GVN.cpp revision a723d1e48f4a261512c28845c53eda569fa5218c
1//===- GVN.cpp - Eliminate redundant values and loads ------------===// 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// This pass performs global value numbering to eliminate fully redundant 11// instructions. It also performs simple dead load elimination. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "gvn" 16#include "llvm/Transforms/Scalar.h" 17#include "llvm/BasicBlock.h" 18#include "llvm/Constants.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Function.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/Instructions.h" 23#include "llvm/ParameterAttributes.h" 24#include "llvm/Value.h" 25#include "llvm/ADT/DenseMap.h" 26#include "llvm/ADT/DepthFirstIterator.h" 27#include "llvm/ADT/SmallPtrSet.h" 28#include "llvm/ADT/SmallVector.h" 29#include "llvm/ADT/SparseBitVector.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Analysis/AliasAnalysis.h" 33#include "llvm/Analysis/MemoryDependenceAnalysis.h" 34#include "llvm/Support/CFG.h" 35#include "llvm/Support/CommandLine.h" 36#include "llvm/Support/Compiler.h" 37#include "llvm/Support/Debug.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Target/TargetData.h" 40#include <list> 41using namespace llvm; 42 43STATISTIC(NumGVNInstr, "Number of instructions deleted"); 44STATISTIC(NumGVNLoad, "Number of loads deleted"); 45 46//===----------------------------------------------------------------------===// 47// ValueTable Class 48//===----------------------------------------------------------------------===// 49 50/// This class holds the mapping between values and value numbers. It is used 51/// as an efficient mechanism to determine the expression-wise equivalence of 52/// two values. 53namespace { 54 struct VISIBILITY_HIDDEN Expression { 55 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM, 56 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 57 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 58 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 59 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 60 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 61 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 62 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 63 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 64 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, EMPTY, 65 TOMBSTONE }; 66 67 ExpressionOpcode opcode; 68 const Type* type; 69 uint32_t firstVN; 70 uint32_t secondVN; 71 uint32_t thirdVN; 72 SmallVector<uint32_t, 4> varargs; 73 Value* function; 74 75 Expression() { } 76 Expression(ExpressionOpcode o) : opcode(o) { } 77 78 bool operator==(const Expression &other) const { 79 if (opcode != other.opcode) 80 return false; 81 else if (opcode == EMPTY || opcode == TOMBSTONE) 82 return true; 83 else if (type != other.type) 84 return false; 85 else if (function != other.function) 86 return false; 87 else if (firstVN != other.firstVN) 88 return false; 89 else if (secondVN != other.secondVN) 90 return false; 91 else if (thirdVN != other.thirdVN) 92 return false; 93 else { 94 if (varargs.size() != other.varargs.size()) 95 return false; 96 97 for (size_t i = 0; i < varargs.size(); ++i) 98 if (varargs[i] != other.varargs[i]) 99 return false; 100 101 return true; 102 } 103 } 104 105 bool operator!=(const Expression &other) const { 106 if (opcode != other.opcode) 107 return true; 108 else if (opcode == EMPTY || opcode == TOMBSTONE) 109 return false; 110 else if (type != other.type) 111 return true; 112 else if (function != other.function) 113 return true; 114 else if (firstVN != other.firstVN) 115 return true; 116 else if (secondVN != other.secondVN) 117 return true; 118 else if (thirdVN != other.thirdVN) 119 return true; 120 else { 121 if (varargs.size() != other.varargs.size()) 122 return true; 123 124 for (size_t i = 0; i < varargs.size(); ++i) 125 if (varargs[i] != other.varargs[i]) 126 return true; 127 128 return false; 129 } 130 } 131 }; 132 133 class VISIBILITY_HIDDEN ValueTable { 134 private: 135 DenseMap<Value*, uint32_t> valueNumbering; 136 DenseMap<Expression, uint32_t> expressionNumbering; 137 AliasAnalysis* AA; 138 139 uint32_t nextValueNumber; 140 141 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 142 Expression::ExpressionOpcode getOpcode(CmpInst* C); 143 Expression::ExpressionOpcode getOpcode(CastInst* C); 