GVN.cpp revision c25e7581b9b8088910da31702d4ca21c4734c6d7
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/Transforms/Utils/BasicBlockUtils.h" 42#include "llvm/Transforms/Utils/Local.h" 43#include <cstdio> 44using namespace llvm; 45 46STATISTIC(NumGVNInstr, "Number of instructions deleted"); 47STATISTIC(NumGVNLoad, "Number of loads deleted"); 48STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 49STATISTIC(NumGVNBlocks, "Number of blocks merged"); 50STATISTIC(NumPRELoad, "Number of loads PRE'd"); 51 52static cl::opt<bool> EnablePRE("enable-pre", 53 cl::init(true), cl::Hidden); 54static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 55 56//===----------------------------------------------------------------------===// 57// ValueTable Class 58//===----------------------------------------------------------------------===// 59 60/// This class holds the mapping between values and value numbers. It is used 61/// as an efficient mechanism to determine the expression-wise equivalence of 62/// two values. 63namespace { 64 struct VISIBILITY_HIDDEN Expression { 65 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 66 UDIV, SDIV, FDIV, UREM, SREM, 67 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 68 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 69 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 70 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 71 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 72 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 73 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 74 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 75 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 76 EMPTY, TOMBSTONE }; 77 78 ExpressionOpcode opcode; 79 const Type* type; 80 uint32_t firstVN; 81 uint32_t secondVN; 82 uint32_t thirdVN; 83 SmallVector<uint32_t, 4> varargs; 84 Value* function; 85 86 Expression() { } 87 Expression(ExpressionOpcode o) : opcode(o) { } 88 89 bool operator==(const Expression &other) const { 90 if (opcode != other.opcode) 91 return false; 92 else if (opcode == EMPTY || opcode == TOMBSTONE) 93 return true; 94 else if (type != other.type) 95 return false; 96 else if (function != other.function) 97 return false; 98 else if (firstVN != other.firstVN) 99 return false; 100 else if (secondVN != other.secondVN) 101 return false; 102 else if (thirdVN != other.thirdVN) 103 return false; 104 else { 105 if (varargs.size() != other.varargs.size()) 106 return false; 107 108 for (size_t i = 0; i < varargs.size(); ++i) 109 if (varargs[i] != other.varargs[i]) 110 return false; 111 112 return true; 113 } 114 } 115 116 bool operator!=(const Expression &other) const { 117 return !(*this == other); 118 } 119 }; 120 121 class VISIBILITY_HIDDEN ValueTable { 122 private: 123 DenseMap<Value*, uint32_t> valueNumbering; 124 DenseMap<Expression, uint32_t> expressionNumbering; 125 AliasAnalysis* AA; 126 MemoryDependenceAnalysis* MD; 127 DominatorTree* DT; 128 129 uint32_t nextValueNumber; 130 131 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 132 Expression::ExpressionOpcode getOpcode(CmpInst* C); 133 Expression::ExpressionOpcode getOpcode(CastInst* C); 134 Expression create_expression(BinaryOperator* BO); 135 Expression create_expression(CmpInst* C); 136 Expression create_expression(ShuffleVectorInst* V); 137 Expression create_expression(ExtractElementInst* C); 138 Expression create_expression(InsertElementInst* V); 139 Expression create_expression(SelectInst* V); 140 Expression create_expression(CastInst* C); 141 Expression create_expression(GetElementPtrInst* G); 142 Expression create_expression(CallInst* C); 143 Expression create_expression(Constant* C); 144 public: 145 ValueTable() : nextValueNumber(1) { } 146 uint32_t lookup_or_add(Value* V); 147 uint32_t lookup(Value* V) const; 148 void add(Value* V, uint32_t num); 149 void clear(); 150 void erase(Value* v); 151 unsigned size(); 152 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 153 AliasAnalysis *getAliasAnalysis() const { return AA; } 154 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 155 void setDomTree(DominatorTree* D) { DT = D; } 156 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 157 void verifyRemoved(const Value *) const; 158 }; 159} 160 161namespace llvm { 162template <> struct DenseMapInfo<Expression> { 163 static inline Expression getEmptyKey() { 164 return Expression(Expression::EMPTY); 165 } 166 167 static inline Expression getTombstoneKey() { 168 return Expression(Expression::TOMBSTONE); 169 } 170 171 static unsigned getHashValue(const Expression e) { 172 unsigned hash = e.opcode; 173 174 hash = e.firstVN + hash * 37; 175 hash = e.secondVN + hash * 37; 176 hash = e.thirdVN + hash * 37; 177 178 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 179 (unsigned)((uintptr_t)e.type >> 9)) + 180 hash * 37; 181 182 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 183 E = e.varargs.end(); I != E; ++I) 184 hash = *I + hash * 37; 185 186 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 187 (unsigned)((uintptr_t)e.function >> 9)) + 188 hash * 37; 189 190 return hash; 191 } 192 static bool isEqual(const Expression &LHS, const Expression &RHS) { 193 return LHS == RHS; 194 } 195 static bool isPod() { return true; } 196}; 197} 198 199//===----------------------------------------------------------------------===// 200// ValueTable Internal Functions 201//===----------------------------------------------------------------------===// 202Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 203 switch(BO->getOpcode()) { 204 default: // THIS SHOULD NEVER HAPPEN 205 LLVM_UNREACHABLE("Binary operator with unknown opcode?"); 206 case Instruction::Add: return Expression::ADD; 207 case Instruction::FAdd: return Expression::FADD; 208 case Instruction::Sub: return Expression::SUB; 209 case Instruction::FSub: return Expression::FSUB; 210 case Instruction::Mul: return Expression::MUL; 211 case Instruction::FMul: return Expression::FMUL; 212 case Instruction::UDiv: return Expression::UDIV; 213 case Instruction::SDiv: return Expression::SDIV; 214 case Instruction::FDiv: return Expression::FDIV; 215 case Instruction::URem: return Expression::UREM; 216 case Instruction::SRem: return Expression::SREM; 217 case Instruction::FRem: return Expression::FREM; 218 case Instruction::Shl: return Expression::SHL; 219 case Instruction::LShr: return Expression::LSHR; 220 case Instruction::AShr: return Expression::ASHR; 221 case Instruction::And: return Expression::AND; 222 case Instruction::Or: return Expression::OR; 223 case Instruction::Xor: return Expression::XOR; 224 } 225} 226 227Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 228 if (isa<ICmpInst>(C)) { 229 switch (C->getPredicate()) { 230 default: // THIS SHOULD NEVER HAPPEN 231 LLVM_UNREACHABLE("Comparison with unknown predicate?"); 232 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 233 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 234 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 235 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 236 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 237 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 238 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 239 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 240 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 241 