GVN.cpp revision 06cb8ed00696eb14d1b831921452e50ec0568ea2
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/IRBuilder.h" 22#include "llvm/IntrinsicInst.h" 23#include "llvm/LLVMContext.h" 24#include "llvm/Metadata.h" 25#include "llvm/ADT/DenseMap.h" 26#include "llvm/ADT/DepthFirstIterator.h" 27#include "llvm/ADT/Hashing.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/ADT/Statistic.h" 30#include "llvm/Analysis/AliasAnalysis.h" 31#include "llvm/Analysis/ConstantFolding.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Analysis/InstructionSimplify.h" 34#include "llvm/Analysis/Loads.h" 35#include "llvm/Analysis/MemoryBuiltins.h" 36#include "llvm/Analysis/MemoryDependenceAnalysis.h" 37#include "llvm/Analysis/PHITransAddr.h" 38#include "llvm/Analysis/ValueTracking.h" 39#include "llvm/Assembly/Writer.h" 40#include "llvm/Support/Allocator.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/PatternMatch.h" 44#include "llvm/Target/TargetData.h" 45#include "llvm/Target/TargetLibraryInfo.h" 46#include "llvm/Transforms/Utils/BasicBlockUtils.h" 47#include "llvm/Transforms/Utils/SSAUpdater.h" 48using namespace llvm; 49using namespace PatternMatch; 50 51STATISTIC(NumGVNInstr, "Number of instructions deleted"); 52STATISTIC(NumGVNLoad, "Number of loads deleted"); 53STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 54STATISTIC(NumGVNBlocks, "Number of blocks merged"); 55STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 56STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 57STATISTIC(NumPRELoad, "Number of loads PRE'd"); 58 59static cl::opt<bool> EnablePRE("enable-pre", 60 cl::init(true), cl::Hidden); 61static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 62 63// Maximum allowed recursion depth. 64static cl::opt<uint32_t> 65MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 66 cl::desc("Max recurse depth (default = 1000)")); 67 68//===----------------------------------------------------------------------===// 69// ValueTable Class 70//===----------------------------------------------------------------------===// 71 72/// This class holds the mapping between values and value numbers. It is used 73/// as an efficient mechanism to determine the expression-wise equivalence of 74/// two values. 75namespace { 76 struct Expression { 77 uint32_t opcode; 78 Type *type; 79 SmallVector<uint32_t, 4> varargs; 80 81 Expression(uint32_t o = ~2U) : opcode(o) { } 82 83 bool operator==(const Expression &other) const { 84 if (opcode != other.opcode) 85 return false; 86 if (opcode == ~0U || opcode == ~1U) 87 return true; 88 if (type != other.type) 89 return false; 90 if (varargs != other.varargs) 91 return false; 92 return true; 93 } 94 95 friend hash_code hash_value(const Expression &Value) { 96 return hash_combine(Value.opcode, Value.type, 97 hash_combine_range(Value.varargs.begin(), 98 Value.varargs.end())); 99 } 100 }; 101 102 class ValueTable { 103 DenseMap<Value*, uint32_t> valueNumbering; 104 DenseMap<Expression, uint32_t> expressionNumbering; 105 AliasAnalysis *AA; 106 MemoryDependenceAnalysis *MD; 107 DominatorTree *DT; 108 109 uint32_t nextValueNumber; 110 111 Expression create_expression(Instruction* I); 112 Expression create_cmp_expression(unsigned Opcode, 113 CmpInst::Predicate Predicate, 114 Value *LHS, Value *RHS); 115 Expression create_extractvalue_expression(ExtractValueInst* EI); 116 uint32_t lookup_or_add_call(CallInst* C); 117 public: 118 ValueTable() : nextValueNumber(1) { } 119 uint32_t lookup_or_add(Value *V); 120 uint32_t lookup(Value *V) const; 121 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 122 Value *LHS, Value *RHS); 123 void add(Value *V, uint32_t num); 124 void clear(); 125 void erase(Value *v); 126 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 127 AliasAnalysis *getAliasAnalysis() const { return AA; } 128 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 129 void setDomTree(DominatorTree* D) { DT = D; } 130 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 131 void verifyRemoved(const Value *) const; 132 }; 133} 134 135namespace llvm { 136template <> struct DenseMapInfo<Expression> { 137 static inline Expression getEmptyKey() { 138 return ~0U; 139 } 140 141 static inline Expression getTombstoneKey() { 142 return ~1U; 143 } 144 145 static unsigned getHashValue(const Expression e) { 146 using llvm::hash_value; 147 return static_cast<unsigned>(hash_value(e)); 148 } 149 static bool isEqual(const Expression &LHS, const Expression &RHS) { 150 return LHS == RHS; 151 } 152}; 153 154} 155 156//===----------------------------------------------------------------------===// 157// ValueTable Internal Functions 158//===----------------------------------------------------------------------===// 159 160Expression ValueTable::create_expression(Instruction *I) { 161 Expression e; 162 e.type = I->getType(); 163 e.opcode = I->getOpcode(); 164 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 165 OI != OE; ++OI) 166 e.varargs.push_back(lookup_or_add(*OI)); 167 if (I->isCommutative()) { 168 // Ensure that commutative instructions that only differ by a permutation 169 // of their operands get the same value number by sorting the operand value 170 // numbers. Since all commutative instructions have two operands it is more 171 // efficient to sort by hand rather than using, say, std::sort. 172 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 173 if (e.varargs[0] > e.varargs[1]) 174 std::swap(e.varargs[0], e.varargs[1]); 175 } 176 177 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 178 // Sort the operand value numbers so x<y and y>x get the same value number. 