GVN.cpp revision 771a5422e167200f8a750449f688215f9abe925a
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/IntrinsicInst.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Value.h" 27#include "llvm/ADT/DenseMap.h" 28#include "llvm/ADT/DepthFirstIterator.h" 29#include "llvm/ADT/PostOrderIterator.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/Analysis/Dominators.h" 34#include "llvm/Analysis/AliasAnalysis.h" 35#include "llvm/Analysis/MallocHelper.h" 36#include "llvm/Analysis/MemoryDependenceAnalysis.h" 37#include "llvm/Support/CFG.h" 38#include "llvm/Support/CommandLine.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/ErrorHandling.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/Target/TargetData.h" 43#include "llvm/Transforms/Utils/BasicBlockUtils.h" 44#include "llvm/Transforms/Utils/Local.h" 45#include <cstdio> 46using namespace llvm; 47 48STATISTIC(NumGVNInstr, "Number of instructions deleted"); 49STATISTIC(NumGVNLoad, "Number of loads deleted"); 50STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 51STATISTIC(NumGVNBlocks, "Number of blocks merged"); 52STATISTIC(NumPRELoad, "Number of loads PRE'd"); 53 54static cl::opt<bool> EnablePRE("enable-pre", 55 cl::init(true), cl::Hidden); 56static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 57 58//===----------------------------------------------------------------------===// 59// ValueTable Class 60//===----------------------------------------------------------------------===// 61 62/// This class holds the mapping between values and value numbers. It is used 63/// as an efficient mechanism to determine the expression-wise equivalence of 64/// two values. 65namespace { 66 struct Expression { 67 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 68 UDIV, SDIV, FDIV, UREM, SREM, 69 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 70 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 71 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 72 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 73 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 74 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 75 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 76 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 77 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 78 EMPTY, TOMBSTONE }; 79 80 ExpressionOpcode opcode; 81 const Type* type; 82 uint32_t firstVN; 83 uint32_t secondVN; 84 uint32_t thirdVN; 85 SmallVector<uint32_t, 4> varargs; 86 Value* function; 87 88 Expression() { } 89 Expression(ExpressionOpcode o) : opcode(o) { } 90 91 bool operator==(const Expression &other) const { 92 if (opcode != other.opcode) 93 return false; 94 else if (opcode == EMPTY || opcode == TOMBSTONE) 95 return true; 96 else if (type != other.type) 97 return false; 98 else if (function != other.function) 99 return false; 100 else if (firstVN != other.firstVN) 101 return false; 102 else if (secondVN != other.secondVN) 103 return false; 104 else if (thirdVN != other.thirdVN) 105 return false; 106 else { 107 if (varargs.size() != other.varargs.size()) 108 return false; 109 110 for (size_t i = 0; i < varargs.size(); ++i) 111 if (varargs[i] != other.varargs[i]) 112 return false; 113 114 return true; 115 } 116 } 117 118 bool operator!=(const Expression &other) const { 119 return !(*this == other); 120 } 121 }; 122 123 class ValueTable { 124 private: 125 DenseMap<Value*, uint32_t> valueNumbering; 126 DenseMap<Expression, uint32_t> expressionNumbering; 127 AliasAnalysis* AA; 128 MemoryDependenceAnalysis* MD; 129 DominatorTree* DT; 130 131 uint32_t nextValueNumber; 132 133 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 134 Expression::ExpressionOpcode getOpcode(CmpInst* C); 135 Expression::ExpressionOpcode getOpcode(CastInst* C); 136 Expression create_expression(BinaryOperator* BO); 137 Expression create_expression(CmpInst* C); 138 Expression create_expression(ShuffleVectorInst* V); 139 Expression create_expression(ExtractElementInst* C); 140 Expression create_expression(InsertElementInst* V); 141 Expression create_expression(SelectInst* V); 142 Expression create_expression(CastInst* C); 143 Expression create_expression(GetElementPtrInst* G); 144 Expression create_expression(CallInst* C); 145 Expression create_expression(Constant* C); 146 public: 147 ValueTable() : nextValueNumber(1) { } 148 uint32_t lookup_or_add(Value* V); 149 uint32_t lookup(Value* V) const; 150 void add(Value* V, uint32_t num); 151 void clear(); 152 void erase(Value* v); 153 unsigned size(); 154 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 155 AliasAnalysis *getAliasAnalysis() const { return AA; } 156 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 157 void setDomTree(DominatorTree* D) { DT = D; } 158 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 159 void verifyRemoved(const Value *) const; 160 }; 161} 162 163namespace llvm { 164template <> struct DenseMapInfo<Expression> { 165 static inline Expression getEmptyKey() { 166 return Expression(Expression::EMPTY); 167 } 168 169 static inline Expression getTombstoneKey() { 170 return Expression(Expression::TOMBSTONE); 171 } 172 173 static unsigned getHashValue(const Expression e) { 174 unsigned hash = e.opcode; 175 176 hash = e.firstVN + hash * 37; 177 hash = e.secondVN + hash * 37; 178 hash = e.thirdVN + hash * 37; 179 180 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 181 (unsigned)((uintptr_t)e.type >> 9)) + 182 hash * 37; 183 184 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 185 E = e.varargs.end(); I != E; ++I) 186 hash = *I + hash * 37; 187 188 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 189 (unsigned)((uintptr_t)e.function >> 9)) + 190 hash * 37; 191 192 return hash; 193 } 194 static bool isEqual(const Expression &LHS, const Expression &RHS) { 195 return LHS == RHS; 196 } 197 static bool isPod() { return true; } 198}; 199} 200 201//===----------------------------------------------------------------------===// 202// ValueTable Internal Functions 203//===----------------------------------------------------------------------===// 204Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 205 switch(BO->getOpcode()) { 206 default: // THIS SHOULD NEVER HAPPEN 207 llvm_unreachable("Binary operator with unknown opcode?"); 208 case Instruction::Add: return Expression::ADD; 209 case Instruction::FAdd: return Expression::FADD; 210 case Instruction::Sub: return Expression::SUB; 211 case Instruction::FSub: return Expression::FSUB; 212 case Instruction::Mul: return Expression::MUL; 213 case Instruction::FMul: return Expression::FMUL; 214 case Instruction::UDiv: return Expression::UDIV; 215 case Instruction::SDiv: return Expression::SDIV; 216 case Instruction::FDiv: return Expression::FDIV; 217 case Instruction::URem: return Expression::UREM; 218 case Instruction::SRem: return Expression::SREM; 219 case Instruction::FRem: return Expression::FREM; 220 case Instruction::Shl: return Expression::SHL; 221 case Instruction::LShr: return Expression::LSHR; 222 case Instruction::AShr: return Expression::ASHR; 223 case Instruction::And: return Expression::AND; 224 case Instruction::Or: return Expression::OR; 225 case Instruction::Xor: return Expression::XOR; 226 } 227} 228 229Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 230 if (isa<ICmpInst>(C)) { 231 switch (C->getPredicate()) { 232 default: // THIS SHOULD NEVER HAPPEN 233 llvm_unreachable("Comparison with unknown predicate?"); 234 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 235 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 236 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 237 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 238 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 239 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 240 