1//===-- DataFlowSanitizer.cpp - dynamic data flow analysis ----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9/// \file
10/// This file is a part of DataFlowSanitizer, a generalised dynamic data flow
11/// analysis.
12///
13/// Unlike other Sanitizer tools, this tool is not designed to detect a specific
14/// class of bugs on its own.  Instead, it provides a generic dynamic data flow
15/// analysis framework to be used by clients to help detect application-specific
16/// issues within their own code.
17///
18/// The analysis is based on automatic propagation of data flow labels (also
19/// known as taint labels) through a program as it performs computation.  Each
20/// byte of application memory is backed by two bytes of shadow memory which
21/// hold the label.  On Linux/x86_64, memory is laid out as follows:
22///
23/// +--------------------+ 0x800000000000 (top of memory)
24/// | application memory |
25/// +--------------------+ 0x700000008000 (kAppAddr)
26/// |                    |
27/// |       unused       |
28/// |                    |
29/// +--------------------+ 0x200200000000 (kUnusedAddr)
30/// |    union table     |
31/// +--------------------+ 0x200000000000 (kUnionTableAddr)
32/// |   shadow memory    |
33/// +--------------------+ 0x000000010000 (kShadowAddr)
34/// | reserved by kernel |
35/// +--------------------+ 0x000000000000
36///
37/// To derive a shadow memory address from an application memory address,
38/// bits 44-46 are cleared to bring the address into the range
39/// [0x000000008000,0x100000000000).  Then the address is shifted left by 1 to
40/// account for the double byte representation of shadow labels and move the
41/// address into the shadow memory range.  See the function
42/// DataFlowSanitizer::getShadowAddress below.
43///
44/// For more information, please refer to the design document:
45/// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html
46
47#include "llvm/Transforms/Instrumentation.h"
48#include "llvm/ADT/DenseMap.h"
49#include "llvm/ADT/DenseSet.h"
50#include "llvm/ADT/DepthFirstIterator.h"
51#include "llvm/ADT/StringExtras.h"
52#include "llvm/Analysis/ValueTracking.h"
53#include "llvm/IR/IRBuilder.h"
54#include "llvm/IR/InlineAsm.h"
55#include "llvm/IR/InstVisitor.h"
56#include "llvm/IR/LLVMContext.h"
57#include "llvm/IR/MDBuilder.h"
58#include "llvm/IR/Type.h"
59#include "llvm/IR/Value.h"
60#include "llvm/Pass.h"
61#include "llvm/Support/CommandLine.h"
62#include "llvm/Support/SpecialCaseList.h"
63#include "llvm/Transforms/Utils/BasicBlockUtils.h"
64#include "llvm/Transforms/Utils/Local.h"
65#include <iterator>
66
67using namespace llvm;
68
69// The -dfsan-preserve-alignment flag controls whether this pass assumes that
70// alignment requirements provided by the input IR are correct.  For example,
71// if the input IR contains a load with alignment 8, this flag will cause
72// the shadow load to have alignment 16.  This flag is disabled by default as
73// we have unfortunately encountered too much code (including Clang itself;
74// see PR14291) which performs misaligned access.
75static cl::opt<bool> ClPreserveAlignment(
76    "dfsan-preserve-alignment",
77    cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
78    cl::init(false));
79
80// The ABI list file controls how shadow parameters are passed.  The pass treats
81// every function labelled "uninstrumented" in the ABI list file as conforming
82// to the "native" (i.e. unsanitized) ABI.  Unless the ABI list contains
83// additional annotations for those functions, a call to one of those functions
84// will produce a warning message, as the labelling behaviour of the function is
85// unknown.  The other supported annotations are "functional" and "discard",
86// which are described below under DataFlowSanitizer::WrapperKind.
87static cl::opt<std::string> ClABIListFile(
88    "dfsan-abilist",
89    cl::desc("File listing native ABI functions and how the pass treats them"),
90    cl::Hidden);
91
92// Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
93// functions (see DataFlowSanitizer::InstrumentedABI below).
94static cl::opt<bool> ClArgsABI(
95    "dfsan-args-abi",
96    cl::desc("Use the argument ABI rather than the TLS ABI"),
97    cl::Hidden);
98
99// Controls whether the pass includes or ignores the labels of pointers in load
100// instructions.
101static cl::opt<bool> ClCombinePointerLabelsOnLoad(
102    "dfsan-combine-pointer-labels-on-load",
103    cl::desc("Combine the label of the pointer with the label of the data when "
104             "loading from memory."),
105    cl::Hidden, cl::init(true));
106
107// Controls whether the pass includes or ignores the labels of pointers in
108// stores instructions.
109static cl::opt<bool> ClCombinePointerLabelsOnStore(
110    "dfsan-combine-pointer-labels-on-store",
111    cl::desc("Combine the label of the pointer with the label of the data when "
112             "storing in memory."),
113    cl::Hidden, cl::init(false));
114
115static cl::opt<bool> ClDebugNonzeroLabels(
116    "dfsan-debug-nonzero-labels",
117    cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
118             "load or return with a nonzero label"),
119    cl::Hidden);
120
121namespace {
122
123StringRef GetGlobalTypeString(const GlobalValue &G) {
124  // Types of GlobalVariables are always pointer types.
125  Type *GType = G.getType()->getElementType();
126  // For now we support blacklisting struct types only.
127  if (StructType *SGType = dyn_cast<StructType>(GType)) {
128    if (!SGType->isLiteral())
129      return SGType->getName();
130  }
131  return "<unknown type>";
132}
133
134class DFSanABIList {
135  std::unique_ptr<SpecialCaseList> SCL;
136
137 public:
138  DFSanABIList(SpecialCaseList *SCL) : SCL(SCL) {}
139
140  /// Returns whether either this function or its source file are listed in the
141  /// given category.
142  bool isIn(const Function &F, const StringRef Category) const {
143    return isIn(*F.getParent(), Category) ||
144           SCL->inSection("fun", F.getName(), Category);
145  }
146
147  /// Returns whether this global alias is listed in the given category.
148  ///
149  /// If GA aliases a function, the alias's name is matched as a function name
150  /// would be.  Similarly, aliases of globals are matched like globals.
151  bool isIn(const GlobalAlias &GA, const StringRef Category) const {
152    if (isIn(*GA.getParent(), Category))
153      return true;
154
155    if (isa<FunctionType>(GA.getType()->getElementType()))
156      return SCL->inSection("fun", GA.getName(), Category);
157
158    return SCL->inSection("global", GA.getName(), Category) ||
159           SCL->inSection("type", GetGlobalTypeString(GA), Category);
160  }
161
162  /// Returns whether this module is listed in the given category.
