| 1 | //===- IR2Vec.cpp - Implementation of IR2Vec -----------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM |
| 4 | // Exceptions. See the LICENSE file for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | /// |
| 9 | /// \file |
| 10 | /// This file implements the IR2Vec algorithm. |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/Analysis/IR2Vec.h" |
| 15 | |
| 16 | #include "llvm/ADT/DepthFirstIterator.h" |
| 17 | #include "llvm/ADT/Statistic.h" |
| 18 | #include "llvm/IR/CFG.h" |
| 19 | #include "llvm/IR/Module.h" |
| 20 | #include "llvm/IR/PassManager.h" |
| 21 | #include "llvm/Support/Debug.h" |
| 22 | #include "llvm/Support/Errc.h" |
| 23 | #include "llvm/Support/Error.h" |
| 24 | #include "llvm/Support/ErrorHandling.h" |
| 25 | #include "llvm/Support/Format.h" |
| 26 | #include "llvm/Support/MemoryBuffer.h" |
| 27 | |
| 28 | using namespace llvm; |
| 29 | using namespace ir2vec; |
| 30 | |
| 31 | #define DEBUG_TYPE "ir2vec" |
| 32 | |
| 33 | STATISTIC(VocabMissCounter, |
| 34 | "Number of lookups to entites not present in the vocabulary" ); |
| 35 | |
| 36 | namespace llvm { |
| 37 | namespace ir2vec { |
| 38 | static cl::OptionCategory IR2VecCategory("IR2Vec Options" ); |
| 39 | |
| 40 | // FIXME: Use a default vocab when not specified |
| 41 | static cl::opt<std::string> |
| 42 | VocabFile("ir2vec-vocab-path" , cl::Optional, |
| 43 | cl::desc("Path to the vocabulary file for IR2Vec" ), cl::init(Val: "" ), |
| 44 | cl::cat(IR2VecCategory)); |
| 45 | cl::opt<float> OpcWeight("ir2vec-opc-weight" , cl::Optional, cl::init(Val: 1.0), |
| 46 | cl::desc("Weight for opcode embeddings" ), |
| 47 | cl::cat(IR2VecCategory)); |
| 48 | cl::opt<float> TypeWeight("ir2vec-type-weight" , cl::Optional, cl::init(Val: 0.5), |
| 49 | cl::desc("Weight for type embeddings" ), |
| 50 | cl::cat(IR2VecCategory)); |
| 51 | cl::opt<float> ArgWeight("ir2vec-arg-weight" , cl::Optional, cl::init(Val: 0.2), |
| 52 | cl::desc("Weight for argument embeddings" ), |
| 53 | cl::cat(IR2VecCategory)); |
| 54 | } // namespace ir2vec |
| 55 | } // namespace llvm |
| 56 | |
| 57 | AnalysisKey IR2VecVocabAnalysis::Key; |
| 58 | |
| 59 | namespace llvm::json { |
| 60 | inline bool fromJSON(const llvm::json::Value &E, Embedding &Out, |
| 61 | llvm::json::Path P) { |
| 62 | std::vector<double> TempOut; |
| 63 | if (!llvm::json::fromJSON(E, Out&: TempOut, P)) |
| 64 | return false; |
| 65 | Out = Embedding(std::move(TempOut)); |
| 66 | return true; |
| 67 | } |
| 68 | } // namespace llvm::json |
| 69 | |
| 70 | // ==----------------------------------------------------------------------===// |
| 71 | // Embedding |
| 72 | //===----------------------------------------------------------------------===// |
| 73 | Embedding &Embedding::operator+=(const Embedding &RHS) { |
| 74 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 75 | std::transform(first1: this->begin(), last1: this->end(), first2: RHS.begin(), result: this->begin(), |
| 76 | binary_op: std::plus<double>()); |
| 77 | return *this; |
| 78 | } |
| 79 | |
| 80 | Embedding Embedding::operator+(const Embedding &RHS) const { |