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 241 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 242 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 243 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 244 } 245 } else { 246 switch (C->getPredicate()) { 247 default: // THIS SHOULD NEVER HAPPEN 248 llvm_unreachable("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} 266 267Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 268 switch(C->getOpcode()) { 269 default: // THIS SHOULD NEVER HAPPEN 270 llvm_unreachable("Cast operator with unknown opcode?"); 271 case Instruction::Trunc: return Expression::TRUNC; 272 case Instruction::ZExt: return Expression::ZEXT; 273 case Instruction::SExt: return Expression::SEXT; 274 case Instruction::FPToUI: return Expression::FPTOUI; 275 case Instruction::FPToSI: return Expression::FPTOSI; 276 case Instruction::UIToFP: return Expression::UITOFP; 277 case Instruction::SIToFP: return Expression::SITOFP; 278 case Instruction::FPTrunc: return Expression::FPTRUNC; 279 case Instruction::FPExt: return Expression::FPEXT; 280 case Instruction::PtrToInt: return Expression::PTRTOINT; 281 case Instruction::IntToPtr: return Expression::INTTOPTR; 282 case Instruction::BitCast: return Expression::BITCAST; 283 } 284} 285 286Expression ValueTable::create_expression(CallInst* C) { 287 Expression e; 288 289 e.type = C->getType(); 290 e.firstVN = 0; 291 e.secondVN = 0; 292 e.thirdVN = 0; 293 e.function = C->getCalledFunction(); 294 e.opcode = Expression::CALL; 295 296 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 297 I != E; ++I) 298 e.varargs.push_back(lookup_or_add(*I)); 299 300 return e; 301} 302 303Expression ValueTable::create_expression(BinaryOperator* BO) { 304 Expression e; 305 306 e.firstVN = lookup_or_add(BO->getOperand(0)); 307 e.secondVN = lookup_or_add(BO->getOperand(1)); 308 e.thirdVN = 0; 309 e.function = 0; 310 e.type = BO->getType(); 311 e.opcode = getOpcode(BO); 312 313 return e; 314} 315 316Expression ValueTable::create_expression(CmpInst* C) { 317 Expression e; 318 319 e.firstVN = lookup_or_add(C->getOperand(0)); 320 e.secondVN = lookup_or_add(C->getOperand(1)); 321 e.thirdVN = 0; 322 e.function = 0; 323 e.type = C->getType(); 324 e.opcode = getOpcode(C); 325 326 return e; 327} 328 329Expression ValueTable::create_expression(CastInst* C) { 330 Expression e; 331 332 e.firstVN = lookup_or_add(C->getOperand(0)); 333 e.secondVN = 0; 334 e.thirdVN = 0; 335 e.function = 0; 336 e.type = C->getType(); 337 e.opcode = getOpcode(C); 338 339 return e; 340} 341 342Expression ValueTable::create_expression(ShuffleVectorInst* S) { 343 Expression e; 344 345 e.firstVN = lookup_or_add(S->getOperand(0)); 346 e.secondVN = lookup_or_add(S->getOperand(1)); 347 e.thirdVN = lookup_or_add(S->getOperand(2)); 348 e.function = 0; 349 e.type = S->getType(); 350 e.opcode = Expression::SHUFFLE; 351 352 return e; 353} 354 355Expression ValueTable::create_expression(ExtractElementInst* E) { 356 Expression e; 357 358 e.firstVN = lookup_or_add(E->getOperand(0)); 359 e.secondVN = lookup_or_add(E->getOperand(1)); 360 e.thirdVN = 0; 361 e.function = 0; 362 e.type = E->getType(); 363 e.opcode = Expression::EXTRACT; 364 365 return e; 366} 367 368Expression ValueTable::create_expression(InsertElementInst* I) { 369 Expression e; 370 371 e.firstVN = lookup_or_add(I->getOperand(0)); 372 e.secondVN = lookup_or_add(I->getOperand(1)); 373 e.thirdVN = lookup_or_add(I->getOperand(2)); 374 e.function = 0; 375 e.type = I->getType(); 376 e.opcode = Expression::INSERT; 377 378 return e; 379} 380 381Expression ValueTable::create_expression(SelectInst* I) { 382 Expression e; 383 384 e.firstVN = lookup_or_add(I->getCondition()); 385 e.secondVN = lookup_or_add(I->getTrueValue()); 386 e.thirdVN = lookup_or_add(I->getFalseValue()); 387 e.function = 0; 388 e.type = I->getType(); 389 e.opcode = Expression::SELECT; 390 391 return e; 392} 393 394Expression ValueTable::create_expression(GetElementPtrInst* G) { 395 Expression e; 396 397 e.firstVN = lookup_or_add(G->getPointerOperand()); 398 e.secondVN = 0; 399 e.thirdVN = 0; 400 e.function = 0; 401 e.type = G->getType(); 402 e.opcode = Expression::GEP; 403 404 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 405 I != E; ++I) 406 e.varargs.push_back(lookup_or_add(*I)); 407 408 return e; 409} 410 411//===----------------------------------------------------------------------===// 412// ValueTable External Functions 413//===----------------------------------------------------------------------===// 414 415/// add - Insert a value into the table with a specified value number. 416void ValueTable::add(Value* V, uint32_t num) { 417 valueNumbering.insert(std::make_pair(V, num)); 418} 419 420/// lookup_or_add - Returns the value number for the specified value, assigning 421/// it a new number if it did not have one before. 422uint32_t ValueTable::lookup_or_add(Value* V) { 423 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 424 if (VI != valueNumbering.end()) 425 return VI->second; 426 427 if (CallInst* C = dyn_cast<CallInst>(V)) { 428 if (AA->doesNotAccessMemory(C)) { 429 Expression e = create_expression(C); 430 431 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 432 if (EI != expressionNumbering.end()) { 433 valueNumbering.insert(std::make_pair(V, EI->second)); 434 return EI->second; 435 } else { 436 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 437 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 438 439 return nextValueNumber++; 440 } 441 } else if (AA->onlyReadsMemory(C)) { 442 Expression e = create_expression(C); 443 444 if (expressionNumbering.find(e) == expressionNumbering.end()) { 445 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 446 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 447 return nextValueNumber++; 448 } 449 450 MemDepResult local_dep = MD->getDependency(C); 451 452 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 453 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 454 return nextValueNumber++; 455 } 456 457 if (local_dep.isDef()) { 458 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 459 460 if (local_cdep->getNumOperands() != C->getNumOperands()) { 461 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 462 return nextValueNumber++; 463 } 464 465 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 // Non-local case. 480 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 481 MD->getNonLocalCallDependency(CallSite(C)); 482 // FIXME: call/call dependencies for readonly calls should return def, not 483 // clobber! Move the checking logic to MemDep! 484 CallInst* cdep = 0; 485 486 // Check to see if we have a single dominating call instruction that is 487 // identical to C. 488 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 489 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 490 // Ignore non-local dependencies. 491 if (I->second.isNonLocal()) 492 continue; 493 494 // We don't handle non-depedencies. If we already have a call, reject 495 // instruction dependencies. 