,
00802                       IRBuilder<> &IRB) {
00803     Value *ShadowLong =
00804       IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
00805                     ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
00806     return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
00807   }
00808 
00809   /// \brief Compute the origin address that corresponds to a given application
00810   /// address.
00811   ///
00812   /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
00813   Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
00814     Value *ShadowLong =
00815       IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
00816                     ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
00817     Value *Add =
00818       IRB.CreateAdd(ShadowLong,
00819                     ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
00820     Value *SecondAnd =
00821       IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
00822     return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
00823   }
00824 
00825   /// \brief Compute the shadow address for a given function argument.
00826   ///
00827   /// Shadow = ParamTLS+ArgOffset.
00828   Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
00829                                  int ArgOffset) {
00830     Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
00831     Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
00832     return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
00833                               "_msarg");
00834   }
00835 
00836   /// \brief Compute the origin address for a given function argument.
00837   Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
00838                                  int ArgOffset) {
00839     if (!MS.TrackOrigins) return nullptr;
00840     Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
00841     Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
00842     return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
00843                               "_msarg_o");
00844   }
00845 
00846   /// \brief Compute the shadow address for a retval.
00847   Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
00848     Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
00849     return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
00850                               "_msret");
00851   }
00852 
00853   /// \brief Compute the origin address for a retval.
00854   Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
00855     // We keep a single origin for the entire retval. Might be too optimistic.
00856     return MS.RetvalOriginTLS;
00857   }
00858 
00859   /// \brief Set SV to be the shadow value for V.
00860   void setShadow(Value *V, Value *SV) {
00861     assert(!ShadowMap.count(V) && "Values may only have one shadow");
00862     ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
00863   }
00864 
00865   /// \brief Set Origin to be the origin value for V.
00866   void setOrigin(Value *V, Value *Origin) {
00867     if (!MS.TrackOrigins) return;
00868     assert(!OriginMap.count(V) && "Values may only have one origin");
00869     DEBUG(dbgs() << "ORIGIN: " << *V << "  ==> " << *Origin << "\n");
00870     OriginMap[V] = Origin;
00871   }
00872 
00873   /// \brief Create a clean shadow value for a given value.
00874   ///
00875   /// Clean shadow (all zeroes) means all bits of the value are defined
00876   /// (initialized).
00877   Constant *getCleanShadow(Value *V) {
00878     Type *ShadowTy = getShadowTy(V);
00879     if (!ShadowTy)
00880       return nullptr;
00881     return Constant::getNullValue(ShadowTy);
00882   }
00883 
00884   /// \brief Create a dirty shadow of a given shadow type.
00885   Constant *getPoisonedShadow(Type *ShadowTy) {
00886     assert(ShadowTy);
00887     if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
00888       return Constant::getAllOnesValue(ShadowTy);
00889     if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
00890       SmallVector<Constant *, 4> Vals(AT->getNumElements(),
00891                                       getPoisonedShadow(AT->getElementType()));
00892       return ConstantArray::get(AT, Vals);
00893     }
00894     if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
00895       SmallVector<Constant *, 4> Vals;
00896       for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
00897         Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
00898       return ConstantStruct::get(ST, Vals);
00899     }
00900     llvm_unreachable("Unexpected shadow type");
00901   }
00902 
00903   /// \brief Create a dirty shadow for a given value.
00904   Constant *getPoisonedShadow(Value *V) {
00905     Type *ShadowTy = getShadowTy(V);
00906     if (!ShadowTy)
00907       return nullptr;
00908     return getPoisonedShadow(ShadowTy);
00909   }
00910 
00911   /// \brief Create a clean (zero) origin.
00912   Value *getCleanOrigin() {
00913     return Constant::getNullValue(MS.OriginTy);
00914   }
00915 
00916   /// \brief Get the shadow value for a given Value.
00917   ///
00918   /// This function either returns the value set earlier with setShadow,
00919   /// or extracts if from ParamTLS (for function arguments).
00920   Value *getShadow(Value *V) {
00921     if (!PropagateShadow) return getCleanShadow(V);
00922     if (Instruction *I = dyn_cast<Instruction>(V)) {
00923       // For instructions the shadow is already stored in the map.
00924       Value *Shadow = ShadowMap[V];
00925       if (!Shadow) {
00926         DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
00927         (void)I;
00928         assert(Shadow && "No shadow for a value");
00929       }
00930       return Shadow;
00931     }
00932     if (UndefValue *U = dyn_cast<UndefValue>(V)) {
00933       Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
00934       DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
00935       (void)U;
00936       return AllOnes;
00937     }
00938     if (Argument *A = dyn_cast<Argument>(V)) {
00939       // For arguments we compute the shadow on demand and store it in the map.
00940       Value **ShadowPtr = &ShadowMap[V];
00941       if (*ShadowPtr)
00942         return *ShadowPtr;
00943       Function *F = A->getParent();
00944       IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
00945       unsigned ArgOffset = 0;
00946       for (auto &FArg : F->args()) {
00947         if (!FArg.getType()->isSized()) {
00948           DEBUG(dbgs() << "Arg is not sized\n");
00949           continue;
00950         }
00951         unsigned Size = FArg.hasByValAttr()
00952           ? MS.DL->getTypeAllocSize(FArg.getType()->getPointerElementType())
00953           : MS.DL->getTypeAllocSize(FArg.getType());
00954         if (A == &FArg) {
00955           Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
00956           if (FArg.hasByValAttr()) {
00957             // ByVal pointer itself has clean shadow. We copy the actual
00958             // argument shadow to the underlying memory.
00959             // Figure out maximal valid memcpy alignment.
00960             unsigned ArgAlign = FArg.getParamAlignment();
00961             if (ArgAlign == 0) {
00962               Type *EltType = A->getType()->getPointerElementType();
00963               ArgAlign = MS.DL->getABITypeAlignment(EltType);
00964             }
00965             unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
00966             Value *Cpy = EntryIRB.CreateMemCpy(
00967                 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
00968                 CopyAlign);
00969             DEBUG(dbgs() << "  ByValCpy: " << *Cpy << "\n");
00970             (void)Cpy;
00971             *ShadowPtr = getCleanShadow(V);
00972           } else {
00973             *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
00974           }
00975           DEBUG(dbgs() << "  ARG:    "  << FArg << " ==> " <<
00976                 **ShadowPtr << "\n");
00977           if (MS.TrackOrigins) {
00978             Value *OriginPtr =
00979                 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
00980             setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
00981           }
00982         }
00983         ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
00984       }
00985       assert(*ShadowPtr && "Could not find shadow for an argument");
00986       return *ShadowPtr;
00987     }
00988     // For everything else the shadow is zero.
00989     return getCleanShadow(V);
00990   }
00991 
00992   /// \brief Get the shadow for i-th argument of the instruction I.
00993   Value *getShadow(Instruction *I, int i) {
00994     return getShadow(I->getOperand(i));
00995   }
00996 
00997   /// \brief Get the origin for a value.
00998   Value *getOrigin(Value *V) {
00999     if (!MS.TrackOrigins) return nullptr;
01000     if (isa<Instruction>(V) || isa<Argument>(V)) {
01001       Value *Origin = OriginMap[V];
01002       if (!Origin) {
01003         DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
01004         Origin = getCleanOrigin();
01005       }
01006       return Origin;
01007     }
01008     return getCleanOrigin();
01009   }
01010 
01011   /// \brief Get the origin for i-th argument of the instruction I.
01012   Value *getOrigin(Instruction *I, int i) {
01013     return getOrigin(I->getOperand(i));
01014   }
01015 
01016   /// \brief Remember the place where a shadow check should be inserted.
01017   ///
01018   /// This location will be later instrumented with a check that will print a
01019   /// UMR warning in runtime if the shadow value is not 0.
01020   void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
01021     assert(Shadow);
01022     if (!InsertChecks) return;
01023 #ifndef NDEBUG
01024     Type *ShadowTy = Shadow->getType();
01025     assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
01026            "Can only insert checks for integer and vector shadow types");
01027 #endif
01028     InstrumentationList.push_back(
01029         ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
01030   }
01031 
01032   /// \brief Remember the place where a shadow check should be inserted.
01033   ///
01034   /// This location will be later instrumented with a check that will print a
01035   /// UMR warning in runtime if the value is not fully defined.
01036   void insertShadowCheck(Value *Val, Instruction *OrigIns) {
01037     assert(Val);
01038     Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
01039     if (!Shadow) return;
01040     Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
01041     insertShadowCheck(Shadow, Origin, OrigIns);
01042   }
01043 
01044   AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
01045     switch (a) {
01046       case NotAtomic:
01047         return NotAtomic;
01048       case Unordered:
01049       case Monotonic:
01050       case Release:
01051         return Release;
01052       case Acquire:
01053       case AcquireRelease:
01054         return AcquireRelease;
01055       case SequentiallyConsistent:
01056         return SequentiallyConsistent;
01057     }
01058     llvm_unreachable("Unknown ordering");
01059   }
01060 
01061   AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
01062     switch (a) {
01063       case NotAtomic:
01064         return NotAtomic;
01065       case Unordered:
01066       case Monotonic:
01067       case Acquire:
01068         return Acquire;
01069       case Release:
01070       case AcquireRelease:
01071         return AcquireRelease;
01072       case SequentiallyConsistent:
01073         return SequentiallyConsistent;
01074     }
01075     llvm_unreachable("Unknown ordering");
01076   }
01077 
01078   // ------------------- Visitors.
01079 
01080   /// \brief Instrument LoadInst
01081   ///
01082   /// Loads the corresponding shadow and (optionally) origin.
01083   /// Optionally, checks that the load address is fully defined.