| 81 | Embedding Result(*this); |
| 82 | Result += RHS; |
| 83 | return Result; |
| 84 | } |
| 85 | |
| 86 | Embedding &Embedding::operator-=(const Embedding &RHS) { |
| 87 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 88 | std::transform(first1: this->begin(), last1: this->end(), first2: RHS.begin(), result: this->begin(), |
| 89 | binary_op: std::minus<double>()); |
| 90 | return *this; |
| 91 | } |
| 92 | |
| 93 | Embedding Embedding::operator-(const Embedding &RHS) const { |
| 94 | Embedding Result(*this); |
| 95 | Result -= RHS; |
| 96 | return Result; |
| 97 | } |
| 98 | |
| 99 | Embedding &Embedding::operator*=(double Factor) { |
| 100 | std::transform(first: this->begin(), last: this->end(), result: this->begin(), |
| 101 | unary_op: [Factor](double Elem) { return Elem * Factor; }); |
| 102 | return *this; |
| 103 | } |
| 104 | |
| 105 | Embedding Embedding::operator*(double Factor) const { |
| 106 | Embedding Result(*this); |
| 107 | Result *= Factor; |
| 108 | return Result; |
| 109 | } |
| 110 | |
| 111 | Embedding &Embedding::scaleAndAdd(const Embedding &Src, float Factor) { |
| 112 | assert(this->size() == Src.size() && "Vectors must have the same dimension" ); |
| 113 | for (size_t Itr = 0; Itr < this->size(); ++Itr) |
| 114 | (*this)[Itr] += Src[Itr] * Factor; |
| 115 | return *this; |
| 116 | } |
| 117 | |
| 118 | bool Embedding::approximatelyEquals(const Embedding &RHS, |
| 119 | double Tolerance) const { |
| 120 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 121 | for (size_t Itr = 0; Itr < this->size(); ++Itr) |
| 122 | if (std::abs(x: (*this)[Itr] - RHS[Itr]) > Tolerance) |
| 123 | return false; |
| 124 | return true; |
| 125 | } |
| 126 | |
| 127 | void Embedding::print(raw_ostream &OS) const { |
| 128 | OS << " [" ; |
| 129 | for (const auto &Elem : Data) |
| 130 | OS << " " << format(Fmt: "%.2f" , Vals: Elem) << " " ; |
| 131 | OS << "]\n" ; |
| 132 | } |
| 133 | |
| 134 | // ==----------------------------------------------------------------------===// |
| 135 | // Embedder and its subclasses |
| 136 | //===----------------------------------------------------------------------===// |
| 137 | |
| 138 | Embedder::Embedder(const Function &F, const Vocab &Vocabulary) |
| 139 | : F(F), Vocabulary(Vocabulary), |
| 140 | Dimension(Vocabulary.begin()->second.size()), OpcWeight(::OpcWeight), |
| 141 | TypeWeight(::TypeWeight), ArgWeight(::ArgWeight) {} |
| 142 | |
| 143 | std::unique_ptr<Embedder> Embedder::create(IR2VecKind Mode, const Function &F, |
| 144 | const Vocab &Vocabulary) { |
| 145 | switch (Mode) { |
| 146 | case IR2VecKind::Symbolic: |
| 147 | return std::make_unique<SymbolicEmbedder>(args: F, args: Vocabulary); |
| 148 | } |
| 149 | return nullptr; |
| 150 | } |
| 151 | |
| 152 | // FIXME: Currently lookups are string based. Use numeric Keys |
| 153 | // for efficiency |
| 154 | Embedding Embedder::lookupVocab(const std::string &Key) const { |
| 155 | Embedding Vec(Dimension, 0); |
| 156 | // FIXME: Use zero vectors in vocab and assert failure for |
| 157 | // unknown entities rather than silently returning zeroes here. |