496 if (I->second.isClobber() || cdep != 0) { 497 cdep = 0; 498 break; 499 } 500 501 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 502 // FIXME: All duplicated with non-local case. 503 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 504 cdep = NonLocalDepCall; 505 continue; 506 } 507 508 cdep = 0; 509 break; 510 } 511 512 if (!cdep) { 513 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 514 return nextValueNumber++; 515 } 516 517 if (cdep->getNumOperands() != C->getNumOperands()) { 518 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 519 return nextValueNumber++; 520 } 521 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 522 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 523 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 524 if (c_vn != cd_vn) { 525 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 526 return nextValueNumber++; 527 } 528 } 529 530 uint32_t v = lookup_or_add(cdep); 531 valueNumbering.insert(std::make_pair(V, v)); 532 return v; 533 534 } else { 535 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 536 return nextValueNumber++; 537 } 538 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 539 Expression e = create_expression(BO); 540 541 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 542 if (EI != expressionNumbering.end()) { 543 valueNumbering.insert(std::make_pair(V, EI->second)); 544 return EI->second; 545 } else { 546 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 547 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 548 549 return nextValueNumber++; 550 } 551 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 552 Expression e = create_expression(C); 553 554 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 555 if (EI != expressionNumbering.end()) { 556 valueNumbering.insert(std::make_pair(V, EI->second)); 557 return EI->second; 558 } else { 559 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 560 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 561 562 return nextValueNumber++; 563 } 564 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 565 Expression e = create_expression(U); 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 (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 578 Expression e = create_expression(U); 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 (InsertElementInst* U = dyn_cast<InsertElementInst>(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 (SelectInst* U = dyn_cast<SelectInst>(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 (CastInst* U = dyn_cast<CastInst>(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 (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(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 { 643 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 644 return nextValueNumber++; 645 } 646} 647 648/// lookup - Returns the value number of the specified value. Fails if 649/// the value has not yet been numbered. 650uint32_t ValueTable::lookup(Value* V) const { 651 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 652 assert(VI != valueNumbering.end() && "Value not numbered?"); 653 return VI->second; 654} 655 656/// clear - Remove all entries from the ValueTable 657void ValueTable::clear() { 658 valueNumbering.clear(); 659 expressionNumbering.clear(); 660 nextValueNumber = 1; 661} 662 663/// erase - Remove a value from the value numbering 664void ValueTable::erase(Value* V) { 665 valueNumbering.erase(V); 666} 667 668/// verifyRemoved - Verify that the value is removed from all internal data 669/// structures. 670void ValueTable::verifyRemoved(const Value *V) const { 671 for (DenseMap<Value*, uint32_t>::iterator 672 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 673 assert(I->first != V && "Inst still occurs in value numbering map!"); 674 } 675} 676 677//===----------------------------------------------------------------------===// 678// GVN Pass 679//===----------------------------------------------------------------------===// 680 681namespace { 682 struct ValueNumberScope { 683 ValueNumberScope* parent; 684 DenseMap<uint32_t, Value*> table; 685 686 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 687 }; 688} 689 690namespace { 691 692 class GVN : public FunctionPass { 693 bool runOnFunction(Function &F); 694 public: 695 static char ID; // Pass identification, replacement for typeid 696 GVN() : FunctionPass(&ID) { } 697 698 private: 699 MemoryDependenceAnalysis *MD; 700 DominatorTree *DT; 701 702 ValueTable VN; 703 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 704 705 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 706 PhiMapType phiMap; 707 708 709 // This transformation requires dominator postdominator info 710 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 711 AU.addRequired<DominatorTree>(); 712 AU.addRequired<MemoryDependenceAnalysis>(); 713 AU.addRequired<AliasAnalysis>(); 714 715 AU.addPreserved<DominatorTree>(); 716 AU.addPreserved<AliasAnalysis>(); 717 } 718 719 // Helper fuctions 720 // FIXME: eliminate or document these better 721 bool processLoad(LoadInst* L, 722 SmallVectorImpl<Instruction*> &toErase); 723 bool processInstruction(Instruction* I, 724 SmallVectorImpl<Instruction*> &toErase); 725 bool processNonLocalLoad(LoadInst* L, 726 SmallVectorImpl<Instruction*> &toErase); 727 bool processBlock(BasicBlock* BB); 728 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig, 729 DenseMap<BasicBlock*, Value*> &Phis, 730 bool top_level = false); 731 void dump(DenseMap<uint32_t, Value*>& d); 732 bool iterateOnFunction(Function &F); 733 Value* CollapsePhi(PHINode* p); 734 bool performPRE(Function& F); 735 Value* lookupNumber(BasicBlock* BB, uint32_t num); 736 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno); 737 void cleanupGlobalSets(); 738 void verifyRemoved(const Instruction *I) const; 739 }; 740 741 char GVN::ID = 0; 742} 743 744// createGVNPass - The public interface to this file... 745FunctionPass *llvm::createGVNPass() { return new GVN(); } 746 747static RegisterPass<GVN> X("gvn", 748 "Global Value Numbering"); 749 750void GVN::dump(DenseMap<uint32_t, Value*>& d) { 751 printf("{\n"); 752 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 753 E = d.end(); I != E; ++I) { 754 printf("%d\n", I->first); 755 I->second->dump(); 756 } 757 printf("}\n"); 758} 759 760static bool isSafeReplacement(PHINode* p, Instruction* inst) { 761 if (!isa<PHINode>(inst)) 762 return true; 763 764 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 765 UI != E; ++UI) 766 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 767 if (use_phi->getParent() == inst->getParent()) 768 return false; 769 770 return true; 771} 772 773Value* GVN::CollapsePhi(PHINode* p) { 774 Value* constVal = p->hasConstantValue(DT); 775 if (!constVal) return 0; 776 777 Instruction* inst = dyn_cast<Instruction>(constVal); 778 if (!inst) 779 return constVal; 780 781 if (DT->dominates(inst, p)) 782 if (isSafeReplacement(p, inst)) 783 return inst; 784 return 0; 785} 786 787/// GetValueForBlock - Get the value to use within the specified basic block. 788/// available values are in Phis. 789Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig, 790 DenseMap<BasicBlock*, Value*> &Phis, 791 bool top_level) { 792 793 // If we have already computed this value, return the previously computed val. 794 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 795 if (V != Phis.end() && !top_level) return V->second; 796 797 // If the block is unreachable, just return undef, since this path 798 // can't actually occur at runtime. 799 if (!DT->isReachableFromEntry(BB)) 800 return Phis[BB] = UndefValue::get(orig->getType()); 801 802 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 803 Value *ret = GetValueForBlock(Pred, orig, Phis); 804 Phis[BB] = ret; 805 return ret; 806 } 807 808 // Get the number of predecessors of this block so we can reserve space later. 809 // If there is already a PHI in it, use the #preds from it, otherwise count. 810 // Getting it from the PHI is constant time. 811 unsigned NumPreds; 812 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin())) 813 NumPreds = ExistingPN->getNumIncomingValues(); 814 else 815 NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 816 817 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 818 // now, then get values to fill in the incoming values for the PHI. 819 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 820 BB->begin()); 821 PN->reserveOperandSpace(NumPreds); 822 823 Phis.insert(std::make_pair(BB, PN)); 824 825 // Fill in the incoming values for the block. 