1//===-- SanitizerCoverage.cpp - coverage instrumentation for sanitizers ---===//
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//
10// Coverage instrumentation that works with AddressSanitizer
11// and potentially with other Sanitizers.
12//
13// We create a Guard variable with the same linkage
14// as the function and inject this code into the entry block (CoverageLevel=1)
15// or all blocks (CoverageLevel>=2):
16// if (Guard < 0) {
17//    __sanitizer_cov(&Guard);
18// }
19// The accesses to Guard are atomic. The rest of the logic is
20// in __sanitizer_cov (it's fine to call it more than once).
21//
22// With CoverageLevel>=3 we also split critical edges this effectively
23// instrumenting all edges.
24//
25// CoverageLevel>=4 add indirect call profiling implented as a function call.
26//
27// This coverage implementation provides very limited data:
28// it only tells if a given function (block) was ever executed. No counters.
29// But for many use cases this is what we need and the added slowdown small.
30//
31//===----------------------------------------------------------------------===//
32
33#include "llvm/Transforms/Instrumentation.h"
34#include "llvm/ADT/ArrayRef.h"
35#include "llvm/ADT/SmallVector.h"
36#include "llvm/IR/CallSite.h"
37#include "llvm/IR/DataLayout.h"
38#include "llvm/IR/Function.h"
39#include "llvm/IR/IRBuilder.h"
40#include "llvm/IR/InlineAsm.h"
41#include "llvm/IR/LLVMContext.h"
42#include "llvm/IR/MDBuilder.h"
43#include "llvm/IR/Module.h"
44#include "llvm/IR/Type.h"
45#include "llvm/Support/CommandLine.h"
46#include "llvm/Support/Debug.h"
47#include "llvm/Support/raw_ostream.h"
48#include "llvm/Transforms/Scalar.h"
49#include "llvm/Transforms/Utils/BasicBlockUtils.h"
50#include "llvm/Transforms/Utils/ModuleUtils.h"
51
52using namespace llvm;
53
54#define DEBUG_TYPE "sancov"
55
56static const char *const kSanCovModuleInitName = "__sanitizer_cov_module_init";
57static const char *const kSanCovName = "__sanitizer_cov";
58static const char *const kSanCovWithCheckName = "__sanitizer_cov_with_check";
59static const char *const kSanCovIndirCallName = "__sanitizer_cov_indir_call16";
60static const char *const kSanCovTraceEnter = "__sanitizer_cov_trace_func_enter";
61static const char *const kSanCovTraceBB = "__sanitizer_cov_trace_basic_block";
62static const char *const kSanCovTraceCmp = "__sanitizer_cov_trace_cmp";
63static const char *const kSanCovModuleCtorName = "sancov.module_ctor";
64static const uint64_t    kSanCtorAndDtorPriority = 2;
65
66static cl::opt<int> ClCoverageLevel("sanitizer-coverage-level",
67       cl::desc("Sanitizer Coverage. 0: none, 1: entry block, 2: all blocks, "
68                "3: all blocks and critical edges, "
69                "4: above plus indirect calls"),
70       cl::Hidden, cl::init(0));
71
72static cl::opt<unsigned> ClCoverageBlockThreshold(
73    "sanitizer-coverage-block-threshold",
74    cl::desc("Use a callback with a guard check inside it if there are"
75             " more than this number of blocks."),
76    cl::Hidden, cl::init(500));
77
78static cl::opt<bool>
79    ClExperimentalTracing("sanitizer-coverage-experimental-tracing",
80                          cl::desc("Experimental basic-block tracing: insert "
81                                   "callbacks at every basic block"),
82                          cl::Hidden, cl::init(false));
83
84static cl::opt<bool>
85    ClExperimentalCMPTracing("sanitizer-coverage-experimental-trace-compares",
86                             cl::desc("Experimental tracing of CMP and similar "
87                                      "instructions"),
88                             cl::Hidden, cl::init(false));
89
90// Experimental 8-bit counters used as an additional search heuristic during
91// coverage-guided fuzzing.
92// The counters are not thread-friendly:
93//   - contention on these counters may cause significant slowdown;
94//   - the counter updates are racy and the results may be inaccurate.