| 158 | auto It = Vocabulary.find(x: Key); |
| 159 | if (It != Vocabulary.end()) |
| 160 | return It->second; |
| 161 | LLVM_DEBUG(errs() << "cannot find key in map : " << Key << "\n" ); |
| 162 | ++VocabMissCounter; |
| 163 | return Vec; |
| 164 | } |
| 165 | |
| 166 | const InstEmbeddingsMap &Embedder::getInstVecMap() const { |
| 167 | if (InstVecMap.empty()) |
| 168 | computeEmbeddings(); |
| 169 | return InstVecMap; |
| 170 | } |
| 171 | |
| 172 | const BBEmbeddingsMap &Embedder::getBBVecMap() const { |
| 173 | if (BBVecMap.empty()) |
| 174 | computeEmbeddings(); |
| 175 | return BBVecMap; |
| 176 | } |
| 177 | |
| 178 | const Embedding &Embedder::getBBVector(const BasicBlock &BB) const { |
| 179 | auto It = BBVecMap.find(Val: &BB); |
| 180 | if (It != BBVecMap.end()) |
| 181 | return It->second; |
| 182 | computeEmbeddings(BB); |
| 183 | return BBVecMap[&BB]; |
| 184 | } |
| 185 | |
| 186 | const Embedding &Embedder::getFunctionVector() const { |
| 187 | // Currently, we always (re)compute the embeddings for the function. |
| 188 | // This is cheaper than caching the vector. |
| 189 | computeEmbeddings(); |
| 190 | return FuncVector; |
| 191 | } |
| 192 | |
| 193 | #define RETURN_LOOKUP_IF(CONDITION, KEY_STR) \ |
| 194 | if (CONDITION) \ |
| 195 | return lookupVocab(KEY_STR); |
| 196 | |
| 197 | Embedding SymbolicEmbedder::getTypeEmbedding(const Type *Ty) const { |
| 198 | RETURN_LOOKUP_IF(Ty->isVoidTy(), "voidTy" ); |
| 199 | RETURN_LOOKUP_IF(Ty->isFloatingPointTy(), "floatTy" ); |
| 200 | RETURN_LOOKUP_IF(Ty->isIntegerTy(), "integerTy" ); |
| 201 | RETURN_LOOKUP_IF(Ty->isFunctionTy(), "functionTy" ); |
| 202 | RETURN_LOOKUP_IF(Ty->isStructTy(), "structTy" ); |
| 203 | RETURN_LOOKUP_IF(Ty->isArrayTy(), "arrayTy" ); |
| 204 | RETURN_LOOKUP_IF(Ty->isPointerTy(), "pointerTy" ); |
| 205 | RETURN_LOOKUP_IF(Ty->isVectorTy(), "vectorTy" ); |
| 206 | RETURN_LOOKUP_IF(Ty->isEmptyTy(), "emptyTy" ); |
| 207 | RETURN_LOOKUP_IF(Ty->isLabelTy(), "labelTy" ); |
| 208 | RETURN_LOOKUP_IF(Ty->isTokenTy(), "tokenTy" ); |
| 209 | RETURN_LOOKUP_IF(Ty->isMetadataTy(), "metadataTy" ); |
| 210 | return lookupVocab(Key: "unknownTy" ); |
| 211 | } |
| 212 | |
| 213 | Embedding SymbolicEmbedder::getOperandEmbedding(const Value *Op) const { |
| 214 | RETURN_LOOKUP_IF(isa<Function>(Op), "function" ); |
| 215 | RETURN_LOOKUP_IF(isa<PointerType>(Op->getType()), "pointer" ); |
| 216 | RETURN_LOOKUP_IF(isa<Constant>(Op), "constant" ); |
| 217 | return lookupVocab(Key: "variable" ); |
| 218 | } |
| 219 | |
| 220 | #undef RETURN_LOOKUP_IF |
| 221 | |
| 222 | void SymbolicEmbedder::computeEmbeddings(const BasicBlock &BB) const { |
| 223 | Embedding BBVector(Dimension, 0); |
| 224 | |
| 225 | // We consider only the non-debug and non-pseudo instructions |
| 226 | for (const auto &I : BB.instructionsWithoutDebug()) { |
| 227 | Embedding InstVector(Dimension, 0); |
| 228 | |
| 229 | // FIXME: Currently lookups are string based. Use numeric Keys |
| 230 | // for efficiency. |
| 231 | InstVector += lookupVocab(Key: I.getOpcodeName()); |
| 232 | InstVector += getTypeEmbedding(Ty: I.getType()); |