163  bool isIn(const Module &M, const StringRef Category) const {
164    return SCL->inSection("src", M.getModuleIdentifier(), Category);
165  }
166};
167
168class DataFlowSanitizer : public ModulePass {
169  friend struct DFSanFunction;
170  friend class DFSanVisitor;
171
172  enum {
173    ShadowWidth = 16
174  };
175
176  /// Which ABI should be used for instrumented functions?
177  enum InstrumentedABI {
178    /// Argument and return value labels are passed through additional
179    /// arguments and by modifying the return type.
180    IA_Args,
181
182    /// Argument and return value labels are passed through TLS variables
183    /// __dfsan_arg_tls and __dfsan_retval_tls.
184    IA_TLS
185  };
186
187  /// How should calls to uninstrumented functions be handled?
188  enum WrapperKind {
189    /// This function is present in an uninstrumented form but we don't know
190    /// how it should be handled.  Print a warning and call the function anyway.
191    /// Don't label the return value.
192    WK_Warning,
193
194    /// This function does not write to (user-accessible) memory, and its return
195    /// value is unlabelled.
196    WK_Discard,
197
198    /// This function does not write to (user-accessible) memory, and the label
199    /// of its return value is the union of the label of its arguments.
200    WK_Functional,
201
202    /// Instead of calling the function, a custom wrapper __dfsw_F is called,
203    /// where F is the name of the function.  This function may wrap the
204    /// original function or provide its own implementation.  This is similar to
205    /// the IA_Args ABI, except that IA_Args uses a struct return type to
206    /// pass the return value shadow in a register, while WK_Custom uses an
207    /// extra pointer argument to return the shadow.  This allows the wrapped
208    /// form of the function type to be expressed in C.
209    WK_Custom
210  };
211
212  const DataLayout *DL;
213  Module *Mod;
214  LLVMContext *Ctx;
215  IntegerType *ShadowTy;
216  PointerType *ShadowPtrTy;
217  IntegerType *IntptrTy;
218  ConstantInt *ZeroShadow;
219  ConstantInt *ShadowPtrMask;
220  ConstantInt *ShadowPtrMul;
221  Constant *ArgTLS;
222  Constant *RetvalTLS;
223  void *(*GetArgTLSPtr)();
224  void *(*GetRetvalTLSPtr)();
225  Constant *GetArgTLS;
226  Constant *GetRetvalTLS;
227  FunctionType *DFSanUnionFnTy;
228  FunctionType *DFSanUnionLoadFnTy;
229  FunctionType *DFSanUnimplementedFnTy;
230  FunctionType *DFSanSetLabelFnTy;
231  FunctionType *DFSanNonzeroLabelFnTy;
232  Constant *DFSanUnionFn;
233  Constant *DFSanUnionLoadFn;
234  Constant *DFSanUnimplementedFn;
235  Constant *DFSanSetLabelFn;
236  Constant *DFSanNonzeroLabelFn;
237  MDNode *ColdCallWeights;
238  DFSanABIList ABIList;
239  DenseMap<Value *, Function *> UnwrappedFnMap;
240  AttributeSet ReadOnlyNoneAttrs;
241
242  Value *getShadowAddress(Value *Addr, Instruction *Pos);
243  Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
244  bool isInstrumented(const Function *F);
245  bool isInstrumented(const GlobalAlias *GA);
246  FunctionType *getArgsFunctionType(FunctionType *T);
247  FunctionType *getTrampolineFunctionType(FunctionType *T);
248  FunctionType *getCustomFunctionType(FunctionType *T);
249  InstrumentedABI getInstrumentedABI();
250  WrapperKind getWrapperKind(Function *F);
251  void addGlobalNamePrefix(GlobalValue *GV);
252  Function *buildWrapperFunction(Function *F, StringRef NewFName,
253                                 GlobalValue::LinkageTypes NewFLink,
254                                 FunctionType *NewFT);
255  Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
256
257 public:
258  DataFlowSanitizer(StringRef ABIListFile = StringRef(),
259                    void *(*getArgTLS)() = nullptr,
260                    void *(*getRetValTLS)() = nullptr);
261  static char ID;
262  bool doInitialization(Module &M) override;
263  bool runOnModule(Module &M) override;
264};
265
266struct DFSanFunction {
267  DataFlowSanitizer &DFS;
268  Function *F;
269  DataFlowSanitizer::InstrumentedABI IA;
270  bool IsNativeABI;
271  Value *ArgTLSPtr;
272  Value *RetvalTLSPtr;
273  AllocaInst *LabelReturnAlloca;
274  DenseMap<Value *, Value *> ValShadowMap;
275  DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
276  std::vector<std::pair<PHINode *, PHINode *> > PHIFixups;
277  DenseSet<Instruction *> SkipInsts;
278  DenseSet<Value *> NonZeroChecks;
279
280  DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
281      : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()),
282        IsNativeABI(IsNativeABI), ArgTLSPtr(nullptr), RetvalTLSPtr(nullptr),
283        LabelReturnAlloca(nullptr) {}
284  Value *getArgTLSPtr();
285  Value *getArgTLS(unsigned Index, Instruction *Pos);
286  Value *getRetvalTLS();
287  Value *getShadow(Value *V);
288  void setShadow(Instruction *I, Value *Shadow);
289  Value *combineOperandShadows(Instruction *Inst);
290  Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
291                    Instruction *Pos);
292  void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow,
293                   Instruction *Pos);
294};
295
296class DFSanVisitor : public InstVisitor<DFSanVisitor> {
297 public:
298  DFSanFunction &DFSF;
299  DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
300
301  void visitOperandShadowInst(Instruction &I);
302
303  void visitBinaryOperator(BinaryOperator &BO);
304  void visitCastInst(CastInst &CI);
305  void visitCmpInst(CmpInst &CI);
306  void visitGetElementPtrInst(GetElementPtrInst &GEPI);
307  void visitLoadInst(LoadInst &LI);
308  void visitStoreInst(StoreInst &SI);
309  void visitReturnInst(ReturnInst &RI);
310  void visitCallSite(CallSite CS);
311  void visitPHINode(PHINode &PN);
312  void visitExtractElementInst(ExtractElementInst &I);
313  void visitInsertElementInst(InsertElementInst &I);
314  void visitShuffleVectorInst(ShuffleVectorInst &I);
315  void visitExtractValueInst(ExtractValueInst &I);
316  void visitInsertValueInst(InsertValueInst &I);
317  void visitAllocaInst(AllocaInst &I);
318  void visitSelectInst(SelectInst &I);
319  void visitMemSetInst(MemSetInst &I);
320  void visitMemTransferInst(MemTransferInst &I);
321};
