GVN.cpp revision 669011f50b8234bb4775d52a2d1e1ba5f6311e62
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/GlobalVariable.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/LLVMContext.h" 23#include "llvm/Analysis/AliasAnalysis.h" 24#include "llvm/Analysis/ConstantFolding.h" 25#include "llvm/Analysis/Dominators.h" 26#include "llvm/Analysis/InstructionSimplify.h" 27#include "llvm/Analysis/Loads.h" 28#include "llvm/Analysis/MemoryBuiltins.h" 29#include "llvm/Analysis/MemoryDependenceAnalysis.h" 30#include "llvm/Analysis/PHITransAddr.h" 31#include "llvm/Analysis/ValueTracking.h" 32#include "llvm/Assembly/Writer.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Target/TargetLibraryInfo.h" 35#include "llvm/Transforms/Utils/BasicBlockUtils.h" 36#include "llvm/Transforms/Utils/SSAUpdater.h" 37#include "llvm/ADT/DenseMap.h" 38#include "llvm/ADT/DepthFirstIterator.h" 39#include "llvm/ADT/SmallPtrSet.h" 40#include "llvm/ADT/Statistic.h" 41#include "llvm/Support/Allocator.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/IRBuilder.h" 45#include "llvm/Support/PatternMatch.h" 46using namespace llvm; 47using namespace PatternMatch; 48 49STATISTIC(NumGVNInstr, "Number of instructions deleted"); 50STATISTIC(NumGVNLoad, "Number of loads deleted"); 51STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 52STATISTIC(NumGVNBlocks, "Number of blocks merged"); 53STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 54STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 55STATISTIC(NumPRELoad, "Number of loads PRE'd"); 56 57static cl::opt<bool> EnablePRE("enable-pre", 58 cl::init(true), cl::Hidden); 59static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 60 61//===----------------------------------------------------------------------===// 62// ValueTable Class 63//===----------------------------------------------------------------------===// 64 65/// This class holds the mapping between values and value numbers. It is used 66/// as an efficient mechanism to determine the expression-wise equivalence of 67/// two values. 68namespace { 69 struct Expression { 70 uint32_t opcode; 71 Type *type; 72 SmallVector<uint32_t, 4> varargs; 73 74 Expression(uint32_t o = ~2U) : opcode(o) { } 75 76 bool operator==(const Expression &other) const { 77 if (opcode != other.opcode) 78 return false; 79 if (opcode == ~0U || opcode == ~1U) 80 return true; 81 if (type != other.type) 82 return false; 83 if (varargs != other.varargs) 84 return false; 85 return true; 86 } 87 }; 88 89 class ValueTable { 90 DenseMap<Value*, uint32_t> valueNumbering; 91 DenseMap<Expression, uint32_t> expressionNumbering; 92 AliasAnalysis *AA; 93 MemoryDependenceAnalysis *MD; 94 DominatorTree *DT; 95 96 uint32_t nextValueNumber; 97 98 Expression create_expression(Instruction* I); 99 Expression create_cmp_expression(unsigned Opcode, 100 CmpInst::Predicate Predicate, 101 Value *LHS, Value *RHS); 102 Expression create_extractvalue_expression(ExtractValueInst* EI); 103 uint32_t lookup_or_add_call(CallInst* C); 104 public: 105 ValueTable() : nextValueNumber(1) { } 106 uint32_t lookup_or_add(Value *V); 107 uint32_t lookup(Value *V) const; 108 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 109 Value *LHS, Value *RHS); 110 void add(Value *V, uint32_t num); 111 void clear(); 112 void erase(Value *v); 113 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 114 AliasAnalysis *getAliasAnalysis() const { return AA; } 115 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 116 void setDomTree(DominatorTree* D) { DT = D; } 117 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 118 void verifyRemoved(const Value *) const; 119 }; 120} 121 122namespace llvm { 123template <> struct DenseMapInfo<Expression> { 124 static inline Expression getEmptyKey() { 125 return ~0U; 126 } 127 128 static inline Expression getTombstoneKey() { 129 return ~1U; 130 } 131 132 static unsigned getHashValue(const Expression e) { 133 unsigned hash = e.opcode; 134 135 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 136 (unsigned)((uintptr_t)e.type >> 9)); 137 138 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 139 E = e.varargs.end(); I != E; ++I) 140 hash = *I + hash * 37; 141 142 return hash; 143 } 144 static bool isEqual(const Expression &LHS, const Expression &RHS) { 145 return LHS == RHS; 146 } 147}; 148 149} 150 151//===----------------------------------------------------------------------===// 152// ValueTable Internal Functions 153//===----------------------------------------------------------------------===// 154 155Expression ValueTable::create_expression(Instruction *I) { 156 Expression e; 157 e.type = I->getType(); 158 e.opcode = I->getOpcode(); 159 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 160 OI != OE; ++OI) 161 e.varargs.push_back(lookup_or_add(*OI)); 162 if (I->isCommutative()) { 163 // Ensure that commutative instructions that only differ by a permutation 164 // of their operands get the same value number by sorting the operand value 165 // numbers. Since all commutative instructions have two operands it is more 166 // efficient to sort by hand rather than using, say, std::sort. 167 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 168 if (e.varargs[0] > e.varargs[1]) 169 std::swap(e.varargs[0], e.varargs[1]); 170 } 171 172 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 173 // Sort the operand value numbers so x<y and y>x get the same value number. 174 CmpInst::Predicate Predicate = C->getPredicate(); 175 if (e.varargs[0] > e.varargs[1]) { 176 std::swap(e.varargs[0], e.varargs[1]); 177 Predicate = CmpInst::getSwappedPredicate(Predicate); 178 } 179 e.opcode = (C->getOpcode() << 8) | Predicate; 180 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 181 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 182 II != IE; ++II) 183 e.varargs.push_back(*II); 184 } 185 186 return e; 187} 188 189Expression ValueTable::create_cmp_expression(unsigned Opcode, 190 CmpInst::Predicate Predicate, 191 Value *LHS, Value *RHS) { 192 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 193 "Not a comparison!"); 194 Expression e; 195 e.type = CmpInst::makeCmpResultType(LHS->getType()); 196 e.varargs.push_back(lookup_or_add(LHS)); 197 e.varargs.push_back(lookup_or_add(RHS)); 198 199 // Sort the operand value numbers so x<y and y>x get the same value number. 200 if (e.varargs[0] > e.varargs[1]) { 201 std::swap(e.varargs[0], e.varargs[1]); 202 Predicate = CmpInst::getSwappedPredicate(Predicate); 203 } 204 e.opcode = (Opcode << 8) | Predicate; 205 return e; 206} 207 208Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 209 assert(EI != 0 && "Not an ExtractValueInst?"); 210 Expression e; 211 e.type = EI->getType(); 212 e.opcode = 0; 213 214 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 215 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 216 // EI might be an extract from one of our recognised intrinsics. If it 217 // is we'll synthesize a semantically equivalent expression instead on 218 // an extract value expression. 219 switch (I->getIntrinsicID()) { 220 case Intrinsic::sadd_with_overflow: 221 case Intrinsic::uadd_with_overflow: 222 e.opcode = Instruction::Add; 223 break; 224 case Intrinsic::ssub_with_overflow: 225 case Intrinsic::usub_with_overflow: 226 e.opcode = Instruction::Sub; 227 break; 228 case Intrinsic::smul_with_overflow: 229 case Intrinsic::umul_with_overflow: 230 e.opcode = Instruction::Mul; 231 break; 232 default: 233 break; 234 } 235 236 if (e.opcode != 0) { 237 // Intrinsic recognized. Grab its args to finish building the expression. 238 assert(I->getNumArgOperands() == 2 && 239 "Expect two args for recognised intrinsics."); 240 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 241 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 242 return e; 243 } 244 } 245 246 // Not a recognised intrinsic. Fall back to producing an extract value 247 // expression. 248 e.opcode = EI->getOpcode(); 249 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 250 OI != OE; ++OI) 251 e.varargs.push_back(lookup_or_add(*OI)); 252 253 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 254 II != IE; ++II) 255 e.varargs.push_back(*II); 256 257 return e; 258} 259 260//===----------------------------------------------------------------------===// 261// ValueTable External Functions 262//===----------------------------------------------------------------------===// 263 264/// add - Insert a value into the table with a specified value number. 265void ValueTable::add(Value *V, uint32_t num) { 266 valueNumbering.insert(std::make_pair(V, num)); 267} 268 269uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 270 if (AA->doesNotAccessMemory(C)) { 271 Expression exp = create_expression(C); 272 uint32_t& e = expressionNumbering[exp]; 273 if (!e) e = nextValueNumber++; 274 valueNumbering[C] = e; 275 return e; 276 } else if (AA->onlyReadsMemory(C)) { 277 Expression exp = create_expression(C); 278 uint32_t& e = expressionNumbering[exp]; 279 if (!e) { 280 e = nextValueNumber++; 281 valueNumbering[C] = e; 282 return e; 283 } 284 if (!MD) { 285 e = nextValueNumber++; 286 valueNumbering[C] = e; 287 return e; 288 } 289 290 MemDepResult local_dep = MD->getDependency(C); 291 292 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 293 valueNumbering[C] = nextValueNumber; 294 return nextValueNumber++; 295 } 296 297 if (local_dep.isDef()) { 298 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 299 300 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 301 valueNumbering[C] = nextValueNumber; 302 return nextValueNumber++; 303 } 304 305 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 306 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 307 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 308 if (c_vn != cd_vn) { 309 valueNumbering[C] = nextValueNumber; 310 return nextValueNumber++; 311 } 312 } 313 314 uint32_t v = lookup_or_add(local_cdep); 315 valueNumbering[C] = v; 316 return v; 317 } 318 319 // Non-local case. 320 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 321 MD->getNonLocalCallDependency(CallSite(C)); 322 // FIXME: Move the checking logic to MemDep! 323 CallInst* cdep = 0; 324 325 // Check to see if we have a single dominating call instruction that is 326 // identical to C. 327 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 328 const NonLocalDepEntry *I = &deps[i]; 329 if (I->getResult().isNonLocal()) 330 continue; 331 332 // We don't handle non-definitions. If we already have a call, reject 333 // instruction dependencies. 