GVN.cpp revision 72bc70d499f64216defd971ef140cbaa2fad0fbd
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// Note that this pass does the value numbering itself, it does not use the 14// ValueNumbering analysis passes. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "gvn" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/BasicBlock.h" 21#include "llvm/Constants.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/Function.h" 24#include "llvm/Instructions.h" 25#include "llvm/Value.h" 26#include "llvm/ADT/DenseMap.h" 27#include "llvm/ADT/DepthFirstIterator.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/ADT/SmallVector.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/Transforms/Utils/BasicBlockUtils.h" 39#include <cstdio> 40using namespace llvm; 41 42STATISTIC(NumGVNInstr, "Number of instructions deleted"); 43STATISTIC(NumGVNLoad, "Number of loads deleted"); 44STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 45STATISTIC(NumGVNBlocks, "Number of blocks merged"); 46STATISTIC(NumPRELoad, "Number of loads PRE'd"); 47 48static cl::opt<bool> EnablePRE("enable-pre", 49 cl::init(true), cl::Hidden); 50cl::opt<bool> EnableLoadPRE("enable-load-pre"/*, cl::init(true)*/); 51 52//===----------------------------------------------------------------------===// 53// ValueTable Class 54//===----------------------------------------------------------------------===// 55 56/// This class holds the mapping between values and value numbers. It is used 57/// as an efficient mechanism to determine the expression-wise equivalence of 58/// two values. 59namespace { 60 struct VISIBILITY_HIDDEN Expression { 61 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM, 62 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 63 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 64 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 65 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 66 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 67 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 68 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 69 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 70 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 71 EMPTY, TOMBSTONE }; 72 73 ExpressionOpcode opcode; 74 const Type* type; 75 uint32_t firstVN; 76 uint32_t secondVN; 77 uint32_t thirdVN; 78 SmallVector<uint32_t, 4> varargs; 79 Value* function; 80 81 Expression() { } 82 Expression(ExpressionOpcode o) : opcode(o) { } 83 84 bool operator==(const Expression &other) const { 85 if (opcode != other.opcode) 86 return false; 87 else if (opcode == EMPTY || opcode == TOMBSTONE) 88 return true; 89 else if (type != other.type) 90 return false; 91 else if (function != other.function) 92 return false; 93 else if (firstVN != other.firstVN) 94 return false; 95 else if (secondVN != other.secondVN) 96 return false; 97 else if (thirdVN != other.thirdVN) 98 return false; 99 else { 100 if (varargs.size() != other.varargs.size()) 101 return false; 102 103 for (size_t i = 0; i < varargs.size(); ++i) 104 if (varargs[i] != other.varargs[i]) 105 return false; 106 107 return true; 108 } 109 } 110 111 bool operator!=(const Expression &other) const { 112 if (opcode != other.opcode) 113 return true; 114 else if (opcode == EMPTY || opcode == TOMBSTONE) 115 return false; 116 else if (type != other.type) 117 return true; 118 else if (function != other.function) 119 return true; 120 else if (firstVN != other.firstVN) 121 return true; 122 else if (secondVN != other.secondVN) 123 return true; 124 else if (thirdVN != other.thirdVN) 125 return true; 126 else { 127 if (varargs.size() != other.varargs.size()) 128 return true; 129 130 for (size_t i = 0; i < varargs.size(); ++i) 131 if (varargs[i] != other.varargs[i]) 132 return true; 133 134 return false; 135 } 136 } 137 }; 138 139 class VISIBILITY_HIDDEN ValueTable { 140 private: 141 DenseMap<Value*, uint32_t> valueNumbering; 142 DenseMap<Expression, uint32_t> expressionNumbering; 143 AliasAnalysis* AA; 144 MemoryDependenceAnalysis* MD; 145 DominatorTree* DT; 146 147 uint32_t nextValueNumber; 148 149 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 150 Expression::ExpressionOpcode getOpcode(CmpInst* C); 151 Expression::ExpressionOpcode getOpcode(CastInst* C); 152 Expression create_expression(BinaryOperator* BO); 153 Expression create_expression(CmpInst* C); 154 Expression create_expression(ShuffleVectorInst* V); 155 Expression create_expression(ExtractElementInst* C); 156 Expression create_expression(InsertElementInst* V); 157 Expression create_expression(SelectInst* V); 158 Expression create_expression(CastInst* C); 159 Expression create_expression(GetElementPtrInst* G); 160 Expression create_expression(CallInst* C); 161 Expression create_expression(Constant* C); 162 public: 163 ValueTable() : nextValueNumber(1) { } 164 uint32_t lookup_or_add(Value* V); 165 uint32_t lookup(Value* V) const; 166 void add(Value* V, uint32_t num); 167 void clear(); 168 void erase(Value* v); 169 unsigned size(); 170 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 171 AliasAnalysis *getAliasAnalysis() const { return AA; } 172 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 173 void setDomTree(DominatorTree* D) { DT = D; } 174 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 175 }; 176} 177 178namespace llvm { 179template <> struct DenseMapInfo<Expression> { 180 static inline Expression getEmptyKey() { 181 return Expression(Expression::EMPTY); 182 } 183 184 static inline Expression getTombstoneKey() { 185 return Expression(Expression::TOMBSTONE); 186 } 187 188 static unsigned getHashValue(const Expression e) { 189 unsigned hash = e.opcode; 190 191 hash = e.firstVN + hash * 37; 192 hash = e.secondVN + hash * 37; 193 hash = e.thirdVN + hash * 37; 194 195 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 196 (unsigned)((uintptr_t)e.type >> 9)) + 197 hash * 37; 198 199 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 200 E = e.varargs.end(); I != E; ++I) 201 hash = *I + hash * 37; 202 203 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 204 (unsigned)((uintptr_t)e.function >> 9)) + 205 hash * 37; 206 207 return hash; 208 } 209 static bool isEqual(const Expression &LHS, const Expression &RHS) { 210 return LHS == RHS; 211 } 212 static bool isPod() { return true; } 213}; 214} 215 216//===----------------------------------------------------------------------===// 217// ValueTable Internal Functions 218//===----------------------------------------------------------------------===// 219Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 220 switch(BO->getOpcode()) { 221 default: // THIS SHOULD NEVER HAPPEN 222 assert(0 && "Binary operator with unknown opcode?"); 223 case Instruction::Add: return Expression::ADD; 224 case Instruction::Sub: return Expression::SUB; 225 case Instruction::Mul: return Expression::MUL; 226 case Instruction::UDiv: return Expression::UDIV; 227 case Instruction::SDiv: return Expression::SDIV; 228 case Instruction::FDiv: return Expression::FDIV; 229 case Instruction::URem: return Expression::UREM; 230 case Instruction::SRem: return Expression::SREM; 231 case Instruction::FRem: return Expression::FREM; 232 case Instruction::Shl: return Expression::SHL; 233 case Instruction::LShr: return Expression::LSHR; 234 case Instruction::AShr: return Expression::ASHR; 235 case Instruction::And: return Expression::AND; 236 case Instruction::Or: return Expression::OR; 237 case Instruction::Xor: return Expression::XOR; 238 } 239} 240 241Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 242 if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) { 243 switch (C->getPredicate()) { 244 default: // THIS SHOULD NEVER HAPPEN 245 assert(0 && "Comparison with unknown predicate?"); 246 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 247 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 248 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 249 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 250 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 251 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 252 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 253 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 254 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 255 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 256 } 257 } 258 assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare"); 259 switch (C->getPredicate()) { 260 default: // THIS SHOULD NEVER HAPPEN 261 assert(0 && "Comparison with unknown predicate?"); 262 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 263 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 264 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 265 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 266 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 267 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 268 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 269 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 270 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 271 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 272 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 273 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 274 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 275 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 276 } 277} 278 279Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 280 switch(C->getOpcode()) { 281 default: // THIS SHOULD NEVER HAPPEN 282 assert(0 && "Cast operator with unknown opcode?"); 283 case Instruction::Trunc: return Expression::TRUNC; 284 case Instruction::ZExt: return Expression::ZEXT; 285 case Instruction::SExt: return Expression::SEXT; 286 case Instruction::FPToUI: return Expression::FPTOUI; 287 case Instruction::FPToSI: return Expression::FPTOSI; 288 case Instruction::UIToFP: return Expression::UITOFP; 289 case Instruction::SIToFP: return Expression::SITOFP; 290 case Instruction::FPTrunc: return Expression::FPTRUNC; 291 case Instruction::FPExt: return Expression::FPEXT; 292 case Instruction::PtrToInt: return Expression::PTRTOINT; 293 case Instruction::IntToPtr: return Expression::INTTOPTR; 294 case Instruction::BitCast: return Expression::BITCAST; 295 } 296} 297 298Expression ValueTable::create_expression(CallInst* C) { 299 Expression e; 300 301 e.type = C->getType(); 302 e.firstVN = 0; 303 e.secondVN = 0; 304 e.thirdVN = 0; 305 e.function = C->getCalledFunction(); 306 e.opcode = Expression::CALL; 307 308 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 309 I != E; ++I) 310 e.varargs.push_back(lookup_or_add(*I)); 311 312 return e; 313} 314 315Expression ValueTable::create_expression(BinaryOperator* BO) { 316 Expression e; 317 318 e.firstVN = lookup_or_add(BO->getOperand(0)); 319 e.secondVN = lookup_or_add(BO->getOperand(1)); 320 e.thirdVN = 0; 321 e.function = 0; 322 e.type = BO->getType(); 323 e.opcode = getOpcode(BO); 324 325 return e; 326} 327 328Expression ValueTable::create_expression(CmpInst* C) { 329 Expression e; 330 331 e.firstVN = lookup_or_add(C->getOperand(0)); 332 e.secondVN = lookup_or_add(C->getOperand(1)); 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(CastInst* C) { 342 Expression e; 343 344 e.firstVN = lookup_or_add(C->getOperand(0)); 345 e.secondVN = 0; 346 e.thirdVN = 0; 347 e.function = 0; 348 e.type = C->getType(); 349 e.opcode = getOpcode(C); 350 351 return e; 352} 353 354Expression ValueTable::create_expression(ShuffleVectorInst* S) { 355 Expression e; 356 357 e.firstVN = lookup_or_add(S->getOperand(0)); 358 e.secondVN = lookup_or_add(S->getOperand(1)); 359 e.thirdVN = lookup_or_add(S->getOperand(2)); 360 e.function = 0; 361 e.type = S->getType(); 362 e.opcode = Expression::SHUFFLE; 363 364 return e; 365} 366 367Expression ValueTable::create_expression(ExtractElementInst* E) { 368 Expression e; 369 370 e.firstVN = lookup_or_add(E->getOperand(0)); 371 e.secondVN = lookup_or_add(E->getOperand(1)); 372 e.thirdVN = 0; 373 e.function = 0; 374 e.type = E->getType(); 375 e.opcode = Expression::EXTRACT; 376 377 return e; 378} 379 380Expression ValueTable::create_expression(InsertElementInst* I) { 381 Expression e; 382 383 e.firstVN = lookup_or_add(I->getOperand(0)); 384 e.secondVN = lookup_or_add(I->getOperand(1)); 385 e.thirdVN = lookup_or_add(I->getOperand(2)); 386 e.function = 0; 387 e.type = I->getType(); 388 e.opcode = Expression::INSERT; 389 390 return e; 391} 392 393Expression ValueTable::create_expression(SelectInst* I) { 394 Expression e; 395 396 e.firstVN = lookup_or_add(I->getCondition()); 397 e.secondVN = lookup_or_add(I->getTrueValue()); 398 e.thirdVN = lookup_or_add(I->getFalseValue()); 399 e.function = 0; 400 e.type = I->getType(); 401 e.opcode = Expression::SELECT; 402 403 return e; 404} 405 406Expression ValueTable::create_expression(GetElementPtrInst* G) { 407 Expression e; 408 409 e.firstVN = lookup_or_add(G->getPointerOperand()); 410 e.secondVN = 0; 411 e.thirdVN = 0; 412 e.function = 0; 413 e.type = G->getType(); 414 e.opcode = Expression::GEP; 415 416 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 417 I != E; ++I) 418 e.varargs.push_back(lookup_or_add(*I)); 419 420 return e; 421} 422 423//===----------------------------------------------------------------------===// 424// ValueTable External Functions 425//===----------------------------------------------------------------------===// 426 427/// add - Insert a value into the table with a specified value number. 