322
323}
324
325char DataFlowSanitizer::ID;
326INITIALIZE_PASS(DataFlowSanitizer, "dfsan",
327                "DataFlowSanitizer: dynamic data flow analysis.", false, false)
328
329ModulePass *llvm::createDataFlowSanitizerPass(StringRef ABIListFile,
330                                              void *(*getArgTLS)(),
331                                              void *(*getRetValTLS)()) {
332  return new DataFlowSanitizer(ABIListFile, getArgTLS, getRetValTLS);
333}
334
335DataFlowSanitizer::DataFlowSanitizer(StringRef ABIListFile,
336                                     void *(*getArgTLS)(),
337                                     void *(*getRetValTLS)())
338    : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS),
339      ABIList(SpecialCaseList::createOrDie(ABIListFile.empty() ? ClABIListFile
340                                                               : ABIListFile)) {
341}
342
343FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
344  llvm::SmallVector<Type *, 4> ArgTypes;
345  std::copy(T->param_begin(), T->param_end(), std::back_inserter(ArgTypes));
346  for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
347    ArgTypes.push_back(ShadowTy);
348  if (T->isVarArg())
349    ArgTypes.push_back(ShadowPtrTy);
350  Type *RetType = T->getReturnType();
351  if (!RetType->isVoidTy())
352    RetType = StructType::get(RetType, ShadowTy, (Type *)nullptr);
353  return FunctionType::get(RetType, ArgTypes, T->isVarArg());
354}
355
356FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
357  assert(!T->isVarArg());
358  llvm::SmallVector<Type *, 4> ArgTypes;
359  ArgTypes.push_back(T->getPointerTo());
360  std::copy(T->param_begin(), T->param_end(), std::back_inserter(ArgTypes));
361  for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
362    ArgTypes.push_back(ShadowTy);
363  Type *RetType = T->getReturnType();
364  if (!RetType->isVoidTy())
365    ArgTypes.push_back(ShadowPtrTy);
366  return FunctionType::get(T->getReturnType(), ArgTypes, false);
367}
368
369FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
370  assert(!T->isVarArg());
371  llvm::SmallVector<Type *, 4> ArgTypes;
372  for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end();
373       i != e; ++i) {
374    FunctionType *FT;
375    if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>(
376                                     *i)->getElementType()))) {
377      ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
378      ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
379    } else {
380      ArgTypes.push_back(*i);
381    }
382  }
383  for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
384    ArgTypes.push_back(ShadowTy);
385  Type *RetType = T->getReturnType();
386  if (!RetType->isVoidTy())
387    ArgTypes.push_back(ShadowPtrTy);
388  return FunctionType::get(T->getReturnType(), ArgTypes, false);
389}
390
391bool DataFlowSanitizer::doInitialization(Module &M) {
392  DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
393  if (!DLP)
394    report_fatal_error("data layout missing");
395  DL = &DLP->getDataLayout();
396
397  Mod = &M;
398  Ctx = &M.getContext();
399  ShadowTy = IntegerType::get(*Ctx, ShadowWidth);
400  ShadowPtrTy = PointerType::getUnqual(ShadowTy);
401  IntptrTy = DL->getIntPtrType(*Ctx);
402  ZeroShadow = ConstantInt::getSigned(ShadowTy, 0);
403  ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL);
404  ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8);
405
406  Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy };
407  DFSanUnionFnTy =
408      FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false);
409  Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy };
410  DFSanUnionLoadFnTy =
411      FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false);
412  DFSanUnimplementedFnTy = FunctionType::get(
413      Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
414  Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy };
415  DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
416                                        DFSanSetLabelArgs, /*isVarArg=*/false);
417  DFSanNonzeroLabelFnTy = FunctionType::get(
418      Type::getVoidTy(*Ctx), ArrayRef<Type *>(), /*isVarArg=*/false);
419
420  if (GetArgTLSPtr) {
421    Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
422    ArgTLS = nullptr;
423    GetArgTLS = ConstantExpr::getIntToPtr(
424        ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)),
425        PointerType::getUnqual(
426            FunctionType::get(PointerType::getUnqual(ArgTLSTy),
427                              (Type *)nullptr)));
428  }
429  if (GetRetvalTLSPtr) {
430    RetvalTLS = nullptr;
431    GetRetvalTLS = ConstantExpr::getIntToPtr(
432        ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)),
433        PointerType::getUnqual(
434            FunctionType::get(PointerType::getUnqual(ShadowTy),
435                              (Type *)nullptr)));
436  }
437
438  ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
439  return true;
440}
441
442bool DataFlowSanitizer::isInstrumented(const Function *F) {
443  return !ABIList.isIn(*F, "uninstrumented");
444}
445
446bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
447  return !ABIList.isIn(*GA, "uninstrumented");
448}
449
450DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
451  return ClArgsABI ? IA_Args : IA_TLS;
452}
453
454DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
455  if (ABIList.isIn(*F, "functional"))
456    return WK_Functional;
457  if (ABIList.isIn(*F, "discard"))
458    return WK_Discard;
459  if (ABIList.isIn(*F, "custom"))
460    return WK_Custom;
461
462  return WK_Warning;
463}
464
465void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) {
466  std::string GVName = GV->getName(), Prefix = "dfs$";
467  GV->setName(Prefix + GVName);
468
469  // Try to change the name of the function in module inline asm.  We only do
470  // this for specific asm directives, currently only ".symver", to try to avoid
471  // corrupting asm which happens to contain the symbol name as a substring.
472  // Note that the substitution for .symver assumes that the versioned symbol
473  // also has an instrumented name.
