GVN.cpp revision 5dde20bfacdad9bc3ddc99eff93d52ed54312ed9
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/Hashing.h" 40#include "llvm/ADT/SmallPtrSet.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/Support/Allocator.h" 43#include "llvm/Support/CommandLine.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/Support/IRBuilder.h" 46#include "llvm/Support/PatternMatch.h" 47using namespace llvm; 48using namespace PatternMatch; 49 50STATISTIC(NumGVNInstr, "Number of instructions deleted"); 51STATISTIC(NumGVNLoad, "Number of loads deleted"); 52STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 53STATISTIC(NumGVNBlocks, "Number of blocks merged"); 54STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 55STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 56STATISTIC(NumPRELoad, "Number of loads PRE'd"); 57 58static cl::opt<bool> EnablePRE("enable-pre", 59 cl::init(true), cl::Hidden); 60static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 61 62// Maximum allowed recursion depth. 63static cl::opt<int> 64MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 65 cl::desc("Max recurse depth (default = 1000)")); 66 67//===----------------------------------------------------------------------===// 68// ValueTable Class 69//===----------------------------------------------------------------------===// 70 71/// This class holds the mapping between values and value numbers. It is used 72/// as an efficient mechanism to determine the expression-wise equivalence of 73/// two values. 74namespace { 75 struct Expression { 76 uint32_t opcode; 77 Type *type; 78 SmallVector<uint32_t, 4> varargs; 79 80 Expression(uint32_t o = ~2U) : opcode(o) { } 81 82 bool operator==(const Expression &other) const { 83 if (opcode != other.opcode) 84 return false; 85 if (opcode == ~0U || opcode == ~1U) 86 return true; 87 if (type != other.type) 88 return false; 89 if (varargs != other.varargs) 90 return false; 91 return true; 92 } 93 94 friend hash_code hash_value(const Expression &Value) { 95 return hash_combine(Value.opcode, Value.type, 96 hash_combine_range(Value.varargs.begin(), 97 Value.varargs.end())); 98 } 99 }; 100 101 class ValueTable { 102 DenseMap<Value*, uint32_t> valueNumbering; 103 DenseMap<Expression, uint32_t> expressionNumbering; 104 AliasAnalysis *AA; 105 MemoryDependenceAnalysis *MD; 106 DominatorTree *DT; 107 108 uint32_t nextValueNumber; 109 110 Expression create_expression(Instruction* I); 111 Expression create_cmp_expression(unsigned Opcode, 112 CmpInst::Predicate Predicate, 113 Value *LHS, Value *RHS); 114 Expression create_extractvalue_expression(ExtractValueInst* EI); 115 uint32_t lookup_or_add_call(CallInst* C); 116 public: 117 ValueTable() : nextValueNumber(1) { } 118 uint32_t lookup_or_add(Value *V); 119 uint32_t lookup(Value *V) const; 120 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 121 Value *LHS, Value *RHS); 122 void add(Value *V, uint32_t num); 123 void clear(); 124 void erase(Value *v); 125 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 126 AliasAnalysis *getAliasAnalysis() const { return AA; } 127 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 128 void setDomTree(DominatorTree* D) { DT = D; } 129 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 130 void verifyRemoved(const Value *) const; 131 }; 132} 133 134namespace llvm { 135template <> struct DenseMapInfo<Expression> { 136 static inline Expression getEmptyKey() { 137 return ~0U; 138 } 139 140 static inline Expression getTombstoneKey() { 141 return ~1U; 142 } 143 144 static unsigned getHashValue(const Expression e) { 145 using llvm::hash_value; 146 return static_cast<unsigned>(hash_value(e)); 147 } 148 static bool isEqual(const Expression &LHS, const Expression &RHS) { 149 return LHS == RHS; 150 } 151}; 152 153} 154 155//===----------------------------------------------------------------------===// 156// ValueTable Internal Functions 157//===----------------------------------------------------------------------===// 158 159Expression ValueTable::create_expression(Instruction *I) { 160 Expression e; 161 e.type = I->getType(); 162 e.opcode = I->getOpcode(); 163 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 164 OI != OE; ++OI) 165 e.varargs.push_back(lookup_or_add(*OI)); 166 if (I->isCommutative()) { 167 // Ensure that commutative instructions that only differ by a permutation 168 // of their operands get the same value number by sorting the operand value 169 // numbers. Since all commutative instructions have two operands it is more 170 // efficient to sort by hand rather than using, say, std::sort. 171 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 172 if (e.varargs[0] > e.varargs[1]) 173 std::swap(e.varargs[0], e.varargs[1]); 174 } 175 176 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 177 // Sort the operand value numbers so x<y and y>x get the same value number. 178 CmpInst::Predicate Predicate = C->getPredicate(); 179 if (e.varargs[0] > e.varargs[1]) { 180 std::swap(e.varargs[0], e.varargs[1]); 181 Predicate = CmpInst::getSwappedPredicate(Predicate); 182 } 183 e.opcode = (C->getOpcode() << 8) | Predicate; 184 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 185 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 186 II != IE; ++II) 187 e.varargs.push_back(*II); 188 } 189 190 return e; 191} 192 193Expression ValueTable::create_cmp_expression(unsigned Opcode, 194 CmpInst::Predicate Predicate, 195 Value *LHS, Value *RHS) { 196 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 197 "Not a comparison!"); 198 Expression e; 199 e.type = CmpInst::makeCmpResultType(LHS->getType()); 200 e.varargs.push_back(lookup_or_add(LHS)); 201 e.varargs.push_back(lookup_or_add(RHS)); 202 203 // Sort the operand value numbers so x<y and y>x get the same value number. 204 if (e.varargs[0] > e.varargs[1]) { 205 std::swap(e.varargs[0], e.varargs[1]); 206 Predicate = CmpInst::getSwappedPredicate(Predicate); 207 } 208 e.opcode = (Opcode << 8) | Predicate; 209 return e; 210} 211 212Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 213 assert(EI != 0 && "Not an ExtractValueInst?"); 214 Expression e; 215 e.type = EI->getType(); 216 e.opcode = 0; 217 218 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 219 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 220 // EI might be an extract from one of our recognised intrinsics. If it 221 // is we'll synthesize a semantically equivalent expression instead on 222 // an extract value expression. 223 switch (I->getIntrinsicID()) { 224 case Intrinsic::sadd_with_overflow: 225 case Intrinsic::uadd_with_overflow: 226 e.opcode = Instruction::Add; 227 break; 228 case Intrinsic::ssub_with_overflow: 229 case Intrinsic::usub_with_overflow: 230 e.opcode = Instruction::Sub; 231 break; 232 case Intrinsic::smul_with_overflow: 233 case Intrinsic::umul_with_overflow: 234 e.opcode = Instruction::Mul; 235 break; 236 default: 237 break; 238 } 239 240 if (e.opcode != 0) { 241 // Intrinsic recognized. Grab its args to finish building the expression. 242 assert(I->getNumArgOperands() == 2 && 243 "Expect two args for recognised intrinsics."); 244 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 245 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 246 return e; 247 } 248 } 249 250 // Not a recognised intrinsic. Fall back to producing an extract value 251 // expression. 252 e.opcode = EI->getOpcode(); 253 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 254 OI != OE; ++OI) 255 e.varargs.push_back(lookup_or_add(*OI)); 256 257 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 258 II != IE; ++II) 259 e.varargs.push_back(*II); 260 261 return e; 262} 263 264//===----------------------------------------------------------------------===// 265// ValueTable External Functions 266//===----------------------------------------------------------------------===// 267 268/// add - Insert a value into the table with a specified value number. 269void ValueTable::add(Value *V, uint32_t num) { 270 valueNumbering.insert(std::make_pair(V, num)); 271} 272 273uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 274 if (AA->doesNotAccessMemory(C)) { 275 Expression exp = create_expression(C); 276 uint32_t& e = expressionNumbering[exp]; 277 if (!e) e = nextValueNumber++; 278 valueNumbering[C] = e; 279 return e; 280 } else if (AA->onlyReadsMemory(C)) { 281 Expression exp = create_expression(C); 282 uint32_t& e = expressionNumbering[exp]; 283 if (!e) { 284 e = nextValueNumber++; 285 valueNumbering[C] = e; 286 return e; 287 } 288 if (!MD) { 289 e = nextValueNumber++; 290 valueNumbering[C] = e; 291 return e; 292 } 293 294 MemDepResult local_dep = MD->getDependency(C); 295 296 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 297 valueNumbering[C] = nextValueNumber; 298 return nextValueNumber++; 299 } 300 301 if (local_dep.isDef()) { 302 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 303 304 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 305 valueNumbering[C] = nextValueNumber; 306 return nextValueNumber++; 307 } 308 309 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 310 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 311 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 312 if (c_vn != cd_vn) { 313 valueNumbering[C] = nextValueNumber; 314 return nextValueNumber++; 315 } 316 } 317 318 uint32_t v = lookup_or_add(local_cdep); 319 valueNumbering[C] = v; 320 return v; 321 } 322 323 // Non-local case. 324 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 325 MD->getNonLocalCallDependency(CallSite(C)); 326 // FIXME: Move the checking logic to MemDep! 327 CallInst* cdep = 0; 328 329 // Check to see if we have a single dominating call instruction that is 330 // identical to C. 331 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 332 const NonLocalDepEntry *I = &deps[i]; 333 if (I->getResult().isNonLocal()) 334 continue; 335 336 // We don't handle non-definitions. If we already have a call, reject 337 // instruction dependencies. 338 if (!I->getResult().isDef() || cdep != 0) { 339 cdep = 0; 340 break; 341 } 342 343 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 344 // FIXME: All duplicated with non-local case. 345 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 346 cdep = NonLocalDepCall; 347 continue; 348 } 349 350 cdep = 0; 351 break; 352 } 353 354 if (!cdep) { 355 valueNumbering[C] = nextValueNumber; 356 return nextValueNumber++; 357 } 358 359 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 360 valueNumbering[C] = nextValueNumber; 361 return nextValueNumber++; 362 } 363 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 364 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 365 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 366 if (c_vn != cd_vn) { 367 valueNumbering[C] = nextValueNumber; 368 return nextValueNumber++; 369 } 370 } 371 372 uint32_t v = lookup_or_add(cdep); 373 valueNumbering[C] = v; 374 return v; 375 376 } else { 377 valueNumbering[C] = nextValueNumber; 378 return nextValueNumber++; 379 } 380} 381 382/// lookup_or_add - Returns the value number for the specified value, assigning 383/// it a new number if it did not have one before. 