01084   void visitLoadInst(LoadInst &I) {
01085     assert(I.getType()->isSized() && "Load type must have size");
01086     IRBuilder<> IRB(I.getNextNode());
01087     Type *ShadowTy = getShadowTy(&I);
01088     Value *Addr = I.getPointerOperand();
01089     if (PropagateShadow) {
01090       Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
01091       setShadow(&I,
01092                 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
01093     } else {
01094       setShadow(&I, getCleanShadow(&I));
01095     }
01096 
01097     if (ClCheckAccessAddress)
01098       insertShadowCheck(I.getPointerOperand(), &I);
01099 
01100     if (I.isAtomic())
01101       I.setOrdering(addAcquireOrdering(I.getOrdering()));
01102 
01103     if (MS.TrackOrigins) {
01104       if (PropagateShadow) {
01105         unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
01106         setOrigin(&I,
01107                   IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
01108       } else {
01109         setOrigin(&I, getCleanOrigin());
01110       }
01111     }
01112   }
01113 
01114   /// \brief Instrument StoreInst
01115   ///
01116   /// Stores the corresponding shadow and (optionally) origin.
01117   /// Optionally, checks that the store address is fully defined.
01118   void visitStoreInst(StoreInst &I) {
01119     StoreList.push_back(&I);
01120   }
01121 
01122   void handleCASOrRMW(Instruction &I) {
01123     assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
01124 
01125     IRBuilder<> IRB(&I);
01126     Value *Addr = I.getOperand(0);
01127     Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
01128 
01129     if (ClCheckAccessAddress)
01130       insertShadowCheck(Addr, &I);
01131 
01132     // Only test the conditional argument of cmpxchg instruction.
01133     // The other argument can potentially be uninitialized, but we can not
01134     // detect this situation reliably without possible false positives.
01135     if (isa<AtomicCmpXchgInst>(I))
01136       insertShadowCheck(I.getOperand(1), &I);
01137 
01138     IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
01139 
01140     setShadow(&I, getCleanShadow(&I));
01141   }
01142 
01143   void visitAtomicRMWInst(AtomicRMWInst &I) {
01144     handleCASOrRMW(I);
01145     I.setOrdering(addReleaseOrdering(I.getOrdering()));
01146   }
01147 
01148   void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
01149     handleCASOrRMW(I);
01150     I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
01151   }
01152 
01153   // Vector manipulation.
01154   void visitExtractElementInst(ExtractElementInst &I) {
01155     insertShadowCheck(I.getOperand(1), &I);
01156     IRBuilder<> IRB(&I);
01157     setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
01158               "_msprop"));
01159     setOrigin(&I, getOrigin(&I, 0));
01160   }
01161 
01162   void visitInsertElementInst(InsertElementInst &I) {
01163     insertShadowCheck(I.getOperand(2), &I);
01164     IRBuilder<> IRB(&I);
01165     setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
01166               I.getOperand(2), "_msprop"));
01167     setOriginForNaryOp(I);
01168   }
01169 
01170   void visitShuffleVectorInst(ShuffleVectorInst &I) {
01171     insertShadowCheck(I.getOperand(2), &I);
01172     IRBuilder<> IRB(&I);
01173     setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
01174               I.getOperand(2), "_msprop"));
01175     setOriginForNaryOp(I);
01176   }
01177 
01178   // Casts.
01179   void visitSExtInst(SExtInst &I) {
01180     IRBuilder<> IRB(&I);
01181     setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
01182     setOrigin(&I, getOrigin(&I, 0));
01183   }
01184 
01185   void visitZExtInst(ZExtInst &I) {
01186     IRBuilder<> IRB(&I);
01187     setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
01188     setOrigin(&I, getOrigin(&I, 0));
01189   }
01190 
01191   void visitTruncInst(TruncInst &I) {
01192     IRBuilder<> IRB(&I);
01193     setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
01194     setOrigin(&I, getOrigin(&I, 0));
01195   }
01196 
01197   void visitBitCastInst(BitCastInst &I) {
01198     IRBuilder<> IRB(&I);
01199     setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
01200     setOrigin(&I, getOrigin(&I, 0));
01201   }
01202 
01203   void visitPtrToIntInst(PtrToIntInst &I) {
01204     IRBuilder<> IRB(&I);
01205     setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
01206              "_msprop_ptrtoint"));
01207     setOrigin(&I, getOrigin(&I, 0));
01208   }
01209 
01210   void visitIntToPtrInst(IntToPtrInst &I) {
01211     IRBuilder<> IRB(&I);
01212     setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
01213              "_msprop_inttoptr"));
01214     setOrigin(&I, getOrigin(&I, 0));
01215   }
01216 
01217   void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
01218   void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
01219   void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
01220   void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
01221   void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
01222   void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
01223 
01224   /// \brief Propagate shadow for bitwise AND.
01225   ///
01226   /// This code is exact, i.e. if, for example, a bit in the left argument
01227   /// is defined and 0, then neither the value not definedness of the
01228   /// corresponding bit in B don't affect the resulting shadow.
01229   void visitAnd(BinaryOperator &I) {
01230     IRBuilder<> IRB(&I);
01231     //  "And" of 0 and a poisoned value results in unpoisoned value.
01232     //  1&1 => 1;     0&1 => 0;     p&1 => p;
01233     //  1&0 => 0;     0&0 => 0;     p&0 => 0;
01234     //  1&p => p;     0&p => 0;     p&p => p;
01235     //  S = (S1 & S2) | (V1 & S2) | (S1 & V2)
01236     Value *S1 = getShadow(&I, 0);
01237     Value *S2 = getShadow(&I, 1);
01238     Value *V1 = I.getOperand(0);
01239     Value *V2 = I.getOperand(1);
01240     if (V1->getType() != S1->getType()) {
01241       V1 = IRB.CreateIntCast(V1, S1->getType(), false);
01242       V2 = IRB.CreateIntCast(V2, S2->getType(), false);
01243     }
01244     Value *S1S2 = IRB.CreateAnd(S1, S2);
01245     Value *V1S2 = IRB.CreateAnd(V1, S2);
01246     Value *S1V2 = IRB.CreateAnd(S1, V2);
01247     setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
01248     setOriginForNaryOp(I);
01249   }
01250 
01251   void visitOr(BinaryOperator &I) {
01252     IRBuilder<> IRB(&I);
01253     //  "Or" of 1 and a poisoned value results in unpoisoned value.
01254     //  1|1 => 1;     0|1 => 1;     p|1 => 1;
01255     //  1|0 => 1;     0|0 => 0;     p|0 => p;
01256     //  1|p => 1;     0|p => p;     p|p => p;
01257     //  S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
01258     Value *S1 = getShadow(&I, 0);
01259     Value *S2 = getShadow(&I, 1);
01260     Value *V1 = IRB.CreateNot(I.getOperand(0));
01261     Value *V2 = IRB.CreateNot(I.getOperand(1));
01262     if (V1->getType() != S1->getType()) {
01263       V1 = IRB.CreateIntCast(V1, S1->getType(), false);
01264       V2 = IRB.CreateIntCast(V2, S2->getType(), false);
01265     }
01266     Value *S1S2 = IRB.CreateAnd(S1, S2);
01267     Value *V1S2 = IRB.CreateAnd(V1, S2);
01268     Value *S1V2 = IRB.CreateAnd(S1, V2);
01269     setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
01270     setOriginForNaryOp(I);
01271   }
01272 
01273   /// \brief Default propagation of shadow and/or origin.
01274   ///
01275   /// This class implements the general case of shadow propagation, used in all
01276   /// cases where we don't know and/or don't care about what the operation
01277   /// actually does. It converts all input shadow values to a common type
01278   /// (extending or truncating as necessary), and bitwise OR's them.
01279   ///
01280   /// This is much cheaper than inserting checks (i.e. requiring inputs to be
01281   /// fully initialized), and less prone to false positives.
01282   ///
01283   /// This class also implements the general case of origin propagation. For a
01284   /// Nary operation, result origin is set to the origin of an argument that is
01285   /// not entirely initialized. If there is more than one such arguments, the
01286   /// rightmost of them is picked. It does not matter which one is picked if all
01287   /// arguments are initialized.
01288   template <bool CombineShadow>
01289   class Combiner {
01290     Value *Shadow;
01291     Value *Origin;
01292     IRBuilder<> &IRB;
01293     MemorySanitizerVisitor *MSV;
01294 
01295   public:
01296     Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
01297       Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
01298 
01299     /// \brief Add a pair of shadow and origin values to the mix.
01300     Combiner &Add(Value *OpShadow, Value *OpOrigin) {
01301       if (CombineShadow) {
01302         assert(OpShadow);
01303         if (!Shadow)
01304           Shadow = OpShadow;
01305         else {
01306           OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
01307           Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
01308         }
01309       }
01310 
01311       if (MSV->MS.TrackOrigins) {
01312         assert(OpOrigin);
01313         if (!Origin) {
01314           Origin = OpOrigin;
01315         } else {
01316           Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
01317           // No point in adding something that might result in 0 origin value.
01318           if (!ConstOrigin || !ConstOrigin->isNullValue()) {
01319             Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
01320             Value *Cond =
01321                 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
01322             Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
01323           }
01324         }
01325       }
01326       return *this;
01327     }
01328 
01329     /// \brief Add an application value to the mix.
01330     Combiner &Add(Value *V) {
01331       Value *OpShadow = MSV->getShadow(V);
01332       Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
01333       return Add(OpShadow, OpOrigin);
01334     }
01335 
01336     /// \brief Set the current combined values as the given instruction's shadow
01337     /// and origin.