826 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 827 Value* val = GetValueForBlock(*PI, orig, Phis); 828 PN->addIncoming(val, *PI); 829 } 830 831 VN.getAliasAnalysis()->copyValue(orig, PN); 832 833 // Attempt to collapse PHI nodes that are trivially redundant 834 Value* v = CollapsePhi(PN); 835 if (!v) { 836 // Cache our phi construction results 837 if (LoadInst* L = dyn_cast<LoadInst>(orig)) 838 phiMap[L->getPointerOperand()].insert(PN); 839 else 840 phiMap[orig].insert(PN); 841 842 return PN; 843 } 844 845 PN->replaceAllUsesWith(v); 846 if (isa<PointerType>(v->getType())) 847 MD->invalidateCachedPointerInfo(v); 848 849 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 850 E = Phis.end(); I != E; ++I) 851 if (I->second == PN) 852 I->second = v; 853 854 DEBUG(errs() << "GVN removed: " << *PN << '\n'); 855 MD->removeInstruction(PN); 856 PN->eraseFromParent(); 857 DEBUG(verifyRemoved(PN)); 858 859 Phis[BB] = v; 860 return v; 861} 862 863/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 864/// we're analyzing is fully available in the specified block. As we go, keep 865/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 866/// map is actually a tri-state map with the following values: 867/// 0) we know the block *is not* fully available. 868/// 1) we know the block *is* fully available. 869/// 2) we do not know whether the block is fully available or not, but we are 870/// currently speculating that it will be. 871/// 3) we are speculating for this block and have used that to speculate for 872/// other blocks. 873static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 874 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 875 // Optimistically assume that the block is fully available and check to see 876 // if we already know about this block in one lookup. 877 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 878 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 879 880 // If the entry already existed for this block, return the precomputed value. 881 if (!IV.second) { 882 // If this is a speculative "available" value, mark it as being used for 883 // speculation of other blocks. 884 if (IV.first->second == 2) 885 IV.first->second = 3; 886 return IV.first->second != 0; 887 } 888 889 // Otherwise, see if it is fully available in all predecessors. 890 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 891 892 // If this block has no predecessors, it isn't live-in here. 893 if (PI == PE) 894 goto SpeculationFailure; 895 896 for (; PI != PE; ++PI) 897 // If the value isn't fully available in one of our predecessors, then it 898 // isn't fully available in this block either. Undo our previous 899 // optimistic assumption and bail out. 900 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 901 goto SpeculationFailure; 902 903 return true; 904 905// SpeculationFailure - If we get here, we found out that this is not, after 906// all, a fully-available block. We have a problem if we speculated on this and 907// used the speculation to mark other blocks as available. 908SpeculationFailure: 909 char &BBVal = FullyAvailableBlocks[BB]; 910 911 // If we didn't speculate on this, just return with it set to false. 912 if (BBVal == 2) { 913 BBVal = 0; 914 return false; 915 } 916 917 // If we did speculate on this value, we could have blocks set to 1 that are 918 // incorrect. Walk the (transitive) successors of this block and mark them as 919 // 0 if set to one. 920 SmallVector<BasicBlock*, 32> BBWorklist; 921 BBWorklist.push_back(BB); 922 923 while (!BBWorklist.empty()) { 924 BasicBlock *Entry = BBWorklist.pop_back_val(); 925 // Note that this sets blocks to 0 (unavailable) if they happen to not 926 // already be in FullyAvailableBlocks. This is safe. 927 char &EntryVal = FullyAvailableBlocks[Entry]; 928 if (EntryVal == 0) continue; // Already unavailable. 929 930 // Mark as unavailable. 931 EntryVal = 0; 932 933 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 934 BBWorklist.push_back(*I); 935 } 936 937 return false; 938} 939 940 941/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 942/// then a load from a must-aliased pointer of a different type, try to coerce 943/// the stored value. LoadedTy is the type of the load we want to replace and 944/// InsertPt is the place to insert new instructions. 945/// 946/// If we can't do it, return null. 947static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 948 const Type *LoadedTy, 949 Instruction *InsertPt, 950 const TargetData &TD) { 951 const Type *StoredValTy = StoredVal->getType(); 952 953 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 954 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 955 956 // If the store and reload are the same size, we can always reuse it. 957 if (StoreSize == LoadSize) { 958 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) { 959 // Pointer to Pointer -> use bitcast. 960 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 961 } 962 963 // Convert source pointers to integers, which can be bitcast. 964 if (isa<PointerType>(StoredValTy)) { 965 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 966 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 967 } 968 969 const Type *TypeToCastTo = LoadedTy; 970 if (isa<PointerType>(TypeToCastTo)) 971 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 972 973 if (StoredValTy != TypeToCastTo) 974 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 975 976 // Cast to pointer if the load needs a pointer type. 977 if (isa<PointerType>(LoadedTy)) 978 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 979 980 return StoredVal; 981 } 982 983 // If the loaded value is smaller than the available value, then we can 984 // extract out a piece from it. If the available value is too small, then we 985 // can't do anything. 986 if (StoreSize < LoadSize) 987 return 0; 988 989 // Convert source pointers to integers, which can be manipulated. 990 if (isa<PointerType>(StoredValTy)) { 991 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 992 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 993 } 994 995 // Convert vectors and fp to integer, which can be manipulated. 996 if (!isa<IntegerType>(StoredValTy)) { 997 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 998 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 999 } 1000 1001 // If this is a big-endian system, we need to shift the value down to the low 1002 // bits so that a truncate will work. 1003 if (TD.isBigEndian()) { 1004 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 1005 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 1006 } 1007 1008 // Truncate the integer to the right size now. 1009 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 1010 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 1011 1012 if (LoadedTy == NewIntTy) 1013 return StoredVal; 1014 1015 // If the result is a pointer, inttoptr. 1016 if (isa<PointerType>(LoadedTy)) 1017 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 1018 1019 // Otherwise, bitcast. 1020 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 1021} 1022 1023static void 1024GetAvailableBlockValues(DenseMap<BasicBlock*, Value*> &BlockReplValues, 1025 SmallVector<std::pair<BasicBlock*, 1026 Value*>, 16> &ValuesPerBlock, 1027 const Type *LoadTy, 1028 const TargetData *TD) { 1029 1030 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1031 BasicBlock *BB = ValuesPerBlock[i].first; 1032 Value *AvailableVal = ValuesPerBlock[i].second; 1033 1034 Value *&BlockEntry = BlockReplValues[BB]; 1035 if (BlockEntry) continue; 1036 1037 if (AvailableVal->getType() != LoadTy) { 1038 assert(TD && "Need target data to handle type mismatch case"); 1039 AvailableVal = CoerceAvailableValueToLoadType(AvailableVal, LoadTy, 1040 BB->getTerminator(), *TD); 1041 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1042 << *ValuesPerBlock[i].second << '\n' 1043 << *AvailableVal << '\n' << "\n\n\n"); 1044 } 1045 BlockEntry = AvailableVal; 1046 } 1047} 1048 1049/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1050/// non-local by performing PHI construction. 