384uint32_t ValueTable::lookup_or_add(Value *V) { 385 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 386 if (VI != valueNumbering.end()) 387 return VI->second; 388 389 if (!isa<Instruction>(V)) { 390 valueNumbering[V] = nextValueNumber; 391 return nextValueNumber++; 392 } 393 394 Instruction* I = cast<Instruction>(V); 395 Expression exp; 396 switch (I->getOpcode()) { 397 case Instruction::Call: 398 return lookup_or_add_call(cast<CallInst>(I)); 399 case Instruction::Add: 400 case Instruction::FAdd: 401 case Instruction::Sub: 402 case Instruction::FSub: 403 case Instruction::Mul: 404 case Instruction::FMul: 405 case Instruction::UDiv: 406 case Instruction::SDiv: 407 case Instruction::FDiv: 408 case Instruction::URem: 409 case Instruction::SRem: 410 case Instruction::FRem: 411 case Instruction::Shl: 412 case Instruction::LShr: 413 case Instruction::AShr: 414 case Instruction::And: 415 case Instruction::Or : 416 case Instruction::Xor: 417 case Instruction::ICmp: 418 case Instruction::FCmp: 419 case Instruction::Trunc: 420 case Instruction::ZExt: 421 case Instruction::SExt: 422 case Instruction::FPToUI: 423 case Instruction::FPToSI: 424 case Instruction::UIToFP: 425 case Instruction::SIToFP: 426 case Instruction::FPTrunc: 427 case Instruction::FPExt: 428 case Instruction::PtrToInt: 429 case Instruction::IntToPtr: 430 case Instruction::BitCast: 431 case Instruction::Select: 432 case Instruction::ExtractElement: 433 case Instruction::InsertElement: 434 case Instruction::ShuffleVector: 435 case Instruction::InsertValue: 436 case Instruction::GetElementPtr: 437 exp = create_expression(I); 438 break; 439 case Instruction::ExtractValue: 440 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 441 break; 442 default: 443 valueNumbering[V] = nextValueNumber; 444 return nextValueNumber++; 445 } 446 447 uint32_t& e = expressionNumbering[exp]; 448 if (!e) e = nextValueNumber++; 449 valueNumbering[V] = e; 450 return e; 451} 452 453/// lookup - Returns the value number of the specified value. Fails if 454/// the value has not yet been numbered. 455uint32_t ValueTable::lookup(Value *V) const { 456 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 457 assert(VI != valueNumbering.end() && "Value not numbered?"); 458 return VI->second; 459} 460 461/// lookup_or_add_cmp - Returns the value number of the given comparison, 462/// assigning it a new number if it did not have one before. Useful when 463/// we deduced the result of a comparison, but don't immediately have an 464/// instruction realizing that comparison to hand. 465uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 466 CmpInst::Predicate Predicate, 467 Value *LHS, Value *RHS) { 468 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 469 uint32_t& e = expressionNumbering[exp]; 470 if (!e) e = nextValueNumber++; 471 return e; 472} 473 474/// clear - Remove all entries from the ValueTable. 475void ValueTable::clear() { 476 valueNumbering.clear(); 477 expressionNumbering.clear(); 478 nextValueNumber = 1; 479} 480 481/// erase - Remove a value from the value numbering. 482void ValueTable::erase(Value *V) { 483 valueNumbering.erase(V); 484} 485 486/// verifyRemoved - Verify that the value is removed from all internal data 487/// structures. 488void ValueTable::verifyRemoved(const Value *V) const { 489 for (DenseMap<Value*, uint32_t>::const_iterator 490 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 491 assert(I->first != V && "Inst still occurs in value numbering map!"); 492 } 493} 494 495//===----------------------------------------------------------------------===// 496// GVN Pass 497//===----------------------------------------------------------------------===// 498 499namespace { 500 501 class GVN : public FunctionPass { 502 bool NoLoads; 503 MemoryDependenceAnalysis *MD; 504 DominatorTree *DT; 505 const TargetData *TD; 506 const TargetLibraryInfo *TLI; 507 508 ValueTable VN; 509 510 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 511 /// have that value number. Use findLeader to query it. 512 struct LeaderTableEntry { 513 Value *Val; 514 BasicBlock *BB; 515 LeaderTableEntry *Next; 516 }; 517 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 518 BumpPtrAllocator TableAllocator; 519 520 SmallVector<Instruction*, 8> InstrsToErase; 521 public: 522 static char ID; // Pass identification, replacement for typeid 523 explicit GVN(bool noloads = false) 524 : FunctionPass(ID), NoLoads(noloads), MD(0) { 525 initializeGVNPass(*PassRegistry::getPassRegistry()); 526 } 527 528 bool runOnFunction(Function &F); 529 530 /// markInstructionForDeletion - This removes the specified instruction from 531 /// our various maps and marks it for deletion. 532 void markInstructionForDeletion(Instruction *I) { 533 VN.erase(I); 534 InstrsToErase.push_back(I); 535 } 536 537 const TargetData *getTargetData() const { return TD; } 538 DominatorTree &getDominatorTree() const { return *DT; } 539 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 540 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 541 private: 542 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 543 /// its value number. 544 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 545 LeaderTableEntry &Curr = LeaderTable[N]; 546 if (!Curr.Val) { 547 Curr.Val = V; 548 Curr.BB = BB; 549 return; 550 } 551 552 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 553 Node->Val = V; 554 Node->BB = BB; 555 Node->Next = Curr.Next; 556 Curr.Next = Node; 557 } 558 559 /// removeFromLeaderTable - Scan the list of values corresponding to a given 560 /// value number, and remove the given value if encountered. 561 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 562 LeaderTableEntry* Prev = 0; 563 LeaderTableEntry* Curr = &LeaderTable[N]; 564 565 while (Curr->Val != V || Curr->BB != BB) { 566 Prev = Curr; 567 Curr = Curr->Next; 568 } 569 570 if (Prev) { 571 Prev->Next = Curr->Next; 572 } else { 573 if (!Curr->Next) { 574 Curr->Val = 0; 575 Curr->BB = 0; 576 } else { 577 LeaderTableEntry* Next = Curr->Next; 578 Curr->Val = Next->Val; 579 Curr->BB = Next->BB; 580 Curr->Next = Next->Next; 581 } 582 } 583 } 584 585 // List of critical edges to be split between iterations. 586 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 587 588 // This transformation requires dominator postdominator info 589 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 590 AU.addRequired<DominatorTree>(); 591 AU.addRequired<TargetLibraryInfo>(); 592 if (!NoLoads) 593 AU.addRequired<MemoryDependenceAnalysis>(); 594 AU.addRequired<AliasAnalysis>(); 595 596 AU.addPreserved<DominatorTree>(); 597 AU.addPreserved<AliasAnalysis>(); 598 } 599 600 601 // Helper fuctions 602 // FIXME: eliminate or document these better 603 bool processLoad(LoadInst *L); 604 bool processInstruction(Instruction *I); 605 bool processNonLocalLoad(LoadInst *L); 606 bool processBlock(BasicBlock *BB); 607 void dump(DenseMap<uint32_t, Value*> &d); 608 bool iterateOnFunction(Function &F); 609 bool performPRE(Function &F); 610 Value *findLeader(BasicBlock *BB, uint32_t num); 611 void cleanupGlobalSets(); 612 void verifyRemoved(const Instruction *I) const; 613 bool splitCriticalEdges(); 614 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 615 BasicBlock *Root); 616 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root); 617 }; 618 619 char GVN::ID = 0; 620} 621 622// createGVNPass - The public interface to this file... 623FunctionPass *llvm::createGVNPass(bool NoLoads) { 624 return new GVN(NoLoads); 625} 626 627INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 628INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 629INITIALIZE_PASS_DEPENDENCY(DominatorTree) 630INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 631INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 632INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 633 634void GVN::dump(DenseMap<uint32_t, Value*>& d) { 635 errs() << "{\n"; 636 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 637 E = d.end(); I != E; ++I) { 638 errs() << I->first << "\n"; 639 I->second->dump(); 640 } 641 errs() << "}\n"; 642} 643 644/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 645/// we're analyzing is fully available in the specified block. As we go, keep 646/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 647/// map is actually a tri-state map with the following values: 648/// 0) we know the block *is not* fully available. 649/// 1) we know the block *is* fully available. 650/// 2) we do not know whether the block is fully available or not, but we are 651/// currently speculating that it will be. 652/// 3) we are speculating for this block and have used that to speculate for 653/// other blocks. 654static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 655 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 656 uint32_t RecurseDepth) { 657 if (RecurseDepth > MaxRecurseDepth) 658 return false; 659 660 // Optimistically assume that the block is fully available and check to see 661 // if we already know about this block in one lookup. 662 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 663 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 664 665 // If the entry already existed for this block, return the precomputed value. 666 if (!IV.second) { 667 // If this is a speculative "available" value, mark it as being used for 668 // speculation of other blocks. 669 if (IV.first->second == 2) 670 IV.first->second = 3; 671 return IV.first->second != 0; 672 } 673 674 // Otherwise, see if it is fully available in all predecessors. 675 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 676 677 // If this block has no predecessors, it isn't live-in here. 678 if (PI == PE) 679 goto SpeculationFailure; 680 681 for (; PI != PE; ++PI) 682 // If the value isn't fully available in one of our predecessors, then it 683 // isn't fully available in this block either. Undo our previous 684 // optimistic assumption and bail out. 685 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 686 goto SpeculationFailure; 687 688 return true; 689 690// SpeculationFailure - If we get here, we found out that this is not, after 691// all, a fully-available block. We have a problem if we speculated on this and 692// used the speculation to mark other blocks as available. 693SpeculationFailure: 694 char &BBVal = FullyAvailableBlocks[BB]; 695 696 // If we didn't speculate on this, just return with it set to false. 697 if (BBVal == 2) { 698 BBVal = 0; 699 return false; 700 } 701 702 // If we did speculate on this value, we could have blocks set to 1 that are 703 // incorrect. Walk the (transitive) successors of this block and mark them as 704 // 0 if set to one. 705 SmallVector<BasicBlock*, 32> BBWorklist; 706 BBWorklist.push_back(BB); 707 708 do { 709 BasicBlock *Entry = BBWorklist.pop_back_val(); 710 // Note that this sets blocks to 0 (unavailable) if they happen to not 711 // already be in FullyAvailableBlocks. This is safe. 712 char &EntryVal = FullyAvailableBlocks[Entry]; 713 if (EntryVal == 0) continue; // Already unavailable. 