| 233 | for (const auto &Op : I.operands()) { |
| 234 | InstVector += getOperandEmbedding(Op: Op.get()); |
| 235 | } |
| 236 | InstVecMap[&I] = InstVector; |
| 237 | BBVector += InstVector; |
| 238 | } |
| 239 | BBVecMap[&BB] = BBVector; |
| 240 | } |
| 241 | |
| 242 | void SymbolicEmbedder::computeEmbeddings() const { |
| 243 | if (F.isDeclaration()) |
| 244 | return; |
| 245 | |
| 246 | // Consider only the basic blocks that are reachable from entry |
| 247 | for (const BasicBlock *BB : depth_first(G: &F)) { |
| 248 | computeEmbeddings(BB: *BB); |
| 249 | FuncVector += BBVecMap[BB]; |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | // ==----------------------------------------------------------------------===// |
| 254 | // IR2VecVocabResult and IR2VecVocabAnalysis |
| 255 | //===----------------------------------------------------------------------===// |
| 256 | |
| 257 | IR2VecVocabResult::IR2VecVocabResult(ir2vec::Vocab &&Vocabulary) |
| 258 | : Vocabulary(std::move(Vocabulary)), Valid(true) {} |
| 259 | |
| 260 | const ir2vec::Vocab &IR2VecVocabResult::getVocabulary() const { |
| 261 | assert(Valid && "IR2Vec Vocabulary is invalid" ); |
| 262 | return Vocabulary; |
| 263 | } |
| 264 | |
| 265 | unsigned IR2VecVocabResult::getDimension() const { |
| 266 | assert(Valid && "IR2Vec Vocabulary is invalid" ); |
| 267 | return Vocabulary.begin()->second.size(); |
| 268 | } |
| 269 | |
| 270 | // For now, assume vocabulary is stable unless explicitly invalidated. |
| 271 | bool IR2VecVocabResult::invalidate( |
| 272 | Module &M, const PreservedAnalyses &PA, |
| 273 | ModuleAnalysisManager::Invalidator &Inv) const { |
| 274 | auto PAC = PA.getChecker<IR2VecVocabAnalysis>(); |
| 275 | return !(PAC.preservedWhenStateless()); |
| 276 | } |
| 277 | |
| 278 | Error IR2VecVocabAnalysis::parseVocabSection( |
| 279 | StringRef Key, const json::Value &ParsedVocabValue, |
| 280 | ir2vec::Vocab &TargetVocab, unsigned &Dim) { |
| 281 | json::Path::Root Path("" ); |
| 282 | const json::Object *RootObj = ParsedVocabValue.getAsObject(); |
| 283 | if (!RootObj) |
| 284 | return createStringError(EC: errc::invalid_argument, |
| 285 | S: "JSON root is not an object" ); |
| 286 | |
| 287 | const json::Value *SectionValue = RootObj->get(K: Key); |
| 288 | if (!SectionValue) |
| 289 | return createStringError(EC: errc::invalid_argument, |
| 290 | S: "Missing '" + std::string(Key) + |
| 291 | "' section in vocabulary file" ); |
| 292 | if (!json::fromJSON(E: *SectionValue, Out&: TargetVocab, P: Path)) |
| 293 | return createStringError(EC: errc::illegal_byte_sequence, |
| 294 | S: "Unable to parse '" + std::string(Key) + |
| 295 | "' section from vocabulary" ); |
| 296 | |
| 297 | Dim = TargetVocab.begin()->second.size(); |
| 298 | if (Dim == 0) |
| 299 | return createStringError(EC: errc::illegal_byte_sequence, |
| 300 | S: "Dimension of '" + std::string(Key) + |
| 301 | "' section of the vocabulary is zero" ); |
| 302 | |
| 303 | if (!std::all_of(first: TargetVocab.begin(), last: TargetVocab.end(), |
| 304 | pred: [Dim](const std::pair<StringRef, Embedding> &Entry) { |