179 CmpInst::Predicate Predicate = C->getPredicate(); 180 if (e.varargs[0] > e.varargs[1]) { 181 std::swap(e.varargs[0], e.varargs[1]); 182 Predicate = CmpInst::getSwappedPredicate(Predicate); 183 } 184 e.opcode = (C->getOpcode() << 8) | Predicate; 185 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 186 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 187 II != IE; ++II) 188 e.varargs.push_back(*II); 189 } 190 191 return e; 192} 193 194Expression ValueTable::create_cmp_expression(unsigned Opcode, 195 CmpInst::Predicate Predicate, 196 Value *LHS, Value *RHS) { 197 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 198 "Not a comparison!"); 199 Expression e; 200 e.type = CmpInst::makeCmpResultType(LHS->getType()); 201 e.varargs.push_back(lookup_or_add(LHS)); 202 e.varargs.push_back(lookup_or_add(RHS)); 203 204 // Sort the operand value numbers so x<y and y>x get the same value number. 205 if (e.varargs[0] > e.varargs[1]) { 206 std::swap(e.varargs[0], e.varargs[1]); 207 Predicate = CmpInst::getSwappedPredicate(Predicate); 208 } 209 e.opcode = (Opcode << 8) | Predicate; 210 return e; 211} 212 213Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 214 assert(EI != 0 && "Not an ExtractValueInst?"); 215 Expression e; 216 e.type = EI->getType(); 217 e.opcode = 0; 218 219 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 220 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 221 // EI might be an extract from one of our recognised intrinsics. If it 222 // is we'll synthesize a semantically equivalent expression instead on 223 // an extract value expression. 224 switch (I->getIntrinsicID()) { 225 case Intrinsic::sadd_with_overflow: 226 case Intrinsic::uadd_with_overflow: 227 e.opcode = Instruction::Add; 228 break; 229 case Intrinsic::ssub_with_overflow: 230 case Intrinsic::usub_with_overflow: 231 e.opcode = Instruction::Sub; 232 break; 233 case Intrinsic::smul_with_overflow: 234 case Intrinsic::umul_with_overflow: 235 e.opcode = Instruction::Mul; 236 break; 237 default: 238 break; 239 } 240 241 if (e.opcode != 0) { 242 // Intrinsic recognized. Grab its args to finish building the expression. 243 assert(I->getNumArgOperands() == 2 && 244 "Expect two args for recognised intrinsics."); 245 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 246 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 247 return e; 248 } 249 } 250 251 // Not a recognised intrinsic. Fall back to producing an extract value 252 // expression. 253 e.opcode = EI->getOpcode(); 254 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 255 OI != OE; ++OI) 256 e.varargs.push_back(lookup_or_add(*OI)); 257 258 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 259 II != IE; ++II) 260 e.varargs.push_back(*II); 261 262 return e; 263} 264 265//===----------------------------------------------------------------------===// 266// ValueTable External Functions 267//===----------------------------------------------------------------------===// 268 269/// add - Insert a value into the table with a specified value number. 270void ValueTable::add(Value *V, uint32_t num) { 271 valueNumbering.insert(std::make_pair(V, num)); 272} 273 274uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 275 if (AA->doesNotAccessMemory(C)) { 276 Expression exp = create_expression(C); 277 uint32_t& e = expressionNumbering[exp]; 278 if (!e) e = nextValueNumber++; 279 valueNumbering[C] = e; 280 return e; 281 } else if (AA->onlyReadsMemory(C)) { 282 Expression exp = create_expression(C); 283 uint32_t& e = expressionNumbering[exp]; 284 if (!e) { 285 e = nextValueNumber++; 286 valueNumbering[C] = e; 287 return e; 288 } 289 if (!MD) { 290 e = nextValueNumber++; 291 valueNumbering[C] = e; 292 return e; 293 } 294 295 MemDepResult local_dep = MD->getDependency(C); 296 297 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 298 valueNumbering[C] = nextValueNumber; 299 return nextValueNumber++; 300 } 301 302 if (local_dep.isDef()) { 303 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 304 305 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 306 valueNumbering[C] = nextValueNumber; 307 return nextValueNumber++; 308 } 309 310 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 311 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 312 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 313 if (c_vn != cd_vn) { 314 valueNumbering[C] = nextValueNumber; 315 return nextValueNumber++; 316 } 317 } 318 319 uint32_t v = lookup_or_add(local_cdep); 320 valueNumbering[C] = v; 321 return v; 322 } 323 324 // Non-local case. 325 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 326 MD->getNonLocalCallDependency(CallSite(C)); 327 // FIXME: Move the checking logic to MemDep! 328 CallInst* cdep = 0; 329 330 // Check to see if we have a single dominating call instruction that is 331 // identical to C. 332 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 333 const NonLocalDepEntry *I = &deps[i]; 334 if (I->getResult().isNonLocal()) 335 continue; 336 337 // We don't handle non-definitions. If we already have a call, reject 338 // instruction dependencies. 