GVN.cpp revision 333c40096561218bc3597cf153c0a3895274414c
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/Transforms/Utils/BasicBlockUtils.h" 41#include "llvm/Transforms/Utils/Local.h" 42#include <cstdio> 43using namespace llvm; 44 45STATISTIC(NumGVNInstr, "Number of instructions deleted"); 46STATISTIC(NumGVNLoad, "Number of loads deleted"); 47STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 48STATISTIC(NumGVNBlocks, "Number of blocks merged"); 49STATISTIC(NumPRELoad, "Number of loads PRE'd"); 50 51static cl::opt<bool> EnablePRE("enable-pre", 52 cl::init(true), cl::Hidden); 53static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 54 55//===----------------------------------------------------------------------===// 56// ValueTable Class 57//===----------------------------------------------------------------------===// 58 59/// This class holds the mapping between values and value numbers. It is used 60/// as an efficient mechanism to determine the expression-wise equivalence of 61/// two values. 62namespace { 63 struct VISIBILITY_HIDDEN Expression { 64 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 65 UDIV, SDIV, FDIV, UREM, SREM, 66 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 67 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 68 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 69 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 70 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 71 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 72 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 73 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 74 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 75 EMPTY, TOMBSTONE }; 76 77 ExpressionOpcode opcode; 78 const Type* type; 79 uint32_t firstVN; 80 uint32_t secondVN; 81 uint32_t thirdVN; 82 SmallVector<uint32_t, 4> varargs; 83 Value* function; 84 85 Expression() { } 86 Expression(ExpressionOpcode o) : opcode(o) { } 87 88 bool operator==(const Expression &other) const { 89 if (opcode != other.opcode) 90 return false; 91 else if (opcode == EMPTY || opcode == TOMBSTONE) 92 return true; 93 else if (type != other.type) 94 return false; 95 else if (function != other.function) 96 return false; 97 else if (firstVN != other.firstVN) 98 return false; 99 else if (secondVN != other.secondVN) 100 return false; 101 else if (thirdVN != other.thirdVN) 102 return false; 103 else { 104 if (varargs.size() != other.varargs.size()) 105 return false; 106 107 for (size_t i = 0; i < varargs.size(); ++i) 108 if (varargs[i] != other.varargs[i]) 109 return false; 110 111 return true; 112 } 113 } 114 115 bool operator!=(const Expression &other) const { 116 return !(*this == other); 117 } 118 }; 119 120 class VISIBILITY_HIDDEN ValueTable { 121 private: 122 DenseMap<Value*, uint32_t> valueNumbering; 123 DenseMap<Expression, uint32_t> expressionNumbering; 124 AliasAnalysis* AA; 125 MemoryDependenceAnalysis* MD; 126 DominatorTree* DT; 127 128 uint32_t nextValueNumber; 129 130 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 131 Expression::ExpressionOpcode getOpcode(CmpInst* C); 132 Expression::ExpressionOpcode getOpcode(CastInst* C); 133 Expression create_expression(BinaryOperator* BO); 134 Expression create_expression(CmpInst* C); 135 Expression create_expression(ShuffleVectorInst* V); 136 Expression create_expression(ExtractElementInst* C); 137 Expression create_expression(InsertElementInst* V); 138 Expression create_expression(SelectInst* V); 139 Expression create_expression(CastInst* C); 140 Expression create_expression(GetElementPtrInst* G); 141 Expression create_expression(CallInst* C); 142 Expression create_expression(Constant* C); 143 public: 144 ValueTable() : nextValueNumber(1) { } 145 uint32_t lookup_or_add(Value* V); 146 uint32_t lookup(Value* V) const; 147 void add(Value* V, uint32_t num); 148 void clear(); 149 void erase(Value* v); 150 unsigned size(); 151 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 152 AliasAnalysis *getAliasAnalysis() const { return AA; } 153 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 154 void setDomTree(DominatorTree* D) { DT = D; } 155 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 156 void verifyRemoved(const Value *) const; 157 }; 158} 159 160namespace llvm { 161template <> struct DenseMapInfo<Expression> { 162 static inline Expression getEmptyKey() { 163 return Expression(Expression::EMPTY); 164 } 165 166 static inline Expression getTombstoneKey() { 167 return Expression(Expression::TOMBSTONE); 168 } 169 170 static unsigned getHashValue(const Expression e) { 171 unsigned hash = e.opcode; 172 173 hash = e.firstVN + hash * 37; 174 hash = e.secondVN + hash * 37; 175 hash = e.thirdVN + hash * 37; 176 177 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 178 (unsigned)((uintptr_t)e.type >> 9)) + 179 hash * 37; 180 181 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 182 E = e.varargs.end(); I != E; ++I) 183 hash = *I + hash * 37; 184 185 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 186 (unsigned)((uintptr_t)e.function >> 9)) + 187 hash * 37; 188 189 return hash; 190 } 191 static bool isEqual(const Expression &LHS, const Expression &RHS) { 192 return LHS == RHS; 193 } 194 static bool isPod() { return true; } 195}; 196} 197 198//===----------------------------------------------------------------------===// 199// ValueTable Internal Functions 200//===----------------------------------------------------------------------===// 201Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 202 switch(BO->getOpcode()) { 203 default: // THIS SHOULD NEVER HAPPEN 204 assert(0 && "Binary operator with unknown opcode?"); 205 case Instruction::Add: return Expression::ADD; 206 case Instruction::FAdd: return Expression::FADD; 207 case Instruction::Sub: return Expression::SUB; 208 case Instruction::FSub: return Expression::FSUB; 209 case Instruction::Mul: return Expression::MUL; 210 case Instruction::FMul: return Expression::FMUL; 211 case Instruction::UDiv: return Expression::UDIV; 212 case Instruction::SDiv: return Expression::SDIV; 213 case Instruction::FDiv: return Expression::FDIV; 214 case Instruction::URem: return Expression::UREM; 215 case Instruction::SRem: return Expression::SREM; 216 case Instruction::FRem: return Expression::FREM; 217 case Instruction::Shl: return Expression::SHL; 218 case Instruction::LShr: return Expression::LSHR; 219 case Instruction::AShr: return Expression::ASHR; 220 case Instruction::And: return Expression::AND; 221 case Instruction::Or: return Expression::OR; 222 case Instruction::Xor: return Expression::XOR; 223 } 224} 225 226Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 227 if (isa<ICmpInst>(C)) { 228 switch (C->getPredicate()) { 229 default: // THIS SHOULD NEVER HAPPEN 230 assert(0 && "Comparison with unknown predicate?"); 231 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 232 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 233 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 234 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 235 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 236 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 237 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 238 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 239 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 240 