| 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/Sequence.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/ADT/Statistic.h" |
| 20 | #include "llvm/IR/CFG.h" |
| 21 | #include "llvm/IR/Module.h" |
| 22 | #include "llvm/IR/PassManager.h" |
| 23 | #include "llvm/Support/Debug.h" |
| 24 | #include "llvm/Support/Errc.h" |
| 25 | #include "llvm/Support/Error.h" |
| 26 | #include "llvm/Support/ErrorHandling.h" |
| 27 | #include "llvm/Support/Format.h" |
| 28 | #include "llvm/Support/MemoryBuffer.h" |
| 29 | |
| 30 | using namespace llvm; |
| 31 | using namespace ir2vec; |
| 32 | |
| 33 | #define DEBUG_TYPE "ir2vec" |
| 34 | |
| 35 | STATISTIC(VocabMissCounter, |
| 36 | "Number of lookups to entities not present in the vocabulary" ); |
| 37 | |
| 38 | namespace llvm { |
| 39 | namespace ir2vec { |
| 40 | cl::OptionCategory IR2VecCategory("IR2Vec Options" ); |
| 41 | |
| 42 | // FIXME: Use a default vocab when not specified |
| 43 | cl::opt<std::string> |
| 44 | VocabFile("ir2vec-vocab-path" , cl::Optional, |
| 45 | cl::desc("Path to the vocabulary file for IR2Vec" ), cl::init(Val: "" ), |
| 46 | cl::cat(IR2VecCategory)); |
| 47 | cl::opt<float> OpcWeight("ir2vec-opc-weight" , cl::Optional, cl::init(Val: 1.0), |
| 48 | cl::desc("Weight for opcode embeddings" ), |
| 49 | cl::cat(IR2VecCategory)); |
| 50 | cl::opt<float> TypeWeight("ir2vec-type-weight" , cl::Optional, cl::init(Val: 0.5), |
| 51 | cl::desc("Weight for type embeddings" ), |
| 52 | cl::cat(IR2VecCategory)); |
| 53 | cl::opt<float> ArgWeight("ir2vec-arg-weight" , cl::Optional, cl::init(Val: 0.2), |
| 54 | cl::desc("Weight for argument embeddings" ), |
| 55 | cl::cat(IR2VecCategory)); |
| 56 | cl::opt<IR2VecKind> IR2VecEmbeddingKind( |
| 57 | "ir2vec-kind" , cl::Optional, |
| 58 | cl::values(clEnumValN(IR2VecKind::Symbolic, "symbolic" , |
| 59 | "Generate symbolic embeddings" ), |
| 60 | clEnumValN(IR2VecKind::FlowAware, "flow-aware" , |
| 61 | "Generate flow-aware embeddings" )), |
| 62 | cl::init(Val: IR2VecKind::Symbolic), cl::desc("IR2Vec embedding kind" ), |
| 63 | cl::cat(IR2VecCategory)); |
| 64 | |
| 65 | } // namespace ir2vec |
| 66 | } // namespace llvm |
| 67 | |
| 68 | AnalysisKey IR2VecVocabAnalysis::Key; |
| 69 | |
| 70 | // ==----------------------------------------------------------------------===// |
| 71 | // Local helper functions |
| 72 | //===----------------------------------------------------------------------===// |
| 73 | namespace llvm::json { |
| 74 | inline bool fromJSON(const llvm::json::Value &E, Embedding &Out, |
| 75 | llvm::json::Path P) { |
| 76 | std::vector<double> TempOut; |
| 77 | if (!llvm::json::fromJSON(E, Out&: TempOut, P)) |
| 78 | return false; |
| 79 | Out = Embedding(std::move(TempOut)); |
| 80 | return true; |
| 81 | } |
| 82 | } // namespace llvm::json |
| 83 | |
| 84 | // ==----------------------------------------------------------------------===// |
| 85 | // Embedding |
| 86 | //===----------------------------------------------------------------------===// |
| 87 | Embedding &Embedding::operator+=(const Embedding &RHS) { |
| 88 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 89 | std::transform(first1: this->begin(), last1: this->end(), first2: RHS.begin(), result: this->begin(), |
| 90 | binary_op: std::plus<double>()); |
| 91 | return *this; |
| 92 | } |
| 93 | |
| 94 | Embedding Embedding::operator+(const Embedding &RHS) const { |
| 95 | Embedding Result(*this); |
| 96 | Result += RHS; |
| 97 | return Result; |
| 98 | } |
| 99 | |
| 100 | Embedding &Embedding::operator-=(const Embedding &RHS) { |
| 101 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 102 | std::transform(first1: this->begin(), last1: this->end(), first2: RHS.begin(), result: this->begin(), |
| 103 | binary_op: std::minus<double>()); |
| 104 | return *this; |