339 if (!I->getResult().isDef() || cdep != 0) { 340 cdep = 0; 341 break; 342 } 343 344 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 345 // FIXME: All duplicated with non-local case. 346 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 347 cdep = NonLocalDepCall; 348 continue; 349 } 350 351 cdep = 0; 352 break; 353 } 354 355 if (!cdep) { 356 valueNumbering[C] = nextValueNumber; 357 return nextValueNumber++; 358 } 359 360 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 361 valueNumbering[C] = nextValueNumber; 362 return nextValueNumber++; 363 } 364 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 365 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 366 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 367 if (c_vn != cd_vn) { 368 valueNumbering[C] = nextValueNumber; 369 return nextValueNumber++; 370 } 371 } 372 373 uint32_t v = lookup_or_add(cdep); 374 valueNumbering[C] = v; 375 return v; 376 377 } else { 378 valueNumbering[C] = nextValueNumber; 379 return nextValueNumber++; 380 } 381} 382 383/// lookup_or_add - Returns the value number for the specified value, assigning 384/// it a new number if it did not have one before. 385uint32_t ValueTable::lookup_or_add(Value *V) { 386 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 387 if (VI != valueNumbering.end()) 388 return VI->second; 389 390 if (!isa<Instruction>(V)) { 391 valueNumbering[V] = nextValueNumber; 392 return nextValueNumber++; 393 } 394 395 Instruction* I = cast<Instruction>(V); 396 Expression exp; 397 switch (I->getOpcode()) { 398 case Instruction::Call: 399 return lookup_or_add_call(cast<CallInst>(I)); 400 case Instruction::Add: 401 case Instruction::FAdd: 402 case Instruction::Sub: 403 case Instruction::FSub: 404 case Instruction::Mul: 405 case Instruction::FMul: 406 case Instruction::UDiv: 407 case Instruction::SDiv: 408 case Instruction::FDiv: 409 case Instruction::URem: 410 case Instruction::SRem: 411 case Instruction::FRem: 412 case Instruction::Shl: 413 case Instruction::LShr: 414 case Instruction::AShr: 415 case Instruction::And: 416 case Instruction::Or : 417 case Instruction::Xor: 418 case Instruction::ICmp: 419 case Instruction::FCmp: 420 case Instruction::Trunc: 421 case Instruction::ZExt: 422 case Instruction::SExt: 423 case Instruction::FPToUI: 424 case Instruction::FPToSI: 425 case Instruction::UIToFP: 426 case Instruction::SIToFP: 427 case Instruction::FPTrunc: 428 case Instruction::FPExt: 429 case Instruction::PtrToInt: 430 case Instruction::IntToPtr: 431 case Instruction::BitCast: 432 case Instruction::Select: 433 case Instruction::ExtractElement: 434 case Instruction::InsertElement: 435 case Instruction::ShuffleVector: 436 case Instruction::InsertValue: 437 case Instruction::GetElementPtr: 438 exp = create_expression(I); 439 break; 440 case Instruction::ExtractValue: 441 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 442 break; 443 default: 444 valueNumbering[V] = nextValueNumber; 445 return nextValueNumber++; 446 } 447 448 uint32_t& e = expressionNumbering[exp]; 449 if (!e) e = nextValueNumber++; 450 valueNumbering[V] = e; 451 return e; 452} 453 454/// lookup - Returns the value number of the specified value. Fails if 455/// the value has not yet been numbered. 456uint32_t ValueTable::lookup(Value *V) const { 457 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 458 assert(VI != valueNumbering.end() && "Value not numbered?"); 459 return VI->second; 460} 461 462/// lookup_or_add_cmp - Returns the value number of the given comparison, 463/// assigning it a new number if it did not have one before. Useful when 464/// we deduced the result of a comparison, but don't immediately have an 465/// instruction realizing that comparison to hand. 466uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 467 CmpInst::Predicate Predicate, 468 Value *LHS, Value *RHS) { 469 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 470 uint32_t& e = expressionNumbering[exp]; 471 if (!e) e = nextValueNumber++; 472 return e; 473} 474 475/// clear - Remove all entries from the ValueTable. 476void ValueTable::clear() { 477 valueNumbering.clear(); 478 expressionNumbering.clear(); 479 nextValueNumber = 1; 480} 481 482/// erase - Remove a value from the value numbering. 483void ValueTable::erase(Value *V) { 484 valueNumbering.erase(V); 485} 486 487/// verifyRemoved - Verify that the value is removed from all internal data 488/// structures. 