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 241 } 242 } else { 243 switch (C->getPredicate()) { 244 default: // THIS SHOULD NEVER HAPPEN 245 assert(0 && "Comparison with unknown predicate?"); 246 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 247 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 248 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 249 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 250 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 251 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 252 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 253 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 254 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 255 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 256 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 257 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 258 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 259 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 260 } 261 } 262} 263 264Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 265 switch(C->getOpcode()) { 266 default: // THIS SHOULD NEVER HAPPEN 267 assert(0 && "Cast operator with unknown opcode?"); 268 case Instruction::Trunc: return Expression::TRUNC; 269 case Instruction::ZExt: return Expression::ZEXT; 270 case Instruction::SExt: return Expression::SEXT; 271 case Instruction::FPToUI: return Expression::FPTOUI; 272 case Instruction::FPToSI: return Expression::FPTOSI; 273 case Instruction::UIToFP: return Expression::UITOFP; 274 case Instruction::SIToFP: return Expression::SITOFP; 275 case Instruction::FPTrunc: return Expression::FPTRUNC; 276 case Instruction::FPExt: return Expression::FPEXT; 277 case Instruction::PtrToInt: return Expression::PTRTOINT; 278 case Instruction::IntToPtr: return Expression::INTTOPTR; 279 case Instruction::BitCast: return Expression::BITCAST; 280 } 281} 282 283Expression ValueTable::create_expression(CallInst* C) { 284 Expression e; 285 286 e.type = C->getType(); 287 e.firstVN = 0; 288 e.secondVN = 0; 289 e.thirdVN = 0; 290 e.function = C->getCalledFunction(); 291 e.opcode = Expression::CALL; 292 293 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 294 I != E; ++I) 295 e.varargs.push_back(lookup_or_add(*I)); 296 297 return e; 298} 299 300Expression ValueTable::create_expression(BinaryOperator* BO) { 301 Expression e; 302 303 e.firstVN = lookup_or_add(BO->getOperand(0)); 304 e.secondVN = lookup_or_add(BO->getOperand(1)); 305 e.thirdVN = 0; 306 e.function = 0; 307 e.type = BO->getType(); 308 e.opcode = getOpcode(BO); 309 310 return e; 311} 312 313Expression ValueTable::create_expression(CmpInst* C) { 314 Expression e; 315 316 e.firstVN = lookup_or_add(C->getOperand(0)); 317 e.secondVN = lookup_or_add(C->getOperand(1)); 318 e.thirdVN = 0; 319 e.function = 0; 320 e.type = C->getType(); 321 e.opcode = getOpcode(C); 322 323 return e; 324} 325 326Expression ValueTable::create_expression(CastInst* C) { 327 Expression e; 328 329 e.firstVN = lookup_or_add(C->getOperand(0)); 330 e.secondVN = 0; 331 e.thirdVN = 0; 332 e.function = 0; 333 e.type = C->getType(); 334 e.opcode = getOpcode(C); 335 336 return e; 337} 338 339Expression ValueTable::create_expression(ShuffleVectorInst* S) { 340 Expression e; 341 342 e.firstVN = lookup_or_add(S->getOperand(0)); 343 e.secondVN = lookup_or_add(S->getOperand(1)); 344 e.thirdVN = lookup_or_add(S->getOperand(2)); 345 e.function = 0; 346 e.type = S->getType(); 347 e.opcode = Expression::SHUFFLE; 348 349 return e; 350} 351 352Expression ValueTable::create_expression(ExtractElementInst* E) { 353 Expression e; 354 355 e.firstVN = lookup_or_add(E->getOperand(0)); 356 e.secondVN = lookup_or_add(E->getOperand(1)); 357 e.thirdVN = 0; 358 e.function = 0; 359 e.type = E->getType(); 360 e.opcode = Expression::EXTRACT; 361 362 return e; 363} 364 365Expression ValueTable::create_expression(InsertElementInst* I) { 366 Expression e; 367 368 e.firstVN = lookup_or_add(I->getOperand(0)); 369 e.secondVN = lookup_or_add(I->getOperand(1)); 370 e.thirdVN = lookup_or_add(I->getOperand(2)); 371 e.function = 0; 372 e.type = I->getType(); 373 e.opcode = Expression::INSERT; 374 375 return e; 376} 377 378Expression ValueTable::create_expression(SelectInst* I) { 379 Expression e; 380 381 e.firstVN = lookup_or_add(I->getCondition()); 382 e.secondVN = lookup_or_add(I->getTrueValue()); 383 e.thirdVN = lookup_or_add(I->getFalseValue()); 384 e.function = 0; 385 e.type = I->getType(); 386 e.opcode = Expression::SELECT; 387 388 return e; 389} 390 391Expression ValueTable::create_expression(GetElementPtrInst* G) { 392 Expression e; 393 394 e.firstVN = lookup_or_add(G->getPointerOperand()); 395 e.secondVN = 0; 396 e.thirdVN = 0; 397 e.function = 0; 398 e.type = G->getType(); 399 e.opcode = Expression::GEP; 400 401 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 402 I != E; ++I) 403 e.varargs.push_back(lookup_or_add(*I)); 404 405 return e; 406} 407 408//===----------------------------------------------------------------------===// 409// ValueTable External Functions 410//===----------------------------------------------------------------------===// 411 412/// add - Insert a value into the table with a specified value number. 413void ValueTable::add(Value* V, uint32_t num) { 414 valueNumbering.insert(std::make_pair(V, num)); 415} 416 417/// lookup_or_add - Returns the value number for the specified value, assigning 418/// it a new number if it did not have one before. 419uint32_t ValueTable::lookup_or_add(Value* V) { 420 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 421 if (VI != valueNumbering.end()) 422 return VI->second; 423 424 if (CallInst* C = dyn_cast<CallInst>(V)) { 425 if (AA->doesNotAccessMemory(C)) { 426 Expression e = create_expression(C); 427 428 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 429 if (EI != expressionNumbering.end()) { 430 valueNumbering.insert(std::make_pair(V, EI->second)); 431 return EI->second; 432 } else { 433 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 434 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 435 436 return nextValueNumber++; 437 } 438 } else if (AA->onlyReadsMemory(C)) { 439 Expression e = create_expression(C); 440 441 if (expressionNumbering.find(e) == expressionNumbering.end()) { 442 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 443 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 444 return nextValueNumber++; 445 } 446 447 MemDepResult local_dep = MD->getDependency(C); 448 449 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 450 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 451 return nextValueNumber++; 452 } 453 454 if (local_dep.isDef()) { 455 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 456 457 if (local_cdep->getNumOperands() != C->getNumOperands()) { 458 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 459 return nextValueNumber++; 460 } 461 462 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 463 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 464 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 465 if (c_vn != cd_vn) { 466 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 467 return nextValueNumber++; 468 } 469 } 470 471 uint32_t v = lookup_or_add(local_cdep); 472 valueNumbering.insert(std::make_pair(V, v)); 473 return v; 474 } 475 476 // Non-local case. 477 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 478 MD->getNonLocalCallDependency(CallSite(C)); 479 // FIXME: call/call dependencies for readonly calls should return def, not 480 // clobber! Move the checking logic to MemDep! 