GVN.cpp revision 5defacc6e605f4651c6300237cef8e9bb2eb6d0e
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/MemoryDependenceAnalysis.h" 36#include "llvm/Support/CFG.h" 37#include "llvm/Support/CommandLine.h" 38#include "llvm/Support/Compiler.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/ErrorHandling.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/Transforms/Utils/BasicBlockUtils.h" 43#include "llvm/Transforms/Utils/Local.h" 44#include <cstdio> 45using namespace llvm; 46 47STATISTIC(NumGVNInstr, "Number of instructions deleted"); 48STATISTIC(NumGVNLoad, "Number of loads deleted"); 49STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 50STATISTIC(NumGVNBlocks, "Number of blocks merged"); 51STATISTIC(NumPRELoad, "Number of loads PRE'd"); 52 53static cl::opt<bool> EnablePRE("enable-pre", 54 cl::init(true), cl::Hidden); 55static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 56 57//===----------------------------------------------------------------------===// 58// ValueTable Class 59//===----------------------------------------------------------------------===// 60 61/// This class holds the mapping between values and value numbers. It is used 62/// as an efficient mechanism to determine the expression-wise equivalence of 63/// two values. 64namespace { 65 struct VISIBILITY_HIDDEN Expression { 66 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 67 UDIV, SDIV, FDIV, UREM, SREM, 68 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 69 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 70 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 71 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 72 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 73 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 74 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 75 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 76 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 77 EMPTY, TOMBSTONE }; 78 79 ExpressionOpcode opcode; 80 const Type* type; 81 uint32_t firstVN; 82 uint32_t secondVN; 83 uint32_t thirdVN; 84 SmallVector<uint32_t, 4> varargs; 85 Value* function; 86 87 Expression() { } 88 Expression(ExpressionOpcode o) : opcode(o) { } 89 90 bool operator==(const Expression &other) const { 91 if (opcode != other.opcode) 92 return false; 93 else if (opcode == EMPTY || opcode == TOMBSTONE) 94 return true; 95 else if (type != other.type) 96 return false; 97 else if (function != other.function) 98 return false; 99 else if (firstVN != other.firstVN) 100 return false; 101 else if (secondVN != other.secondVN) 102 return false; 103 else if (thirdVN != other.thirdVN) 104 return false; 105 else { 106 if (varargs.size() != other.varargs.size()) 107 return false; 108 109 for (size_t i = 0; i < varargs.size(); ++i) 110 if (varargs[i] != other.varargs[i]) 111 return false; 112 113 return true; 114 } 115 } 116 117 bool operator!=(const Expression &other) const { 118 return !(*this == other); 119 } 120 }; 121 122 class VISIBILITY_HIDDEN ValueTable { 123 private: 124 DenseMap<Value*, uint32_t> valueNumbering; 125 DenseMap<Expression, uint32_t> expressionNumbering; 126 AliasAnalysis* AA; 127 MemoryDependenceAnalysis* MD; 128 DominatorTree* DT; 129 130 uint32_t nextValueNumber; 131 132 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 133 Expression::ExpressionOpcode getOpcode(CmpInst* C); 134 Expression::ExpressionOpcode getOpcode(CastInst* C); 135 Expression create_expression(BinaryOperator* BO); 136 Expression create_expression(CmpInst* C); 137 Expression create_expression(ShuffleVectorInst* V); 138 Expression create_expression(ExtractElementInst* C); 139 Expression create_expression(InsertElementInst* V); 140 Expression create_expression(SelectInst* V); 141 Expression create_expression(CastInst* C); 142 Expression create_expression(GetElementPtrInst* G); 143 Expression create_expression(CallInst* C); 144 Expression create_expression(Constant* C); 145 public: 146 ValueTable() : nextValueNumber(1) { } 147 uint32_t lookup_or_add(Value* V); 148 uint32_t lookup(Value* V) const; 149 void add(Value* V, uint32_t num); 150 void clear(); 151 void erase(Value* v); 152 unsigned size(); 153 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 154 AliasAnalysis *getAliasAnalysis() const { return AA; } 155 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 156 void setDomTree(DominatorTree* D) { DT = D; } 157 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 158 void verifyRemoved(const Value *) const; 159 }; 160} 161 162namespace llvm { 163template <> struct DenseMapInfo<Expression> { 164 static inline Expression getEmptyKey() { 165 return Expression(Expression::EMPTY); 166 } 167 168 static inline Expression getTombstoneKey() { 169 return Expression(Expression::TOMBSTONE); 170 } 171 172 static unsigned getHashValue(const Expression e) { 173 unsigned hash = e.opcode; 174 175 hash = e.firstVN + hash * 37; 176 hash = e.secondVN + hash * 37; 177 hash = e.thirdVN + hash * 37; 178 179 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 180 (unsigned)((uintptr_t)e.type >> 9)) + 181 hash * 37; 182 183 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 184 E = e.varargs.end(); I != E; ++I) 185 hash = *I + hash * 37; 186 187 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 188 (unsigned)((uintptr_t)e.function >> 9)) + 189 hash * 37; 190 191 return hash; 192 } 193 static bool isEqual(const Expression &LHS, const Expression &RHS) { 194 return LHS == RHS; 195 } 196 static bool isPod() { return true; } 197}; 198} 199 200//===----------------------------------------------------------------------===// 201// ValueTable Internal Functions 202//===----------------------------------------------------------------------===// 203Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 204 switch(BO->getOpcode()) { 205 default: // THIS SHOULD NEVER HAPPEN 206 llvm_unreachable("Binary operator with unknown opcode?"); 207 case Instruction::Add: return Expression::ADD; 208 case Instruction::FAdd: return Expression::FADD; 209 case Instruction::Sub: return Expression::SUB; 210 case Instruction::FSub: return Expression::FSUB; 211 case Instruction::Mul: return Expression::MUL; 212 case Instruction::FMul: return Expression::FMUL; 213 case Instruction::UDiv: return Expression::UDIV; 214 case Instruction::SDiv: return Expression::SDIV; 215 case Instruction::FDiv: return Expression::FDIV; 216 case Instruction::URem: return Expression::UREM; 217 case Instruction::SRem: return Expression::SREM; 218 case Instruction::FRem: return Expression::FREM; 219 case Instruction::Shl: return Expression::SHL; 220 case Instruction::LShr: return Expression::LSHR; 221 case Instruction::AShr: return Expression::ASHR; 222 case Instruction::And: return Expression::AND; 223 case Instruction::Or: return Expression::OR; 224 case Instruction::Xor: return Expression::XOR; 225 } 226} 227 228Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 229 if (isa<ICmpInst>(C)) { 230 switch (C->getPredicate()) { 231 default: // THIS SHOULD NEVER HAPPEN 232 llvm_unreachable("Comparison with unknown predicate?"); 233 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 234 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 235 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 236 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 237 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 238 