714 715 // Mark as unavailable. 716 EntryVal = 0; 717 718 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 719 BBWorklist.push_back(*I); 720 } while (!BBWorklist.empty()); 721 722 return false; 723} 724 725 726/// CanCoerceMustAliasedValueToLoad - Return true if 727/// CoerceAvailableValueToLoadType will succeed. 728static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 729 Type *LoadTy, 730 const TargetData &TD) { 731 // If the loaded or stored value is an first class array or struct, don't try 732 // to transform them. We need to be able to bitcast to integer. 733 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 734 StoredVal->getType()->isStructTy() || 735 StoredVal->getType()->isArrayTy()) 736 return false; 737 738 // The store has to be at least as big as the load. 739 if (TD.getTypeSizeInBits(StoredVal->getType()) < 740 TD.getTypeSizeInBits(LoadTy)) 741 return false; 742 743 return true; 744} 745 746 747/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 748/// then a load from a must-aliased pointer of a different type, try to coerce 749/// the stored value. LoadedTy is the type of the load we want to replace and 750/// InsertPt is the place to insert new instructions. 751/// 752/// If we can't do it, return null. 753static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 754 Type *LoadedTy, 755 Instruction *InsertPt, 756 const TargetData &TD) { 757 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 758 return 0; 759 760 // If this is already the right type, just return it. 761 Type *StoredValTy = StoredVal->getType(); 762 763 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 764 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 765 766 // If the store and reload are the same size, we can always reuse it. 767 if (StoreSize == LoadSize) { 768 // Pointer to Pointer -> use bitcast. 769 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 770 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 771 772 // Convert source pointers to integers, which can be bitcast. 773 if (StoredValTy->isPointerTy()) { 774 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 775 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 776 } 777 778 Type *TypeToCastTo = LoadedTy; 779 if (TypeToCastTo->isPointerTy()) 780 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 781 782 if (StoredValTy != TypeToCastTo) 783 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 784 785 // Cast to pointer if the load needs a pointer type. 786 if (LoadedTy->isPointerTy()) 787 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 788 789 return StoredVal; 790 } 791 792 // If the loaded value is smaller than the available value, then we can 793 // extract out a piece from it. If the available value is too small, then we 794 // can't do anything. 795 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 796 797 // Convert source pointers to integers, which can be manipulated. 798 if (StoredValTy->isPointerTy()) { 799 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 800 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 801 } 802 803 // Convert vectors and fp to integer, which can be manipulated. 804 if (!StoredValTy->isIntegerTy()) { 805 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 806 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 807 } 808 809 // If this is a big-endian system, we need to shift the value down to the low 810 // bits so that a truncate will work. 811 if (TD.isBigEndian()) { 812 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 813 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 814 } 815 816 // Truncate the integer to the right size now. 817 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 818 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 819 820 if (LoadedTy == NewIntTy) 821 return StoredVal; 822 823 // If the result is a pointer, inttoptr. 824 if (LoadedTy->isPointerTy()) 825 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 826 827 // Otherwise, bitcast. 828 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 829} 830 831/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 832/// memdep query of a load that ends up being a clobbering memory write (store, 833/// memset, memcpy, memmove). This means that the write *may* provide bits used 834/// by the load but we can't be sure because the pointers don't mustalias. 835/// 836/// Check this case to see if there is anything more we can do before we give 837/// up. This returns -1 if we have to give up, or a byte number in the stored 838/// value of the piece that feeds the load. 839static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 840 Value *WritePtr, 841 uint64_t WriteSizeInBits, 842 const TargetData &TD) { 843 // If the loaded or stored value is a first class array or struct, don't try 844 // to transform them. We need to be able to bitcast to integer. 845 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 846 return -1; 847 848 int64_t StoreOffset = 0, LoadOffset = 0; 849 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 850 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 851 if (StoreBase != LoadBase) 852 return -1; 853 854 // If the load and store are to the exact same address, they should have been 855 // a must alias. AA must have gotten confused. 856 // FIXME: Study to see if/when this happens. One case is forwarding a memset 857 // to a load from the base of the memset. 858#if 0 859 if (LoadOffset == StoreOffset) { 860 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 861 << "Base = " << *StoreBase << "\n" 862 << "Store Ptr = " << *WritePtr << "\n" 863 << "Store Offs = " << StoreOffset << "\n" 864 << "Load Ptr = " << *LoadPtr << "\n"; 865 abort(); 866 } 867#endif 868 869 // If the load and store don't overlap at all, the store doesn't provide 870 // anything to the load. In this case, they really don't alias at all, AA 871 // must have gotten confused. 872 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 873 874 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 875 return -1; 876 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 877 LoadSize >>= 3; 878 879 880 bool isAAFailure = false; 881 if (StoreOffset < LoadOffset) 882 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 883 else 884 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 885 886 if (isAAFailure) { 887#if 0 888 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 889 << "Base = " << *StoreBase << "\n" 890 << "Store Ptr = " << *WritePtr << "\n" 891 << "Store Offs = " << StoreOffset << "\n" 892 << "Load Ptr = " << *LoadPtr << "\n"; 893 abort(); 894#endif 895 return -1; 896 } 897 898 // If the Load isn't completely contained within the stored bits, we don't 899 // have all the bits to feed it. We could do something crazy in the future 900 // (issue a smaller load then merge the bits in) but this seems unlikely to be 901 // valuable. 902 if (StoreOffset > LoadOffset || 903 StoreOffset+StoreSize < LoadOffset+LoadSize) 904 return -1; 905 906 // Okay, we can do this transformation. Return the number of bytes into the 907 // store that the load is. 908 return LoadOffset-StoreOffset; 909} 910 911/// AnalyzeLoadFromClobberingStore - This function is called when we have a 912/// memdep query of a load that ends up being a clobbering store. 913static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 914 StoreInst *DepSI, 915 const TargetData &TD) { 916 // Cannot handle reading from store of first-class aggregate yet. 917 if (DepSI->getValueOperand()->getType()->isStructTy() || 918 DepSI->getValueOperand()->getType()->isArrayTy()) 919 return -1; 920 921 Value *StorePtr = DepSI->getPointerOperand(); 922 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 923 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 924 StorePtr, StoreSize, TD); 925} 926 927/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 928/// memdep query of a load that ends up being clobbered by another load. See if 929/// the other load can feed into the second load. 930static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 931 LoadInst *DepLI, const TargetData &TD){ 932 // Cannot handle reading from store of first-class aggregate yet. 933 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 934 return -1; 935 936 Value *DepPtr = DepLI->getPointerOperand(); 937 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 938 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 939 if (R != -1) return R; 940 941 // If we have a load/load clobber an DepLI can be widened to cover this load, 942 // then we should widen it! 943 int64_t LoadOffs = 0; 944 const Value *LoadBase = 945 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 946 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 947 948 unsigned Size = MemoryDependenceAnalysis:: 949 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 950 if (Size == 0) return -1; 951 952 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 953} 954 955 956 957static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 958 MemIntrinsic *MI, 959 const TargetData &TD) { 960 // If the mem operation is a non-constant size, we can't handle it. 961 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 962 if (SizeCst == 0) return -1; 963 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 964 965 // If this is memset, we just need to see if the offset is valid in the size 966 // of the memset.. 967 if (MI->getIntrinsicID() == Intrinsic::memset) 968 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 969 MemSizeInBits, TD); 970 971 // If we have a memcpy/memmove, the only case we can handle is if this is a 972 // copy from constant memory. In that case, we can read directly from the 973 // constant memory. 974 MemTransferInst *MTI = cast<MemTransferInst>(MI); 975 976 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 977 if (Src == 0) return -1; 978 979 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 980 if (GV == 0 || !GV->isConstant()) return -1; 981 982 // See if the access is within the bounds of the transfer. 983 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 984 MI->getDest(), MemSizeInBits, TD); 985 if (Offset == -1) 986 return Offset; 987 988 // Otherwise, see if we can constant fold a load from the constant with the 989 // offset applied as appropriate. 