95// They are also inaccurate due to 8-bit integer overflow.
96static cl::opt<bool> ClUse8bitCounters("sanitizer-coverage-8bit-counters",
97                                       cl::desc("Experimental 8-bit counters"),
98                                       cl::Hidden, cl::init(false));
99
100namespace {
101
102class SanitizerCoverageModule : public ModulePass {
103 public:
104   SanitizerCoverageModule(int CoverageLevel = 0)
105       : ModulePass(ID),
106         CoverageLevel(std::max(CoverageLevel, (int)ClCoverageLevel)) {}
107  bool runOnModule(Module &M) override;
108  bool runOnFunction(Function &F);
109  static char ID;  // Pass identification, replacement for typeid
110  const char *getPassName() const override {
111    return "SanitizerCoverageModule";
112  }
113
114 private:
115  void InjectCoverageForIndirectCalls(Function &F,
116                                      ArrayRef<Instruction *> IndirCalls);
117  void InjectTraceForCmp(Function &F, ArrayRef<Instruction *> CmpTraceTargets);
118  bool InjectCoverage(Function &F, ArrayRef<BasicBlock *> AllBlocks);
119  void SetNoSanitizeMetada(Instruction *I);
120  void InjectCoverageAtBlock(Function &F, BasicBlock &BB, bool UseCalls);
121  unsigned NumberOfInstrumentedBlocks() {
122    return SanCovFunction->getNumUses() + SanCovWithCheckFunction->getNumUses();
123  }
124  Function *SanCovFunction;
125  Function *SanCovWithCheckFunction;
126  Function *SanCovIndirCallFunction;
127  Function *SanCovModuleInit;
128  Function *SanCovTraceEnter, *SanCovTraceBB;
129  Function *SanCovTraceCmpFunction;
130  InlineAsm *EmptyAsm;
131  Type *IntptrTy, *Int64Ty;
132  LLVMContext *C;
133  const DataLayout *DL;
134
135  GlobalVariable *GuardArray;
136  GlobalVariable *EightBitCounterArray;
137
138  int CoverageLevel;
139};
140
141}  // namespace
142
143bool SanitizerCoverageModule::runOnModule(Module &M) {
144  if (!CoverageLevel) return false;
145  C = &(M.getContext());
146  DL = &M.getDataLayout();
147  IntptrTy = Type::getIntNTy(*C, DL->getPointerSizeInBits());
148  Type *VoidTy = Type::getVoidTy(*C);
149  IRBuilder<> IRB(*C);
150  Type *Int8PtrTy = PointerType::getUnqual(IRB.getInt8Ty());
151  Type *Int32PtrTy = PointerType::getUnqual(IRB.getInt32Ty());
152  Int64Ty = IRB.getInt64Ty();
153
154  Function *CtorFunc =
155      Function::Create(FunctionType::get(VoidTy, false),
156                       GlobalValue::InternalLinkage, kSanCovModuleCtorName, &M);
157  ReturnInst::Create(*C, BasicBlock::Create(*C, "", CtorFunc));
158  appendToGlobalCtors(M, CtorFunc, kSanCtorAndDtorPriority);
159
160  SanCovFunction = checkSanitizerInterfaceFunction(
161      M.getOrInsertFunction(kSanCovName, VoidTy, Int32PtrTy, nullptr));
162  SanCovWithCheckFunction = checkSanitizerInterfaceFunction(
163      M.getOrInsertFunction(kSanCovWithCheckName, VoidTy, Int32PtrTy, nullptr));
164  SanCovIndirCallFunction =
165      checkSanitizerInterfaceFunction(M.getOrInsertFunction(
166          kSanCovIndirCallName, VoidTy, IntptrTy, IntptrTy, nullptr));
167  SanCovTraceCmpFunction =
168      checkSanitizerInterfaceFunction(M.getOrInsertFunction(
169          kSanCovTraceCmp, VoidTy, Int64Ty, Int64Ty, Int64Ty, nullptr));
170
171  SanCovModuleInit = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
172      kSanCovModuleInitName, Type::getVoidTy(*C), Int32PtrTy, IntptrTy,
173      Int8PtrTy, Int8PtrTy, nullptr));
174  SanCovModuleInit->setLinkage(Function::ExternalLinkage);
175  // We insert an empty inline asm after cov callbacks to avoid callback merge.