481 CallInst* cdep = 0; 482 483 // Check to see if we have a single dominating call instruction that is 484 // identical to C. 485 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 486 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 487 // Ignore non-local dependencies. 488 if (I->second.isNonLocal()) 489 continue; 490 491 // We don't handle non-depedencies. If we already have a call, reject 492 // instruction dependencies. 493 if (I->second.isClobber() || cdep != 0) { 494 cdep = 0; 495 break; 496 } 497 498 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 499 // FIXME: All duplicated with non-local case. 500 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 501 cdep = NonLocalDepCall; 502 continue; 503 } 504 505 cdep = 0; 506 break; 507 } 508 509 if (!cdep) { 510 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 511 return nextValueNumber++; 512 } 513 514 if (cdep->getNumOperands() != C->getNumOperands()) { 515 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 516 return nextValueNumber++; 517 } 518 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 519 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 520 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 521 if (c_vn != cd_vn) { 522 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 523 return nextValueNumber++; 524 } 525 } 526 527 uint32_t v = lookup_or_add(cdep); 528 valueNumbering.insert(std::make_pair(V, v)); 529 return v; 530 531 } else { 532 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 533 return nextValueNumber++; 534 } 535 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 536 Expression e = create_expression(BO); 537 538 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 539 if (EI != expressionNumbering.end()) { 540 valueNumbering.insert(std::make_pair(V, EI->second)); 541 return EI->second; 542 } else { 543 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 544 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 545 546 return nextValueNumber++; 547 } 548 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 549 Expression e = create_expression(C); 550 551 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 552 if (EI != expressionNumbering.end()) { 553 valueNumbering.insert(std::make_pair(V, EI->second)); 554 return EI->second; 555 } else { 556 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 557 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 558 559 return nextValueNumber++; 560 } 561 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 562 Expression e = create_expression(U); 563 564 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 565 if (EI != expressionNumbering.end()) { 566 valueNumbering.insert(std::make_pair(V, EI->second)); 567 return EI->second; 568 } else { 569 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 570 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 571 572 return nextValueNumber++; 573 } 574 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 575 Expression e = create_expression(U); 576 577 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 578 if (EI != expressionNumbering.end()) { 579 valueNumbering.insert(std::make_pair(V, EI->second)); 580 return EI->second; 581 } else { 582 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 583 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 584 585 return nextValueNumber++; 586 } 587 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 588 Expression e = create_expression(U); 589 590 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 591 if (EI != expressionNumbering.end()) { 592 valueNumbering.insert(std::make_pair(V, EI->second)); 593 return EI->second; 594 } else { 595 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 596 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 597 598 return nextValueNumber++; 599 } 600 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 601 Expression e = create_expression(U); 602 603 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 604 if (EI != expressionNumbering.end()) { 605 valueNumbering.insert(std::make_pair(V, EI->second)); 606 return EI->second; 607 } else { 608 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 609 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 610 611 return nextValueNumber++; 612 } 613 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 614 Expression e = create_expression(U); 615 616 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 617 if (EI != expressionNumbering.end()) { 618 valueNumbering.insert(std::make_pair(V, EI->second)); 619 return EI->second; 620 } else { 621 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 622 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 623 624 return nextValueNumber++; 625 } 626 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 627 Expression e = create_expression(U); 628 629 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 630 if (EI != expressionNumbering.end()) { 631 valueNumbering.insert(std::make_pair(V, EI->second)); 632 return EI->second; 633 } else { 634 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 635 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 636 637 return nextValueNumber++; 638 } 639 } else { 640 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 641 return nextValueNumber++; 642 } 643} 644 645/// lookup - Returns the value number of the specified value. Fails if 646/// the value has not yet been numbered. 647uint32_t ValueTable::lookup(Value* V) const { 648 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 649 assert(VI != valueNumbering.end() && "Value not numbered?"); 650 return VI->second; 651} 652 653/// clear - Remove all entries from the ValueTable 654void ValueTable::clear() { 655 valueNumbering.clear(); 656 expressionNumbering.clear(); 657 nextValueNumber = 1; 658} 659 660/// erase - Remove a value from the value numbering 661void ValueTable::erase(Value* V) { 662 valueNumbering.erase(V); 663} 664 665/// verifyRemoved - Verify that the value is removed from all internal data 666/// structures. 667void ValueTable::verifyRemoved(const Value *V) const { 668 for (DenseMap<Value*, uint32_t>::iterator 669 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 670 assert(I->first != V && "Inst still occurs in value numbering map!"); 671 } 672} 673 674//===----------------------------------------------------------------------===// 675// GVN Pass 676//===----------------------------------------------------------------------===// 677 678namespace { 679 struct VISIBILITY_HIDDEN ValueNumberScope { 680 ValueNumberScope* parent; 681 DenseMap<uint32_t, Value*> table; 682 683 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 684 }; 685} 686 687namespace { 688 689 class VISIBILITY_HIDDEN GVN : public FunctionPass { 690 bool runOnFunction(Function &F); 691 public: 692 static char ID; // Pass identification, replacement for typeid 693 GVN() : FunctionPass(&ID) { } 694 695 private: 696 MemoryDependenceAnalysis *MD; 697 DominatorTree *DT; 698 699 ValueTable VN; 700 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 701 702 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 703 PhiMapType phiMap; 704 705 