990 Src = ConstantExpr::getBitCast(Src, 991 llvm::Type::getInt8PtrTy(Src->getContext())); 992 Constant *OffsetCst = 993 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 994 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 995 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 996 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 997 return Offset; 998 return -1; 999} 1000 1001 1002/// GetStoreValueForLoad - This function is called when we have a 1003/// memdep query of a load that ends up being a clobbering store. This means 1004/// that the store provides bits used by the load but we the pointers don't 1005/// mustalias. Check this case to see if there is anything more we can do 1006/// before we give up. 1007static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1008 Type *LoadTy, 1009 Instruction *InsertPt, const TargetData &TD){ 1010 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1011 1012 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1013 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1014 1015 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1016 1017 // Compute which bits of the stored value are being used by the load. Convert 1018 // to an integer type to start with. 1019 if (SrcVal->getType()->isPointerTy()) 1020 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx)); 1021 if (!SrcVal->getType()->isIntegerTy()) 1022 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1023 1024 // Shift the bits to the least significant depending on endianness. 1025 unsigned ShiftAmt; 1026 if (TD.isLittleEndian()) 1027 ShiftAmt = Offset*8; 1028 else 1029 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1030 1031 if (ShiftAmt) 1032 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1033 1034 if (LoadSize != StoreSize) 1035 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1036 1037 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1038} 1039 1040/// GetLoadValueForLoad - This function is called when we have a 1041/// memdep query of a load that ends up being a clobbering load. This means 1042/// that the load *may* provide bits used by the load but we can't be sure 1043/// because the pointers don't mustalias. Check this case to see if there is 1044/// anything more we can do before we give up. 1045static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1046 Type *LoadTy, Instruction *InsertPt, 1047 GVN &gvn) { 1048 const TargetData &TD = *gvn.getTargetData(); 1049 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1050 // widen SrcVal out to a larger load. 1051 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1052 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1053 if (Offset+LoadSize > SrcValSize) { 1054 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1055 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1056 // If we have a load/load clobber an DepLI can be widened to cover this 1057 // load, then we should widen it to the next power of 2 size big enough! 1058 unsigned NewLoadSize = Offset+LoadSize; 1059 if (!isPowerOf2_32(NewLoadSize)) 1060 NewLoadSize = NextPowerOf2(NewLoadSize); 1061 1062 Value *PtrVal = SrcVal->getPointerOperand(); 1063 1064 // Insert the new load after the old load. This ensures that subsequent 1065 // memdep queries will find the new load. We can't easily remove the old 1066 // load completely because it is already in the value numbering table. 1067 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1068 Type *DestPTy = 1069 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1070 DestPTy = PointerType::get(DestPTy, 1071 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1072 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1073 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1074 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1075 NewLoad->takeName(SrcVal); 1076 NewLoad->setAlignment(SrcVal->getAlignment()); 1077 1078 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1079 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1080 1081 // Replace uses of the original load with the wider load. On a big endian 1082 // system, we need to shift down to get the relevant bits. 1083 Value *RV = NewLoad; 1084 if (TD.isBigEndian()) 1085 RV = Builder.CreateLShr(RV, 1086 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1087 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1088 SrcVal->replaceAllUsesWith(RV); 1089 1090 // We would like to use gvn.markInstructionForDeletion here, but we can't 1091 // because the load is already memoized into the leader map table that GVN 1092 // tracks. It is potentially possible to remove the load from the table, 1093 // but then there all of the operations based on it would need to be 1094 // rehashed. Just leave the dead load around. 1095 gvn.getMemDep().removeInstruction(SrcVal); 1096 SrcVal = NewLoad; 1097 } 1098 1099 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1100} 1101 1102 1103/// GetMemInstValueForLoad - This function is called when we have a 1104/// memdep query of a load that ends up being a clobbering mem intrinsic. 1105static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1106 Type *LoadTy, Instruction *InsertPt, 1107 const TargetData &TD){ 1108 LLVMContext &Ctx = LoadTy->getContext(); 1109 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1110 1111 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1112 1113 // We know that this method is only called when the mem transfer fully 1114 // provides the bits for the load. 1115 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1116 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1117 // independently of what the offset is. 1118 Value *Val = MSI->getValue(); 1119 if (LoadSize != 1) 1120 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1121 1122 Value *OneElt = Val; 1123 1124 // Splat the value out to the right number of bits. 1125 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1126 // If we can double the number of bytes set, do it. 1127 if (NumBytesSet*2 <= LoadSize) { 1128 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1129 Val = Builder.CreateOr(Val, ShVal); 1130 NumBytesSet <<= 1; 1131 continue; 1132 } 1133 1134 // Otherwise insert one byte at a time. 1135 Value *ShVal = Builder.CreateShl(Val, 1*8); 1136 Val = Builder.CreateOr(OneElt, ShVal); 1137 ++NumBytesSet; 1138 } 1139 1140 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1141 } 1142 1143 // Otherwise, this is a memcpy/memmove from a constant global. 1144 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1145 Constant *Src = cast<Constant>(MTI->getSource()); 1146 1147 // Otherwise, see if we can constant fold a load from the constant with the 1148 // offset applied as appropriate. 1149 Src = ConstantExpr::getBitCast(Src, 1150 llvm::Type::getInt8PtrTy(Src->getContext())); 1151 Constant *OffsetCst = 1152 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1153 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1154 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1155 return ConstantFoldLoadFromConstPtr(Src, &TD); 1156} 1157 1158namespace { 1159 1160struct AvailableValueInBlock { 1161 /// BB - The basic block in question. 1162 BasicBlock *BB; 1163 enum ValType { 1164 SimpleVal, // A simple offsetted value that is accessed. 1165 LoadVal, // A value produced by a load. 1166 MemIntrin // A memory intrinsic which is loaded from. 1167 }; 1168 1169 /// V - The value that is live out of the block. 1170 PointerIntPair<Value *, 2, ValType> Val; 1171 1172 /// Offset - The byte offset in Val that is interesting for the load query. 1173 unsigned Offset; 1174 1175 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1176 unsigned Offset = 0) { 1177 AvailableValueInBlock Res; 1178 Res.BB = BB; 1179 Res.Val.setPointer(V); 1180 Res.Val.setInt(SimpleVal); 1181 Res.Offset = Offset; 1182 return Res; 1183 } 1184 1185 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1186 unsigned Offset = 0) { 1187 AvailableValueInBlock Res; 1188 Res.BB = BB; 1189 Res.Val.setPointer(MI); 1190 Res.Val.setInt(MemIntrin); 1191 Res.Offset = Offset; 1192 return Res; 1193 } 1194 1195 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1196 unsigned Offset = 0) { 1197 AvailableValueInBlock Res; 1198 Res.BB = BB; 1199 Res.Val.setPointer(LI); 1200 Res.Val.setInt(LoadVal); 1201 Res.Offset = Offset; 1202 return Res; 1203 } 1204 1205 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1206 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1207 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1208 1209 Value *getSimpleValue() const { 1210 assert(isSimpleValue() && "Wrong accessor"); 1211 return Val.getPointer(); 1212 } 1213 1214 LoadInst *getCoercedLoadValue() const { 1215 assert(isCoercedLoadValue() && "Wrong accessor"); 1216 return cast<LoadInst>(Val.getPointer()); 1217 } 1218 1219 MemIntrinsic *getMemIntrinValue() const { 1220 assert(isMemIntrinValue() && "Wrong accessor"); 1221 return cast<MemIntrinsic>(Val.getPointer()); 1222 } 1223 1224 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1225 /// defined here to the specified type. This handles various coercion cases. 1226 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1227 Value *Res; 1228 if (isSimpleValue()) { 1229 Res = getSimpleValue(); 1230 if (Res->getType() != LoadTy) { 1231 const TargetData *TD = gvn.getTargetData(); 1232 assert(TD && "Need target data to handle type mismatch case"); 1233 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1234 *TD); 1235 1236 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1237 << *getSimpleValue() << '\n' 1238 << *Res << '\n' << "\n\n\n"); 1239 } 1240 } else if (isCoercedLoadValue()) { 1241 LoadInst *Load = getCoercedLoadValue(); 1242 if (Load->getType() == LoadTy && Offset == 0) { 1243 Res = Load; 1244 } else { 1245 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1246 gvn); 1247 1248 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1249 << *getCoercedLoadValue() << '\n' 1250 << *Res << '\n' << "\n\n\n"); 1251 } 1252 } else { 1253 const TargetData *TD = gvn.getTargetData(); 1254 assert(TD && "Need target data to handle type mismatch case"); 1255 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1256 LoadTy, BB->getTerminator(), *TD); 1257 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1258 << " " << *getMemIntrinValue() << '\n' 1259 << *Res << '\n' << "\n\n\n"); 1260 } 1261 return Res; 1262 } 1263}; 1264 1265} // end anonymous namespace 1266 1267/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1268/// construct SSA form, allowing us to eliminate LI. This returns the value 1269/// that should be used at LI's definition site. 1270static Value *ConstructSSAForLoadSet(LoadInst *LI, 1271 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1272 GVN &gvn) { 1273 // Check for the fully redundant, dominating load case. In this case, we can 1274 // just use the dominating value directly. 1275 if (ValuesPerBlock.size() == 1 && 1276 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1277 LI->getParent())) 1278 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1279 1280 // Otherwise, we have to construct SSA form. 1281 SmallVector<PHINode*, 8> NewPHIs; 1282 SSAUpdater SSAUpdate(&NewPHIs); 1283 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1284 1285 Type *LoadTy = LI->getType(); 1286 1287 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1288 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1289 BasicBlock *BB = AV.BB; 1290 1291 if (SSAUpdate.HasValueForBlock(BB)) 1292 continue; 1293 1294 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1295 } 1296 1297 // Perform PHI construction. 1298 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1299 1300 // If new PHI nodes were created, notify alias analysis. 