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 239 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 240 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 241 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 242 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 243 } 244 } else { 245 switch (C->getPredicate()) { 246 default: // THIS SHOULD NEVER HAPPEN 247 llvm_unreachable("Comparison with unknown predicate?"); 248 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 249 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 250 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 251 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 252 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 253 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 254 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 255 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 256 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 257 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 258 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 259 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 260 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 261 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 262 } 263 } 264} 265 266Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 267 switch(C->getOpcode()) { 268 default: // THIS SHOULD NEVER HAPPEN 269 llvm_unreachable("Cast operator with unknown opcode?"); 270 case Instruction::Trunc: return Expression::TRUNC; 271 case Instruction::ZExt: return Expression::ZEXT; 272 case Instruction::SExt: return Expression::SEXT; 273 case Instruction::FPToUI: return Expression::FPTOUI; 274 case Instruction::FPToSI: return Expression::FPTOSI; 275 case Instruction::UIToFP: return Expression::UITOFP; 276 case Instruction::SIToFP: return Expression::SITOFP; 277 case Instruction::FPTrunc: return Expression::FPTRUNC; 278 case Instruction::FPExt: return Expression::FPEXT; 279 case Instruction::PtrToInt: return Expression::PTRTOINT; 280 case Instruction::IntToPtr: return Expression::INTTOPTR; 281 case Instruction::BitCast: return Expression::BITCAST; 282 } 283} 284 285Expression ValueTable::create_expression(CallInst* C) { 286 Expression e; 287 288 e.type = C->getType(); 289 e.firstVN = 0; 290 e.secondVN = 0; 291 e.thirdVN = 0; 292 e.function = C->getCalledFunction(); 293 e.opcode = Expression::CALL; 294 295 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 296 I != E; ++I) 297 e.varargs.push_back(lookup_or_add(*I)); 298 299 return e; 300} 301 302Expression ValueTable::create_expression(BinaryOperator* BO) { 303 Expression e; 304 305 e.firstVN = lookup_or_add(BO->getOperand(0)); 306 e.secondVN = lookup_or_add(BO->getOperand(1)); 307 e.thirdVN = 0; 308 e.function = 0; 309 e.type = BO->getType(); 310 e.opcode = getOpcode(BO); 311 312 return e; 313} 314 315Expression ValueTable::create_expression(CmpInst* C) { 316 Expression e; 317 318 e.firstVN = lookup_or_add(C->getOperand(0)); 319 e.secondVN = lookup_or_add(C->getOperand(1)); 320 e.thirdVN = 0; 321 e.function = 0; 322 e.type = C->getType(); 323 e.opcode = getOpcode(C); 324 325 return e; 326} 327 328Expression ValueTable::create_expression(CastInst* C) { 329 Expression e; 330 331 e.firstVN = lookup_or_add(C->getOperand(0)); 332 e.secondVN = 0; 333 e.thirdVN = 0; 334 e.function = 0; 335 e.type = C->getType(); 336 e.opcode = getOpcode(C); 337 338 return e; 339} 340 341Expression ValueTable::create_expression(ShuffleVectorInst* S) { 342 Expression e; 343 344 e.firstVN = lookup_or_add(S->getOperand(0)); 345 e.secondVN = lookup_or_add(S->getOperand(1)); 346 e.thirdVN = lookup_or_add(S->getOperand(2)); 347 e.function = 0; 348 e.type = S->getType(); 349 e.opcode = Expression::SHUFFLE; 350 351 return e; 352} 353 354Expression ValueTable::create_expression(ExtractElementInst* E) { 355 Expression e; 356 357 e.firstVN = lookup_or_add(E->getOperand(0)); 358 e.secondVN = lookup_or_add(E->getOperand(1)); 359 e.thirdVN = 0; 360 e.function = 0; 361 e.type = E->getType(); 362 e.opcode = Expression::EXTRACT; 363 364 return e; 365} 366 367Expression ValueTable::create_expression(InsertElementInst* I) { 368 Expression e; 369 370 e.firstVN = lookup_or_add(I->getOperand(0)); 371 e.secondVN = lookup_or_add(I->getOperand(1)); 372 e.thirdVN = lookup_or_add(I->getOperand(2)); 373 e.function = 0; 374 e.type = I->getType(); 375 e.opcode = Expression::INSERT; 376 377 return e; 378} 379 380Expression ValueTable::create_expression(SelectInst* I) { 381 Expression e; 382 383 e.firstVN = lookup_or_add(I->getCondition()); 384 e.secondVN = lookup_or_add(I->getTrueValue()); 385 e.thirdVN = lookup_or_add(I->getFalseValue()); 386 e.function = 0; 387 e.type = I->getType(); 388 e.opcode = Expression::SELECT; 389 390 return e; 391} 392 393Expression ValueTable::create_expression(GetElementPtrInst* G) { 394 Expression e; 395 396 e.firstVN = lookup_or_add(G->getPointerOperand()); 397 e.secondVN = 0; 398 e.thirdVN = 0; 399 e.function = 0; 400 e.type = G->getType(); 401 e.opcode = Expression::GEP; 402 403 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 404 I != E; ++I) 405 e.varargs.push_back(lookup_or_add(*I)); 406 407 return e; 408} 409 410//===----------------------------------------------------------------------===// 411// ValueTable External Functions 412//===----------------------------------------------------------------------===// 413 414/// add - Insert a value into the table with a specified value number. 415void ValueTable::add(Value* V, uint32_t num) { 416 valueNumbering.insert(std::make_pair(V, num)); 417} 418 419/// lookup_or_add - Returns the value number for the specified value, assigning 420/// it a new number if it did not have one before. 421uint32_t ValueTable::lookup_or_add(Value* V) { 422 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 423 if (VI != valueNumbering.end()) 424 return VI->second; 425 426 if (CallInst* C = dyn_cast<CallInst>(V)) { 427 if (AA->doesNotAccessMemory(C)) { 428 Expression e = create_expression(C); 429 430 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 431 if (EI != expressionNumbering.end()) { 432 valueNumbering.insert(std::make_pair(V, EI->second)); 433 return EI->second; 434 } else { 435 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 436 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 437 438 return nextValueNumber++; 439 } 440 } else if (AA->onlyReadsMemory(C)) { 441 Expression e = create_expression(C); 442 443 if (expressionNumbering.find(e) == expressionNumbering.end()) { 444 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 445 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 446 return nextValueNumber++; 447 } 448 449 MemDepResult local_dep = MD->getDependency(C); 450 451 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 452 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 453 return nextValueNumber++; 454 } 455 456 if (local_dep.isDef()) { 457 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 458 459 if (local_cdep->getNumOperands() != C->getNumOperands()) { 460 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 461 return nextValueNumber++; 462 } 463 464 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 465 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 466 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 467 if (c_vn != cd_vn) { 468 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 469 return nextValueNumber++; 470 } 471 } 472 473 uint32_t v = lookup_or_add(local_cdep); 474 valueNumbering.insert(std::make_pair(V, v)); 475 return v; 476 } 477 478 // Non-local case. 479 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 480 MD->getNonLocalCallDependency(CallSite(C)); 481 // FIXME: call/call dependencies for readonly calls should return def, not 482 // clobber! Move the checking logic to MemDep! 