GVN.cpp revision c4f406e736d2cb5f313eef50abe1c0c935c4090e
12a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)//===- GVN.cpp - Eliminate redundant values and loads ------------===// 22a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// 32a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// The LLVM Compiler Infrastructure 42a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// 55821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This file is distributed under the University of Illinois Open Source 65821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// License. See LICENSE.TXT for details. 75821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 85821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===----------------------------------------------------------------------===// 95821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 105821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This pass performs global value numbering to eliminate fully redundant 115821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// instructions. It also performs simple dead load elimination. 122a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// 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/Instructions.h" 22#include "llvm/Value.h" 23#include "llvm/ADT/DenseMap.h" 24#include "llvm/ADT/DepthFirstIterator.h" 25#include "llvm/ADT/SmallPtrSet.h" 26#include "llvm/ADT/SmallVector.h" 27#include "llvm/ADT/SparseBitVector.h" 28#include "llvm/ADT/Statistic.h" 29#include "llvm/Analysis/Dominators.h" 30#include "llvm/Analysis/AliasAnalysis.h" 31#include "llvm/Analysis/MemoryDependenceAnalysis.h" 32#include "llvm/Support/CFG.h" 33#include "llvm/Support/Compiler.h" 34#include "llvm/Support/Debug.h" 35#include <list> 36using namespace llvm; 37 38STATISTIC(NumGVNInstr, "Number of instructions deleted"); 39STATISTIC(NumGVNLoad, "Number of loads deleted"); 40 41//===----------------------------------------------------------------------===// 42// ValueTable Class 43//===----------------------------------------------------------------------===// 44 45/// This class holds the mapping between values and value numbers. It is used 46/// as an efficient mechanism to determine the expression-wise equivalence of 47/// two values. 48namespace { 49 struct VISIBILITY_HIDDEN Expression { 50 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM, 51 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 52 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 53 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 54 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 55 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 56 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 57 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 58 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 59 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 60 EMPTY, TOMBSTONE }; 61 62 ExpressionOpcode opcode; 63 const Type* type; 64 uint32_t firstVN; 65 uint32_t secondVN; 66 uint32_t thirdVN; 67 SmallVector<uint32_t, 4> varargs; 68 Value* function; 69 70 Expression() { } 71 Expression(ExpressionOpcode o) : opcode(o) { } 72 73 bool operator==(const Expression &other) const { 74 if (opcode != other.opcode) 75 return false; 76 else if (opcode == EMPTY || opcode == TOMBSTONE) 77 return true; 78 else if (type != other.type) 79 return false; 80 else if (function != other.function) 81 return false; 82 else if (firstVN != other.firstVN) 83 return false; 84 else if (secondVN != other.secondVN) 85 return false; 86 else if (thirdVN != other.thirdVN) 87 return false; 88 else { 89 if (varargs.size() != other.varargs.size()) 90 return false; 91 92 for (size_t i = 0; i < varargs.size(); ++i) 93 if (varargs[i] != other.varargs[i]) 94 return false; 95 96 return true; 97 } 98 } 99 100 bool operator!=(const Expression &other) const { 101 if (opcode != other.opcode) 102 return true; 103 else if (opcode == EMPTY || opcode == TOMBSTONE) 104 return false; 105 else if (type != other.type) 106 return true; 107 else if (function != other.function) 108 return true; 109 else if (firstVN != other.firstVN) 110 return true; 111 else if (secondVN != other.secondVN) 112 return true; 113 else if (thirdVN != other.thirdVN) 114 return true; 115 else { 116 if (varargs.size() != other.varargs.size()) 117 return true; 118 119 for (size_t i = 0; i < varargs.size(); ++i) 120 if (varargs[i] != other.varargs[i]) 121 return true; 122 123 return false; 124 } 125 } 126 }; 127 128 class VISIBILITY_HIDDEN ValueTable { 129 private: 130 DenseMap<Value*, uint32_t> valueNumbering; 131 DenseMap<Expression, uint32_t> expressionNumbering; 132 AliasAnalysis* AA; 133 MemoryDependenceAnalysis* MD; 134 DominatorTree* DT; 135 136 uint32_t