1051bool GVN::processNonLocalLoad(LoadInst *LI, 1052 SmallVectorImpl<Instruction*> &toErase) { 1053 // Find the non-local dependencies of the load. 1054 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 1055 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1056 Deps); 1057 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: " 1058 // << Deps.size() << *LI << '\n'); 1059 1060 // If we had to process more than one hundred blocks to find the 1061 // dependencies, this load isn't worth worrying about. Optimizing 1062 // it will be too expensive. 1063 if (Deps.size() > 100) 1064 return false; 1065 1066 // If we had a phi translation failure, we'll have a single entry which is a 1067 // clobber in the current block. Reject this early. 1068 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 1069 DEBUG( 1070 errs() << "GVN: non-local load "; 1071 WriteAsOperand(errs(), LI); 1072 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n'; 1073 ); 1074 return false; 1075 } 1076 1077 // Filter out useless results (non-locals, etc). Keep track of the blocks 1078 // where we have a value available in repl, also keep track of whether we see 1079 // dependencies that produce an unknown value for the load (such as a call 1080 // that could potentially clobber the load). 1081 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 1082 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1083 1084 const TargetData *TD = 0; 1085 1086 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1087 BasicBlock *DepBB = Deps[i].first; 1088 MemDepResult DepInfo = Deps[i].second; 1089 1090 if (DepInfo.isClobber()) { 1091 UnavailableBlocks.push_back(DepBB); 1092 continue; 1093 } 1094 1095 Instruction *DepInst = DepInfo.getInst(); 1096 1097 // Loading the allocation -> undef. 1098 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1099 ValuesPerBlock.push_back(std::make_pair(DepBB, 1100 UndefValue::get(LI->getType()))); 1101 continue; 1102 } 1103 1104 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 1105 // Reject loads and stores that are to the same address but are of 1106 // different types if we have to. 1107 if (S->getOperand(0)->getType() != LI->getType()) { 1108 if (TD == 0) 1109 TD = getAnalysisIfAvailable<TargetData>(); 1110 1111 // If the stored value is larger or equal to the loaded value, we can 1112 // reuse it. 1113 if (TD == 0 || 1114 TD->getTypeSizeInBits(S->getOperand(0)->getType()) < 1115 TD->getTypeSizeInBits(LI->getType())) { 1116 UnavailableBlocks.push_back(DepBB); 1117 continue; 1118 } 1119 } 1120 1121 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1122 1123 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) { 1124 // If the types mismatch and we can't handle it, reject reuse of the load. 1125 if (LD->getType() != LI->getType()) { 1126 if (TD == 0) 1127 TD = getAnalysisIfAvailable<TargetData>(); 1128 1129 // If the stored value is larger or equal to the loaded value, we can 1130 // reuse it. 1131 if (TD == 0 || 1132 TD->getTypeSizeInBits(LD->getType()) < 1133 TD->getTypeSizeInBits(LI->getType())) { 1134 UnavailableBlocks.push_back(DepBB); 1135 continue; 1136 } 1137 } 1138 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1139 } else { 1140 // FIXME: Handle memset/memcpy. 1141 UnavailableBlocks.push_back(DepBB); 1142 continue; 1143 } 1144 } 1145 1146 // If we have no predecessors that produce a known value for this load, exit 1147 // early. 1148 if (ValuesPerBlock.empty()) return false; 1149 1150 // If all of the instructions we depend on produce a known value for this 1151 // load, then it is fully redundant and we can use PHI insertion to compute 1152 // its value. Insert PHIs and remove the fully redundant value now. 1153 if (UnavailableBlocks.empty()) { 1154 // Use cached PHI construction information from previous runs 1155 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1156 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated? 1157 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1158 I != E; ++I) { 1159 if ((*I)->getParent() == LI->getParent()) { 1160 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD #1: " << *LI << '\n'); 1161 LI->replaceAllUsesWith(*I); 1162 if (isa<PointerType>((*I)->getType())) 1163 MD->invalidateCachedPointerInfo(*I); 1164 toErase.push_back(LI); 1165 NumGVNLoad++; 1166 return true; 1167 } 1168 1169 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1170 } 1171 1172 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1173 1174 // Convert the block information to a map, and insert coersions as needed. 1175 DenseMap<BasicBlock*, Value*> BlockReplValues; 1176 GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD); 1177 1178 // Perform PHI construction. 1179 Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1180 LI->replaceAllUsesWith(V); 1181 1182 if (isa<PHINode>(V)) 1183 V->takeName(LI); 1184 if (isa<PointerType>(V->getType())) 1185 MD->invalidateCachedPointerInfo(V); 1186 toErase.push_back(LI); 1187 NumGVNLoad++; 1188 return true; 1189 } 1190 1191 if (!EnablePRE || !EnableLoadPRE) 1192 return false; 1193 1194 // Okay, we have *some* definitions of the value. This means that the value 1195 // is available in some of our (transitive) predecessors. Lets think about 1196 // doing PRE of this load. This will involve inserting a new load into the 1197 // predecessor when it's not available. We could do this in general, but 1198 // prefer to not increase code size. As such, we only do this when we know 1199 // that we only have to insert *one* load (which means we're basically moving 1200 // the load, not inserting a new one). 1201 1202 SmallPtrSet<BasicBlock *, 4> Blockers; 1203 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1204 Blockers.insert(UnavailableBlocks[i]); 1205 1206 // Lets find first basic block with more than one predecessor. Walk backwards 1207 // through predecessors if needed. 1208 BasicBlock *LoadBB = LI->getParent(); 1209 BasicBlock *TmpBB = LoadBB; 1210 1211 bool isSinglePred = false; 1212 bool allSingleSucc = true; 1213 while (TmpBB->getSinglePredecessor()) { 1214 isSinglePred = true; 1215 TmpBB = TmpBB->getSinglePredecessor(); 1216 if (!TmpBB) // If haven't found any, bail now. 1217 return false; 1218 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1219 return false; 1220 if (Blockers.count(TmpBB)) 1221 return false; 1222 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1223 allSingleSucc = false; 1224 } 1225 1226 assert(TmpBB); 1227 LoadBB = TmpBB; 1228 1229 // If we have a repl set with LI itself in it, this means we have a loop where 1230 // at least one of the values is LI. Since this means that we won't be able 1231 // to eliminate LI even if we insert uses in the other predecessors, we will 1232 // end up increasing code size. Reject this by scanning for LI. 1233 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1234 if (ValuesPerBlock[i].second == LI) 1235 return false; 1236 1237 if (isSinglePred) { 1238 bool isHot = false; 1239 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1240 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second)) 1241 // "Hot" Instruction is in some loop (because it dominates its dep. 1242 // instruction). 1243 if (DT->dominates(LI, I)) { 1244 isHot = true; 1245 break; 1246 } 1247 1248 // We are interested only in "hot" instructions. We don't want to do any 1249 // mis-optimizations here. 1250 if (!isHot) 1251 return false; 1252 } 1253 1254 // Okay, we have some hope :). Check to see if the loaded value is fully 1255 // available in all but one predecessor. 