176  EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
177                            StringRef(""), StringRef(""),
178                            /*hasSideEffects=*/true);
179
180  if (ClExperimentalTracing) {
181    SanCovTraceEnter = checkSanitizerInterfaceFunction(
182        M.getOrInsertFunction(kSanCovTraceEnter, VoidTy, Int32PtrTy, nullptr));
183    SanCovTraceBB = checkSanitizerInterfaceFunction(
184        M.getOrInsertFunction(kSanCovTraceBB, VoidTy, Int32PtrTy, nullptr));
185  }
186
187  // At this point we create a dummy array of guards because we don't
188  // know how many elements we will need.
189  Type *Int32Ty = IRB.getInt32Ty();
190  Type *Int8Ty = IRB.getInt8Ty();
191
192  GuardArray =
193      new GlobalVariable(M, Int32Ty, false, GlobalValue::ExternalLinkage,
194                         nullptr, "__sancov_gen_cov_tmp");
195  if (ClUse8bitCounters)
196    EightBitCounterArray =
197        new GlobalVariable(M, Int8Ty, false, GlobalVariable::ExternalLinkage,
198                           nullptr, "__sancov_gen_cov_tmp");
199
200  for (auto &F : M)
201    runOnFunction(F);
202
203  auto N = NumberOfInstrumentedBlocks();
204
205  // Now we know how many elements we need. Create an array of guards
206  // with one extra element at the beginning for the size.
207  Type *Int32ArrayNTy = ArrayType::get(Int32Ty, N + 1);
208  GlobalVariable *RealGuardArray = new GlobalVariable(
209      M, Int32ArrayNTy, false, GlobalValue::PrivateLinkage,
210      Constant::getNullValue(Int32ArrayNTy), "__sancov_gen_cov");
211
212
213  // Replace the dummy array with the real one.
214  GuardArray->replaceAllUsesWith(
215      IRB.CreatePointerCast(RealGuardArray, Int32PtrTy));
216  GuardArray->eraseFromParent();
217
218  GlobalVariable *RealEightBitCounterArray;
219  if (ClUse8bitCounters) {
220    // Make sure the array is 16-aligned.
221    static const int kCounterAlignment = 16;
222    Type *Int8ArrayNTy =
223        ArrayType::get(Int8Ty, RoundUpToAlignment(N, kCounterAlignment));
224    RealEightBitCounterArray = new GlobalVariable(
225        M, Int8ArrayNTy, false, GlobalValue::PrivateLinkage,
226        Constant::getNullValue(Int8ArrayNTy), "__sancov_gen_cov_counter");
227    RealEightBitCounterArray->setAlignment(kCounterAlignment);
228    EightBitCounterArray->replaceAllUsesWith(
229        IRB.CreatePointerCast(RealEightBitCounterArray, Int8PtrTy));
230    EightBitCounterArray->eraseFromParent();
231  }
232
233  // Create variable for module (compilation unit) name
234  Constant *ModNameStrConst =
235      ConstantDataArray::getString(M.getContext(), M.getName(), true);
236  GlobalVariable *ModuleName =
237      new GlobalVariable(M, ModNameStrConst->getType(), true,
238                         GlobalValue::PrivateLinkage, ModNameStrConst);
239
240  // Call __sanitizer_cov_module_init
241  IRB.SetInsertPoint(CtorFunc->getEntryBlock().getTerminator());
242  IRB.CreateCall4(
243      SanCovModuleInit, IRB.CreatePointerCast(RealGuardArray, Int32PtrTy),
244      ConstantInt::get(IntptrTy, N),
245      ClUse8bitCounters
246          ? IRB.CreatePointerCast(RealEightBitCounterArray, Int8PtrTy)
247          : Constant::getNullValue(Int8PtrTy),
248      IRB.CreatePointerCast(ModuleName, Int8PtrTy));
249  return true;
250}
251
252bool SanitizerCoverageModule::runOnFunction(Function &F) {
253  if (F.empty()) return false;
254  if (F.getName().find(".module_ctor") != std::string::npos)
255    return false;  // Should not instrument sanitizer init functions.