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 242 } 243 } else { 244 switch (C->getPredicate()) { 245 default: // THIS SHOULD NEVER HAPPEN 246 LLVM_UNREACHABLE("Comparison with unknown predicate?"); 247 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 248 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 249 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 250 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 251 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 252 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 253 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 254 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 255 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 256 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 257 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 258 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 259 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 260 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 261 } 262 } 263} 264 265Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 266 switch(C->getOpcode()) { 267 default: // THIS SHOULD NEVER HAPPEN 268 LLVM_UNREACHABLE("Cast operator with unknown opcode?"); 269 case Instruction::Trunc: return Expression::TRUNC; 270 case Instruction::ZExt: return Expression::ZEXT; 271 case Instruction::SExt: return Expression::SEXT; 272 case Instruction::FPToUI: return Expression::FPTOUI; 273 case Instruction::FPToSI: return Expression::FPTOSI; 274 case Instruction::UIToFP: return Expression::UITOFP; 275 case Instruction::SIToFP: return Expression::SITOFP; 276 case Instruction::FPTrunc: return Expression::FPTRUNC; 277 case Instruction::FPExt: return Expression::FPEXT; 278 case Instruction::PtrToInt: return Expression::PTRTOINT; 279 case Instruction::IntToPtr: return Expression::INTTOPTR; 280 case Instruction::BitCast: return Expression::BITCAST; 281 } 282} 283 284Expression ValueTable::create_expression(CallInst* C) { 285 Expression e; 286 287 e.type = C->getType(); 288 e.firstVN = 0; 289 e.secondVN = 0; 290 e.thirdVN = 0; 291 e.function = C->getCalledFunction(); 292 e.opcode = Expression::CALL; 293 294 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 295 I != E; ++I) 296 e.varargs.push_back(lookup_or_add(*I)); 297 298 return e; 299} 300 301Expression ValueTable::create_expression(BinaryOperator* BO) { 302 Expression e; 303 304 e.firstVN = lookup_or_add(BO->getOperand(0)); 305 e.secondVN = lookup_or_add(BO->getOperand(1)); 306 e.thirdVN = 0; 307 e.function = 0; 308 e.type = BO->getType(); 309 e.opcode = getOpcode(BO); 310 311 return e; 312} 313 314Expression ValueTable::create_expression(CmpInst* C) { 315 Expression e; 316 317 e.firstVN = lookup_or_add(C->getOperand(0)); 318 e.secondVN = lookup_or_add(C->getOperand(1)); 319 e.thirdVN = 0; 320 e.function = 0; 321 e.type = C->getType(); 322 e.opcode = getOpcode(C); 323 324 return e; 325} 326 327Expression ValueTable::create_expression(CastInst* C) { 328 Expression e; 329 330 e.firstVN = lookup_or_add(C->getOperand(0)); 331 e.secondVN = 0; 332 e.thirdVN = 0; 333 e.function = 0; 334 e.type = C->getType(); 335 e.opcode = getOpcode(C); 336 337 return e; 338} 339 340Expression ValueTable::create_expression(ShuffleVectorInst* S) { 341 Expression e; 342 343 e.firstVN = lookup_or_add(S->getOperand(0)); 344 e.secondVN = lookup_or_add(S->getOperand(1)); 345 e.thirdVN = lookup_or_add(S->getOperand(2)); 346 e.function = 0; 347 e.type = S->getType(); 348 e.opcode = Expression::SHUFFLE; 349 350 return e; 351} 352 353Expression ValueTable::create_expression(ExtractElementInst* E) { 354 Expression e; 355 356 e.firstVN = lookup_or_add(E->getOperand(0)); 357 e.secondVN = lookup_or_add(E->getOperand(1)); 358 e.thirdVN = 0; 359 e.function = 0; 360 e.type = E->getType(); 361 e.opcode = Expression::EXTRACT; 362 363 return e; 364} 365 366Expression ValueTable::create_expression(InsertElementInst* I) { 367 Expression e; 368 369 e.firstVN = lookup_or_add(I->getOperand(0)); 370 e.secondVN = lookup_or_add(I->getOperand(1)); 371 e.thirdVN = lookup_or_add(I->getOperand(2)); 372 e.function = 0; 373 e.type = I->getType(); 374 e.opcode = Expression::INSERT; 375 376 return e; 377} 378 379Expression ValueTable::create_expression(SelectInst* I) { 380 Expression e; 381 382 e.firstVN = lookup_or_add(I->getCondition()); 383 e.secondVN = lookup_or_add(I->getTrueValue()); 384 e.thirdVN = lookup_or_add(I->getFalseValue()); 385 e.function = 0; 386 e.type = I->getType(); 387 e.opcode = Expression::SELECT; 388 389 return e; 390} 391 392Expression ValueTable::create_expression(GetElementPtrInst* G) { 393 Expression e; 394 395 e.firstVN = lookup_or_add(G->getPointerOperand()); 396 e.secondVN = 0; 397 e.thirdVN = 0; 398 e.function = 0; 399 e.type = G->getType(); 400 e.opcode = Expression::GEP; 401 402 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 403 I != E; ++I) 404 e.varargs.push_back(lookup_or_add(*I)); 405 406 return e; 407} 408 409//===----------------------------------------------------------------------===// 410// ValueTable External Functions 411//===----------------------------------------------------------------------===// 412 413/// add - Insert a value into the table with a specified value number. 414void ValueTable::add(Value* V, uint32_t num) { 415 valueNumbering.insert(std::make_pair(V, num)); 416} 417 418/// lookup_or_add - Returns the value number for the specified value, assigning 419/// it a new number if it did not have one before. 420uint32_t ValueTable::lookup_or_add(Value* V) { 421 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 422 if (VI != valueNumbering.end()) 423 return VI->second; 424 425 if (CallInst* C = dyn_cast<CallInst>(V)) { 426 if (AA->doesNotAccessMemory(C)) { 427 Expression e = create_expression(C); 428 429 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 430 if (EI != expressionNumbering.end()) { 431 valueNumbering.insert(std::make_pair(V, EI->second)); 432 return EI->second; 433 } else { 434 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 435 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 436 437 return nextValueNumber++; 438 } 439 } else if (AA->onlyReadsMemory(C)) { 440 Expression e = create_expression(C); 441 442 if (expressionNumbering.find(e) == expressionNumbering.end()) { 443 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 444 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 445 return nextValueNumber++; 446 } 447 448 MemDepResult local_dep = MD->getDependency(C); 449 450 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 451 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 452 return nextValueNumber++; 453 } 454 455 if (local_dep.isDef()) { 456 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 457 458 if (local_cdep->getNumOperands() != C->getNumOperands()) { 459 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 460 return nextValueNumber++; 461 } 462 463 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 464 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 465 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 466 if (c_vn != cd_vn) { 467 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 468 return nextValueNumber++; 469 } 470 } 471 472 uint32_t v = lookup_or_add(local_cdep); 473 valueNumbering.insert(std::make_pair(V, v)); 474 return v; 475 } 476 477 // Non-local case. 478 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 479 MD->getNonLocalCallDependency(CallSite(C)); 480 // FIXME: call/call dependencies for readonly calls should return def, not 481 // clobber! Move the checking logic to MemDep! 