489void ValueTable::verifyRemoved(const Value *V) const { 490 for (DenseMap<Value*, uint32_t>::const_iterator 491 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 492 assert(I->first != V && "Inst still occurs in value numbering map!"); 493 } 494} 495 496//===----------------------------------------------------------------------===// 497// GVN Pass 498//===----------------------------------------------------------------------===// 499 500namespace { 501 502 class GVN : public FunctionPass { 503 bool NoLoads; 504 MemoryDependenceAnalysis *MD; 505 DominatorTree *DT; 506 const TargetData *TD; 507 const TargetLibraryInfo *TLI; 508 509 ValueTable VN; 510 511 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 512 /// have that value number. Use findLeader to query it. 513 struct LeaderTableEntry { 514 Value *Val; 515 BasicBlock *BB; 516 LeaderTableEntry *Next; 517 }; 518 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 519 BumpPtrAllocator TableAllocator; 520 521 SmallVector<Instruction*, 8> InstrsToErase; 522 public: 523 static char ID; // Pass identification, replacement for typeid 524 explicit GVN(bool noloads = false) 525 : FunctionPass(ID), NoLoads(noloads), MD(0) { 526 initializeGVNPass(*PassRegistry::getPassRegistry()); 527 } 528 529 bool runOnFunction(Function &F); 530 531 /// markInstructionForDeletion - This removes the specified instruction from 532 /// our various maps and marks it for deletion. 533 void markInstructionForDeletion(Instruction *I) { 534 VN.erase(I); 535 InstrsToErase.push_back(I); 536 } 537 538 const TargetData *getTargetData() const { return TD; } 539 DominatorTree &getDominatorTree() const { return *DT; } 540 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 541 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 542 private: 543 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 544 /// its value number. 545 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 546 LeaderTableEntry &Curr = LeaderTable[N]; 547 if (!Curr.Val) { 548 Curr.Val = V; 549 Curr.BB = BB; 550 return; 551 } 552 553 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 554 Node->Val = V; 555 Node->BB = BB; 556 Node->Next = Curr.Next; 557 Curr.Next = Node; 558 } 559 560 /// removeFromLeaderTable - Scan the list of values corresponding to a given 561 /// value number, and remove the given instruction if encountered. 562 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 563 LeaderTableEntry* Prev = 0; 564 LeaderTableEntry* Curr = &LeaderTable[N]; 565 566 while (Curr->Val != I || Curr->BB != BB) { 567 Prev = Curr; 568 Curr = Curr->Next; 569 } 570 571 if (Prev) { 572 Prev->Next = Curr->Next; 573 } else { 574 if (!Curr->Next) { 575 Curr->Val = 0; 576 Curr->BB = 0; 577 } else { 578 LeaderTableEntry* Next = Curr->Next; 579 Curr->Val = Next->Val; 580 Curr->BB = Next->BB; 581 Curr->Next = Next->Next; 582 } 583 } 584 } 585 586 // List of critical edges to be split between iterations. 587 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 588 589 // This transformation requires dominator postdominator info 590 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 591 AU.addRequired<DominatorTree>(); 592 AU.addRequired<TargetLibraryInfo>(); 593 if (!NoLoads) 594 AU.addRequired<MemoryDependenceAnalysis>(); 595 AU.addRequired<AliasAnalysis>(); 596 597 AU.addPreserved<DominatorTree>(); 598 AU.addPreserved<AliasAnalysis>(); 599 } 600 601 602 // Helper fuctions 603 // FIXME: eliminate or document these better 604 bool processLoad(LoadInst *L); 605 bool processInstruction(Instruction *I); 606 bool processNonLocalLoad(LoadInst *L); 607 bool processBlock(BasicBlock *BB); 608 void dump(DenseMap<uint32_t, Value*> &d); 609 bool iterateOnFunction(Function &F); 610 bool performPRE(Function &F); 611 Value *findLeader(BasicBlock *BB, uint32_t num); 612 void cleanupGlobalSets(); 613 void verifyRemoved(const Instruction *I) const; 614 bool splitCriticalEdges(); 615 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 616 BasicBlock *Root); 617 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root); 618 }; 619 620 char GVN::ID = 0; 621} 622 623// createGVNPass - The public interface to this file... 624FunctionPass *llvm::createGVNPass(bool NoLoads) { 625 return new GVN(NoLoads); 626} 627 628INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 629INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 630INITIALIZE_PASS_DEPENDENCY(DominatorTree) 631INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 632INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 633INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 634 635void GVN::dump(DenseMap<uint32_t, Value*>& d) { 636 errs() << "{\n"; 637 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 638 E = d.end(); I != E; ++I) { 639 errs() << I->first << "\n"; 640 I->second->dump(); 641 } 642 errs() << "}\n"; 643} 644 645/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 646/// we're analyzing is fully available in the specified block. As we go, keep 647/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 648/// map is actually a tri-state map with the following values: 649/// 0) we know the block *is not* fully available. 650/// 1) we know the block *is* fully available. 651/// 2) we do not know whether the block is fully available or not, but we are 652/// currently speculating that it will be. 653/// 3) we are speculating for this block and have used that to speculate for 654/// other blocks. 655static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 656 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 657 uint32_t RecurseDepth) { 658 if (RecurseDepth > MaxRecurseDepth) 659 return false; 660 661 // Optimistically assume that the block is fully available and check to see 662 // if we already know about this block in one lookup. 663 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 664 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 665 666 // If the entry already existed for this block, return the precomputed value. 