483 CallInst* cdep = 0; 484 485 // Check to see if we have a single dominating call instruction that is 486 // identical to C. 487 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 488 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 489 // Ignore non-local dependencies. 490 if (I->second.isNonLocal()) 491 continue; 492 493 // We don't handle non-depedencies. If we already have a call, reject 494 // instruction dependencies. 495 if (I->second.isClobber() || cdep != 0) { 496 cdep = 0; 497 break; 498 } 499 500 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 501 // FIXME: All duplicated with non-local case. 502 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 503 cdep = NonLocalDepCall; 504 continue; 505 } 506 507 cdep = 0; 508 break; 509 } 510 511 if (!cdep) { 512 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 513 return nextValueNumber++; 514 } 515 516 if (cdep->getNumOperands() != C->getNumOperands()) { 517 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 518 return nextValueNumber++; 519 } 520 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 521 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 522 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 523 if (c_vn != cd_vn) { 524 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 525 return nextValueNumber++; 526 } 527 } 528 529 uint32_t v = lookup_or_add(cdep); 530 valueNumbering.insert(std::make_pair(V, v)); 531 return v; 532 533 } else { 534 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 535 return nextValueNumber++; 536 } 537 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 538 Expression e = create_expression(BO); 539 540 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 541 if (EI != expressionNumbering.end()) { 542 valueNumbering.insert(std::make_pair(V, EI->second)); 543 return EI->second; 544 } else { 545 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 546 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 547 548 return nextValueNumber++; 549 } 550 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 551 Expression e = create_expression(C); 552 553 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 554 if (EI != expressionNumbering.end()) { 555 valueNumbering.insert(std::make_pair(V, EI->second)); 556 return EI->second; 557 } else { 558 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 559 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 560 561 return nextValueNumber++; 562 } 563 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 564 Expression e = create_expression(U); 565 566 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 567 if (EI != expressionNumbering.end()) { 568 valueNumbering.insert(std::make_pair(V, EI->second)); 569 return EI->second; 570 } else { 571 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 572 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 573 574 return nextValueNumber++; 575 } 576 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 577 Expression e = create_expression(U); 578 579 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 580 if (EI != expressionNumbering.end()) { 581 valueNumbering.insert(std::make_pair(V, EI->second)); 582 return EI->second; 583 } else { 584 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 585 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 586 587 return nextValueNumber++; 588 } 589 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 590 Expression e = create_expression(U); 591 592 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 593 if (EI != expressionNumbering.end()) { 594 valueNumbering.insert(std::make_pair(V, EI->second)); 595 return EI->second; 596 } else { 597 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 598 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 599 600 return nextValueNumber++; 601 } 602 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 603 Expression e = create_expression(U); 604 605 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 606 if (EI != expressionNumbering.end()) { 607 valueNumbering.insert(std::make_pair(V, EI->second)); 608 return EI->second; 609 } else { 610 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 611 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 612 613 return nextValueNumber++; 614 } 615 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 616 Expression e = create_expression(U); 617 618 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 619 if (EI != expressionNumbering.end()) { 620 valueNumbering.insert(std::make_pair(V, EI->second)); 621 return EI->second; 622 } else { 623 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 624 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 625 626 return nextValueNumber++; 627 } 628 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 629 Expression e = create_expression(U); 630 631 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 632 if (EI != expressionNumbering.end()) { 633 valueNumbering.insert(std::make_pair(V, EI->second)); 634 return EI->second; 635 } else { 636 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 637 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 638 639 return nextValueNumber++; 640 } 641 } else { 642 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 643 return nextValueNumber++; 644 } 645} 646 647/// lookup - Returns the value number of the specified value. Fails if 648/// the value has not yet been numbered. 649uint32_t ValueTable::lookup(Value* V) const { 650 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 651 assert(VI != valueNumbering.end() && "Value not numbered?"); 652 return VI->second; 653} 654 655/// clear - Remove all entries from the ValueTable 656void ValueTable::clear() { 657 valueNumbering.clear(); 658 expressionNumbering.clear(); 659 nextValueNumber = 1; 660} 661 662/// erase - Remove a value from the value numbering 663void ValueTable::erase(Value* V) { 664 valueNumbering.erase(V); 665} 666 667/// verifyRemoved - Verify that the value is removed from all internal data 668/// structures. 