706 // This transformation requires dominator postdominator info 707 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 708 AU.addRequired<DominatorTree>(); 709 AU.addRequired<MemoryDependenceAnalysis>(); 710 AU.addRequired<AliasAnalysis>(); 711 712 AU.addPreserved<DominatorTree>(); 713 AU.addPreserved<AliasAnalysis>(); 714 } 715 716 // Helper fuctions 717 // FIXME: eliminate or document these better 718 bool processLoad(LoadInst* L, 719 SmallVectorImpl<Instruction*> &toErase); 720 bool processInstruction(Instruction* I, 721 SmallVectorImpl<Instruction*> &toErase); 722 bool processNonLocalLoad(LoadInst* L, 723 SmallVectorImpl<Instruction*> &toErase); 724 bool processBlock(BasicBlock* BB); 725 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig, 726 DenseMap<BasicBlock*, Value*> &Phis, 727 bool top_level = false); 728 void dump(DenseMap<uint32_t, Value*>& d); 729 bool iterateOnFunction(Function &F); 730 Value* CollapsePhi(PHINode* p); 731 bool isSafeReplacement(PHINode* p, Instruction* inst); 732 bool performPRE(Function& F); 733 Value* lookupNumber(BasicBlock* BB, uint32_t num); 734 bool mergeBlockIntoPredecessor(BasicBlock* BB); 735 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno); 736 void cleanupGlobalSets(); 737 void verifyRemoved(const Instruction *I) const; 738 }; 739 740 char GVN::ID = 0; 741} 742 743// createGVNPass - The public interface to this file... 744FunctionPass *llvm::createGVNPass() { return new GVN(); } 745 746static RegisterPass<GVN> X("gvn", 747 "Global Value Numbering"); 748 749void GVN::dump(DenseMap<uint32_t, Value*>& d) { 750 printf("{\n"); 751 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 752 E = d.end(); I != E; ++I) { 753 printf("%d\n", I->first); 754 I->second->dump(); 755 } 756 printf("}\n"); 757} 758 759Value* GVN::CollapsePhi(PHINode* p) { 760 Value* constVal = p->hasConstantValue(); 761 if (!constVal) return 0; 762 763 Instruction* inst = dyn_cast<Instruction>(constVal); 764 if (!inst) 765 return constVal; 766 767 if (DT->dominates(inst, p)) 768 if (isSafeReplacement(p, inst)) 769 return inst; 770 return 0; 771} 772 773bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 774 if (!isa<PHINode>(inst)) 775 return true; 776 777 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 778 UI != E; ++UI) 779 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 780 if (use_phi->getParent() == inst->getParent()) 781 return false; 782 783 return true; 784} 785 786/// GetValueForBlock - Get the value to use within the specified basic block. 787/// available values are in Phis. 788Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig, 789 DenseMap<BasicBlock*, Value*> &Phis, 790 bool top_level) { 791 792 // If we have already computed this value, return the previously computed val. 793 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 794 if (V != Phis.end() && !top_level) return V->second; 795 796 // If the block is unreachable, just return undef, since this path 797 // can't actually occur at runtime. 798 if (!DT->isReachableFromEntry(BB)) 799 return Phis[BB] = Context->getUndef(orig->getType()); 800 801 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 802 Value *ret = GetValueForBlock(Pred, orig, Phis); 803 Phis[BB] = ret; 804 return ret; 805 } 806 807 // Get the number of predecessors of this block so we can reserve space later. 808 // If there is already a PHI in it, use the #preds from it, otherwise count. 809 // Getting it from the PHI is constant time. 810 unsigned NumPreds; 811 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin())) 812 NumPreds = ExistingPN->getNumIncomingValues(); 813 else 814 NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 815 816 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 817 // now, then get values to fill in the incoming values for the PHI. 818 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 819 BB->begin()); 820 PN->reserveOperandSpace(NumPreds); 821 822 Phis.insert(std::make_pair(BB, PN)); 823 824 // Fill in the incoming values for the block. 825 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 826 Value* val = GetValueForBlock(*PI, orig, Phis); 827 PN->addIncoming(val, *PI); 828 } 829 830 VN.getAliasAnalysis()->copyValue(orig, PN); 831 832 // Attempt to collapse PHI nodes that are trivially redundant 833 Value* v = CollapsePhi(PN); 834 if (!v) { 835 // Cache our phi construction results 836 if (LoadInst* L = dyn_cast<LoadInst>(orig)) 837 phiMap[L->getPointerOperand()].insert(PN); 838 else 839 phiMap[orig].insert(PN); 840 841 return PN; 842 } 843 844 PN->replaceAllUsesWith(v); 845 if (isa<PointerType>(v->getType())) 846 MD->invalidateCachedPointerInfo(v); 847 848 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 849 E = Phis.end(); I != E; ++I) 850 if (I->second == PN) 851 I->second = v; 852 853 DEBUG(cerr << "GVN removed: " << *PN); 854 MD->removeInstruction(PN); 855 PN->eraseFromParent(); 856 DEBUG(verifyRemoved(PN)); 857 858 Phis[BB] = v; 859 return v; 860} 861 862/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 863/// we're analyzing is fully available in the specified block. As we go, keep 864/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 865/// map is actually a tri-state map with the following values: 866/// 0) we know the block *is not* fully available. 867/// 1) we know the block *is* fully available. 868/// 2) we do not know whether the block is fully available or not, but we are 869/// currently speculating that it will be. 870/// 3) we are speculating for this block and have used that to speculate for 871/// other blocks. 872static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 873 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 874 // Optimistically assume that the block is fully available and check to see 875 // if we already know about this block in one lookup. 876 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 877 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 878 879 // If the entry already existed for this block, return the precomputed value. 880 if (!IV.second) { 881 // If this is a speculative "available" value, mark it as being used for 882 // speculation of other blocks. 883 if (IV.first->second == 2) 884 IV.first->second = 3; 885 return IV.first->second != 0; 886 } 887 888 // Otherwise, see if it is fully available in all predecessors. 889 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 890 891 // If this block has no predecessors, it isn't live-in here. 892 if (PI == PE) 893 goto SpeculationFailure; 894 895 for (; PI != PE; ++PI) 896 // If the value isn't fully available in one of our predecessors, then it 897 // isn't fully available in this block either. Undo our previous 898 // optimistic assumption and bail out. 899 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 900 goto SpeculationFailure; 901 902 return true; 903 904// SpeculationFailure - If we get here, we found out that this is not, after 905// all, a fully-available block. We have a problem if we speculated on this and 906// used the speculation to mark other blocks as available. 907SpeculationFailure: 908 char &BBVal = FullyAvailableBlocks[BB]; 909 910 // If we didn't speculate on this, just return with it set to false. 