428void ValueTable::add(Value* V, uint32_t num) { 429 valueNumbering.insert(std::make_pair(V, num)); 430} 431 432/// lookup_or_add - Returns the value number for the specified value, assigning 433/// it a new number if it did not have one before. 434uint32_t ValueTable::lookup_or_add(Value* V) { 435 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 436 if (VI != valueNumbering.end()) 437 return VI->second; 438 439 if (CallInst* C = dyn_cast<CallInst>(V)) { 440 if (AA->doesNotAccessMemory(C)) { 441 Expression e = create_expression(C); 442 443 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 444 if (EI != expressionNumbering.end()) { 445 valueNumbering.insert(std::make_pair(V, EI->second)); 446 return EI->second; 447 } else { 448 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 449 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 450 451 return nextValueNumber++; 452 } 453 } else if (AA->onlyReadsMemory(C)) { 454 Expression e = create_expression(C); 455 456 if (expressionNumbering.find(e) == expressionNumbering.end()) { 457 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 458 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 459 return nextValueNumber++; 460 } 461 462 MemDepResult local_dep = MD->getDependency(C); 463 464 if (local_dep.isNone()) { 465 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 466 return nextValueNumber++; 467 } 468 469 if (Instruction *LocalDepInst = local_dep.getInst()) { 470 if (!isa<CallInst>(LocalDepInst)) { 471 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 472 return nextValueNumber++; 473 } 474 475 CallInst* local_cdep = cast<CallInst>(LocalDepInst); 476 477 if (local_cdep->getCalledFunction() != C->getCalledFunction() || 478 local_cdep->getNumOperands() != C->getNumOperands()) { 479 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 480 return nextValueNumber++; 481 } 482 483 if (!C->getCalledFunction()) { 484 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 485 return nextValueNumber++; 486 } 487 488 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 489 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 490 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 491 if (c_vn != cd_vn) { 492 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 493 return nextValueNumber++; 494 } 495 } 496 497 uint32_t v = lookup_or_add(local_cdep); 498 valueNumbering.insert(std::make_pair(V, v)); 499 return v; 500 } 501 502 503 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 504 MD->getNonLocalDependency(C); 505 CallInst* cdep = 0; 506 507 // Check to see if we have a single dominating call instruction that is 508 // identical to C. 509 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 510 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 511 // Ignore non-local dependencies. 512 if (I->second.isNonLocal()) 513 continue; 514 515 // We don't handle non-depedencies. If we already have a call, reject 516 // instruction dependencies. 517 if (I->second.isNone() || cdep != 0) { 518 cdep = 0; 519 break; 520 } 521 522 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 523 // FIXME: All duplicated with non-local case. 524 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 525 cdep = NonLocalDepCall; 526 continue; 527 } 528 529 cdep = 0; 530 break; 531 } 532 533 if (!cdep) { 534 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 535 return nextValueNumber++; 536 } 537 538 if (cdep->getCalledFunction() != C->getCalledFunction() || 539 cdep->getNumOperands() != C->getNumOperands()) { 540 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 541 return nextValueNumber++; 542 } 543 if (!C->getCalledFunction()) { 544 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 545 return nextValueNumber++; 546 } 547 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 548 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 549 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 550 if (c_vn != cd_vn) { 551 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 552 return nextValueNumber++; 553 } 554 } 555 556 uint32_t v = lookup_or_add(cdep); 557 valueNumbering.insert(std::make_pair(V, v)); 558 return v; 559 560 } else { 561 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 562 return nextValueNumber++; 563 } 564 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 565 Expression e = create_expression(BO); 566 567 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 568 if (EI != expressionNumbering.end()) { 569 valueNumbering.insert(std::make_pair(V, EI->second)); 570 return EI->second; 571 } else { 572 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 573 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 574 575 return nextValueNumber++; 576 } 577 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 578 Expression e = create_expression(C); 579 580 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 581 if (EI != expressionNumbering.end()) { 582 valueNumbering.insert(std::make_pair(V, EI->second)); 583 return EI->second; 584 } else { 585 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 586 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 587 588 return nextValueNumber++; 589 } 590 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 591 Expression e = create_expression(U); 592 593 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 594 if (EI != expressionNumbering.end()) { 595 valueNumbering.insert(std::make_pair(V, EI->second)); 596 return EI->second; 597 } else { 598 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 599 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 600 601 return nextValueNumber++; 602 } 603 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 604 Expression e = create_expression(U); 605 606 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 607 if (EI != expressionNumbering.end()) { 608 valueNumbering.insert(std::make_pair(V, EI->second)); 609 return EI->second; 610 } else { 611 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 612 