669void ValueTable::verifyRemoved(const Value *V) const { 670 for (DenseMap<Value*, uint32_t>::iterator 671 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 672 assert(I->first != V && "Inst still occurs in value numbering map!"); 673 } 674} 675 676//===----------------------------------------------------------------------===// 677// GVN Pass 678//===----------------------------------------------------------------------===// 679 680namespace { 681 struct VISIBILITY_HIDDEN ValueNumberScope { 682 ValueNumberScope* parent; 683 DenseMap<uint32_t, Value*> table; 684 685 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 686 }; 687} 688 689namespace { 690 691 class VISIBILITY_HIDDEN GVN : public FunctionPass { 692 bool runOnFunction(Function &F); 693 public: 694 static char ID; // Pass identification, replacement for typeid 695 GVN() : FunctionPass(&ID) { } 696 697 private: 698 MemoryDependenceAnalysis *MD; 699 DominatorTree *DT; 700 701 ValueTable VN; 702 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 703 704 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 705 PhiMapType phiMap; 706 707 708 // This transformation requires dominator postdominator info 709 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 710 AU.addRequired<DominatorTree>(); 711 AU.addRequired<MemoryDependenceAnalysis>(); 712 AU.addRequired<AliasAnalysis>(); 713 714 AU.addPreserved<DominatorTree>(); 715 AU.addPreserved<AliasAnalysis>(); 716 } 717 718 // Helper fuctions 719 // FIXME: eliminate or document these better 720 bool processLoad(LoadInst* L, 721 SmallVectorImpl<Instruction*> &toErase); 722 bool processInstruction(Instruction* I, 723 SmallVectorImpl<Instruction*> &toErase); 724 bool processNonLocalLoad(LoadInst* L, 725 SmallVectorImpl<Instruction*> &toErase); 726 bool processBlock(BasicBlock* BB); 727 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig, 728 DenseMap<BasicBlock*, Value*> &Phis, 729 bool top_level = false); 730 void dump(DenseMap<uint32_t, Value*>& d); 731 bool iterateOnFunction(Function &F); 732 Value* CollapsePhi(PHINode* p); 733 bool isSafeReplacement(PHINode* p, Instruction* inst); 734 bool performPRE(Function& F); 735 Value* lookupNumber(BasicBlock* BB, uint32_t num); 736 bool mergeBlockIntoPredecessor(BasicBlock* BB); 737 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno); 738 void cleanupGlobalSets(); 739 void verifyRemoved(const Instruction *I) const; 740 }; 741 742 char GVN::ID = 0; 743} 744 745// createGVNPass - The public interface to this file... 746FunctionPass *llvm::createGVNPass() { return new GVN(); } 747 748static RegisterPass<GVN> X("gvn", 749 "Global Value Numbering"); 750 751void GVN::dump(DenseMap<uint32_t, Value*>& d) { 752 printf("{\n"); 753 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 754 E = d.end(); I != E; ++I) { 755 printf("%d\n", I->first); 756 I->second->dump(); 757 } 758 printf("}\n"); 759} 760 761Value* GVN::CollapsePhi(PHINode* p) { 762 Value* constVal = p->hasConstantValue(); 763 if (!constVal) return 0; 764 765 Instruction* inst = dyn_cast<Instruction>(constVal); 766 if (!inst) 767 return constVal; 768 769 if (DT->dominates(inst, p)) 770 if (isSafeReplacement(p, inst)) 771 return inst; 772 return 0; 773} 774 775bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 776 if (!isa<PHINode>(inst)) 777 return true; 778 779 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 780 UI != E; ++UI) 781 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 782 if (use_phi->getParent() == inst->getParent()) 783 return false; 784 785 return true; 786} 787 788/// GetValueForBlock - Get the value to use within the specified basic block. 789/// available values are in Phis. 790Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig, 791 DenseMap<BasicBlock*, Value*> &Phis, 792 bool top_level) { 793 794 // If we have already computed this value, return the previously computed val. 795 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 796 if (V != Phis.end() && !top_level) return V->second; 797 798 // If the block is unreachable, just return undef, since this path 799 // can't actually occur at runtime. 800 if (!DT->isReachableFromEntry(BB)) 801 return Phis[BB] = UndefValue::get(orig->getType()); 802 803 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 804 Value *ret = GetValueForBlock(Pred, orig, Phis); 805 Phis[BB] = ret; 806 return ret; 807 } 808 809 // Get the number of predecessors of this block so we can reserve space later. 810 // If there is already a PHI in it, use the #preds from it, otherwise count. 811 // Getting it from the PHI is constant time. 812 unsigned NumPreds; 813 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin())) 814 NumPreds = ExistingPN->getNumIncomingValues(); 815 else 816 NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 817 818 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 819 // now, then get values to fill in the incoming values for the PHI. 820 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 821 BB->begin()); 822 PN->reserveOperandSpace(NumPreds); 823 824 Phis.insert(std::make_pair(BB, PN)); 825 826 // Fill in the incoming values for the block. 827 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 828 Value* val = GetValueForBlock(*PI, orig, Phis); 829 PN->addIncoming(val, *PI); 830 } 831 832 VN.getAliasAnalysis()->copyValue(orig, PN); 833 834 // Attempt to collapse PHI nodes that are trivially redundant 835 Value* v = CollapsePhi(PN); 836 if (!v) { 837 // Cache our phi construction results 838 if (LoadInst* L = dyn_cast<LoadInst>(orig)) 839 phiMap[L->getPointerOperand()].insert(PN); 840 else 841 phiMap[orig].insert(PN); 842 843 return PN; 844 } 845 846 PN->replaceAllUsesWith(v); 847 if (isa<PointerType>(v->getType())) 848 MD->invalidateCachedPointerInfo(v); 849 850 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 851 E = Phis.end(); I != E; ++I) 852 if (I->second == PN) 853 I->second = v; 854 855 DEBUG(errs() << "GVN removed: " << *PN); 856 MD->removeInstruction(PN); 857 PN->eraseFromParent(); 858 DEBUG(verifyRemoved(PN)); 859 860 Phis[BB] = v; 861 return v; 862} 863 864/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 865/// we're analyzing is fully available in the specified block. As we go, keep 866/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 867/// map is actually a tri-state map with the following values: 868/// 0) we know the block *is not* fully available. 869/// 1) we know the block *is* fully available. 870/// 2) we do not know whether the block is fully available or not, but we are 871/// currently speculating that it will be. 872/// 3) we are speculating for this block and have used that to speculate for 873/// other blocks. 874static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 875 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 876 // Optimistically assume that the block is fully available and check to see 877 // if we already know about this block in one lookup. 878 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 879 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 880 881 // If the entry already existed for this block, return the precomputed value. 882 if (!IV.second) { 883 // If this is a speculative "available" value, mark it as being used for 884 // speculation of other blocks. 885 if (IV.first->second == 2) 886 IV.first->second = 3; 887 return IV.first->second != 0; 888 } 889 890 // Otherwise, see if it is fully available in all predecessors. 891 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 892 893 // If this block has no predecessors, it isn't live-in here. 894 if (PI == PE) 895 goto SpeculationFailure; 896 897 for (; PI != PE; ++PI) 898 // If the value isn't fully available in one of our predecessors, then it 899 // isn't fully available in this block either. Undo our previous 900 // optimistic assumption and bail out. 901 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 902 goto SpeculationFailure; 903 904 return true; 905 906// SpeculationFailure - If we get here, we found out that this is not, after 907// all, a fully-available block. We have a problem if we speculated on this and 908// used the speculation to mark other blocks as available. 909SpeculationFailure: 910 char &BBVal = FullyAvailableBlocks[BB]; 911 912 // If we didn't speculate on this, just return with it set to false. 