474  std::string Asm = GV->getParent()->getModuleInlineAsm();
475  std::string SearchStr = ".symver " + GVName + ",";
476  size_t Pos = Asm.find(SearchStr);
477  if (Pos != std::string::npos) {
478    Asm.replace(Pos, SearchStr.size(),
479                ".symver " + Prefix + GVName + "," + Prefix);
480    GV->getParent()->setModuleInlineAsm(Asm);
481  }
482}
483
484Function *
485DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
486                                        GlobalValue::LinkageTypes NewFLink,
487                                        FunctionType *NewFT) {
488  FunctionType *FT = F->getFunctionType();
489  Function *NewF = Function::Create(NewFT, NewFLink, NewFName,
490                                    F->getParent());
491  NewF->copyAttributesFrom(F);
492  NewF->removeAttributes(
493      AttributeSet::ReturnIndex,
494      AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
495                                       AttributeSet::ReturnIndex));
496
497  BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
498  std::vector<Value *> Args;
499  unsigned n = FT->getNumParams();
500  for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n)
501    Args.push_back(&*ai);
502  CallInst *CI = CallInst::Create(F, Args, "", BB);
503  if (FT->getReturnType()->isVoidTy())
504    ReturnInst::Create(*Ctx, BB);
505  else
506    ReturnInst::Create(*Ctx, CI, BB);
507
508  return NewF;
509}
510
511Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
512                                                          StringRef FName) {
513  FunctionType *FTT = getTrampolineFunctionType(FT);
514  Constant *C = Mod->getOrInsertFunction(FName, FTT);
515  Function *F = dyn_cast<Function>(C);
516  if (F && F->isDeclaration()) {
517    F->setLinkage(GlobalValue::LinkOnceODRLinkage);
518    BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
519    std::vector<Value *> Args;
520    Function::arg_iterator AI = F->arg_begin(); ++AI;
521    for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
522      Args.push_back(&*AI);
523    CallInst *CI =
524        CallInst::Create(&F->getArgumentList().front(), Args, "", BB);
525    ReturnInst *RI;
526    if (FT->getReturnType()->isVoidTy())
527      RI = ReturnInst::Create(*Ctx, BB);
528    else
529      RI = ReturnInst::Create(*Ctx, CI, BB);
530
531    DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
532    Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI;
533    for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N)
534      DFSF.ValShadowMap[ValAI] = ShadowAI;
535    DFSanVisitor(DFSF).visitCallInst(*CI);
536    if (!FT->getReturnType()->isVoidTy())
537      new StoreInst(DFSF.getShadow(RI->getReturnValue()),
538                    &F->getArgumentList().back(), RI);
539  }
540
541  return C;
542}
543
544bool DataFlowSanitizer::runOnModule(Module &M) {
545  if (!DL)
546    return false;
547
548  if (ABIList.isIn(M, "skip"))
549    return false;
550
551  if (!GetArgTLSPtr) {
552    Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
553    ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy);
554    if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS))
555      G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
556  }
557  if (!GetRetvalTLSPtr) {
558    RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy);
559    if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS))
560      G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
561  }
562
563  DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy);
564  if (Function *F = dyn_cast<Function>(DFSanUnionFn)) {
565    F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
566    F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
567    F->addAttribute(1, Attribute::ZExt);
568    F->addAttribute(2, Attribute::ZExt);
569  }
570  DFSanUnionLoadFn =
571      Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy);
572  if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) {
573    F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadOnly);
574    F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
575  }
576  DFSanUnimplementedFn =
577      Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
578  DFSanSetLabelFn =
579      Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy);
580  if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) {
581    F->addAttribute(1, Attribute::ZExt);
582  }
583  DFSanNonzeroLabelFn =
584      Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
585
586  std::vector<Function *> FnsToInstrument;
587  llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI;
588  for (Module::iterator i = M.begin(), e = M.end(); i != e; ++i) {
589    if (!i->isIntrinsic() &&
590        i != DFSanUnionFn &&
591        i != DFSanUnionLoadFn &&
592        i != DFSanUnimplementedFn &&
593        i != DFSanSetLabelFn &&
594        i != DFSanNonzeroLabelFn)
595      FnsToInstrument.push_back(&*i);
596  }
597
598  // Give function aliases prefixes when necessary, and build wrappers where the
599  // instrumentedness is inconsistent.
600  for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) {
601    GlobalAlias *GA = &*i;
602    ++i;
603    // Don't stop on weak.  We assume people aren't playing games with the
604    // instrumentedness of overridden weak aliases.
605    if (auto F = dyn_cast<Function>(GA->getBaseObject())) {
606      bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
607      if (GAInst && FInst) {
608        addGlobalNamePrefix(GA);
609      } else if (GAInst != FInst) {
610        // Non-instrumented alias of an instrumented function, or vice versa.
611        // Replace the alias with a native-ABI wrapper of the aliasee.  The pass
612        // below will take care of instrumenting it.
613        Function *NewF =
614            buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
615        GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
616        NewF->takeName(GA);
617        GA->eraseFromParent();
618        FnsToInstrument.push_back(NewF);
619      }
620    }
621  }
622
623  AttrBuilder B;
624  B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
625  ReadOnlyNoneAttrs = AttributeSet::get(*Ctx, AttributeSet::FunctionIndex, B);
626
627  // First, change the ABI of every function in the module.  ABI-listed
628  // functions keep their original ABI and get a wrapper function.
629  for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
630                                         e = FnsToInstrument.end();
631       i != e; ++i) {
632    Function &F = **i;
633    FunctionType *FT = F.getFunctionType();
634
635    bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
636                              FT->getReturnType()->isVoidTy());
637
638    if (isInstrumented(&F)) {
639      // Instrumented functions get a 'dfs$' prefix.  This allows us to more
640      // easily identify cases of mismatching ABIs.
641      if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
642        FunctionType *NewFT = getArgsFunctionType(FT);
643        Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M);
644        NewF->copyAttributesFrom(&F);
645        NewF->removeAttributes(
646            AttributeSet::ReturnIndex,
647            AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
648                                             AttributeSet::ReturnIndex));
649        for (Function::arg_iterator FArg = F.arg_begin(),
650                                    NewFArg = NewF->arg_begin(),
651                                    FArgEnd = F.arg_end();
652             FArg != FArgEnd; ++FArg, ++NewFArg) {
653          FArg->replaceAllUsesWith(NewFArg);
654        }
655        NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
656
657        for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
658             UI != UE;) {
659          BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
660          ++UI;
661          if (BA) {
662            BA->replaceAllUsesWith(
663                BlockAddress::get(NewF, BA->getBasicBlock()));
664            delete BA;
665          }
666        }
667        F.replaceAllUsesWith(
668            ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
669        NewF->takeName(&F);
670        F.eraseFromParent();
671        *i = NewF;
672        addGlobalNamePrefix(NewF);
673      } else {
674        addGlobalNamePrefix(&F);
675      }
676               // Hopefully, nobody will try to indirectly call a vararg
677               // function... yet.
678    } else if (FT->isVarArg()) {
679      UnwrappedFnMap[&F] = &F;
680      *i = nullptr;
681    } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
682      // Build a wrapper function for F.  The wrapper simply calls F, and is
683      // added to FnsToInstrument so that any instrumentation according to its
684      // WrapperKind is done in the second pass below.
685      FunctionType *NewFT = getInstrumentedABI() == IA_Args
686                                ? getArgsFunctionType(FT)
687                                : FT;
688      Function *NewF = buildWrapperFunction(
689          &F, std::string("dfsw$") + std::string(F.getName()),
690          GlobalValue::LinkOnceODRLinkage, NewFT);
691      if (getInstrumentedABI() == IA_TLS)
692        NewF->removeAttributes(AttributeSet::FunctionIndex, ReadOnlyNoneAttrs);
693
694      Value *WrappedFnCst =
695          ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
696      F.replaceAllUsesWith(WrappedFnCst);
697      UnwrappedFnMap[WrappedFnCst] = &F;
698      *i = NewF;
699
700      if (!F.isDeclaration()) {
701        // This function is probably defining an interposition of an
702        // uninstrumented function and hence needs to keep the original ABI.