1256 // FIXME: If we could restructure the CFG, we could make a common pred with 1257 // all the preds that don't have an available LI and insert a new load into 1258 // that one block. 1259 BasicBlock *UnavailablePred = 0; 1260 1261 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1262 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1263 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1264 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1265 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1266 1267 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1268 PI != E; ++PI) { 1269 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1270 continue; 1271 1272 // If this load is not available in multiple predecessors, reject it. 1273 if (UnavailablePred && UnavailablePred != *PI) 1274 return false; 1275 UnavailablePred = *PI; 1276 } 1277 1278 assert(UnavailablePred != 0 && 1279 "Fully available value should be eliminated above!"); 1280 1281 // If the loaded pointer is PHI node defined in this block, do PHI translation 1282 // to get its value in the predecessor. 1283 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1284 1285 // Make sure the value is live in the predecessor. If it was defined by a 1286 // non-PHI instruction in this block, we don't know how to recompute it above. 1287 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1288 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1289 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1290 << *LPInst << '\n' << *LI << "\n"); 1291 return false; 1292 } 1293 1294 // We don't currently handle critical edges :( 1295 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1296 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1297 << UnavailablePred->getName() << "': " << *LI << '\n'); 1298 return false; 1299 } 1300 1301 // Make sure it is valid to move this load here. We have to watch out for: 1302 // @1 = getelementptr (i8* p, ... 1303 // test p and branch if == 0 1304 // load @1 1305 // It is valid to have the getelementptr before the test, even if p can be 0, 1306 // as getelementptr only does address arithmetic. 1307 // If we are not pushing the value through any multiple-successor blocks 1308 // we do not have this case. Otherwise, check that the load is safe to 1309 // put anywhere; this can be improved, but should be conservatively safe. 1310 if (!allSingleSucc && 1311 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1312 return false; 1313 1314 // Okay, we can eliminate this load by inserting a reload in the predecessor 1315 // and using PHI construction to get the value in the other predecessors, do 1316 // it. 1317 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1318 1319 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1320 LI->getAlignment(), 1321 UnavailablePred->getTerminator()); 1322 1323 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1324 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1325 I != E; ++I) 1326 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1327 1328 DenseMap<BasicBlock*, Value*> BlockReplValues; 1329 GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD); 1330 BlockReplValues[UnavailablePred] = NewLoad; 1331 1332 // Perform PHI construction. 1333 Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1334 LI->replaceAllUsesWith(V); 1335 if (isa<PHINode>(V)) 1336 V->takeName(LI); 1337 if (isa<PointerType>(V->getType())) 1338 MD->invalidateCachedPointerInfo(V); 1339 toErase.push_back(LI); 1340 NumPRELoad++; 1341 return true; 1342} 1343 1344/// processLoad - Attempt to eliminate a load, first by eliminating it 1345/// locally, and then attempting non-local elimination if that fails. 1346bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1347 if (L->isVolatile()) 1348 return false; 1349 1350 // ... to a pointer that has been loaded from before... 1351 MemDepResult dep = MD->getDependency(L); 1352 1353 // If the value isn't available, don't do anything! 1354 if (dep.isClobber()) { 1355 // FIXME: In the future, we should handle things like: 1356 // store i32 123, i32* %P 1357 // %A = bitcast i32* %P to i8* 1358 // %B = gep i8* %A, i32 1 1359 // %C = load i8* %B 1360 // 1361 // We could do that by recognizing if the clobber instructions are obviously 1362 // a common base + constant offset, and if the previous store (or memset) 1363 // completely covers this load. This sort of thing can happen in bitfield 1364 // access code. 1365 DEBUG( 1366 // fast print dep, using operator<< on instruction would be too slow 1367 errs() << "GVN: load "; 1368 WriteAsOperand(errs(), L); 1369 Instruction *I = dep.getInst(); 1370 errs() << " is clobbered by " << *I << '\n'; 1371 ); 1372 return false; 1373 } 1374 1375 // If it is defined in another block, try harder. 1376 if (dep.isNonLocal()) 1377 return processNonLocalLoad(L, toErase); 1378 1379 Instruction *DepInst = dep.getInst(); 1380 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1381 Value *StoredVal = DepSI->getOperand(0); 1382 1383 // The store and load are to a must-aliased pointer, but they may not 1384 // actually have the same type. See if we know how to reuse the stored 1385 // value (depending on its type). 1386 const TargetData *TD = 0; 1387 if (StoredVal->getType() != L->getType() && 1388 (TD = getAnalysisIfAvailable<TargetData>())) { 1389 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), L, *TD); 1390 if (StoredVal == 0) 1391 return false; 1392 1393 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1394 << '\n' << *L << "\n\n\n"); 1395 } 1396 1397 // Remove it! 1398 L->replaceAllUsesWith(StoredVal); 1399 if (isa<PointerType>(StoredVal->getType())) 1400 MD->invalidateCachedPointerInfo(StoredVal); 1401 toErase.push_back(L); 1402 NumGVNLoad++; 1403 return true; 1404 } 1405 1406 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1407 Value *AvailableVal = DepLI; 1408 1409 // The loads are of a must-aliased pointer, but they may not actually have 1410 // the same type. See if we know how to reuse the previously loaded value 1411 // (depending on its type). 1412 const TargetData *TD = 0; 1413 if (DepLI->getType() != L->getType() && 1414 (TD = getAnalysisIfAvailable<TargetData>())) { 1415 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L, *TD); 1416 if (AvailableVal == 0) 1417 return false; 1418 1419 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1420 << "\n" << *L << "\n\n\n"); 1421 } 1422 1423 // Remove it! 1424 L->replaceAllUsesWith(AvailableVal); 1425 if (isa<PointerType>(DepLI->getType())) 1426 MD->invalidateCachedPointerInfo(DepLI); 1427 toErase.push_back(L); 1428 NumGVNLoad++; 1429 return true; 1430 } 1431 1432 // FIXME: We should handle memset/memcpy/memmove as dependent instructions to 1433 // forward the value if available. 1434 //if (isa<MemIntrinsic>(DepInst)) 1435 // errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *DepInst << "\n\n"; 1436 1437 1438 // If this load really doesn't depend on anything, then we must be loading an 1439 // undef value. This can happen when loading for a fresh allocation with no 1440 // intervening stores, for example. 1441 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1442 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1443 toErase.push_back(L); 1444 NumGVNLoad++; 1445 return true; 1446 } 1447 1448 return false; 1449} 1450 1451Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) { 1452 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1453 if (I == localAvail.end()) 1454 return 0; 1455 1456 ValueNumberScope* locals = I->second; 1457 1458 while (locals) { 1459 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num); 1460 if (I != locals->table.end()) 1461 return I->second; 1462 else 1463 locals = locals->parent; 1464 } 1465 1466 return 0; 1467} 1468 1469/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination 1470/// by inheritance from the dominator fails, see if we can perform phi 1471/// construction to eliminate the redundancy. 