144 Expression create_expression(BinaryOperator* BO); 145 Expression create_expression(CmpInst* C); 146 Expression create_expression(ShuffleVectorInst* V); 147 Expression create_expression(ExtractElementInst* C); 148 Expression create_expression(InsertElementInst* V); 149 Expression create_expression(SelectInst* V); 150 Expression create_expression(CastInst* C); 151 Expression create_expression(GetElementPtrInst* G); 152 Expression create_expression(CallInst* C); 153 public: 154 ValueTable() : nextValueNumber(1) { } 155 uint32_t lookup_or_add(Value* V); 156 uint32_t lookup(Value* V) const; 157 void add(Value* V, uint32_t num); 158 void clear(); 159 void erase(Value* v); 160 unsigned size(); 161 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 162 uint32_t hash_operand(Value* v); 163 }; 164} 165 166namespace llvm { 167template <> struct DenseMapInfo<Expression> { 168 static inline Expression getEmptyKey() { 169 return Expression(Expression::EMPTY); 170 } 171 172 static inline Expression getTombstoneKey() { 173 return Expression(Expression::TOMBSTONE); 174 } 175 176 static unsigned getHashValue(const Expression e) { 177 unsigned hash = e.opcode; 178 179 hash = e.firstVN + hash * 37; 180 hash = e.secondVN + hash * 37; 181 hash = e.thirdVN + hash * 37; 182 183 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 184 (unsigned)((uintptr_t)e.type >> 9)) + 185 hash * 37; 186 187 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 188 E = e.varargs.end(); I != E; ++I) 189 hash = *I + hash * 37; 190 191 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 192 (unsigned)((uintptr_t)e.function >> 9)) + 193 hash * 37; 194 195 return hash; 196 } 197 static bool isEqual(const Expression &LHS, const Expression &RHS) { 198 return LHS == RHS; 199 } 200 static bool isPod() { return true; } 201}; 202} 203 204//===----------------------------------------------------------------------===// 205// ValueTable Internal Functions 206//===----------------------------------------------------------------------===// 207Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 208 switch(BO->getOpcode()) { 209 default: // THIS SHOULD NEVER HAPPEN 210 assert(0 && "Binary operator with unknown opcode?"); 211 case Instruction::Add: return Expression::ADD; 212 case Instruction::Sub: return Expression::SUB; 213 case Instruction::Mul: return Expression::MUL; 214 case Instruction::UDiv: return Expression::UDIV; 215 case Instruction::SDiv: return Expression::SDIV; 216 case Instruction::FDiv: return Expression::FDIV; 217 case Instruction::URem: return Expression::UREM; 218 case Instruction::SRem: return Expression::SREM; 219 case Instruction::FRem: return Expression::FREM; 220 case Instruction::Shl: return Expression::SHL; 221 case Instruction::LShr: return Expression::LSHR; 222 case Instruction::AShr: return Expression::ASHR; 223 case Instruction::And: return Expression::AND; 224 case Instruction::Or: return Expression::OR; 225 case Instruction::Xor: return Expression::XOR; 226 } 227} 228 229Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 230 if (isa<ICmpInst>(C)) { 231 switch (C->getPredicate()) { 232 default: // THIS SHOULD NEVER HAPPEN 233 assert(0 && "Comparison with unknown predicate?"); 234 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 235 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 236 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 237 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 238 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 239 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 240 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 241 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 242 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 243 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 244 } 245 } 246 assert(isa<FCmpInst>(C) && "Unknown compare"); 247 switch (C->getPredicate()) { 248 default: // THIS SHOULD NEVER HAPPEN 249 assert(0 && "Comparison with unknown predicate?"); 250 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 251 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 252 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 253 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 254 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 255 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 256 