482 CallInst* cdep = 0; 483 484 // Check to see if we have a single dominating call instruction that is 485 // identical to C. 486 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 487 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 488 // Ignore non-local dependencies. 489 if (I->second.isNonLocal()) 490 continue; 491 492 // We don't handle non-depedencies. If we already have a call, reject 493 // instruction dependencies. 494 if (I->second.isClobber() || cdep != 0) { 495 cdep = 0; 496 break; 497 } 498 499 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 500 // FIXME: All duplicated with non-local case. 501 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 502 cdep = NonLocalDepCall; 503 continue; 504 } 505 506 cdep = 0; 507 break; 508 } 509 510 if (!cdep) { 511 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 512 return nextValueNumber++; 513 } 514 515 if (cdep->getNumOperands() != C->getNumOperands()) { 516 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 517 return nextValueNumber++; 518 } 519 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 520 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 521 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 522 if (c_vn != cd_vn) { 523 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 524 return nextValueNumber++; 525 } 526 } 527 528 uint32_t v = lookup_or_add(cdep); 529 valueNumbering.insert(std::make_pair(V, v)); 530 return v; 531 532 } else { 533 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 534 return nextValueNumber++; 535 } 536 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 537 Expression e = create_expression(BO); 538 539 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 540 if (EI != expressionNumbering.end()) { 541 valueNumbering.insert(std::make_pair(V, EI->second)); 542 return EI->second; 543 } else { 544 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 545 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 546 547 return nextValueNumber++; 548 } 549 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 550 Expression e = create_expression(C); 551 552 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 553 if (EI != expressionNumbering.end()) { 554 valueNumbering.insert(std::make_pair(V, EI->second)); 555 return EI->second; 556 } else { 557 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 558 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 559 560 return nextValueNumber++; 561 } 562 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 563 Expression e = create_expression(U); 564 565 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 566 if (EI != expressionNumbering.end()) { 567 valueNumbering.insert(std::make_pair(V, EI->second)); 568 return EI->second; 569 } else { 570 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 571 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 572 573 return nextValueNumber++; 574 } 575 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 576 Expression e = create_expression(U); 577 578 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 579 if (EI != expressionNumbering.end()) { 580 valueNumbering.insert(std::make_pair(V, EI->second)); 581 return EI->second; 582 } else { 583 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 584 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 585 586 return nextValueNumber++; 587 } 588 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 589 Expression e = create_expression(U); 590 591 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 592 if (EI != expressionNumbering.end()) { 593 valueNumbering.insert(std::make_pair(V, EI->second)); 594 return EI->second; 595 } else { 596 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 597 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 598 599 return nextValueNumber++; 600 } 601 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 602 Expression e = create_expression(U); 603 604 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 605 if (EI != expressionNumbering.end()) { 606 valueNumbering.insert(std::make_pair(V, EI->second)); 607 return EI->second; 608 } else { 609 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 610 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 611 612 return nextValueNumber++; 613 } 614 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 615 Expression e = create_expression(U); 616 617 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 618 if (EI != expressionNumbering.end()) { 619 valueNumbering.insert(std::make_pair(V, EI->second)); 620 return EI->second; 621 } else { 622 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 623 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 624 625 return nextValueNumber++; 626 } 627 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 628 Expression e = create_expression(U); 629 630 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 631 if (EI != expressionNumbering.end()) { 632 valueNumbering.insert(std::make_pair(V, EI->second)); 633 return EI->second; 634 } else { 635 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 636 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 637 638 return nextValueNumber++; 639 } 640 } else { 641 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 642 return nextValueNumber++; 643 } 644} 645 646/// lookup - Returns the value number of the specified value. Fails if 647/// the value has not yet been numbered. 648uint32_t ValueTable::lookup(Value* V) const { 649 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 650 assert(VI != valueNumbering.end() && "Value not numbered?"); 651 return VI->second; 652} 653 654/// clear - Remove all entries from the ValueTable 655void ValueTable::clear() { 656 valueNumbering.clear(); 657 expressionNumbering.clear(); 658 nextValueNumber = 1; 659} 660 661/// erase - Remove a value from the value numbering 662void ValueTable::erase(Value* V) { 663 valueNumbering.erase(V); 664} 665 666/// verifyRemoved - Verify that the value is removed from all internal data 667/// structures. 668void ValueTable::verifyRemoved(const Value *V) const { 669 for (DenseMap<Value*, uint32_t>::iterator 670 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 671 assert(I->first != V && "Inst still occurs in value numbering map!"); 672 } 673} 674 675//===----------------------------------------------------------------------===// 676// GVN Pass 677//===----------------------------------------------------------------------===// 678 679namespace { 680 struct VISIBILITY_HIDDEN ValueNumberScope { 681 ValueNumberScope* parent; 682 DenseMap<uint32_t, Value*> table; 683 684 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 685 }; 686} 687 688namespace { 689 690 class VISIBILITY_HIDDEN GVN : public FunctionPass { 691 bool runOnFunction(Function &F); 692 public: 693 static char ID; // Pass identification, replacement for typeid 694 GVN() : FunctionPass(&ID) { } 695 696 private: 697 MemoryDependenceAnalysis *MD; 698 DominatorTree *DT; 699 700 ValueTable VN; 701 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 702 703 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 704 PhiMapType phiMap; 