911 if (BBVal == 2) { 912 BBVal = 0; 913 return false; 914 } 915 916 // If we did speculate on this value, we could have blocks set to 1 that are 917 // incorrect. Walk the (transitive) successors of this block and mark them as 918 // 0 if set to one. 919 SmallVector<BasicBlock*, 32> BBWorklist; 920 BBWorklist.push_back(BB); 921 922 while (!BBWorklist.empty()) { 923 BasicBlock *Entry = BBWorklist.pop_back_val(); 924 // Note that this sets blocks to 0 (unavailable) if they happen to not 925 // already be in FullyAvailableBlocks. This is safe. 926 char &EntryVal = FullyAvailableBlocks[Entry]; 927 if (EntryVal == 0) continue; // Already unavailable. 928 929 // Mark as unavailable. 930 EntryVal = 0; 931 932 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 933 BBWorklist.push_back(*I); 934 } 935 936 return false; 937} 938 939/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 940/// non-local by performing PHI construction. 941bool GVN::processNonLocalLoad(LoadInst *LI, 942 SmallVectorImpl<Instruction*> &toErase) { 943 // Find the non-local dependencies of the load. 944 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 945 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 946 Deps); 947 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI); 948 949 // If we had to process more than one hundred blocks to find the 950 // dependencies, this load isn't worth worrying about. Optimizing 951 // it will be too expensive. 952 if (Deps.size() > 100) 953 return false; 954 955 // If we had a phi translation failure, we'll have a single entry which is a 956 // clobber in the current block. Reject this early. 957 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 958 DEBUG( 959 DOUT << "GVN: non-local load "; 960 WriteAsOperand(*DOUT.stream(), LI); 961 DOUT << " is clobbered by " << *Deps[0].second.getInst(); 962 ); 963 return false; 964 } 965 966 // Filter out useless results (non-locals, etc). Keep track of the blocks 967 // where we have a value available in repl, also keep track of whether we see 968 // dependencies that produce an unknown value for the load (such as a call 969 // that could potentially clobber the load). 970 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 971 SmallVector<BasicBlock*, 16> UnavailableBlocks; 972 973 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 974 BasicBlock *DepBB = Deps[i].first; 975 MemDepResult DepInfo = Deps[i].second; 976 977 if (DepInfo.isClobber()) { 978 UnavailableBlocks.push_back(DepBB); 979 continue; 980 } 981 982 Instruction *DepInst = DepInfo.getInst(); 983 984 // Loading the allocation -> undef. 985 if (isa<AllocationInst>(DepInst)) { 986 ValuesPerBlock.push_back(std::make_pair(DepBB, 987 Context->getUndef(LI->getType()))); 988 continue; 989 } 990 991 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 992 // Reject loads and stores that are to the same address but are of 993 // different types. 994 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because 995 // of bitfield access, it would be interesting to optimize for it at some 996 // point. 997 if (S->getOperand(0)->getType() != LI->getType()) { 998 UnavailableBlocks.push_back(DepBB); 999 continue; 1000 } 1001 1002 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1003 1004 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) { 1005 if (LD->getType() != LI->getType()) { 1006 UnavailableBlocks.push_back(DepBB); 1007 continue; 1008 } 1009 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1010 } else { 1011 UnavailableBlocks.push_back(DepBB); 1012 continue; 1013 } 1014 } 1015 1016 // If we have no predecessors that produce a known value for this load, exit 1017 // early. 1018 if (ValuesPerBlock.empty()) return false; 1019 1020 // If all of the instructions we depend on produce a known value for this 1021 // load, then it is fully redundant and we can use PHI insertion to compute 1022 // its value. Insert PHIs and remove the fully redundant value now. 1023 if (UnavailableBlocks.empty()) { 1024 // Use cached PHI construction information from previous runs 1025 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1026 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated? 1027 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1028 I != E; ++I) { 1029 if ((*I)->getParent() == LI->getParent()) { 1030 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI); 1031 LI->replaceAllUsesWith(*I); 1032 if (isa<PointerType>((*I)->getType())) 1033 MD->invalidateCachedPointerInfo(*I); 1034 toErase.push_back(LI); 1035 NumGVNLoad++; 1036 return true; 1037 } 1038 1039 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1040 } 1041 1042 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI); 1043 1044 DenseMap<BasicBlock*, Value*> BlockReplValues; 1045 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1046 // Perform PHI construction. 1047 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1048 LI->replaceAllUsesWith(v); 1049 1050 if (isa<PHINode>(v)) 1051 v->takeName(LI); 1052 if (isa<PointerType>(v->getType())) 1053 MD->invalidateCachedPointerInfo(v); 1054 toErase.push_back(LI); 1055 NumGVNLoad++; 1056 return true; 1057 } 1058 1059 if (!EnablePRE || !EnableLoadPRE) 1060 return false; 1061 1062 // Okay, we have *some* definitions of the value. This means that the value 1063 // is available in some of our (transitive) predecessors. Lets think about 1064 // doing PRE of this load. This will involve inserting a new load into the 1065 // predecessor when it's not available. We could do this in general, but 1066 // prefer to not increase code size. As such, we only do this when we know 1067 // that we only have to insert *one* load (which means we're basically moving 1068 // the load, not inserting a new one). 1069 1070 SmallPtrSet<BasicBlock *, 4> Blockers; 1071 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1072 Blockers.insert(UnavailableBlocks[i]); 1073 1074 // Lets find first basic block with more than one predecessor. Walk backwards 1075 // through predecessors if needed. 1076 BasicBlock *LoadBB = LI->getParent(); 1077 BasicBlock *TmpBB = LoadBB; 1078 1079 bool isSinglePred = false; 1080 bool allSingleSucc = true; 1081 while (TmpBB->getSinglePredecessor()) { 1082 isSinglePred = true; 1083 TmpBB = TmpBB->getSinglePredecessor(); 1084 if (!TmpBB) // If haven't found any, bail now. 1085 return false; 1086 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1087 return false; 1088 if (Blockers.count(TmpBB)) 1089 return false; 1090 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1091 allSingleSucc = false; 1092 } 1093 1094 assert(TmpBB); 1095 LoadBB = TmpBB; 1096 1097 // If we have a repl set with LI itself in it, this means we have a loop where 1098 // at least one of the values is LI. Since this means that we won't be able 1099 // to eliminate LI even if we insert uses in the other predecessors, we will 1100 // end up increasing code size. Reject this by scanning for LI. 1101 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1102 if (ValuesPerBlock[i].second == LI) 1103 return false; 1104 1105 if (isSinglePred) { 1106 bool isHot = false; 1107 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1108 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second)) 1109 // "Hot" Instruction is in some loop (because it dominates its dep. 1110 // instruction). 1111 if (DT->dominates(LI, I)) { 1112 isHot = true; 1113 break; 1114 } 1115 1116 // We are interested only in "hot" instructions. We don't want to do any 1117 // mis-optimizations here. 1118 if (!isHot) 1119 return false; 1120 } 1121 1122 // Okay, we have some hope :). Check to see if the loaded value is fully 1123 // available in all but one predecessor. 