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 257 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 258 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 259 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 260 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 261 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 262 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 263 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 264 } 265} 266 267Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 268 switch(C->getOpcode()) { 269 default: // THIS SHOULD NEVER HAPPEN 270 assert(0 && "Cast operator with unknown opcode?"); 271 case Instruction::Trunc: return Expression::TRUNC; 272 case Instruction::ZExt: return Expression::ZEXT; 273 case Instruction::SExt: return Expression::SEXT; 274 case Instruction::FPToUI: return Expression::FPTOUI; 275 case Instruction::FPToSI: return Expression::FPTOSI; 276 case Instruction::UIToFP: return Expression::UITOFP; 277 case Instruction::SIToFP: return Expression::SITOFP; 278 case Instruction::FPTrunc: return Expression::FPTRUNC; 279 case Instruction::FPExt: return Expression::FPEXT; 280 case Instruction::PtrToInt: return Expression::PTRTOINT; 281 case Instruction::IntToPtr: return Expression::INTTOPTR; 282 case Instruction::BitCast: return Expression::BITCAST; 283 } 284} 285 286uint32_t ValueTable::hash_operand(Value* v) { 287 if (CallInst* CI = dyn_cast<CallInst>(v)) 288 if (!AA->doesNotAccessMemory(CI)) 289 return nextValueNumber++; 290 291 return lookup_or_add(v); 292} 293 294Expression ValueTable::create_expression(CallInst* C) { 295 Expression e; 296 297 e.type = C->getType(); 298 e.firstVN = 0; 299 e.secondVN = 0; 300 e.thirdVN = 0; 301 e.function = C->getCalledFunction(); 302 e.opcode = Expression::CALL; 303 304 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 305 I != E; ++I) 306 e.varargs.push_back(hash_operand(*I)); 307 308 return e; 309} 310 311Expression ValueTable::create_expression(BinaryOperator* BO) { 312 Expression e; 313 314 e.firstVN = hash_operand(BO->getOperand(0)); 315 e.secondVN = hash_operand(BO->getOperand(1)); 316 e.thirdVN = 0; 317 e.function = 0; 318 e.type = BO->getType(); 319 e.opcode = getOpcode(BO); 320 321 return e; 322} 323 324Expression ValueTable::create_expression(CmpInst* C) { 325 Expression e; 326 327 e.firstVN = hash_operand(C->getOperand(0)); 328 e.secondVN = hash_operand(C->getOperand(1)); 329 e.thirdVN = 0; 330 e.function = 0; 331 e.type = C->getType(); 332 e.opcode = getOpcode(C); 333 334 return e; 335} 336 337Expression ValueTable::create_expression(CastInst* C) { 338 Expression e; 339 340 e.firstVN = hash_operand(C->getOperand(0)); 341 e.secondVN = 0; 342 e.thirdVN = 0; 343 e.function = 0; 344 e.type = C->getType(); 345 e.opcode = getOpcode(C); 346 347 return e; 348} 349 350Expression ValueTable::create_expression(ShuffleVectorInst* S) { 351 Expression e; 352 353 e.firstVN = hash_operand(S->getOperand(0)); 354 e.secondVN = hash_operand(S->getOperand(1)); 355 e.thirdVN = hash_operand(S->getOperand(2)); 356 e.function = 0; 357 e.type = S->getType(); 358 e.opcode = Expression::SHUFFLE; 359 360 return e; 361} 362 363Expression ValueTable::create_expression(ExtractElementInst* E) { 364 Expression e; 365 366 e.firstVN = hash_operand(E->getOperand(0)); 367 e.secondVN = hash_operand(E->getOperand(1)); 368 e.thirdVN = 0; 369 e.function = 0; 370 e.type = E->getType(); 371 e.opcode = Expression::EXTRACT; 372 373 return e; 374} 375 376Expression ValueTable::create_expression(InsertElementInst* I) { 377 Expression e; 378 379 e.firstVN = hash_operand(I->getOperand(0)); 380 e.secondVN = hash_operand(I->getOperand(1)); 381 e.thirdVN = hash_operand(I->getOperand(2)); 382 e.function = 0; 383 e.type = I->getType(); 384 e.opcode = Expression::INSERT; 385 386 return e; 387} 388 389Expression ValueTable::create_expression(SelectInst* I) { 390 Expression e; 391 392 e.firstVN = hash_operand(I->getCondition()); 393 e.secondVN = hash_operand(I->getTrueValue()); 394 e.thirdVN = hash_operand(I->getFalseValue()); 395 e.function = 0; 396 e.type = I->getType(); 397 e.opcode = Expression::SELECT; 398 399 return e; 400} 401 402Expression ValueTable::create_expression(GetElementPtrInst* G) { 403 Expression e; 404 405 e.firstVN = hash_operand(G->getPointerOperand()); 406 e.secondVN = 0; 407 e.thirdVN = 0; 408 e.function = 0; 409 e.type = G->getType(); 410 e.opcode = Expression::GEP; 411 412 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 413 I != E; ++I) 414 e.varargs.push_back(hash_operand(*I)); 415 416 return e; 417} 418 419//===----------------------------------------------------------------------===// 420// ValueTable External Functions 421//===----------------------------------------------------------------------===// 422 423/// lookup_or_add - Returns the value number for the specified value, assigning 424/// it a new number if it did not have one before. 