| 105 | } |
| 106 | |
| 107 | Embedding Embedding::operator-(const Embedding &RHS) const { |
| 108 | Embedding Result(*this); |
| 109 | Result -= RHS; |
| 110 | return Result; |
| 111 | } |
| 112 | |
| 113 | Embedding &Embedding::operator*=(double Factor) { |
| 114 | std::transform(first: this->begin(), last: this->end(), result: this->begin(), |
| 115 | unary_op: [Factor](double Elem) { return Elem * Factor; }); |
| 116 | return *this; |
| 117 | } |
| 118 | |
| 119 | Embedding Embedding::operator*(double Factor) const { |
| 120 | Embedding Result(*this); |
| 121 | Result *= Factor; |
| 122 | return Result; |
| 123 | } |
| 124 | |
| 125 | Embedding &Embedding::scaleAndAdd(const Embedding &Src, float Factor) { |
| 126 | assert(this->size() == Src.size() && "Vectors must have the same dimension" ); |
| 127 | for (size_t Itr = 0; Itr < this->size(); ++Itr) |
| 128 | (*this)[Itr] += Src[Itr] * Factor; |
| 129 | return *this; |
| 130 | } |
| 131 | |
| 132 | bool Embedding::approximatelyEquals(const Embedding &RHS, |
| 133 | double Tolerance) const { |
| 134 | assert(this->size() == RHS.size() && "Vectors must have the same dimension" ); |
| 135 | for (size_t Itr = 0; Itr < this->size(); ++Itr) |
| 136 | if (std::abs(x: (*this)[Itr] - RHS[Itr]) > Tolerance) { |
| 137 | LLVM_DEBUG(errs() << "Embedding mismatch at index " << Itr << ": " |
| 138 | << (*this)[Itr] << " vs " << RHS[Itr] |
| 139 | << "; Tolerance: " << Tolerance << "\n" ); |
| 140 | return false; |
| 141 | } |
| 142 | return true; |
| 143 | } |
| 144 | |
| 145 | void Embedding::print(raw_ostream &OS) const { |
| 146 | OS << " [" ; |
| 147 | for (const auto &Elem : Data) |
| 148 | OS << " " << format(Fmt: "%.2f" , Vals: Elem) << " " ; |
| 149 | OS << "]\n" ; |
| 150 | } |
| 151 | |
| 152 | // ==----------------------------------------------------------------------===// |
| 153 | // Embedder and its subclasses |
| 154 | //===----------------------------------------------------------------------===// |
| 155 | |
| 156 | std::unique_ptr<Embedder> Embedder::create(IR2VecKind Mode, const Function &F, |
| 157 | const Vocabulary &Vocab) { |
| 158 | switch (Mode) { |
| 159 | case IR2VecKind::Symbolic: |
| 160 | return std::make_unique<SymbolicEmbedder>(args: F, args: Vocab); |
| 161 | case IR2VecKind::FlowAware: |
| 162 | return std::make_unique<FlowAwareEmbedder>(args: F, args: Vocab); |
| 163 | } |
| 164 | return nullptr; |
| 165 | } |
| 166 | |
| 167 | Embedding Embedder::computeEmbeddings() const { |
| 168 | Embedding FuncVector(Dimension, 0.0); |
| 169 | |
| 170 | if (F.isDeclaration()) |
| 171 | return FuncVector; |
| 172 | |
| 173 | // Consider only the basic blocks that are reachable from entry |
| 174 | for (const BasicBlock *BB : depth_first(G: &F)) |
| 175 | FuncVector += computeEmbeddings(BB: *BB); |
| 176 | return FuncVector; |
| 177 | } |
| 178 | |
| 179 | Embedding Embedder::computeEmbeddings(const BasicBlock &BB) const { |
| 180 | Embedding BBVector(Dimension, 0); |
| 181 | |
| 182 | // We consider only the non-debug and non-pseudo instructions |
| 183 | for (const auto &I : BB.instructionsWithoutDebug()) |
| 184 | BBVector += computeEmbeddings(I); |
| 185 | return BBVector; |
| 186 | } |
| 187 | |
| 188 | Embedding SymbolicEmbedder::computeEmbeddings(const Instruction &I) const { |
| 189 | // Currently, we always (re)compute the embeddings for symbolic embedder. |
| 190 | // This is cheaper than caching the vectors. |
| 191 | Embedding ArgEmb(Dimension, 0); |
| 192 | for (const auto &Op : I.operands()) |
| 193 | ArgEmb += Vocab[*Op]; |
| 194 | auto InstVector = |
| 195 | Vocab[I.getOpcode()] + Vocab[I.getType()->getTypeID()] + ArgEmb; |
| 196 | if (const auto *IC = dyn_cast<CmpInst>(Val: &I)) |
| 197 | InstVector += Vocab[IC->getPredicate()]; |
| 198 | return InstVector; |