nextValueNumber; 137 138 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 139 Expression::ExpressionOpcode getOpcode(CmpInst* C); 140 Expression::ExpressionOpcode getOpcode(CastInst* C); 141 Expression create_expression(BinaryOperator* BO); 142 Expression create_expression(CmpInst* C); 143 Expression create_expression(ShuffleVectorInst* V); 144 Expression create_expression(ExtractElementInst* C); 145 Expression create_expression(InsertElementInst* V); 146 Expression create_expression(SelectInst* V); 147 Expression create_expression(CastInst* C); 148 Expression create_expression(GetElementPtrInst* G); 149 Expression create_expression(CallInst* C); 150 Expression create_expression(Constant* C); 151 public: 152 ValueTable() : nextValueNumber(1) { } 153 uint32_t lookup_or_add(Value* V); 154 uint32_t lookup(Value* V) const; 155 void add(Value* V, uint32_t num); 156 void clear(); 157 void erase(Value* v); 158 unsigned size(); 159 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 160 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 161 void setDomTree(DominatorTree* D) { DT = D; } 162 }; 163} 164 165namespace llvm { 166template <> struct DenseMapInfo<Expression> { 167 static inline Expression getEmptyKey() { 168 return Expression(Expression::EMPTY); 169 } 170 171 static inline Expression getTombstoneKey() { 172 return Expression(Expression::TOMBSTONE); 173 } 174 175 static unsigned getHashValue(const Expression e) { 176 unsigned hash = e.opcode; 177 178 hash = e.firstVN + hash * 37; 179 hash = e.secondVN + hash * 37; 180 hash = e.thirdVN + hash * 37; 181 182 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 183 (unsigned)((uintptr_t)e.type >> 9)) + 184 hash * 37; 185 186 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 187 E = e.varargs.end(); I != E; ++I) 188 hash = *I + hash * 37; 189 190 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 191 (unsigned)((uintptr_t)e.function >> 9)) + 192 hash * 37; 193 194 return hash; 195 } 196 static bool isEqual(const Expression &LHS, const Expression &RHS) { 197 return LHS == RHS; 198 } 199 static bool isPod() { return true; } 200}; 201} 202 203//===----------------------------------------------------------------------===// 204// ValueTable Internal Functions 205//===----------------------------------------------------------------------===// 206Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 207 switch(BO->getOpcode()) { 208 default: // THIS SHOULD NEVER HAPPEN 209 assert(0 && "Binary operator with unknown opcode?"); 210 case Instruction::Add: return Expression::ADD; 211 case Instruction::Sub: return Expression::SUB; 212 case Instruction::Mul: return Expression::MUL; 213 case Instruction::UDiv: return Expression::UDIV; 214 case Instruction::SDiv: return Expression::SDIV; 215 case Instruction::FDiv: return Expression::FDIV; 216 case Instruction::URem: return Expression::UREM; 217 case Instruction::SRem: return Expression::SREM; 218 case Instruction::FRem: return Expression::FREM; 219 case Instruction::Shl: return Expression::SHL; 220 case Instruction::LShr: return Expression::LSHR; 221 case Instruction::AShr: return Expression::ASHR; 222 case Instruction::And: return Expression::AND; 223 case Instruction::Or: return Expression::OR; 224 case Instruction::Xor: return Expression::XOR; 225 } 226} 227 228Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 229 if (isa<ICmpInst>(C)) { 230 switch (C->getPredicate()) { 231 default: // THIS SHOULD NEVER HAPPEN 232 assert(0 && "Comparison with unknown predicate?"); 233 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 234 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 235 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 236 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 237 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 238 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 239 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 240 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 241 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 242 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 243 } 244 } 245 assert(isa<FCmpInst>(C) && "Unknown compare"); 246 switch (C->getPredicate()) { 247 default: // THIS SHOULD NEVER HAPPEN 248 assert(0 && "Comparison with unknown predicate?"); 249 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 250 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 251 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 252 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 253 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 254 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 255 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 256 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 