667 if (!IV.second) { 668 // If this is a speculative "available" value, mark it as being used for 669 // speculation of other blocks. 670 if (IV.first->second == 2) 671 IV.first->second = 3; 672 return IV.first->second != 0; 673 } 674 675 // Otherwise, see if it is fully available in all predecessors. 676 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 677 678 // If this block has no predecessors, it isn't live-in here. 679 if (PI == PE) 680 goto SpeculationFailure; 681 682 for (; PI != PE; ++PI) 683 // If the value isn't fully available in one of our predecessors, then it 684 // isn't fully available in this block either. Undo our previous 685 // optimistic assumption and bail out. 686 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 687 goto SpeculationFailure; 688 689 return true; 690 691// SpeculationFailure - If we get here, we found out that this is not, after 692// all, a fully-available block. We have a problem if we speculated on this and 693// used the speculation to mark other blocks as available. 694SpeculationFailure: 695 char &BBVal = FullyAvailableBlocks[BB]; 696 697 // If we didn't speculate on this, just return with it set to false. 698 if (BBVal == 2) { 699 BBVal = 0; 700 return false; 701 } 702 703 // If we did speculate on this value, we could have blocks set to 1 that are 704 // incorrect. Walk the (transitive) successors of this block and mark them as 705 // 0 if set to one. 706 SmallVector<BasicBlock*, 32> BBWorklist; 707 BBWorklist.push_back(BB); 708 709 do { 710 BasicBlock *Entry = BBWorklist.pop_back_val(); 711 // Note that this sets blocks to 0 (unavailable) if they happen to not 712 // already be in FullyAvailableBlocks. This is safe. 713 char &EntryVal = FullyAvailableBlocks[Entry]; 714 if (EntryVal == 0) continue; // Already unavailable. 715 716 // Mark as unavailable. 717 EntryVal = 0; 718 719 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 720 BBWorklist.push_back(*I); 721 } while (!BBWorklist.empty()); 722 723 return false; 724} 725 726 727/// CanCoerceMustAliasedValueToLoad - Return true if 728/// CoerceAvailableValueToLoadType will succeed. 729static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 730 Type *LoadTy, 731 const TargetData &TD) { 732 // If the loaded or stored value is an first class array or struct, don't try 733 // to transform them. We need to be able to bitcast to integer. 734 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 735 StoredVal->getType()->isStructTy() || 736 StoredVal->getType()->isArrayTy()) 737 return false; 738 739 // The store has to be at least as big as the load. 740 if (TD.getTypeSizeInBits(StoredVal->getType()) < 741 TD.getTypeSizeInBits(LoadTy)) 742 return false; 743 744 return true; 745} 746 747 748/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 749/// then a load from a must-aliased pointer of a different type, try to coerce 750/// the stored value. LoadedTy is the type of the load we want to replace and 751/// InsertPt is the place to insert new instructions. 752/// 753/// If we can't do it, return null. 754static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 755 Type *LoadedTy, 756 Instruction *InsertPt, 757 const TargetData &TD) { 758 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 759 return 0; 760 761 // If this is already the right type, just return it. 762 Type *StoredValTy = StoredVal->getType(); 763 764 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 765 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 766 767 // If the store and reload are the same size, we can always reuse it. 768 if (StoreSize == LoadSize) { 769 // Pointer to Pointer -> use bitcast. 770 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 771 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 772 773 // Convert source pointers to integers, which can be bitcast. 774 if (StoredValTy->isPointerTy()) { 775 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 776 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 777 } 778 779 Type *TypeToCastTo = LoadedTy; 780 if (TypeToCastTo->isPointerTy()) 781 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 782 783 if (StoredValTy != TypeToCastTo) 784 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 785 786 // Cast to pointer if the load needs a pointer type. 787 if (LoadedTy->isPointerTy()) 788 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 789 790 return StoredVal; 791 } 792 793 // If the loaded value is smaller than the available value, then we can 794 // extract out a piece from it. If the available value is too small, then we 795 // can't do anything. 796 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 797 798 // Convert source pointers to integers, which can be manipulated. 799 if (StoredValTy->isPointerTy()) { 800 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 801 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 802 } 803 804 // Convert vectors and fp to integer, which can be manipulated. 