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 613 614 return nextValueNumber++; 615 } 616 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 617 Expression e = create_expression(U); 618 619 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 620 if (EI != expressionNumbering.end()) { 621 valueNumbering.insert(std::make_pair(V, EI->second)); 622 return EI->second; 623 } else { 624 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 625 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 626 627 return nextValueNumber++; 628 } 629 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 630 Expression e = create_expression(U); 631 632 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 633 if (EI != expressionNumbering.end()) { 634 valueNumbering.insert(std::make_pair(V, EI->second)); 635 return EI->second; 636 } else { 637 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 638 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 639 640 return nextValueNumber++; 641 } 642 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 643 Expression e = create_expression(U); 644 645 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 646 if (EI != expressionNumbering.end()) { 647 valueNumbering.insert(std::make_pair(V, EI->second)); 648 return EI->second; 649 } else { 650 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 651 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 652 653 return nextValueNumber++; 654 } 655 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 656 Expression e = create_expression(U); 657 658 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 659 if (EI != expressionNumbering.end()) { 660 valueNumbering.insert(std::make_pair(V, EI->second)); 661 return EI->second; 662 } else { 663 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 664 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 665 666 return nextValueNumber++; 667 } 668 } else { 669 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 670 return nextValueNumber++; 671 } 672} 673 674/// lookup - Returns the value number of the specified value. Fails if 675/// the value has not yet been numbered. 676uint32_t ValueTable::lookup(Value* V) const { 677 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 678 assert(VI != valueNumbering.end() && "Value not numbered?"); 679 return VI->second; 680} 681 682/// clear - Remove all entries from the ValueTable 683void ValueTable::clear() { 684 valueNumbering.clear(); 685 expressionNumbering.clear(); 686 nextValueNumber = 1; 687} 688 689/// erase - Remove a value from the value numbering 690void ValueTable::erase(Value* V) { 691 valueNumbering.erase(V); 692} 693 694//===----------------------------------------------------------------------===// 695// GVN Pass 696//===----------------------------------------------------------------------===// 697 698namespace { 699 struct VISIBILITY_HIDDEN ValueNumberScope { 700 ValueNumberScope* parent; 701 DenseMap<uint32_t, Value*> table; 702 703 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 704 }; 705} 706 707namespace { 708 709 class VISIBILITY_HIDDEN GVN : public FunctionPass { 710 bool runOnFunction(Function &F); 711 public: 712 static char ID; // Pass identification, replacement for typeid 713 GVN() : FunctionPass(&ID) { } 714 715 private: 716 MemoryDependenceAnalysis *MD; 717 DominatorTree *DT; 718 719 ValueTable VN; 720 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 721 722 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 723 PhiMapType phiMap; 724 725 726 // This transformation requires dominator postdominator info 727 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 728 AU.addRequired<DominatorTree>(); 729 AU.addRequired<MemoryDependenceAnalysis>(); 730 AU.addRequired<AliasAnalysis>(); 731 732 AU.addPreserved<DominatorTree>(); 733 AU.addPreserved<AliasAnalysis>(); 734 } 735 736 // Helper fuctions 737 // FIXME: eliminate or document these better 738 bool processLoad(LoadInst* L, 739 DenseMap<Value*, LoadInst*> &lastLoad, 740 SmallVectorImpl<Instruction*> &toErase); 741 bool processInstruction(Instruction* I, 742 DenseMap<Value*, LoadInst*>& lastSeenLoad, 743 SmallVectorImpl<Instruction*> &toErase); 744 bool processNonLocalLoad(LoadInst* L, 745 SmallVectorImpl<Instruction*> &toErase); 746 bool processBlock(DomTreeNode* DTN); 747 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig, 748 DenseMap<BasicBlock*, Value*> &Phis, 749 bool top_level = false); 750 void dump(DenseMap<uint32_t, Value*>& d); 751 bool iterateOnFunction(Function &F); 752 Value* CollapsePhi(PHINode* p); 753 bool isSafeReplacement(PHINode* p, Instruction* inst); 754 bool performPRE(Function& F); 755 Value* lookupNumber(BasicBlock* BB, uint32_t num); 756 bool mergeBlockIntoPredecessor(BasicBlock* BB); 757 void cleanupGlobalSets(); 758 }; 759 760 char GVN::ID = 0; 761} 762 763// createGVNPass - The public interface to this file... 764FunctionPass *llvm::createGVNPass() { return new GVN(); } 765 766static RegisterPass<GVN> X("gvn", 767 "Global Value Numbering"); 768 769void GVN::dump(DenseMap<uint32_t, Value*>& d) { 770 printf("{\n"); 771 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 772 E = d.end(); I != E; ++I) { 773 printf("%d\n", I->first); 774 I->second->dump(); 775 } 776 printf("}\n"); 777} 778 779Value* GVN::CollapsePhi(PHINode* p) { 780 Value* constVal = p->hasConstantValue(); 781 if (!constVal) return 0; 782 783 Instruction* inst = dyn_cast<Instruction>(constVal); 784 if (!inst) 785 return constVal; 786 787 if (DT->dominates(inst, p)) 788 if (isSafeReplacement(p, inst)) 789 return inst; 790 return 0; 791} 792 793bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 794 if (!isa<PHINode>(inst)) 795 return true; 796 797 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 798 UI != E; ++UI) 799 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 800 if (use_phi->getParent() == inst->getParent()) 801 return false; 802 803 return true; 804} 805 806/// GetValueForBlock - Get the value to use within the specified basic block. 807/// available values are in Phis. 808Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig, 809 DenseMap<BasicBlock*, Value*> &Phis, 810 bool top_level) { 811 812 // If we have already computed this value, return the previously computed val. 813 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 814 if (V != Phis.end() && !top_level) return V->second; 815 816 // If the block is unreachable, just return undef, since this path 817 // can't actually occur at runtime. 