1 //===--- llvm-opt-fuzzer.cpp - Fuzzer for instruction selection ----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Tool to fuzz optimization passes using libFuzzer.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Analysis/AliasAnalysis.h"
14 #include "llvm/Bitcode/BitcodeReader.h"
15 #include "llvm/Bitcode/BitcodeWriter.h"
16 #include "llvm/CodeGen/CommandFlags.h"
17 #include "llvm/FuzzMutate/FuzzerCLI.h"
18 #include "llvm/FuzzMutate/IRMutator.h"
19 #include "llvm/IR/Verifier.h"
20 #include "llvm/InitializePasses.h"
21 #include "llvm/Passes/PassBuilder.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/SourceMgr.h"
24 #include "llvm/Support/TargetRegistry.h"
25 #include "llvm/Support/TargetSelect.h"
26 #include "llvm/Target/TargetMachine.h"
27 
28 using namespace llvm;
29 
30 static codegen::RegisterCodeGenFlags CGF;
31 
32 static cl::opt<std::string>
33     TargetTripleStr("mtriple", cl::desc("Override target triple for module"));
34 
35 // Passes to run for this fuzzer instance. Expects new pass manager syntax.
36 static cl::opt<std::string> PassPipeline(
37     "passes",
38     cl::desc("A textual description of the pass pipeline for testing"));
39 
40 static std::unique_ptr<IRMutator> Mutator;
41 static std::unique_ptr<TargetMachine> TM;
42 
createOptMutator()43 std::unique_ptr<IRMutator> createOptMutator() {
44   std::vector<TypeGetter> Types{
45       Type::getInt1Ty,  Type::getInt8Ty,  Type::getInt16Ty, Type::getInt32Ty,
46       Type::getInt64Ty, Type::getFloatTy, Type::getDoubleTy};
47 
48   std::vector<std::unique_ptr<IRMutationStrategy>> Strategies;
49   Strategies.push_back(
50       std::make_unique<InjectorIRStrategy>(
51           InjectorIRStrategy::getDefaultOps()));
52   Strategies.push_back(
53       std::make_unique<InstDeleterIRStrategy>());
54   Strategies.push_back(std::make_unique<InstModificationIRStrategy>());
55 
56   return std::make_unique<IRMutator>(std::move(Types), std::move(Strategies));
57 }
58 
LLVMFuzzerCustomMutator(uint8_t * Data,size_t Size,size_t MaxSize,unsigned int Seed)59 extern "C" LLVM_ATTRIBUTE_USED size_t LLVMFuzzerCustomMutator(
60     uint8_t *Data, size_t Size, size_t MaxSize, unsigned int Seed) {
61 
62   assert(Mutator &&
63       "IR mutator should have been created during fuzzer initialization");
64 
65   LLVMContext Context;
66   auto M = parseAndVerify(Data, Size, Context);
67   if (!M) {
68     errs() << "error: mutator input module is broken!\n";
69     return 0;
70   }
71 
72   Mutator->mutateModule(*M, Seed, Size, MaxSize);
73 
74   if (verifyModule(*M, &errs())) {
75     errs() << "mutation result doesn't pass verification\n";
76 #ifndef NDEBUG
77     M->dump();
78 #endif
79     // Avoid adding incorrect test cases to the corpus.
80     return 0;
81   }
82 
83   std::string Buf;
84   {
85     raw_string_ostream OS(Buf);
86     WriteBitcodeToFile(*M, OS);
87   }
88   if (Buf.size() > MaxSize)
89     return 0;
90 
91   // There are some invariants which are not checked by the verifier in favor
92   // of having them checked by the parser. They may be considered as bugs in the
93   // verifier and should be fixed there. However until all of those are covered
94   // we want to check for them explicitly. Otherwise we will add incorrect input
95   // to the corpus and this is going to confuse the fuzzer which will start
96   // exploration of the bitcode reader error handling code.
97   auto NewM = parseAndVerify(
98       reinterpret_cast<const uint8_t*>(Buf.data()), Buf.size(), Context);
99   if (!NewM) {
100     errs() << "mutator failed to re-read the module\n";
101 #ifndef NDEBUG
102     M->dump();
103 #endif
104     return 0;
105   }
106 
107   memcpy(Data, Buf.data(), Buf.size());
108   return Buf.size();
109 }
110 
LLVMFuzzerTestOneInput(const uint8_t * Data,size_t Size)111 extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
112   assert(TM && "Should have been created during fuzzer initialization");
113 
114   if (Size <= 1)
115     // We get bogus data given an empty corpus - ignore it.