425uint32_t ValueTable::lookup_or_add(Value* V) { 426 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 427 if (VI != valueNumbering.end()) 428 return VI->second; 429 430 if (CallInst* C = dyn_cast<CallInst>(V)) { 431 if (AA->onlyReadsMemory(C)) { // includes doesNotAccessMemory 432 Expression e = create_expression(C); 433 434 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 435 if (EI != expressionNumbering.end()) { 436 valueNumbering.insert(std::make_pair(V, EI->second)); 437 return EI->second; 438 } else { 439 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 440 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 441 442 return nextValueNumber++; 443 } 444 } else { 445 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 446 return nextValueNumber++; 447 } 448 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 449 Expression e = create_expression(BO); 450 451 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 452 if (EI != expressionNumbering.end()) { 453 valueNumbering.insert(std::make_pair(V, EI->second)); 454 return EI->second; 455 } else { 456 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 457 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 458 459 return nextValueNumber++; 460 } 461 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 462 Expression e = create_expression(C); 463 464 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 465 if (EI != expressionNumbering.end()) { 466 valueNumbering.insert(std::make_pair(V, EI->second)); 467 return EI->second; 468 } else { 469 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 470 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 471 472 return nextValueNumber++; 473 } 474 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 475 Expression e = create_expression(U); 476 477 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 478 if (EI != expressionNumbering.end()) { 479 valueNumbering.insert(std::make_pair(V, EI->second)); 480 return EI->second; 481 } else { 482 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 483 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 484 485 return nextValueNumber++; 486 } 487 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 488 Expression e = create_expression(U); 489 490 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 491 if (EI != expressionNumbering.end()) { 492 valueNumbering.insert(std::make_pair(V, EI->second)); 493 return EI->second; 494 } else { 495 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 496 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 497 498 return nextValueNumber++; 499 } 500 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 501 Expression e = create_expression(U); 502 503 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 504 if (EI != expressionNumbering.end()) { 505 valueNumbering.insert(std::make_pair(V, EI->second)); 506 return EI->second; 507 } else { 508 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 509 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 510 511 return nextValueNumber++; 512 } 513 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 514 Expression e = create_expression(U); 515 516 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 517 if (EI != expressionNumbering.end()) { 518 valueNumbering.insert(std::make_pair(V, EI->second)); 519 return EI->second; 520 } else { 521 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 522 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 523 524 return nextValueNumber++; 525 } 526 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 527 Expression e = create_expression(U); 528 529 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 530 if (EI != expressionNumbering.end()) { 531 valueNumbering.insert(std::make_pair(V, EI->second)); 532 return EI->second; 533 } else { 