818 if (!DT->isReachableFromEntry(BB)) 819 return Phis[BB] = UndefValue::get(orig->getType()); 820 821 BasicBlock* singlePred = BB->getSinglePredecessor(); 822 if (singlePred) { 823 Value *ret = GetValueForBlock(singlePred, orig, Phis); 824 Phis[BB] = ret; 825 return ret; 826 } 827 828 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 829 // now, then get values to fill in the incoming values for the PHI. 830 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 831 BB->begin()); 832 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB))); 833 834 if (Phis.count(BB) == 0) 835 Phis.insert(std::make_pair(BB, PN)); 836 837 // Fill in the incoming values for the block. 838 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 839 Value* val = GetValueForBlock(*PI, orig, Phis); 840 PN->addIncoming(val, *PI); 841 } 842 843 VN.getAliasAnalysis()->copyValue(orig, PN); 844 845 // Attempt to collapse PHI nodes that are trivially redundant 846 Value* v = CollapsePhi(PN); 847 if (!v) { 848 // Cache our phi construction results 849 phiMap[orig->getPointerOperand()].insert(PN); 850 return PN; 851 } 852 853 PN->replaceAllUsesWith(v); 854 855 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 856 E = Phis.end(); I != E; ++I) 857 if (I->second == PN) 858 I->second = v; 859 860 DEBUG(cerr << "GVN removed: " << *PN); 861 MD->removeInstruction(PN); 862 PN->eraseFromParent(); 863 864 Phis[BB] = v; 865 return v; 866} 867 868/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 869/// we're analyzing is fully available in the specified block. As we go, keep 870/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 871/// map is actually a tri-state map with the following values: 872/// 0) we know the block *is not* fully available. 873/// 1) we know the block *is* fully available. 874/// 2) we do not know whether the block is fully available or not, but we are 875/// currently speculating that it will be. 876/// 3) we are speculating for this block and have used that to speculate for 877/// other blocks. 878static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 879 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 880 // Optimistically assume that the block is fully available and check to see 881 // if we already know about this block in one lookup. 882 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 883 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 884 885 // If the entry already existed for this block, return the precomputed value. 886 if (!IV.second) { 887 // If this is a speculative "available" value, mark it as being used for 888 // speculation of other blocks. 889 if (IV.first->second == 2) 890 IV.first->second = 3; 891 return IV.first->second != 0; 892 } 893 894 // Otherwise, see if it is fully available in all predecessors. 895 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 896 897 // If this block has no predecessors, it isn't live-in here. 898 if (PI == PE) 899 goto SpeculationFailure; 900 901 for (; PI != PE; ++PI) 902 // If the value isn't fully available in one of our predecessors, then it 903 // isn't fully available in this block either. Undo our previous 904 // optimistic assumption and bail out. 905 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 906 goto SpeculationFailure; 907 908 return true; 909 910// SpeculationFailure - If we get here, we found out that this is not, after 911// all, a fully-available block. We have a problem if we speculated on this and 912// used the speculation to mark other blocks as available. 913SpeculationFailure: 914 char &BBVal = FullyAvailableBlocks[BB]; 915 916 // If we didn't speculate on this, just return with it set to false. 917 if (BBVal == 2) { 918 BBVal = 0; 919 return false; 920 } 921 922 // If we did speculate on this value, we could have blocks set to 1 that are 923 // incorrect. Walk the (transitive) successors of this block and mark them as 924 // 0 if set to one. 925 SmallVector<BasicBlock*, 32> BBWorklist; 926 BBWorklist.push_back(BB); 927 928 while (!BBWorklist.empty()) { 929 BasicBlock *Entry = BBWorklist.pop_back_val(); 930 // Note that this sets blocks to 0 (unavailable) if they happen to not 931 // already be in FullyAvailableBlocks. This is safe. 932 char &EntryVal = FullyAvailableBlocks[Entry]; 933 if (EntryVal == 0) continue; // Already unavailable. 934 935 // Mark as unavailable. 936 EntryVal = 0; 937 938 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 939 BBWorklist.push_back(*I); 940 } 941 942 return false; 943} 944 945/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 946/// non-local by performing PHI construction. 947bool GVN::processNonLocalLoad(LoadInst *LI, 948 SmallVectorImpl<Instruction*> &toErase) { 949 // Find the non-local dependencies of the load. 950 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 951 MD->getNonLocalDependency(LI); 952 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << deps.size() << *LI); 953 954 // If we had to process more than one hundred blocks to find the 955 // dependencies, this load isn't worth worrying about. Optimizing 956 // it will be too expensive. 957 if (deps.size() > 100) 958 return false; 959 960 BasicBlock *EntryBlock = &LI->getParent()->getParent()->getEntryBlock(); 961 962 // Filter out useless results (non-locals, etc). Keep track of the blocks 963 // where we have a value available in repl, also keep track of whether we see 964 // dependencies that produce an unknown value for the load (such as a call 965 // that could potentially clobber the load). 966 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 967 SmallVector<BasicBlock*, 16> UnavailableBlocks; 968 969 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 970 BasicBlock *DepBB = deps[i].first; 971 MemDepResult DepInfo = deps[i].second; 972 973 if (DepInfo.isNonLocal()) { 974 // If this is a non-local dependency in the entry block, then we depend on 975 // the value live-in at the start of the function. We could insert a load 976 // in the entry block to get this, but for now we'll just bail out. 977 if (DepBB == EntryBlock) 978 UnavailableBlocks.push_back(DepBB); 979 continue; 980 } 981 982 if (DepInfo.isNone()) { 983 ValuesPerBlock.push_back(std::make_pair(DepBB, 984 UndefValue::get(LI->getType()))); 985 continue; 986 } 987 988 if (StoreInst* S = dyn_cast<StoreInst>(DepInfo.getInst())) { 989 // Reject loads and stores that are to the same address but are of 990 // different types. 