703        // But any functions it may call need to use the instrumented ABI, so
704        // we instrument it in a mode which preserves the original ABI.
705        FnsWithNativeABI.insert(&F);
706
707        // This code needs to rebuild the iterators, as they may be invalidated
708        // by the push_back, taking care that the new range does not include
709        // any functions added by this code.
710        size_t N = i - FnsToInstrument.begin(),
711               Count = e - FnsToInstrument.begin();
712        FnsToInstrument.push_back(&F);
713        i = FnsToInstrument.begin() + N;
714        e = FnsToInstrument.begin() + Count;
715      }
716    }
717  }
718
719  for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
720                                         e = FnsToInstrument.end();
721       i != e; ++i) {
722    if (!*i || (*i)->isDeclaration())
723      continue;
724
725    removeUnreachableBlocks(**i);
726
727    DFSanFunction DFSF(*this, *i, FnsWithNativeABI.count(*i));
728
729    // DFSanVisitor may create new basic blocks, which confuses df_iterator.
730    // Build a copy of the list before iterating over it.
731    llvm::SmallVector<BasicBlock *, 4> BBList(
732        depth_first(&(*i)->getEntryBlock()));
733
734    for (llvm::SmallVector<BasicBlock *, 4>::iterator i = BBList.begin(),
735                                                      e = BBList.end();
736         i != e; ++i) {
737      Instruction *Inst = &(*i)->front();
738      while (1) {
739        // DFSanVisitor may split the current basic block, changing the current
740        // instruction's next pointer and moving the next instruction to the
741        // tail block from which we should continue.
742        Instruction *Next = Inst->getNextNode();
743        // DFSanVisitor may delete Inst, so keep track of whether it was a
744        // terminator.
745        bool IsTerminator = isa<TerminatorInst>(Inst);
746        if (!DFSF.SkipInsts.count(Inst))
747          DFSanVisitor(DFSF).visit(Inst);
748        if (IsTerminator)
749          break;
750        Inst = Next;
751      }
752    }
753
754    // We will not necessarily be able to compute the shadow for every phi node
755    // until we have visited every block.  Therefore, the code that handles phi
756    // nodes adds them to the PHIFixups list so that they can be properly
757    // handled here.
758    for (std::vector<std::pair<PHINode *, PHINode *> >::iterator
759             i = DFSF.PHIFixups.begin(),
760             e = DFSF.PHIFixups.end();
761         i != e; ++i) {
762      for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n;
763           ++val) {
764        i->second->setIncomingValue(
765            val, DFSF.getShadow(i->first->getIncomingValue(val)));
766      }
767    }
768
769    // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
770    // places (i.e. instructions in basic blocks we haven't even begun visiting
771    // yet).  To make our life easier, do this work in a pass after the main
772    // instrumentation.
773    if (ClDebugNonzeroLabels) {
774      for (DenseSet<Value *>::iterator i = DFSF.NonZeroChecks.begin(),
775                                       e = DFSF.NonZeroChecks.end();
776           i != e; ++i) {
777        Instruction *Pos;
778        if (Instruction *I = dyn_cast<Instruction>(*i))
779          Pos = I->getNextNode();
780        else
781          Pos = DFSF.F->getEntryBlock().begin();
782        while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
783          Pos = Pos->getNextNode();
784        IRBuilder<> IRB(Pos);
785        Value *Ne = IRB.CreateICmpNE(*i, DFSF.DFS.ZeroShadow);
786        BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
787            Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
788        IRBuilder<> ThenIRB(BI);
789        ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn);
790      }
791    }
792  }
793
794  return false;
795}
796
797Value *DFSanFunction::getArgTLSPtr() {
798  if (ArgTLSPtr)
799    return ArgTLSPtr;
800  if (DFS.ArgTLS)
801    return ArgTLSPtr = DFS.ArgTLS;
802
803  IRBuilder<> IRB(F->getEntryBlock().begin());
804  return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS);
805}
806
807Value *DFSanFunction::getRetvalTLS() {
808  if (RetvalTLSPtr)
809    return RetvalTLSPtr;
810  if (DFS.RetvalTLS)
811    return RetvalTLSPtr = DFS.RetvalTLS;
812
813  IRBuilder<> IRB(F->getEntryBlock().begin());
814  return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS);
815}
816
817Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) {
818  IRBuilder<> IRB(Pos);
819  return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx);
820}
821
822Value *DFSanFunction::getShadow(Value *V) {
823  if (!isa<Argument>(V) && !isa<Instruction>(V))
824    return DFS.ZeroShadow;
825  Value *&Shadow = ValShadowMap[V];
826  if (!Shadow) {
827    if (Argument *A = dyn_cast<Argument>(V)) {
828      if (IsNativeABI)
829        return DFS.ZeroShadow;
830      switch (IA) {
831      case DataFlowSanitizer::IA_TLS: {
832        Value *ArgTLSPtr = getArgTLSPtr();
833        Instruction *ArgTLSPos =
834            DFS.ArgTLS ? &*F->getEntryBlock().begin()
835                       : cast<Instruction>(ArgTLSPtr)->getNextNode();
836        IRBuilder<> IRB(ArgTLSPos);
837        Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos));
838        break;
839      }
840      case DataFlowSanitizer::IA_Args: {
841        unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2;
842        Function::arg_iterator i = F->arg_begin();
843        while (ArgIdx--)
844          ++i;
845        Shadow = i;
846        assert(Shadow->getType() == DFS.ShadowTy);
847        break;
848      }
849      }
850      NonZeroChecks.insert(Shadow);
851    } else {
852      Shadow = DFS.ZeroShadow;
853    }
854  }
855  return Shadow;
856}
857
858void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
859  assert(!ValShadowMap.count(I));
860  assert(Shadow->getType() == DFS.ShadowTy);
861  ValShadowMap[I] = Shadow;
862}
863
864Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
865  assert(Addr != RetvalTLS && "Reinstrumenting?");
866  IRBuilder<> IRB(Pos);
867  return IRB.CreateIntToPtr(
868      IRB.CreateMul(
869          IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), ShadowPtrMask),
870          ShadowPtrMul),
871      ShadowPtrTy);
872}
873
874// Generates IR to compute the union of the two given shadows, inserting it
875// before Pos.  Returns the computed union Value.