01338     void Done(Instruction *I) {
01339       if (CombineShadow) {
01340         assert(Shadow);
01341         Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
01342         MSV->setShadow(I, Shadow);
01343       }
01344       if (MSV->MS.TrackOrigins) {
01345         assert(Origin);
01346         MSV->setOrigin(I, Origin);
01347       }
01348     }
01349   };
01350 
01351   typedef Combiner<true> ShadowAndOriginCombiner;
01352   typedef Combiner<false> OriginCombiner;
01353 
01354   /// \brief Propagate origin for arbitrary operation.
01355   void setOriginForNaryOp(Instruction &I) {
01356     if (!MS.TrackOrigins) return;
01357     IRBuilder<> IRB(&I);
01358     OriginCombiner OC(this, IRB);
01359     for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
01360       OC.Add(OI->get());
01361     OC.Done(&I);
01362   }
01363 
01364   size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
01365     assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
01366            "Vector of pointers is not a valid shadow type");
01367     return Ty->isVectorTy() ?
01368       Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
01369       Ty->getPrimitiveSizeInBits();
01370   }
01371 
01372   /// \brief Cast between two shadow types, extending or truncating as
01373   /// necessary.
01374   Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
01375                           bool Signed = false) {
01376     Type *srcTy = V->getType();
01377     if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
01378       return IRB.CreateIntCast(V, dstTy, Signed);
01379     if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
01380         dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
01381       return IRB.CreateIntCast(V, dstTy, Signed);
01382     size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
01383     size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
01384     Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
01385     Value *V2 =
01386       IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
01387     return IRB.CreateBitCast(V2, dstTy);
01388     // TODO: handle struct types.
01389   }
01390 
01391   /// \brief Cast an application value to the type of its own shadow.
01392   Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
01393     Type *ShadowTy = getShadowTy(V);
01394     if (V->getType() == ShadowTy)
01395       return V;
01396     if (V->getType()->isPtrOrPtrVectorTy())
01397       return IRB.CreatePtrToInt(V, ShadowTy);
01398     else
01399       return IRB.CreateBitCast(V, ShadowTy);
01400   }
01401 
01402   /// \brief Propagate shadow for arbitrary operation.
01403   void handleShadowOr(Instruction &I) {
01404     IRBuilder<> IRB(&I);
01405     ShadowAndOriginCombiner SC(this, IRB);
01406     for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
01407       SC.Add(OI->get());
01408     SC.Done(&I);
01409   }
01410 
01411   // \brief Handle multiplication by constant.
01412   //
01413   // Handle a special case of multiplication by constant that may have one or
01414   // more zeros in the lower bits. This makes corresponding number of lower bits
01415   // of the result zero as well. We model it by shifting the other operand
01416   // shadow left by the required number of bits. Effectively, we transform
01417   // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
01418   // We use multiplication by 2**N instead of shift to cover the case of
01419   // multiplication by 0, which may occur in some elements of a vector operand.
01420   void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
01421                            Value *OtherArg) {
01422     Constant *ShadowMul;
01423     Type *Ty = ConstArg->getType();
01424     if (Ty->isVectorTy()) {
01425       unsigned NumElements = Ty->getVectorNumElements();
01426       Type *EltTy = Ty->getSequentialElementType();
01427       SmallVector<Constant *, 16> Elements;
01428       for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
01429         ConstantInt *Elt =
01430             dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
01431         APInt V = Elt->getValue();
01432         APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
01433         Elements.push_back(ConstantInt::get(EltTy, V2));
01434       }
01435       ShadowMul = ConstantVector::get(Elements);
01436     } else {
01437       ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
01438       APInt V = Elt->getValue();
01439       APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
01440       ShadowMul = ConstantInt::get(Elt->getType(), V2);
01441     }
01442 
01443     IRBuilder<> IRB(&I);
01444     setShadow(&I,
01445               IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
01446     setOrigin(&I, getOrigin(OtherArg));
01447   }
01448 
01449   void visitMul(BinaryOperator &I) {
01450     Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
01451     Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
01452     if (constOp0 && !constOp1)
01453       handleMulByConstant(I, constOp0, I.getOperand(1));
01454     else if (constOp1 && !constOp0)
01455       handleMulByConstant(I, constOp1, I.getOperand(0));
01456     else
01457       handleShadowOr(I);
01458   }
01459 
01460   void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
01461   void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
01462   void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
01463   void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
01464   void visitSub(BinaryOperator &I) { handleShadowOr(I); }
01465   void visitXor(BinaryOperator &I) { handleShadowOr(I); }
01466 
01467   void handleDiv(Instruction &I) {
01468     IRBuilder<> IRB(&I);
01469     // Strict on the second argument.
01470     insertShadowCheck(I.getOperand(1), &I);
01471     setShadow(&I, getShadow(&I, 0));
01472     setOrigin(&I, getOrigin(&I, 0));
01473   }
01474 
01475   void visitUDiv(BinaryOperator &I) { handleDiv(I); }
01476   void visitSDiv(BinaryOperator &I) { handleDiv(I); }
01477   void visitFDiv(BinaryOperator &I) { handleDiv(I); }
01478   void visitURem(BinaryOperator &I) { handleDiv(I); }
01479   void visitSRem(BinaryOperator &I) { handleDiv(I); }
01480   void visitFRem(BinaryOperator &I) { handleDiv(I); }
01481 
01482   /// \brief Instrument == and != comparisons.
01483   ///
01484   /// Sometimes the comparison result is known even if some of the bits of the
01485   /// arguments are not.
01486   void handleEqualityComparison(ICmpInst &I) {
01487     IRBuilder<> IRB(&I);
01488     Value *A = I.getOperand(0);
01489     Value *B = I.getOperand(1);
01490     Value *Sa = getShadow(A);
01491     Value *Sb = getShadow(B);
01492 
01493     // Get rid of pointers and vectors of pointers.
01494     // For ints (and vectors of ints), types of A and Sa match,
01495     // and this is a no-op.
01496     A = IRB.CreatePointerCast(A, Sa->getType());
01497     B = IRB.CreatePointerCast(B, Sb->getType());
01498 
01499     // A == B  <==>  (C = A^B) == 0
01500     // A != B  <==>  (C = A^B) != 0
01501     // Sc = Sa | Sb
01502     Value *C = IRB.CreateXor(A, B);
01503     Value *Sc = IRB.CreateOr(Sa, Sb);
01504     // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
01505     // Result is defined if one of the following is true
01506     // * there is a defined 1 bit in C
01507     // * C is fully defined
01508     // Si = !(C & ~Sc) && Sc
01509     Value *Zero = Constant::getNullValue(Sc->getType());
01510     Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
01511     Value *Si =
01512       IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
01513                     IRB.CreateICmpEQ(
01514                       IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
01515     Si->setName("_msprop_icmp");
01516     setShadow(&I, Si);
01517     setOriginForNaryOp(I);
01518   }
01519 
01520   /// \brief Build the lowest possible value of V, taking into account V's
01521   ///        uninitialized bits.
01522   Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
01523                                 bool isSigned) {
01524     if (isSigned) {
01525       // Split shadow into sign bit and other bits.
01526       Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
01527       Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
01528       // Maximise the undefined shadow bit, minimize other undefined bits.
01529       return
01530         IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
01531     } else {
01532       // Minimize undefined bits.
01533       return IRB.CreateAnd(A, IRB.CreateNot(Sa));
01534     }
01535   }
01536 
01537   /// \brief Build the highest possible value of V, taking into account V's
01538   ///        uninitialized bits.
01539   Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
01540                                 bool isSigned) {
01541     if (isSigned) {
01542       // Split shadow into sign bit and other bits.
01543       Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
01544       Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
01545       // Minimise the undefined shadow bit, maximise other undefined bits.
01546       return
01547         IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
01548     } else {
01549       // Maximize undefined bits.
01550       return IRB.CreateOr(A, Sa);
01551     }
01552   }
01553 
01554   /// \brief Instrument relational comparisons.
01555   ///
01556   /// This function does exact shadow propagation for all relational
01557   /// comparisons of integers, pointers and vectors of those.
01558   /// FIXME: output seems suboptimal when one of the operands is a constant
01559   void handleRelationalComparisonExact(ICmpInst &I) {
01560     IRBuilder<> IRB(&I);
01561     Value *A = I.getOperand(0);
01562     Value *B = I.getOperand(1);
01563     Value *Sa = getShadow(A);
01564     Value *Sb = getShadow(B);
01565 
01566     // Get rid of pointers and vectors of pointers.
01567     // For ints (and vectors of ints), types of A and Sa match,
01568     // and this is a no-op.
01569     A = IRB.CreatePointerCast(A, Sa->getType());
01570     B = IRB.CreatePointerCast(B, Sb->getType());
01571 
01572     // Let [a0, a1] be the interval of possible values of A, taking into account
01573     // its undefined bits. Let [b0, b1] be the interval of possible values of B.
01574     // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
01575     bool IsSigned = I.isSigned();
01576     Value *S1 = IRB.CreateICmp(I.getPredicate(),
01577                                getLowestPossibleValue(IRB, A, Sa, IsSigned),
01578                                getHighestPossibleValue(IRB, B, Sb, IsSigned));
01579     Value *S2 = IRB.CreateICmp(I.getPredicate(),
01580                                getHighestPossibleValue(IRB, A, Sa, IsSigned),
01581                                getLowestPossibleValue(IRB, B, Sb, IsSigned));
01582     Value *Si = IRB.CreateXor(S1, S2);
01583     setShadow(&I, Si);
01584     setOriginForNaryOp(I);
01585   }
01586 
01587   /// \brief Instrument signed relational comparisons.
01588   ///
01589   /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
01590   /// propagating the highest bit of the shadow. Everything else is delegated
01591   /// to handleShadowOr().