334 if (!I->getResult().isDef() || cdep != 0) { 335 cdep = 0; 336 break; 337 } 338 339 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 340 // FIXME: All duplicated with non-local case. 341 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 342 cdep = NonLocalDepCall; 343 continue; 344 } 345 346 cdep = 0; 347 break; 348 } 349 350 if (!cdep) { 351 valueNumbering[C] = nextValueNumber; 352 return nextValueNumber++; 353 } 354 355 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 356 valueNumbering[C] = nextValueNumber; 357 return nextValueNumber++; 358 } 359 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 360 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 361 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 362 if (c_vn != cd_vn) { 363 valueNumbering[C] = nextValueNumber; 364 return nextValueNumber++; 365 } 366 } 367 368 uint32_t v = lookup_or_add(cdep); 369 valueNumbering[C] = v; 370 return v; 371 372 } else { 373 valueNumbering[C] = nextValueNumber; 374 return nextValueNumber++; 375 } 376} 377 378/// lookup_or_add - Returns the value number for the specified value, assigning 379/// it a new number if it did not have one before. 380uint32_t ValueTable::lookup_or_add(Value *V) { 381 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 382 if (VI != valueNumbering.end()) 383 return VI->second; 384 385 if (!isa<Instruction>(V)) { 386 valueNumbering[V] = nextValueNumber; 387 return nextValueNumber++; 388 } 389 390 Instruction* I = cast<Instruction>(V); 391 Expression exp; 392 switch (I->getOpcode()) { 393 case Instruction::Call: 394 return lookup_or_add_call(cast<CallInst>(I)); 395 case Instruction::Add: 396 case Instruction::FAdd: 397 case Instruction::Sub: 398 case Instruction::FSub: 399 case Instruction::Mul: 400 case Instruction::FMul: 401 case Instruction::UDiv: 402 case Instruction::SDiv: 403 case Instruction::FDiv: 404 case Instruction::URem: 405 case Instruction::SRem: 406 case Instruction::FRem: 407 case Instruction::Shl: 408 case Instruction::LShr: 409 case Instruction::AShr: 410 case Instruction::And: 411 case Instruction::Or : 412 case Instruction::Xor: 413 case Instruction::ICmp: 414 case Instruction::FCmp: 415 case Instruction::Trunc: 416 case Instruction::ZExt: 417 case Instruction::SExt: 418 case Instruction::FPToUI: 419 case Instruction::FPToSI: 420 case Instruction::UIToFP: 421 case Instruction::SIToFP: 422 case Instruction::FPTrunc: 423 case Instruction::FPExt: 424 case Instruction::PtrToInt: 425 case Instruction::IntToPtr: 426 case Instruction::BitCast: 427 case Instruction::Select: 428 case Instruction::ExtractElement: 429 case Instruction::InsertElement: 430 case Instruction::ShuffleVector: 431 case Instruction::InsertValue: 432 case Instruction::GetElementPtr: 433 exp = create_expression(I); 434 break; 435 case Instruction::ExtractValue: 436 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 437 break; 438 default: 439 valueNumbering[V] = nextValueNumber; 440 return nextValueNumber++; 441 } 442 443 uint32_t& e = expressionNumbering[exp]; 444 if (!e) e = nextValueNumber++; 445 valueNumbering[V] = e; 446 return e; 447} 448 449/// lookup - Returns the value number of the specified value. Fails if 450/// the value has not yet been numbered. 451uint32_t ValueTable::lookup(Value *V) const { 452 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 453 assert(VI != valueNumbering.end() && "Value not numbered?"); 454 return VI->second; 455} 456 457/// lookup_or_add_cmp - Returns the value number of the given comparison, 458/// assigning it a new number if it did not have one before. Useful when 459/// we deduced the result of a comparison, but don't immediately have an 460/// instruction realizing that comparison to hand. 461uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 462 CmpInst::Predicate Predicate, 463 Value *LHS, Value *RHS) { 464 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 465 uint32_t& e = expressionNumbering[exp]; 466 if (!e) e = nextValueNumber++; 467 return e; 468} 469 470/// clear - Remove all entries from the ValueTable. 471void ValueTable::clear() { 472 valueNumbering.clear(); 473 expressionNumbering.clear(); 474 nextValueNumber = 1; 475} 476 477/// erase - Remove a value from the value numbering. 478void ValueTable::erase(Value *V) { 479 valueNumbering.erase(V); 480} 481 482/// verifyRemoved - Verify that the value is removed from all internal data 483/// structures. 484void ValueTable::verifyRemoved(const Value *V) const { 485 for (DenseMap<Value*, uint32_t>::const_iterator 486 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 487 assert(I->first != V && "Inst still occurs in value numbering map!"); 488 } 489} 490 491//===----------------------------------------------------------------------===// 492// GVN Pass 493//===----------------------------------------------------------------------===// 494 495namespace { 496 497 class GVN : public FunctionPass { 498 bool NoLoads; 499 MemoryDependenceAnalysis *MD; 500 DominatorTree *DT; 501 const TargetData *TD; 502 const TargetLibraryInfo *TLI; 503 504 ValueTable VN; 505 506 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 507 /// have that value number. Use findLeader to query it. 508 struct LeaderTableEntry { 509 Value *Val; 510 BasicBlock *BB; 511 LeaderTableEntry *Next; 512 }; 513 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 514 BumpPtrAllocator TableAllocator; 515 516 SmallVector<Instruction*, 8> InstrsToErase; 517 public: 518 static char ID; // Pass identification, replacement for typeid 519 explicit GVN(bool noloads = false) 520 : FunctionPass(ID), NoLoads(noloads), MD(0) { 521 initializeGVNPass(*PassRegistry::getPassRegistry()); 522 } 523 524 bool runOnFunction(Function &F); 525 526 /// markInstructionForDeletion - This removes the specified instruction from 527 /// our various maps and marks it for deletion. 528 void markInstructionForDeletion(Instruction *I) { 529 VN.erase(I); 530 InstrsToErase.push_back(I); 531 } 532 533 const TargetData *getTargetData() const { return TD; } 534 DominatorTree &getDominatorTree() const { return *DT; } 535 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 536 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 537 private: 538 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 539 /// its value number. 540 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 541 LeaderTableEntry &Curr = LeaderTable[N]; 542 if (!Curr.Val) { 543 Curr.Val = V; 544 Curr.BB = BB; 545 return; 546 } 547 548 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 549 Node->Val = V; 550 Node->BB = BB; 551 Node->Next = Curr.Next; 552 Curr.Next = Node; 553 } 554 555 /// removeFromLeaderTable - Scan the list of values corresponding to a given 556 /// value number, and remove the given value if encountered. 557 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 558 LeaderTableEntry* Prev = 0; 559 LeaderTableEntry* Curr = &LeaderTable[N]; 560 561 while (Curr->Val != V || Curr->BB != BB) { 562 Prev = Curr; 563 Curr = Curr->Next; 564 } 565 566 if (Prev) { 567 Prev->Next = Curr->Next; 568 } else { 569 if (!Curr->Next) { 570 Curr->Val = 0; 571 Curr->BB = 0; 572 } else { 573 LeaderTableEntry* Next = Curr->Next; 574 Curr->Val = Next->Val; 575 Curr->BB = Next->BB; 576 Curr->Next = Next->Next; 577 } 578 } 579 } 580 581 // List of critical edges to be split between iterations. 582 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 583 584 // This transformation requires dominator postdominator info 585 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 586 AU.addRequired<DominatorTree>(); 587 AU.addRequired<TargetLibraryInfo>(); 588 if (!NoLoads) 589 AU.addRequired<MemoryDependenceAnalysis>(); 590 AU.addRequired<AliasAnalysis>(); 591 592 AU.addPreserved<DominatorTree>(); 593 AU.addPreserved<AliasAnalysis>(); 594 } 595 596 597 // Helper fuctions 598 // FIXME: eliminate or document these better 599 bool processLoad(LoadInst *L); 600 bool processInstruction(Instruction *I); 601 bool processNonLocalLoad(LoadInst *L); 602 bool processBlock(BasicBlock *BB); 603 void dump(DenseMap<uint32_t, Value*> &d); 604 bool iterateOnFunction(Function &F); 605 bool performPRE(Function &F); 606 Value *findLeader(BasicBlock *BB, uint32_t num); 607 void cleanupGlobalSets(); 608 void verifyRemoved(const Instruction *I) const; 609 bool splitCriticalEdges(); 610 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 611 BasicBlock *Root); 612 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root); 613 }; 614 615 char GVN::ID = 0; 616} 617 618// createGVNPass - The public interface to this file... 619FunctionPass *llvm::createGVNPass(bool NoLoads) { 620 return new GVN(NoLoads); 621} 622 623INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 624INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 625INITIALIZE_PASS_DEPENDENCY(DominatorTree) 626INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 627INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 628INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 629 630void GVN::dump(DenseMap<uint32_t, Value*>& d) { 631 errs() << "{\n"; 632 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 633 E = d.end(); I != E; ++I) { 634 errs() << I->first << "\n"; 635 I->second->dump(); 636 } 637 errs() << "}\n"; 638} 639 640/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 641/// we're analyzing is fully available in the specified block. As we go, keep 642/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 643/// map is actually a tri-state map with the following values: 644/// 0) we know the block *is not* fully available. 645/// 1) we know the block *is* fully available. 646/// 2) we do not know whether the block is fully available or not, but we are 647/// currently speculating that it will be. 648/// 3) we are speculating for this block and have used that to speculate for 649/// other blocks. 650static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 651 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 652 // Optimistically assume that the block is fully available and check to see 653 // if we already know about this block in one lookup. 654 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 655 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 656 657 // If the entry already existed for this block, return the precomputed value. 658 if (!IV.second) { 659 // If this is a speculative "available" value, mark it as being used for 660 // speculation of other blocks. 661 if (IV.first->second == 2) 662 IV.first->second = 3; 663 return IV.first->second != 0; 664 } 665 666 // Otherwise, see if it is fully available in all predecessors. 667 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 668 669 // If this block has no predecessors, it isn't live-in here. 670 if (PI == PE) 671 goto SpeculationFailure; 672 673 for (; PI != PE; ++PI) 674 // If the value isn't fully available in one of our predecessors, then it 675 // isn't fully available in this block either. Undo our previous 676 // optimistic assumption and bail out. 677 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 678 goto SpeculationFailure; 679 680 return true; 681 682// SpeculationFailure - If we get here, we found out that this is not, after 683// all, a fully-available block. We have a problem if we speculated on this and 684// used the speculation to mark other blocks as available. 