805 if (!StoredValTy->isIntegerTy()) { 806 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 807 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 808 } 809 810 // If this is a big-endian system, we need to shift the value down to the low 811 // bits so that a truncate will work. 812 if (TD.isBigEndian()) { 813 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 814 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 815 } 816 817 // Truncate the integer to the right size now. 818 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 819 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 820 821 if (LoadedTy == NewIntTy) 822 return StoredVal; 823 824 // If the result is a pointer, inttoptr. 825 if (LoadedTy->isPointerTy()) 826 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 827 828 // Otherwise, bitcast. 829 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 830} 831 832/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 833/// memdep query of a load that ends up being a clobbering memory write (store, 834/// memset, memcpy, memmove). This means that the write *may* provide bits used 835/// by the load but we can't be sure because the pointers don't mustalias. 836/// 837/// Check this case to see if there is anything more we can do before we give 838/// up. This returns -1 if we have to give up, or a byte number in the stored 839/// value of the piece that feeds the load. 840static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 841 Value *WritePtr, 842 uint64_t WriteSizeInBits, 843 const TargetData &TD) { 844 // If the loaded or stored value is a first class array or struct, don't try 845 // to transform them. We need to be able to bitcast to integer. 846 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 847 return -1; 848 849 int64_t StoreOffset = 0, LoadOffset = 0; 850 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 851 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 852 if (StoreBase != LoadBase) 853 return -1; 854 855 // If the load and store are to the exact same address, they should have been 856 // a must alias. AA must have gotten confused. 857 // FIXME: Study to see if/when this happens. One case is forwarding a memset 858 // to a load from the base of the memset. 859#if 0 860 if (LoadOffset == StoreOffset) { 861 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 862 << "Base = " << *StoreBase << "\n" 863 << "Store Ptr = " << *WritePtr << "\n" 864 << "Store Offs = " << StoreOffset << "\n" 865 << "Load Ptr = " << *LoadPtr << "\n"; 866 abort(); 867 } 868#endif 869 870 // If the load and store don't overlap at all, the store doesn't provide 871 // anything to the load. In this case, they really don't alias at all, AA 872 // must have gotten confused. 873 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 874 875 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 876 return -1; 877 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 878 LoadSize >>= 3; 879 880 881 bool isAAFailure = false; 882 if (StoreOffset < LoadOffset) 883 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 884 else 885 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 886 887 if (isAAFailure) { 888#if 0 889 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 890 << "Base = " << *StoreBase << "\n" 891 << "Store Ptr = " << *WritePtr << "\n" 892 << "Store Offs = " << StoreOffset << "\n" 893 << "Load Ptr = " << *LoadPtr << "\n"; 894 abort(); 895#endif 896 return -1; 897 } 898 899 // If the Load isn't completely contained within the stored bits, we don't 900 // have all the bits to feed it. We could do something crazy in the future 901 // (issue a smaller load then merge the bits in) but this seems unlikely to be 902 // valuable. 903 if (StoreOffset > LoadOffset || 904 StoreOffset+StoreSize < LoadOffset+LoadSize) 905 return -1; 906 907 // Okay, we can do this transformation. Return the number of bytes into the 908 // store that the load is. 909 return LoadOffset-StoreOffset; 910} 911 912/// AnalyzeLoadFromClobberingStore - This function is called when we have a 913/// memdep query of a load that ends up being a clobbering store. 914static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 915 StoreInst *DepSI, 916 const TargetData &TD) { 917 // Cannot handle reading from store of first-class aggregate yet. 918 if (DepSI->getValueOperand()->getType()->isStructTy() || 919 DepSI->getValueOperand()->getType()->isArrayTy()) 920 return -1; 921 922 Value *StorePtr = DepSI->getPointerOperand(); 923 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 924 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 925 StorePtr, StoreSize, TD); 926} 927 928/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 929/// memdep query of a load that ends up being clobbered by another load. See if 930/// the other load can feed into the second load. 931static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 932 LoadInst *DepLI, const TargetData &TD){ 933 // Cannot handle reading from store of first-class aggregate yet. 934 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 935 return -1; 936 937 Value *DepPtr = DepLI->getPointerOperand(); 938 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 939 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 940 if (R != -1) return R; 941 942 // If we have a load/load clobber an DepLI can be widened to cover this load, 943 // then we should widen it! 944 int64_t LoadOffs = 0; 945 const Value *LoadBase = 946 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 947 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 948 949 unsigned Size = MemoryDependenceAnalysis:: 950 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 951 if (Size == 0) return -1; 952 953 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 954} 955 956 957 958static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 959 MemIntrinsic *MI, 960 const TargetData &TD) { 961 // If the mem operation is a non-constant size, we can't handle it. 962 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 963 if (SizeCst == 0) return -1; 964 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 965 966 // If this is memset, we just need to see if the offset is valid in the size 967 // of the memset.. 