1301 if (V->getType()->isPointerTy()) { 1302 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1303 1304 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1305 AA->copyValue(LI, NewPHIs[i]); 1306 1307 // Now that we've copied information to the new PHIs, scan through 1308 // them again and inform alias analysis that we've added potentially 1309 // escaping uses to any values that are operands to these PHIs. 1310 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1311 PHINode *P = NewPHIs[i]; 1312 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1313 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1314 AA->addEscapingUse(P->getOperandUse(jj)); 1315 } 1316 } 1317 } 1318 1319 return V; 1320} 1321 1322static bool isLifetimeStart(const Instruction *Inst) { 1323 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1324 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1325 return false; 1326} 1327 1328/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1329/// non-local by performing PHI construction. 1330bool GVN::processNonLocalLoad(LoadInst *LI) { 1331 // Find the non-local dependencies of the load. 1332 SmallVector<NonLocalDepResult, 64> Deps; 1333 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1334 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1335 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1336 // << Deps.size() << *LI << '\n'); 1337 1338 // If we had to process more than one hundred blocks to find the 1339 // dependencies, this load isn't worth worrying about. Optimizing 1340 // it will be too expensive. 1341 unsigned NumDeps = Deps.size(); 1342 if (NumDeps > 100) 1343 return false; 1344 1345 // If we had a phi translation failure, we'll have a single entry which is a 1346 // clobber in the current block. Reject this early. 1347 if (NumDeps == 1 && 1348 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1349 DEBUG( 1350 dbgs() << "GVN: non-local load "; 1351 WriteAsOperand(dbgs(), LI); 1352 dbgs() << " has unknown dependencies\n"; 1353 ); 1354 return false; 1355 } 1356 1357 // Filter out useless results (non-locals, etc). Keep track of the blocks 1358 // where we have a value available in repl, also keep track of whether we see 1359 // dependencies that produce an unknown value for the load (such as a call 1360 // that could potentially clobber the load). 1361 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock; 1362 SmallVector<BasicBlock*, 64> UnavailableBlocks; 1363 1364 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1365 BasicBlock *DepBB = Deps[i].getBB(); 1366 MemDepResult DepInfo = Deps[i].getResult(); 1367 1368 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1369 UnavailableBlocks.push_back(DepBB); 1370 continue; 1371 } 1372 1373 if (DepInfo.isClobber()) { 1374 // The address being loaded in this non-local block may not be the same as 1375 // the pointer operand of the load if PHI translation occurs. Make sure 1376 // to consider the right address. 1377 Value *Address = Deps[i].getAddress(); 1378 1379 // If the dependence is to a store that writes to a superset of the bits 1380 // read by the load, we can extract the bits we need for the load from the 1381 // stored value. 1382 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1383 if (TD && Address) { 1384 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1385 DepSI, *TD); 1386 if (Offset != -1) { 1387 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1388 DepSI->getValueOperand(), 1389 Offset)); 1390 continue; 1391 } 1392 } 1393 } 1394 1395 // Check to see if we have something like this: 1396 // load i32* P 1397 // load i8* (P+1) 1398 // if we have this, replace the later with an extraction from the former. 1399 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1400 // If this is a clobber and L is the first instruction in its block, then 1401 // we have the first instruction in the entry block. 1402 if (DepLI != LI && Address && TD) { 1403 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1404 LI->getPointerOperand(), 1405 DepLI, *TD); 1406 1407 if (Offset != -1) { 1408 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1409 Offset)); 1410 continue; 1411 } 1412 } 1413 } 1414 1415 // If the clobbering value is a memset/memcpy/memmove, see if we can 1416 // forward a value on from it. 1417 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1418 if (TD && Address) { 1419 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1420 DepMI, *TD); 1421 if (Offset != -1) { 1422 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1423 Offset)); 1424 continue; 1425 } 1426 } 1427 } 1428 1429 UnavailableBlocks.push_back(DepBB); 1430 continue; 1431 } 1432 1433 // DepInfo.isDef() here 1434 1435 Instruction *DepInst = DepInfo.getInst(); 1436 1437 // Loading the allocation -> undef. 1438 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1439 // Loading immediately after lifetime begin -> undef. 1440 isLifetimeStart(DepInst)) { 1441 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1442 UndefValue::get(LI->getType()))); 1443 continue; 1444 } 1445 1446 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1447 // Reject loads and stores that are to the same address but are of 1448 // different types if we have to. 1449 if (S->getValueOperand()->getType() != LI->getType()) { 1450 // If the stored value is larger or equal to the loaded value, we can 1451 // reuse it. 1452 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1453 LI->getType(), *TD)) { 1454 UnavailableBlocks.push_back(DepBB); 1455 continue; 1456 } 1457 } 1458 1459 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1460 S->getValueOperand())); 1461 continue; 1462 } 1463 1464 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1465 // If the types mismatch and we can't handle it, reject reuse of the load. 1466 if (LD->getType() != LI->getType()) { 1467 // If the stored value is larger or equal to the loaded value, we can 1468 // reuse it. 1469 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1470 UnavailableBlocks.push_back(DepBB); 1471 continue; 1472 } 1473 } 1474 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1475 continue; 1476 } 1477 1478 UnavailableBlocks.push_back(DepBB); 1479 continue; 1480 } 1481 1482 // If we have no predecessors that produce a known value for this load, exit 1483 // early. 1484 if (ValuesPerBlock.empty()) return false; 1485 1486 // If all of the instructions we depend on produce a known value for this 1487 // load, then it is fully redundant and we can use PHI insertion to compute 1488 // its value. Insert PHIs and remove the fully redundant value now. 1489 if (UnavailableBlocks.empty()) { 1490 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1491 1492 // Perform PHI construction. 1493 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1494 LI->replaceAllUsesWith(V); 1495 1496 if (isa<PHINode>(V)) 1497 V->takeName(LI); 1498 if (V->getType()->isPointerTy()) 1499 MD->invalidateCachedPointerInfo(V); 1500 markInstructionForDeletion(LI); 1501 ++NumGVNLoad; 1502 return true; 1503 } 1504 1505 if (!EnablePRE || !EnableLoadPRE) 1506 return false; 1507 1508 // Okay, we have *some* definitions of the value. This means that the value 1509 // is available in some of our (transitive) predecessors. Lets think about 1510 // doing PRE of this load. This will involve inserting a new load into the 1511 // predecessor when it's not available. We could do this in general, but 1512 // prefer to not increase code size. As such, we only do this when we know 1513 // that we only have to insert *one* load (which means we're basically moving 1514 // the load, not inserting a new one). 1515 1516 SmallPtrSet<BasicBlock *, 4> Blockers; 1517 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1518 Blockers.insert(UnavailableBlocks[i]); 1519 1520 // Let's find the first basic block with more than one predecessor. Walk 1521 // backwards through predecessors if needed. 1522 BasicBlock *LoadBB = LI->getParent(); 1523 BasicBlock *TmpBB = LoadBB; 1524 1525 bool isSinglePred = false; 1526 bool allSingleSucc = true; 1527 while (TmpBB->getSinglePredecessor()) { 1528 isSinglePred = true; 1529 TmpBB = TmpBB->getSinglePredecessor(); 1530 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1531 return false; 1532 if (Blockers.count(TmpBB)) 1533 return false; 1534 1535 // If any of these blocks has more than one successor (i.e. if the edge we 1536 // just traversed was critical), then there are other paths through this 1537 // block along which the load may not be anticipated. Hoisting the load 1538 // above this block would be adding the load to execution paths along 1539 // which it was not previously executed. 1540 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1541 return false; 1542 } 1543 1544 assert(TmpBB); 1545 LoadBB = TmpBB; 1546 1547 // FIXME: It is extremely unclear what this loop is doing, other than 1548 // artificially restricting loadpre. 1549 if (isSinglePred) { 1550 bool isHot = false; 1551 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1552 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1553 if (AV.isSimpleValue()) 1554 // "Hot" Instruction is in some loop (because it dominates its dep. 1555 // instruction). 1556 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1557 if (DT->dominates(LI, I)) { 1558 isHot = true; 1559 break; 1560 } 1561 } 1562 1563 // We are interested only in "hot" instructions. We don't want to do any 1564 // mis-optimizations here. 1565 if (!isHot) 1566 return false; 1567 } 1568 1569 // Check to see how many predecessors have the loaded value fully 1570 // available. 1571 DenseMap<BasicBlock*, Value*> PredLoads; 1572 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1573 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1574 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1575 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1576 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1577 1578 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1579 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1580 PI != E; ++PI) { 1581 BasicBlock *Pred = *PI; 1582 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1583 continue; 1584 } 1585 PredLoads[Pred] = 0; 1586 1587 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1588 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1589 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1590 << Pred->getName() << "': " << *LI << '\n'); 1591 return false; 1592 } 1593 1594 if (LoadBB->isLandingPad()) { 1595 DEBUG(dbgs() 1596 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1597 << Pred->getName() << "': " << *LI << '\n'); 1598 return false; 1599 } 1600 1601 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1602 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1603 } 1604 } 1605 1606 if (!NeedToSplit.empty()) { 1607 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1608 return false; 1609 } 1610 1611 // Decide whether PRE is profitable for this load. 1612 unsigned NumUnavailablePreds = PredLoads.size(); 1613 assert(NumUnavailablePreds != 0 && 1614 "Fully available value should be eliminated above!"); 1615 1616 // If this load is unavailable in multiple predecessors, reject it. 1617 // FIXME: If we could restructure the CFG, we could make a common pred with 1618 // all the preds that don't have an available LI and insert a new load into 1619 // that one block. 