534 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 535 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 536 537 return nextValueNumber++; 538 } 539 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 540 Expression e = create_expression(U); 541 542 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 543 if (EI != expressionNumbering.end()) { 544 valueNumbering.insert(std::make_pair(V, EI->second)); 545 return EI->second; 546 } else { 547 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 548 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 549 550 return nextValueNumber++; 551 } 552 } else { 553 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 554 return nextValueNumber++; 555 } 556} 557 558/// lookup - Returns the value number of the specified value. Fails if 559/// the value has not yet been numbered. 560uint32_t ValueTable::lookup(Value* V) const { 561 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 562 assert(VI != valueNumbering.end() && "Value not numbered?"); 563 return VI->second; 564} 565 566/// clear - Remove all entries from the ValueTable 567void ValueTable::clear() { 568 valueNumbering.clear(); 569 expressionNumbering.clear(); 570 nextValueNumber = 1; 571} 572 573/// erase - Remove a value from the value numbering 574void ValueTable::erase(Value* V) { 575 valueNumbering.erase(V); 576} 577 578//===----------------------------------------------------------------------===// 579// ValueNumberedSet Class 580//===----------------------------------------------------------------------===// 581namespace { 582class VISIBILITY_HIDDEN ValueNumberedSet { 583 private: 584 SmallPtrSet<Value*, 8> contents; 585 SparseBitVector<64> numbers; 586 public: 587 ValueNumberedSet() { } 588 ValueNumberedSet(const ValueNumberedSet& other) { 589 numbers = other.numbers; 590 contents = other.contents; 591 } 592 593 typedef SmallPtrSet<Value*, 8>::iterator iterator; 594 595 iterator begin() { return contents.begin(); } 596 iterator end() { return contents.end(); } 597 598 bool insert(Value* v) { return contents.insert(v); } 599 void insert(iterator I, iterator E) { contents.insert(I, E); } 600 void erase(Value* v) { contents.erase(v); } 601 unsigned count(Value* v) { return contents.count(v); } 602 size_t size() { return contents.size(); } 603 604 void set(unsigned i) { 605 numbers.set(i); 606 } 607 608 void operator=(const ValueNumberedSet& other) { 609 contents = other.contents; 610 numbers = other.numbers; 611 } 612 613 void reset(unsigned i) { 614 numbers.reset(i); 615 } 616 617 bool test(unsigned i) { 618 return numbers.test(i); 619 } 620}; 621} 622 623//===----------------------------------------------------------------------===// 624// GVN Pass 625//===----------------------------------------------------------------------===// 626 627namespace { 628 629 class VISIBILITY_HIDDEN GVN : public FunctionPass { 630 bool runOnFunction(Function &F); 631 public: 632 static char ID; // Pass identification, replacement for typeid 633 GVN() : FunctionPass((intptr_t)&ID) { } 634 635 private: 636 ValueTable VN; 637 638 DenseMap<BasicBlock*, ValueNumberedSet> availableOut; 639 640 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 641 PhiMapType phiMap; 642 643 644 // This transformation requires dominator postdominator info 645 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 646 AU.setPreservesCFG(); 647 AU.addRequired<DominatorTree>(); 648 AU.addRequired<MemoryDependenceAnalysis>(); 649 AU.addRequired<AliasAnalysis>(); 650 AU.addRequired<TargetData>(); 651 AU.addPreserved<AliasAnalysis>(); 652 AU.addPreserved<MemoryDependenceAnalysis>(); 653 AU.addPreserved<TargetData>(); 654 } 655 656 // Helper fuctions 657 // FIXME: eliminate or document these better 658 Value* find_leader(ValueNumberedSet& vals, uint32_t v) ; 659 void val_insert(ValueNumberedSet& s, Value* v); 660 bool processLoad(LoadInst* L, 661 DenseMap<Value*, LoadInst*> &lastLoad, 662 SmallVectorImpl<Instruction*> &toErase); 663 bool processInstruction(Instruction* I, 664 ValueNumberedSet& currAvail, 665 DenseMap<Value*, LoadInst*>& lastSeenLoad, 666 SmallVectorImpl<Instruction*> &toErase); 667 bool processNonLocalLoad(LoadInst* L, 668 SmallVectorImpl<Instruction*> &toErase); 669 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig, 670 DenseMap<BasicBlock*, Value*> &Phis, 671 bool top_level = false); 