1124 // FIXME: If we could restructure the CFG, we could make a common pred with 1125 // all the preds that don't have an available LI and insert a new load into 1126 // that one block. 1127 BasicBlock *UnavailablePred = 0; 1128 1129 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1130 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1131 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1132 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1133 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1134 1135 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1136 PI != E; ++PI) { 1137 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1138 continue; 1139 1140 // If this load is not available in multiple predecessors, reject it. 1141 if (UnavailablePred && UnavailablePred != *PI) 1142 return false; 1143 UnavailablePred = *PI; 1144 } 1145 1146 assert(UnavailablePred != 0 && 1147 "Fully available value should be eliminated above!"); 1148 1149 // If the loaded pointer is PHI node defined in this block, do PHI translation 1150 // to get its value in the predecessor. 1151 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1152 1153 // Make sure the value is live in the predecessor. If it was defined by a 1154 // non-PHI instruction in this block, we don't know how to recompute it above. 1155 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1156 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1157 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1158 << *LPInst << *LI << "\n"); 1159 return false; 1160 } 1161 1162 // We don't currently handle critical edges :( 1163 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1164 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1165 << UnavailablePred->getName() << "': " << *LI); 1166 return false; 1167 } 1168 1169 // Make sure it is valid to move this load here. We have to watch out for: 1170 // @1 = getelementptr (i8* p, ... 1171 // test p and branch if == 0 1172 // load @1 1173 // It is valid to have the getelementptr before the test, even if p can be 0, 1174 // as getelementptr only does address arithmetic. 1175 // If we are not pushing the value through any multiple-successor blocks 1176 // we do not have this case. Otherwise, check that the load is safe to 1177 // put anywhere; this can be improved, but should be conservatively safe. 1178 if (!allSingleSucc && 1179 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1180 return false; 1181 1182 // Okay, we can eliminate this load by inserting a reload in the predecessor 1183 // and using PHI construction to get the value in the other predecessors, do 1184 // it. 1185 DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI); 1186 1187 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1188 LI->getAlignment(), 1189 UnavailablePred->getTerminator()); 1190 1191 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1192 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1193 I != E; ++I) 1194 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1195 1196 DenseMap<BasicBlock*, Value*> BlockReplValues; 1197 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1198 BlockReplValues[UnavailablePred] = NewLoad; 1199 1200 // Perform PHI construction. 1201 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1202 LI->replaceAllUsesWith(v); 1203 if (isa<PHINode>(v)) 1204 v->takeName(LI); 1205 if (isa<PointerType>(v->getType())) 1206 MD->invalidateCachedPointerInfo(v); 1207 toErase.push_back(LI); 1208 NumPRELoad++; 1209 return true; 1210} 1211 1212/// processLoad - Attempt to eliminate a load, first by eliminating it 1213/// locally, and then attempting non-local elimination if that fails. 1214bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1215 if (L->isVolatile()) 1216 return false; 1217 1218 Value* pointer = L->getPointerOperand(); 1219 1220 // ... to a pointer that has been loaded from before... 1221 MemDepResult dep = MD->getDependency(L); 1222 1223 // If the value isn't available, don't do anything! 1224 if (dep.isClobber()) { 1225 DEBUG( 1226 // fast print dep, using operator<< on instruction would be too slow 1227 DOUT << "GVN: load "; 1228 WriteAsOperand(*DOUT.stream(), L); 1229 Instruction *I = dep.getInst(); 1230 DOUT << " is clobbered by " << *I; 1231 ); 1232 return false; 1233 } 1234 1235 // If it is defined in another block, try harder. 1236 if (dep.isNonLocal()) 1237 return processNonLocalLoad(L, toErase); 1238 1239 Instruction *DepInst = dep.getInst(); 1240 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1241 // Only forward substitute stores to loads of the same type. 1242 // FIXME: Could do better! 1243 if (DepSI->getPointerOperand()->getType() != pointer->getType()) 1244 return false; 1245 1246 // Remove it! 1247 L->replaceAllUsesWith(DepSI->getOperand(0)); 1248 if (isa<PointerType>(DepSI->getOperand(0)->getType())) 1249 MD->invalidateCachedPointerInfo(DepSI->getOperand(0)); 1250 toErase.push_back(L); 1251 NumGVNLoad++; 1252 return true; 1253 } 1254 1255 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1256 // Only forward substitute stores to loads of the same type. 1257 // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian. 1258 if (DepLI->getType() != L->getType()) 1259 return false; 1260 1261 // Remove it! 1262 L->replaceAllUsesWith(DepLI); 1263 if (isa<PointerType>(DepLI->getType())) 1264 MD->invalidateCachedPointerInfo(DepLI); 1265 toErase.push_back(L); 1266 NumGVNLoad++; 1267 return true; 1268 } 1269 1270 // If this load really doesn't depend on anything, then we must be loading an 1271 // undef value. This can happen when loading for a fresh allocation with no 1272 // intervening stores, for example. 1273 if (isa<AllocationInst>(DepInst)) { 1274 L->replaceAllUsesWith(Context->getUndef(L->getType())); 1275 toErase.push_back(L); 1276 NumGVNLoad++; 1277 return true; 1278 } 1279 1280 return false; 1281} 1282 1283Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) { 1284 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1285 if (I == localAvail.end()) 1286 return 0; 1287 1288 ValueNumberScope* locals = I->second; 1289 1290 while (locals) { 1291 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num); 1292 if (I != locals->table.end()) 1293 return I->second; 1294 else 1295 locals = locals->parent; 1296 } 1297 1298 return 0; 1299} 1300 1301/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination 1302/// by inheritance from the dominator fails, see if we can perform phi 1303/// construction to eliminate the redundancy. 1304Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) { 1305 BasicBlock* BaseBlock = orig->getParent(); 1306 1307 SmallPtrSet<BasicBlock*, 4> Visited; 1308 SmallVector<BasicBlock*, 8> Stack; 1309 Stack.push_back(BaseBlock); 1310 1311 DenseMap<BasicBlock*, Value*> Results; 1312 1313 // Walk backwards through our predecessors, looking for instances of the 1314 // value number we're looking for. Instances are recorded in the Results 1315 // map, which is then used to perform phi construction. 