| 1 | //==-- handle_llvm.cpp - Helper function for Clang fuzzers -----------------==// |
| 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 | // Implements HandleLLVM for use by the Clang fuzzers. First runs a loop |
| 10 | // vectorizer optimization pass over the given IR code. Then mimics lli on both |
| 11 | // versions to JIT the generated code and execute it. Currently, functions are |
| 12 | // executed on dummy inputs. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #include "handle_llvm.h" |
| 17 | #include "input_arrays.h" |
| 18 | |
| 19 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 20 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 21 | #include "llvm/CodeGen/CommandFlags.h" |
| 22 | #include "llvm/CodeGen/MachineModuleInfo.h" |
| 23 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 24 | #include "llvm/ExecutionEngine/JITEventListener.h" |
| 25 | #include "llvm/ExecutionEngine/JITSymbol.h" |
| 26 | #include "llvm/ExecutionEngine/MCJIT.h" |
| 27 | #include "llvm/ExecutionEngine/ObjectCache.h" |
| 28 | #include "llvm/ExecutionEngine/RTDyldMemoryManager.h" |
| 29 | #include "llvm/ExecutionEngine/SectionMemoryManager.h" |
| 30 | #include "llvm/IR/IRPrintingPasses.h" |
| 31 | #include "llvm/IR/LLVMContext.h" |
| 32 | #include "llvm/IR/Module.h" |
| 33 | #include "llvm/IR/Verifier.h" |
| 34 | #include "llvm/IRPrinter/IRPrintingPasses.h" |
| 35 | #include "llvm/IRReader/IRReader.h" |
| 36 | #include "llvm/MC/TargetRegistry.h" |
| 37 | #include "llvm/Passes/OptimizationLevel.h" |
| 38 | #include "llvm/Passes/PassBuilder.h" |
| 39 | #include "llvm/Support/MemoryBuffer.h" |
| 40 | #include "llvm/Support/SourceMgr.h" |
| 41 | #include "llvm/Support/TargetSelect.h" |
| 42 | #include "llvm/Target/TargetMachine.h" |
| 43 | #include "llvm/TargetParser/Triple.h" |
| 44 | |
| 45 | using namespace llvm; |
| 46 | |
| 47 | // Define a type for the functions that are compiled and executed |
| 48 | typedef void (*LLVMFunc)(int*, int*, int*, int); |
| 49 | |
| 50 | // Helper function to parse command line args and find the optimization level |
| 51 | static CodeGenOptLevel getOptLevel(const std::vector<const char *> &) { |
| 52 | // Find the optimization level from the command line args |
| 53 | CodeGenOptLevel OLvl = CodeGenOptLevel::Default; |
| 54 | for (auto &A : ExtraArgs) { |
| 55 | if (A[0] == '-' && A[1] == 'O') { |
| 56 | if (auto Level = CodeGenOpt::parseLevel(C: A[2])) { |
| 57 | OLvl = *Level; |
| 58 | } else { |
| 59 | errs() << "error: opt level must be between 0 and 3.\n" ; |
| 60 | std::exit(status: 1); |
| 61 | } |
| 62 | } |
| 63 | } |
| 64 | return OLvl; |
| 65 | } |
| 66 | |
| 67 | static void ErrorAndExit(std::string message) { |
| 68 | errs()<< "ERROR: " << message << "\n" ; |
| 69 | std::exit(status: 1); |
| 70 | } |
| 71 | |
| 72 | // Helper function to add optimization passes to the TargetMachine at the |
| 73 | // specified optimization level, OptLevel |
| 74 | static void RunOptimizationPasses(raw_ostream &OS, Module &M, |