257 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 258 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 259 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 260 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 261 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 262 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 263 } 264} 265 266Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 267 switch(C->getOpcode()) { 268 default: // THIS SHOULD NEVER HAPPEN 269 assert(0 && "Cast operator with unknown opcode?"); 270 case Instruction::Trunc: return Expression::TRUNC; 271 case Instruction::ZExt: return Expression::ZEXT; 272 case Instruction::SExt: return Expression::SEXT; 273 case Instruction::FPToUI: return Expression::FPTOUI; 274 case Instruction::FPToSI: return Expression::FPTOSI; 275 case Instruction::UIToFP: return Expression::UITOFP; 276 case Instruction::SIToFP: return Expression::SITOFP; 277 case Instruction::FPTrunc: return Expression::FPTRUNC; 278 case Instruction::FPExt: return Expression::FPEXT; 279 case Instruction::PtrToInt: return Expression::PTRTOINT; 280 case Instruction::IntToPtr: return Expression::INTTOPTR; 281 case Instruction::BitCast: return Expression::BITCAST; 282 } 283} 284 285Expression ValueTable::create_expression(CallInst* C) { 286 Expression e; 287 288 e.type = C->getType(); 289 e.firstVN = 0; 290 e.secondVN = 0; 291 e.thirdVN = 0; 292 e.function = C->getCalledFunction(); 293 e.opcode = Expression::CALL; 294 295 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 296 I != E; ++I) 297 e.varargs.push_back(lookup_or_add(*I)); 298 299 return e; 300} 301 302Expression ValueTable::create_expression(BinaryOperator* BO) { 303 Expression e; 304 305 e.firstVN = lookup_or_add(BO->getOperand(0)); 306 e.secondVN = lookup_or_add(BO->getOperand(1)); 307 e.thirdVN = 0; 308 e.function = 0; 309 e.type = BO->getType(); 310 e.opcode = getOpcode(BO); 311 312 return e; 313} 314 315Expression ValueTable::create_expression(CmpInst* C) { 316 Expression e; 317 318 e.firstVN = lookup_or_add(C->getOperand(0)); 319 e.secondVN = lookup_or_add(C->getOperand(1)); 320 e.thirdVN = 0; 321 e.function = 0; 322 e.type = C->getType(); 323 e.opcode = getOpcode(C); 324 325 return e; 326} 327 328Expression ValueTable::create_expression(CastInst* C) { 329 Expression e; 330 331 e.firstVN = lookup_or_add(C->getOperand(0)); 332 e.secondVN = 0; 333 e.thirdVN = 0; 334 e.function = 0; 335 e.type = C->getType(); 336 e.opcode = getOpcode(C); 337 338 return e; 339} 340 341Expression ValueTable::create_expression(ShuffleVectorInst* S) { 342 Expression e; 343 344 e.firstVN = lookup_or_add(S->getOperand(0)); 345 e.secondVN = lookup_or_add(S->getOperand(1)); 346 e.thirdVN = lookup_or_add(S->getOperand(2)); 347 e.function = 0; 348 e.type = S->getType(); 349 e.opcode = Expression::SHUFFLE; 350 351 return e; 352} 353 354Expression ValueTable::create_expression(ExtractElementInst* E) { 355 Expression e; 356 357 e.firstVN = lookup_or_add(E->getOperand(0)); 358 e.secondVN = lookup_or_add(E->getOperand(1)); 359 e.thirdVN = 0; 360 e.function = 0; 361 e.type = E->getType(); 362 e.opcode = Expression::EXTRACT; 363 364 return e; 365} 366 367Expression ValueTable::create_expression(InsertElementInst* I) { 368 Expression e; 369 370 e.firstVN = lookup_or_add(I->getOperand(0)); 371 e.secondVN = lookup_or_add(I->getOperand(1)); 372 e.thirdVN = lookup_or_add(I->getOperand(2)); 373 e.function = 0; 374 e.type = I->getType(); 375 e.opcode = Expression::INSERT; 376 377 return e; 378} 379 380Expression ValueTable::create_expression(SelectInst* I) { 381 Expression e; 382 383 e.firstVN = lookup_or_add(I->getCondition()); 384 e.secondVN = lookup_or_add(I->getTrueValue()); 385 e.thirdVN = lookup_or_add(I->getFalseValue()); 386 e.function = 0; 387 e.type = I->getType(); 388 e.opcode = Expression::SELECT; 389 390 return e; 391} 392 393Expression ValueTable::create_expression(GetElementPtrInst* G) { 394 Expression e; 395 396 e.firstVN = lookup_or_add(G->getPointerOperand()); 397 e.secondVN = 0; 398 e.thirdVN = 0; 399 e.function = 0; 400 e.type = G->getType(); 401 e.opcode = Expression::GEP; 402 403 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 404 I != E; ++I) 405 e.varargs.push_back(lookup_or_add(*I)); 406 407 return e; 408} 409 410//===----------------------------------------------------------------------===// 411// ValueTable External Functions 412//===----------------------------------------------------------------------===// 413 414/// lookup_or_add - Returns the value number for the specified value, assigning 415/// it a new number if it did not have one before. 