| 305 | return Entry.second.size() == Dim; |
| 306 | })) |
| 307 | return createStringError( |
| 308 | EC: errc::illegal_byte_sequence, |
| 309 | S: "All vectors in the '" + std::string(Key) + |
| 310 | "' section of the vocabulary are not of the same dimension" ); |
| 311 | |
| 312 | return Error::success(); |
| 313 | } |
| 314 | |
| 315 | // FIXME: Make this optional. We can avoid file reads |
| 316 | // by auto-generating a default vocabulary during the build time. |
| 317 | Error IR2VecVocabAnalysis::readVocabulary() { |
| 318 | auto BufOrError = MemoryBuffer::getFileOrSTDIN(Filename: VocabFile, /*IsText=*/true); |
| 319 | if (!BufOrError) |
| 320 | return createFileError(F: VocabFile, EC: BufOrError.getError()); |
| 321 | |
| 322 | auto Content = BufOrError.get()->getBuffer(); |
| 323 | |
| 324 | Expected<json::Value> ParsedVocabValue = json::parse(JSON: Content); |
| 325 | if (!ParsedVocabValue) |
| 326 | return ParsedVocabValue.takeError(); |
| 327 | |
| 328 | ir2vec::Vocab OpcodeVocab, TypeVocab, ArgVocab; |
| 329 | unsigned OpcodeDim = 0, TypeDim = 0, ArgDim = 0; |
| 330 | if (auto Err = parseVocabSection(Key: "Opcodes" , ParsedVocabValue: *ParsedVocabValue, TargetVocab&: OpcodeVocab, |
| 331 | Dim&: OpcodeDim)) |
| 332 | return Err; |
| 333 | |
| 334 | if (auto Err = |
| 335 | parseVocabSection(Key: "Types" , ParsedVocabValue: *ParsedVocabValue, TargetVocab&: TypeVocab, Dim&: TypeDim)) |
| 336 | return Err; |
| 337 | |
| 338 | if (auto Err = |
| 339 | parseVocabSection(Key: "Arguments" , ParsedVocabValue: *ParsedVocabValue, TargetVocab&: ArgVocab, Dim&: ArgDim)) |
| 340 | return Err; |
| 341 | |
| 342 | if (!(OpcodeDim == TypeDim && TypeDim == ArgDim)) |
| 343 | return createStringError(EC: errc::illegal_byte_sequence, |
| 344 | S: "Vocabulary sections have different dimensions" ); |
| 345 | |
| 346 | auto scaleVocabSection = [](ir2vec::Vocab &Vocab, double Weight) { |
| 347 | for (auto &Entry : Vocab) |
| 348 | Entry.second *= Weight; |
| 349 | }; |
| 350 | scaleVocabSection(OpcodeVocab, OpcWeight); |
| 351 | scaleVocabSection(TypeVocab, TypeWeight); |
| 352 | scaleVocabSection(ArgVocab, ArgWeight); |
| 353 | |
| 354 | Vocabulary.insert(first: OpcodeVocab.begin(), last: OpcodeVocab.end()); |
| 355 | Vocabulary.insert(first: TypeVocab.begin(), last: TypeVocab.end()); |
| 356 | Vocabulary.insert(first: ArgVocab.begin(), last: ArgVocab.end()); |
| 357 | |
| 358 | return Error::success(); |
| 359 | } |
| 360 | |
| 361 | IR2VecVocabAnalysis::IR2VecVocabAnalysis(const Vocab &Vocabulary) |
| 362 | : Vocabulary(Vocabulary) {} |
| 363 | |
| 364 | IR2VecVocabAnalysis::IR2VecVocabAnalysis(Vocab &&Vocabulary) |
| 365 | : Vocabulary(std::move(Vocabulary)) {} |
| 366 | |
| 367 | void IR2VecVocabAnalysis::emitError(Error Err, LLVMContext &Ctx) { |
| 368 | handleAllErrors(E: std::move(Err), Handlers: [&](const ErrorInfoBase &EI) { |
| 369 | Ctx.emitError(ErrorStr: "Error reading vocabulary: " + EI.message()); |
| 370 | }); |
| 371 | } |
| 372 | |
| 373 | IR2VecVocabAnalysis::Result |
| 374 | IR2VecVocabAnalysis::run(Module &M, ModuleAnalysisManager &AM) { |