685SpeculationFailure: 686 char &BBVal = FullyAvailableBlocks[BB]; 687 688 // If we didn't speculate on this, just return with it set to false. 689 if (BBVal == 2) { 690 BBVal = 0; 691 return false; 692 } 693 694 // If we did speculate on this value, we could have blocks set to 1 that are 695 // incorrect. Walk the (transitive) successors of this block and mark them as 696 // 0 if set to one. 697 SmallVector<BasicBlock*, 32> BBWorklist; 698 BBWorklist.push_back(BB); 699 700 do { 701 BasicBlock *Entry = BBWorklist.pop_back_val(); 702 // Note that this sets blocks to 0 (unavailable) if they happen to not 703 // already be in FullyAvailableBlocks. This is safe. 704 char &EntryVal = FullyAvailableBlocks[Entry]; 705 if (EntryVal == 0) continue; // Already unavailable. 706 707 // Mark as unavailable. 708 EntryVal = 0; 709 710 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 711 BBWorklist.push_back(*I); 712 } while (!BBWorklist.empty()); 713 714 return false; 715} 716 717 718/// CanCoerceMustAliasedValueToLoad - Return true if 719/// CoerceAvailableValueToLoadType will succeed. 720static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 721 Type *LoadTy, 722 const TargetData &TD) { 723 // If the loaded or stored value is an first class array or struct, don't try 724 // to transform them. We need to be able to bitcast to integer. 725 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 726 StoredVal->getType()->isStructTy() || 727 StoredVal->getType()->isArrayTy()) 728 return false; 729 730 // The store has to be at least as big as the load. 731 if (TD.getTypeSizeInBits(StoredVal->getType()) < 732 TD.getTypeSizeInBits(LoadTy)) 733 return false; 734 735 return true; 736} 737 738 739/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 740/// then a load from a must-aliased pointer of a different type, try to coerce 741/// the stored value. LoadedTy is the type of the load we want to replace and 742/// InsertPt is the place to insert new instructions. 743/// 744/// If we can't do it, return null. 745static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 746 Type *LoadedTy, 747 Instruction *InsertPt, 748 const TargetData &TD) { 749 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 750 return 0; 751 752 // If this is already the right type, just return it. 753 Type *StoredValTy = StoredVal->getType(); 754 755 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 756 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 757 758 // If the store and reload are the same size, we can always reuse it. 759 if (StoreSize == LoadSize) { 760 // Pointer to Pointer -> use bitcast. 761 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 762 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 763 764 // Convert source pointers to integers, which can be bitcast. 765 if (StoredValTy->isPointerTy()) { 766 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 767 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 768 } 769 770 Type *TypeToCastTo = LoadedTy; 771 if (TypeToCastTo->isPointerTy()) 772 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 773 774 if (StoredValTy != TypeToCastTo) 775 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 776 777 // Cast to pointer if the load needs a pointer type. 778 if (LoadedTy->isPointerTy()) 779 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 780 781 return StoredVal; 782 } 783 784 // If the loaded value is smaller than the available value, then we can 785 // extract out a piece from it. If the available value is too small, then we 786 // can't do anything. 787 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 788 789 // Convert source pointers to integers, which can be manipulated. 790 if (StoredValTy->isPointerTy()) { 791 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 792 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 793 } 794 795 // Convert vectors and fp to integer, which can be manipulated. 796 if (!StoredValTy->isIntegerTy()) { 797 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 798 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 799 } 800 801 // If this is a big-endian system, we need to shift the value down to the low 802 // bits so that a truncate will work. 803 if (TD.isBigEndian()) { 804 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 805 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 806 } 807 808 // Truncate the integer to the right size now. 809 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 810 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 811 812 if (LoadedTy == NewIntTy) 813 return StoredVal; 814 815 // If the result is a pointer, inttoptr. 816 if (LoadedTy->isPointerTy()) 817 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 818 819 // Otherwise, bitcast. 820 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 821} 822 823/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 824/// memdep query of a load that ends up being a clobbering memory write (store, 825/// memset, memcpy, memmove). This means that the write *may* provide bits used 826/// by the load but we can't be sure because the pointers don't mustalias. 827/// 828/// Check this case to see if there is anything more we can do before we give 829/// up. This returns -1 if we have to give up, or a byte number in the stored 830/// value of the piece that feeds the load. 831static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 832 Value *WritePtr, 833 uint64_t WriteSizeInBits, 834 const TargetData &TD) { 835 // If the loaded or stored value is a first class array or struct, don't try 836 // to transform them. We need to be able to bitcast to integer. 837 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 838 return -1; 839 840 int64_t StoreOffset = 0, LoadOffset = 0; 841 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 842 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 843 if (StoreBase != LoadBase) 844 return -1; 845 846 // If the load and store are to the exact same address, they should have been 847 // a must alias. AA must have gotten confused. 848 // FIXME: Study to see if/when this happens. One case is forwarding a memset 849 // to a load from the base of the memset. 850#if 0 851 if (LoadOffset == StoreOffset) { 852 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 853 << "Base = " << *StoreBase << "\n" 854 << "Store Ptr = " << *WritePtr << "\n" 855 << "Store Offs = " << StoreOffset << "\n" 856 << "Load Ptr = " << *LoadPtr << "\n"; 857 abort(); 858 } 859#endif 860 861 // If the load and store don't overlap at all, the store doesn't provide 862 // anything to the load. In this case, they really don't alias at all, AA 863 // must have gotten confused. 864 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 865 866 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 867 return -1; 868 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 869 LoadSize >>= 3; 870 871 872 bool isAAFailure = false; 873 if (StoreOffset < LoadOffset) 874 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 875 else 876 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 877 878 if (isAAFailure) { 879#if 0 880 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 881 << "Base = " << *StoreBase << "\n" 882 << "Store Ptr = " << *WritePtr << "\n" 883 << "Store Offs = " << StoreOffset << "\n" 884 << "Load Ptr = " << *LoadPtr << "\n"; 885 abort(); 886#endif 887 return -1; 888 } 889 890 // If the Load isn't completely contained within the stored bits, we don't 891 // have all the bits to feed it. We could do something crazy in the future 892 // (issue a smaller load then merge the bits in) but this seems unlikely to be 893 // valuable. 894 if (StoreOffset > LoadOffset || 895 StoreOffset+StoreSize < LoadOffset+LoadSize) 896 return -1; 897 898 // Okay, we can do this transformation. Return the number of bytes into the 899 // store that the load is. 900 return LoadOffset-StoreOffset; 901} 902 903/// AnalyzeLoadFromClobberingStore - This function is called when we have a 904/// memdep query of a load that ends up being a clobbering store. 905static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 906 StoreInst *DepSI, 907 const TargetData &TD) { 908 // Cannot handle reading from store of first-class aggregate yet. 909 if (DepSI->getValueOperand()->getType()->isStructTy() || 910 DepSI->getValueOperand()->getType()->isArrayTy()) 911 return -1; 912 913 Value *StorePtr = DepSI->getPointerOperand(); 914 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 915 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 916 StorePtr, StoreSize, TD); 917} 918 919/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 920/// memdep query of a load that ends up being clobbered by another load. See if 921/// the other load can feed into the second load. 922static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 923 LoadInst *DepLI, const TargetData &TD){ 924 // Cannot handle reading from store of first-class aggregate yet. 925 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 926 return -1; 927 928 Value *DepPtr = DepLI->getPointerOperand(); 929 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 930 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 931 if (R != -1) return R; 932 933 // If we have a load/load clobber an DepLI can be widened to cover this load, 934 // then we should widen it! 935 int64_t LoadOffs = 0; 936 const Value *LoadBase = 937 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 938 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 939 940 unsigned Size = MemoryDependenceAnalysis:: 941 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 942 if (Size == 0) return -1; 943 944 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 945} 946 947 948 949static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 950 MemIntrinsic *MI, 951 const TargetData &TD) { 952 // If the mem operation is a non-constant size, we can't handle it. 953 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 954 if (SizeCst == 0) return -1; 955 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 956 957 // If this is memset, we just need to see if the offset is valid in the size 958 // of the memset.. 959 if (MI->getIntrinsicID() == Intrinsic::memset) 960 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 961 MemSizeInBits, TD); 962 963 // If we have a memcpy/memmove, the only case we can handle is if this is a 964 // copy from constant memory. In that case, we can read directly from the 965 // constant memory. 966 MemTransferInst *MTI = cast<MemTransferInst>(MI); 967 968 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 969 if (Src == 0) return -1; 970 971 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 972 if (GV == 0 || !GV->isConstant()) return -1; 973 974 // See if the access is within the bounds of the transfer. 