| 199 | } |
| 200 | |
| 201 | Embedding FlowAwareEmbedder::computeEmbeddings(const Instruction &I) const { |
| 202 | // If we have already computed the embedding for this instruction, return it |
| 203 | auto It = InstVecMap.find(Val: &I); |
| 204 | if (It != InstVecMap.end()) |
| 205 | return It->second; |
| 206 | |
| 207 | // TODO: Handle call instructions differently. |
| 208 | // For now, we treat them like other instructions |
| 209 | Embedding ArgEmb(Dimension, 0); |
| 210 | for (const auto &Op : I.operands()) { |
| 211 | // If the operand is defined elsewhere, we use its embedding |
| 212 | if (const auto *DefInst = dyn_cast<Instruction>(Val: Op)) { |
| 213 | auto DefIt = InstVecMap.find(Val: DefInst); |
| 214 | // Fixme (#159171): Ideally we should never miss an instruction |
| 215 | // embedding here. |
| 216 | // But when we have cyclic dependencies (e.g., phi |
| 217 | // nodes), we might miss the embedding. In such cases, we fall back to |
| 218 | // using the vocabulary embedding. This can be fixed by iterating to a |
| 219 | // fixed-point, or by using a simple solver for the set of simultaneous |
| 220 | // equations. |
| 221 | // Another case when we might miss an instruction embedding is when |
| 222 | // the operand instruction is in a different basic block that has not |
| 223 | // been processed yet. This can be fixed by processing the basic blocks |
| 224 | // in a topological order. |
| 225 | if (DefIt != InstVecMap.end()) |
| 226 | ArgEmb += DefIt->second; |
| 227 | else |
| 228 | ArgEmb += Vocab[*Op]; |
| 229 | } |
| 230 | // If the operand is not defined by an instruction, we use the |
| 231 | // vocabulary |
| 232 | else { |
| 233 | LLVM_DEBUG(errs() << "Using embedding from vocabulary for operand: " |
| 234 | << *Op << "=" << Vocab[*Op][0] << "\n" ); |
| 235 | ArgEmb += Vocab[*Op]; |
| 236 | } |
| 237 | } |
| 238 | // Create the instruction vector by combining opcode, type, and arguments |
| 239 | // embeddings |
| 240 | auto InstVector = |
| 241 | Vocab[I.getOpcode()] + Vocab[I.getType()->getTypeID()] + ArgEmb; |
| 242 | if (const auto *IC = dyn_cast<CmpInst>(Val: &I)) |
| 243 | InstVector += Vocab[IC->getPredicate()]; |
| 244 | InstVecMap[&I] = InstVector; |
| 245 | return InstVector; |
| 246 | } |
| 247 | |
| 248 | // ==----------------------------------------------------------------------===// |
| 249 | // VocabStorage |
| 250 | //===----------------------------------------------------------------------===// |
| 251 | |
| 252 | VocabStorage::VocabStorage(std::vector<std::vector<Embedding>> &&SectionData) |
| 253 | : Sections(std::move(SectionData)), TotalSize([&] { |
| 254 | assert(!Sections.empty() && "Vocabulary has no sections" ); |
| 255 | // Compute total size across all sections |
| 256 | size_t Size = 0; |
| 257 | for (const auto &Section : Sections) { |
| 258 | assert(!Section.empty() && "Vocabulary section is empty" ); |
| 259 | Size += Section.size(); |
| 260 | } |
| 261 | return Size; |
| 262 | }()), |
| 263 | Dimension([&] { |
| 264 | // Get dimension from the first embedding in the first section - all |
| 265 | // embeddings must have the same dimension |
| 266 | assert(!Sections.empty() && "Vocabulary has no sections" ); |
| 267 | assert(!Sections[0].empty() && "First section of vocabulary is empty" ); |
| 268 | unsigned ExpectedDim = static_cast<unsigned>(Sections[0][0].size()); |
| 269 | |
| 270 | // Verify that all embeddings across all sections have the same |
| 271 | // dimension |
| 272 | [[maybe_unused]] auto allSameDim = |
| 273 | [ExpectedDim](const std::vector<Embedding> &Section) { |
| 274 | return std::all_of(first: Section.begin(), last: Section.end(), |
| 275 | pred: [ExpectedDim](const Embedding &Emb) { |
| 276 | return Emb.size() == ExpectedDim; |
| 277 | }); |
| 278 | }; |