256  if (CoverageLevel >= 3)
257    SplitAllCriticalEdges(F);
258  SmallVector<Instruction*, 8> IndirCalls;
259  SmallVector<BasicBlock*, 16> AllBlocks;
260  SmallVector<Instruction*, 8> CmpTraceTargets;
261  for (auto &BB : F) {
262    AllBlocks.push_back(&BB);
263    for (auto &Inst : BB) {
264      if (CoverageLevel >= 4) {
265        CallSite CS(&Inst);
266        if (CS && !CS.getCalledFunction())
267          IndirCalls.push_back(&Inst);
268      }
269      if (ClExperimentalCMPTracing)
270        if (isa<ICmpInst>(&Inst))
271          CmpTraceTargets.push_back(&Inst);
272    }
273  }
274  InjectCoverage(F, AllBlocks);
275  InjectCoverageForIndirectCalls(F, IndirCalls);
276  InjectTraceForCmp(F, CmpTraceTargets);
277  return true;
278}
279
280bool SanitizerCoverageModule::InjectCoverage(Function &F,
281                                             ArrayRef<BasicBlock *> AllBlocks) {
282  if (!CoverageLevel) return false;
283
284  if (CoverageLevel == 1) {
285    InjectCoverageAtBlock(F, F.getEntryBlock(), false);
286  } else {
287    for (auto BB : AllBlocks)
288      InjectCoverageAtBlock(F, *BB,
289                            ClCoverageBlockThreshold < AllBlocks.size());
290  }
291  return true;
292}
293
294// On every indirect call we call a run-time function
295// __sanitizer_cov_indir_call* with two parameters:
296//   - callee address,
297//   - global cache array that contains kCacheSize pointers (zero-initialized).
298//     The cache is used to speed up recording the caller-callee pairs.
299// The address of the caller is passed implicitly via caller PC.
300// kCacheSize is encoded in the name of the run-time function.
301void SanitizerCoverageModule::InjectCoverageForIndirectCalls(
302    Function &F, ArrayRef<Instruction *> IndirCalls) {
303  if (IndirCalls.empty()) return;
304  const int kCacheSize = 16;
305  const int kCacheAlignment = 64;  // Align for better performance.
306  Type *Ty = ArrayType::get(IntptrTy, kCacheSize);
307  for (auto I : IndirCalls) {
308    IRBuilder<> IRB(I);
309    CallSite CS(I);
310    Value *Callee = CS.getCalledValue();
311    if (isa<InlineAsm>(Callee)) continue;
312    GlobalVariable *CalleeCache = new GlobalVariable(
313        *F.getParent(), Ty, false, GlobalValue::PrivateLinkage,
314        Constant::getNullValue(Ty), "__sancov_gen_callee_cache");
315    CalleeCache->setAlignment(kCacheAlignment);
316    IRB.CreateCall2(SanCovIndirCallFunction,
317                    IRB.CreatePointerCast(Callee, IntptrTy),
318                    IRB.CreatePointerCast(CalleeCache, IntptrTy));
319  }
320}
321
322void SanitizerCoverageModule::InjectTraceForCmp(
323    Function &F, ArrayRef<Instruction *> CmpTraceTargets) {
324  if (!ClExperimentalCMPTracing) return;
325  for (auto I : CmpTraceTargets) {
326    if (ICmpInst *ICMP = dyn_cast<ICmpInst>(I)) {
327      IRBuilder<> IRB(ICMP);
328      Value *A0 = ICMP->getOperand(0);
329      Value *A1 = ICMP->getOperand(1);
330      if (!A0->getType()->isIntegerTy()) continue;
331      uint64_t TypeSize = DL->getTypeStoreSizeInBits(A0->getType());
332      // __sanitizer_cov_indir_call((type_size << 32) | predicate, A0, A1);
333      IRB.CreateCall3(
334          SanCovTraceCmpFunction,
335          ConstantInt::get(Int64Ty, (TypeSize << 32) | ICMP->getPredicate()),
336          IRB.CreateIntCast(A0, Int64Ty, true),
337          IRB.CreateIntCast(A1, Int64Ty, true));
338    }
339  }
340}
341
342void SanitizerCoverageModule::SetNoSanitizeMetada(Instruction *I) {
343  I->setMetadata(
344      I->getParent()->getParent()->getParent()->getMDKindID("nosanitize"),
345      MDNode::get(*C, None));
346}
347
348void SanitizerCoverageModule::InjectCoverageAtBlock(Function &F, BasicBlock &BB,
349                                                    bool UseCalls) {
350  BasicBlock::iterator IP = BB.getFirstInsertionPt(), BE = BB.end();
351  // Skip static allocas at the top of the entry block so they don't become
352  // dynamic when we split the block.  If we used our optimized stack layout,
353  // then there will only be one alloca and it will come first.