913 if (BBVal == 2) { 914 BBVal = 0; 915 return false; 916 } 917 918 // If we did speculate on this value, we could have blocks set to 1 that are 919 // incorrect. Walk the (transitive) successors of this block and mark them as 920 // 0 if set to one. 921 SmallVector<BasicBlock*, 32> BBWorklist; 922 BBWorklist.push_back(BB); 923 924 while (!BBWorklist.empty()) { 925 BasicBlock *Entry = BBWorklist.pop_back_val(); 926 // Note that this sets blocks to 0 (unavailable) if they happen to not 927 // already be in FullyAvailableBlocks. This is safe. 928 char &EntryVal = FullyAvailableBlocks[Entry]; 929 if (EntryVal == 0) continue; // Already unavailable. 930 931 // Mark as unavailable. 932 EntryVal = 0; 933 934 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 935 BBWorklist.push_back(*I); 936 } 937 938 return false; 939} 940 941/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 942/// non-local by performing PHI construction. 943bool GVN::processNonLocalLoad(LoadInst *LI, 944 SmallVectorImpl<Instruction*> &toErase) { 945 // Find the non-local dependencies of the load. 946 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 947 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 948 Deps); 949 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI); 950 951 // If we had to process more than one hundred blocks to find the 952 // dependencies, this load isn't worth worrying about. Optimizing 953 // it will be too expensive. 954 if (Deps.size() > 100) 955 return false; 956 957 // If we had a phi translation failure, we'll have a single entry which is a 958 // clobber in the current block. Reject this early. 959 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 960 DEBUG( 961 errs() << "GVN: non-local load "; 962 WriteAsOperand(errs(), LI); 963 errs() << " is clobbered by " << *Deps[0].second.getInst(); 964 ); 965 return false; 966 } 967 968 // Filter out useless results (non-locals, etc). Keep track of the blocks 969 // where we have a value available in repl, also keep track of whether we see 970 // dependencies that produce an unknown value for the load (such as a call 971 // that could potentially clobber the load). 972 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 973 SmallVector<BasicBlock*, 16> UnavailableBlocks; 974 975 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 976 BasicBlock *DepBB = Deps[i].first; 977 MemDepResult DepInfo = Deps[i].second; 978 979 if (DepInfo.isClobber()) { 980 UnavailableBlocks.push_back(DepBB); 981 continue; 982 } 983 984 Instruction *DepInst = DepInfo.getInst(); 985 986 // Loading the allocation -> undef. 987 if (isa<AllocationInst>(DepInst)) { 988 ValuesPerBlock.push_back(std::make_pair(DepBB, 989 UndefValue::get(LI->getType()))); 990 continue; 991 } 992 993 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 994 // Reject loads and stores that are to the same address but are of 995 // different types. 996 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because 997 // of bitfield access, it would be interesting to optimize for it at some 998 // point. 999 if (S->getOperand(0)->getType() != LI->getType()) { 1000 UnavailableBlocks.push_back(DepBB); 1001 continue; 1002 } 1003 1004 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1005 1006 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) { 1007 if (LD->getType() != LI->getType()) { 1008 UnavailableBlocks.push_back(DepBB); 1009 continue; 1010 } 1011 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1012 } else { 1013 UnavailableBlocks.push_back(DepBB); 1014 continue; 1015 } 1016 } 1017 1018 // If we have no predecessors that produce a known value for this load, exit 1019 // early. 1020 if (ValuesPerBlock.empty()) return false; 1021 1022 // If all of the instructions we depend on produce a known value for this 1023 // load, then it is fully redundant and we can use PHI insertion to compute 1024 // its value. Insert PHIs and remove the fully redundant value now. 1025 if (UnavailableBlocks.empty()) { 1026 // Use cached PHI construction information from previous runs 1027 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1028 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated? 1029 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1030 I != E; ++I) { 1031 if ((*I)->getParent() == LI->getParent()) { 1032 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD #1: " << *LI); 1033 LI->replaceAllUsesWith(*I); 1034 if (isa<PointerType>((*I)->getType())) 1035 MD->invalidateCachedPointerInfo(*I); 1036 toErase.push_back(LI); 1037 NumGVNLoad++; 1038 return true; 1039 } 1040 1041 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1042 } 1043 1044 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI); 1045 1046 DenseMap<BasicBlock*, Value*> BlockReplValues; 1047 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1048 // Perform PHI construction. 1049 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1050 LI->replaceAllUsesWith(v); 1051 1052 if (isa<PHINode>(v)) 1053 v->takeName(LI); 1054 if (isa<PointerType>(v->getType())) 1055 MD->invalidateCachedPointerInfo(v); 1056 toErase.push_back(LI); 1057 NumGVNLoad++; 1058 return true; 1059 } 1060 1061 if (!EnablePRE || !EnableLoadPRE) 1062 return false; 1063 1064 // Okay, we have *some* definitions of the value. This means that the value 1065 // is available in some of our (transitive) predecessors. Lets think about 1066 // doing PRE of this load. This will involve inserting a new load into the 1067 // predecessor when it's not available. We could do this in general, but 1068 // prefer to not increase code size. As such, we only do this when we know 1069 // that we only have to insert *one* load (which means we're basically moving 1070 // the load, not inserting a new one). 1071 1072 SmallPtrSet<BasicBlock *, 4> Blockers; 1073 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1074 Blockers.insert(UnavailableBlocks[i]); 1075 1076 // Lets find first basic block with more than one predecessor. Walk backwards 1077 // through predecessors if needed. 1078 BasicBlock *LoadBB = LI->getParent(); 1079 BasicBlock *TmpBB = LoadBB; 1080 1081 bool isSinglePred = false; 1082 bool allSingleSucc = true; 1083 while (TmpBB->getSinglePredecessor()) { 1084 isSinglePred = true; 1085 TmpBB = TmpBB->getSinglePredecessor(); 1086 if (!TmpBB) // If haven't found any, bail now. 1087 return false; 1088 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1089 return false; 1090 if (Blockers.count(TmpBB)) 1091 return false; 1092 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1093 allSingleSucc = false; 1094 } 1095 1096 assert(TmpBB); 1097 LoadBB = TmpBB; 1098 1099 // If we have a repl set with LI itself in it, this means we have a loop where 1100 // at least one of the values is LI. Since this means that we won't be able 1101 // to eliminate LI even if we insert uses in the other predecessors, we will 1102 // end up increasing code size. Reject this by scanning for LI. 1103 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1104 if (ValuesPerBlock[i].second == LI) 1105 return false; 1106 1107 if (isSinglePred) { 1108 bool isHot = false; 1109 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1110 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second)) 1111 // "Hot" Instruction is in some loop (because it dominates its dep. 1112 // instruction). 