416uint32_t ValueTable::lookup_or_add(Value* V) { 417 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 418 if (VI != valueNumbering.end()) 419 return VI->second; 420 421 if (CallInst* C = dyn_cast<CallInst>(V)) { 422 if (AA->doesNotAccessMemory(C)) { 423 Expression e = create_expression(C); 424 425 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 426 if (EI != expressionNumbering.end()) { 427 valueNumbering.insert(std::make_pair(V, EI->second)); 428 return EI->second; 429 } else { 430 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 431 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 432 433 return nextValueNumber++; 434 } 435 } else if (AA->onlyReadsMemory(C)) { 436 Expression e = create_expression(C); 437 438 if (expressionNumbering.find(e) == expressionNumbering.end()) { 439 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 440 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 441 return nextValueNumber++; 442 } 443 444 Instruction* local_dep = MD->getDependency(C); 445 446 if (local_dep == MemoryDependenceAnalysis::None) { 447 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 448 return nextValueNumber++; 449 } else if (local_dep != MemoryDependenceAnalysis::NonLocal) { 450 if (!isa<CallInst>(local_dep)) { 451 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 452 return nextValueNumber++; 453 } 454 455 CallInst* local_cdep = cast<CallInst>(local_dep); 456 457 if (local_cdep->getCalledFunction() != C->getCalledFunction() || 458 local_cdep->getNumOperands() != C->getNumOperands()) { 459 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 460 return nextValueNumber++; 461 } else if (!C->getCalledFunction()) { 462 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 463 return nextValueNumber++; 464 } else { 465 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 466 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 467 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 468 if (c_vn != cd_vn) { 469 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 470 return nextValueNumber++; 471 } 472 } 473 474 uint32_t v = lookup_or_add(local_cdep); 475 valueNumbering.insert(std::make_pair(V, v)); 476 return v; 477 } 478 } 479 480 481 DenseMap<BasicBlock*, Value*> deps; 482 MD->getNonLocalDependency(C, deps); 483 CallInst* cdep = 0; 484 485 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), 486 E = deps.end(); I != E; ++I) { 487 if (I->second == MemoryDependenceAnalysis::None) { 488 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 489 490 return nextValueNumber++; 491 } else if (I->second != MemoryDependenceAnalysis::NonLocal) { 492 if (DT->dominates(I->first, C->getParent())) { 493 if (CallInst* CD = dyn_cast<CallInst>(I->second)) 494 cdep = CD; 495 else { 496 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 497 return nextValueNumber++; 498 } 499 } else { 500 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 501 return nextValueNumber++; 502 } 503 } 504 } 505 506 if (!cdep) { 507 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 508 return nextValueNumber++; 509 } 510 511 if (cdep->getCalledFunction() != C->getCalledFunction() || 512 cdep->getNumOperands() != C->getNumOperands()) { 513 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 514 return nextValueNumber++; 515 } else if (!C->getCalledFunction()) { 516 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 517 return nextValueNumber++; 518 } else { 519 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 520 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 521 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 522 if (c_vn != cd_vn) { 523 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 524 return nextValueNumber++; 525 } 526 } 527 528 uint32_t v = lookup_or_add(cdep); 529 valueNumbering.insert(std::make_pair(V, v)); 530 return v; 531 } 532 533 } else { 534 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 535 return nextValueNumber++; 536 } 537 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 538 Expression e = create_expression(BO); 539 540 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 541 if (EI != expressionNumbering.end()) { 542 valueNumbering.insert(std::make_pair(V, EI->second)); 543 return EI->second; 544 } else { 545 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 546 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 547 548 return nextValueNumber++; 549 } 550 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 551 Expression e = create_expression(C); 552 553 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 554 if (EI != expressionNumbering.end()) { 555 valueNumbering.insert(std::make_pair(V, EI->second)); 556 return EI->second; 557 } else { 558 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 559 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 560 561 return nextValueNumber++; 562 } 563 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 