876Value *DataFlowSanitizer::combineShadows(Value *V1, Value *V2,
877                                         Instruction *Pos) {
878  if (V1 == ZeroShadow)
879    return V2;
880  if (V2 == ZeroShadow)
881    return V1;
882  if (V1 == V2)
883    return V1;
884  IRBuilder<> IRB(Pos);
885  BasicBlock *Head = Pos->getParent();
886  Value *Ne = IRB.CreateICmpNE(V1, V2);
887  BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
888      Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
889  IRBuilder<> ThenIRB(BI);
890  CallInst *Call = ThenIRB.CreateCall2(DFSanUnionFn, V1, V2);
891  Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
892  Call->addAttribute(1, Attribute::ZExt);
893  Call->addAttribute(2, Attribute::ZExt);
894
895  BasicBlock *Tail = BI->getSuccessor(0);
896  PHINode *Phi = PHINode::Create(ShadowTy, 2, "", Tail->begin());
897  Phi->addIncoming(Call, Call->getParent());
898  Phi->addIncoming(V1, Head);
899  return Phi;
900}
901
902// A convenience function which folds the shadows of each of the operands
903// of the provided instruction Inst, inserting the IR before Inst.  Returns
904// the computed union Value.
905Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
906  if (Inst->getNumOperands() == 0)
907    return DFS.ZeroShadow;
908
909  Value *Shadow = getShadow(Inst->getOperand(0));
910  for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) {
911    Shadow = DFS.combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst);
912  }
913  return Shadow;
914}
915
916void DFSanVisitor::visitOperandShadowInst(Instruction &I) {
917  Value *CombinedShadow = DFSF.combineOperandShadows(&I);
918  DFSF.setShadow(&I, CombinedShadow);
919}
920
921// Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
922// Addr has alignment Align, and take the union of each of those shadows.
923Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
924                                 Instruction *Pos) {
925  if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
926    llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
927        AllocaShadowMap.find(AI);
928    if (i != AllocaShadowMap.end()) {
929      IRBuilder<> IRB(Pos);
930      return IRB.CreateLoad(i->second);
931    }
932  }
933
934  uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
935  SmallVector<Value *, 2> Objs;
936  GetUnderlyingObjects(Addr, Objs, DFS.DL);
937  bool AllConstants = true;
938  for (SmallVector<Value *, 2>::iterator i = Objs.begin(), e = Objs.end();
939       i != e; ++i) {
940    if (isa<Function>(*i) || isa<BlockAddress>(*i))
941      continue;
942    if (isa<GlobalVariable>(*i) && cast<GlobalVariable>(*i)->isConstant())
943      continue;
944
945    AllConstants = false;
946    break;
947  }
948  if (AllConstants)
949    return DFS.ZeroShadow;
950
951  Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
952  switch (Size) {
953  case 0:
954    return DFS.ZeroShadow;
955  case 1: {
956    LoadInst *LI = new LoadInst(ShadowAddr, "", Pos);
957    LI->setAlignment(ShadowAlign);
958    return LI;
959  }
960  case 2: {
961    IRBuilder<> IRB(Pos);
962    Value *ShadowAddr1 =
963        IRB.CreateGEP(ShadowAddr, ConstantInt::get(DFS.IntptrTy, 1));
964    return DFS.combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign),
965                              IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign),
966                              Pos);
967  }
968  }
969  if (Size % (64 / DFS.ShadowWidth) == 0) {
970    // Fast path for the common case where each byte has identical shadow: load
971    // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
972    // shadow is non-equal.
973    BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F);
974    IRBuilder<> FallbackIRB(FallbackBB);
975    CallInst *FallbackCall = FallbackIRB.CreateCall2(
976        DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
977    FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
978
979    // Compare each of the shadows stored in the loaded 64 bits to each other,
980    // by computing (WideShadow rotl ShadowWidth) == WideShadow.
981    IRBuilder<> IRB(Pos);
982    Value *WideAddr =
983        IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
984    Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign);
985    Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy);
986    Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth);
987    Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth);
988    Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow);
989    Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow);
990
991    BasicBlock *Head = Pos->getParent();
992    BasicBlock *Tail = Head->splitBasicBlock(Pos);
993    // In the following code LastBr will refer to the previous basic block's
994    // conditional branch instruction, whose true successor is fixed up to point
995    // to the next block during the loop below or to the tail after the final
996    // iteration.
997    BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq);
998    ReplaceInstWithInst(Head->getTerminator(), LastBr);
999
1000    for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size;
1001         Ofs += 64 / DFS.ShadowWidth) {
1002      BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F);
1003      IRBuilder<> NextIRB(NextBB);
1004      WideAddr = NextIRB.CreateGEP(WideAddr, ConstantInt::get(DFS.IntptrTy, 1));
1005      Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1006      ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow);
1007      LastBr->setSuccessor(0, NextBB);
1008      LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB);
1009    }
1010
1011    LastBr->setSuccessor(0, Tail);
1012    FallbackIRB.CreateBr(Tail);
1013    PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1014    Shadow->addIncoming(FallbackCall, FallbackBB);
1015    Shadow->addIncoming(TruncShadow, LastBr->getParent());
1016    return Shadow;
1017  }
1018
1019  IRBuilder<> IRB(Pos);
1020  CallInst *FallbackCall = IRB.CreateCall2(
1021      DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
1022  FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
1023  return FallbackCall;
1024}
1025
1026void DFSanVisitor::visitLoadInst(LoadInst &LI) {
1027  uint64_t Size = DFSF.DFS.DL->getTypeStoreSize(LI.getType());
1028  uint64_t Align;
1029  if (ClPreserveAlignment) {
1030    Align = LI.getAlignment();
1031    if (Align == 0)
1032      Align = DFSF.DFS.DL->getABITypeAlignment(LI.getType());
1033  } else {
1034    Align = 1;
1035  }
1036  IRBuilder<> IRB(&LI);
1037  Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI);
1038  if (ClCombinePointerLabelsOnLoad) {
1039    Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
1040    Shadow = DFSF.DFS.combineShadows(Shadow, PtrShadow, &LI);
1041  }
1042  if (Shadow != DFSF.DFS.ZeroShadow)
1043    DFSF.NonZeroChecks.insert(Shadow);
1044
1045  DFSF.setShadow(&LI, Shadow);
1046}
1047
1048void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align,
1049                                Value *Shadow, Instruction *Pos) {
1050  if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1051    llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1052        AllocaShadowMap.find(AI);
1053    if (i != AllocaShadowMap.end()) {
1054      IRBuilder<> IRB(Pos);
1055      IRB.CreateStore(Shadow, i->second);
1056      return;
1057    }
1058  }
1059
1060  uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1061  IRBuilder<> IRB(Pos);
1062  Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1063  if (Shadow == DFS.ZeroShadow) {
1064    IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth);
1065    Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
1066    Value *ExtShadowAddr =