975 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 976 MI->getDest(), MemSizeInBits, TD); 977 if (Offset == -1) 978 return Offset; 979 980 // Otherwise, see if we can constant fold a load from the constant with the 981 // offset applied as appropriate. 982 Src = ConstantExpr::getBitCast(Src, 983 llvm::Type::getInt8PtrTy(Src->getContext())); 984 Constant *OffsetCst = 985 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 986 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 987 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 988 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 989 return Offset; 990 return -1; 991} 992 993 994/// GetStoreValueForLoad - This function is called when we have a 995/// memdep query of a load that ends up being a clobbering store. This means 996/// that the store provides bits used by the load but we the pointers don't 997/// mustalias. Check this case to see if there is anything more we can do 998/// before we give up. 999static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1000 Type *LoadTy, 1001 Instruction *InsertPt, const TargetData &TD){ 1002 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1003 1004 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1005 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1006 1007 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1008 1009 // Compute which bits of the stored value are being used by the load. Convert 1010 // to an integer type to start with. 1011 if (SrcVal->getType()->isPointerTy()) 1012 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx)); 1013 if (!SrcVal->getType()->isIntegerTy()) 1014 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1015 1016 // Shift the bits to the least significant depending on endianness. 1017 unsigned ShiftAmt; 1018 if (TD.isLittleEndian()) 1019 ShiftAmt = Offset*8; 1020 else 1021 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1022 1023 if (ShiftAmt) 1024 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1025 1026 if (LoadSize != StoreSize) 1027 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1028 1029 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1030} 1031 1032/// GetLoadValueForLoad - This function is called when we have a 1033/// memdep query of a load that ends up being a clobbering load. This means 1034/// that the load *may* provide bits used by the load but we can't be sure 1035/// because the pointers don't mustalias. Check this case to see if there is 1036/// anything more we can do before we give up. 1037static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1038 Type *LoadTy, Instruction *InsertPt, 1039 GVN &gvn) { 1040 const TargetData &TD = *gvn.getTargetData(); 1041 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1042 // widen SrcVal out to a larger load. 1043 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1044 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1045 if (Offset+LoadSize > SrcValSize) { 1046 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1047 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1048 // If we have a load/load clobber an DepLI can be widened to cover this 1049 // load, then we should widen it to the next power of 2 size big enough! 1050 unsigned NewLoadSize = Offset+LoadSize; 1051 if (!isPowerOf2_32(NewLoadSize)) 1052 NewLoadSize = NextPowerOf2(NewLoadSize); 1053 1054 Value *PtrVal = SrcVal->getPointerOperand(); 1055 1056 // Insert the new load after the old load. This ensures that subsequent 1057 // memdep queries will find the new load. We can't easily remove the old 1058 // load completely because it is already in the value numbering table. 1059 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1060 Type *DestPTy = 1061 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1062 DestPTy = PointerType::get(DestPTy, 1063 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1064 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1065 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1066 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1067 NewLoad->takeName(SrcVal); 1068 NewLoad->setAlignment(SrcVal->getAlignment()); 1069 1070 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1071 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1072 1073 // Replace uses of the original load with the wider load. On a big endian 1074 // system, we need to shift down to get the relevant bits. 1075 Value *RV = NewLoad; 1076 if (TD.isBigEndian()) 1077 RV = Builder.CreateLShr(RV, 1078 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1079 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1080 SrcVal->replaceAllUsesWith(RV); 1081 1082 // We would like to use gvn.markInstructionForDeletion here, but we can't 1083 // because the load is already memoized into the leader map table that GVN 1084 // tracks. It is potentially possible to remove the load from the table, 1085 // but then there all of the operations based on it would need to be 1086 // rehashed. Just leave the dead load around. 1087 gvn.getMemDep().removeInstruction(SrcVal); 1088 SrcVal = NewLoad; 1089 } 1090 1091 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1092} 1093 1094 1095/// GetMemInstValueForLoad - This function is called when we have a 1096/// memdep query of a load that ends up being a clobbering mem intrinsic. 1097static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1098 Type *LoadTy, Instruction *InsertPt, 1099 const TargetData &TD){ 1100 LLVMContext &Ctx = LoadTy->getContext(); 1101 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1102 1103 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1104 1105 // We know that this method is only called when the mem transfer fully 1106 // provides the bits for the load. 1107 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1108 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1109 // independently of what the offset is. 1110 Value *Val = MSI->getValue(); 1111 if (LoadSize != 1) 1112 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1113 1114 Value *OneElt = Val; 1115 1116 // Splat the value out to the right number of bits. 1117 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1118 // If we can double the number of bytes set, do it. 1119 if (NumBytesSet*2 <= LoadSize) { 1120 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1121 Val = Builder.CreateOr(Val, ShVal); 1122 NumBytesSet <<= 1; 1123 continue; 1124 } 1125 1126 // Otherwise insert one byte at a time. 1127 Value *ShVal = Builder.CreateShl(Val, 1*8); 1128 Val = Builder.CreateOr(OneElt, ShVal); 1129 ++NumBytesSet; 1130 } 1131 1132 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1133 } 1134 1135 // Otherwise, this is a memcpy/memmove from a constant global. 1136 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1137 Constant *Src = cast<Constant>(MTI->getSource()); 1138 1139 // Otherwise, see if we can constant fold a load from the constant with the 1140 // offset applied as appropriate. 1141 Src = ConstantExpr::getBitCast(Src, 1142 llvm::Type::getInt8PtrTy(Src->getContext())); 1143 Constant *OffsetCst = 1144 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1145 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1146 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1147 return ConstantFoldLoadFromConstPtr(Src, &TD); 1148} 1149 1150namespace { 1151 1152struct AvailableValueInBlock { 1153 /// BB - The basic block in question. 1154 BasicBlock *BB; 1155 enum ValType { 1156 SimpleVal, // A simple offsetted value that is accessed. 1157 LoadVal, // A value produced by a load. 1158 MemIntrin // A memory intrinsic which is loaded from. 1159 }; 1160 1161 /// V - The value that is live out of the block. 1162 PointerIntPair<Value *, 2, ValType> Val; 1163 1164 /// Offset - The byte offset in Val that is interesting for the load query. 1165 unsigned Offset; 1166 1167 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1168 unsigned Offset = 0) { 1169 AvailableValueInBlock Res; 1170 Res.BB = BB; 1171 Res.Val.setPointer(V); 1172 Res.Val.setInt(SimpleVal); 1173 Res.Offset = Offset; 1174 return Res; 1175 } 1176 1177 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1178 unsigned Offset = 0) { 1179 AvailableValueInBlock Res; 1180 Res.BB = BB; 1181 Res.Val.setPointer(MI); 1182 Res.Val.setInt(MemIntrin); 1183 Res.Offset = Offset; 1184 return Res; 1185 } 1186 1187 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1188 unsigned Offset = 0) { 1189 AvailableValueInBlock Res; 1190 Res.BB = BB; 1191 Res.Val.setPointer(LI); 1192 Res.Val.setInt(LoadVal); 1193 Res.Offset = Offset; 1194 return Res; 1195 } 1196 1197 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1198 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1199 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1200 1201 Value *getSimpleValue() const { 1202 assert(isSimpleValue() && "Wrong accessor"); 1203 return Val.getPointer(); 1204 } 1205 1206 LoadInst *getCoercedLoadValue() const { 1207 assert(isCoercedLoadValue() && "Wrong accessor"); 1208 return cast<LoadInst>(Val.getPointer()); 1209 } 1210 1211 MemIntrinsic *getMemIntrinValue() const { 1212 assert(isMemIntrinValue() && "Wrong accessor"); 1213 return cast<MemIntrinsic>(Val.getPointer()); 1214 } 1215 1216 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1217 /// defined here to the specified type. This handles various coercion cases. 1218 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1219 Value *Res; 1220 if (isSimpleValue()) { 1221 Res = getSimpleValue(); 1222 if (Res->getType() != LoadTy) { 1223 const TargetData *TD = gvn.getTargetData(); 1224 assert(TD && "Need target data to handle type mismatch case"); 1225 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1226 *TD); 1227 1228 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1229 << *getSimpleValue() << '\n' 1230 << *Res << '\n' << "\n\n\n"); 1231 } 1232 } else if (isCoercedLoadValue()) { 1233 LoadInst *Load = getCoercedLoadValue(); 1234 if (Load->getType() == LoadTy && Offset == 0) { 1235 Res = Load; 1236 } else { 1237 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1238 gvn); 1239 1240 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1241 << *getCoercedLoadValue() << '\n' 1242 << *Res << '\n' << "\n\n\n"); 1243 } 1244 } else { 1245 const TargetData *TD = gvn.getTargetData(); 1246 assert(TD && "Need target data to handle type mismatch case"); 1247 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1248 LoadTy, BB->getTerminator(), *TD); 1249 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1250 << " " << *getMemIntrinValue() << '\n' 1251 << *Res << '\n' << "\n\n\n"); 1252 } 1253 return Res; 1254 } 1255}; 1256 1257} // end anonymous namespace 1258 1259/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1260/// construct SSA form, allowing us to eliminate LI. This returns the value 1261/// that should be used at LI's definition site. 1262static Value *ConstructSSAForLoadSet(LoadInst *LI, 1263 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1264 GVN &gvn) { 1265 // Check for the fully redundant, dominating load case. In this case, we can 1266 // just use the dominating value directly. 