991 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because 992 // of bitfield access, it would be interesting to optimize for it at some 993 // point. 994 if (S->getOperand(0)->getType() != LI->getType()) { 995 UnavailableBlocks.push_back(DepBB); 996 continue; 997 } 998 999 if (S->getPointerOperand() != LI->getPointerOperand() && 1000 VN.getAliasAnalysis()->alias(S->getPointerOperand(), 1, 1001 LI->getPointerOperand(), 1) 1002 != AliasAnalysis::MustAlias) { 1003 UnavailableBlocks.push_back(DepBB); 1004 continue; 1005 } 1006 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1007 1008 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInfo.getInst())) { 1009 if (LD->getType() != LI->getType()) { 1010 UnavailableBlocks.push_back(DepBB); 1011 continue; 1012 } 1013 1014 if (LD->getPointerOperand() != LI->getPointerOperand() && 1015 VN.getAliasAnalysis()->alias(LD->getPointerOperand(), 1, 1016 LI->getPointerOperand(), 1) 1017 != AliasAnalysis::MustAlias) { 1018 UnavailableBlocks.push_back(DepBB); 1019 continue; 1020 } 1021 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1022 } else { 1023 UnavailableBlocks.push_back(DepBB); 1024 continue; 1025 } 1026 } 1027 1028 // If we have no predecessors that produce a known value for this load, exit 1029 // early. 1030 if (ValuesPerBlock.empty()) return false; 1031 1032 // If all of the instructions we depend on produce a known value for this 1033 // load, then it is fully redundant and we can use PHI insertion to compute 1034 // its value. Insert PHIs and remove the fully redundant value now. 1035 if (UnavailableBlocks.empty()) { 1036 // Use cached PHI construction information from previous runs 1037 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1038 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1039 I != E; ++I) { 1040 if ((*I)->getParent() == LI->getParent()) { 1041 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI); 1042 LI->replaceAllUsesWith(*I); 1043 toErase.push_back(LI); 1044 NumGVNLoad++; 1045 return true; 1046 } 1047 1048 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1049 } 1050 1051 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI); 1052 1053 DenseMap<BasicBlock*, Value*> BlockReplValues; 1054 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1055 // Perform PHI construction. 1056 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1057 LI->replaceAllUsesWith(v); 1058 toErase.push_back(LI); 1059 NumGVNLoad++; 1060 return true; 1061 } 1062 1063 if (!EnablePRE || !EnableLoadPRE) 1064 return false; 1065 1066 // Okay, we have *some* definitions of the value. This means that the value 1067 // is available in some of our (transitive) predecessors. Lets think about 1068 // doing PRE of this load. This will involve inserting a new load into the 1069 // predecessor when it's not available. We could do this in general, but 1070 // prefer to not increase code size. As such, we only do this when we know 1071 // that we only have to insert *one* load (which means we're basically moving 1072 // the load, not inserting a new one). 1073 1074 // Everything we do here is based on local predecessors of LI's block. If it 1075 // only has one predecessor, bail now. 1076 BasicBlock *LoadBB = LI->getParent(); 1077 if (LoadBB->getSinglePredecessor()) 1078 return false; 1079 1080 // If we have a repl set with LI itself in it, this means we have a loop where 1081 // at least one of the values is LI. Since this means that we won't be able 1082 // to eliminate LI even if we insert uses in the other predecessors, we will 1083 // end up increasing code size. Reject this by scanning for LI. 1084 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1085 if (ValuesPerBlock[i].second == LI) 1086 return false; 1087 1088 // Okay, we have some hope :). Check to see if the loaded value is fully 1089 // available in all but one predecessor. 1090 // FIXME: If we could restructure the CFG, we could make a common pred with 1091 // all the preds that don't have an available LI and insert a new load into 1092 // that one block. 1093 BasicBlock *UnavailablePred = 0; 1094 1095 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1096 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1097 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1098 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1099 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1100 1101 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1102 PI != E; ++PI) { 1103 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1104 continue; 1105 1106 // If this load is not available in multiple predecessors, reject it. 1107 if (UnavailablePred && UnavailablePred != *PI) 1108 return false; 1109 UnavailablePred = *PI; 1110 } 1111 1112 assert(UnavailablePred != 0 && 1113 "Fully available value should be eliminated above!"); 1114 1115 // If the loaded pointer is PHI node defined in this block, do PHI translation 1116 // to get its value in the predecessor. 1117 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1118 1119 // Make sure the value is live in the predecessor. If it was defined by a 1120 // non-PHI instruction in this block, we don't know how to recompute it above. 1121 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1122 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1123 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1124 << *LPInst << *LI << "\n"); 1125 return false; 1126 } 1127 1128 // We don't currently handle critical edges :( 1129 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1130 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1131 << UnavailablePred->getName() << "': " << *LI); 1132 return false; 1133 } 1134 1135 // Okay, we can eliminate this load by inserting a reload in the predecessor 1136 // and using PHI construction to get the value in the other predecessors, do 1137 // it. 1138 /*DEBUG*/(cerr << "GVN REMOVING PRE LOAD: " << *LI); 1139 1140 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1141 LI->getAlignment(), 1142 UnavailablePred->getTerminator()); 1143 1144 DenseMap<BasicBlock*, Value*> BlockReplValues; 1145 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1146 BlockReplValues[UnavailablePred] = NewLoad; 1147 1148 // Perform PHI construction. 1149 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1150 LI->replaceAllUsesWith(v); 1151 v->takeName(LI); 1152 toErase.push_back(LI); 1153 NumPRELoad++; 1154 return true; 1155} 1156 1157/// processLoad - Attempt to eliminate a load, first by eliminating it 1158/// locally, and then attempting non-local elimination if that fails. 