1067        IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
1068    IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
1069    return;
1070  }
1071
1072  const unsigned ShadowVecSize = 128 / DFS.ShadowWidth;
1073  uint64_t Offset = 0;
1074  if (Size >= ShadowVecSize) {
1075    VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize);
1076    Value *ShadowVec = UndefValue::get(ShadowVecTy);
1077    for (unsigned i = 0; i != ShadowVecSize; ++i) {
1078      ShadowVec = IRB.CreateInsertElement(
1079          ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i));
1080    }
1081    Value *ShadowVecAddr =
1082        IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
1083    do {
1084      Value *CurShadowVecAddr = IRB.CreateConstGEP1_32(ShadowVecAddr, Offset);
1085      IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
1086      Size -= ShadowVecSize;
1087      ++Offset;
1088    } while (Size >= ShadowVecSize);
1089    Offset *= ShadowVecSize;
1090  }
1091  while (Size > 0) {
1092    Value *CurShadowAddr = IRB.CreateConstGEP1_32(ShadowAddr, Offset);
1093    IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign);
1094    --Size;
1095    ++Offset;
1096  }
1097}
1098
1099void DFSanVisitor::visitStoreInst(StoreInst &SI) {
1100  uint64_t Size =
1101      DFSF.DFS.DL->getTypeStoreSize(SI.getValueOperand()->getType());
1102  uint64_t Align;
1103  if (ClPreserveAlignment) {
1104    Align = SI.getAlignment();
1105    if (Align == 0)
1106      Align = DFSF.DFS.DL->getABITypeAlignment(SI.getValueOperand()->getType());
1107  } else {
1108    Align = 1;
1109  }
1110
1111  Value* Shadow = DFSF.getShadow(SI.getValueOperand());
1112  if (ClCombinePointerLabelsOnStore) {
1113    Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
1114    Shadow = DFSF.DFS.combineShadows(Shadow, PtrShadow, &SI);
1115  }
1116  DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI);
1117}
1118
1119void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
1120  visitOperandShadowInst(BO);
1121}
1122
1123void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); }
1124
1125void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); }
1126
1127void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
1128  visitOperandShadowInst(GEPI);
1129}
1130
1131void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
1132  visitOperandShadowInst(I);
1133}
1134
1135void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
1136  visitOperandShadowInst(I);
1137}
1138
1139void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
1140  visitOperandShadowInst(I);
1141}
1142
1143void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
1144  visitOperandShadowInst(I);
1145}
1146
1147void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
1148  visitOperandShadowInst(I);
1149}
1150
1151void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
1152  bool AllLoadsStores = true;
1153  for (User *U : I.users()) {
1154    if (isa<LoadInst>(U))
1155      continue;
1156
1157    if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1158      if (SI->getPointerOperand() == &I)
1159        continue;
1160    }
1161
1162    AllLoadsStores = false;
1163    break;
1164  }
1165  if (AllLoadsStores) {
1166    IRBuilder<> IRB(&I);
1167    DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy);
1168  }
1169  DFSF.setShadow(&I, DFSF.DFS.ZeroShadow);
1170}
1171
1172void DFSanVisitor::visitSelectInst(SelectInst &I) {
1173  Value *CondShadow = DFSF.getShadow(I.getCondition());
1174  Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
1175  Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
1176
1177  if (isa<VectorType>(I.getCondition()->getType())) {
1178    DFSF.setShadow(
1179        &I, DFSF.DFS.combineShadows(
1180                CondShadow,
1181                DFSF.DFS.combineShadows(TrueShadow, FalseShadow, &I), &I));
1182  } else {
1183    Value *ShadowSel;
1184    if (TrueShadow == FalseShadow) {
1185      ShadowSel = TrueShadow;
1186    } else {
1187      ShadowSel =
1188          SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
1189    }
1190    DFSF.setShadow(&I, DFSF.DFS.combineShadows(CondShadow, ShadowSel, &I));
1191  }
1192}
1193
1194void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
1195  IRBuilder<> IRB(&I);
1196  Value *ValShadow = DFSF.getShadow(I.getValue());
1197  IRB.CreateCall3(
1198      DFSF.DFS.DFSanSetLabelFn, ValShadow,
1199      IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(*DFSF.DFS.Ctx)),
1200      IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy));
1201}
1202
1203void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
1204  IRBuilder<> IRB(&I);
1205  Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
1206  Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
1207  Value *LenShadow = IRB.CreateMul(
1208      I.getLength(),
1209      ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8));
1210  Value *AlignShadow;
1211  if (ClPreserveAlignment) {
1212    AlignShadow = IRB.CreateMul(I.getAlignmentCst(),
1213                                ConstantInt::get(I.getAlignmentCst()->getType(),
1214                                                 DFSF.DFS.ShadowWidth / 8));
1215  } else {
1216    AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(),
1217                                   DFSF.DFS.ShadowWidth / 8);
1218  }
1219  Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
1220  DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr);
1221  SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
1222  IRB.CreateCall5(I.getCalledValue(), DestShadow, SrcShadow, LenShadow,
1223                  AlignShadow, I.getVolatileCst());
1224}
1225
1226void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
1227  if (!DFSF.IsNativeABI && RI.getReturnValue()) {
1228    switch (DFSF.IA) {
1229    case DataFlowSanitizer::IA_TLS: {
1230      Value *S = DFSF.getShadow(RI.getReturnValue());
1231      IRBuilder<> IRB(&RI);
1232      IRB.CreateStore(S, DFSF.getRetvalTLS());
1233      break;
1234    }
1235    case DataFlowSanitizer::IA_Args: {
1236      IRBuilder<> IRB(&RI);
1237      Type *RT = DFSF.F->getFunctionType()->getReturnType();
1238      Value *InsVal =
1239          IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
1240      Value *InsShadow =
1241          IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
1242      RI.setOperand(0, InsShadow);
1243      break;
1244    }
1245    }
1246  }
1247}
1248
1249void DFSanVisitor::visitCallSite(CallSite CS) {
1250  Function *F = CS.getCalledFunction();
1251  if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) {
1252    visitOperandShadowInst(*CS.getInstruction());
1253    return;
1254  }
1255
1256  IRBuilder<> IRB(CS.getInstruction());
1257
1258  DenseMap<Value *, Function *>::iterator i =
1259      DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue());
1260  if (i != DFSF.DFS.UnwrappedFnMap.end()) {
1261    Function *F = i->second;
1262    switch (DFSF.DFS.getWrapperKind(F)) {
1263    case DataFlowSanitizer::WK_Warning: {
1264      CS.setCalledFunction(F);
1265      IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
1266                     IRB.CreateGlobalStringPtr(F->getName()));
1267      DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1268      return;
1269    }
1270    case DataFlowSanitizer::WK_Discard: {
1271      CS.setCalledFunction(F);
1272      DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1273      return;
1274    }
1275    case DataFlowSanitizer::WK_Functional: {
1276      CS.setCalledFunction(F);
1277      visitOperandShadowInst(*CS.getInstruction());
1278      return;
1279    }
1280    case DataFlowSanitizer::WK_Custom: {
1281      // Don't try to handle invokes of custom functions, it's too complicated.