1472Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) { 1473 BasicBlock* BaseBlock = orig->getParent(); 1474 1475 SmallPtrSet<BasicBlock*, 4> Visited; 1476 SmallVector<BasicBlock*, 8> Stack; 1477 Stack.push_back(BaseBlock); 1478 1479 DenseMap<BasicBlock*, Value*> Results; 1480 1481 // Walk backwards through our predecessors, looking for instances of the 1482 // value number we're looking for. Instances are recorded in the Results 1483 // map, which is then used to perform phi construction. 1484 while (!Stack.empty()) { 1485 BasicBlock* Current = Stack.back(); 1486 Stack.pop_back(); 1487 1488 // If we've walked all the way to a proper dominator, then give up. Cases 1489 // where the instance is in the dominator will have been caught by the fast 1490 // path, and any cases that require phi construction further than this are 1491 // probably not worth it anyways. Note that this is a SIGNIFICANT compile 1492 // time improvement. 1493 if (DT->properlyDominates(Current, orig->getParent())) return 0; 1494 1495 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA = 1496 localAvail.find(Current); 1497 if (LA == localAvail.end()) return 0; 1498 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno); 1499 1500 if (V != LA->second->table.end()) { 1501 // Found an instance, record it. 1502 Results.insert(std::make_pair(Current, V->second)); 1503 continue; 1504 } 1505 1506 // If we reach the beginning of the function, then give up. 1507 if (pred_begin(Current) == pred_end(Current)) 1508 return 0; 1509 1510 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current); 1511 PI != PE; ++PI) 1512 if (Visited.insert(*PI)) 1513 Stack.push_back(*PI); 1514 } 1515 1516 // If we didn't find instances, give up. Otherwise, perform phi construction. 1517 if (Results.size() == 0) 1518 return 0; 1519 else 1520 return GetValueForBlock(BaseBlock, orig, Results, true); 1521} 1522 1523/// processInstruction - When calculating availability, handle an instruction 1524/// by inserting it into the appropriate sets 1525bool GVN::processInstruction(Instruction *I, 1526 SmallVectorImpl<Instruction*> &toErase) { 1527 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 1528 bool changed = processLoad(L, toErase); 1529 1530 if (!changed) { 1531 unsigned num = VN.lookup_or_add(L); 1532 localAvail[I->getParent()]->table.insert(std::make_pair(num, L)); 1533 } 1534 1535 return changed; 1536 } 1537 1538 uint32_t nextNum = VN.getNextUnusedValueNumber(); 1539 unsigned num = VN.lookup_or_add(I); 1540 1541 if (BranchInst* BI = dyn_cast<BranchInst>(I)) { 1542 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1543 1544 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1545 return false; 1546 1547 Value* branchCond = BI->getCondition(); 1548 uint32_t condVN = VN.lookup_or_add(branchCond); 1549 1550 BasicBlock* trueSucc = BI->getSuccessor(0); 1551 BasicBlock* falseSucc = BI->getSuccessor(1); 1552 1553 if (trueSucc->getSinglePredecessor()) 1554 localAvail[trueSucc]->table[condVN] = 1555 ConstantInt::getTrue(trueSucc->getContext()); 1556 if (falseSucc->getSinglePredecessor()) 1557 localAvail[falseSucc]->table[condVN] = 1558 ConstantInt::getFalse(trueSucc->getContext()); 1559 1560 return false; 1561 1562 // Allocations are always uniquely numbered, so we can save time and memory 1563 // by fast failing them. 1564 } else if (isa<AllocationInst>(I) || isMalloc(I) || isa<TerminatorInst>(I)) { 1565 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1566 return false; 1567 } 1568 1569 // Collapse PHI nodes 1570 if (PHINode* p = dyn_cast<PHINode>(I)) { 1571 Value* constVal = CollapsePhi(p); 1572 1573 if (constVal) { 1574 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1575 PI != PE; ++PI) 1576 PI->second.erase(p); 1577 1578 p->replaceAllUsesWith(constVal); 1579 if (isa<PointerType>(constVal->getType())) 1580 MD->invalidateCachedPointerInfo(constVal); 1581 VN.erase(p); 1582 1583 toErase.push_back(p); 1584 } else { 1585 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1586 } 1587 1588 // If the number we were assigned was a brand new VN, then we don't 1589 // need to do a lookup to see if the number already exists 1590 // somewhere in the domtree: it can't! 1591 } else if (num == nextNum) { 1592 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1593 1594 // Perform fast-path value-number based elimination of values inherited from 1595 // dominators. 1596 } else if (Value* repl = lookupNumber(I->getParent(), num)) { 1597 // Remove it! 1598 VN.erase(I); 1599 I->replaceAllUsesWith(repl); 1600 if (isa<PointerType>(repl->getType())) 1601 MD->invalidateCachedPointerInfo(repl); 1602 toErase.push_back(I); 1603 return true; 1604 1605#if 0 1606 // Perform slow-pathvalue-number based elimination with phi construction. 1607 } else if (Value* repl = AttemptRedundancyElimination(I, num)) { 1608 // Remove it! 1609 VN.erase(I); 1610 I->replaceAllUsesWith(repl); 1611 if (isa<PointerType>(repl->getType())) 1612 MD->invalidateCachedPointerInfo(repl); 1613 toErase.push_back(I); 1614 return true; 1615#endif 1616 } else { 1617 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1618 } 1619 1620 return false; 1621} 1622 1623/// runOnFunction - This is the main transformation entry point for a function. 1624bool GVN::runOnFunction(Function& F) { 1625 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1626 DT = &getAnalysis<DominatorTree>(); 1627 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1628 VN.setMemDep(MD); 1629 VN.setDomTree(DT); 1630 1631 bool changed = false; 1632 bool shouldContinue = true; 1633 1634 // Merge unconditional branches, allowing PRE to catch more 1635 // optimization opportunities. 1636 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1637 BasicBlock* BB = FI; 1638 ++FI; 1639 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1640 if (removedBlock) NumGVNBlocks++; 1641 1642 changed |= removedBlock; 1643 } 1644 1645 unsigned Iteration = 0; 1646 1647 while (shouldContinue) { 1648 DEBUG(errs() << "GVN iteration: " << Iteration << "\n"); 1649 shouldContinue = iterateOnFunction(F); 1650 changed |= shouldContinue; 1651 ++Iteration; 1652 } 1653 1654 if (EnablePRE) { 1655 bool PREChanged = true; 1656 while (PREChanged) { 1657 PREChanged = performPRE(F); 1658 changed |= PREChanged; 1659 } 1660 } 1661 // FIXME: Should perform GVN again after PRE does something. PRE can move 1662 // computations into blocks where they become fully redundant. Note that 1663 // we can't do this until PRE's critical edge splitting updates memdep. 1664 // Actually, when this happens, we should just fully integrate PRE into GVN. 1665 1666 cleanupGlobalSets(); 1667 1668 return changed; 1669} 1670 1671 1672bool GVN::processBlock(BasicBlock* BB) { 1673 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1674 // incrementing BI before processing an instruction). 1675 SmallVector<Instruction*, 8> toErase; 1676 bool changed_function = false; 1677 1678 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1679 BI != BE;) { 1680 changed_function |= processInstruction(BI, toErase); 1681 if (toErase.empty()) { 1682 ++BI; 1683 continue; 1684 } 1685 1686 // If we need some instructions deleted, do it now. 1687 NumGVNInstr += toErase.size(); 1688 1689 // Avoid iterator invalidation. 1690 bool AtStart = BI == BB->begin(); 1691 if (!AtStart) 1692 --BI; 1693 1694 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1695 E = toErase.end(); I != E; ++I) { 1696 DEBUG(errs() << "GVN removed: " << **I << '\n'); 1697 MD->removeInstruction(*I); 1698 (*I)->eraseFromParent(); 1699 DEBUG(verifyRemoved(*I)); 1700 } 1701 toErase.clear(); 1702 1703 if (AtStart) 1704 BI = BB->begin(); 1705 else 1706 ++BI; 1707 } 1708 1709 return changed_function; 1710} 1711 1712/// performPRE - Perform a purely local form of PRE that looks for diamond 1713/// control flow patterns and attempts to perform simple PRE at the join point. 