1620 if (NumUnavailablePreds != 1) 1621 return false; 1622 1623 // Check if the load can safely be moved to all the unavailable predecessors. 1624 bool CanDoPRE = true; 1625 SmallVector<Instruction*, 8> NewInsts; 1626 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1627 E = PredLoads.end(); I != E; ++I) { 1628 BasicBlock *UnavailablePred = I->first; 1629 1630 // Do PHI translation to get its value in the predecessor if necessary. The 1631 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1632 1633 // If all preds have a single successor, then we know it is safe to insert 1634 // the load on the pred (?!?), so we can insert code to materialize the 1635 // pointer if it is not available. 1636 PHITransAddr Address(LI->getPointerOperand(), TD); 1637 Value *LoadPtr = 0; 1638 if (allSingleSucc) { 1639 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1640 *DT, NewInsts); 1641 } else { 1642 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1643 LoadPtr = Address.getAddr(); 1644 } 1645 1646 // If we couldn't find or insert a computation of this phi translated value, 1647 // we fail PRE. 1648 if (LoadPtr == 0) { 1649 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1650 << *LI->getPointerOperand() << "\n"); 1651 CanDoPRE = false; 1652 break; 1653 } 1654 1655 // Make sure it is valid to move this load here. We have to watch out for: 1656 // @1 = getelementptr (i8* p, ... 1657 // test p and branch if == 0 1658 // load @1 1659 // It is valid to have the getelementptr before the test, even if p can 1660 // be 0, as getelementptr only does address arithmetic. 1661 // If we are not pushing the value through any multiple-successor blocks 1662 // we do not have this case. Otherwise, check that the load is safe to 1663 // put anywhere; this can be improved, but should be conservatively safe. 1664 if (!allSingleSucc && 1665 // FIXME: REEVALUTE THIS. 1666 !isSafeToLoadUnconditionally(LoadPtr, 1667 UnavailablePred->getTerminator(), 1668 LI->getAlignment(), TD)) { 1669 CanDoPRE = false; 1670 break; 1671 } 1672 1673 I->second = LoadPtr; 1674 } 1675 1676 if (!CanDoPRE) { 1677 while (!NewInsts.empty()) { 1678 Instruction *I = NewInsts.pop_back_val(); 1679 if (MD) MD->removeInstruction(I); 1680 I->eraseFromParent(); 1681 } 1682 return false; 1683 } 1684 1685 // Okay, we can eliminate this load by inserting a reload in the predecessor 1686 // and using PHI construction to get the value in the other predecessors, do 1687 // it. 1688 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1689 DEBUG(if (!NewInsts.empty()) 1690 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1691 << *NewInsts.back() << '\n'); 1692 1693 // Assign value numbers to the new instructions. 1694 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1695 // FIXME: We really _ought_ to insert these value numbers into their 1696 // parent's availability map. However, in doing so, we risk getting into 1697 // ordering issues. If a block hasn't been processed yet, we would be 1698 // marking a value as AVAIL-IN, which isn't what we intend. 1699 VN.lookup_or_add(NewInsts[i]); 1700 } 1701 1702 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1703 E = PredLoads.end(); I != E; ++I) { 1704 BasicBlock *UnavailablePred = I->first; 1705 Value *LoadPtr = I->second; 1706 1707 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1708 LI->getAlignment(), 1709 UnavailablePred->getTerminator()); 1710 1711 // Transfer the old load's TBAA tag to the new load. 1712 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1713 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1714 1715 // Transfer DebugLoc. 1716 NewLoad->setDebugLoc(LI->getDebugLoc()); 1717 1718 // Add the newly created load. 1719 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1720 NewLoad)); 1721 MD->invalidateCachedPointerInfo(LoadPtr); 1722 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1723 } 1724 1725 // Perform PHI construction. 1726 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1727 LI->replaceAllUsesWith(V); 1728 if (isa<PHINode>(V)) 1729 V->takeName(LI); 1730 if (V->getType()->isPointerTy()) 1731 MD->invalidateCachedPointerInfo(V); 1732 markInstructionForDeletion(LI); 1733 ++NumPRELoad; 1734 return true; 1735} 1736 1737/// processLoad - Attempt to eliminate a load, first by eliminating it 1738/// locally, and then attempting non-local elimination if that fails. 1739bool GVN::processLoad(LoadInst *L) { 1740 if (!MD) 1741 return false; 1742 1743 if (!L->isSimple()) 1744 return false; 1745 1746 if (L->use_empty()) { 1747 markInstructionForDeletion(L); 1748 return true; 1749 } 1750 1751 // ... to a pointer that has been loaded from before... 1752 MemDepResult Dep = MD->getDependency(L); 1753 1754 // If we have a clobber and target data is around, see if this is a clobber 1755 // that we can fix up through code synthesis. 1756 if (Dep.isClobber() && TD) { 1757 // Check to see if we have something like this: 1758 // store i32 123, i32* %P 1759 // %A = bitcast i32* %P to i8* 1760 // %B = gep i8* %A, i32 1 1761 // %C = load i8* %B 1762 // 1763 // We could do that by recognizing if the clobber instructions are obviously 1764 // a common base + constant offset, and if the previous store (or memset) 1765 // completely covers this load. This sort of thing can happen in bitfield 1766 // access code. 1767 Value *AvailVal = 0; 1768 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1769 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1770 L->getPointerOperand(), 1771 DepSI, *TD); 1772 if (Offset != -1) 1773 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1774 L->getType(), L, *TD); 1775 } 1776 1777 // Check to see if we have something like this: 1778 // load i32* P 1779 // load i8* (P+1) 1780 // if we have this, replace the later with an extraction from the former. 1781 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1782 // If this is a clobber and L is the first instruction in its block, then 1783 // we have the first instruction in the entry block. 1784 if (DepLI == L) 1785 return false; 1786 1787 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1788 L->getPointerOperand(), 1789 DepLI, *TD); 1790 if (Offset != -1) 1791 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1792 } 1793 1794 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1795 // a value on from it. 1796 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1797 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1798 L->getPointerOperand(), 1799 DepMI, *TD); 1800 if (Offset != -1) 1801 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1802 } 1803 1804 if (AvailVal) { 1805 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1806 << *AvailVal << '\n' << *L << "\n\n\n"); 1807 1808 // Replace the load! 1809 L->replaceAllUsesWith(AvailVal); 1810 if (AvailVal->getType()->isPointerTy()) 1811 MD->invalidateCachedPointerInfo(AvailVal); 1812 markInstructionForDeletion(L); 1813 ++NumGVNLoad; 1814 return true; 1815 } 1816 } 1817 1818 // If the value isn't available, don't do anything! 1819 if (Dep.isClobber()) { 1820 DEBUG( 1821 // fast print dep, using operator<< on instruction is too slow. 1822 dbgs() << "GVN: load "; 1823 WriteAsOperand(dbgs(), L); 1824 Instruction *I = Dep.getInst(); 1825 dbgs() << " is clobbered by " << *I << '\n'; 1826 ); 1827 return false; 1828 } 1829 1830 // If it is defined in another block, try harder. 1831 if (Dep.isNonLocal()) 1832 return processNonLocalLoad(L); 1833 1834 if (!Dep.isDef()) { 1835 DEBUG( 1836 // fast print dep, using operator<< on instruction is too slow. 1837 dbgs() << "GVN: load "; 1838 WriteAsOperand(dbgs(), L); 1839 dbgs() << " has unknown dependence\n"; 1840 ); 1841 return false; 1842 } 1843 1844 Instruction *DepInst = Dep.getInst(); 1845 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1846 Value *StoredVal = DepSI->getValueOperand(); 1847 1848 // The store and load are to a must-aliased pointer, but they may not 1849 // actually have the same type. See if we know how to reuse the stored 1850 // value (depending on its type). 1851 if (StoredVal->getType() != L->getType()) { 1852 if (TD) { 1853 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1854 L, *TD); 1855 if (StoredVal == 0) 1856 return false; 1857 1858 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1859 << '\n' << *L << "\n\n\n"); 1860 } 1861 else 1862 return false; 1863 } 1864 1865 // Remove it! 1866 L->replaceAllUsesWith(StoredVal); 1867 if (StoredVal->getType()->isPointerTy()) 1868 MD->invalidateCachedPointerInfo(StoredVal); 1869 markInstructionForDeletion(L); 1870 ++NumGVNLoad; 1871 return true; 1872 } 1873 1874 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1875 Value *AvailableVal = DepLI; 1876 1877 // The loads are of a must-aliased pointer, but they may not actually have 1878 // the same type. See if we know how to reuse the previously loaded value 1879 // (depending on its type). 1880 if (DepLI->getType() != L->getType()) { 1881 if (TD) { 1882 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1883 L, *TD); 1884 if (AvailableVal == 0) 1885 return false; 1886 1887 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1888 << "\n" << *L << "\n\n\n"); 1889 } 1890 else 1891 return false; 1892 } 1893 1894 // Remove it! 1895 L->replaceAllUsesWith(AvailableVal); 1896 if (DepLI->getType()->isPointerTy()) 1897 MD->invalidateCachedPointerInfo(DepLI); 1898 markInstructionForDeletion(L); 1899 ++NumGVNLoad; 1900 return true; 1901 } 1902 1903 // If this load really doesn't depend on anything, then we must be loading an 1904 // undef value. This can happen when loading for a fresh allocation with no 1905 // intervening stores, for example. 1906 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1907 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1908 markInstructionForDeletion(L); 1909 ++NumGVNLoad; 1910 return true; 1911 } 1912 1913 // If this load occurs either right after a lifetime begin, 1914 // then the loaded value is undefined. 1915 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1916 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1917 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1918 markInstructionForDeletion(L); 1919 ++NumGVNLoad; 1920 return true; 1921 } 1922 } 1923 1924 return false; 1925} 1926 1927// findLeader - In order to find a leader for a given value number at a 1928// specific basic block, we first obtain the list of all Values for that number, 1929// and then scan the list to find one whose block dominates the block in 1930// question. This is fast because dominator tree queries consist of only 1931// a few comparisons of DFS numbers. 