1113 if (DT->dominates(LI, I)) { 1114 isHot = true; 1115 break; 1116 } 1117 1118 // We are interested only in "hot" instructions. We don't want to do any 1119 // mis-optimizations here. 1120 if (!isHot) 1121 return false; 1122 } 1123 1124 // Okay, we have some hope :). Check to see if the loaded value is fully 1125 // available in all but one predecessor. 1126 // FIXME: If we could restructure the CFG, we could make a common pred with 1127 // all the preds that don't have an available LI and insert a new load into 1128 // that one block. 1129 BasicBlock *UnavailablePred = 0; 1130 1131 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1132 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1133 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1134 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1135 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1136 1137 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1138 PI != E; ++PI) { 1139 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1140 continue; 1141 1142 // If this load is not available in multiple predecessors, reject it. 1143 if (UnavailablePred && UnavailablePred != *PI) 1144 return false; 1145 UnavailablePred = *PI; 1146 } 1147 1148 assert(UnavailablePred != 0 && 1149 "Fully available value should be eliminated above!"); 1150 1151 // If the loaded pointer is PHI node defined in this block, do PHI translation 1152 // to get its value in the predecessor. 1153 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1154 1155 // Make sure the value is live in the predecessor. If it was defined by a 1156 // non-PHI instruction in this block, we don't know how to recompute it above. 1157 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1158 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1159 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1160 << *LPInst << *LI << "\n"); 1161 return false; 1162 } 1163 1164 // We don't currently handle critical edges :( 1165 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1166 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1167 << UnavailablePred->getName() << "': " << *LI); 1168 return false; 1169 } 1170 1171 // Make sure it is valid to move this load here. We have to watch out for: 1172 // @1 = getelementptr (i8* p, ... 1173 // test p and branch if == 0 1174 // load @1 1175 // It is valid to have the getelementptr before the test, even if p can be 0, 1176 // as getelementptr only does address arithmetic. 1177 // If we are not pushing the value through any multiple-successor blocks 1178 // we do not have this case. Otherwise, check that the load is safe to 1179 // put anywhere; this can be improved, but should be conservatively safe. 1180 if (!allSingleSucc && 1181 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1182 return false; 1183 1184 // Okay, we can eliminate this load by inserting a reload in the predecessor 1185 // and using PHI construction to get the value in the other predecessors, do 1186 // it. 1187 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI); 1188 1189 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1190 LI->getAlignment(), 1191 UnavailablePred->getTerminator()); 1192 1193 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1194 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1195 I != E; ++I) 1196 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1197 1198 DenseMap<BasicBlock*, Value*> BlockReplValues; 1199 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1200 BlockReplValues[UnavailablePred] = NewLoad; 1201 1202 // Perform PHI construction. 1203 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1204 LI->replaceAllUsesWith(v); 1205 if (isa<PHINode>(v)) 1206 v->takeName(LI); 1207 if (isa<PointerType>(v->getType())) 1208 MD->invalidateCachedPointerInfo(v); 1209 toErase.push_back(LI); 1210 NumPRELoad++; 1211 return true; 1212} 1213 1214/// processLoad - Attempt to eliminate a load, first by eliminating it 1215/// locally, and then attempting non-local elimination if that fails. 1216bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1217 if (L->isVolatile()) 1218 return false; 1219 1220 Value* pointer = L->getPointerOperand(); 1221 1222 // ... to a pointer that has been loaded from before... 1223 MemDepResult dep = MD->getDependency(L); 1224 1225 // If the value isn't available, don't do anything! 1226 if (dep.isClobber()) { 1227 DEBUG( 1228 // fast print dep, using operator<< on instruction would be too slow 1229 errs() << "GVN: load "; 1230 WriteAsOperand(errs(), L); 1231 Instruction *I = dep.getInst(); 1232 errs() << " is clobbered by " << *I; 1233 ); 1234 return false; 1235 } 1236 1237 // If it is defined in another block, try harder. 1238 if (dep.isNonLocal()) 1239 return processNonLocalLoad(L, toErase); 1240 1241 Instruction *DepInst = dep.getInst(); 1242 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1243 // Only forward substitute stores to loads of the same type. 1244 // FIXME: Could do better! 1245 if (DepSI->getPointerOperand()->getType() != pointer->getType()) 1246 return false; 1247 1248 // Remove it! 1249 L->replaceAllUsesWith(DepSI->getOperand(0)); 1250 if (isa<PointerType>(DepSI->getOperand(0)->getType())) 1251 MD->invalidateCachedPointerInfo(DepSI->getOperand(0)); 1252 toErase.push_back(L); 1253 NumGVNLoad++; 1254 return true; 1255 } 1256 1257 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1258 // Only forward substitute stores to loads of the same type. 1259 // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian. 1260 if (DepLI->getType() != L->getType()) 1261 return false; 1262 1263 // Remove it! 1264 L->replaceAllUsesWith(DepLI); 1265 if (isa<PointerType>(DepLI->getType())) 1266 MD->invalidateCachedPointerInfo(DepLI); 1267 toErase.push_back(L); 1268 NumGVNLoad++; 1269 return true; 1270 } 1271 1272 // If this load really doesn't depend on anything, then we must be loading an 1273 // undef value. This can happen when loading for a fresh allocation with no 1274 // intervening stores, for example. 1275 if (isa<AllocationInst>(DepInst)) { 1276 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1277 toErase.push_back(L); 1278 NumGVNLoad++; 1279 return true; 1280 } 1281 1282 return false; 1283} 1284 1285Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) { 1286 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1287 if (I == localAvail.end()) 1288 return 0; 1289 1290 ValueNumberScope* locals = I->second; 1291 1292 while (locals) { 1293 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num); 1294 if (I != locals->table.end()) 1295 return I->second; 1296 else 1297 locals = locals->parent; 1298 } 1299 1300 return 0; 1301} 1302 1303/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination 1304/// by inheritance from the dominator fails, see if we can perform phi 1305/// construction to eliminate the redundancy. 1306Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) { 1307 BasicBlock* BaseBlock = orig->getParent(); 1308 1309 SmallPtrSet<BasicBlock*, 4> Visited; 1310 SmallVector<BasicBlock*, 8> Stack; 1311 Stack.push_back(BaseBlock); 1312 1313 DenseMap<BasicBlock*, Value*> Results; 1314 1315 // Walk backwards through our predecessors, looking for instances of the 1316 // value number we're looking for. Instances are recorded in the Results 1317 // map, which is then used to perform phi construction. 