1316 while (!Stack.empty()) { 1317 BasicBlock* Current = Stack.back(); 1318 Stack.pop_back(); 1319 1320 // If we've walked all the way to a proper dominator, then give up. Cases 1321 // where the instance is in the dominator will have been caught by the fast 1322 // path, and any cases that require phi construction further than this are 1323 // probably not worth it anyways. Note that this is a SIGNIFICANT compile 1324 // time improvement. 1325 if (DT->properlyDominates(Current, orig->getParent())) return 0; 1326 1327 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA = 1328 localAvail.find(Current); 1329 if (LA == localAvail.end()) return 0; 1330 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno); 1331 1332 if (V != LA->second->table.end()) { 1333 // Found an instance, record it. 1334 Results.insert(std::make_pair(Current, V->second)); 1335 continue; 1336 } 1337 1338 // If we reach the beginning of the function, then give up. 1339 if (pred_begin(Current) == pred_end(Current)) 1340 return 0; 1341 1342 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current); 1343 PI != PE; ++PI) 1344 if (Visited.insert(*PI)) 1345 Stack.push_back(*PI); 1346 } 1347 1348 // If we didn't find instances, give up. Otherwise, perform phi construction. 1349 if (Results.size() == 0) 1350 return 0; 1351 else 1352 return GetValueForBlock(BaseBlock, orig, Results, true); 1353} 1354 1355/// processInstruction - When calculating availability, handle an instruction 1356/// by inserting it into the appropriate sets 1357bool GVN::processInstruction(Instruction *I, 1358 SmallVectorImpl<Instruction*> &toErase) { 1359 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 1360 bool changed = processLoad(L, toErase); 1361 1362 if (!changed) { 1363 unsigned num = VN.lookup_or_add(L); 1364 localAvail[I->getParent()]->table.insert(std::make_pair(num, L)); 1365 } 1366 1367 return changed; 1368 } 1369 1370 uint32_t nextNum = VN.getNextUnusedValueNumber(); 1371 unsigned num = VN.lookup_or_add(I); 1372 1373 if (BranchInst* BI = dyn_cast<BranchInst>(I)) { 1374 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1375 1376 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1377 return false; 1378 1379 Value* branchCond = BI->getCondition(); 1380 uint32_t condVN = VN.lookup_or_add(branchCond); 1381 1382 BasicBlock* trueSucc = BI->getSuccessor(0); 1383 BasicBlock* falseSucc = BI->getSuccessor(1); 1384 1385 if (trueSucc->getSinglePredecessor()) 1386 localAvail[trueSucc]->table[condVN] = Context->getConstantIntTrue(); 1387 if (falseSucc->getSinglePredecessor()) 1388 localAvail[falseSucc]->table[condVN] = Context->getConstantIntFalse(); 1389 1390 return false; 1391 1392 // Allocations are always uniquely numbered, so we can save time and memory 1393 // by fast failing them. 1394 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) { 1395 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1396 return false; 1397 } 1398 1399 // Collapse PHI nodes 1400 if (PHINode* p = dyn_cast<PHINode>(I)) { 1401 Value* constVal = CollapsePhi(p); 1402 1403 if (constVal) { 1404 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1405 PI != PE; ++PI) 1406 PI->second.erase(p); 1407 1408 p->replaceAllUsesWith(constVal); 1409 if (isa<PointerType>(constVal->getType())) 1410 MD->invalidateCachedPointerInfo(constVal); 1411 VN.erase(p); 1412 1413 toErase.push_back(p); 1414 } else { 1415 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1416 } 1417 1418 // If the number we were assigned was a brand new VN, then we don't 1419 // need to do a lookup to see if the number already exists 1420 // somewhere in the domtree: it can't! 1421 } else if (num == nextNum) { 1422 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1423 1424 // Perform fast-path value-number based elimination of values inherited from 1425 // dominators. 1426 } else if (Value* repl = lookupNumber(I->getParent(), num)) { 1427 // Remove it! 1428 VN.erase(I); 1429 I->replaceAllUsesWith(repl); 1430 if (isa<PointerType>(repl->getType())) 1431 MD->invalidateCachedPointerInfo(repl); 1432 toErase.push_back(I); 1433 return true; 1434 1435#if 0 1436 // Perform slow-pathvalue-number based elimination with phi construction. 1437 } else if (Value* repl = AttemptRedundancyElimination(I, num)) { 1438 // Remove it! 1439 VN.erase(I); 1440 I->replaceAllUsesWith(repl); 1441 if (isa<PointerType>(repl->getType())) 1442 MD->invalidateCachedPointerInfo(repl); 1443 toErase.push_back(I); 1444 return true; 1445#endif 1446 } else { 1447 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1448 } 1449 1450 return false; 1451} 1452 1453/// runOnFunction - This is the main transformation entry point for a function. 1454bool GVN::runOnFunction(Function& F) { 1455 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1456 DT = &getAnalysis<DominatorTree>(); 1457 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1458 VN.setMemDep(MD); 1459 VN.setDomTree(DT); 1460 1461 bool changed = false; 1462 bool shouldContinue = true; 1463 1464 // Merge unconditional branches, allowing PRE to catch more 1465 // optimization opportunities. 1466 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1467 BasicBlock* BB = FI; 1468 ++FI; 1469 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1470 if (removedBlock) NumGVNBlocks++; 1471 1472 changed |= removedBlock; 1473 } 1474 1475 unsigned Iteration = 0; 1476 1477 while (shouldContinue) { 1478 DEBUG(cerr << "GVN iteration: " << Iteration << "\n"); 1479 shouldContinue = iterateOnFunction(F); 1480 changed |= shouldContinue; 1481 ++Iteration; 1482 } 1483 1484 if (EnablePRE) { 1485 bool PREChanged = true; 1486 while (PREChanged) { 1487 PREChanged = performPRE(F); 1488 changed |= PREChanged; 1489 } 1490 } 1491 // FIXME: Should perform GVN again after PRE does something. PRE can move 1492 // computations into blocks where they become fully redundant. Note that 1493 // we can't do this until PRE's critical edge splitting updates memdep. 1494 // Actually, when this happens, we should just fully integrate PRE into GVN. 1495 1496 cleanupGlobalSets(); 1497 1498 return changed; 1499} 1500 1501 1502bool GVN::processBlock(BasicBlock* BB) { 1503 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1504 // incrementing BI before processing an instruction). 1505 SmallVector<Instruction*, 8> toErase; 1506 bool changed_function = false; 1507 1508 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1509 BI != BE;) { 1510 changed_function |= processInstruction(BI, toErase); 1511 if (toErase.empty()) { 1512 ++BI; 1513 continue; 1514 } 1515 1516 // If we need some instructions deleted, do it now. 1517 NumGVNInstr += toErase.size(); 1518 1519 // Avoid iterator invalidation. 1520 bool AtStart = BI == BB->begin(); 1521 if (!AtStart) 1522 --BI; 1523 1524 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1525 E = toErase.end(); I != E; ++I) { 1526 DEBUG(cerr << "GVN removed: " << **I); 1527 MD->removeInstruction(*I); 1528 (*I)->eraseFromParent(); 1529 DEBUG(verifyRemoved(*I)); 1530 } 1531 toErase.clear(); 1532 1533 if (AtStart) 1534 BI = BB->begin(); 1535 else 1536 ++BI; 1537 } 1538 1539 return changed_function; 1540} 1541 1542/// performPRE - Perform a purely local form of PRE that looks for diamond 1543/// control flow patterns and attempts to perform simple PRE at the join point. 1544bool GVN::performPRE(Function& F) { 1545 bool Changed = false; 1546 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1547 DenseMap<BasicBlock*, Value*> predMap; 1548 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1549 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1550 BasicBlock* CurrentBlock = *DI; 1551 1552 // Nothing to PRE in the entry block. 