01592   void handleSignedRelationalComparison(ICmpInst &I) {
01593     Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
01594     Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
01595     Value* op = nullptr;
01596     CmpInst::Predicate pre = I.getPredicate();
01597     if (constOp0 && constOp0->isNullValue() &&
01598         (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
01599       op = I.getOperand(1);
01600     } else if (constOp1 && constOp1->isNullValue() &&
01601                (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
01602       op = I.getOperand(0);
01603     }
01604     if (op) {
01605       IRBuilder<> IRB(&I);
01606       Value* Shadow =
01607         IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
01608       setShadow(&I, Shadow);
01609       setOrigin(&I, getOrigin(op));
01610     } else {
01611       handleShadowOr(I);
01612     }
01613   }
01614 
01615   void visitICmpInst(ICmpInst &I) {
01616     if (!ClHandleICmp) {
01617       handleShadowOr(I);
01618       return;
01619     }
01620     if (I.isEquality()) {
01621       handleEqualityComparison(I);
01622       return;
01623     }
01624 
01625     assert(I.isRelational());
01626     if (ClHandleICmpExact) {
01627       handleRelationalComparisonExact(I);
01628       return;
01629     }
01630     if (I.isSigned()) {
01631       handleSignedRelationalComparison(I);
01632       return;
01633     }
01634 
01635     assert(I.isUnsigned());
01636     if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
01637       handleRelationalComparisonExact(I);
01638       return;
01639     }
01640 
01641     handleShadowOr(I);
01642   }
01643 
01644   void visitFCmpInst(FCmpInst &I) {
01645     handleShadowOr(I);
01646   }
01647 
01648   void handleShift(BinaryOperator &I) {
01649     IRBuilder<> IRB(&I);
01650     // If any of the S2 bits are poisoned, the whole thing is poisoned.
01651     // Otherwise perform the same shift on S1.
01652     Value *S1 = getShadow(&I, 0);
01653     Value *S2 = getShadow(&I, 1);
01654     Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
01655                                    S2->getType());
01656     Value *V2 = I.getOperand(1);
01657     Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
01658     setShadow(&I, IRB.CreateOr(Shift, S2Conv));
01659     setOriginForNaryOp(I);
01660   }
01661 
01662   void visitShl(BinaryOperator &I) { handleShift(I); }
01663   void visitAShr(BinaryOperator &I) { handleShift(I); }
01664   void visitLShr(BinaryOperator &I) { handleShift(I); }
01665 
01666   /// \brief Instrument llvm.memmove
01667   ///
01668   /// At this point we don't know if llvm.memmove will be inlined or not.
01669   /// If we don't instrument it and it gets inlined,
01670   /// our interceptor will not kick in and we will lose the memmove.
01671   /// If we instrument the call here, but it does not get inlined,
01672   /// we will memove the shadow twice: which is bad in case
01673   /// of overlapping regions. So, we simply lower the intrinsic to a call.
01674   ///
01675   /// Similar situation exists for memcpy and memset.
01676   void visitMemMoveInst(MemMoveInst &I) {
01677     IRBuilder<> IRB(&I);
01678     IRB.CreateCall3(
01679       MS.MemmoveFn,
01680       IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
01681       IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
01682       IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
01683     I.eraseFromParent();
01684   }
01685 
01686   // Similar to memmove: avoid copying shadow twice.
01687   // This is somewhat unfortunate as it may slowdown small constant memcpys.
01688   // FIXME: consider doing manual inline for small constant sizes and proper
01689   // alignment.
01690   void visitMemCpyInst(MemCpyInst &I) {
01691     IRBuilder<> IRB(&I);
01692     IRB.CreateCall3(
01693       MS.MemcpyFn,
01694       IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
01695       IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
01696       IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
01697     I.eraseFromParent();
01698   }
01699 
01700   // Same as memcpy.
01701   void visitMemSetInst(MemSetInst &I) {
01702     IRBuilder<> IRB(&I);
01703     IRB.CreateCall3(
01704       MS.MemsetFn,
01705       IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
01706       IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
01707       IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
01708     I.eraseFromParent();
01709   }
01710 
01711   void visitVAStartInst(VAStartInst &I) {
01712     VAHelper->visitVAStartInst(I);
01713   }
01714 
01715   void visitVACopyInst(VACopyInst &I) {
01716     VAHelper->visitVACopyInst(I);
01717   }
01718 
01719   enum IntrinsicKind {
01720     IK_DoesNotAccessMemory,
01721     IK_OnlyReadsMemory,
01722     IK_WritesMemory
01723   };
01724 
01725   static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
01726     const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
01727     const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
01728     const int OnlyReadsMemory = IK_OnlyReadsMemory;
01729     const int OnlyAccessesArgumentPointees = IK_WritesMemory;
01730     const int UnknownModRefBehavior = IK_WritesMemory;
01731 #define GET_INTRINSIC_MODREF_BEHAVIOR
01732 #define ModRefBehavior IntrinsicKind
01733 #include "llvm/IR/Intrinsics.gen"
01734 #undef ModRefBehavior
01735 #undef GET_INTRINSIC_MODREF_BEHAVIOR
01736   }
01737 
01738   /// \brief Handle vector store-like intrinsics.
01739   ///
01740   /// Instrument intrinsics that look like a simple SIMD store: writes memory,
01741   /// has 1 pointer argument and 1 vector argument, returns void.
01742   bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
01743     IRBuilder<> IRB(&I);
01744     Value* Addr = I.getArgOperand(0);
01745     Value *Shadow = getShadow(&I, 1);
01746     Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
01747 
01748     // We don't know the pointer alignment (could be unaligned SSE store!).
01749     // Have to assume to worst case.
01750     IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
01751 
01752     if (ClCheckAccessAddress)
01753       insertShadowCheck(Addr, &I);
01754 
01755     // FIXME: use ClStoreCleanOrigin
01756     // FIXME: factor out common code from materializeStores
01757     if (MS.TrackOrigins)
01758       IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
01759     return true;
01760   }
01761 
01762   /// \brief Handle vector load-like intrinsics.
01763   ///
01764   /// Instrument intrinsics that look like a simple SIMD load: reads memory,
01765   /// has 1 pointer argument, returns a vector.
01766   bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
01767     IRBuilder<> IRB(&I);
01768     Value *Addr = I.getArgOperand(0);
01769 
01770     Type *ShadowTy = getShadowTy(&I);
01771     if (PropagateShadow) {
01772       Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
01773       // We don't know the pointer alignment (could be unaligned SSE load!).
01774       // Have to assume to worst case.
01775       setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
01776     } else {
01777       setShadow(&I, getCleanShadow(&I));
01778     }
01779 
01780     if (ClCheckAccessAddress)
01781       insertShadowCheck(Addr, &I);
01782 
01783     if (MS.TrackOrigins) {
01784       if (PropagateShadow)
01785         setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
01786       else
01787         setOrigin(&I, getCleanOrigin());
01788     }
01789     return true;
01790   }
01791 
01792   /// \brief Handle (SIMD arithmetic)-like intrinsics.
01793   ///
01794   /// Instrument intrinsics with any number of arguments of the same type,
01795   /// equal to the return type. The type should be simple (no aggregates or
01796   /// pointers; vectors are fine).
01797   /// Caller guarantees that this intrinsic does not access memory.
01798   bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
01799     Type *RetTy = I.getType();
01800     if (!(RetTy->isIntOrIntVectorTy() ||
01801           RetTy->isFPOrFPVectorTy() ||
01802           RetTy->isX86_MMXTy()))
01803       return false;
01804 
01805     unsigned NumArgOperands = I.getNumArgOperands();
01806 
01807     for (unsigned i = 0; i < NumArgOperands; ++i) {
01808       Type *Ty = I.getArgOperand(i)->getType();
01809       if (Ty != RetTy)
01810         return false;
01811     }
01812 
01813     IRBuilder<> IRB(&I);
01814     ShadowAndOriginCombiner SC(this, IRB);
01815     for (unsigned i = 0; i < NumArgOperands; ++i)
01816       SC.Add(I.getArgOperand(i));
01817     SC.Done(&I);
01818 
01819     return true;
01820   }
01821 
01822   /// \brief Heuristically instrument unknown intrinsics.
01823   ///
01824   /// The main purpose of this code is to do something reasonable with all
01825   /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
01826   /// We recognize several classes of intrinsics by their argument types and
01827   /// ModRefBehaviour and apply special intrumentation when we are reasonably
01828   /// sure that we know what the intrinsic does.
01829   ///
01830   /// We special-case intrinsics where this approach fails. See llvm.bswap
01831   /// handling as an example of that.
01832   bool handleUnknownIntrinsic(IntrinsicInst &I) {
01833     unsigned NumArgOperands = I.getNumArgOperands();
01834     if (NumArgOperands == 0)
01835       return false;
01836 
01837     Intrinsic::ID iid = I.getIntrinsicID();
01838     IntrinsicKind IK = getIntrinsicKind(iid);
01839     bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
01840     bool WritesMemory = IK == IK_WritesMemory;
01841     assert(!(OnlyReadsMemory && WritesMemory));
01842 
01843     if (NumArgOperands == 2 &&
01844         I.getArgOperand(0)->getType()->isPointerTy() &&
01845         I.getArgOperand(1)->getType()->isVectorTy() &&
01846         I.getType()->isVoidTy() &&
01847         WritesMemory) {
01848       // This looks like a vector store.
01849       return handleVectorStoreIntrinsic(I);
01850     }
01851 
01852     if (NumArgOperands == 1 &&
01853         I.getArgOperand(0)->getType()->isPointerTy() &&
01854         I.getType()->isVectorTy() &&
01855         OnlyReadsMemory) {
01856       // This looks like a vector load.