| 375 | auto Ctx = &M.getContext(); |
| 376 | // If vocabulary is already populated by the constructor, use it. |
| 377 | if (!Vocabulary.empty()) |
| 378 | return IR2VecVocabResult(std::move(Vocabulary)); |
| 379 | |
| 380 | // Otherwise, try to read from the vocabulary file. |
| 381 | if (VocabFile.empty()) { |
| 382 | // FIXME: Use default vocabulary |
| 383 | Ctx->emitError(ErrorStr: "IR2Vec vocabulary file path not specified" ); |
| 384 | return IR2VecVocabResult(); // Return invalid result |
| 385 | } |
| 386 | if (auto Err = readVocabulary()) { |
| 387 | emitError(Err: std::move(Err), Ctx&: *Ctx); |
| 388 | return IR2VecVocabResult(); |
| 389 | } |
| 390 | return IR2VecVocabResult(std::move(Vocabulary)); |
| 391 | } |
| 392 | |
| 393 | // ==----------------------------------------------------------------------===// |
| 394 | // Printer Passes |
| 395 | //===----------------------------------------------------------------------===// |
| 396 | |
| 397 | PreservedAnalyses IR2VecPrinterPass::run(Module &M, |
| 398 | ModuleAnalysisManager &MAM) { |
| 399 | auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(IR&: M); |
| 400 | assert(IR2VecVocabResult.isValid() && "IR2Vec Vocabulary is invalid" ); |
| 401 | |
| 402 | auto Vocab = IR2VecVocabResult.getVocabulary(); |
| 403 | for (Function &F : M) { |
| 404 | std::unique_ptr<Embedder> Emb = |
| 405 | Embedder::create(Mode: IR2VecKind::Symbolic, F, Vocabulary: Vocab); |
| 406 | if (!Emb) { |
| 407 | OS << "Error creating IR2Vec embeddings \n" ; |
| 408 | continue; |
| 409 | } |
| 410 | |
| 411 | OS << "IR2Vec embeddings for function " << F.getName() << ":\n" ; |
| 412 | OS << "Function vector: " ; |
| 413 | Emb->getFunctionVector().print(OS); |
| 414 | |
| 415 | OS << "Basic block vectors:\n" ; |
| 416 | const auto &BBMap = Emb->getBBVecMap(); |
| 417 | for (const BasicBlock &BB : F) { |
| 418 | auto It = BBMap.find(Val: &BB); |
| 419 | if (It != BBMap.end()) { |
| 420 | OS << "Basic block: " << BB.getName() << ":\n" ; |
| 421 | It->second.print(OS); |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | OS << "Instruction vectors:\n" ; |
| 426 | const auto &InstMap = Emb->getInstVecMap(); |
| 427 | for (const BasicBlock &BB : F) { |
| 428 | for (const Instruction &I : BB) { |
| 429 | auto It = InstMap.find(Val: &I); |
| 430 | if (It != InstMap.end()) { |
| 431 | OS << "Instruction: " ; |
| 432 | I.print(O&: OS); |
| 433 | It->second.print(OS); |
| 434 | } |
| 435 | } |
| 436 | } |
| 437 | } |
| 438 | return PreservedAnalyses::all(); |
| 439 | } |
| 440 | |
| 441 | PreservedAnalyses IR2VecVocabPrinterPass::run(Module &M, |
| 442 | ModuleAnalysisManager &MAM) { |
| 443 | auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(IR&: M); |
| 444 | assert(IR2VecVocabResult.isValid() && "IR2Vec Vocabulary is invalid" ); |
| 445 | |
| 446 | auto Vocab = IR2VecVocabResult.getVocabulary(); |
| 447 | for (const auto &Entry : Vocab) { |
| 448 | OS << "Key: " << Entry.first << ": " ; |
| 449 | Entry.second.print(OS); |
| 450 | } |
| 451 | |
| 452 | return PreservedAnalyses::all(); |
| 453 | } |
| 454 | |