1282      // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
1283      // wrapper.
1284      if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1285        FunctionType *FT = F->getFunctionType();
1286        FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT);
1287        std::string CustomFName = "__dfsw_";
1288        CustomFName += F->getName();
1289        Constant *CustomF =
1290            DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT);
1291        if (Function *CustomFn = dyn_cast<Function>(CustomF)) {
1292          CustomFn->copyAttributesFrom(F);
1293
1294          // Custom functions returning non-void will write to the return label.
1295          if (!FT->getReturnType()->isVoidTy()) {
1296            CustomFn->removeAttributes(AttributeSet::FunctionIndex,
1297                                       DFSF.DFS.ReadOnlyNoneAttrs);
1298          }
1299        }
1300
1301        std::vector<Value *> Args;
1302
1303        CallSite::arg_iterator i = CS.arg_begin();
1304        for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) {
1305          Type *T = (*i)->getType();
1306          FunctionType *ParamFT;
1307          if (isa<PointerType>(T) &&
1308              (ParamFT = dyn_cast<FunctionType>(
1309                   cast<PointerType>(T)->getElementType()))) {
1310            std::string TName = "dfst";
1311            TName += utostr(FT->getNumParams() - n);
1312            TName += "$";
1313            TName += F->getName();
1314            Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
1315            Args.push_back(T);
1316            Args.push_back(
1317                IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
1318          } else {
1319            Args.push_back(*i);
1320          }
1321        }
1322
1323        i = CS.arg_begin();
1324        for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1325          Args.push_back(DFSF.getShadow(*i));
1326
1327        if (!FT->getReturnType()->isVoidTy()) {
1328          if (!DFSF.LabelReturnAlloca) {
1329            DFSF.LabelReturnAlloca =
1330                new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn",
1331                               DFSF.F->getEntryBlock().begin());
1332          }
1333          Args.push_back(DFSF.LabelReturnAlloca);
1334        }
1335
1336        CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
1337        CustomCI->setCallingConv(CI->getCallingConv());
1338        CustomCI->setAttributes(CI->getAttributes());
1339
1340        if (!FT->getReturnType()->isVoidTy()) {
1341          LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca);
1342          DFSF.setShadow(CustomCI, LabelLoad);
1343        }
1344
1345        CI->replaceAllUsesWith(CustomCI);
1346        CI->eraseFromParent();
1347        return;
1348      }
1349      break;
1350    }
1351    }
1352  }
1353
1354  FunctionType *FT = cast<FunctionType>(
1355      CS.getCalledValue()->getType()->getPointerElementType());
1356  if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1357    for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) {
1358      IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)),
1359                      DFSF.getArgTLS(i, CS.getInstruction()));
1360    }
1361  }
1362
1363  Instruction *Next = nullptr;
1364  if (!CS.getType()->isVoidTy()) {
1365    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1366      if (II->getNormalDest()->getSinglePredecessor()) {
1367        Next = II->getNormalDest()->begin();
1368      } else {
1369        BasicBlock *NewBB =
1370            SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DFS);
1371        Next = NewBB->begin();
1372      }
1373    } else {
1374      Next = CS->getNextNode();
1375    }
1376
1377    if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1378      IRBuilder<> NextIRB(Next);
1379      LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS());
1380      DFSF.SkipInsts.insert(LI);
1381      DFSF.setShadow(CS.getInstruction(), LI);
1382      DFSF.NonZeroChecks.insert(LI);
1383    }
1384  }
1385
1386  // Do all instrumentation for IA_Args down here to defer tampering with the
1387  // CFG in a way that SplitEdge may be able to detect.
1388  if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
1389    FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
1390    Value *Func =
1391        IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT));
1392    std::vector<Value *> Args;
1393
1394    CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1395    for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1396      Args.push_back(*i);
1397
1398    i = CS.arg_begin();
1399    for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1400      Args.push_back(DFSF.getShadow(*i));
1401
1402    if (FT->isVarArg()) {
1403      unsigned VarArgSize = CS.arg_size() - FT->getNumParams();
1404      ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize);
1405      AllocaInst *VarArgShadow =
1406          new AllocaInst(VarArgArrayTy, "", DFSF.F->getEntryBlock().begin());
1407      Args.push_back(IRB.CreateConstGEP2_32(VarArgShadow, 0, 0));
1408      for (unsigned n = 0; i != e; ++i, ++n) {
1409        IRB.CreateStore(DFSF.getShadow(*i),
1410                        IRB.CreateConstGEP2_32(VarArgShadow, 0, n));
1411        Args.push_back(*i);
1412      }
1413    }
1414
1415    CallSite NewCS;
1416    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1417      NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(),
1418                               Args);
1419    } else {
1420      NewCS = IRB.CreateCall(Func, Args);
1421    }
1422    NewCS.setCallingConv(CS.getCallingConv());
1423    NewCS.setAttributes(CS.getAttributes().removeAttributes(
1424        *DFSF.DFS.Ctx, AttributeSet::ReturnIndex,
1425        AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType(),
1426                                         AttributeSet::ReturnIndex)));
1427
1428    if (Next) {
1429      ExtractValueInst *ExVal =
1430          ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next);
1431      DFSF.SkipInsts.insert(ExVal);
1432      ExtractValueInst *ExShadow =
1433          ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next);
1434      DFSF.SkipInsts.insert(ExShadow);
1435      DFSF.setShadow(ExVal, ExShadow);
1436      DFSF.NonZeroChecks.insert(ExShadow);
1437
1438      CS.getInstruction()->replaceAllUsesWith(ExVal);
1439    }
1440
1441    CS.getInstruction()->eraseFromParent();
1442  }
1443}
1444
1445void DFSanVisitor::visitPHINode(PHINode &PN) {
1446  PHINode *ShadowPN =
1447      PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN);
1448
1449  // Give the shadow phi node valid predecessors to fool SplitEdge into working.
1450  Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy);
1451  for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e;
1452       ++i) {
1453    ShadowPN->addIncoming(UndefShadow, *i);
1454  }
1455
1456  DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN));
1457  DFSF.setShadow(&PN, ShadowPN);
1458}
1459