| 279 | assert(std::all_of(Sections.begin(), Sections.end(), allSameDim) && |
| 280 | "All embeddings must have the same dimension" ); |
| 281 | |
| 282 | return ExpectedDim; |
| 283 | }()) {} |
| 284 | |
| 285 | const Embedding &VocabStorage::const_iterator::operator*() const { |
| 286 | assert(SectionId < Storage->Sections.size() && "Invalid section ID" ); |
| 287 | assert(LocalIndex < Storage->Sections[SectionId].size() && |
| 288 | "Local index out of range" ); |
| 289 | return Storage->Sections[SectionId][LocalIndex]; |
| 290 | } |
| 291 | |
| 292 | VocabStorage::const_iterator &VocabStorage::const_iterator::operator++() { |
| 293 | ++LocalIndex; |
| 294 | // Check if we need to move to the next section |
| 295 | if (SectionId < Storage->getNumSections() && |
| 296 | LocalIndex >= Storage->Sections[SectionId].size()) { |
| 297 | assert(LocalIndex == Storage->Sections[SectionId].size() && |
| 298 | "Local index should be at the end of the current section" ); |
| 299 | LocalIndex = 0; |
| 300 | ++SectionId; |
| 301 | } |
| 302 | return *this; |
| 303 | } |
| 304 | |
| 305 | bool VocabStorage::const_iterator::operator==( |
| 306 | const const_iterator &Other) const { |
| 307 | return Storage == Other.Storage && SectionId == Other.SectionId && |
| 308 | LocalIndex == Other.LocalIndex; |
| 309 | } |
| 310 | |
| 311 | bool VocabStorage::const_iterator::operator!=( |
| 312 | const const_iterator &Other) const { |
| 313 | return !(*this == Other); |
| 314 | } |
| 315 | |
| 316 | Error VocabStorage::parseVocabSection(StringRef Key, |
| 317 | const json::Value &ParsedVocabValue, |
| 318 | VocabMap &TargetVocab, unsigned &Dim) { |
| 319 | json::Path::Root Path("" ); |
| 320 | const json::Object *RootObj = ParsedVocabValue.getAsObject(); |
| 321 | if (!RootObj) |
| 322 | return createStringError(EC: errc::invalid_argument, |
| 323 | S: "JSON root is not an object" ); |
| 324 | |
| 325 | const json::Value *SectionValue = RootObj->get(K: Key); |
| 326 | if (!SectionValue) |
| 327 | return createStringError(EC: errc::invalid_argument, |
| 328 | S: "Missing '" + std::string(Key) + |
| 329 | "' section in vocabulary file" ); |
| 330 | if (!json::fromJSON(E: *SectionValue, Out&: TargetVocab, P: Path)) |
| 331 | return createStringError(EC: errc::illegal_byte_sequence, |
| 332 | S: "Unable to parse '" + std::string(Key) + |
| 333 | "' section from vocabulary" ); |
| 334 | |
| 335 | Dim = TargetVocab.begin()->second.size(); |
| 336 | if (Dim == 0) |
| 337 | return createStringError(EC: errc::illegal_byte_sequence, |
| 338 | S: "Dimension of '" + std::string(Key) + |
| 339 | "' section of the vocabulary is zero" ); |
| 340 | |
| 341 | if (!std::all_of(first: TargetVocab.begin(), last: TargetVocab.end(), |
| 342 | pred: [Dim](const std::pair<StringRef, Embedding> &Entry) { |
| 343 | return Entry.second.size() == Dim; |
| 344 | })) |
| 345 | return createStringError( |
| 346 | EC: errc::illegal_byte_sequence, |
| 347 | S: "All vectors in the '" + std::string(Key) + |
| 348 | "' section of the vocabulary are not of the same dimension" ); |
| 349 | |
| 350 | return Error::success(); |
| 351 | } |
| 352 | |
| 353 | // ==----------------------------------------------------------------------===// |
| 354 | // Vocabulary |
| 355 | //===----------------------------------------------------------------------===// |
| 356 | |
| 357 | StringRef Vocabulary::getVocabKeyForOpcode(unsigned Opcode) { |
| 358 | assert(Opcode >= 1 && Opcode <= MaxOpcodes && "Invalid opcode" ); |
| 359 | #define HANDLE_INST(NUM, OPCODE, CLASS) \ |
| 360 | if (Opcode == NUM) { \ |
| 361 | return #OPCODE; \ |
| 362 | } |
| 363 | #include "llvm/IR/Instruction.def" |
| 364 | #undef HANDLE_INST |
| 365 | return "UnknownOpcode" ; |
| 366 | } |
| 367 | |
| 368 | // Helper function to classify an operand into OperandKind |