1267 if (ValuesPerBlock.size() == 1 && 1268 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1269 LI->getParent())) 1270 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1271 1272 // Otherwise, we have to construct SSA form. 1273 SmallVector<PHINode*, 8> NewPHIs; 1274 SSAUpdater SSAUpdate(&NewPHIs); 1275 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1276 1277 Type *LoadTy = LI->getType(); 1278 1279 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1280 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1281 BasicBlock *BB = AV.BB; 1282 1283 if (SSAUpdate.HasValueForBlock(BB)) 1284 continue; 1285 1286 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1287 } 1288 1289 // Perform PHI construction. 1290 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1291 1292 // If new PHI nodes were created, notify alias analysis. 1293 if (V->getType()->isPointerTy()) { 1294 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1295 1296 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1297 AA->copyValue(LI, NewPHIs[i]); 1298 1299 // Now that we've copied information to the new PHIs, scan through 1300 // them again and inform alias analysis that we've added potentially 1301 // escaping uses to any values that are operands to these PHIs. 1302 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1303 PHINode *P = NewPHIs[i]; 1304 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1305 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1306 AA->addEscapingUse(P->getOperandUse(jj)); 1307 } 1308 } 1309 } 1310 1311 return V; 1312} 1313 1314static bool isLifetimeStart(const Instruction *Inst) { 1315 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1316 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1317 return false; 1318} 1319 1320/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1321/// non-local by performing PHI construction. 1322bool GVN::processNonLocalLoad(LoadInst *LI) { 1323 // Find the non-local dependencies of the load. 1324 SmallVector<NonLocalDepResult, 64> Deps; 1325 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1326 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1327 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1328 // << Deps.size() << *LI << '\n'); 1329 1330 // If we had to process more than one hundred blocks to find the 1331 // dependencies, this load isn't worth worrying about. Optimizing 1332 // it will be too expensive. 1333 unsigned NumDeps = Deps.size(); 1334 if (NumDeps > 100) 1335 return false; 1336 1337 // If we had a phi translation failure, we'll have a single entry which is a 1338 // clobber in the current block. Reject this early. 1339 if (NumDeps == 1 && 1340 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1341 DEBUG( 1342 dbgs() << "GVN: non-local load "; 1343 WriteAsOperand(dbgs(), LI); 1344 dbgs() << " has unknown dependencies\n"; 1345 ); 1346 return false; 1347 } 1348 1349 // Filter out useless results (non-locals, etc). Keep track of the blocks 1350 // where we have a value available in repl, also keep track of whether we see 1351 // dependencies that produce an unknown value for the load (such as a call 1352 // that could potentially clobber the load). 1353 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock; 1354 SmallVector<BasicBlock*, 64> UnavailableBlocks; 1355 1356 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1357 BasicBlock *DepBB = Deps[i].getBB(); 1358 MemDepResult DepInfo = Deps[i].getResult(); 1359 1360 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1361 UnavailableBlocks.push_back(DepBB); 1362 continue; 1363 } 1364 1365 if (DepInfo.isClobber()) { 1366 // The address being loaded in this non-local block may not be the same as 1367 // the pointer operand of the load if PHI translation occurs. Make sure 1368 // to consider the right address. 1369 Value *Address = Deps[i].getAddress(); 1370 1371 // If the dependence is to a store that writes to a superset of the bits 1372 // read by the load, we can extract the bits we need for the load from the 1373 // stored value. 1374 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1375 if (TD && Address) { 1376 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1377 DepSI, *TD); 1378 if (Offset != -1) { 1379 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1380 DepSI->getValueOperand(), 1381 Offset)); 1382 continue; 1383 } 1384 } 1385 } 1386 1387 // Check to see if we have something like this: 1388 // load i32* P 1389 // load i8* (P+1) 1390 // if we have this, replace the later with an extraction from the former. 1391 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1392 // If this is a clobber and L is the first instruction in its block, then 1393 // we have the first instruction in the entry block. 1394 if (DepLI != LI && Address && TD) { 1395 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1396 LI->getPointerOperand(), 1397 DepLI, *TD); 1398 1399 if (Offset != -1) { 1400 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1401 Offset)); 1402 continue; 1403 } 1404 } 1405 } 1406 1407 // If the clobbering value is a memset/memcpy/memmove, see if we can 1408 // forward a value on from it. 1409 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1410 if (TD && Address) { 1411 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1412 DepMI, *TD); 1413 if (Offset != -1) { 1414 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1415 Offset)); 1416 continue; 1417 } 1418 } 1419 } 1420 1421 UnavailableBlocks.push_back(DepBB); 1422 continue; 1423 } 1424 1425 // DepInfo.isDef() here 1426 1427 Instruction *DepInst = DepInfo.getInst(); 1428 1429 // Loading the allocation -> undef. 1430 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1431 // Loading immediately after lifetime begin -> undef. 1432 isLifetimeStart(DepInst)) { 1433 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1434 UndefValue::get(LI->getType()))); 1435 continue; 1436 } 1437 1438 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1439 // Reject loads and stores that are to the same address but are of 1440 // different types if we have to. 1441 if (S->getValueOperand()->getType() != LI->getType()) { 1442 // If the stored value is larger or equal to the loaded value, we can 1443 // reuse it. 1444 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1445 LI->getType(), *TD)) { 1446 UnavailableBlocks.push_back(DepBB); 1447 continue; 1448 } 1449 } 1450 1451 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1452 S->getValueOperand())); 1453 continue; 1454 } 1455 1456 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1457 // If the types mismatch and we can't handle it, reject reuse of the load. 1458 if (LD->getType() != LI->getType()) { 1459 // If the stored value is larger or equal to the loaded value, we can 1460 // reuse it. 1461 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1462 UnavailableBlocks.push_back(DepBB); 1463 continue; 1464 } 1465 } 1466 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1467 continue; 1468 } 1469 1470 UnavailableBlocks.push_back(DepBB); 1471 continue; 1472 } 1473 1474 // If we have no predecessors that produce a known value for this load, exit 1475 // early. 1476 if (ValuesPerBlock.empty()) return false; 1477 1478 // If all of the instructions we depend on produce a known value for this 1479 // load, then it is fully redundant and we can use PHI insertion to compute 1480 // its value. Insert PHIs and remove the fully redundant value now. 1481 if (UnavailableBlocks.empty()) { 1482 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1483 1484 // Perform PHI construction. 1485 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1486 LI->replaceAllUsesWith(V); 1487 1488 if (isa<PHINode>(V)) 1489 V->takeName(LI); 1490 if (V->getType()->isPointerTy()) 1491 MD->invalidateCachedPointerInfo(V); 1492 markInstructionForDeletion(LI); 1493 ++NumGVNLoad; 1494 return true; 1495 } 1496 1497 if (!EnablePRE || !EnableLoadPRE) 1498 return false; 1499 1500 // Okay, we have *some* definitions of the value. This means that the value 1501 // is available in some of our (transitive) predecessors. Lets think about 1502 // doing PRE of this load. This will involve inserting a new load into the 1503 // predecessor when it's not available. We could do this in general, but 1504 // prefer to not increase code size. As such, we only do this when we know 1505 // that we only have to insert *one* load (which means we're basically moving 1506 // the load, not inserting a new one). 1507 1508 SmallPtrSet<BasicBlock *, 4> Blockers; 1509 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1510 Blockers.insert(UnavailableBlocks[i]); 1511 1512 // Let's find the first basic block with more than one predecessor. Walk 1513 // backwards through predecessors if needed. 1514 BasicBlock *LoadBB = LI->getParent(); 1515 BasicBlock *TmpBB = LoadBB; 1516 1517 bool isSinglePred = false; 1518 bool allSingleSucc = true; 1519 while (TmpBB->getSinglePredecessor()) { 1520 isSinglePred = true; 1521 TmpBB = TmpBB->getSinglePredecessor(); 1522 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1523 return false; 1524 if (Blockers.count(TmpBB)) 1525 return false; 1526 1527 // If any of these blocks has more than one successor (i.e. if the edge we 1528 // just traversed was critical), then there are other paths through this 1529 // block along which the load may not be anticipated. Hoisting the load 1530 // above this block would be adding the load to execution paths along 1531 // which it was not previously executed. 1532 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1533 return false; 1534 } 1535 1536 assert(TmpBB); 1537 LoadBB = TmpBB; 1538 1539 // FIXME: It is extremely unclear what this loop is doing, other than 1540 // artificially restricting loadpre. 1541 if (isSinglePred) { 1542 bool isHot = false; 1543 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1544 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1545 if (AV.isSimpleValue()) 1546 // "Hot" Instruction is in some loop (because it dominates its dep. 1547 // instruction). 1548 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1549 if (DT->dominates(LI, I)) { 1550 isHot = true; 1551 break; 1552 } 1553 } 1554 1555 // We are interested only in "hot" instructions. We don't want to do any 1556 // mis-optimizations here. 1557 if (!isHot) 1558 return false; 1559 } 1560 1561 // Check to see how many predecessors have the loaded value fully 1562 // available. 