968 if (MI->getIntrinsicID() == Intrinsic::memset) 969 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 970 MemSizeInBits, TD); 971 972 // If we have a memcpy/memmove, the only case we can handle is if this is a 973 // copy from constant memory. In that case, we can read directly from the 974 // constant memory. 975 MemTransferInst *MTI = cast<MemTransferInst>(MI); 976 977 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 978 if (Src == 0) return -1; 979 980 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 981 if (GV == 0 || !GV->isConstant()) return -1; 982 983 // See if the access is within the bounds of the transfer. 984 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 985 MI->getDest(), MemSizeInBits, TD); 986 if (Offset == -1) 987 return Offset; 988 989 // Otherwise, see if we can constant fold a load from the constant with the 990 // offset applied as appropriate. 991 Src = ConstantExpr::getBitCast(Src, 992 llvm::Type::getInt8PtrTy(Src->getContext())); 993 Constant *OffsetCst = 994 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 995 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 996 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 997 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 998 return Offset; 999 return -1; 1000} 1001 1002 1003/// GetStoreValueForLoad - This function is called when we have a 1004/// memdep query of a load that ends up being a clobbering store. This means 1005/// that the store provides bits used by the load but we the pointers don't 1006/// mustalias. Check this case to see if there is anything more we can do 1007/// before we give up. 1008static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1009 Type *LoadTy, 1010 Instruction *InsertPt, const TargetData &TD){ 1011 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1012 1013 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1014 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1015 1016 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1017 1018 // Compute which bits of the stored value are being used by the load. Convert 1019 // to an integer type to start with. 1020 if (SrcVal->getType()->isPointerTy()) 1021 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx)); 1022 if (!SrcVal->getType()->isIntegerTy()) 1023 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1024 1025 // Shift the bits to the least significant depending on endianness. 1026 unsigned ShiftAmt; 1027 if (TD.isLittleEndian()) 1028 ShiftAmt = Offset*8; 1029 else 1030 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1031 1032 if (ShiftAmt) 1033 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1034 1035 if (LoadSize != StoreSize) 1036 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1037 1038 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1039} 1040 1041/// GetLoadValueForLoad - This function is called when we have a 1042/// memdep query of a load that ends up being a clobbering load. This means 1043/// that the load *may* provide bits used by the load but we can't be sure 1044/// because the pointers don't mustalias. Check this case to see if there is 1045/// anything more we can do before we give up. 1046static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1047 Type *LoadTy, Instruction *InsertPt, 1048 GVN &gvn) { 1049 const TargetData &TD = *gvn.getTargetData(); 1050 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1051 // widen SrcVal out to a larger load. 1052 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1053 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1054 if (Offset+LoadSize > SrcValSize) { 1055 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1056 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1057 // If we have a load/load clobber an DepLI can be widened to cover this 1058 // load, then we should widen it to the next power of 2 size big enough! 1059 unsigned NewLoadSize = Offset+LoadSize; 1060 if (!isPowerOf2_32(NewLoadSize)) 1061 NewLoadSize = NextPowerOf2(NewLoadSize); 1062 1063 Value *PtrVal = SrcVal->getPointerOperand(); 1064 1065 // Insert the new load after the old load. This ensures that subsequent 1066 // memdep queries will find the new load. We can't easily remove the old 1067 // load completely because it is already in the value numbering table. 