354  for (; IP != BE; ++IP) {
355    AllocaInst *AI = dyn_cast<AllocaInst>(IP);
356    if (!AI || !AI->isStaticAlloca())
357      break;
358  }
359
360  bool IsEntryBB = &BB == &F.getEntryBlock();
361  DebugLoc EntryLoc = IsEntryBB && IP->getDebugLoc()
362                          ? IP->getDebugLoc().getFnDebugLoc()
363                          : IP->getDebugLoc();
364  IRBuilder<> IRB(IP);
365  IRB.SetCurrentDebugLocation(EntryLoc);
366  SmallVector<Value *, 1> Indices;
367  Value *GuardP = IRB.CreateAdd(
368      IRB.CreatePointerCast(GuardArray, IntptrTy),
369      ConstantInt::get(IntptrTy, (1 + NumberOfInstrumentedBlocks()) * 4));
370  Type *Int32PtrTy = PointerType::getUnqual(IRB.getInt32Ty());
371  GuardP = IRB.CreateIntToPtr(GuardP, Int32PtrTy);
372  if (UseCalls) {
373    IRB.CreateCall(SanCovWithCheckFunction, GuardP);
374  } else {
375    LoadInst *Load = IRB.CreateLoad(GuardP);
376    Load->setAtomic(Monotonic);
377    Load->setAlignment(4);
378    SetNoSanitizeMetada(Load);
379    Value *Cmp = IRB.CreateICmpSGE(Constant::getNullValue(Load->getType()), Load);
380    Instruction *Ins = SplitBlockAndInsertIfThen(
381        Cmp, IP, false, MDBuilder(*C).createBranchWeights(1, 100000));
382    IRB.SetInsertPoint(Ins);
383    IRB.SetCurrentDebugLocation(EntryLoc);
384    // __sanitizer_cov gets the PC of the instruction using GET_CALLER_PC.
385    IRB.CreateCall(SanCovFunction, GuardP);
386    IRB.CreateCall(EmptyAsm);  // Avoids callback merge.
387  }
388
389  if(ClUse8bitCounters) {
390    IRB.SetInsertPoint(IP);
391    Value *P = IRB.CreateAdd(
392        IRB.CreatePointerCast(EightBitCounterArray, IntptrTy),
393        ConstantInt::get(IntptrTy, NumberOfInstrumentedBlocks() - 1));
394    P = IRB.CreateIntToPtr(P, IRB.getInt8PtrTy());
395    LoadInst *LI = IRB.CreateLoad(P);
396    Value *Inc = IRB.CreateAdd(LI, ConstantInt::get(IRB.getInt8Ty(), 1));
397    StoreInst *SI = IRB.CreateStore(Inc, P);
398    SetNoSanitizeMetada(LI);
399    SetNoSanitizeMetada(SI);
400  }
401
402  if (ClExperimentalTracing) {
403    // Experimental support for tracing.
404    // Insert a callback with the same guard variable as used for coverage.
405    IRB.SetInsertPoint(IP);
406    IRB.CreateCall(IsEntryBB ? SanCovTraceEnter : SanCovTraceBB, GuardP);
407  }
408}
409
410char SanitizerCoverageModule::ID = 0;
411INITIALIZE_PASS(SanitizerCoverageModule, "sancov",
412    "SanitizerCoverage: TODO."
413    "ModulePass", false, false)
414ModulePass *llvm::createSanitizerCoverageModulePass(int CoverageLevel) {
415  return new SanitizerCoverageModule(CoverageLevel);
416}
417