672 void dump(DenseMap<BasicBlock*, Value*>& d); 673 bool iterateOnFunction(Function &F); 674 Value* CollapsePhi(PHINode* p); 675 bool isSafeReplacement(PHINode* p, Instruction* inst); 676 }; 677 678 char GVN::ID = 0; 679} 680 681// createGVNPass - The public interface to this file... 682FunctionPass *llvm::createGVNPass() { return new GVN(); } 683 684static RegisterPass<GVN> X("gvn", 685 "Global Value Numbering"); 686 687/// find_leader - Given a set and a value number, return the first 688/// element of the set with that value number, or 0 if no such element 689/// is present 690Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) { 691 if (!vals.test(v)) 692 return 0; 693 694 for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end(); 695 I != E; ++I) 696 if (v == VN.lookup(*I)) 697 return *I; 698 699 assert(0 && "No leader found, but present bit is set?"); 700 return 0; 701} 702 703/// val_insert - Insert a value into a set only if there is not a value 704/// with the same value number already in the set 705void GVN::val_insert(ValueNumberedSet& s, Value* v) { 706 uint32_t num = VN.lookup(v); 707 if (!s.test(num)) 708 s.insert(v); 709} 710 711void GVN::dump(DenseMap<BasicBlock*, Value*>& d) { 712 printf("{\n"); 713 for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(), 714 E = d.end(); I != E; ++I) { 715 if (I->second == MemoryDependenceAnalysis::None) 716 printf("None\n"); 717 else 718 I->second->dump(); 719 } 720 printf("}\n"); 721} 722 723Value* GVN::CollapsePhi(PHINode* p) { 724 DominatorTree &DT = getAnalysis<DominatorTree>(); 725 Value* constVal = p->hasConstantValue(); 726 727 if (!constVal) return 0; 728 729 Instruction* inst = dyn_cast<Instruction>(constVal); 730 if (!inst) 731 return constVal; 732 733 if (DT.dominates(inst, p)) 734 if (isSafeReplacement(p, inst)) 735 return inst; 736 return 0; 737} 738 739bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 740 if (!isa<PHINode>(inst)) 741 return true; 742 743 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 744 UI != E; ++UI) 745 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 746 if (use_phi->getParent() == inst->getParent()) 747 return false; 748 749 return true; 750} 751 752/// GetValueForBlock - Get the value to use within the specified basic block. 753/// available values are in Phis. 754Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig, 755 DenseMap<BasicBlock*, Value*> &Phis, 756 bool top_level) { 757 758 // If we have already computed this value, return the previously computed val. 759 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 760 if (V != Phis.end() && !top_level) return V->second; 761 762 BasicBlock* singlePred = BB->getSinglePredecessor(); 763 if (singlePred) { 764 Value *ret = GetValueForBlock(singlePred, orig, Phis); 765 Phis[BB] = ret; 766 return ret; 767 } 768 769 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 770 // now, then get values to fill in the incoming values for the PHI. 771 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 772 BB->begin()); 773 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB))); 774 775 if (Phis.count(BB) == 0) 776 Phis.insert(std::make_pair(BB, PN)); 777 778 // Fill in the incoming values for the block. 779 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 780 Value* val = GetValueForBlock(*PI, orig, Phis); 781 PN->addIncoming(val, *PI); 782 } 783 784 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 785 AA.copyValue(orig, PN); 786 787 // Attempt to collapse PHI nodes that are trivially redundant 788 Value* v = CollapsePhi(PN); 789 if (!v) { 790 // Cache our phi construction results 791 phiMap[orig->getPointerOperand()].insert(PN); 792 return PN; 793 } 794 795 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 796 797 MD.removeInstruction(PN); 798 PN->replaceAllUsesWith(v); 799 800 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 801 E = Phis.end(); I != E; ++I) 802 if (I->second == PN) 803 I->second = v; 804 805 PN->eraseFromParent(); 806 807 Phis[BB] = v; 808 return v; 809} 810 811/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 812/// non-local by performing PHI construction. 