| 75 | CodeGenOptLevel OptLevel) { |
| 76 | llvm::OptimizationLevel OL; |
| 77 | switch (OptLevel) { |
| 78 | case CodeGenOptLevel::None: |
| 79 | OL = OptimizationLevel::O0; |
| 80 | break; |
| 81 | case CodeGenOptLevel::Less: |
| 82 | OL = OptimizationLevel::O1; |
| 83 | break; |
| 84 | case CodeGenOptLevel::Default: |
| 85 | OL = OptimizationLevel::O2; |
| 86 | break; |
| 87 | case CodeGenOptLevel::Aggressive: |
| 88 | OL = OptimizationLevel::O3; |
| 89 | break; |
| 90 | } |
| 91 | |
| 92 | LoopAnalysisManager LAM; |
| 93 | FunctionAnalysisManager FAM; |
| 94 | CGSCCAnalysisManager CGAM; |
| 95 | ModuleAnalysisManager MAM; |
| 96 | |
| 97 | PassBuilder PB; |
| 98 | |
| 99 | PB.registerModuleAnalyses(MAM); |
| 100 | PB.registerCGSCCAnalyses(CGAM); |
| 101 | PB.registerFunctionAnalyses(FAM); |
| 102 | PB.registerLoopAnalyses(LAM); |
| 103 | PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); |
| 104 | |
| 105 | ModulePassManager MPM = PB.buildPerModuleDefaultPipeline(Level: OL); |
| 106 | MPM.addPass(Pass: PrintModulePass(OS)); |
| 107 | |
| 108 | MPM.run(IR&: M, AM&: MAM); |
| 109 | } |
| 110 | |
| 111 | // Mimics the opt tool to run an optimization pass over the provided IR |
| 112 | static std::string OptLLVM(const std::string &IR, CodeGenOptLevel OLvl) { |
| 113 | // Create a module that will run the optimization passes |
| 114 | SMDiagnostic Err; |
| 115 | LLVMContext Context; |
| 116 | std::unique_ptr<Module> M = parseIR(Buffer: MemoryBufferRef(IR, "IR" ), Err, Context); |
| 117 | if (!M || verifyModule(M: *M, OS: &errs())) |
| 118 | ErrorAndExit(message: "Could not parse IR" ); |
| 119 | |
| 120 | Triple ModuleTriple(M->getTargetTriple()); |
| 121 | const TargetOptions Options = |
| 122 | codegen::InitTargetOptionsFromCodeGenFlags(TheTriple: ModuleTriple); |
| 123 | std::string E; |
| 124 | const Target *TheTarget = |
| 125 | TargetRegistry::lookupTarget(ArchName: codegen::getMArch(), TheTriple&: ModuleTriple, Error&: E); |
| 126 | if (!TheTarget) |
| 127 | ErrorAndExit(message: E); |
| 128 | |
| 129 | std::unique_ptr<TargetMachine> TM(TheTarget->createTargetMachine( |
| 130 | TT: M->getTargetTriple(), CPU: codegen::getCPUStr(), Features: codegen::getFeaturesStr(), |
| 131 | Options, RM: codegen::getExplicitRelocModel(), |
| 132 | CM: codegen::getExplicitCodeModel(), OL: OLvl)); |
| 133 | if (!TM) |
| 134 | ErrorAndExit(message: "Could not create target machine" ); |
| 135 | |
| 136 | codegen::setFunctionAttributes(CPU: codegen::getCPUStr(), |
| 137 | Features: codegen::getFeaturesStr(), M&: *M); |
| 138 | |
| 139 | // Add a pass that writes the optimized IR to an output stream |
| 140 | std::string outString; |
| 141 | raw_string_ostream OS(outString); |
| 142 | RunOptimizationPasses(OS, M&: *M, OptLevel: OLvl); |
| 143 | |
| 144 | return outString; |
| 145 | } |
| 146 | |
| 147 | // Takes a function and runs it on a set of inputs |
| 148 | // First determines whether f is the optimized or unoptimized function |
| 149 | static void RunFuncOnInputs(LLVMFunc f, int Arr[kNumArrays][kArraySize]) { |