| 369 | Vocabulary::OperandKind Vocabulary::getOperandKind(const Value *Op) { |
| 370 | if (isa<Function>(Val: Op)) |
| 371 | return OperandKind::FunctionID; |
| 372 | if (isa<PointerType>(Val: Op->getType())) |
| 373 | return OperandKind::PointerID; |
| 374 | if (isa<Constant>(Val: Op)) |
| 375 | return OperandKind::ConstantID; |
| 376 | return OperandKind::VariableID; |
| 377 | } |
| 378 | |
| 379 | unsigned Vocabulary::getPredicateLocalIndex(CmpInst::Predicate P) { |
| 380 | if (P >= CmpInst::FIRST_FCMP_PREDICATE && P <= CmpInst::LAST_FCMP_PREDICATE) |
| 381 | return P - CmpInst::FIRST_FCMP_PREDICATE; |
| 382 | else |
| 383 | return P - CmpInst::FIRST_ICMP_PREDICATE + |
| 384 | (CmpInst::LAST_FCMP_PREDICATE - CmpInst::FIRST_FCMP_PREDICATE + 1); |
| 385 | } |
| 386 | |
| 387 | CmpInst::Predicate Vocabulary::getPredicateFromLocalIndex(unsigned LocalIndex) { |
| 388 | unsigned fcmpRange = |
| 389 | CmpInst::LAST_FCMP_PREDICATE - CmpInst::FIRST_FCMP_PREDICATE + 1; |
| 390 | if (LocalIndex < fcmpRange) |
| 391 | return static_cast<CmpInst::Predicate>(CmpInst::FIRST_FCMP_PREDICATE + |
| 392 | LocalIndex); |
| 393 | else |
| 394 | return static_cast<CmpInst::Predicate>(CmpInst::FIRST_ICMP_PREDICATE + |
| 395 | LocalIndex - fcmpRange); |
| 396 | } |
| 397 | |
| 398 | StringRef Vocabulary::getVocabKeyForPredicate(CmpInst::Predicate Pred) { |
| 399 | static SmallString<16> PredNameBuffer; |
| 400 | if (Pred < CmpInst::FIRST_ICMP_PREDICATE) |
| 401 | PredNameBuffer = "FCMP_" ; |
| 402 | else |
| 403 | PredNameBuffer = "ICMP_" ; |
| 404 | PredNameBuffer += CmpInst::getPredicateName(P: Pred); |
| 405 | return PredNameBuffer; |
| 406 | } |
| 407 | |
| 408 | StringRef Vocabulary::getStringKey(unsigned Pos) { |
| 409 | assert(Pos < NumCanonicalEntries && "Position out of bounds in vocabulary" ); |
| 410 | // Opcode |
| 411 | if (Pos < MaxOpcodes) |
| 412 | return getVocabKeyForOpcode(Opcode: Pos + 1); |
| 413 | // Type |
| 414 | if (Pos < OperandBaseOffset) |
| 415 | return getVocabKeyForCanonicalTypeID( |
| 416 | CType: static_cast<CanonicalTypeID>(Pos - MaxOpcodes)); |
| 417 | // Operand |
| 418 | if (Pos < PredicateBaseOffset) |
| 419 | return getVocabKeyForOperandKind( |
| 420 | Kind: static_cast<OperandKind>(Pos - OperandBaseOffset)); |
| 421 | // Predicates |
| 422 | return getVocabKeyForPredicate(Pred: getPredicate(Index: Pos - PredicateBaseOffset)); |
| 423 | } |
| 424 | |
| 425 | // For now, assume vocabulary is stable unless explicitly invalidated. |
| 426 | bool Vocabulary::invalidate(Module &M, const PreservedAnalyses &PA, |
| 427 | ModuleAnalysisManager::Invalidator &Inv) const { |
| 428 | auto PAC = PA.getChecker<IR2VecVocabAnalysis>(); |
| 429 | return !(PAC.preservedWhenStateless()); |
| 430 | } |
| 431 | |
| 432 | VocabStorage Vocabulary::createDummyVocabForTest(unsigned Dim) { |
| 433 | float DummyVal = 0.1f; |
| 434 | |
| 435 | // Create sections for opcodes, types, operands, and predicates |
| 436 | // Order must match Vocabulary::Section enum |
| 437 | std::vector<std::vector<Embedding>> Sections; |
| 438 | Sections.reserve(n: 4); |
| 439 | |
| 440 | // Opcodes section |
| 441 | std::vector<Embedding> OpcodeSec; |
| 442 | OpcodeSec.reserve(n: MaxOpcodes); |
| 443 | for (unsigned I = 0; I < MaxOpcodes; ++I) { |
| 444 | OpcodeSec.emplace_back(args&: Dim, args&: DummyVal); |
| 445 | DummyVal += 0.1f; |
| 446 | } |
| 447 | Sections.push_back(x: std::move(OpcodeSec)); |
| 448 | |
| 449 | // Types section |
| 450 | std::vector<Embedding> TypeSec; |
| 451 | TypeSec.reserve(n: MaxCanonicalTypeIDs); |
| 452 | for (unsigned I = 0; I < MaxCanonicalTypeIDs; ++I) { |
| 453 | TypeSec.emplace_back(args&: Dim, args&: DummyVal); |
| 454 | DummyVal += 0.1f; |
| 455 | } |
| 456 | Sections.push_back(x: std::move(TypeSec)); |
| 457 | |