705 706 707 // This transformation requires dominator postdominator info 708 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 709 AU.addRequired<DominatorTree>(); 710 AU.addRequired<MemoryDependenceAnalysis>(); 711 AU.addRequired<AliasAnalysis>(); 712 713 AU.addPreserved<DominatorTree>(); 714 AU.addPreserved<AliasAnalysis>(); 715 } 716 717 // Helper fuctions 718 // FIXME: eliminate or document these better 719 bool processLoad(LoadInst* L, 720 SmallVectorImpl<Instruction*> &toErase); 721 bool processInstruction(Instruction* I, 722 SmallVectorImpl<Instruction*> &toErase); 723 bool processNonLocalLoad(LoadInst* L, 724 SmallVectorImpl<Instruction*> &toErase); 725 bool processBlock(BasicBlock* BB); 726 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig, 727 DenseMap<BasicBlock*, Value*> &Phis, 728 bool top_level = false); 729 void dump(DenseMap<uint32_t, Value*>& d); 730 bool iterateOnFunction(Function &F); 731 Value* CollapsePhi(PHINode* p); 732 bool isSafeReplacement(PHINode* p, Instruction* inst); 733 bool performPRE(Function& F); 734 Value* lookupNumber(BasicBlock* BB, uint32_t num); 735 bool mergeBlockIntoPredecessor(BasicBlock* BB); 736 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno); 737 void cleanupGlobalSets(); 738 void verifyRemoved(const Instruction *I) const; 739 }; 740 741 char GVN::ID = 0; 742} 743 744// createGVNPass - The public interface to this file... 745FunctionPass *llvm::createGVNPass() { return new GVN(); } 746 747static RegisterPass<GVN> X("gvn", 748 "Global Value Numbering"); 749 750void GVN::dump(DenseMap<uint32_t, Value*>& d) { 751 printf("{\n"); 752 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 753 E = d.end(); I != E; ++I) { 754 printf("%d\n", I->first); 755 I->second->dump(); 756 } 757 printf("}\n"); 758} 759 760Value* GVN::CollapsePhi(PHINode* p) { 761 Value* constVal = p->hasConstantValue(); 762 if (!constVal) return 0; 763 764 Instruction* inst = dyn_cast<Instruction>(constVal); 765 if (!inst) 766 return constVal; 767 768 if (DT->dominates(inst, p)) 769 if (isSafeReplacement(p, inst)) 770 return inst; 771 return 0; 772} 773 774bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) { 775 if (!isa<PHINode>(inst)) 776 return true; 777 778 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 779 UI != E; ++UI) 780 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 781 if (use_phi->getParent() == inst->getParent()) 782 return false; 783 784 return true; 785} 786 787/// GetValueForBlock - Get the value to use within the specified basic block. 788/// available values are in Phis. 789Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig, 790 DenseMap<BasicBlock*, Value*> &Phis, 791 bool top_level) { 792 793 // If we have already computed this value, return the previously computed val. 794 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 795 if (V != Phis.end() && !top_level) return V->second; 796 797 // If the block is unreachable, just return undef, since this path 798 // can't actually occur at runtime. 799 if (!DT->isReachableFromEntry(BB)) 800 return Phis[BB] = Context->getUndef(orig->getType()); 801 802 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 803 Value *ret = GetValueForBlock(Pred, orig, Phis); 804 Phis[BB] = ret; 805 return ret; 806 } 807 808 // Get the number of predecessors of this block so we can reserve space later. 809 // If there is already a PHI in it, use the #preds from it, otherwise count. 810 // Getting it from the PHI is constant time. 811 unsigned NumPreds; 812 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin())) 813 NumPreds = ExistingPN->getNumIncomingValues(); 814 else 815 NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 816 817 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 818 // now, then get values to fill in the incoming values for the PHI. 819 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle", 820 BB->begin()); 821 PN->reserveOperandSpace(NumPreds); 822 823 Phis.insert(std::make_pair(BB, PN)); 824 825 // Fill in the incoming values for the block. 826 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 827 Value* val = GetValueForBlock(*PI, orig, Phis); 828 PN->addIncoming(val, *PI); 829 } 830 831 VN.getAliasAnalysis()->copyValue(orig, PN); 832 833 // Attempt to collapse PHI nodes that are trivially redundant 834 Value* v = CollapsePhi(PN); 835 if (!v) { 836 // Cache our phi construction results 837 if (LoadInst* L = dyn_cast<LoadInst>(orig)) 838 phiMap[L->getPointerOperand()].insert(PN); 839 else 840 phiMap[orig].insert(PN); 841 842 return PN; 843 } 844 845 PN->replaceAllUsesWith(v); 846 if (isa<PointerType>(v->getType())) 847 MD->invalidateCachedPointerInfo(v); 848 849 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 850 E = Phis.end(); I != E; ++I) 851 if (I->second == PN) 852 I->second = v; 853 854 DEBUG(cerr << "GVN removed: " << *PN); 855 MD->removeInstruction(PN); 856 PN->eraseFromParent(); 857 DEBUG(verifyRemoved(PN)); 858 859 Phis[BB] = v; 860 return v; 861} 862 863/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 864/// we're analyzing is fully available in the specified block. As we go, keep 865/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 866/// map is actually a tri-state map with the following values: 867/// 0) we know the block *is not* fully available. 868/// 1) we know the block *is* fully available. 869/// 2) we do not know whether the block is fully available or not, but we are 870/// currently speculating that it will be. 871/// 3) we are speculating for this block and have used that to speculate for 872/// other blocks. 873static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 874 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 875 // Optimistically assume that the block is fully available and check to see 876 // if we already know about this block in one lookup. 877 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 878 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 879 880 // If the entry already existed for this block, return the precomputed value. 881 if (!IV.second) { 882 // If this is a speculative "available" value, mark it as being used for 883 // speculation of other blocks. 884 if (IV.first->second == 2) 885 IV.first->second = 3; 886 return IV.first->second != 0; 887 } 888 889 // Otherwise, see if it is fully available in all predecessors. 890 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 891 892 // If this block has no predecessors, it isn't live-in here. 893 if (PI == PE) 894 goto SpeculationFailure; 895 896 for (; PI != PE; ++PI) 897 // If the value isn't fully available in one of our predecessors, then it 898 // isn't fully available in this block either. Undo our previous 899 // optimistic assumption and bail out. 900 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 901 goto SpeculationFailure; 902 903 return true; 904 905// SpeculationFailure - If we get here, we found out that this is not, after 906// all, a fully-available block. We have a problem if we speculated on this and 907// used the speculation to mark other blocks as available. 908SpeculationFailure: 909 char &BBVal = FullyAvailableBlocks[BB]; 910 911 // If we didn't speculate on this, just return with it set to false. 