1318 while (!Stack.empty()) { 1319 BasicBlock* Current = Stack.back(); 1320 Stack.pop_back(); 1321 1322 // If we've walked all the way to a proper dominator, then give up. Cases 1323 // where the instance is in the dominator will have been caught by the fast 1324 // path, and any cases that require phi construction further than this are 1325 // probably not worth it anyways. Note that this is a SIGNIFICANT compile 1326 // time improvement. 1327 if (DT->properlyDominates(Current, orig->getParent())) return 0; 1328 1329 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA = 1330 localAvail.find(Current); 1331 if (LA == localAvail.end()) return 0; 1332 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno); 1333 1334 if (V != LA->second->table.end()) { 1335 // Found an instance, record it. 1336 Results.insert(std::make_pair(Current, V->second)); 1337 continue; 1338 } 1339 1340 // If we reach the beginning of the function, then give up. 1341 if (pred_begin(Current) == pred_end(Current)) 1342 return 0; 1343 1344 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current); 1345 PI != PE; ++PI) 1346 if (Visited.insert(*PI)) 1347 Stack.push_back(*PI); 1348 } 1349 1350 // If we didn't find instances, give up. Otherwise, perform phi construction. 1351 if (Results.size() == 0) 1352 return 0; 1353 else 1354 return GetValueForBlock(BaseBlock, orig, Results, true); 1355} 1356 1357/// processInstruction - When calculating availability, handle an instruction 1358/// by inserting it into the appropriate sets 1359bool GVN::processInstruction(Instruction *I, 1360 SmallVectorImpl<Instruction*> &toErase) { 1361 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 1362 bool changed = processLoad(L, toErase); 1363 1364 if (!changed) { 1365 unsigned num = VN.lookup_or_add(L); 1366 localAvail[I->getParent()]->table.insert(std::make_pair(num, L)); 1367 } 1368 1369 return changed; 1370 } 1371 1372 uint32_t nextNum = VN.getNextUnusedValueNumber(); 1373 unsigned num = VN.lookup_or_add(I); 1374 1375 if (BranchInst* BI = dyn_cast<BranchInst>(I)) { 1376 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1377 1378 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1379 return false; 1380 1381 Value* branchCond = BI->getCondition(); 1382 uint32_t condVN = VN.lookup_or_add(branchCond); 1383 1384 BasicBlock* trueSucc = BI->getSuccessor(0); 1385 BasicBlock* falseSucc = BI->getSuccessor(1); 1386 1387 if (trueSucc->getSinglePredecessor()) 1388 localAvail[trueSucc]->table[condVN] = 1389 ConstantInt::getTrue(trueSucc->getContext()); 1390 if (falseSucc->getSinglePredecessor()) 1391 localAvail[falseSucc]->table[condVN] = 1392 ConstantInt::getFalse(trueSucc->getContext()); 1393 1394 return false; 1395 1396 // Allocations are always uniquely numbered, so we can save time and memory 1397 // by fast failing them. 1398 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) { 1399 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1400 return false; 1401 } 1402 1403 // Collapse PHI nodes 1404 if (PHINode* p = dyn_cast<PHINode>(I)) { 1405 Value* constVal = CollapsePhi(p); 1406 1407 if (constVal) { 1408 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1409 PI != PE; ++PI) 1410 PI->second.erase(p); 1411 1412 p->replaceAllUsesWith(constVal); 1413 if (isa<PointerType>(constVal->getType())) 1414 MD->invalidateCachedPointerInfo(constVal); 1415 VN.erase(p); 1416 1417 toErase.push_back(p); 1418 } else { 1419 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1420 } 1421 1422 // If the number we were assigned was a brand new VN, then we don't 1423 // need to do a lookup to see if the number already exists 1424 // somewhere in the domtree: it can't! 1425 } else if (num == nextNum) { 1426 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1427 1428 // Perform fast-path value-number based elimination of values inherited from 1429 // dominators. 1430 } else if (Value* repl = lookupNumber(I->getParent(), num)) { 1431 // Remove it! 1432 VN.erase(I); 1433 I->replaceAllUsesWith(repl); 1434 if (isa<PointerType>(repl->getType())) 1435 MD->invalidateCachedPointerInfo(repl); 1436 toErase.push_back(I); 1437 return true; 1438 1439#if 0 1440 // Perform slow-pathvalue-number based elimination with phi construction. 1441 } else if (Value* repl = AttemptRedundancyElimination(I, num)) { 1442 // Remove it! 1443 VN.erase(I); 1444 I->replaceAllUsesWith(repl); 1445 if (isa<PointerType>(repl->getType())) 1446 MD->invalidateCachedPointerInfo(repl); 1447 toErase.push_back(I); 1448 return true; 1449#endif 1450 } else { 1451 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1452 } 1453 1454 return false; 1455} 1456 1457/// runOnFunction - This is the main transformation entry point for a function. 1458bool GVN::runOnFunction(Function& F) { 1459 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1460 DT = &getAnalysis<DominatorTree>(); 1461 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1462 VN.setMemDep(MD); 1463 VN.setDomTree(DT); 1464 1465 bool changed = false; 1466 bool shouldContinue = true; 1467 1468 // Merge unconditional branches, allowing PRE to catch more 1469 // optimization opportunities. 1470 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1471 BasicBlock* BB = FI; 1472 ++FI; 1473 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1474 if (removedBlock) NumGVNBlocks++; 1475 1476 changed |= removedBlock; 1477 } 1478 1479 unsigned Iteration = 0; 1480 1481 while (shouldContinue) { 1482 DEBUG(errs() << "GVN iteration: " << Iteration << "\n"); 1483 shouldContinue = iterateOnFunction(F); 1484 changed |= shouldContinue; 1485 ++Iteration; 1486 } 1487 1488 if (EnablePRE) { 1489 bool PREChanged = true; 1490 while (PREChanged) { 1491 PREChanged = performPRE(F); 1492 changed |= PREChanged; 1493 } 1494 } 1495 // FIXME: Should perform GVN again after PRE does something. PRE can move 1496 // computations into blocks where they become fully redundant. Note that 1497 // we can't do this until PRE's critical edge splitting updates memdep. 1498 // Actually, when this happens, we should just fully integrate PRE into GVN. 1499 1500 cleanupGlobalSets(); 1501 1502 return changed; 1503} 1504 1505 1506bool GVN::processBlock(BasicBlock* BB) { 1507 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1508 // incrementing BI before processing an instruction). 1509 SmallVector<Instruction*, 8> toErase; 1510 bool changed_function = false; 1511 1512 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1513 BI != BE;) { 1514 changed_function |= processInstruction(BI, toErase); 1515 if (toErase.empty()) { 1516 ++BI; 1517 continue; 1518 } 1519 1520 // If we need some instructions deleted, do it now. 1521 NumGVNInstr += toErase.size(); 1522 1523 // Avoid iterator invalidation. 1524 bool AtStart = BI == BB->begin(); 1525 if (!AtStart) 1526 --BI; 1527 1528 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1529 E = toErase.end(); I != E; ++I) { 1530 DEBUG(errs() << "GVN removed: " << **I); 1531 MD->removeInstruction(*I); 1532 (*I)->eraseFromParent(); 1533 DEBUG(verifyRemoved(*I)); 1534 } 1535 toErase.clear(); 1536 1537 if (AtStart) 1538 BI = BB->begin(); 1539 else 1540 ++BI; 1541 } 1542 1543 return changed_function; 1544} 1545 1546/// performPRE - Perform a purely local form of PRE that looks for diamond 1547/// control flow patterns and attempts to perform simple PRE at the join point. 1548bool GVN::performPRE(Function& F) { 1549 bool Changed = false; 1550 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1551 DenseMap<BasicBlock*, Value*> predMap; 1552 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1553 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1554 BasicBlock* CurrentBlock = *DI; 1555 1556 // Nothing to PRE in the entry block. 