564 Expression e = create_expression(U); 565 566 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 567 if (EI != expressionNumbering.end()) { 568 valueNumbering.insert(std::make_pair(V, EI->second)); 569 return EI->second; 570 } else { 571 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 572 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 573 574 return nextValueNumber++; 575 } 576 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 577 Expression e = create_expression(U); 578 579 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 580 if (EI != expressionNumbering.end()) { 581 valueNumbering.insert(std::make_pair(V, EI->second)); 582 return EI->second; 583 } else { 584 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 585 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 586 587 return nextValueNumber++; 588 } 589 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 590 Expression e = create_expression(U); 591 592 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 593 if (EI != expressionNumbering.end()) { 594 valueNumbering.insert(std::make_pair(V, EI->second)); 595 return EI->second; 596 } else { 597 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 598 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 599 600 return nextValueNumber++; 601 } 602 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 603 Expression e = create_expression(U); 604 605 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 606 if (EI != expressionNumbering.end()) { 607 valueNumbering.insert(std::make_pair(V, EI->second)); 608 return EI->second; 609 } else { 610 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 611 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 612 613 return nextValueNumber++; 614 } 615 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 616 Expression e = create_expression(U); 617 618 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 619 if (EI != expressionNumbering.end()) { 620 valueNumbering.insert(std::make_pair(V, EI->second)); 621 return EI->second; 622 } else { 623 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 624 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 625 626 return nextValueNumber++; 627 } 628 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 629 Expression e = create_expression(U); 630 631 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 632 if (EI != expressionNumbering.end()) { 633 valueNumbering.insert(std::make_pair(V, EI->second)); 634 return EI->second; 635 } else { 636 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 637 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 638 639 return nextValueNumber++; 640 } 641 } else { 642 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 643 return nextValueNumber++; 644 } 645} 646 647/// lookup - Returns the value number of the specified value. Fails if 648/// the value has not yet been numbered. 649uint32_t ValueTable::lookup(Value* V) const { 650 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 651 assert(VI != valueNumbering.end() && "Value not numbered?"); 652 return VI->second; 653} 654 655/// clear - Remove all entries from the ValueTable 656void ValueTable::clear() { 657 valueNumbering.clear(); 658 expressionNumbering.clear(); 659 nextValueNumber = 1; 660} 661 662/// erase - Remove a value from the value numbering 663void ValueTable::erase(Value* V) { 664 valueNumbering.erase(V); 665} 666 667//===----------------------------------------------------------------------===// 668// ValueNumberedSet Class 669//===----------------------------------------------------------------------===// 670namespace { 671class VISIBILITY_HIDDEN ValueNumberedSet { 672 private: 673 SmallPtrSet<Value*, 8> contents; 674 SparseBitVector<64> numbers; 675 public: 676 ValueNumberedSet() { } 677 ValueNumberedSet(const ValueNumberedSet& other) { 678 numbers = other.numbers; 679 contents = other.contents; 680 } 681 682 typedef SmallPtrSet<Value*, 8>::iterator iterator; 683 684 iterator begin() { return contents.begin(); } 685 iterator end() { return contents.end(); } 686 687 bool insert(Value* v) { return contents.insert(v); } 688 void insert(iterator I, iterator E) { contents.insert(I, E); } 689 void erase(Value* v) { contents.erase(v); } 690 unsigned count(Value* v) { return contents.count(v); } 691 size_t size() { return contents.size(); } 692 693 void set(unsigned i) { 694 numbers.set(i); 695 } 696 697 void operator=(const ValueNumberedSet& other) { 698 contents = other.contents; 699 numbers = other.numbers; 700 } 701 702 void reset(unsigned i) { 703 numbers.reset(i); 704 } 705 706 bool test(unsigned i) { 707 return numbers.test(i); 708 } 709}; 710} 711 712//===----------------------------------------------------------------------===// 