1159bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad, 1160 SmallVectorImpl<Instruction*> &toErase) { 1161 if (L->isVolatile()) { 1162 lastLoad[L->getPointerOperand()] = L; 1163 return false; 1164 } 1165 1166 Value* pointer = L->getPointerOperand(); 1167 LoadInst*& last = lastLoad[pointer]; 1168 1169 // ... to a pointer that has been loaded from before... 1170 bool removedNonLocal = false; 1171 MemDepResult dep = MD->getDependency(L); 1172 if (dep.isNonLocal() && 1173 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) { 1174 removedNonLocal = processNonLocalLoad(L, toErase); 1175 1176 if (!removedNonLocal) 1177 last = L; 1178 1179 return removedNonLocal; 1180 } 1181 1182 1183 bool deletedLoad = false; 1184 1185 // Walk up the dependency chain until we either find 1186 // a dependency we can use, or we can't walk any further 1187 while (Instruction *DepInst = dep.getInst()) { 1188 // ... that depends on a store ... 1189 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 1190 if (S->getPointerOperand() == pointer) { 1191 // Remove it! 1192 L->replaceAllUsesWith(S->getOperand(0)); 1193 toErase.push_back(L); 1194 deletedLoad = true; 1195 NumGVNLoad++; 1196 } 1197 1198 // Whether we removed it or not, we can't 1199 // go any further 1200 break; 1201 } else if (!isa<LoadInst>(DepInst)) { 1202 // Only want to handle loads below. 1203 break; 1204 } else if (!last) { 1205 // If we don't depend on a store, and we haven't 1206 // been loaded before, bail. 1207 break; 1208 } else if (DepInst == last) { 1209 // Remove it! 1210 L->replaceAllUsesWith(last); 1211 toErase.push_back(L); 1212 deletedLoad = true; 1213 NumGVNLoad++; 1214 break; 1215 } else { 1216 dep = MD->getDependencyFrom(L, DepInst, DepInst->getParent()); 1217 } 1218 } 1219 1220 // If this load really doesn't depend on anything, then we must be loading an 1221 // undef value. This can happen when loading for a fresh allocation with no 1222 // intervening stores, for example. 1223 if (dep.isNone()) { 1224 // If this load depends directly on an allocation, there isn't 1225 // anything stored there; therefore, we can optimize this load 1226 // to undef. 1227 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1228 toErase.push_back(L); 1229 deletedLoad = true; 1230 NumGVNLoad++; 1231 } 1232 1233 if (!deletedLoad) 1234 last = L; 1235 1236 return deletedLoad; 1237} 1238 1239Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) { 1240 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1241 if (I == localAvail.end()) 1242 return 0; 1243 1244 ValueNumberScope* locals = I->second; 1245 1246 while (locals) { 1247 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num); 1248 if (I != locals->table.end()) 1249 return I->second; 1250 else 1251 locals = locals->parent; 1252 } 1253 1254 return 0; 1255} 1256 1257/// processInstruction - When calculating availability, handle an instruction 1258/// by inserting it into the appropriate sets 1259bool GVN::processInstruction(Instruction *I, 1260 DenseMap<Value*, LoadInst*> &lastSeenLoad, 1261 SmallVectorImpl<Instruction*> &toErase) { 1262 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 1263 bool changed = processLoad(L, lastSeenLoad, toErase); 1264 1265 if (!changed) { 1266 unsigned num = VN.lookup_or_add(L); 1267 localAvail[I->getParent()]->table.insert(std::make_pair(num, L)); 1268 } 1269 1270 return changed; 1271 } 1272 1273 uint32_t nextNum = VN.getNextUnusedValueNumber(); 1274 unsigned num = VN.lookup_or_add(I); 1275 1276 // Allocations are always uniquely numbered, so we can save time and memory 1277 // by fast failing them. 1278 if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) { 1279 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1280 return false; 1281 } 1282 1283 // Collapse PHI nodes 1284 if (PHINode* p = dyn_cast<PHINode>(I)) { 1285 Value* constVal = CollapsePhi(p); 1286 1287 if (constVal) { 1288 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1289 PI != PE; ++PI) 1290 if (PI->second.count(p)) 1291 PI->second.erase(p); 1292 1293 p->replaceAllUsesWith(constVal); 1294 toErase.push_back(p); 1295 } else { 1296 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1297 } 1298 1299 // If the number we were assigned was a brand new VN, then we don't 1300 // need to do a lookup to see if the number already exists 1301 // somewhere in the domtree: it can't! 1302 } else if (num == nextNum) { 1303 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1304 1305 // Perform value-number based elimination 1306 } else if (Value* repl = lookupNumber(I->getParent(), num)) { 1307 // Remove it! 1308 VN.erase(I); 1309 I->replaceAllUsesWith(repl); 1310 toErase.push_back(I); 1311 return true; 1312 } else { 1313 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1314 } 1315 1316 return false; 1317} 1318 1319// GVN::runOnFunction - This is the main transformation entry point for a 1320// function. 1321// 1322bool GVN::runOnFunction(Function& F) { 1323 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1324 DT = &getAnalysis<DominatorTree>(); 1325 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1326 VN.setMemDep(MD); 1327 VN.setDomTree(DT); 1328 1329 bool changed = false; 1330 bool shouldContinue = true; 1331 1332 // Merge unconditional branches, allowing PRE to catch more 1333 // optimization opportunities. 1334 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1335 BasicBlock* BB = FI; 1336 ++FI; 1337 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1338 if (removedBlock) NumGVNBlocks++; 1339 1340 changed |= removedBlock; 1341 } 1342 1343 while (shouldContinue) { 1344 shouldContinue = iterateOnFunction(F); 1345 changed |= shouldContinue; 1346 } 1347 1348 if (EnablePRE) { 1349 bool PREChanged = true; 1350 while (PREChanged) { 1351 PREChanged = performPRE(F); 1352 changed |= PREChanged; 1353 } 1354 } 1355 1356 cleanupGlobalSets(); 1357 1358 return changed; 1359} 1360 1361 1362bool GVN::processBlock(DomTreeNode* DTN) { 1363 BasicBlock* BB = DTN->getBlock(); 1364 SmallVector<Instruction*, 8> toErase; 1365 DenseMap<Value*, LoadInst*> lastSeenLoad; 1366 bool changed_function = false; 1367 1368 if (DTN->getIDom()) 1369 localAvail[BB] = 1370 new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]); 1371 else 1372 localAvail[BB] = new ValueNumberScope(0); 1373 1374 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1375 BI != BE;) { 1376 changed_function |= processInstruction(BI, lastSeenLoad, toErase); 1377 if (toErase.empty()) { 1378 ++BI; 1379 continue; 1380 } 1381 1382 // If we need some instructions deleted, do it now. 1383 NumGVNInstr += toErase.size(); 1384 1385 // Avoid iterator invalidation. 