1553 if (CurrentBlock == &F.getEntryBlock()) continue; 1554 1555 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1556 BE = CurrentBlock->end(); BI != BE; ) { 1557 Instruction *CurInst = BI++; 1558 1559 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) || 1560 isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) || 1561 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1562 isa<DbgInfoIntrinsic>(CurInst)) 1563 continue; 1564 1565 uint32_t valno = VN.lookup(CurInst); 1566 1567 // Look for the predecessors for PRE opportunities. We're 1568 // only trying to solve the basic diamond case, where 1569 // a value is computed in the successor and one predecessor, 1570 // but not the other. We also explicitly disallow cases 1571 // where the successor is its own predecessor, because they're 1572 // more complicated to get right. 1573 unsigned numWith = 0; 1574 unsigned numWithout = 0; 1575 BasicBlock* PREPred = 0; 1576 predMap.clear(); 1577 1578 for (pred_iterator PI = pred_begin(CurrentBlock), 1579 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1580 // We're not interested in PRE where the block is its 1581 // own predecessor, on in blocks with predecessors 1582 // that are not reachable. 1583 if (*PI == CurrentBlock) { 1584 numWithout = 2; 1585 break; 1586 } else if (!localAvail.count(*PI)) { 1587 numWithout = 2; 1588 break; 1589 } 1590 1591 DenseMap<uint32_t, Value*>::iterator predV = 1592 localAvail[*PI]->table.find(valno); 1593 if (predV == localAvail[*PI]->table.end()) { 1594 PREPred = *PI; 1595 numWithout++; 1596 } else if (predV->second == CurInst) { 1597 numWithout = 2; 1598 } else { 1599 predMap[*PI] = predV->second; 1600 numWith++; 1601 } 1602 } 1603 1604 // Don't do PRE when it might increase code size, i.e. when 1605 // we would need to insert instructions in more than one pred. 1606 if (numWithout != 1 || numWith == 0) 1607 continue; 1608 1609 // We can't do PRE safely on a critical edge, so instead we schedule 1610 // the edge to be split and perform the PRE the next time we iterate 1611 // on the function. 1612 unsigned succNum = 0; 1613 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1614 i != e; ++i) 1615 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1616 succNum = i; 1617 break; 1618 } 1619 1620 if (isCriticalEdge(PREPred->getTerminator(), succNum)) { 1621 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum)); 1622 continue; 1623 } 1624 1625 // Instantiate the expression the in predecessor that lacked it. 1626 // Because we are going top-down through the block, all value numbers 1627 // will be available in the predecessor by the time we need them. Any 1628 // that weren't original present will have been instantiated earlier 1629 // in this loop. 1630 Instruction* PREInstr = CurInst->clone(*Context); 1631 bool success = true; 1632 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1633 Value *Op = PREInstr->getOperand(i); 1634 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1635 continue; 1636 1637 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1638 PREInstr->setOperand(i, V); 1639 } else { 1640 success = false; 1641 break; 1642 } 1643 } 1644 1645 // Fail out if we encounter an operand that is not available in 1646 // the PRE predecessor. This is typically because of loads which 1647 // are not value numbered precisely. 1648 if (!success) { 1649 delete PREInstr; 1650 DEBUG(verifyRemoved(PREInstr)); 1651 continue; 1652 } 1653 1654 PREInstr->insertBefore(PREPred->getTerminator()); 1655 PREInstr->setName(CurInst->getName() + ".pre"); 1656 predMap[PREPred] = PREInstr; 1657 VN.add(PREInstr, valno); 1658 NumGVNPRE++; 1659 1660 // Update the availability map to include the new instruction. 1661 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr)); 1662 1663 // Create a PHI to make the value available in this block. 1664 PHINode* Phi = PHINode::Create(CurInst->getType(), 1665 CurInst->getName() + ".pre-phi", 1666 CurrentBlock->begin()); 1667 for (pred_iterator PI = pred_begin(CurrentBlock), 1668 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1669 Phi->addIncoming(predMap[*PI], *PI); 1670 1671 VN.add(Phi, valno); 1672 localAvail[CurrentBlock]->table[valno] = Phi; 1673 1674 CurInst->replaceAllUsesWith(Phi); 1675 if (isa<PointerType>(Phi->getType())) 1676 MD->invalidateCachedPointerInfo(Phi); 1677 VN.erase(CurInst); 1678 1679 DEBUG(cerr << "GVN PRE removed: " << *CurInst); 1680 MD->removeInstruction(CurInst); 1681 CurInst->eraseFromParent(); 1682 DEBUG(verifyRemoved(CurInst)); 1683 Changed = true; 1684 } 1685 } 1686 1687 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1688 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1689 SplitCriticalEdge(I->first, I->second, this); 1690 1691 return Changed || toSplit.size(); 1692} 1693 1694/// iterateOnFunction - Executes one iteration of GVN 1695bool GVN::iterateOnFunction(Function &F) { 1696 cleanupGlobalSets(); 1697 1698 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1699 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1700 if (DI->getIDom()) 1701 localAvail[DI->getBlock()] = 1702 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1703 else 1704 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1705 } 1706 1707 // Top-down walk of the dominator tree 1708 bool changed = false; 1709#if 0 1710 // Needed for value numbering with phi construction to work. 1711 ReversePostOrderTraversal<Function*> RPOT(&F); 1712 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1713 RE = RPOT.end(); RI != RE; ++RI) 1714 changed |= processBlock(*RI); 1715#else 1716 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1717 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1718 changed |= processBlock(DI->getBlock()); 1719#endif 1720 1721 return changed; 1722} 1723 1724void GVN::cleanupGlobalSets() { 1725 VN.clear(); 1726 phiMap.clear(); 1727 1728 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1729 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1730 delete I->second; 1731 localAvail.clear(); 1732} 1733 1734/// verifyRemoved - Verify that the specified instruction does not occur in our 1735/// internal data structures. 1736void GVN::verifyRemoved(const Instruction *Inst) const { 1737 VN.verifyRemoved(Inst); 1738 1739 // Walk through the PHI map to make sure the instruction isn't hiding in there 1740 // somewhere. 1741 for (PhiMapType::iterator 1742 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) { 1743 assert(I->first != Inst && "Inst is still a key in PHI map!"); 1744 1745 for (SmallPtrSet<Instruction*, 4>::iterator 1746 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) { 1747 assert(*II != Inst && "Inst is still a value in PHI map!"); 1748 } 1749 } 1750 1751 // Walk through the value number scope to make sure the instruction isn't 1752 // ferreted away in it. 1753 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1754 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1755 const ValueNumberScope *VNS = I->second; 1756 1757 while (VNS) { 1758 for (DenseMap<uint32_t, Value*>::iterator 1759 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1760 assert(II->second != Inst && "Inst still in value numbering scope!"); 1761 } 1762 1763 VNS = VNS->parent; 1764 } 1765 } 1766} 1767