116     return 0;
117 
118   // Parse module
119   //
120 
121   LLVMContext Context;
122   auto M = parseAndVerify(Data, Size, Context);
123   if (!M) {
124     errs() << "error: input module is broken!\n";
125     return 0;
126   }
127 
128   // Set up target dependant options
129   //
130 
131   M->setTargetTriple(TM->getTargetTriple().normalize());
132   M->setDataLayout(TM->createDataLayout());
133   codegen::setFunctionAttributes(TM->getTargetCPU(),
134                                  TM->getTargetFeatureString(), *M);
135 
136   // Create pass pipeline
137   //
138 
139   PassBuilder PB(false, TM.get());
140 
141   LoopAnalysisManager LAM;
142   FunctionAnalysisManager FAM;
143   CGSCCAnalysisManager CGAM;
144   ModulePassManager MPM;
145   ModuleAnalysisManager MAM;
146 
147   FAM.registerPass([&] { return PB.buildDefaultAAPipeline(); });
148   PB.registerModuleAnalyses(MAM);
149   PB.registerCGSCCAnalyses(CGAM);
150   PB.registerFunctionAnalyses(FAM);
151   PB.registerLoopAnalyses(LAM);
152   PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
153 
154   auto Err = PB.parsePassPipeline(MPM, PassPipeline);
155   assert(!Err && "Should have been checked during fuzzer initialization");
156   // Only fail with assert above, otherwise ignore the parsing error.
157   consumeError(std::move(Err));
158 
159   // Run passes which we need to test
160   //
161 
162   MPM.run(*M, MAM);
163 
164   // Check that passes resulted in a correct code
165   if (verifyModule(*M, &errs())) {
166     errs() << "Transformation resulted in an invalid module\n";
167     abort();
168   }
169 
170   return 0;
171 }
172 
handleLLVMFatalError(void *,const std::string & Message,bool)173 static void handleLLVMFatalError(void *, const std::string &Message, bool) {
174   // TODO: Would it be better to call into the fuzzer internals directly?
175   dbgs() << "LLVM ERROR: " << Message << "\n"
176          << "Aborting to trigger fuzzer exit handling.\n";
177   abort();
178 }
179 
LLVMFuzzerInitialize(int * argc,char *** argv)180 extern "C" LLVM_ATTRIBUTE_USED int LLVMFuzzerInitialize(
181     int *argc, char ***argv) {
182   EnableDebugBuffering = true;
183 
184   // Make sure we print the summary and the current unit when LLVM errors out.
185   install_fatal_error_handler(handleLLVMFatalError, nullptr);
186 
187   // Initialize llvm
188   //
189 
190   InitializeAllTargets();
191   InitializeAllTargetMCs();
192 
193   PassRegistry &Registry = *PassRegistry::getPassRegistry();
194   initializeCore(Registry);
195   initializeCoroutines(Registry);
196   initializeScalarOpts(Registry);
197   initializeObjCARCOpts(Registry);
198   initializeVectorization(Registry);
199   initializeIPO(Registry);
200   initializeAnalysis(Registry);
201   initializeTransformUtils(Registry);
202   initializeInstCombine(Registry);
203   initializeAggressiveInstCombine(Registry);
204   initializeInstrumentation(Registry);
205   initializeTarget(Registry);
206 
207   // Parse input options
208   //
209 
210   handleExecNameEncodedOptimizerOpts(*argv[0]);
211   parseFuzzerCLOpts(*argc, *argv);
212 
213   // Create TargetMachine
214   //
215 
216   if (TargetTripleStr.empty()) {
217     errs() << *argv[0] << ": -mtriple must be specified\n";
218     exit(1);
219   }
220   Triple TargetTriple = Triple(Triple::normalize(TargetTripleStr));
221 
222   std::string Error;
223   const Target *TheTarget =
224       TargetRegistry::lookupTarget(codegen::getMArch(), TargetTriple, Error);
225   if (!TheTarget) {
226     errs() << *argv[0] << ": " << Error;
227     exit(1);
228   }
229 
230   TargetOptions Options =
231       codegen::InitTargetOptionsFromCodeGenFlags(TargetTriple);
232   TM.reset(TheTarget->createTargetMachine(
233       TargetTriple.getTriple(), codegen::getCPUStr(), codegen::getFeaturesStr(),
234       Options, codegen::getExplicitRelocModel(),
235       codegen::getExplicitCodeModel(), CodeGenOpt::Default));
236   assert(TM && "Could not allocate target machine!");
237 
238   // Check that pass pipeline is specified and correct
239   //
240 
241   if (PassPipeline.empty()) {
242     errs() << *argv[0] << ": at least one pass should be specified\n";
243     exit(1);
244   }
245 
246   PassBuilder PB(false, TM.get());
247   ModulePassManager MPM;
248   if (auto Err = PB.parsePassPipeline(MPM, PassPipeline)) {
249     errs() << *argv[0] << ": " << toString(std::move(Err)) << "\n";
250     exit(1);
251   }
252 
253   // Create mutator
254   //
255 
256   Mutator = createOptMutator();
257 
258   return 0;
259 }
260