813bool GVN::processNonLocalLoad(LoadInst* L, 814 SmallVectorImpl<Instruction*> &toErase) { 815 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 816 817 // Find the non-local dependencies of the load 818 DenseMap<BasicBlock*, Value*> deps; 819 MD.getNonLocalDependency(L, deps); 820 821 DenseMap<BasicBlock*, Value*> repl; 822 823 // Filter out useless results (non-locals, etc) 824 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end(); 825 I != E; ++I) { 826 if (I->second == MemoryDependenceAnalysis::None) 827 return false; 828 829 if (I->second == MemoryDependenceAnalysis::NonLocal) 830 continue; 831 832 if (StoreInst* S = dyn_cast<StoreInst>(I->second)) { 833 if (S->getPointerOperand() != L->getPointerOperand()) 834 return false; 835 repl[I->first] = S->getOperand(0); 836 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) { 837 if (LD->getPointerOperand() != L->getPointerOperand()) 838 return false; 839 repl[I->first] = LD; 840 } else { 841 return false; 842 } 843 } 844 845 // Use cached PHI construction information from previous runs 846 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()]; 847 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 848 I != E; ++I) { 849 if ((*I)->getParent() == L->getParent()) { 850 MD.removeInstruction(L); 851 L->replaceAllUsesWith(*I); 852 toErase.push_back(L); 853 NumGVNLoad++; 854 return true; 855 } 856 857 repl.insert(std::make_pair((*I)->getParent(), *I)); 858 } 859 860 // Perform PHI construction 861 SmallPtrSet<BasicBlock*, 4> visited; 862 Value* v = GetValueForBlock(L->getParent(), L, repl, true); 863 864 MD.removeInstruction(L); 865 L->replaceAllUsesWith(v); 866 toErase.push_back(L); 867 NumGVNLoad++; 868 869 return true; 870} 871 872/// processLoad - Attempt to eliminate a load, first by eliminating it 873/// locally, and then attempting non-local elimination if that fails. 874bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad, 875 SmallVectorImpl<Instruction*> &toErase) { 876 if (L->isVolatile()) { 877 lastLoad[L->getPointerOperand()] = L; 878 return false; 879 } 880 881 Value* pointer = L->getPointerOperand(); 882 LoadInst*& last = lastLoad[pointer]; 883 884 // ... to a pointer that has been loaded from before... 885 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 886 bool removedNonLocal = false; 887 Instruction* dep = MD.getDependency(L); 888 if (dep == MemoryDependenceAnalysis::NonLocal && 889 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) { 890 removedNonLocal = processNonLocalLoad(L, toErase); 891 892 if (!removedNonLocal) 893 last = L; 894 895 return removedNonLocal; 896 } 897 898 899 bool deletedLoad = false; 900 901 // Walk up the dependency chain until we either find 902 // a dependency we can use, or we can't walk any further 903 while (dep != MemoryDependenceAnalysis::None && 904 dep != MemoryDependenceAnalysis::NonLocal && 905 (isa<LoadInst>(dep) || isa<StoreInst>(dep))) { 906 // ... that depends on a store ... 907 if (StoreInst* S = dyn_cast<StoreInst>(dep)) { 908 if (S->getPointerOperand() == pointer) { 909 // Remove it! 910 MD.removeInstruction(L); 911 912 L->replaceAllUsesWith(S->getOperand(0)); 913 toErase.push_back(L); 914 deletedLoad = true; 915 NumGVNLoad++; 916 } 917 918 // Whether we removed it or not, we can't 919 // go any further 920 break; 921 } else if (!last) { 922 // If we don't depend on a store, and we haven't 923 // been loaded before, bail. 924 break; 925 } else if (dep == last) { 926 // Remove it! 927 MD.removeInstruction(L); 928 929 L->replaceAllUsesWith(last); 930 toErase.push_back(L); 931 deletedLoad = true; 932 NumGVNLoad++; 933 934 break; 935 } else { 936 dep = MD.getDependency(L, dep); 937 } 938 } 939 940 if (dep != MemoryDependenceAnalysis::None && 941 dep != MemoryDependenceAnalysis::NonLocal && 942 isa<AllocationInst>(dep)) { 943 // Check that this load is actually from the 944 // allocation we found 945 Value* v = L->getOperand(0); 946 while (true) { 947 if (BitCastInst *BC = dyn_cast<BitCastInst>(v)) 948 v = BC->getOperand(0); 949 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v)) 950 v = GEP->getOperand(0); 951 else 952 break; 953 } 954 if (v == dep) { 955 // If this load depends directly on an allocation, there isn't 956 // anything stored there; therefore, we can optimize this load 957 // to undef. 