| 458 | // Operands section |
| 459 | std::vector<Embedding> OperandSec; |
| 460 | OperandSec.reserve(n: MaxOperandKinds); |
| 461 | for (unsigned I = 0; I < MaxOperandKinds; ++I) { |
| 462 | OperandSec.emplace_back(args&: Dim, args&: DummyVal); |
| 463 | DummyVal += 0.1f; |
| 464 | } |
| 465 | Sections.push_back(x: std::move(OperandSec)); |
| 466 | |
| 467 | // Predicates section |
| 468 | std::vector<Embedding> PredicateSec; |
| 469 | PredicateSec.reserve(n: MaxPredicateKinds); |
| 470 | for (unsigned I = 0; I < MaxPredicateKinds; ++I) { |
| 471 | PredicateSec.emplace_back(args&: Dim, args&: DummyVal); |
| 472 | DummyVal += 0.1f; |
| 473 | } |
| 474 | Sections.push_back(x: std::move(PredicateSec)); |
| 475 | |
| 476 | return VocabStorage(std::move(Sections)); |
| 477 | } |
| 478 | |
| 479 | namespace { |
| 480 | using VocabMap = std::map<std::string, Embedding>; |
| 481 | |
| 482 | /// Read vocabulary JSON file and populate the section maps. |
| 483 | Error readVocabularyFromFile(StringRef VocabFilePath, VocabMap &OpcVocab, |
| 484 | VocabMap &TypeVocab, VocabMap &ArgVocab) { |
| 485 | auto BufOrError = |
| 486 | MemoryBuffer::getFileOrSTDIN(Filename: VocabFilePath, /*IsText=*/true); |
| 487 | if (!BufOrError) |
| 488 | return createFileError(F: VocabFilePath, EC: BufOrError.getError()); |
| 489 | |
| 490 | auto Content = BufOrError.get()->getBuffer(); |
| 491 | |
| 492 | Expected<json::Value> ParsedVocabValue = json::parse(JSON: Content); |
| 493 | if (!ParsedVocabValue) |
| 494 | return ParsedVocabValue.takeError(); |
| 495 | |
| 496 | unsigned OpcodeDim = 0, TypeDim = 0, ArgDim = 0; |
| 497 | if (auto Err = VocabStorage::parseVocabSection(Key: "Opcodes" , ParsedVocabValue: *ParsedVocabValue, |
| 498 | TargetVocab&: OpcVocab, Dim&: OpcodeDim)) |
| 499 | return Err; |
| 500 | |
| 501 | if (auto Err = VocabStorage::parseVocabSection(Key: "Types" , ParsedVocabValue: *ParsedVocabValue, |
| 502 | TargetVocab&: TypeVocab, Dim&: TypeDim)) |
| 503 | return Err; |
| 504 | |
| 505 | if (auto Err = VocabStorage::parseVocabSection(Key: "Arguments" , ParsedVocabValue: *ParsedVocabValue, |
| 506 | TargetVocab&: ArgVocab, Dim&: ArgDim)) |
| 507 | return Err; |
| 508 | |
| 509 | if (!(OpcodeDim == TypeDim && TypeDim == ArgDim)) |
| 510 | return createStringError(EC: errc::illegal_byte_sequence, |
| 511 | S: "Vocabulary sections have different dimensions" ); |
| 512 | |
| 513 | return Error::success(); |
| 514 | } |
| 515 | } // anonymous namespace |
| 516 | |
| 517 | /// Generate VocabStorage from vocabulary maps. |
| 518 | VocabStorage Vocabulary::buildVocabStorage(const VocabMap &OpcVocab, |
| 519 | const VocabMap &TypeVocab, |
| 520 | const VocabMap &ArgVocab) { |
| 521 | |
| 522 | // Helper for handling missing entities in the vocabulary. |
| 523 | // Currently, we use a zero vector. In the future, we will throw an error to |
| 524 | // ensure that *all* known entities are present in the vocabulary. |
| 525 | auto handleMissingEntity = [](const std::string &Val) { |
| 526 | LLVM_DEBUG(errs() << Val |
| 527 | << " is not in vocabulary, using zero vector; This " |
| 528 | "would result in an error in future.\n" ); |
| 529 | ++VocabMissCounter; |
| 530 | }; |
| 531 | |
| 532 | unsigned Dim = OpcVocab.begin()->second.size(); |
| 533 | assert(Dim > 0 && "Vocabulary dimension must be greater than zero" ); |
| 534 | |
| 535 | // Handle Opcodes |
| 536 | std::vector<Embedding> NumericOpcodeEmbeddings(Vocabulary::MaxOpcodes, |
| 537 | Embedding(Dim)); |
| 538 | for (unsigned Opcode : seq(Begin: 0u, End: Vocabulary::MaxOpcodes)) { |
| 539 | StringRef VocabKey = Vocabulary::getVocabKeyForOpcode(Opcode: Opcode + 1); |
| 540 | auto It = OpcVocab.find(x: VocabKey.str()); |
| 541 | if (It != OpcVocab.end()) |
| 542 | NumericOpcodeEmbeddings[Opcode] = It->second; |