1932Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1933 LeaderTableEntry Vals = LeaderTable[num]; 1934 if (!Vals.Val) return 0; 1935 1936 Value *Val = 0; 1937 if (DT->dominates(Vals.BB, BB)) { 1938 Val = Vals.Val; 1939 if (isa<Constant>(Val)) return Val; 1940 } 1941 1942 LeaderTableEntry* Next = Vals.Next; 1943 while (Next) { 1944 if (DT->dominates(Next->BB, BB)) { 1945 if (isa<Constant>(Next->Val)) return Next->Val; 1946 if (!Val) Val = Next->Val; 1947 } 1948 1949 Next = Next->Next; 1950 } 1951 1952 return Val; 1953} 1954 1955/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1956/// use is dominated by the given basic block. Returns the number of uses that 1957/// were replaced. 1958unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1959 BasicBlock *Root) { 1960 unsigned Count = 0; 1961 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1962 UI != UE; ) { 1963 Use &U = (UI++).getUse(); 1964 1965 // If From occurs as a phi node operand then the use implicitly lives in the 1966 // corresponding incoming block. Otherwise it is the block containing the 1967 // user that must be dominated by Root. 1968 BasicBlock *UsingBlock; 1969 if (PHINode *PN = dyn_cast<PHINode>(U.getUser())) 1970 UsingBlock = PN->getIncomingBlock(U); 1971 else 1972 UsingBlock = cast<Instruction>(U.getUser())->getParent(); 1973 1974 if (DT->dominates(Root, UsingBlock)) { 1975 U.set(To); 1976 ++Count; 1977 } 1978 } 1979 return Count; 1980} 1981 1982/// propagateEquality - The given values are known to be equal in every block 1983/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 1984/// 'RHS' everywhere in the scope. Returns whether a change was made. 1985bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) { 1986 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 1987 Worklist.push_back(std::make_pair(LHS, RHS)); 1988 bool Changed = false; 1989 1990 while (!Worklist.empty()) { 1991 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 1992 LHS = Item.first; RHS = Item.second; 1993 1994 if (LHS == RHS) continue; 1995 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 1996 1997 // Don't try to propagate equalities between constants. 1998 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 1999 2000 // Prefer a constant on the right-hand side, or an Argument if no constants. 2001 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2002 std::swap(LHS, RHS); 2003 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2004 2005 // If there is no obvious reason to prefer the left-hand side over the right- 2006 // hand side, ensure the longest lived term is on the right-hand side, so the 2007 // shortest lived term will be replaced by the longest lived. This tends to 2008 // expose more simplifications. 2009 uint32_t LVN = VN.lookup_or_add(LHS); 2010 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2011 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2012 // Move the 'oldest' value to the right-hand side, using the value number as 2013 // a proxy for age. 2014 uint32_t RVN = VN.lookup_or_add(RHS); 2015 if (LVN < RVN) { 2016 std::swap(LHS, RHS); 2017 LVN = RVN; 2018 } 2019 } 2020 assert((!isa<Instruction>(RHS) || 2021 DT->properlyDominates(cast<Instruction>(RHS)->getParent(), Root)) && 2022 "Instruction doesn't dominate scope!"); 2023 2024 // If value numbering later deduces that an instruction in the scope is equal 2025 // to 'LHS' then ensure it will be turned into 'RHS'. 2026 addToLeaderTable(LVN, RHS, Root); 2027 2028 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2029 // LHS always has at least one use that is not dominated by Root, this will 2030 // never do anything if LHS has only one use. 2031 if (!LHS->hasOneUse()) { 2032 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2033 Changed |= NumReplacements > 0; 2034 NumGVNEqProp += NumReplacements; 2035 } 2036 2037 // Now try to deduce additional equalities from this one. For example, if the 2038 // known equality was "(A != B)" == "false" then it follows that A and B are 2039 // equal in the scope. Only boolean equalities with an explicit true or false 2040 // RHS are currently supported. 2041 if (!RHS->getType()->isIntegerTy(1)) 2042 // Not a boolean equality - bail out. 2043 continue; 2044 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2045 if (!CI) 2046 // RHS neither 'true' nor 'false' - bail out. 2047 continue; 2048 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2049 bool isKnownTrue = CI->isAllOnesValue(); 2050 bool isKnownFalse = !isKnownTrue; 2051 2052 // If "A && B" is known true then both A and B are known true. If "A || B" 2053 // is known false then both A and B are known false. 2054 Value *A, *B; 2055 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2056 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2057 Worklist.push_back(std::make_pair(A, RHS)); 2058 Worklist.push_back(std::make_pair(B, RHS)); 2059 continue; 2060 } 2061 2062 // If we are propagating an equality like "(A == B)" == "true" then also 2063 // propagate the equality A == B. When propagating a comparison such as 2064 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2065 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2066 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2067 2068 // If "A == B" is known true, or "A != B" is known false, then replace 2069 // A with B everywhere in the scope. 2070 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2071 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2072 Worklist.push_back(std::make_pair(Op0, Op1)); 2073 2074 // If "A >= B" is known true, replace "A < B" with false everywhere. 2075 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2076 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2077 // Since we don't have the instruction "A < B" immediately to hand, work out 2078 // the value number that it would have and use that to find an appropriate 2079 // instruction (if any). 2080 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2081 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2082 // If the number we were assigned was brand new then there is no point in 2083 // looking for an instruction realizing it: there cannot be one! 2084 if (Num < NextNum) { 2085 Value *NotCmp = findLeader(Root, Num); 2086 if (NotCmp && isa<Instruction>(NotCmp)) { 2087 unsigned NumReplacements = 2088 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2089 Changed |= NumReplacements > 0; 2090 NumGVNEqProp += NumReplacements; 2091 } 2092 } 2093 // Ensure that any instruction in scope that gets the "A < B" value number 2094 // is replaced with false. 2095 addToLeaderTable(Num, NotVal, Root); 2096 2097 continue; 2098 } 2099 } 2100 2101 return Changed; 2102} 2103 2104/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2105/// true if every path from the entry block to 'Dst' passes via this edge. In 2106/// particular 'Dst' must not be reachable via another edge from 'Src'. 2107static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst, 2108 DominatorTree *DT) { 2109 // While in theory it is interesting to consider the case in which Dst has 2110 // more than one predecessor, because Dst might be part of a loop which is 2111 // only reachable from Src, in practice it is pointless since at the time 2112 // GVN runs all such loops have preheaders, which means that Dst will have 2113 // been changed to have only one predecessor, namely Src. 2114 BasicBlock *Pred = Dst->getSinglePredecessor(); 2115 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2116 (void)Src; 2117 return Pred != 0; 2118} 2119 2120/// processInstruction - When calculating availability, handle an instruction 2121/// by inserting it into the appropriate sets 2122bool GVN::processInstruction(Instruction *I) { 2123 // Ignore dbg info intrinsics. 2124 if (isa<DbgInfoIntrinsic>(I)) 2125 return false; 2126 2127 // If the instruction can be easily simplified then do so now in preference 2128 // to value numbering it. Value numbering often exposes redundancies, for 2129 // example if it determines that %y is equal to %x then the instruction 2130 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2131 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2132 I->replaceAllUsesWith(V); 2133 if (MD && V->getType()->isPointerTy()) 2134 MD->invalidateCachedPointerInfo(V); 2135 markInstructionForDeletion(I); 2136 ++NumGVNSimpl; 2137 return true; 2138 } 2139 2140 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2141 if (processLoad(LI)) 2142 return true; 2143 2144 unsigned Num = VN.lookup_or_add(LI); 2145 addToLeaderTable(Num, LI, LI->getParent()); 2146 return false; 2147 } 2148 2149 // For conditional branches, we can perform simple conditional propagation on 2150 // the condition value itself. 2151 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2152 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2153 return false; 2154 2155 Value *BranchCond = BI->getCondition(); 2156 2157 BasicBlock *TrueSucc = BI->getSuccessor(0); 2158 BasicBlock *FalseSucc = BI->getSuccessor(1); 2159 BasicBlock *Parent = BI->getParent(); 2160 bool Changed = false; 2161 2162 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT)) 2163 Changed |= propagateEquality(BranchCond, 2164 ConstantInt::getTrue(TrueSucc->getContext()), 2165 TrueSucc); 2166 2167 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT)) 2168 Changed |= propagateEquality(BranchCond, 2169 ConstantInt::getFalse(FalseSucc->getContext()), 2170 FalseSucc); 2171 2172 return Changed; 2173 } 2174 2175 // For switches, propagate the case values into the case destinations. 2176 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2177 Value *SwitchCond = SI->getCondition(); 2178 BasicBlock *Parent = SI->getParent(); 2179 bool Changed = false; 2180 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2181 i != e; ++i) { 2182 BasicBlock *Dst = i.getCaseSuccessor(); 2183 if (isOnlyReachableViaThisEdge(Parent, Dst, DT)) 2184 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), Dst); 2185 } 2186 return Changed; 2187 } 2188 2189 // Instructions with void type don't return a value, so there's 2190 // no point in trying to find redundancies in them. 2191 if (I->getType()->isVoidTy()) return false; 2192 2193 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2194 unsigned Num = VN.lookup_or_add(I); 2195 2196 // Allocations are always uniquely numbered, so we can save time and memory 2197 // by fast failing them. 2198 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2199 addToLeaderTable(Num, I, I->getParent()); 2200 return false; 2201 } 2202 2203 // If the number we were assigned was a brand new VN, then we don't 2204 // need to do a lookup to see if the number already exists 2205 // somewhere in the domtree: it can't! 2206 if (Num >= NextNum) { 2207 addToLeaderTable(Num, I, I->getParent()); 2208 return false; 2209 } 2210 2211 // Perform fast-path value-number based elimination of values inherited from 2212 // dominators. 