1563 DenseMap<BasicBlock*, Value*> PredLoads; 1564 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1565 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1566 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1567 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1568 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1569 1570 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1571 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1572 PI != E; ++PI) { 1573 BasicBlock *Pred = *PI; 1574 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1575 continue; 1576 } 1577 PredLoads[Pred] = 0; 1578 1579 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1580 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1581 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1582 << Pred->getName() << "': " << *LI << '\n'); 1583 return false; 1584 } 1585 1586 if (LoadBB->isLandingPad()) { 1587 DEBUG(dbgs() 1588 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1589 << Pred->getName() << "': " << *LI << '\n'); 1590 return false; 1591 } 1592 1593 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1594 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1595 } 1596 } 1597 1598 if (!NeedToSplit.empty()) { 1599 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1600 return false; 1601 } 1602 1603 // Decide whether PRE is profitable for this load. 1604 unsigned NumUnavailablePreds = PredLoads.size(); 1605 assert(NumUnavailablePreds != 0 && 1606 "Fully available value should be eliminated above!"); 1607 1608 // If this load is unavailable in multiple predecessors, reject it. 1609 // FIXME: If we could restructure the CFG, we could make a common pred with 1610 // all the preds that don't have an available LI and insert a new load into 1611 // that one block. 1612 if (NumUnavailablePreds != 1) 1613 return false; 1614 1615 // Check if the load can safely be moved to all the unavailable predecessors. 1616 bool CanDoPRE = true; 1617 SmallVector<Instruction*, 8> NewInsts; 1618 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1619 E = PredLoads.end(); I != E; ++I) { 1620 BasicBlock *UnavailablePred = I->first; 1621 1622 // Do PHI translation to get its value in the predecessor if necessary. The 1623 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1624 1625 // If all preds have a single successor, then we know it is safe to insert 1626 // the load on the pred (?!?), so we can insert code to materialize the 1627 // pointer if it is not available. 1628 PHITransAddr Address(LI->getPointerOperand(), TD); 1629 Value *LoadPtr = 0; 1630 if (allSingleSucc) { 1631 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1632 *DT, NewInsts); 1633 } else { 1634 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1635 LoadPtr = Address.getAddr(); 1636 } 1637 1638 // If we couldn't find or insert a computation of this phi translated value, 1639 // we fail PRE. 1640 if (LoadPtr == 0) { 1641 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1642 << *LI->getPointerOperand() << "\n"); 1643 CanDoPRE = false; 1644 break; 1645 } 1646 1647 // Make sure it is valid to move this load here. We have to watch out for: 1648 // @1 = getelementptr (i8* p, ... 1649 // test p and branch if == 0 1650 // load @1 1651 // It is valid to have the getelementptr before the test, even if p can 1652 // be 0, as getelementptr only does address arithmetic. 1653 // If we are not pushing the value through any multiple-successor blocks 1654 // we do not have this case. Otherwise, check that the load is safe to 1655 // put anywhere; this can be improved, but should be conservatively safe. 1656 if (!allSingleSucc && 1657 // FIXME: REEVALUTE THIS. 1658 !isSafeToLoadUnconditionally(LoadPtr, 1659 UnavailablePred->getTerminator(), 1660 LI->getAlignment(), TD)) { 1661 CanDoPRE = false; 1662 break; 1663 } 1664 1665 I->second = LoadPtr; 1666 } 1667 1668 if (!CanDoPRE) { 1669 while (!NewInsts.empty()) { 1670 Instruction *I = NewInsts.pop_back_val(); 1671 if (MD) MD->removeInstruction(I); 1672 I->eraseFromParent(); 1673 } 1674 return false; 1675 } 1676 1677 // Okay, we can eliminate this load by inserting a reload in the predecessor 1678 // and using PHI construction to get the value in the other predecessors, do 1679 // it. 1680 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1681 DEBUG(if (!NewInsts.empty()) 1682 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1683 << *NewInsts.back() << '\n'); 1684 1685 // Assign value numbers to the new instructions. 1686 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1687 // FIXME: We really _ought_ to insert these value numbers into their 1688 // parent's availability map. However, in doing so, we risk getting into 1689 // ordering issues. If a block hasn't been processed yet, we would be 1690 // marking a value as AVAIL-IN, which isn't what we intend. 1691 VN.lookup_or_add(NewInsts[i]); 1692 } 1693 1694 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1695 E = PredLoads.end(); I != E; ++I) { 1696 BasicBlock *UnavailablePred = I->first; 1697 Value *LoadPtr = I->second; 1698 1699 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1700 LI->getAlignment(), 1701 UnavailablePred->getTerminator()); 1702 1703 // Transfer the old load's TBAA tag to the new load. 1704 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1705 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1706 1707 // Transfer DebugLoc. 1708 NewLoad->setDebugLoc(LI->getDebugLoc()); 1709 1710 // Add the newly created load. 1711 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1712 NewLoad)); 1713 MD->invalidateCachedPointerInfo(LoadPtr); 1714 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1715 } 1716 1717 // Perform PHI construction. 1718 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1719 LI->replaceAllUsesWith(V); 1720 if (isa<PHINode>(V)) 1721 V->takeName(LI); 1722 if (V->getType()->isPointerTy()) 1723 MD->invalidateCachedPointerInfo(V); 1724 markInstructionForDeletion(LI); 1725 ++NumPRELoad; 1726 return true; 1727} 1728 1729/// processLoad - Attempt to eliminate a load, first by eliminating it 1730/// locally, and then attempting non-local elimination if that fails. 1731bool GVN::processLoad(LoadInst *L) { 1732 if (!MD) 1733 return false; 1734 1735 if (!L->isSimple()) 1736 return false; 1737 1738 if (L->use_empty()) { 1739 markInstructionForDeletion(L); 1740 return true; 1741 } 1742 1743 // ... to a pointer that has been loaded from before... 1744 MemDepResult Dep = MD->getDependency(L); 1745 1746 // If we have a clobber and target data is around, see if this is a clobber 1747 // that we can fix up through code synthesis. 1748 if (Dep.isClobber() && TD) { 1749 // Check to see if we have something like this: 1750 // store i32 123, i32* %P 1751 // %A = bitcast i32* %P to i8* 1752 // %B = gep i8* %A, i32 1 1753 // %C = load i8* %B 1754 // 1755 // We could do that by recognizing if the clobber instructions are obviously 1756 // a common base + constant offset, and if the previous store (or memset) 1757 // completely covers this load. This sort of thing can happen in bitfield 1758 // access code. 1759 Value *AvailVal = 0; 1760 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1761 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1762 L->getPointerOperand(), 1763 DepSI, *TD); 1764 if (Offset != -1) 1765 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1766 L->getType(), L, *TD); 1767 } 1768 1769 // Check to see if we have something like this: 1770 // load i32* P 1771 // load i8* (P+1) 1772 // if we have this, replace the later with an extraction from the former. 1773 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1774 // If this is a clobber and L is the first instruction in its block, then 1775 // we have the first instruction in the entry block. 1776 if (DepLI == L) 1777 return false; 1778 1779 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1780 L->getPointerOperand(), 1781 DepLI, *TD); 1782 if (Offset != -1) 1783 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1784 } 1785 1786 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1787 // a value on from it. 1788 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1789 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1790 L->getPointerOperand(), 1791 DepMI, *TD); 1792 if (Offset != -1) 1793 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1794 } 1795 1796 if (AvailVal) { 1797 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1798 << *AvailVal << '\n' << *L << "\n\n\n"); 1799 1800 // Replace the load! 1801 L->replaceAllUsesWith(AvailVal); 1802 if (AvailVal->getType()->isPointerTy()) 1803 MD->invalidateCachedPointerInfo(AvailVal); 1804 markInstructionForDeletion(L); 1805 ++NumGVNLoad; 1806 return true; 1807 } 1808 } 1809 1810 // If the value isn't available, don't do anything! 1811 if (Dep.isClobber()) { 1812 DEBUG( 1813 // fast print dep, using operator<< on instruction is too slow. 1814 dbgs() << "GVN: load "; 1815 WriteAsOperand(dbgs(), L); 1816 Instruction *I = Dep.getInst(); 1817 dbgs() << " is clobbered by " << *I << '\n'; 1818 ); 1819 return false; 1820 } 1821 1822 // If it is defined in another block, try harder. 1823 if (Dep.isNonLocal()) 1824 return processNonLocalLoad(L); 1825 1826 if (!Dep.isDef()) { 1827 DEBUG( 1828 // fast print dep, using operator<< on instruction is too slow. 1829 dbgs() << "GVN: load "; 1830 WriteAsOperand(dbgs(), L); 1831 dbgs() << " has unknown dependence\n"; 1832 ); 1833 return false; 1834 } 1835 1836 Instruction *DepInst = Dep.getInst(); 1837 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1838 Value *StoredVal = DepSI->getValueOperand(); 1839 1840 // The store and load are to a must-aliased pointer, but they may not 1841 // actually have the same type. See if we know how to reuse the stored 1842 // value (depending on its type). 1843 if (StoredVal->getType() != L->getType()) { 1844 if (TD) { 1845 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1846 L, *TD); 1847 if (StoredVal == 0) 1848 return false; 1849 1850 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1851 << '\n' << *L << "\n\n\n"); 1852 } 1853 else 1854 return false; 1855 } 1856 1857 // Remove it! 1858 L->replaceAllUsesWith(StoredVal); 1859 if (StoredVal->getType()->isPointerTy()) 1860 MD->invalidateCachedPointerInfo(StoredVal); 1861 markInstructionForDeletion(L); 1862 ++NumGVNLoad; 1863 return true; 1864 } 1865 1866 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1867 Value *AvailableVal = DepLI; 1868 1869 // The loads are of a must-aliased pointer, but they may not actually have 1870 // the same type. See if we know how to reuse the previously loaded value 1871 // (depending on its type). 1872 if (DepLI->getType() != L->getType()) { 1873 if (TD) { 1874 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1875 L, *TD); 1876 if (AvailableVal == 0) 1877 return false; 1878 1879 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1880 << "\n" << *L << "\n\n\n"); 1881 } 1882 else 1883 return false; 1884 } 1885 1886 // Remove it! 1887 L->replaceAllUsesWith(AvailableVal); 1888 if (DepLI->getType()->isPointerTy()) 1889 MD->invalidateCachedPointerInfo(DepLI); 1890 markInstructionForDeletion(L); 1891 ++NumGVNLoad; 1892 return true; 1893 } 1894 1895 // If this load really doesn't depend on anything, then we must be loading an 1896 // undef value. This can happen when loading for a fresh allocation with no 1897 // intervening stores, for example. 