713// GVN Pass 714//===----------------------------------------------------------------------===// 715 716namespace { 717 718 class VISIBILITY_HIDDEN GVN : public FunctionPass { 719 bool runOnFunction(Function &F); 720 public: 721 static char ID; // Pass identification, replacement for typeid 722 GVN() : FunctionPass((intptr_t)&ID) { } 723 724 private: 725 ValueTable VN; 726 727 DenseMap<BasicBlock*, ValueNumberedSet> availableOut; 728 729 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 730 PhiMapType phiMap; 731 732 733 // This transformation requires dominator postdominator info 734 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 735 AU.setPreservesCFG(); 736 AU.addRequired<DominatorTree>(); 737 AU.addRequired<MemoryDependenceAnalysis>(); 738 AU.addRequired<AliasAnalysis>(); 739 AU.addPreserved<AliasAnalysis>(); 740 AU.addPreserved<MemoryDependenceAnalysis>(); 741 } 742 743 // Helper fuctions 744 // FIXME: eliminate or document these better 745 Value* find_leader(ValueNumberedSet& vals, uint32_t v) ; 746 void val_insert(ValueNumberedSet& s, Value* v); 747 bool processLoad(LoadInst* L, 748 DenseMap<Value*, LoadInst*> &lastLoad, 749 SmallVectorImpl<Instruction*> &toErase); 750 bool processInstruction(Instruction* I, 751 ValueNumberedSet& currAvail, 752 DenseMap<Value*, LoadInst*>& lastSeenLoad, 753 SmallVectorImpl<Instruction*> &toErase); 754 bool processNonLocalLoad(LoadInst* L, 755 SmallVectorImpl<Instruction*> &toErase); 756 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig, 757 DenseMap<BasicBlock*, Value*> &Phis, 758 bool top_level = false); 759 void dump(DenseMap<BasicBlock*, Value*>& d); 760 bool iterateOnFunction(Function &F); 761 Value* CollapsePhi(PHINode* p); 762 bool isSafeReplacement(PHINode* p, Instruction* inst); 763 }; 764 765 char GVN::ID = 0; 766} 767 768// createGVNPass - The public interface to this file... 769FunctionPass *llvm::createGVNPass() { return new GVN(); } 770 771static RegisterPass<GVN> X("gvn", 772 "Global Value Numbering"); 773 774/// find_leader - Given a set and a value number, return the first 775/// element of the set with that value number, or 0 if no such element 776/// is present 777Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) { 778 if (!vals.test(v)) 779 return 0; 780 781 for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end(); 782 I != E; ++I) 783 if (v == VN.lookup(*I)) 784 return *I; 785 786 assert(0 && "No leader found, but present bit is set?"); 787 return 0; 788} 789 790/// val_insert - Insert a value into a set only if there is not a value 791/// with the same value number already in the set 792void GVN::val_insert(ValueNumberedSet& s, Value* v) { 793 uint32_t num = VN.lookup(v); 794 if (!s.test(num)) 795 s.insert(v); 796} 797 798void GVN::dump(DenseMap<BasicBlock*, Value*>& d) { 799 printf("{\n"); 800 for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(), 801 E = d.end(); I != E; ++I) { 802 if (I->second == MemoryDependenceAnalysis::None) 803 printf("None\n"); 804 else 805 I->second->dump(); 806 } 807 printf("}\n"); 808} 809 810Value* GVN::CollapsePhi(PHINode* p) { 811 DominatorTree &DT = getAnalysis<DominatorTree>(); 812 Value* constVal = p->hasConstantValue(); 813 814 if (!constVal) return 0; 815 816 Instruction* inst = dyn_cast<Instruction>(constVal); 817 if (!inst) 818 return constVal; 819 820 if (DT.dominates(inst, p)) 821 if (isSafeReplacement(p, inst)) 822 return inst; 823 return 0; 824} 825 826bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 827 if (!isa<PHINode>(inst)) 828 return true; 829 830 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 831 UI != E; ++UI) 832 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 833 if (use_phi->getParent() == inst->getParent()) 834 return false; 835 836 return true; 837} 838 839/// GetValueForBlock - Get the value to use within the specified basic block. 840/// available values are in Phis. 841Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig, 842 DenseMap<BasicBlock*, Value*> &Phis, 843 bool top_level) { 844 845 // If we have already computed this value, return the previously computed val. 846 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 847 if (V != Phis.end() && !top_level) return V->second; 848 849 BasicBlock* singlePred = BB->getSinglePredecessor(); 850 if (singlePred) { 851 Value *ret = GetValueForBlock(singlePred, orig, Phis); 852 Phis[BB] = ret; 853 return ret; 854 } 855 856 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 857 // now, then get values to fill in the incoming values for the PHI. 858 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 859 BB->begin()); 860 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB))); 861 862 if (Phis.count(BB) == 0) 863 Phis.insert(std::make_pair(BB, PN)); 864 865 // Fill in the incoming values for the block. 