1557 if (CurrentBlock == &F.getEntryBlock()) continue; 1558 1559 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1560 BE = CurrentBlock->end(); BI != BE; ) { 1561 Instruction *CurInst = BI++; 1562 1563 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) || 1564 isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) || 1565 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1566 isa<DbgInfoIntrinsic>(CurInst)) 1567 continue; 1568 1569 uint32_t valno = VN.lookup(CurInst); 1570 1571 // Look for the predecessors for PRE opportunities. We're 1572 // only trying to solve the basic diamond case, where 1573 // a value is computed in the successor and one predecessor, 1574 // but not the other. We also explicitly disallow cases 1575 // where the successor is its own predecessor, because they're 1576 // more complicated to get right. 1577 unsigned numWith = 0; 1578 unsigned numWithout = 0; 1579 BasicBlock* PREPred = 0; 1580 predMap.clear(); 1581 1582 for (pred_iterator PI = pred_begin(CurrentBlock), 1583 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1584 // We're not interested in PRE where the block is its 1585 // own predecessor, on in blocks with predecessors 1586 // that are not reachable. 1587 if (*PI == CurrentBlock) { 1588 numWithout = 2; 1589 break; 1590 } else if (!localAvail.count(*PI)) { 1591 numWithout = 2; 1592 break; 1593 } 1594 1595 DenseMap<uint32_t, Value*>::iterator predV = 1596 localAvail[*PI]->table.find(valno); 1597 if (predV == localAvail[*PI]->table.end()) { 1598 PREPred = *PI; 1599 numWithout++; 1600 } else if (predV->second == CurInst) { 1601 numWithout = 2; 1602 } else { 1603 predMap[*PI] = predV->second; 1604 numWith++; 1605 } 1606 } 1607 1608 // Don't do PRE when it might increase code size, i.e. when 1609 // we would need to insert instructions in more than one pred. 1610 if (numWithout != 1 || numWith == 0) 1611 continue; 1612 1613 // We can't do PRE safely on a critical edge, so instead we schedule 1614 // the edge to be split and perform the PRE the next time we iterate 1615 // on the function. 1616 unsigned succNum = 0; 1617 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1618 i != e; ++i) 1619 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1620 succNum = i; 1621 break; 1622 } 1623 1624 if (isCriticalEdge(PREPred->getTerminator(), succNum)) { 1625 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum)); 1626 continue; 1627 } 1628 1629 // Instantiate the expression the in predecessor that lacked it. 1630 // Because we are going top-down through the block, all value numbers 1631 // will be available in the predecessor by the time we need them. Any 1632 // that weren't original present will have been instantiated earlier 1633 // in this loop. 1634 Instruction* PREInstr = CurInst->clone(CurInst->getContext()); 1635 bool success = true; 1636 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1637 Value *Op = PREInstr->getOperand(i); 1638 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1639 continue; 1640 1641 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1642 PREInstr->setOperand(i, V); 1643 } else { 1644 success = false; 1645 break; 1646 } 1647 } 1648 1649 // Fail out if we encounter an operand that is not available in 1650 // the PRE predecessor. This is typically because of loads which 1651 // are not value numbered precisely. 1652 if (!success) { 1653 delete PREInstr; 1654 DEBUG(verifyRemoved(PREInstr)); 1655 continue; 1656 } 1657 1658 PREInstr->insertBefore(PREPred->getTerminator()); 1659 PREInstr->setName(CurInst->getName() + ".pre"); 1660 predMap[PREPred] = PREInstr; 1661 VN.add(PREInstr, valno); 1662 NumGVNPRE++; 1663 1664 // Update the availability map to include the new instruction. 1665 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr)); 1666 1667 // Create a PHI to make the value available in this block. 1668 PHINode* Phi = PHINode::Create(CurInst->getType(), 1669 CurInst->getName() + ".pre-phi", 1670 CurrentBlock->begin()); 1671 for (pred_iterator PI = pred_begin(CurrentBlock), 1672 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1673 Phi->addIncoming(predMap[*PI], *PI); 1674 1675 VN.add(Phi, valno); 1676 localAvail[CurrentBlock]->table[valno] = Phi; 1677 1678 CurInst->replaceAllUsesWith(Phi); 1679 if (isa<PointerType>(Phi->getType())) 1680 MD->invalidateCachedPointerInfo(Phi); 1681 VN.erase(CurInst); 1682 1683 DEBUG(errs() << "GVN PRE removed: " << *CurInst); 1684 MD->removeInstruction(CurInst); 1685 CurInst->eraseFromParent(); 1686 DEBUG(verifyRemoved(CurInst)); 1687 Changed = true; 1688 } 1689 } 1690 1691 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1692 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1693 SplitCriticalEdge(I->first, I->second, this); 1694 1695 return Changed || toSplit.size(); 1696} 1697 1698/// iterateOnFunction - Executes one iteration of GVN 1699bool GVN::iterateOnFunction(Function &F) { 1700 cleanupGlobalSets(); 1701 1702 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1703 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1704 if (DI->getIDom()) 1705 localAvail[DI->getBlock()] = 1706 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1707 else 1708 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1709 } 1710 1711 // Top-down walk of the dominator tree 1712 bool changed = false; 1713#if 0 1714 // Needed for value numbering with phi construction to work. 1715 ReversePostOrderTraversal<Function*> RPOT(&F); 1716 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1717 RE = RPOT.end(); RI != RE; ++RI) 1718 changed |= processBlock(*RI); 1719#else 1720 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1721 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1722 changed |= processBlock(DI->getBlock()); 1723#endif 1724 1725 return changed; 1726} 1727 1728void GVN::cleanupGlobalSets() { 1729 VN.clear(); 1730 phiMap.clear(); 1731 1732 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1733 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1734 delete I->second; 1735 localAvail.clear(); 1736} 1737 1738/// verifyRemoved - Verify that the specified instruction does not occur in our 1739/// internal data structures. 1740void GVN::verifyRemoved(const Instruction *Inst) const { 1741 VN.verifyRemoved(Inst); 1742 1743 // Walk through the PHI map to make sure the instruction isn't hiding in there 1744 // somewhere. 1745 for (PhiMapType::iterator 1746 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) { 1747 assert(I->first != Inst && "Inst is still a key in PHI map!"); 1748 1749 for (SmallPtrSet<Instruction*, 4>::iterator 1750 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) { 1751 assert(*II != Inst && "Inst is still a value in PHI map!"); 1752 } 1753 } 1754 1755 // Walk through the value number scope to make sure the instruction isn't 1756 // ferreted away in it. 1757 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1758 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1759 const ValueNumberScope *VNS = I->second; 1760 1761 while (VNS) { 1762 for (DenseMap<uint32_t, Value*>::iterator 1763 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1764 assert(II->second != Inst && "Inst still in value numbering scope!"); 1765 } 1766 1767 VNS = VNS->parent; 1768 } 1769 } 1770} 1771