| 150 | for (int i = 0; i < kNumArrays / 3; i++) |
| 151 | f(Arr[i], Arr[i + (kNumArrays / 3)], Arr[i + (2 * kNumArrays / 3)], |
| 152 | kArraySize); |
| 153 | } |
| 154 | |
| 155 | // Takes a string of IR and compiles it using LLVM's JIT Engine |
| 156 | static void CreateAndRunJITFunc(const std::string &IR, CodeGenOptLevel OLvl) { |
| 157 | SMDiagnostic Err; |
| 158 | LLVMContext Context; |
| 159 | std::unique_ptr<Module> M = parseIR(Buffer: MemoryBufferRef(IR, "IR" ), Err, Context); |
| 160 | if (!M) |
| 161 | ErrorAndExit(message: "Could not parse IR" ); |
| 162 | |
| 163 | Function *EntryFunc = M->getFunction(Name: "foo" ); |
| 164 | if (!EntryFunc) |
| 165 | ErrorAndExit(message: "Function not found in module" ); |
| 166 | |
| 167 | std::string ErrorMsg; |
| 168 | Triple ModuleTriple(M->getTargetTriple()); |
| 169 | |
| 170 | EngineBuilder builder(std::move(M)); |
| 171 | builder.setMArch(codegen::getMArch()); |
| 172 | builder.setMCPU(codegen::getCPUStr()); |
| 173 | builder.setMAttrs(codegen::getFeatureList()); |
| 174 | builder.setErrorStr(&ErrorMsg); |
| 175 | builder.setEngineKind(EngineKind::JIT); |
| 176 | builder.setMCJITMemoryManager(std::make_unique<SectionMemoryManager>()); |
| 177 | builder.setOptLevel(OLvl); |
| 178 | builder.setTargetOptions( |
| 179 | codegen::InitTargetOptionsFromCodeGenFlags(TheTriple: ModuleTriple)); |
| 180 | |
| 181 | std::unique_ptr<ExecutionEngine> EE(builder.create()); |
| 182 | if (!EE) |
| 183 | ErrorAndExit(message: "Could not create execution engine" ); |
| 184 | |
| 185 | EE->finalizeObject(); |
| 186 | EE->runStaticConstructorsDestructors(isDtors: false); |
| 187 | |
| 188 | LLVMFunc f = reinterpret_cast<LLVMFunc>(EE->getPointerToFunction(F: EntryFunc)); |
| 189 | |
| 190 | // Figure out if we are running the optimized func or the unoptimized func |
| 191 | RunFuncOnInputs(f, Arr: (OLvl == CodeGenOptLevel::None) ? UnoptArrays : OptArrays); |
| 192 | |
| 193 | EE->runStaticConstructorsDestructors(isDtors: true); |
| 194 | } |
| 195 | |
| 196 | // Main fuzz target called by ExampleClangLLVMProtoFuzzer.cpp |
| 197 | // Mimics the lli tool to JIT the LLVM IR code and execute it |
| 198 | void clang_fuzzer::HandleLLVM(const std::string &IR, |
| 199 | const std::vector<const char *> &) { |
| 200 | // Populate OptArrays and UnoptArrays with the arrays from InputArrays |
| 201 | memcpy(dest: OptArrays, src: InputArrays, n: kTotalSize); |
| 202 | memcpy(dest: UnoptArrays, src: InputArrays, n: kTotalSize); |
| 203 | |
| 204 | // Parse ExtraArgs to set the optimization level |
| 205 | CodeGenOptLevel OLvl = getOptLevel(ExtraArgs); |
| 206 | |
| 207 | // First we optimize the IR by running a loop vectorizer pass |
| 208 | std::string OptIR = OptLLVM(IR, OLvl); |
| 209 | |
| 210 | CreateAndRunJITFunc(IR: OptIR, OLvl); |
| 211 | CreateAndRunJITFunc(IR, OLvl: CodeGenOptLevel::None); |
| 212 | |
| 213 | if (memcmp(s1: OptArrays, s2: UnoptArrays, n: kTotalSize)) |
| 214 | ErrorAndExit(message: "!!!BUG!!!" ); |
| 215 | } |
| 216 | |