866 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 867 Value* val = GetValueForBlock(*PI, orig, Phis); 868 PN->addIncoming(val, *PI); 869 } 870 871 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 872 AA.copyValue(orig, PN); 873 874 // Attempt to collapse PHI nodes that are trivially redundant 875 Value* v = CollapsePhi(PN); 876 if (!v) { 877 // Cache our phi construction results 878 phiMap[orig->getPointerOperand()].insert(PN); 879 return PN; 880 } 881 882 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 883 884 MD.removeInstruction(PN); 885 PN->replaceAllUsesWith(v); 886 887 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 888 E = Phis.end(); I != E; ++I) 889 if (I->second == PN) 890 I->second = v; 891 892 PN->eraseFromParent(); 893 894 Phis[BB] = v; 895 return v; 896} 897 898/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 899/// non-local by performing PHI construction. 900bool GVN::processNonLocalLoad(LoadInst* L, 901 SmallVectorImpl<Instruction*> &toErase) { 902 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 903 904 // Find the non-local dependencies of the load 905 DenseMap<BasicBlock*, Value*> deps; 906 MD.getNonLocalDependency(L, deps); 907 908 DenseMap<BasicBlock*, Value*> repl; 909 910 // Filter out useless results (non-locals, etc) 911 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end(); 912 I != E; ++I) { 913 if (I->second == MemoryDependenceAnalysis::None) 914 return false; 915 916 if (I->second == MemoryDependenceAnalysis::NonLocal) 917 continue; 918 919 if (StoreInst* S = dyn_cast<StoreInst>(I->second)) { 920 if (S->getPointerOperand() != L->getPointerOperand()) 921 return false; 922 repl[I->first] = S->getOperand(0); 923 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) { 924 if (LD->getPointerOperand() != L->getPointerOperand()) 925 return false; 926 repl[I->first] = LD; 927 } else { 928 return false; 929 } 930 } 931 932 // Use cached PHI construction information from previous runs 933 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()]; 934 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 935 I != E; ++I) { 936 if ((*I)->getParent() == L->getParent()) { 937 MD.removeInstruction(L); 938 L->replaceAllUsesWith(*I); 939 toErase.push_back(L); 940 NumGVNLoad++; 941 return true; 942 } 943 944 repl.insert(std::make_pair((*I)->getParent(), *I)); 945 } 946 947 // Perform PHI construction 948 SmallPtrSet<BasicBlock*, 4> visited; 949 Value* v = GetValueForBlock(L->getParent(), L, repl, true); 950 951 MD.removeInstruction(L); 952 L->replaceAllUsesWith(v); 953 toErase.push_back(L); 954 NumGVNLoad++; 955 956 return true; 957} 958 959/// processLoad - Attempt to eliminate a load, first by eliminating it 960/// locally, and then attempting non-local elimination if that fails. 961bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad, 962 SmallVectorImpl<Instruction*> &toErase) { 963 if (L->isVolatile()) { 964 lastLoad[L->getPointerOperand()] = L; 965 return false; 966 } 967 968 Value* pointer = L->getPointerOperand(); 969 LoadInst*& last = lastLoad[pointer]; 970 971 // ... to a pointer that has been loaded from before... 972 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 973 bool removedNonLocal = false; 974 Instruction* dep = MD.getDependency(L); 975 if (dep == MemoryDependenceAnalysis::NonLocal && 976 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) { 977 removedNonLocal = processNonLocalLoad(L, toErase); 978 979 if (!removedNonLocal) 980 last = L; 981 982 return removedNonLocal; 983 } 984 985 986 bool deletedLoad = false; 987 988 // Walk up the dependency chain until we either find 989 // a dependency we can use, or we can't walk any further 990 while (dep != MemoryDependenceAnalysis::None && 991 dep != MemoryDependenceAnalysis::NonLocal && 992 (isa<LoadInst>(dep) || isa<StoreInst>(dep))) { 993 // ... that depends on a store ... 994 if (StoreInst* S = dyn_cast<StoreInst>(dep)) { 995 if (S->getPointerOperand() == pointer) { 996 // Remove it! 997 MD.removeInstruction(L); 998 999 L->replaceAllUsesWith(S->getOperand(0)); 1000 toErase.push_back(L); 1001 deletedLoad = true; 1002 NumGVNLoad++; 1003 } 1004 1005 // Whether we removed it or not, we can't 1006 // go any further 1007 break; 1008 } else if (!last) { 1009 // If we don't depend on a store, and we haven't 1010 // been loaded before, bail. 1011 break; 1012 } else if (dep == last) { 1013 // Remove it! 1014 MD.removeInstruction(L); 1015 1016 L->replaceAllUsesWith(last); 1017 toErase.push_back(L); 1018 deletedLoad = true; 1019 NumGVNLoad++; 1020 1021 break; 1022 } else { 1023 dep = MD.getDependency(L, dep); 1024 } 1025 } 1026 1027 if (dep != MemoryDependenceAnalysis::None && 1028 dep != MemoryDependenceAnalysis::NonLocal && 1029 isa<AllocationInst>(dep)) { 1030 // Check that this load is actually from the 1031 // allocation we found 1032 Value* v = L->getOperand(0); 1033 while (true) { 1034 if (BitCastInst *BC = dyn_cast<BitCastInst>(v)) 1035 v = BC->getOperand(0); 1036 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v)) 1037 v = GEP->getOperand(0); 1038 else 1039 break; 1040 } 1041 if (v == dep) { 1042 // If this load depends directly on an allocation, there isn't 1043 // anything stored there; therefore, we can optimize this load 1044 // to undef. 