2213 Value *repl = findLeader(I->getParent(), Num); 2214 if (repl == 0) { 2215 // Failure, just remember this instance for future use. 2216 addToLeaderTable(Num, I, I->getParent()); 2217 return false; 2218 } 2219 2220 // Remove it! 2221 I->replaceAllUsesWith(repl); 2222 if (MD && repl->getType()->isPointerTy()) 2223 MD->invalidateCachedPointerInfo(repl); 2224 markInstructionForDeletion(I); 2225 return true; 2226} 2227 2228/// runOnFunction - This is the main transformation entry point for a function. 2229bool GVN::runOnFunction(Function& F) { 2230 if (!NoLoads) 2231 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2232 DT = &getAnalysis<DominatorTree>(); 2233 TD = getAnalysisIfAvailable<TargetData>(); 2234 TLI = &getAnalysis<TargetLibraryInfo>(); 2235 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2236 VN.setMemDep(MD); 2237 VN.setDomTree(DT); 2238 2239 bool Changed = false; 2240 bool ShouldContinue = true; 2241 2242 // Merge unconditional branches, allowing PRE to catch more 2243 // optimization opportunities. 2244 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2245 BasicBlock *BB = FI++; 2246 2247 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2248 if (removedBlock) ++NumGVNBlocks; 2249 2250 Changed |= removedBlock; 2251 } 2252 2253 unsigned Iteration = 0; 2254 while (ShouldContinue) { 2255 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2256 ShouldContinue = iterateOnFunction(F); 2257 if (splitCriticalEdges()) 2258 ShouldContinue = true; 2259 Changed |= ShouldContinue; 2260 ++Iteration; 2261 } 2262 2263 if (EnablePRE) { 2264 bool PREChanged = true; 2265 while (PREChanged) { 2266 PREChanged = performPRE(F); 2267 Changed |= PREChanged; 2268 } 2269 } 2270 // FIXME: Should perform GVN again after PRE does something. PRE can move 2271 // computations into blocks where they become fully redundant. Note that 2272 // we can't do this until PRE's critical edge splitting updates memdep. 2273 // Actually, when this happens, we should just fully integrate PRE into GVN. 2274 2275 cleanupGlobalSets(); 2276 2277 return Changed; 2278} 2279 2280 2281bool GVN::processBlock(BasicBlock *BB) { 2282 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2283 // (and incrementing BI before processing an instruction). 2284 assert(InstrsToErase.empty() && 2285 "We expect InstrsToErase to be empty across iterations"); 2286 bool ChangedFunction = false; 2287 2288 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2289 BI != BE;) { 2290 ChangedFunction |= processInstruction(BI); 2291 if (InstrsToErase.empty()) { 2292 ++BI; 2293 continue; 2294 } 2295 2296 // If we need some instructions deleted, do it now. 2297 NumGVNInstr += InstrsToErase.size(); 2298 2299 // Avoid iterator invalidation. 2300 bool AtStart = BI == BB->begin(); 2301 if (!AtStart) 2302 --BI; 2303 2304 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2305 E = InstrsToErase.end(); I != E; ++I) { 2306 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2307 if (MD) MD->removeInstruction(*I); 2308 (*I)->eraseFromParent(); 2309 DEBUG(verifyRemoved(*I)); 2310 } 2311 InstrsToErase.clear(); 2312 2313 if (AtStart) 2314 BI = BB->begin(); 2315 else 2316 ++BI; 2317 } 2318 2319 return ChangedFunction; 2320} 2321 2322/// performPRE - Perform a purely local form of PRE that looks for diamond 2323/// control flow patterns and attempts to perform simple PRE at the join point. 2324bool GVN::performPRE(Function &F) { 2325 bool Changed = false; 2326 DenseMap<BasicBlock*, Value*> predMap; 2327 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2328 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2329 BasicBlock *CurrentBlock = *DI; 2330 2331 // Nothing to PRE in the entry block. 2332 if (CurrentBlock == &F.getEntryBlock()) continue; 2333 2334 // Don't perform PRE on a landing pad. 2335 if (CurrentBlock->isLandingPad()) continue; 2336 2337 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2338 BE = CurrentBlock->end(); BI != BE; ) { 2339 Instruction *CurInst = BI++; 2340 2341 if (isa<AllocaInst>(CurInst) || 2342 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2343 CurInst->getType()->isVoidTy() || 2344 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2345 isa<DbgInfoIntrinsic>(CurInst)) 2346 continue; 2347 2348 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2349 // sinking the compare again, and it would force the code generator to 2350 // move the i1 from processor flags or predicate registers into a general 2351 // purpose register. 2352 if (isa<CmpInst>(CurInst)) 2353 continue; 2354 2355 // We don't currently value number ANY inline asm calls. 2356 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2357 if (CallI->isInlineAsm()) 2358 continue; 2359 2360 uint32_t ValNo = VN.lookup(CurInst); 2361 2362 // Look for the predecessors for PRE opportunities. We're 2363 // only trying to solve the basic diamond case, where 2364 // a value is computed in the successor and one predecessor, 2365 // but not the other. We also explicitly disallow cases 2366 // where the successor is its own predecessor, because they're 2367 // more complicated to get right. 2368 unsigned NumWith = 0; 2369 unsigned NumWithout = 0; 2370 BasicBlock *PREPred = 0; 2371 predMap.clear(); 2372 2373 for (pred_iterator PI = pred_begin(CurrentBlock), 2374 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2375 BasicBlock *P = *PI; 2376 // We're not interested in PRE where the block is its 2377 // own predecessor, or in blocks with predecessors 2378 // that are not reachable. 2379 if (P == CurrentBlock) { 2380 NumWithout = 2; 2381 break; 2382 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2383 NumWithout = 2; 2384 break; 2385 } 2386 2387 Value* predV = findLeader(P, ValNo); 2388 if (predV == 0) { 2389 PREPred = P; 2390 ++NumWithout; 2391 } else if (predV == CurInst) { 2392 NumWithout = 2; 2393 } else { 2394 predMap[P] = predV; 2395 ++NumWith; 2396 } 2397 } 2398 2399 // Don't do PRE when it might increase code size, i.e. when 2400 // we would need to insert instructions in more than one pred. 2401 if (NumWithout != 1 || NumWith == 0) 2402 continue; 2403 2404 // Don't do PRE across indirect branch. 2405 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2406 continue; 2407 2408 // We can't do PRE safely on a critical edge, so instead we schedule 2409 // the edge to be split and perform the PRE the next time we iterate 2410 // on the function. 2411 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2412 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2413 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2414 continue; 2415 } 2416 2417 // Instantiate the expression in the predecessor that lacked it. 2418 // Because we are going top-down through the block, all value numbers 2419 // will be available in the predecessor by the time we need them. Any 2420 // that weren't originally present will have been instantiated earlier 2421 // in this loop. 2422 Instruction *PREInstr = CurInst->clone(); 2423 bool success = true; 2424 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2425 Value *Op = PREInstr->getOperand(i); 2426 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2427 continue; 2428 2429 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2430 PREInstr->setOperand(i, V); 2431 } else { 2432 success = false; 2433 break; 2434 } 2435 } 2436 2437 // Fail out if we encounter an operand that is not available in 2438 // the PRE predecessor. This is typically because of loads which 2439 // are not value numbered precisely. 2440 if (!success) { 2441 delete PREInstr; 2442 DEBUG(verifyRemoved(PREInstr)); 2443 continue; 2444 } 2445 2446 PREInstr->insertBefore(PREPred->getTerminator()); 2447 PREInstr->setName(CurInst->getName() + ".pre"); 2448 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2449 predMap[PREPred] = PREInstr; 2450 VN.add(PREInstr, ValNo); 2451 ++NumGVNPRE; 2452 2453 // Update the availability map to include the new instruction. 2454 addToLeaderTable(ValNo, PREInstr, PREPred); 2455 2456 // Create a PHI to make the value available in this block. 2457 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2458 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2459 CurInst->getName() + ".pre-phi", 2460 CurrentBlock->begin()); 2461 for (pred_iterator PI = PB; PI != PE; ++PI) { 2462 BasicBlock *P = *PI; 2463 Phi->addIncoming(predMap[P], P); 2464 } 2465 2466 VN.add(Phi, ValNo); 2467 addToLeaderTable(ValNo, Phi, CurrentBlock); 2468 Phi->setDebugLoc(CurInst->getDebugLoc()); 2469 CurInst->replaceAllUsesWith(Phi); 2470 if (Phi->getType()->isPointerTy()) { 2471 // Because we have added a PHI-use of the pointer value, it has now 2472 // "escaped" from alias analysis' perspective. We need to inform 2473 // AA of this. 2474 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2475 ++ii) { 2476 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2477 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2478 } 2479 2480 if (MD) 2481 MD->invalidateCachedPointerInfo(Phi); 2482 } 2483 VN.erase(CurInst); 2484 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2485 2486 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2487 if (MD) MD->removeInstruction(CurInst); 2488 CurInst->eraseFromParent(); 2489 DEBUG(verifyRemoved(CurInst)); 2490 Changed = true; 2491 } 2492 } 2493 2494 if (splitCriticalEdges()) 2495 Changed = true; 2496 2497 return Changed; 2498} 2499 2500/// splitCriticalEdges - Split critical edges found during the previous 2501/// iteration that may enable further optimization. 2502bool GVN::splitCriticalEdges() { 2503 if (toSplit.empty()) 2504 return false; 2505 do { 2506 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2507 SplitCriticalEdge(Edge.first, Edge.second, this); 2508 } while (!toSplit.empty()); 2509 if (MD) MD->invalidateCachedPredecessors(); 2510 return true; 2511} 2512 2513/// iterateOnFunction - Executes one iteration of GVN 2514bool GVN::iterateOnFunction(Function &F) { 2515 cleanupGlobalSets(); 2516 2517 // Top-down walk of the dominator tree 2518 bool Changed = false; 2519#if 0 2520 // Needed for value numbering with phi construction to work. 2521 ReversePostOrderTraversal<Function*> RPOT(&F); 2522 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2523 RE = RPOT.end(); RI != RE; ++RI) 2524 Changed |= processBlock(*RI); 2525#else 2526 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2527 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2528 Changed |= processBlock(DI->getBlock()); 2529#endif 2530 2531 return Changed; 2532} 2533 2534void GVN::cleanupGlobalSets() { 2535 VN.clear(); 2536 LeaderTable.clear(); 2537 TableAllocator.Reset(); 2538} 2539 2540/// verifyRemoved - Verify that the specified instruction does not occur in our 2541/// internal data structures. 2542void GVN::verifyRemoved(const Instruction *Inst) const { 2543 VN.verifyRemoved(Inst); 2544 2545 // Walk through the value number scope to make sure the instruction isn't 2546 // ferreted away in it. 2547 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2548 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2549 const LeaderTableEntry *Node = &I->second; 2550 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2551 2552 while (Node->Next) { 2553 Node = Node->Next; 2554 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2555 } 2556 } 2557} 2558