958 MD.removeInstruction(L); 959 960 L->replaceAllUsesWith(UndefValue::get(L->getType())); 961 toErase.push_back(L); 962 deletedLoad = true; 963 NumGVNLoad++; 964 } 965 } 966 967 if (!deletedLoad) 968 last = L; 969 970 return deletedLoad; 971} 972 973/// processInstruction - When calculating availability, handle an instruction 974/// by inserting it into the appropriate sets 975bool GVN::processInstruction(Instruction *I, ValueNumberedSet &currAvail, 976 DenseMap<Value*, LoadInst*> &lastSeenLoad, 977 SmallVectorImpl<Instruction*> &toErase) { 978 if (LoadInst* L = dyn_cast<LoadInst>(I)) 979 return processLoad(L, lastSeenLoad, toErase); 980 981 // Allocations are always uniquely numbered, so we can save time and memory 982 // by fast failing them. 983 if (isa<AllocationInst>(I)) 984 return false; 985 986 unsigned num = VN.lookup_or_add(I); 987 988 // Collapse PHI nodes 989 if (PHINode* p = dyn_cast<PHINode>(I)) { 990 Value* constVal = CollapsePhi(p); 991 992 if (constVal) { 993 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 994 PI != PE; ++PI) 995 if (PI->second.count(p)) 996 PI->second.erase(p); 997 998 p->replaceAllUsesWith(constVal); 999 toErase.push_back(p); 1000 } 1001 // Perform value-number based elimination 1002 } else if (currAvail.test(num)) { 1003 Value* repl = find_leader(currAvail, num); 1004 1005 if (CallInst* CI = dyn_cast<CallInst>(I)) { 1006 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 1007 if (!AA.doesNotAccessMemory(CI)) { 1008 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 1009 if (cast<Instruction>(repl)->getParent() != CI->getParent() || 1010 MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) { 1011 // There must be an intervening may-alias store, so nothing from 1012 // this point on will be able to be replaced with the preceding call 1013 currAvail.erase(repl); 1014 currAvail.insert(I); 1015 1016 return false; 1017 } 1018 } 1019 } 1020 1021 // Remove it! 1022 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 1023 MD.removeInstruction(I); 1024 1025 VN.erase(I); 1026 I->replaceAllUsesWith(repl); 1027 toErase.push_back(I); 1028 return true; 1029 } else if (!I->isTerminator()) { 1030 currAvail.set(num); 1031 currAvail.insert(I); 1032 } 1033 1034 return false; 1035} 1036 1037// GVN::runOnFunction - This is the main transformation entry point for a 1038// function. 1039// 1040bool GVN::runOnFunction(Function& F) { 1041 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1042 1043 bool changed = false; 1044 bool shouldContinue = true; 1045 1046 while (shouldContinue) { 1047 shouldContinue = iterateOnFunction(F); 1048 changed |= shouldContinue; 1049 } 1050 1051 return changed; 1052} 1053 1054 1055// GVN::iterateOnFunction - Executes one iteration of GVN 1056bool GVN::iterateOnFunction(Function &F) { 1057 // Clean out global sets from any previous functions 1058 VN.clear(); 1059 availableOut.clear(); 1060 phiMap.clear(); 1061 1062 bool changed_function = false; 1063 1064 DominatorTree &DT = getAnalysis<DominatorTree>(); 1065 1066 SmallVector<Instruction*, 8> toErase; 1067 DenseMap<Value*, LoadInst*> lastSeenLoad; 1068 DenseMap<DomTreeNode*, size_t> numChildrenVisited; 1069 1070 // Top-down walk of the dominator tree 1071 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()), 1072 E = df_end(DT.getRootNode()); DI != E; ++DI) { 1073 1074 // Get the set to update for this block 1075 ValueNumberedSet& currAvail = availableOut[DI->getBlock()]; 1076 lastSeenLoad.clear(); 1077 1078 BasicBlock* BB = DI->getBlock(); 1079 1080 // A block inherits AVAIL_OUT from its dominator 1081 if (DI->getIDom() != 0) { 1082 currAvail = availableOut[DI->getIDom()->getBlock()]; 1083 1084 numChildrenVisited[DI->getIDom()]++; 1085 1086 if (numChildrenVisited[DI->getIDom()] == DI->getIDom()->getNumChildren()) { 1087 availableOut.erase(DI->getIDom()->getBlock()); 1088 numChildrenVisited.erase(DI->getIDom()); 1089 } 1090 } 1091 1092 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1093 BI != BE;) { 1094 changed_function |= processInstruction(BI, currAvail, 1095 lastSeenLoad, toErase); 1096 if (toErase.empty()) { 1097 ++BI; 1098 continue; 1099 } 1100 1101 // If we need some instructions deleted, do it now. 1102 NumGVNInstr += toErase.size(); 1103 1104 // Avoid iterator invalidation. 1105 bool AtStart = BI == BB->begin(); 1106 if (!AtStart) 1107 --BI; 1108 1109 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1110 E = toErase.end(); I != E; ++I) 1111 (*I)->eraseFromParent(); 1112 1113 if (AtStart) 1114 BI = BB->begin(); 1115 else 1116 ++BI; 1117 1118 toErase.clear(); 1119 } 1120 } 1121 1122 return changed_function; 1123} 1124