| 543 | else |
| 544 | handleMissingEntity(VocabKey.str()); |
| 545 | } |
| 546 | |
| 547 | // Handle Types - only canonical types are present in vocabulary |
| 548 | std::vector<Embedding> NumericTypeEmbeddings(Vocabulary::MaxCanonicalTypeIDs, |
| 549 | Embedding(Dim)); |
| 550 | for (unsigned CTypeID : seq(Begin: 0u, End: Vocabulary::MaxCanonicalTypeIDs)) { |
| 551 | StringRef VocabKey = Vocabulary::getVocabKeyForCanonicalTypeID( |
| 552 | CType: static_cast<Vocabulary::CanonicalTypeID>(CTypeID)); |
| 553 | if (auto It = TypeVocab.find(x: VocabKey.str()); It != TypeVocab.end()) { |
| 554 | NumericTypeEmbeddings[CTypeID] = It->second; |
| 555 | continue; |
| 556 | } |
| 557 | handleMissingEntity(VocabKey.str()); |
| 558 | } |
| 559 | |
| 560 | // Handle Arguments/Operands |
| 561 | std::vector<Embedding> NumericArgEmbeddings(Vocabulary::MaxOperandKinds, |
| 562 | Embedding(Dim)); |
| 563 | for (unsigned OpKind : seq(Begin: 0u, End: Vocabulary::MaxOperandKinds)) { |
| 564 | Vocabulary::OperandKind Kind = static_cast<Vocabulary::OperandKind>(OpKind); |
| 565 | StringRef VocabKey = Vocabulary::getVocabKeyForOperandKind(Kind); |
| 566 | auto It = ArgVocab.find(x: VocabKey.str()); |
| 567 | if (It != ArgVocab.end()) { |
| 568 | NumericArgEmbeddings[OpKind] = It->second; |
| 569 | continue; |
| 570 | } |
| 571 | handleMissingEntity(VocabKey.str()); |
| 572 | } |
| 573 | |
| 574 | // Handle Predicates: part of Operands section. We look up predicate keys |
| 575 | // in ArgVocab. |
| 576 | std::vector<Embedding> NumericPredEmbeddings(Vocabulary::MaxPredicateKinds, |
| 577 | Embedding(Dim, 0)); |
| 578 | for (unsigned PK : seq(Begin: 0u, End: Vocabulary::MaxPredicateKinds)) { |
| 579 | StringRef VocabKey = |
| 580 | Vocabulary::getVocabKeyForPredicate(Pred: Vocabulary::getPredicate(Index: PK)); |
| 581 | auto It = ArgVocab.find(x: VocabKey.str()); |
| 582 | if (It != ArgVocab.end()) { |
| 583 | NumericPredEmbeddings[PK] = It->second; |
| 584 | continue; |
| 585 | } |
| 586 | handleMissingEntity(VocabKey.str()); |
| 587 | } |
| 588 | |
| 589 | // Create section-based storage instead of flat vocabulary |
| 590 | // Order must match Vocabulary::Section enum |
| 591 | std::vector<std::vector<Embedding>> Sections(4); |
| 592 | Sections[static_cast<unsigned>(Section::Opcodes)] = |
| 593 | std::move(NumericOpcodeEmbeddings); // Section::Opcodes |
| 594 | Sections[static_cast<unsigned>(Section::CanonicalTypes)] = |
| 595 | std::move(NumericTypeEmbeddings); // Section::CanonicalTypes |
| 596 | Sections[static_cast<unsigned>(Section::Operands)] = |
| 597 | std::move(NumericArgEmbeddings); // Section::Operands |
| 598 | Sections[static_cast<unsigned>(Section::Predicates)] = |
| 599 | std::move(NumericPredEmbeddings); // Section::Predicates |
| 600 | |
| 601 | // Create VocabStorage from organized sections |
| 602 | return VocabStorage(std::move(Sections)); |
| 603 | } |
| 604 | |
| 605 | // ==----------------------------------------------------------------------===// |
| 606 | // Vocabulary |
| 607 | //===----------------------------------------------------------------------===// |
| 608 | |
| 609 | Expected<Vocabulary> Vocabulary::fromFile(StringRef VocabFilePath, |
| 610 | float OpcWeight, float TypeWeight, |
| 611 | float ArgWeight) { |
| 612 | VocabMap OpcVocab, TypeVocab, ArgVocab; |
| 613 | if (auto Err = |
| 614 | readVocabularyFromFile(VocabFilePath, OpcVocab, TypeVocab, ArgVocab)) |
| 615 | return std::move(Err); |
| 616 | |
| 617 | // Scale the vocabulary sections based on the provided weights |
| 618 | auto scaleVocabSection = [](VocabMap &Vocab, float Weight) { |
| 619 | for (auto &Entry : Vocab) |
| 620 | Entry.second *= Weight; |
| 621 | }; |
| 622 | scaleVocabSection(OpcVocab, OpcWeight); |
| 623 | scaleVocabSection(TypeVocab, TypeWeight); |