01857       return handleVectorLoadIntrinsic(I);
01858     }
01859 
01860     if (!OnlyReadsMemory && !WritesMemory)
01861       if (maybeHandleSimpleNomemIntrinsic(I))
01862         return true;
01863 
01864     // FIXME: detect and handle SSE maskstore/maskload
01865     return false;
01866   }
01867 
01868   void handleBswap(IntrinsicInst &I) {
01869     IRBuilder<> IRB(&I);
01870     Value *Op = I.getArgOperand(0);
01871     Type *OpType = Op->getType();
01872     Function *BswapFunc = Intrinsic::getDeclaration(
01873       F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
01874     setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
01875     setOrigin(&I, getOrigin(Op));
01876   }
01877 
01878   // \brief Instrument vector convert instrinsic.
01879   //
01880   // This function instruments intrinsics like cvtsi2ss:
01881   // %Out = int_xxx_cvtyyy(%ConvertOp)
01882   // or
01883   // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
01884   // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
01885   // number \p Out elements, and (if has 2 arguments) copies the rest of the
01886   // elements from \p CopyOp.
01887   // In most cases conversion involves floating-point value which may trigger a
01888   // hardware exception when not fully initialized. For this reason we require
01889   // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
01890   // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
01891   // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
01892   // return a fully initialized value.
01893   void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
01894     IRBuilder<> IRB(&I);
01895     Value *CopyOp, *ConvertOp;
01896 
01897     switch (I.getNumArgOperands()) {
01898     case 2:
01899       CopyOp = I.getArgOperand(0);
01900       ConvertOp = I.getArgOperand(1);
01901       break;
01902     case 1:
01903       ConvertOp = I.getArgOperand(0);
01904       CopyOp = nullptr;
01905       break;
01906     default:
01907       llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
01908     }
01909 
01910     // The first *NumUsedElements* elements of ConvertOp are converted to the
01911     // same number of output elements. The rest of the output is copied from
01912     // CopyOp, or (if not available) filled with zeroes.
01913     // Combine shadow for elements of ConvertOp that are used in this operation,
01914     // and insert a check.
01915     // FIXME: consider propagating shadow of ConvertOp, at least in the case of
01916     // int->any conversion.
01917     Value *ConvertShadow = getShadow(ConvertOp);
01918     Value *AggShadow = nullptr;
01919     if (ConvertOp->getType()->isVectorTy()) {
01920       AggShadow = IRB.CreateExtractElement(
01921           ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
01922       for (int i = 1; i < NumUsedElements; ++i) {
01923         Value *MoreShadow = IRB.CreateExtractElement(
01924             ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
01925         AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
01926       }
01927     } else {
01928       AggShadow = ConvertShadow;
01929     }
01930     assert(AggShadow->getType()->isIntegerTy());
01931     insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
01932 
01933     // Build result shadow by zero-filling parts of CopyOp shadow that come from
01934     // ConvertOp.
01935     if (CopyOp) {
01936       assert(CopyOp->getType() == I.getType());
01937       assert(CopyOp->getType()->isVectorTy());
01938       Value *ResultShadow = getShadow(CopyOp);
01939       Type *EltTy = ResultShadow->getType()->getVectorElementType();
01940       for (int i = 0; i < NumUsedElements; ++i) {
01941         ResultShadow = IRB.CreateInsertElement(
01942             ResultShadow, ConstantInt::getNullValue(EltTy),
01943             ConstantInt::get(IRB.getInt32Ty(), i));
01944       }
01945       setShadow(&I, ResultShadow);
01946       setOrigin(&I, getOrigin(CopyOp));
01947     } else {
01948       setShadow(&I, getCleanShadow(&I));
01949     }
01950   }
01951 
01952   // Given a scalar or vector, extract lower 64 bits (or less), and return all
01953   // zeroes if it is zero, and all ones otherwise.
01954   Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
01955     if (S->getType()->isVectorTy())
01956       S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
01957     assert(S->getType()->getPrimitiveSizeInBits() <= 64);
01958     Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
01959     return CreateShadowCast(IRB, S2, T, /* Signed */ true);
01960   }
01961 
01962   Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
01963     Type *T = S->getType();
01964     assert(T->isVectorTy());
01965     Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
01966     return IRB.CreateSExt(S2, T);
01967   }
01968 
01969   // \brief Instrument vector shift instrinsic.
01970   //
01971   // This function instruments intrinsics like int_x86_avx2_psll_w.
01972   // Intrinsic shifts %In by %ShiftSize bits.
01973   // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
01974   // size, and the rest is ignored. Behavior is defined even if shift size is
01975   // greater than register (or field) width.
01976   void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
01977     assert(I.getNumArgOperands() == 2);
01978     IRBuilder<> IRB(&I);
01979     // If any of the S2 bits are poisoned, the whole thing is poisoned.
01980     // Otherwise perform the same shift on S1.
01981     Value *S1 = getShadow(&I, 0);
01982     Value *S2 = getShadow(&I, 1);
01983     Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
01984                              : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
01985     Value *V1 = I.getOperand(0);
01986     Value *V2 = I.getOperand(1);
01987     Value *Shift = IRB.CreateCall2(I.getCalledValue(),
01988                                    IRB.CreateBitCast(S1, V1->getType()), V2);
01989     Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
01990     setShadow(&I, IRB.CreateOr(Shift, S2Conv));
01991     setOriginForNaryOp(I);
01992   }
01993 
01994   // \brief Get an X86_MMX-sized vector type.
01995   Type *getMMXVectorTy(unsigned EltSizeInBits) {
01996     const unsigned X86_MMXSizeInBits = 64;
01997     return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
01998                            X86_MMXSizeInBits / EltSizeInBits);
01999   }
02000 
02001   // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
02002   // intrinsic.
02003   Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
02004     switch (id) {
02005       case llvm::Intrinsic::x86_sse2_packsswb_128:
02006       case llvm::Intrinsic::x86_sse2_packuswb_128:
02007         return llvm::Intrinsic::x86_sse2_packsswb_128;
02008 
02009       case llvm::Intrinsic::x86_sse2_packssdw_128:
02010       case llvm::Intrinsic::x86_sse41_packusdw:
02011         return llvm::Intrinsic::x86_sse2_packssdw_128;
02012 
02013       case llvm::Intrinsic::x86_avx2_packsswb:
02014       case llvm::Intrinsic::x86_avx2_packuswb:
02015         return llvm::Intrinsic::x86_avx2_packsswb;
02016 
02017       case llvm::Intrinsic::x86_avx2_packssdw:
02018       case llvm::Intrinsic::x86_avx2_packusdw:
02019         return llvm::Intrinsic::x86_avx2_packssdw;
02020 
02021       case llvm::Intrinsic::x86_mmx_packsswb:
02022       case llvm::Intrinsic::x86_mmx_packuswb:
02023         return llvm::Intrinsic::x86_mmx_packsswb;
02024 
02025       case llvm::Intrinsic::x86_mmx_packssdw:
02026         return llvm::Intrinsic::x86_mmx_packssdw;
02027       default:
02028         llvm_unreachable("unexpected intrinsic id");
02029     }
02030   }
02031 
02032   // \brief Instrument vector pack instrinsic.
02033   //
02034   // This function instruments intrinsics like x86_mmx_packsswb, that
02035   // packs elements of 2 input vectors into half as many bits with saturation.
02036   // Shadow is propagated with the signed variant of the same intrinsic applied
02037   // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
02038   // EltSizeInBits is used only for x86mmx arguments.
02039   void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
02040     assert(I.getNumArgOperands() == 2);
02041     bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
02042     IRBuilder<> IRB(&I);
02043     Value *S1 = getShadow(&I, 0);
02044     Value *S2 = getShadow(&I, 1);
02045     assert(isX86_MMX || S1->getType()->isVectorTy());
02046 
02047     // SExt and ICmpNE below must apply to individual elements of input vectors.
02048     // In case of x86mmx arguments, cast them to appropriate vector types and
02049     // back.
02050     Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
02051     if (isX86_MMX) {
02052       S1 = IRB.CreateBitCast(S1, T);
02053       S2 = IRB.CreateBitCast(S2, T);
02054     }
02055     Value *S1_ext = IRB.CreateSExt(
02056         IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
02057     Value *S2_ext = IRB.CreateSExt(
02058         IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
02059     if (isX86_MMX) {
02060       Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
02061       S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
02062       S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
02063     }
02064 
02065     Function *ShadowFn = Intrinsic::getDeclaration(
02066         F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
02067 
02068     Value *S = IRB.CreateCall2(ShadowFn, S1_ext, S2_ext, "_msprop_vector_pack");
02069     if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
02070     setShadow(&I, S);
02071     setOriginForNaryOp(I);
02072   }
02073 
02074   // \brief Instrument sum-of-absolute-differencies intrinsic.
02075   void handleVectorSadIntrinsic(IntrinsicInst &I) {
02076     const unsigned SignificantBitsPerResultElement = 16;
02077     bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
02078     Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
02079     unsigned ZeroBitsPerResultElement =
02080         ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
02081 
02082     IRBuilder<> IRB(&I);
02083     Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
02084     S = IRB.CreateBitCast(S, ResTy);
02085     S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
02086                        ResTy);
02087     S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
02088     S = IRB.CreateBitCast(S, getShadowTy(&I));
02089     setShadow(&I, S);
02090     setOriginForNaryOp(I);
02091   }
02092 
02093   // \brief Instrument multiply-add intrinsic.