912 if (BBVal == 2) { 913 BBVal = 0; 914 return false; 915 } 916 917 // If we did speculate on this value, we could have blocks set to 1 that are 918 // incorrect. Walk the (transitive) successors of this block and mark them as 919 // 0 if set to one. 920 SmallVector<BasicBlock*, 32> BBWorklist; 921 BBWorklist.push_back(BB); 922 923 while (!BBWorklist.empty()) { 924 BasicBlock *Entry = BBWorklist.pop_back_val(); 925 // Note that this sets blocks to 0 (unavailable) if they happen to not 926 // already be in FullyAvailableBlocks. This is safe. 927 char &EntryVal = FullyAvailableBlocks[Entry]; 928 if (EntryVal == 0) continue; // Already unavailable. 929 930 // Mark as unavailable. 931 EntryVal = 0; 932 933 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 934 BBWorklist.push_back(*I); 935 } 936 937 return false; 938} 939 940/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 941/// non-local by performing PHI construction. 942bool GVN::processNonLocalLoad(LoadInst *LI, 943 SmallVectorImpl<Instruction*> &toErase) { 944 // Find the non-local dependencies of the load. 945 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 946 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 947 Deps); 948 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI); 949 950 // If we had to process more than one hundred blocks to find the 951 // dependencies, this load isn't worth worrying about. Optimizing 952 // it will be too expensive. 953 if (Deps.size() > 100) 954 return false; 955 956 // If we had a phi translation failure, we'll have a single entry which is a 957 // clobber in the current block. Reject this early. 958 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 959 DEBUG( 960 DOUT << "GVN: non-local load "; 961 WriteAsOperand(*DOUT.stream(), LI); 962 DOUT << " is clobbered by " << *Deps[0].second.getInst(); 963 ); 964 return false; 965 } 966 967 // Filter out useless results (non-locals, etc). Keep track of the blocks 968 // where we have a value available in repl, also keep track of whether we see 969 // dependencies that produce an unknown value for the load (such as a call 970 // that could potentially clobber the load). 971 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 972 SmallVector<BasicBlock*, 16> UnavailableBlocks; 973 974 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 975 BasicBlock *DepBB = Deps[i].first; 976 MemDepResult DepInfo = Deps[i].second; 977 978 if (DepInfo.isClobber()) { 979 UnavailableBlocks.push_back(DepBB); 980 continue; 981 } 982 983 Instruction *DepInst = DepInfo.getInst(); 984 985 // Loading the allocation -> undef. 986 if (isa<AllocationInst>(DepInst)) { 987 ValuesPerBlock.push_back(std::make_pair(DepBB, 988 Context->getUndef(LI->getType()))); 989 continue; 990 } 991 992 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 993 // Reject loads and stores that are to the same address but are of 994 // different types. 995 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because 996 // of bitfield access, it would be interesting to optimize for it at some 997 // point. 998 if (S->getOperand(0)->getType() != LI->getType()) { 999 UnavailableBlocks.push_back(DepBB); 1000 continue; 1001 } 1002 1003 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1004 1005 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) { 1006 if (LD->getType() != LI->getType()) { 1007 UnavailableBlocks.push_back(DepBB); 1008 continue; 1009 } 1010 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1011 } else { 1012 UnavailableBlocks.push_back(DepBB); 1013 continue; 1014 } 1015 } 1016 1017 // If we have no predecessors that produce a known value for this load, exit 1018 // early. 1019 if (ValuesPerBlock.empty()) return false; 1020 1021 // If all of the instructions we depend on produce a known value for this 1022 // load, then it is fully redundant and we can use PHI insertion to compute 1023 // its value. Insert PHIs and remove the fully redundant value now. 1024 if (UnavailableBlocks.empty()) { 1025 // Use cached PHI construction information from previous runs 1026 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1027 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated? 1028 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1029 I != E; ++I) { 1030 if ((*I)->getParent() == LI->getParent()) { 1031 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI); 1032 LI->replaceAllUsesWith(*I); 1033 if (isa<PointerType>((*I)->getType())) 1034 MD->invalidateCachedPointerInfo(*I); 1035 toErase.push_back(LI); 1036 NumGVNLoad++; 1037 return true; 1038 } 1039 1040 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1041 } 1042 1043 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI); 1044 1045 DenseMap<BasicBlock*, Value*> BlockReplValues; 1046 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1047 // Perform PHI construction. 1048 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1049 LI->replaceAllUsesWith(v); 1050 1051 if (isa<PHINode>(v)) 1052 v->takeName(LI); 1053 if (isa<PointerType>(v->getType())) 1054 MD->invalidateCachedPointerInfo(v); 1055 toErase.push_back(LI); 1056 NumGVNLoad++; 1057 return true; 1058 } 1059 1060 if (!EnablePRE || !EnableLoadPRE) 1061 return false; 1062 1063 // Okay, we have *some* definitions of the value. This means that the value 1064 // is available in some of our (transitive) predecessors. Lets think about 1065 // doing PRE of this load. This will involve inserting a new load into the 1066 // predecessor when it's not available. We could do this in general, but 1067 // prefer to not increase code size. As such, we only do this when we know 1068 // that we only have to insert *one* load (which means we're basically moving 1069 // the load, not inserting a new one). 1070 1071 SmallPtrSet<BasicBlock *, 4> Blockers; 1072 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1073 Blockers.insert(UnavailableBlocks[i]); 1074 1075 // Lets find first basic block with more than one predecessor. Walk backwards 1076 // through predecessors if needed. 1077 BasicBlock *LoadBB = LI->getParent(); 1078 BasicBlock *TmpBB = LoadBB; 1079 1080 bool isSinglePred = false; 1081 bool allSingleSucc = true; 1082 while (TmpBB->getSinglePredecessor()) { 1083 isSinglePred = true; 1084 TmpBB = TmpBB->getSinglePredecessor(); 1085 if (!TmpBB) // If haven't found any, bail now. 1086 return false; 1087 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1088 return false; 1089 if (Blockers.count(TmpBB)) 1090 return false; 1091 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1092 allSingleSucc = false; 1093 } 1094 1095 assert(TmpBB); 1096 LoadBB = TmpBB; 1097 1098 // If we have a repl set with LI itself in it, this means we have a loop where 1099 // at least one of the values is LI. Since this means that we won't be able 1100 // to eliminate LI even if we insert uses in the other predecessors, we will 1101 // end up increasing code size. Reject this by scanning for LI. 1102 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1103 if (ValuesPerBlock[i].second == LI) 1104 return false; 1105 1106 if (isSinglePred) { 1107 bool isHot = false; 1108 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1109 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second)) 1110 // "Hot" Instruction is in some loop (because it dominates its dep. 1111 // instruction). 1112 if (DT->dominates(LI, I)) { 1113 isHot = true; 1114 break; 1115 } 1116 1117 // We are interested only in "hot" instructions. We don't want to do any 1118 // mis-optimizations here. 1119 if (!isHot) 1120 return false; 1121 } 1122 1123 // Okay, we have some hope :). Check to see if the loaded value is fully 1124 // available in all but one predecessor. 