1714bool GVN::performPRE(Function& F) { 1715 bool Changed = false; 1716 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1717 DenseMap<BasicBlock*, Value*> predMap; 1718 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1719 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1720 BasicBlock* CurrentBlock = *DI; 1721 1722 // Nothing to PRE in the entry block. 1723 if (CurrentBlock == &F.getEntryBlock()) continue; 1724 1725 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1726 BE = CurrentBlock->end(); BI != BE; ) { 1727 Instruction *CurInst = BI++; 1728 1729 if (isa<AllocationInst>(CurInst) || isMalloc(CurInst) || 1730 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1731 (CurInst->getType() == Type::getVoidTy(F.getContext())) || 1732 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1733 isa<DbgInfoIntrinsic>(CurInst)) 1734 continue; 1735 1736 uint32_t valno = VN.lookup(CurInst); 1737 1738 // Look for the predecessors for PRE opportunities. We're 1739 // only trying to solve the basic diamond case, where 1740 // a value is computed in the successor and one predecessor, 1741 // but not the other. We also explicitly disallow cases 1742 // where the successor is its own predecessor, because they're 1743 // more complicated to get right. 1744 unsigned numWith = 0; 1745 unsigned numWithout = 0; 1746 BasicBlock* PREPred = 0; 1747 predMap.clear(); 1748 1749 for (pred_iterator PI = pred_begin(CurrentBlock), 1750 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1751 // We're not interested in PRE where the block is its 1752 // own predecessor, on in blocks with predecessors 1753 // that are not reachable. 1754 if (*PI == CurrentBlock) { 1755 numWithout = 2; 1756 break; 1757 } else if (!localAvail.count(*PI)) { 1758 numWithout = 2; 1759 break; 1760 } 1761 1762 DenseMap<uint32_t, Value*>::iterator predV = 1763 localAvail[*PI]->table.find(valno); 1764 if (predV == localAvail[*PI]->table.end()) { 1765 PREPred = *PI; 1766 numWithout++; 1767 } else if (predV->second == CurInst) { 1768 numWithout = 2; 1769 } else { 1770 predMap[*PI] = predV->second; 1771 numWith++; 1772 } 1773 } 1774 1775 // Don't do PRE when it might increase code size, i.e. when 1776 // we would need to insert instructions in more than one pred. 1777 if (numWithout != 1 || numWith == 0) 1778 continue; 1779 1780 // We can't do PRE safely on a critical edge, so instead we schedule 1781 // the edge to be split and perform the PRE the next time we iterate 1782 // on the function. 1783 unsigned succNum = 0; 1784 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1785 i != e; ++i) 1786 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1787 succNum = i; 1788 break; 1789 } 1790 1791 if (isCriticalEdge(PREPred->getTerminator(), succNum)) { 1792 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum)); 1793 continue; 1794 } 1795 1796 // Instantiate the expression the in predecessor that lacked it. 1797 // Because we are going top-down through the block, all value numbers 1798 // will be available in the predecessor by the time we need them. Any 1799 // that weren't original present will have been instantiated earlier 1800 // in this loop. 1801 Instruction* PREInstr = CurInst->clone(CurInst->getContext()); 1802 bool success = true; 1803 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1804 Value *Op = PREInstr->getOperand(i); 1805 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1806 continue; 1807 1808 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1809 PREInstr->setOperand(i, V); 1810 } else { 1811 success = false; 1812 break; 1813 } 1814 } 1815 1816 // Fail out if we encounter an operand that is not available in 1817 // the PRE predecessor. This is typically because of loads which 1818 // are not value numbered precisely. 1819 if (!success) { 1820 delete PREInstr; 1821 DEBUG(verifyRemoved(PREInstr)); 1822 continue; 1823 } 1824 1825 PREInstr->insertBefore(PREPred->getTerminator()); 1826 PREInstr->setName(CurInst->getName() + ".pre"); 1827 predMap[PREPred] = PREInstr; 1828 VN.add(PREInstr, valno); 1829 NumGVNPRE++; 1830 1831 // Update the availability map to include the new instruction. 1832 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr)); 1833 1834 // Create a PHI to make the value available in this block. 1835 PHINode* Phi = PHINode::Create(CurInst->getType(), 1836 CurInst->getName() + ".pre-phi", 1837 CurrentBlock->begin()); 1838 for (pred_iterator PI = pred_begin(CurrentBlock), 1839 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1840 Phi->addIncoming(predMap[*PI], *PI); 1841 1842 VN.add(Phi, valno); 1843 localAvail[CurrentBlock]->table[valno] = Phi; 1844 1845 CurInst->replaceAllUsesWith(Phi); 1846 if (isa<PointerType>(Phi->getType())) 1847 MD->invalidateCachedPointerInfo(Phi); 1848 VN.erase(CurInst); 1849 1850 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n'); 1851 MD->removeInstruction(CurInst); 1852 CurInst->eraseFromParent(); 1853 DEBUG(verifyRemoved(CurInst)); 1854 Changed = true; 1855 } 1856 } 1857 1858 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1859 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1860 SplitCriticalEdge(I->first, I->second, this); 1861 1862 return Changed || toSplit.size(); 1863} 1864 1865/// iterateOnFunction - Executes one iteration of GVN 1866bool GVN::iterateOnFunction(Function &F) { 1867 cleanupGlobalSets(); 1868 1869 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1870 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1871 if (DI->getIDom()) 1872 localAvail[DI->getBlock()] = 1873 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1874 else 1875 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1876 } 1877 1878 // Top-down walk of the dominator tree 1879 bool changed = false; 1880#if 0 1881 // Needed for value numbering with phi construction to work. 1882 ReversePostOrderTraversal<Function*> RPOT(&F); 1883 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1884 RE = RPOT.end(); RI != RE; ++RI) 1885 changed |= processBlock(*RI); 1886#else 1887 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1888 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1889 changed |= processBlock(DI->getBlock()); 1890#endif 1891 1892 return changed; 1893} 1894 1895void GVN::cleanupGlobalSets() { 1896 VN.clear(); 1897 phiMap.clear(); 1898 1899 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1900 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1901 delete I->second; 1902 localAvail.clear(); 1903} 1904 1905/// verifyRemoved - Verify that the specified instruction does not occur in our 1906/// internal data structures. 1907void GVN::verifyRemoved(const Instruction *Inst) const { 1908 VN.verifyRemoved(Inst); 1909 1910 // Walk through the PHI map to make sure the instruction isn't hiding in there 1911 // somewhere. 1912 for (PhiMapType::iterator 1913 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) { 1914 assert(I->first != Inst && "Inst is still a key in PHI map!"); 1915 1916 for (SmallPtrSet<Instruction*, 4>::iterator 1917 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) { 1918 assert(*II != Inst && "Inst is still a value in PHI map!"); 1919 } 1920 } 1921 1922 // Walk through the value number scope to make sure the instruction isn't 1923 // ferreted away in it. 1924 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1925 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1926 const ValueNumberScope *VNS = I->second; 1927 1928 while (VNS) { 1929 for (DenseMap<uint32_t, Value*>::iterator 1930 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1931 assert(II->second != Inst && "Inst still in value numbering scope!"); 1932 } 1933 1934 VNS = VNS->parent; 1935 } 1936 } 1937} 1938