02094   void handleVectorPmaddIntrinsic(IntrinsicInst &I,
02095                                   unsigned EltSizeInBits = 0) {
02096     bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
02097     Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
02098     IRBuilder<> IRB(&I);
02099     Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
02100     S = IRB.CreateBitCast(S, ResTy);
02101     S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
02102                        ResTy);
02103     S = IRB.CreateBitCast(S, getShadowTy(&I));
02104     setShadow(&I, S);
02105     setOriginForNaryOp(I);
02106   }
02107 
02108   void visitIntrinsicInst(IntrinsicInst &I) {
02109     switch (I.getIntrinsicID()) {
02110     case llvm::Intrinsic::bswap:
02111       handleBswap(I);
02112       break;
02113     case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
02114     case llvm::Intrinsic::x86_avx512_cvtsd2usi:
02115     case llvm::Intrinsic::x86_avx512_cvtss2usi64:
02116     case llvm::Intrinsic::x86_avx512_cvtss2usi:
02117     case llvm::Intrinsic::x86_avx512_cvttss2usi64:
02118     case llvm::Intrinsic::x86_avx512_cvttss2usi:
02119     case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
02120     case llvm::Intrinsic::x86_avx512_cvttsd2usi:
02121     case llvm::Intrinsic::x86_avx512_cvtusi2sd:
02122     case llvm::Intrinsic::x86_avx512_cvtusi2ss:
02123     case llvm::Intrinsic::x86_avx512_cvtusi642sd:
02124     case llvm::Intrinsic::x86_avx512_cvtusi642ss:
02125     case llvm::Intrinsic::x86_sse2_cvtsd2si64:
02126     case llvm::Intrinsic::x86_sse2_cvtsd2si:
02127     case llvm::Intrinsic::x86_sse2_cvtsd2ss:
02128     case llvm::Intrinsic::x86_sse2_cvtsi2sd:
02129     case llvm::Intrinsic::x86_sse2_cvtsi642sd:
02130     case llvm::Intrinsic::x86_sse2_cvtss2sd:
02131     case llvm::Intrinsic::x86_sse2_cvttsd2si64:
02132     case llvm::Intrinsic::x86_sse2_cvttsd2si:
02133     case llvm::Intrinsic::x86_sse_cvtsi2ss:
02134     case llvm::Intrinsic::x86_sse_cvtsi642ss:
02135     case llvm::Intrinsic::x86_sse_cvtss2si64:
02136     case llvm::Intrinsic::x86_sse_cvtss2si:
02137     case llvm::Intrinsic::x86_sse_cvttss2si64:
02138     case llvm::Intrinsic::x86_sse_cvttss2si:
02139       handleVectorConvertIntrinsic(I, 1);
02140       break;
02141     case llvm::Intrinsic::x86_sse2_cvtdq2pd:
02142     case llvm::Intrinsic::x86_sse2_cvtps2pd:
02143     case llvm::Intrinsic::x86_sse_cvtps2pi:
02144     case llvm::Intrinsic::x86_sse_cvttps2pi:
02145       handleVectorConvertIntrinsic(I, 2);
02146       break;
02147     case llvm::Intrinsic::x86_avx512_psll_dq:
02148     case llvm::Intrinsic::x86_avx512_psrl_dq:
02149     case llvm::Intrinsic::x86_avx2_psll_w:
02150     case llvm::Intrinsic::x86_avx2_psll_d:
02151     case llvm::Intrinsic::x86_avx2_psll_q:
02152     case llvm::Intrinsic::x86_avx2_pslli_w:
02153     case llvm::Intrinsic::x86_avx2_pslli_d:
02154     case llvm::Intrinsic::x86_avx2_pslli_q:
02155     case llvm::Intrinsic::x86_avx2_psll_dq:
02156     case llvm::Intrinsic::x86_avx2_psrl_w:
02157     case llvm::Intrinsic::x86_avx2_psrl_d:
02158     case llvm::Intrinsic::x86_avx2_psrl_q:
02159     case llvm::Intrinsic::x86_avx2_psra_w:
02160     case llvm::Intrinsic::x86_avx2_psra_d:
02161     case llvm::Intrinsic::x86_avx2_psrli_w:
02162     case llvm::Intrinsic::x86_avx2_psrli_d:
02163     case llvm::Intrinsic::x86_avx2_psrli_q:
02164     case llvm::Intrinsic::x86_avx2_psrai_w:
02165     case llvm::Intrinsic::x86_avx2_psrai_d:
02166     case llvm::Intrinsic::x86_avx2_psrl_dq:
02167     case llvm::Intrinsic::x86_sse2_psll_w:
02168     case llvm::Intrinsic::x86_sse2_psll_d:
02169     case llvm::Intrinsic::x86_sse2_psll_q:
02170     case llvm::Intrinsic::x86_sse2_pslli_w:
02171     case llvm::Intrinsic::x86_sse2_pslli_d:
02172     case llvm::Intrinsic::x86_sse2_pslli_q:
02173     case llvm::Intrinsic::x86_sse2_psll_dq:
02174     case llvm::Intrinsic::x86_sse2_psrl_w:
02175     case llvm::Intrinsic::x86_sse2_psrl_d:
02176     case llvm::Intrinsic::x86_sse2_psrl_q:
02177     case llvm::Intrinsic::x86_sse2_psra_w:
02178     case llvm::Intrinsic::x86_sse2_psra_d:
02179     case llvm::Intrinsic::x86_sse2_psrli_w:
02180     case llvm::Intrinsic::x86_sse2_psrli_d:
02181     case llvm::Intrinsic::x86_sse2_psrli_q:
02182     case llvm::Intrinsic::x86_sse2_psrai_w:
02183     case llvm::Intrinsic::x86_sse2_psrai_d:
02184     case llvm::Intrinsic::x86_sse2_psrl_dq:
02185     case llvm::Intrinsic::x86_mmx_psll_w:
02186     case llvm::Intrinsic::x86_mmx_psll_d:
02187     case llvm::Intrinsic::x86_mmx_psll_q:
02188     case llvm::Intrinsic::x86_mmx_pslli_w:
02189     case llvm::Intrinsic::x86_mmx_pslli_d:
02190     case llvm::Intrinsic::x86_mmx_pslli_q:
02191     case llvm::Intrinsic::x86_mmx_psrl_w:
02192     case llvm::Intrinsic::x86_mmx_psrl_d:
02193     case llvm::Intrinsic::x86_mmx_psrl_q:
02194     case llvm::Intrinsic::x86_mmx_psra_w:
02195     case llvm::Intrinsic::x86_mmx_psra_d:
02196     case llvm::Intrinsic::x86_mmx_psrli_w:
02197     case llvm::Intrinsic::x86_mmx_psrli_d:
02198     case llvm::Intrinsic::x86_mmx_psrli_q:
02199     case llvm::Intrinsic::x86_mmx_psrai_w:
02200     case llvm::Intrinsic::x86_mmx_psrai_d:
02201       handleVectorShiftIntrinsic(I, /* Variable */ false);
02202       break;
02203     case llvm::Intrinsic::x86_avx2_psllv_d:
02204     case llvm::Intrinsic::x86_avx2_psllv_d_256:
02205     case llvm::Intrinsic::x86_avx2_psllv_q:
02206     case llvm::Intrinsic::x86_avx2_psllv_q_256:
02207     case llvm::Intrinsic::x86_avx2_psrlv_d:
02208     case llvm::Intrinsic::x86_avx2_psrlv_d_256:
02209     case llvm::Intrinsic::x86_avx2_psrlv_q:
02210     case llvm::Intrinsic::x86_avx2_psrlv_q_256:
02211     case llvm::Intrinsic::x86_avx2_psrav_d:
02212     case llvm::Intrinsic::x86_avx2_psrav_d_256:
02213       handleVectorShiftIntrinsic(I, /* Variable */ true);
02214       break;
02215 
02216     // Byte shifts are not implemented.
02217     // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
02218     // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
02219     // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
02220     // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
02221     // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
02222     // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
02223 
02224     case llvm::Intrinsic::x86_sse2_packsswb_128:
02225     case llvm::Intrinsic::x86_sse2_packssdw_128:
02226     case llvm::Intrinsic::x86_sse2_packuswb_128:
02227     case llvm::Intrinsic::x86_sse41_packusdw:
02228     case llvm::Intrinsic::x86_avx2_packsswb:
02229     case llvm::Intrinsic::x86_avx2_packssdw:
02230     case llvm::Intrinsic::x86_avx2_packuswb:
02231     case llvm::Intrinsic::x86_avx2_packusdw:
02232       handleVectorPackIntrinsic(I);
02233       break;
02234 
02235     case llvm::Intrinsic::x86_mmx_packsswb:
02236     case llvm::Intrinsic::x86_mmx_packuswb:
02237       handleVectorPackIntrinsic(I, 16);
02238       break;
02239 
02240     case llvm::Intrinsic::x86_mmx_packssdw:
02241       handleVectorPackIntrinsic(I, 32);
02242       break;
02243 
02244     case llvm::Intrinsic::x86_mmx_psad_bw:
02245     case llvm::Intrinsic::x86_sse2_psad_bw:
02246     case llvm::Intrinsic::x86_avx2_psad_bw:
02247       handleVectorSadIntrinsic(I);
02248       break;
02249 
02250     case llvm::Intrinsic::x86_sse2_pmadd_wd:
02251     case llvm::Intrinsic::x86_avx2_pmadd_wd:
02252     case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
02253     case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
02254       handleVectorPmaddIntrinsic(I);
02255       break;
02256 
02257     case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
02258       handleVectorPmaddIntrinsic(I, 8);
02259       break;
02260 
02261     case llvm::Intrinsic::x86_mmx_pmadd_wd:
02262       handleVectorPmaddIntrinsic(I, 16);
02263       break;
02264 
02265     default:
02266       if (!handleUnknownIntrinsic(I))
02267         visitInstruction(I);
02268       break;
02269     }
02270   }
02271 
02272   void visitCallSite(CallSite CS) {
02273     Instruction &I = *CS.getInstruction();
02274     assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
02275     if (CS.isCall()) {
02276       CallInst *Call = cast<CallInst>(&I);
02277 
02278       // For inline asm, do the usual thing: check argument shadow and mark all
02279       // outputs as clean. Note that any side effects of the inline asm that are
02280       // not immediately visible in its constraints are not handled.