1125 // FIXME: If we could restructure the CFG, we could make a common pred with 1126 // all the preds that don't have an available LI and insert a new load into 1127 // that one block. 1128 BasicBlock *UnavailablePred = 0; 1129 1130 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1131 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1132 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1133 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1134 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1135 1136 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1137 PI != E; ++PI) { 1138 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1139 continue; 1140 1141 // If this load is not available in multiple predecessors, reject it. 1142 if (UnavailablePred && UnavailablePred != *PI) 1143 return false; 1144 UnavailablePred = *PI; 1145 } 1146 1147 assert(UnavailablePred != 0 && 1148 "Fully available value should be eliminated above!"); 1149 1150 // If the loaded pointer is PHI node defined in this block, do PHI translation 1151 // to get its value in the predecessor. 1152 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1153 1154 // Make sure the value is live in the predecessor. If it was defined by a 1155 // non-PHI instruction in this block, we don't know how to recompute it above. 1156 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1157 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1158 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1159 << *LPInst << *LI << "\n"); 1160 return false; 1161 } 1162 1163 // We don't currently handle critical edges :( 1164 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1165 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1166 << UnavailablePred->getName() << "': " << *LI); 1167 return false; 1168 } 1169 1170 // Make sure it is valid to move this load here. We have to watch out for: 1171 // @1 = getelementptr (i8* p, ... 1172 // test p and branch if == 0 1173 // load @1 1174 // It is valid to have the getelementptr before the test, even if p can be 0, 1175 // as getelementptr only does address arithmetic. 1176 // If we are not pushing the value through any multiple-successor blocks 1177 // we do not have this case. Otherwise, check that the load is safe to 1178 // put anywhere; this can be improved, but should be conservatively safe. 1179 if (!allSingleSucc && 1180 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1181 return false; 1182 1183 // Okay, we can eliminate this load by inserting a reload in the predecessor 1184 // and using PHI construction to get the value in the other predecessors, do 1185 // it. 1186 DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI); 1187 1188 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1189 LI->getAlignment(), 1190 UnavailablePred->getTerminator()); 1191 1192 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1193 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1194 I != E; ++I) 1195 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1196 1197 DenseMap<BasicBlock*, Value*> BlockReplValues; 1198 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end()); 1199 BlockReplValues[UnavailablePred] = NewLoad; 1200 1201 // Perform PHI construction. 1202 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1203 LI->replaceAllUsesWith(v); 1204 if (isa<PHINode>(v)) 1205 v->takeName(LI); 1206 if (isa<PointerType>(v->getType())) 1207 MD->invalidateCachedPointerInfo(v); 1208 toErase.push_back(LI); 1209 NumPRELoad++; 1210 return true; 1211} 1212 1213/// processLoad - Attempt to eliminate a load, first by eliminating it 1214/// locally, and then attempting non-local elimination if that fails. 1215bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1216 if (L->isVolatile()) 1217 return false; 1218 1219 Value* pointer = L->getPointerOperand(); 1220 1221 // ... to a pointer that has been loaded from before... 1222 MemDepResult dep = MD->getDependency(L); 1223 1224 // If the value isn't available, don't do anything! 1225 if (dep.isClobber()) { 1226 DEBUG( 1227 // fast print dep, using operator<< on instruction would be too slow 1228 DOUT << "GVN: load "; 1229 WriteAsOperand(*DOUT.stream(), L); 1230 Instruction *I = dep.getInst(); 1231 DOUT << " is clobbered by " << *I; 1232 ); 1233 return false; 1234 } 1235 1236 // If it is defined in another block, try harder. 1237 if (dep.isNonLocal()) 1238 return processNonLocalLoad(L, toErase); 1239 1240 Instruction *DepInst = dep.getInst(); 1241 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1242 // Only forward substitute stores to loads of the same type. 1243 // FIXME: Could do better! 1244 if (DepSI->getPointerOperand()->getType() != pointer->getType()) 1245 return false; 1246 1247 // Remove it! 1248 L->replaceAllUsesWith(DepSI->getOperand(0)); 1249 if (isa<PointerType>(DepSI->getOperand(0)->getType())) 1250 MD->invalidateCachedPointerInfo(DepSI->getOperand(0)); 1251 toErase.push_back(L); 1252 NumGVNLoad++; 1253 return true; 1254 } 1255 1256 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1257 // Only forward substitute stores to loads of the same type. 1258 // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian. 1259 if (DepLI->getType() != L->getType()) 1260 return false; 1261 1262 // Remove it! 1263 L->replaceAllUsesWith(DepLI); 1264 if (isa<PointerType>(DepLI->getType())) 1265 MD->invalidateCachedPointerInfo(DepLI); 1266 toErase.push_back(L); 1267 NumGVNLoad++; 1268 return true; 1269 } 1270 1271 // If this load really doesn't depend on anything, then we must be loading an 1272 // undef value. This can happen when loading for a fresh allocation with no 1273 // intervening stores, for example. 1274 if (isa<AllocationInst>(DepInst)) { 1275 L->replaceAllUsesWith(Context->getUndef(L->getType())); 1276 toErase.push_back(L); 1277 NumGVNLoad++; 1278 return true; 1279 } 1280 1281 return false; 1282} 1283 1284Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) { 1285 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1286 if (I == localAvail.end()) 1287 return 0; 1288 1289 ValueNumberScope* locals = I->second; 1290 1291 while (locals) { 1292 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num); 1293 if (I != locals->table.end()) 1294 return I->second; 1295 else 1296 locals = locals->parent; 1297 } 1298 1299 return 0; 1300} 1301 1302/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination 1303/// by inheritance from the dominator fails, see if we can perform phi 1304/// construction to eliminate the redundancy. 1305Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) { 1306 BasicBlock* BaseBlock = orig->getParent(); 1307 1308 SmallPtrSet<BasicBlock*, 4> Visited; 1309 SmallVector<BasicBlock*, 8> Stack; 1310 Stack.push_back(BaseBlock); 1311 1312 DenseMap<BasicBlock*, Value*> Results; 1313 1314 // Walk backwards through our predecessors, looking for instances of the 1315 // value number we're looking for. Instances are recorded in the Results 1316 // map, which is then used to perform phi construction. 