1898 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1899 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1900 markInstructionForDeletion(L); 1901 ++NumGVNLoad; 1902 return true; 1903 } 1904 1905 // If this load occurs either right after a lifetime begin, 1906 // then the loaded value is undefined. 1907 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1908 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1909 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1910 markInstructionForDeletion(L); 1911 ++NumGVNLoad; 1912 return true; 1913 } 1914 } 1915 1916 return false; 1917} 1918 1919// findLeader - In order to find a leader for a given value number at a 1920// specific basic block, we first obtain the list of all Values for that number, 1921// and then scan the list to find one whose block dominates the block in 1922// question. This is fast because dominator tree queries consist of only 1923// a few comparisons of DFS numbers. 1924Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1925 LeaderTableEntry Vals = LeaderTable[num]; 1926 if (!Vals.Val) return 0; 1927 1928 Value *Val = 0; 1929 if (DT->dominates(Vals.BB, BB)) { 1930 Val = Vals.Val; 1931 if (isa<Constant>(Val)) return Val; 1932 } 1933 1934 LeaderTableEntry* Next = Vals.Next; 1935 while (Next) { 1936 if (DT->dominates(Next->BB, BB)) { 1937 if (isa<Constant>(Next->Val)) return Next->Val; 1938 if (!Val) Val = Next->Val; 1939 } 1940 1941 Next = Next->Next; 1942 } 1943 1944 return Val; 1945} 1946 1947/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1948/// use is dominated by the given basic block. Returns the number of uses that 1949/// were replaced. 1950unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1951 BasicBlock *Root) { 1952 unsigned Count = 0; 1953 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1954 UI != UE; ) { 1955 Use &U = (UI++).getUse(); 1956 if (DT->dominates(Root, cast<Instruction>(U.getUser())->getParent())) { 1957 U.set(To); 1958 ++Count; 1959 } 1960 } 1961 return Count; 1962} 1963 1964/// propagateEquality - The given values are known to be equal in every block 1965/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 1966/// 'RHS' everywhere in the scope. Returns whether a change was made. 1967bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) { 1968 if (LHS == RHS) return false; 1969 assert(LHS->getType() == RHS->getType() && "Equal but types differ!"); 1970 1971 // Don't try to propagate equalities between constants. 1972 if (isa<Constant>(LHS) && isa<Constant>(RHS)) 1973 return false; 1974 1975 // Make sure that any constants are on the right-hand side. In general the 1976 // best results are obtained by placing the longest lived value on the RHS. 1977 if (isa<Constant>(LHS)) 1978 std::swap(LHS, RHS); 1979 1980 // If neither term is constant then bail out. This is not for correctness, 1981 // it's just that the non-constant case is much less useful: it occurs just 1982 // as often as the constant case but handling it hardly ever results in an 1983 // improvement. 1984 if (!isa<Constant>(RHS)) 1985 return false; 1986 1987 // If value numbering later deduces that an instruction in the scope is equal 1988 // to 'LHS' then ensure it will be turned into 'RHS'. 1989 addToLeaderTable(VN.lookup_or_add(LHS), RHS, Root); 1990 1991 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 1992 // LHS always has at least one use that is not dominated by Root, this will 1993 // never do anything if LHS has only one use. 1994 bool Changed = false; 1995 if (!LHS->hasOneUse()) { 1996 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 1997 Changed |= NumReplacements > 0; 1998 NumGVNEqProp += NumReplacements; 1999 } 2000 2001 // Now try to deduce additional equalities from this one. For example, if the 2002 // known equality was "(A != B)" == "false" then it follows that A and B are 2003 // equal in the scope. Only boolean equalities with an explicit true or false 2004 // RHS are currently supported. 2005 if (!RHS->getType()->isIntegerTy(1)) 2006 // Not a boolean equality - bail out. 2007 return Changed; 2008 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2009 if (!CI) 2010 // RHS neither 'true' nor 'false' - bail out. 2011 return Changed; 2012 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2013 bool isKnownTrue = CI->isAllOnesValue(); 2014 bool isKnownFalse = !isKnownTrue; 2015 2016 // If "A && B" is known true then both A and B are known true. If "A || B" 2017 // is known false then both A and B are known false. 2018 Value *A, *B; 2019 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2020 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2021 Changed |= propagateEquality(A, RHS, Root); 2022 Changed |= propagateEquality(B, RHS, Root); 2023 return Changed; 2024 } 2025 2026 // If we are propagating an equality like "(A == B)" == "true" then also 2027 // propagate the equality A == B. When propagating a comparison such as 2028 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2029 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2030 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2031 2032 // If "A == B" is known true, or "A != B" is known false, then replace 2033 // A with B everywhere in the scope. 2034 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2035 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2036 Changed |= propagateEquality(Op0, Op1, Root); 2037 2038 // If "A >= B" is known true, replace "A < B" with false everywhere. 2039 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2040 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2041 // Since we don't have the instruction "A < B" immediately to hand, work out 2042 // the value number that it would have and use that to find an appropriate 2043 // instruction (if any). 2044 unsigned Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2045 Value *NotCmp = findLeader(Root, Num); 2046 if (NotCmp && isa<Instruction>(NotCmp)) { 2047 unsigned NumReplacements = 2048 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2049 Changed |= NumReplacements > 0; 2050 NumGVNEqProp += NumReplacements; 2051 } 2052 // Ensure that any instruction in scope that gets the "A < B" value number 2053 // is replaced with false. 2054 addToLeaderTable(Num, NotVal, Root); 2055 2056 return Changed; 2057 } 2058 2059 return Changed; 2060} 2061 2062/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2063/// true if every path from the entry block to 'Dst' passes via this edge. In 2064/// particular 'Dst' must not be reachable via another edge from 'Src'. 2065static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst, 2066 DominatorTree *DT) { 2067 // While in theory it is interesting to consider the case in which Dst has 2068 // more than one predecessor, because Dst might be part of a loop which is 2069 // only reachable from Src, in practice it is pointless since at the time 2070 // GVN runs all such loops have preheaders, which means that Dst will have 2071 // been changed to have only one predecessor, namely Src. 2072 BasicBlock *Pred = Dst->getSinglePredecessor(); 2073 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2074 (void)Src; 2075 return Pred != 0; 2076} 2077 2078/// processInstruction - When calculating availability, handle an instruction 2079/// by inserting it into the appropriate sets 2080bool GVN::processInstruction(Instruction *I) { 2081 // Ignore dbg info intrinsics. 2082 if (isa<DbgInfoIntrinsic>(I)) 2083 return false; 2084 2085 // If the instruction can be easily simplified then do so now in preference 2086 // to value numbering it. Value numbering often exposes redundancies, for 2087 // example if it determines that %y is equal to %x then the instruction 2088 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2089 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2090 I->replaceAllUsesWith(V); 2091 if (MD && V->getType()->isPointerTy()) 2092 MD->invalidateCachedPointerInfo(V); 2093 markInstructionForDeletion(I); 2094 ++NumGVNSimpl; 2095 return true; 2096 } 2097 2098 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2099 if (processLoad(LI)) 2100 return true; 2101 2102 unsigned Num = VN.lookup_or_add(LI); 2103 addToLeaderTable(Num, LI, LI->getParent()); 2104 return false; 2105 } 2106 2107 // For conditional branches, we can perform simple conditional propagation on 2108 // the condition value itself. 2109 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2110 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2111 return false; 2112 2113 Value *BranchCond = BI->getCondition(); 2114 2115 BasicBlock *TrueSucc = BI->getSuccessor(0); 2116 BasicBlock *FalseSucc = BI->getSuccessor(1); 2117 BasicBlock *Parent = BI->getParent(); 2118 bool Changed = false; 2119 2120 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT)) 2121 Changed |= propagateEquality(BranchCond, 2122 ConstantInt::getTrue(TrueSucc->getContext()), 2123 TrueSucc); 2124 2125 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT)) 2126 Changed |= propagateEquality(BranchCond, 2127 ConstantInt::getFalse(FalseSucc->getContext()), 2128 FalseSucc); 2129 2130 return Changed; 2131 } 2132 2133 // For switches, propagate the case values into the case destinations. 2134 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2135 Value *SwitchCond = SI->getCondition(); 2136 BasicBlock *Parent = SI->getParent(); 2137 bool Changed = false; 2138 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) { 2139 BasicBlock *Dst = SI->getCaseSuccessor(i); 2140 if (isOnlyReachableViaThisEdge(Parent, Dst, DT)) 2141 Changed |= propagateEquality(SwitchCond, SI->getCaseValue(i), Dst); 2142 } 2143 return Changed; 2144 } 2145 2146 // Instructions with void type don't return a value, so there's 2147 // no point in trying to find redudancies in them. 2148 if (I->getType()->isVoidTy()) return false; 2149 2150 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2151 unsigned Num = VN.lookup_or_add(I); 2152 2153 // Allocations are always uniquely numbered, so we can save time and memory 2154 // by fast failing them. 2155 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2156 addToLeaderTable(Num, I, I->getParent()); 2157 return false; 2158 } 2159 2160 // If the number we were assigned was a brand new VN, then we don't 2161 // need to do a lookup to see if the number already exists 2162 // somewhere in the domtree: it can't! 2163 if (Num == NextNum) { 2164 addToLeaderTable(Num, I, I->getParent()); 2165 return false; 2166 } 2167 2168 // Perform fast-path value-number based elimination of values inherited from 2169 // dominators. 