02281       if (Call->isInlineAsm()) {
02282         visitInstruction(I);
02283         return;
02284       }
02285 
02286       assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
02287 
02288       // We are going to insert code that relies on the fact that the callee
02289       // will become a non-readonly function after it is instrumented by us. To
02290       // prevent this code from being optimized out, mark that function
02291       // non-readonly in advance.
02292       if (Function *Func = Call->getCalledFunction()) {
02293         // Clear out readonly/readnone attributes.
02294         AttrBuilder B;
02295         B.addAttribute(Attribute::ReadOnly)
02296           .addAttribute(Attribute::ReadNone);
02297         Func->removeAttributes(AttributeSet::FunctionIndex,
02298                                AttributeSet::get(Func->getContext(),
02299                                                  AttributeSet::FunctionIndex,
02300                                                  B));
02301       }
02302     }
02303     IRBuilder<> IRB(&I);
02304 
02305     if (MS.WrapIndirectCalls && !CS.getCalledFunction())
02306       IndirectCallList.push_back(CS);
02307 
02308     unsigned ArgOffset = 0;
02309     DEBUG(dbgs() << "  CallSite: " << I << "\n");
02310     for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
02311          ArgIt != End; ++ArgIt) {
02312       Value *A = *ArgIt;
02313       unsigned i = ArgIt - CS.arg_begin();
02314       if (!A->getType()->isSized()) {
02315         DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
02316         continue;
02317       }
02318       unsigned Size = 0;
02319       Value *Store = nullptr;
02320       // Compute the Shadow for arg even if it is ByVal, because
02321       // in that case getShadow() will copy the actual arg shadow to
02322       // __msan_param_tls.
02323       Value *ArgShadow = getShadow(A);
02324       Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
02325       DEBUG(dbgs() << "  Arg#" << i << ": " << *A <<
02326             " Shadow: " << *ArgShadow << "\n");
02327       bool ArgIsInitialized = false;
02328       if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
02329         assert(A->getType()->isPointerTy() &&
02330                "ByVal argument is not a pointer!");
02331         Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
02332         unsigned Alignment = CS.getParamAlignment(i + 1);
02333         Store = IRB.CreateMemCpy(ArgShadowBase,
02334                                  getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
02335                                  Size, Alignment);
02336       } else {
02337         Size = MS.DL->getTypeAllocSize(A->getType());
02338         Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
02339                                        kShadowTLSAlignment);
02340         Constant *Cst = dyn_cast<Constant>(ArgShadow);
02341         if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
02342       }
02343       if (MS.TrackOrigins && !ArgIsInitialized)
02344         IRB.CreateStore(getOrigin(A),
02345                         getOriginPtrForArgument(A, IRB, ArgOffset));
02346       (void)Store;
02347       assert(Size != 0 && Store != nullptr);
02348       DEBUG(dbgs() << "  Param:" << *Store << "\n");
02349       ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
02350     }
02351     DEBUG(dbgs() << "  done with call args\n");
02352 
02353     FunctionType *FT =
02354       cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
02355     if (FT->isVarArg()) {
02356       VAHelper->visitCallSite(CS, IRB);
02357     }
02358 
02359     // Now, get the shadow for the RetVal.
02360     if (!I.getType()->isSized()) return;
02361     IRBuilder<> IRBBefore(&I);
02362     // Until we have full dynamic coverage, make sure the retval shadow is 0.
02363     Value *Base = getShadowPtrForRetval(&I, IRBBefore);
02364     IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
02365     Instruction *NextInsn = nullptr;
02366     if (CS.isCall()) {
02367       NextInsn = I.getNextNode();
02368     } else {
02369       BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
02370       if (!NormalDest->getSinglePredecessor()) {
02371         // FIXME: this case is tricky, so we are just conservative here.
02372         // Perhaps we need to split the edge between this BB and NormalDest,
02373         // but a naive attempt to use SplitEdge leads to a crash.
02374         setShadow(&I, getCleanShadow(&I));
02375         setOrigin(&I, getCleanOrigin());
02376         return;
02377       }
02378       NextInsn = NormalDest->getFirstInsertionPt();
02379       assert(NextInsn &&
02380              "Could not find insertion point for retval shadow load");
02381     }
02382     IRBuilder<> IRBAfter(NextInsn);
02383     Value *RetvalShadow =
02384       IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
02385                                  kShadowTLSAlignment, "_msret");
02386     setShadow(&I, RetvalShadow);
02387     if (MS.TrackOrigins)
02388       setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
02389   }
02390 
02391   void visitReturnInst(ReturnInst &I) {
02392     IRBuilder<> IRB(&I);
02393     Value *RetVal = I.getReturnValue();
02394     if (!RetVal) return;
02395     Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
02396     if (CheckReturnValue) {
02397       insertShadowCheck(RetVal, &I);
02398       Value *Shadow = getCleanShadow(RetVal);
02399       IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
02400     } else {
02401       Value *Shadow = getShadow(RetVal);
02402       IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
02403       // FIXME: make it conditional if ClStoreCleanOrigin==0
02404       if (MS.TrackOrigins)
02405         IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
02406     }
02407   }
02408 
02409   void visitPHINode(PHINode &I) {
02410     IRBuilder<> IRB(&I);
02411     if (!PropagateShadow) {
02412       setShadow(&I, getCleanShadow(&I));
02413       return;
02414     }
02415 
02416     ShadowPHINodes.push_back(&I);
02417     setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
02418                                 "_msphi_s"));
02419     if (MS.TrackOrigins)
02420       setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
02421                                   "_msphi_o"));
02422   }
02423 
02424   void visitAllocaInst(AllocaInst &I) {
02425     setShadow(&I, getCleanShadow(&I));
02426     IRBuilder<> IRB(I.getNextNode());
02427     uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
02428     if (PoisonStack && ClPoisonStackWithCall) {
02429       IRB.CreateCall2(MS.MsanPoisonStackFn,
02430                       IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
02431                       ConstantInt::get(MS.IntptrTy, Size));
02432     } else {
02433       Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
02434       Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
02435       IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
02436     }
02437 
02438     if (PoisonStack && MS.TrackOrigins) {
02439       setOrigin(&I, getCleanOrigin());
02440       SmallString<2048> StackDescriptionStorage;
02441       raw_svector_ostream StackDescription(StackDescriptionStorage);
02442       // We create a string with a description of the stack allocation and
02443       // pass it into __msan_set_alloca_origin.
02444       // It will be printed by the run-time if stack-originated UMR is found.
02445       // The first 4 bytes of the string are set to '----' and will be replaced
02446       // by __msan_va_arg_overflow_size_tls at the first call.
02447       StackDescription << "----" << I.getName() << "@" << F.getName();
02448       Value *Descr =
02449           createPrivateNonConstGlobalForString(*F.getParent(),
02450                                                StackDescription.str());
02451 
02452       IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
02453                       IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
02454                       ConstantInt::get(MS.IntptrTy, Size),
02455                       IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
02456                       IRB.CreatePointerCast(&F, MS.IntptrTy));
02457     }
02458   }
02459 
02460   void visitSelectInst(SelectInst& I) {
02461     IRBuilder<> IRB(&I);
02462     // a = select b, c, d
02463     Value *B = I.getCondition();
02464     Value *C = I.getTrueValue();
02465     Value *D = I.getFalseValue();
02466     Value *Sb = getShadow(B);
02467     Value *Sc = getShadow(C);
02468     Value *Sd = getShadow(D);
02469 
02470     // Result shadow if condition shadow is 0.
02471     Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
02472     Value *Sa1;
02473     if (I.getType()->isAggregateType()) {
02474       // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
02475       // an extra "select". This results in much more compact IR.
02476       // Sa = select Sb, poisoned, (select b, Sc, Sd)
02477       Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
02478     } else {
02479       // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
02480       // If Sb (condition is poisoned), look for bits in c and d that are equal
02481       // and both unpoisoned.
02482       // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
02483 
02484       // Cast arguments to shadow-compatible type.
02485       C = CreateAppToShadowCast(IRB, C);
02486       D = CreateAppToShadowCast(IRB, D);
02487 
02488       // Result shadow if condition shadow is 1.
02489       Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
02490     }
02491     Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
02492     setShadow(&I, Sa);
02493     if (MS.TrackOrigins) {
02494       // Origins are always i32, so any vector conditions must be flattened.
02495       // FIXME: consider tracking vector origins for app vectors?
02496       if (B->getType()->isVectorTy()) {
02497         Type *FlatTy = getShadowTyNoVec(B->getType());
02498         B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
02499                                 ConstantInt::getNullValue(FlatTy));
02500         Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
02501                                       ConstantInt::getNullValue(FlatTy));
02502       }
02503       // a = select b, c, d
02504       // Oa = Sb ? Ob : (b ? Oc : Od)
02505       setOrigin(&I, IRB.CreateSelect(
02506                         Sb, getOrigin(I.getCondition()),
02507                         IRB.CreateSelect(B, getOrigin(C), getOrigin(D))));
02508     }
02509   }
02510 
02511   void visitLandingPadInst(LandingPadInst &I) {
02512     // Do nothing.