1386 bool AtStart = BI == BB->begin(); 1387 if (!AtStart) 1388 --BI; 1389 1390 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1391 E = toErase.end(); I != E; ++I) { 1392 DEBUG(cerr << "GVN removed: " << **I); 1393 MD->removeInstruction(*I); 1394 (*I)->eraseFromParent(); 1395 } 1396 1397 if (AtStart) 1398 BI = BB->begin(); 1399 else 1400 ++BI; 1401 1402 toErase.clear(); 1403 } 1404 1405 return changed_function; 1406} 1407 1408/// performPRE - Perform a purely local form of PRE that looks for diamond 1409/// control flow patterns and attempts to perform simple PRE at the join point. 1410bool GVN::performPRE(Function& F) { 1411 bool Changed = false; 1412 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1413 DenseMap<BasicBlock*, Value*> predMap; 1414 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1415 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1416 BasicBlock* CurrentBlock = *DI; 1417 1418 // Nothing to PRE in the entry block. 1419 if (CurrentBlock == &F.getEntryBlock()) continue; 1420 1421 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1422 BE = CurrentBlock->end(); BI != BE; ) { 1423 Instruction *CurInst = BI++; 1424 1425 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) || 1426 isa<PHINode>(CurInst) || CurInst->mayReadFromMemory() || 1427 CurInst->mayWriteToMemory()) 1428 continue; 1429 1430 uint32_t valno = VN.lookup(CurInst); 1431 1432 // Look for the predecessors for PRE opportunities. We're 1433 // only trying to solve the basic diamond case, where 1434 // a value is computed in the successor and one predecessor, 1435 // but not the other. We also explicitly disallow cases 1436 // where the successor is its own predecessor, because they're 1437 // more complicated to get right. 1438 unsigned numWith = 0; 1439 unsigned numWithout = 0; 1440 BasicBlock* PREPred = 0; 1441 predMap.clear(); 1442 1443 for (pred_iterator PI = pred_begin(CurrentBlock), 1444 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1445 // We're not interested in PRE where the block is its 1446 // own predecessor, on in blocks with predecessors 1447 // that are not reachable. 1448 if (*PI == CurrentBlock) { 1449 numWithout = 2; 1450 break; 1451 } else if (!localAvail.count(*PI)) { 1452 numWithout = 2; 1453 break; 1454 } 1455 1456 DenseMap<uint32_t, Value*>::iterator predV = 1457 localAvail[*PI]->table.find(valno); 1458 if (predV == localAvail[*PI]->table.end()) { 1459 PREPred = *PI; 1460 numWithout++; 1461 } else if (predV->second == CurInst) { 1462 numWithout = 2; 1463 } else { 1464 predMap[*PI] = predV->second; 1465 numWith++; 1466 } 1467 } 1468 1469 // Don't do PRE when it might increase code size, i.e. when 1470 // we would need to insert instructions in more than one pred. 1471 if (numWithout != 1 || numWith == 0) 1472 continue; 1473 1474 // We can't do PRE safely on a critical edge, so instead we schedule 1475 // the edge to be split and perform the PRE the next time we iterate 1476 // on the function. 1477 unsigned succNum = 0; 1478 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1479 i != e; ++i) 1480 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1481 succNum = i; 1482 break; 1483 } 1484 1485 if (isCriticalEdge(PREPred->getTerminator(), succNum)) { 1486 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum)); 1487 Changed = true; 1488 continue; 1489 } 1490 1491 // Instantiate the expression the in predecessor that lacked it. 1492 // Because we are going top-down through the block, all value numbers 1493 // will be available in the predecessor by the time we need them. Any 1494 // that weren't original present will have been instantiated earlier 1495 // in this loop. 1496 Instruction* PREInstr = CurInst->clone(); 1497 bool success = true; 1498 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1499 Value *Op = PREInstr->getOperand(i); 1500 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1501 continue; 1502 1503 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1504 PREInstr->setOperand(i, V); 1505 } else { 1506 success = false; 1507 break; 1508 } 1509 } 1510 1511 // Fail out if we encounter an operand that is not available in 1512 // the PRE predecessor. This is typically because of loads which 1513 // are not value numbered precisely. 1514 if (!success) { 1515 delete PREInstr; 1516 continue; 1517 } 1518 1519 PREInstr->insertBefore(PREPred->getTerminator()); 1520 PREInstr->setName(CurInst->getName() + ".pre"); 1521 predMap[PREPred] = PREInstr; 1522 VN.add(PREInstr, valno); 1523 NumGVNPRE++; 1524 1525 // Update the availability map to include the new instruction. 1526 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr)); 1527 1528 // Create a PHI to make the value available in this block. 1529 PHINode* Phi = PHINode::Create(CurInst->getType(), 1530 CurInst->getName() + ".pre-phi", 1531 CurrentBlock->begin()); 1532 for (pred_iterator PI = pred_begin(CurrentBlock), 1533 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1534 Phi->addIncoming(predMap[*PI], *PI); 1535 1536 VN.add(Phi, valno); 1537 localAvail[CurrentBlock]->table[valno] = Phi; 1538 1539 CurInst->replaceAllUsesWith(Phi); 1540 VN.erase(CurInst); 1541 1542 DEBUG(cerr << "GVN PRE removed: " << *CurInst); 1543 MD->removeInstruction(CurInst); 1544 CurInst->eraseFromParent(); 1545 Changed = true; 1546 } 1547 } 1548 1549 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1550 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) { 1551 SplitCriticalEdge(I->first, I->second, this); 1552 BasicBlock* NewBlock = I->first->getSuccessor(I->second); 1553 localAvail[NewBlock] = 1554 new ValueNumberScope(localAvail[I->first->getParent()]); 1555 } 1556 1557 return Changed; 1558} 1559 1560// iterateOnFunction - Executes one iteration of GVN 1561bool GVN::iterateOnFunction(Function &F) { 1562 cleanupGlobalSets(); 1563 1564 // Top-down walk of the dominator tree 1565 bool changed = false; 1566 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1567 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1568 changed |= processBlock(*DI); 1569 1570 return changed; 1571} 1572 1573void GVN::cleanupGlobalSets() { 1574 VN.clear(); 1575 phiMap.clear(); 1576 1577 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1578 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1579 delete I->second; 1580 localAvail.clear(); 1581} 1582