1045 MD.removeInstruction(L); 1046 1047 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1048 toErase.push_back(L); 1049 deletedLoad = true; 1050 NumGVNLoad++; 1051 } 1052 } 1053 1054 if (!deletedLoad) 1055 last = L; 1056 1057 return deletedLoad; 1058} 1059 1060/// processInstruction - When calculating availability, handle an instruction 1061/// by inserting it into the appropriate sets 1062bool GVN::processInstruction(Instruction *I, ValueNumberedSet &currAvail, 1063 DenseMap<Value*, LoadInst*> &lastSeenLoad, 1064 SmallVectorImpl<Instruction*> &toErase) { 1065 if (LoadInst* L = dyn_cast<LoadInst>(I)) 1066 return processLoad(L, lastSeenLoad, toErase); 1067 1068 // Allocations are always uniquely numbered, so we can save time and memory 1069 // by fast failing them. 1070 if (isa<AllocationInst>(I)) 1071 return false; 1072 1073 unsigned num = VN.lookup_or_add(I); 1074 1075 // Collapse PHI nodes 1076 if (PHINode* p = dyn_cast<PHINode>(I)) { 1077 Value* constVal = CollapsePhi(p); 1078 1079 if (constVal) { 1080 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1081 PI != PE; ++PI) 1082 if (PI->second.count(p)) 1083 PI->second.erase(p); 1084 1085 p->replaceAllUsesWith(constVal); 1086 toErase.push_back(p); 1087 } 1088 // Perform value-number based elimination 1089 } else if (currAvail.test(num)) { 1090 Value* repl = find_leader(currAvail, num); 1091 1092 // Remove it! 1093 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 1094 MD.removeInstruction(I); 1095 1096 VN.erase(I); 1097 I->replaceAllUsesWith(repl); 1098 toErase.push_back(I); 1099 return true; 1100 } else if (!I->isTerminator()) { 1101 currAvail.set(num); 1102 currAvail.insert(I); 1103 } 1104 1105 return false; 1106} 1107 1108// GVN::runOnFunction - This is the main transformation entry point for a 1109// function. 1110// 1111bool GVN::runOnFunction(Function& F) { 1112 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1113 VN.setMemDep(&getAnalysis<MemoryDependenceAnalysis>()); 1114 VN.setDomTree(&getAnalysis<DominatorTree>()); 1115 1116 bool changed = false; 1117 bool shouldContinue = true; 1118 1119 while (shouldContinue) { 1120 shouldContinue = iterateOnFunction(F); 1121 changed |= shouldContinue; 1122 } 1123 1124 return changed; 1125} 1126 1127 1128// GVN::iterateOnFunction - Executes one iteration of GVN 1129bool GVN::iterateOnFunction(Function &F) { 1130 // Clean out global sets from any previous functions 1131 VN.clear(); 1132 availableOut.clear(); 1133 phiMap.clear(); 1134 1135 bool changed_function = false; 1136 1137 DominatorTree &DT = getAnalysis<DominatorTree>(); 1138 1139 SmallVector<Instruction*, 8> toErase; 1140 DenseMap<Value*, LoadInst*> lastSeenLoad; 1141 DenseMap<DomTreeNode*, size_t> numChildrenVisited; 1142 1143 // Top-down walk of the dominator tree 1144 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()), 1145 E = df_end(DT.getRootNode()); DI != E; ++DI) { 1146 1147 // Get the set to update for this block 1148 ValueNumberedSet& currAvail = availableOut[DI->getBlock()]; 1149 lastSeenLoad.clear(); 1150 1151 BasicBlock* BB = DI->getBlock(); 1152 1153 // A block inherits AVAIL_OUT from its dominator 1154 if (DI->getIDom() != 0) { 1155 currAvail = availableOut[DI->getIDom()->getBlock()]; 1156 1157 numChildrenVisited[DI->getIDom()]++; 1158 1159 if (numChildrenVisited[DI->getIDom()] == DI->getIDom()->getNumChildren()) { 1160 availableOut.erase(DI->getIDom()->getBlock()); 1161 numChildrenVisited.erase(DI->getIDom()); 1162 } 1163 } 1164 1165 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1166 BI != BE;) { 1167 changed_function |= processInstruction(BI, currAvail, 1168 lastSeenLoad, toErase); 1169 if (toErase.empty()) { 1170 ++BI; 1171 continue; 1172 } 1173 1174 // If we need some instructions deleted, do it now. 1175 NumGVNInstr += toErase.size(); 1176 1177 // Avoid iterator invalidation. 1178 bool AtStart = BI == BB->begin(); 1179 if (!AtStart) 1180 --BI; 1181 1182 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1183 E = toErase.end(); I != E; ++I) 1184 (*I)->eraseFromParent(); 1185 1186 if (AtStart) 1187 BI = BB->begin(); 1188 else 1189 ++BI; 1190 1191 toErase.clear(); 1192 } 1193 } 1194 1195 return changed_function; 1196} 1197