02513     // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
02514     setShadow(&I, getCleanShadow(&I));
02515     setOrigin(&I, getCleanOrigin());
02516   }
02517 
02518   void visitGetElementPtrInst(GetElementPtrInst &I) {
02519     handleShadowOr(I);
02520   }
02521 
02522   void visitExtractValueInst(ExtractValueInst &I) {
02523     IRBuilder<> IRB(&I);
02524     Value *Agg = I.getAggregateOperand();
02525     DEBUG(dbgs() << "ExtractValue:  " << I << "\n");
02526     Value *AggShadow = getShadow(Agg);
02527     DEBUG(dbgs() << "   AggShadow:  " << *AggShadow << "\n");
02528     Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
02529     DEBUG(dbgs() << "   ResShadow:  " << *ResShadow << "\n");
02530     setShadow(&I, ResShadow);
02531     setOriginForNaryOp(I);
02532   }
02533 
02534   void visitInsertValueInst(InsertValueInst &I) {
02535     IRBuilder<> IRB(&I);
02536     DEBUG(dbgs() << "InsertValue:  " << I << "\n");
02537     Value *AggShadow = getShadow(I.getAggregateOperand());
02538     Value *InsShadow = getShadow(I.getInsertedValueOperand());
02539     DEBUG(dbgs() << "   AggShadow:  " << *AggShadow << "\n");
02540     DEBUG(dbgs() << "   InsShadow:  " << *InsShadow << "\n");
02541     Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
02542     DEBUG(dbgs() << "   Res:        " << *Res << "\n");
02543     setShadow(&I, Res);
02544     setOriginForNaryOp(I);
02545   }
02546 
02547   void dumpInst(Instruction &I) {
02548     if (CallInst *CI = dyn_cast<CallInst>(&I)) {
02549       errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
02550     } else {
02551       errs() << "ZZZ " << I.getOpcodeName() << "\n";
02552     }
02553     errs() << "QQQ " << I << "\n";
02554   }
02555 
02556   void visitResumeInst(ResumeInst &I) {
02557     DEBUG(dbgs() << "Resume: " << I << "\n");
02558     // Nothing to do here.
02559   }
02560 
02561   void visitInstruction(Instruction &I) {
02562     // Everything else: stop propagating and check for poisoned shadow.
02563     if (ClDumpStrictInstructions)
02564       dumpInst(I);
02565     DEBUG(dbgs() << "DEFAULT: " << I << "\n");
02566     for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
02567       insertShadowCheck(I.getOperand(i), &I);
02568     setShadow(&I, getCleanShadow(&I));
02569     setOrigin(&I, getCleanOrigin());
02570   }
02571 };
02572 
02573 /// \brief AMD64-specific implementation of VarArgHelper.
02574 struct VarArgAMD64Helper : public VarArgHelper {
02575   // An unfortunate workaround for asymmetric lowering of va_arg stuff.
02576   // See a comment in visitCallSite for more details.
02577   static const unsigned AMD64GpEndOffset = 48;  // AMD64 ABI Draft 0.99.6 p3.5.7
02578   static const unsigned AMD64FpEndOffset = 176;
02579 
02580   Function &F;
02581   MemorySanitizer &MS;
02582   MemorySanitizerVisitor &MSV;
02583   Value *VAArgTLSCopy;
02584   Value *VAArgOverflowSize;
02585 
02586   SmallVector<CallInst*, 16> VAStartInstrumentationList;
02587 
02588   VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
02589                     MemorySanitizerVisitor &MSV)
02590     : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
02591       VAArgOverflowSize(nullptr) {}
02592 
02593   enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
02594 
02595   ArgKind classifyArgument(Value* arg) {
02596     // A very rough approximation of X86_64 argument classification rules.
02597     Type *T = arg->getType();
02598     if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
02599       return AK_FloatingPoint;
02600     if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
02601       return AK_GeneralPurpose;
02602     if (T->isPointerTy())
02603       return AK_GeneralPurpose;
02604     return AK_Memory;
02605   }
02606 
02607   // For VarArg functions, store the argument shadow in an ABI-specific format
02608   // that corresponds to va_list layout.
02609   // We do this because Clang lowers va_arg in the frontend, and this pass
02610   // only sees the low level code that deals with va_list internals.
02611   // A much easier alternative (provided that Clang emits va_arg instructions)
02612   // would have been to associate each live instance of va_list with a copy of
02613   // MSanParamTLS, and extract shadow on va_arg() call in the argument list
02614   // order.
02615   void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
02616     unsigned GpOffset = 0;
02617     unsigned FpOffset = AMD64GpEndOffset;
02618     unsigned OverflowOffset = AMD64FpEndOffset;
02619     for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
02620          ArgIt != End; ++ArgIt) {
02621       Value *A = *ArgIt;
02622       unsigned ArgNo = CS.getArgumentNo(ArgIt);
02623       bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
02624       if (IsByVal) {
02625         // ByVal arguments always go to the overflow area.
02626         assert(A->getType()->isPointerTy());
02627         Type *RealTy = A->getType()->getPointerElementType();
02628         uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
02629         Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
02630         OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
02631         IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
02632                          ArgSize, kShadowTLSAlignment);
02633       } else {
02634         ArgKind AK = classifyArgument(A);
02635         if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
02636           AK = AK_Memory;
02637         if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
02638           AK = AK_Memory;
02639         Value *Base;
02640         switch (AK) {
02641           case AK_GeneralPurpose:
02642             Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
02643             GpOffset += 8;
02644             break;
02645           case AK_FloatingPoint:
02646             Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
02647             FpOffset += 16;
02648             break;
02649           case AK_Memory:
02650             uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
02651             Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
02652             OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
02653         }
02654         IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
02655       }
02656     }
02657     Constant *OverflowSize =
02658       ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
02659     IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
02660   }
02661 
02662   /// \brief Compute the shadow address for a given va_arg.
02663   Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
02664                                    int ArgOffset) {
02665     Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
02666     Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
02667     return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
02668                               "_msarg");
02669   }
02670 
02671   void visitVAStartInst(VAStartInst &I) override {
02672     IRBuilder<> IRB(&I);
02673     VAStartInstrumentationList.push_back(&I);
02674     Value *VAListTag = I.getArgOperand(0);
02675     Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
02676 
02677     // Unpoison the whole __va_list_tag.
02678     // FIXME: magic ABI constants.
02679     IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
02680                      /* size */24, /* alignment */8, false);
02681   }
02682 
02683   void visitVACopyInst(VACopyInst &I) override {
02684     IRBuilder<> IRB(&I);
02685     Value *VAListTag = I.getArgOperand(0);
02686     Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
02687 
02688     // Unpoison the whole __va_list_tag.
02689     // FIXME: magic ABI constants.
02690     IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
02691                      /* size */24, /* alignment */8, false);
02692   }
02693 
02694   void finalizeInstrumentation() override {
02695     assert(!VAArgOverflowSize && !VAArgTLSCopy &&
02696            "finalizeInstrumentation called twice");
02697     if (!VAStartInstrumentationList.empty()) {
02698       // If there is a va_start in this function, make a backup copy of
02699       // va_arg_tls somewhere in the function entry block.
02700       IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
02701       VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
02702       Value *CopySize =
02703         IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
02704                       VAArgOverflowSize);
02705       VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
02706       IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
02707     }
02708 
02709     // Instrument va_start.
02710     // Copy va_list shadow from the backup copy of the TLS contents.
02711     for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
02712       CallInst *OrigInst = VAStartInstrumentationList[i];
02713       IRBuilder<> IRB(OrigInst->getNextNode());
02714       Value *VAListTag = OrigInst->getArgOperand(0);
02715 
02716       Value *RegSaveAreaPtrPtr =
02717         IRB.CreateIntToPtr(
02718           IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
02719                         ConstantInt::get(MS.IntptrTy, 16)),
02720           Type::getInt64PtrTy(*MS.C));
02721       Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
02722       Value *RegSaveAreaShadowPtr =
02723         MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
02724       IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
02725                        AMD64FpEndOffset, 16);
02726 
02727       Value *OverflowArgAreaPtrPtr =
02728         IRB.CreateIntToPtr(
02729           IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
02730                         ConstantInt::get(MS.IntptrTy, 8)),
02731           Type::getInt64PtrTy(*MS.C));
02732       Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
02733       Value *OverflowArgAreaShadowPtr =
02734         MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
02735       Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
02736       IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
02737     }
02738   }
02739 };
02740 
02741 /// \brief A no-op implementation of VarArgHelper.
02742 struct VarArgNoOpHelper : public VarArgHelper {
02743   VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
02744                    MemorySanitizerVisitor &MSV) {}
02745 
02746   void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
02747 
02748   void visitVAStartInst(VAStartInst &I) override {}
02749 
02750   void visitVACopyInst(VACopyInst &I) override {}
02751 
02752   void finalizeInstrumentation() override {}
02753 };
02754 
02755 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
02756                                  MemorySanitizerVisitor &Visitor) {
02757   // VarArg handling is only implemented on AMD64. False positives are possible
02758   // on other platforms.
02759   llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
02760   if (TargetTriple.getArch() == llvm::Triple::x86_64)
02761     return new VarArgAMD64Helper(Func, Msan, Visitor);
02762   else
02763     return new VarArgNoOpHelper(Func, Msan, Visitor);
02764 }
02765 
02766 }  // namespace
02767 
02768 bool MemorySanitizer::runOnFunction(Function &F) {
02769   MemorySanitizerVisitor Visitor(F, *this);
02770 
02771   // Clear out readonly/readnone attributes.
02772   AttrBuilder B;
02773   B.addAttribute(Attribute::ReadOnly)
02774     .addAttribute(Attribute::ReadNone);
02775   F.removeAttributes(AttributeSet::FunctionIndex,
02776                      AttributeSet::get(F.getContext(),
02777                                        AttributeSet::FunctionIndex, B));
02778 
02779   return Visitor.runOnFunction();
02780 }