1317 while (!Stack.empty()) { 1318 BasicBlock* Current = Stack.back(); 1319 Stack.pop_back(); 1320 1321 // If we've walked all the way to a proper dominator, then give up. Cases 1322 // where the instance is in the dominator will have been caught by the fast 1323 // path, and any cases that require phi construction further than this are 1324 // probably not worth it anyways. Note that this is a SIGNIFICANT compile 1325 // time improvement. 1326 if (DT->properlyDominates(Current, orig->getParent())) return 0; 1327 1328 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA = 1329 localAvail.find(Current); 1330 if (LA == localAvail.end()) return 0; 1331 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno); 1332 1333 if (V != LA->second->table.end()) { 1334 // Found an instance, record it. 1335 Results.insert(std::make_pair(Current, V->second)); 1336 continue; 1337 } 1338 1339 // If we reach the beginning of the function, then give up. 1340 if (pred_begin(Current) == pred_end(Current)) 1341 return 0; 1342 1343 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current); 1344 PI != PE; ++PI) 1345 if (Visited.insert(*PI)) 1346 Stack.push_back(*PI); 1347 } 1348 1349 // If we didn't find instances, give up. Otherwise, perform phi construction. 1350 if (Results.size() == 0) 1351 return 0; 1352 else 1353 return GetValueForBlock(BaseBlock, orig, Results, true); 1354} 1355 1356/// processInstruction - When calculating availability, handle an instruction 1357/// by inserting it into the appropriate sets 1358bool GVN::processInstruction(Instruction *I, 1359 SmallVectorImpl<Instruction*> &toErase) { 1360 if (LoadInst* L = dyn_cast<LoadInst>(I)) { 1361 bool changed = processLoad(L, toErase); 1362 1363 if (!changed) { 1364 unsigned num = VN.lookup_or_add(L); 1365 localAvail[I->getParent()]->table.insert(std::make_pair(num, L)); 1366 } 1367 1368 return changed; 1369 } 1370 1371 uint32_t nextNum = VN.getNextUnusedValueNumber(); 1372 unsigned num = VN.lookup_or_add(I); 1373 1374 if (BranchInst* BI = dyn_cast<BranchInst>(I)) { 1375 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1376 1377 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1378 return false; 1379 1380 Value* branchCond = BI->getCondition(); 1381 uint32_t condVN = VN.lookup_or_add(branchCond); 1382 1383 BasicBlock* trueSucc = BI->getSuccessor(0); 1384 BasicBlock* falseSucc = BI->getSuccessor(1); 1385 1386 if (trueSucc->getSinglePredecessor()) 1387 localAvail[trueSucc]->table[condVN] = Context->getConstantIntTrue(); 1388 if (falseSucc->getSinglePredecessor()) 1389 localAvail[falseSucc]->table[condVN] = Context->getConstantIntFalse(); 1390 1391 return false; 1392 1393 // Allocations are always uniquely numbered, so we can save time and memory 1394 // by fast failing them. 1395 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) { 1396 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1397 return false; 1398 } 1399 1400 // Collapse PHI nodes 1401 if (PHINode* p = dyn_cast<PHINode>(I)) { 1402 Value* constVal = CollapsePhi(p); 1403 1404 if (constVal) { 1405 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1406 PI != PE; ++PI) 1407 PI->second.erase(p); 1408 1409 p->replaceAllUsesWith(constVal); 1410 if (isa<PointerType>(constVal->getType())) 1411 MD->invalidateCachedPointerInfo(constVal); 1412 VN.erase(p); 1413 1414 toErase.push_back(p); 1415 } else { 1416 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1417 } 1418 1419 // If the number we were assigned was a brand new VN, then we don't 1420 // need to do a lookup to see if the number already exists 1421 // somewhere in the domtree: it can't! 1422 } else if (num == nextNum) { 1423 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1424 1425 // Perform fast-path value-number based elimination of values inherited from 1426 // dominators. 1427 } else if (Value* repl = lookupNumber(I->getParent(), num)) { 1428 // Remove it! 1429 VN.erase(I); 1430 I->replaceAllUsesWith(repl); 1431 if (isa<PointerType>(repl->getType())) 1432 MD->invalidateCachedPointerInfo(repl); 1433 toErase.push_back(I); 1434 return true; 1435 1436#if 0 1437 // Perform slow-pathvalue-number based elimination with phi construction. 1438 } else if (Value* repl = AttemptRedundancyElimination(I, num)) { 1439 // Remove it! 1440 VN.erase(I); 1441 I->replaceAllUsesWith(repl); 1442 if (isa<PointerType>(repl->getType())) 1443 MD->invalidateCachedPointerInfo(repl); 1444 toErase.push_back(I); 1445 return true; 1446#endif 1447 } else { 1448 localAvail[I->getParent()]->table.insert(std::make_pair(num, I)); 1449 } 1450 1451 return false; 1452} 1453 1454/// runOnFunction - This is the main transformation entry point for a function. 1455bool GVN::runOnFunction(Function& F) { 1456 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1457 DT = &getAnalysis<DominatorTree>(); 1458 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1459 VN.setMemDep(MD); 1460 VN.setDomTree(DT); 1461 1462 bool changed = false; 1463 bool shouldContinue = true; 1464 1465 // Merge unconditional branches, allowing PRE to catch more 1466 // optimization opportunities. 1467 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1468 BasicBlock* BB = FI; 1469 ++FI; 1470 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1471 if (removedBlock) NumGVNBlocks++; 1472 1473 changed |= removedBlock; 1474 } 1475 1476 unsigned Iteration = 0; 1477 1478 while (shouldContinue) { 1479 DEBUG(cerr << "GVN iteration: " << Iteration << "\n"); 1480 shouldContinue = iterateOnFunction(F); 1481 changed |= shouldContinue; 1482 ++Iteration; 1483 } 1484 1485 if (EnablePRE) { 1486 bool PREChanged = true; 1487 while (PREChanged) { 1488 PREChanged = performPRE(F); 1489 changed |= PREChanged; 1490 } 1491 } 1492 // FIXME: Should perform GVN again after PRE does something. PRE can move 1493 // computations into blocks where they become fully redundant. Note that 1494 // we can't do this until PRE's critical edge splitting updates memdep. 1495 // Actually, when this happens, we should just fully integrate PRE into GVN. 1496 1497 cleanupGlobalSets(); 1498 1499 return changed; 1500} 1501 1502 1503bool GVN::processBlock(BasicBlock* BB) { 1504 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1505 // incrementing BI before processing an instruction). 1506 SmallVector<Instruction*, 8> toErase; 1507 bool changed_function = false; 1508 1509 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1510 BI != BE;) { 1511 changed_function |= processInstruction(BI, toErase); 1512 if (toErase.empty()) { 1513 ++BI; 1514 continue; 1515 } 1516 1517 // If we need some instructions deleted, do it now. 1518 NumGVNInstr += toErase.size(); 1519 1520 // Avoid iterator invalidation. 1521 bool AtStart = BI == BB->begin(); 1522 if (!AtStart) 1523 --BI; 1524 1525 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1526 E = toErase.end(); I != E; ++I) { 1527 DEBUG(cerr << "GVN removed: " << **I); 1528 MD->removeInstruction(*I); 1529 (*I)->eraseFromParent(); 1530 DEBUG(verifyRemoved(*I)); 1531 } 1532 toErase.clear(); 1533 1534 if (AtStart) 1535 BI = BB->begin(); 1536 else 1537 ++BI; 1538 } 1539 1540 return changed_function; 1541} 1542 1543/// performPRE - Perform a purely local form of PRE that looks for diamond 1544/// control flow patterns and attempts to perform simple PRE at the join point. 1545bool GVN::performPRE(Function& F) { 1546 bool Changed = false; 1547 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1548 DenseMap<BasicBlock*, Value*> predMap; 1549 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1550 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1551 BasicBlock* CurrentBlock = *DI; 1552 1553 // Nothing to PRE in the entry block. 