2170 Value *repl = findLeader(I->getParent(), Num); 2171 if (repl == 0) { 2172 // Failure, just remember this instance for future use. 2173 addToLeaderTable(Num, I, I->getParent()); 2174 return false; 2175 } 2176 2177 // Remove it! 2178 I->replaceAllUsesWith(repl); 2179 if (MD && repl->getType()->isPointerTy()) 2180 MD->invalidateCachedPointerInfo(repl); 2181 markInstructionForDeletion(I); 2182 return true; 2183} 2184 2185/// runOnFunction - This is the main transformation entry point for a function. 2186bool GVN::runOnFunction(Function& F) { 2187 if (!NoLoads) 2188 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2189 DT = &getAnalysis<DominatorTree>(); 2190 TD = getAnalysisIfAvailable<TargetData>(); 2191 TLI = &getAnalysis<TargetLibraryInfo>(); 2192 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2193 VN.setMemDep(MD); 2194 VN.setDomTree(DT); 2195 2196 bool Changed = false; 2197 bool ShouldContinue = true; 2198 2199 // Merge unconditional branches, allowing PRE to catch more 2200 // optimization opportunities. 2201 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2202 BasicBlock *BB = FI++; 2203 2204 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2205 if (removedBlock) ++NumGVNBlocks; 2206 2207 Changed |= removedBlock; 2208 } 2209 2210 unsigned Iteration = 0; 2211 while (ShouldContinue) { 2212 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2213 ShouldContinue = iterateOnFunction(F); 2214 if (splitCriticalEdges()) 2215 ShouldContinue = true; 2216 Changed |= ShouldContinue; 2217 ++Iteration; 2218 } 2219 2220 if (EnablePRE) { 2221 bool PREChanged = true; 2222 while (PREChanged) { 2223 PREChanged = performPRE(F); 2224 Changed |= PREChanged; 2225 } 2226 } 2227 // FIXME: Should perform GVN again after PRE does something. PRE can move 2228 // computations into blocks where they become fully redundant. Note that 2229 // we can't do this until PRE's critical edge splitting updates memdep. 2230 // Actually, when this happens, we should just fully integrate PRE into GVN. 2231 2232 cleanupGlobalSets(); 2233 2234 return Changed; 2235} 2236 2237 2238bool GVN::processBlock(BasicBlock *BB) { 2239 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2240 // (and incrementing BI before processing an instruction). 2241 assert(InstrsToErase.empty() && 2242 "We expect InstrsToErase to be empty across iterations"); 2243 bool ChangedFunction = false; 2244 2245 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2246 BI != BE;) { 2247 ChangedFunction |= processInstruction(BI); 2248 if (InstrsToErase.empty()) { 2249 ++BI; 2250 continue; 2251 } 2252 2253 // If we need some instructions deleted, do it now. 2254 NumGVNInstr += InstrsToErase.size(); 2255 2256 // Avoid iterator invalidation. 2257 bool AtStart = BI == BB->begin(); 2258 if (!AtStart) 2259 --BI; 2260 2261 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2262 E = InstrsToErase.end(); I != E; ++I) { 2263 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2264 if (MD) MD->removeInstruction(*I); 2265 (*I)->eraseFromParent(); 2266 DEBUG(verifyRemoved(*I)); 2267 } 2268 InstrsToErase.clear(); 2269 2270 if (AtStart) 2271 BI = BB->begin(); 2272 else 2273 ++BI; 2274 } 2275 2276 return ChangedFunction; 2277} 2278 2279/// performPRE - Perform a purely local form of PRE that looks for diamond 2280/// control flow patterns and attempts to perform simple PRE at the join point. 2281bool GVN::performPRE(Function &F) { 2282 bool Changed = false; 2283 DenseMap<BasicBlock*, Value*> predMap; 2284 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2285 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2286 BasicBlock *CurrentBlock = *DI; 2287 2288 // Nothing to PRE in the entry block. 2289 if (CurrentBlock == &F.getEntryBlock()) continue; 2290 2291 // Don't perform PRE on a landing pad. 2292 if (CurrentBlock->isLandingPad()) continue; 2293 2294 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2295 BE = CurrentBlock->end(); BI != BE; ) { 2296 Instruction *CurInst = BI++; 2297 2298 if (isa<AllocaInst>(CurInst) || 2299 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2300 CurInst->getType()->isVoidTy() || 2301 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2302 isa<DbgInfoIntrinsic>(CurInst)) 2303 continue; 2304 2305 // We don't currently value number ANY inline asm calls. 2306 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2307 if (CallI->isInlineAsm()) 2308 continue; 2309 2310 uint32_t ValNo = VN.lookup(CurInst); 2311 2312 // Look for the predecessors for PRE opportunities. We're 2313 // only trying to solve the basic diamond case, where 2314 // a value is computed in the successor and one predecessor, 2315 // but not the other. We also explicitly disallow cases 2316 // where the successor is its own predecessor, because they're 2317 // more complicated to get right. 2318 unsigned NumWith = 0; 2319 unsigned NumWithout = 0; 2320 BasicBlock *PREPred = 0; 2321 predMap.clear(); 2322 2323 for (pred_iterator PI = pred_begin(CurrentBlock), 2324 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2325 BasicBlock *P = *PI; 2326 // We're not interested in PRE where the block is its 2327 // own predecessor, or in blocks with predecessors 2328 // that are not reachable. 2329 if (P == CurrentBlock) { 2330 NumWithout = 2; 2331 break; 2332 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2333 NumWithout = 2; 2334 break; 2335 } 2336 2337 Value* predV = findLeader(P, ValNo); 2338 if (predV == 0) { 2339 PREPred = P; 2340 ++NumWithout; 2341 } else if (predV == CurInst) { 2342 NumWithout = 2; 2343 } else { 2344 predMap[P] = predV; 2345 ++NumWith; 2346 } 2347 } 2348 2349 // Don't do PRE when it might increase code size, i.e. when 2350 // we would need to insert instructions in more than one pred. 2351 if (NumWithout != 1 || NumWith == 0) 2352 continue; 2353 2354 // Don't do PRE across indirect branch. 2355 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2356 continue; 2357 2358 // We can't do PRE safely on a critical edge, so instead we schedule 2359 // the edge to be split and perform the PRE the next time we iterate 2360 // on the function. 2361 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2362 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2363 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2364 continue; 2365 } 2366 2367 // Instantiate the expression in the predecessor that lacked it. 2368 // Because we are going top-down through the block, all value numbers 2369 // will be available in the predecessor by the time we need them. Any 2370 // that weren't originally present will have been instantiated earlier 2371 // in this loop. 2372 Instruction *PREInstr = CurInst->clone(); 2373 bool success = true; 2374 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2375 Value *Op = PREInstr->getOperand(i); 2376 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2377 continue; 2378 2379 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2380 PREInstr->setOperand(i, V); 2381 } else { 2382 success = false; 2383 break; 2384 } 2385 } 2386 2387 // Fail out if we encounter an operand that is not available in 2388 // the PRE predecessor. This is typically because of loads which 2389 // are not value numbered precisely. 2390 if (!success) { 2391 delete PREInstr; 2392 DEBUG(verifyRemoved(PREInstr)); 2393 continue; 2394 } 2395 2396 PREInstr->insertBefore(PREPred->getTerminator()); 2397 PREInstr->setName(CurInst->getName() + ".pre"); 2398 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2399 predMap[PREPred] = PREInstr; 2400 VN.add(PREInstr, ValNo); 2401 ++NumGVNPRE; 2402 2403 // Update the availability map to include the new instruction. 2404 addToLeaderTable(ValNo, PREInstr, PREPred); 2405 2406 // Create a PHI to make the value available in this block. 2407 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2408 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2409 CurInst->getName() + ".pre-phi", 2410 CurrentBlock->begin()); 2411 for (pred_iterator PI = PB; PI != PE; ++PI) { 2412 BasicBlock *P = *PI; 2413 Phi->addIncoming(predMap[P], P); 2414 } 2415 2416 VN.add(Phi, ValNo); 2417 addToLeaderTable(ValNo, Phi, CurrentBlock); 2418 Phi->setDebugLoc(CurInst->getDebugLoc()); 2419 CurInst->replaceAllUsesWith(Phi); 2420 if (Phi->getType()->isPointerTy()) { 2421 // Because we have added a PHI-use of the pointer value, it has now 2422 // "escaped" from alias analysis' perspective. We need to inform 2423 // AA of this. 2424 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2425 ++ii) { 2426 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2427 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2428 } 2429 2430 if (MD) 2431 MD->invalidateCachedPointerInfo(Phi); 2432 } 2433 VN.erase(CurInst); 2434 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2435 2436 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2437 if (MD) MD->removeInstruction(CurInst); 2438 CurInst->eraseFromParent(); 2439 DEBUG(verifyRemoved(CurInst)); 2440 Changed = true; 2441 } 2442 } 2443 2444 if (splitCriticalEdges()) 2445 Changed = true; 2446 2447 return Changed; 2448} 2449 2450/// splitCriticalEdges - Split critical edges found during the previous 2451/// iteration that may enable further optimization. 2452bool GVN::splitCriticalEdges() { 2453 if (toSplit.empty()) 2454 return false; 2455 do { 2456 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2457 SplitCriticalEdge(Edge.first, Edge.second, this); 2458 } while (!toSplit.empty()); 2459 if (MD) MD->invalidateCachedPredecessors(); 2460 return true; 2461} 2462 2463/// iterateOnFunction - Executes one iteration of GVN 2464bool GVN::iterateOnFunction(Function &F) { 2465 cleanupGlobalSets(); 2466 2467 // Top-down walk of the dominator tree 2468 bool Changed = false; 2469#if 0 2470 // Needed for value numbering with phi construction to work. 2471 ReversePostOrderTraversal<Function*> RPOT(&F); 2472 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2473 RE = RPOT.end(); RI != RE; ++RI) 2474 Changed |= processBlock(*RI); 2475#else 2476 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2477 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2478 Changed |= processBlock(DI->getBlock()); 2479#endif 2480 2481 return Changed; 2482} 2483 2484void GVN::cleanupGlobalSets() { 2485 VN.clear(); 2486 LeaderTable.clear(); 2487 TableAllocator.Reset(); 2488} 2489 2490/// verifyRemoved - Verify that the specified instruction does not occur in our 2491/// internal data structures. 2492void GVN::verifyRemoved(const Instruction *Inst) const { 2493 VN.verifyRemoved(Inst); 2494 2495 // Walk through the value number scope to make sure the instruction isn't 2496 // ferreted away in it. 2497 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2498 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2499 const LeaderTableEntry *Node = &I->second; 2500 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2501 2502 while (Node->Next) { 2503 Node = Node->Next; 2504 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2505 } 2506 } 2507} 2508