| 624 | scaleVocabSection(ArgVocab, ArgWeight); |
| 625 | |
| 626 | // Generate the numeric lookup vocabulary |
| 627 | return Vocabulary(buildVocabStorage(OpcVocab, TypeVocab, ArgVocab)); |
| 628 | } |
| 629 | |
| 630 | // ==----------------------------------------------------------------------===// |
| 631 | // IR2VecVocabAnalysis |
| 632 | //===----------------------------------------------------------------------===// |
| 633 | |
| 634 | void IR2VecVocabAnalysis::emitError(Error Err, LLVMContext &Ctx) { |
| 635 | handleAllErrors(E: std::move(Err), Handlers: [&](const ErrorInfoBase &EI) { |
| 636 | Ctx.emitError(ErrorStr: "Error reading vocabulary: " + EI.message()); |
| 637 | }); |
| 638 | } |
| 639 | |
| 640 | IR2VecVocabAnalysis::Result |
| 641 | IR2VecVocabAnalysis::run(Module &M, ModuleAnalysisManager &AM) { |
| 642 | auto Ctx = &M.getContext(); |
| 643 | // If vocabulary is already populated by the constructor, use it. |
| 644 | if (Vocab.has_value()) |
| 645 | return Vocabulary(std::move(Vocab.value())); |
| 646 | |
| 647 | // Otherwise, try to read from the vocabulary file specified via CLI. |
| 648 | if (VocabFile.empty()) { |
| 649 | // FIXME: Use default vocabulary |
| 650 | Ctx->emitError(ErrorStr: "IR2Vec vocabulary file path not specified; You may need to " |
| 651 | "set it using --ir2vec-vocab-path" ); |
| 652 | return Vocabulary(); // Return invalid result |
| 653 | } |
| 654 | |
| 655 | // Use the static factory method to load the vocabulary. |
| 656 | auto VocabOrErr = |
| 657 | Vocabulary::fromFile(VocabFilePath: VocabFile, OpcWeight, TypeWeight, ArgWeight); |
| 658 | if (!VocabOrErr) { |
| 659 | emitError(Err: VocabOrErr.takeError(), Ctx&: *Ctx); |
| 660 | return Vocabulary(); |
| 661 | } |
| 662 | |
| 663 | return std::move(*VocabOrErr); |
| 664 | } |
| 665 | |
| 666 | // ==----------------------------------------------------------------------===// |
| 667 | // Printer Passes |
| 668 | //===----------------------------------------------------------------------===// |
| 669 | |
| 670 | PreservedAnalyses IR2VecPrinterPass::run(Module &M, |
| 671 | ModuleAnalysisManager &MAM) { |
| 672 | auto &Vocabulary = MAM.getResult<IR2VecVocabAnalysis>(IR&: M); |
| 673 | assert(Vocabulary.isValid() && "IR2Vec Vocabulary is invalid" ); |
| 674 | |
| 675 | for (Function &F : M) { |
| 676 | auto Emb = Embedder::create(Mode: IR2VecEmbeddingKind, F, Vocab: Vocabulary); |
| 677 | if (!Emb) { |
| 678 | OS << "Error creating IR2Vec embeddings \n" ; |
| 679 | continue; |
| 680 | } |
| 681 | |
| 682 | OS << "IR2Vec embeddings for function " << F.getName() << ":\n" ; |
| 683 | OS << "Function vector: " ; |
| 684 | Emb->getFunctionVector().print(OS); |
| 685 | |
| 686 | OS << "Basic block vectors:\n" ; |
| 687 | for (const BasicBlock &BB : F) { |
| 688 | OS << "Basic block: " << BB.getName() << ":\n" ; |
| 689 | Emb->getBBVector(BB).print(OS); |
| 690 | } |
| 691 | |
| 692 | OS << "Instruction vectors:\n" ; |
| 693 | for (const BasicBlock &BB : F) { |
| 694 | for (const Instruction &I : BB) { |
| 695 | OS << "Instruction: " ; |
| 696 | I.print(O&: OS); |
| 697 | Emb->getInstVector(I).print(OS); |
| 698 | } |
| 699 | } |
| 700 | } |
| 701 | return PreservedAnalyses::all(); |
| 702 | } |
| 703 | |
| 704 | PreservedAnalyses IR2VecVocabPrinterPass::run(Module &M, |
| 705 | ModuleAnalysisManager &MAM) { |
| 706 | auto &IR2VecVocabulary = MAM.getResult<IR2VecVocabAnalysis>(IR&: M); |
| 707 | assert(IR2VecVocabulary.isValid() && "IR2Vec Vocabulary is invalid" ); |
| 708 | |
| 709 | // Print each entry |
| 710 | unsigned Pos = 0; |
| 711 | for (const auto &Entry : IR2VecVocabulary) { |
| 712 | OS << "Key: " << IR2VecVocabulary.getStringKey(Pos: Pos++) << ": " ; |
| 713 | Entry.print(OS); |
| 714 | } |
| 715 | return PreservedAnalyses::all(); |
| 716 | } |
| 717 | |