1554 if (CurrentBlock == &F.getEntryBlock()) continue; 1555 1556 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1557 BE = CurrentBlock->end(); BI != BE; ) { 1558 Instruction *CurInst = BI++; 1559 1560 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) || 1561 isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) || 1562 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1563 isa<DbgInfoIntrinsic>(CurInst)) 1564 continue; 1565 1566 uint32_t valno = VN.lookup(CurInst); 1567 1568 // Look for the predecessors for PRE opportunities. We're 1569 // only trying to solve the basic diamond case, where 1570 // a value is computed in the successor and one predecessor, 1571 // but not the other. We also explicitly disallow cases 1572 // where the successor is its own predecessor, because they're 1573 // more complicated to get right. 1574 unsigned numWith = 0; 1575 unsigned numWithout = 0; 1576 BasicBlock* PREPred = 0; 1577 predMap.clear(); 1578 1579 for (pred_iterator PI = pred_begin(CurrentBlock), 1580 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1581 // We're not interested in PRE where the block is its 1582 // own predecessor, on in blocks with predecessors 1583 // that are not reachable. 1584 if (*PI == CurrentBlock) { 1585 numWithout = 2; 1586 break; 1587 } else if (!localAvail.count(*PI)) { 1588 numWithout = 2; 1589 break; 1590 } 1591 1592 DenseMap<uint32_t, Value*>::iterator predV = 1593 localAvail[*PI]->table.find(valno); 1594 if (predV == localAvail[*PI]->table.end()) { 1595 PREPred = *PI; 1596 numWithout++; 1597 } else if (predV->second == CurInst) { 1598 numWithout = 2; 1599 } else { 1600 predMap[*PI] = predV->second; 1601 numWith++; 1602 } 1603 } 1604 1605 // Don't do PRE when it might increase code size, i.e. when 1606 // we would need to insert instructions in more than one pred. 1607 if (numWithout != 1 || numWith == 0) 1608 continue; 1609 1610 // We can't do PRE safely on a critical edge, so instead we schedule 1611 // the edge to be split and perform the PRE the next time we iterate 1612 // on the function. 1613 unsigned succNum = 0; 1614 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1615 i != e; ++i) 1616 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1617 succNum = i; 1618 break; 1619 } 1620 1621 if (isCriticalEdge(PREPred->getTerminator(), succNum)) { 1622 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum)); 1623 continue; 1624 } 1625 1626 // Instantiate the expression the in predecessor that lacked it. 1627 // Because we are going top-down through the block, all value numbers 1628 // will be available in the predecessor by the time we need them. Any 1629 // that weren't original present will have been instantiated earlier 1630 // in this loop. 1631 Instruction* PREInstr = CurInst->clone(*Context); 1632 bool success = true; 1633 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1634 Value *Op = PREInstr->getOperand(i); 1635 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1636 continue; 1637 1638 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1639 PREInstr->setOperand(i, V); 1640 } else { 1641 success = false; 1642 break; 1643 } 1644 } 1645 1646 // Fail out if we encounter an operand that is not available in 1647 // the PRE predecessor. This is typically because of loads which 1648 // are not value numbered precisely. 1649 if (!success) { 1650 delete PREInstr; 1651 DEBUG(verifyRemoved(PREInstr)); 1652 continue; 1653 } 1654 1655 PREInstr->insertBefore(PREPred->getTerminator()); 1656 PREInstr->setName(CurInst->getName() + ".pre"); 1657 predMap[PREPred] = PREInstr; 1658 VN.add(PREInstr, valno); 1659 NumGVNPRE++; 1660 1661 // Update the availability map to include the new instruction. 1662 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr)); 1663 1664 // Create a PHI to make the value available in this block. 1665 PHINode* Phi = PHINode::Create(CurInst->getType(), 1666 CurInst->getName() + ".pre-phi", 1667 CurrentBlock->begin()); 1668 for (pred_iterator PI = pred_begin(CurrentBlock), 1669 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1670 Phi->addIncoming(predMap[*PI], *PI); 1671 1672 VN.add(Phi, valno); 1673 localAvail[CurrentBlock]->table[valno] = Phi; 1674 1675 CurInst->replaceAllUsesWith(Phi); 1676 if (isa<PointerType>(Phi->getType())) 1677 MD->invalidateCachedPointerInfo(Phi); 1678 VN.erase(CurInst); 1679 1680 DEBUG(cerr << "GVN PRE removed: " << *CurInst); 1681 MD->removeInstruction(CurInst); 1682 CurInst->eraseFromParent(); 1683 DEBUG(verifyRemoved(CurInst)); 1684 Changed = true; 1685 } 1686 } 1687 1688 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1689 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1690 SplitCriticalEdge(I->first, I->second, this); 1691 1692 return Changed || toSplit.size(); 1693} 1694 1695/// iterateOnFunction - Executes one iteration of GVN 1696bool GVN::iterateOnFunction(Function &F) { 1697 cleanupGlobalSets(); 1698 1699 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1700 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1701 if (DI->getIDom()) 1702 localAvail[DI->getBlock()] = 1703 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1704 else 1705 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1706 } 1707 1708 // Top-down walk of the dominator tree 1709 bool changed = false; 1710#if 0 1711 // Needed for value numbering with phi construction to work. 1712 ReversePostOrderTraversal<Function*> RPOT(&F); 1713 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1714 RE = RPOT.end(); RI != RE; ++RI) 1715 changed |= processBlock(*RI); 1716#else 1717 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1718 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1719 changed |= processBlock(DI->getBlock()); 1720#endif 1721 1722 return changed; 1723} 1724 1725void GVN::cleanupGlobalSets() { 1726 VN.clear(); 1727 phiMap.clear(); 1728 1729 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1730 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1731 delete I->second; 1732 localAvail.clear(); 1733} 1734 1735/// verifyRemoved - Verify that the specified instruction does not occur in our 1736/// internal data structures. 1737void GVN::verifyRemoved(const Instruction *Inst) const { 1738 VN.verifyRemoved(Inst); 1739 1740 // Walk through the PHI map to make sure the instruction isn't hiding in there 1741 // somewhere. 1742 for (PhiMapType::iterator 1743 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) { 1744 assert(I->first != Inst && "Inst is still a key in PHI map!"); 1745 1746 for (SmallPtrSet<Instruction*, 4>::iterator 1747 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) { 1748 assert(*II != Inst && "Inst is still a value in PHI map!"); 1749 } 1750 } 1751 1752 // Walk through the value number scope to make sure the instruction isn't 1753 // ferreted away in it. 1754 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1755 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1756 const ValueNumberScope *VNS = I->second; 1757 1758 while (VNS) { 1759 for (DenseMap<uint32_t, Value*>::iterator 1760 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1761 assert(II->second != Inst && "Inst still in value numbering scope!"); 1762 } 1763 1764 VNS = VNS->parent; 1765 } 1766 } 1767} 1768