1068 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1069 Type *DestPTy = 1070 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1071 DestPTy = PointerType::get(DestPTy, 1072 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1073 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1074 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1075 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1076 NewLoad->takeName(SrcVal); 1077 NewLoad->setAlignment(SrcVal->getAlignment()); 1078 1079 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1080 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1081 1082 // Replace uses of the original load with the wider load. On a big endian 1083 // system, we need to shift down to get the relevant bits. 1084 Value *RV = NewLoad; 1085 if (TD.isBigEndian()) 1086 RV = Builder.CreateLShr(RV, 1087 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1088 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1089 SrcVal->replaceAllUsesWith(RV); 1090 1091 // We would like to use gvn.markInstructionForDeletion here, but we can't 1092 // because the load is already memoized into the leader map table that GVN 1093 // tracks. It is potentially possible to remove the load from the table, 1094 // but then there all of the operations based on it would need to be 1095 // rehashed. Just leave the dead load around. 1096 gvn.getMemDep().removeInstruction(SrcVal); 1097 SrcVal = NewLoad; 1098 } 1099 1100 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1101} 1102 1103 1104/// GetMemInstValueForLoad - This function is called when we have a 1105/// memdep query of a load that ends up being a clobbering mem intrinsic. 1106static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1107 Type *LoadTy, Instruction *InsertPt, 1108 const TargetData &TD){ 1109 LLVMContext &Ctx = LoadTy->getContext(); 1110 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1111 1112 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1113 1114 // We know that this method is only called when the mem transfer fully 1115 // provides the bits for the load. 1116 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1117 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1118 // independently of what the offset is. 1119 Value *Val = MSI->getValue(); 1120 if (LoadSize != 1) 1121 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1122 1123 Value *OneElt = Val; 1124 1125 // Splat the value out to the right number of bits. 1126 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1127 // If we can double the number of bytes set, do it. 1128 if (NumBytesSet*2 <= LoadSize) { 1129 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1130 Val = Builder.CreateOr(Val, ShVal); 1131 NumBytesSet <<= 1; 1132 continue; 1133 } 1134 1135 // Otherwise insert one byte at a time. 1136 Value *ShVal = Builder.CreateShl(Val, 1*8); 1137 Val = Builder.CreateOr(OneElt, ShVal); 1138 ++NumBytesSet; 1139 } 1140 1141 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1142 } 1143 1144 // Otherwise, this is a memcpy/memmove from a constant global. 1145 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1146 Constant *Src = cast<Constant>(MTI->getSource()); 1147 1148 // Otherwise, see if we can constant fold a load from the constant with the 1149 // offset applied as appropriate. 1150 Src = ConstantExpr::getBitCast(Src, 1151 llvm::Type::getInt8PtrTy(Src->getContext())); 1152 Constant *OffsetCst = 1153 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1154 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1155 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1156 return ConstantFoldLoadFromConstPtr(Src, &TD); 1157} 1158 1159namespace { 1160 1161struct AvailableValueInBlock { 1162 /// BB - The basic block in question. 1163 BasicBlock *BB; 1164 enum ValType { 1165 SimpleVal, // A simple offsetted value that is accessed. 1166 LoadVal, // A value produced by a load. 1167 MemIntrin // A memory intrinsic which is loaded from. 1168 }; 1169 1170 /// V - The value that is live out of the block. 1171 PointerIntPair<Value *, 2, ValType> Val; 1172 1173 /// Offset - The byte offset in Val that is interesting for the load query. 1174 unsigned Offset; 1175 1176 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1177 unsigned Offset = 0) { 1178 AvailableValueInBlock Res; 1179 Res.BB = BB; 1180 Res.Val.setPointer(V); 1181 Res.Val.setInt(SimpleVal); 1182 Res.Offset = Offset; 1183 return Res; 1184 } 1185 1186 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1187 unsigned Offset = 0) { 1188 AvailableValueInBlock Res; 1189 Res.BB = BB; 1190 Res.Val.setPointer(MI); 1191 Res.Val.setInt(MemIntrin); 1192 Res.Offset = Offset; 1193 return Res; 1194 } 1195 1196 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1197 unsigned Offset = 0) { 1198 AvailableValueInBlock Res; 1199 Res.BB = BB; 1200 Res.Val.setPointer(LI); 1201 Res.Val.setInt(LoadVal); 1202 Res.Offset = Offset; 1203 return Res; 1204 } 1205 1206 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1207 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1208 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1209 1210 Value *getSimpleValue() const { 1211 assert(isSimpleValue() && "Wrong accessor"); 1212 return Val.getPointer(); 1213 } 1214 1215 LoadInst *getCoercedLoadValue() const { 1216 assert(isCoercedLoadValue() && "Wrong accessor"); 1217 return cast<LoadInst>(Val.getPointer()); 1218 } 1219 1220 MemIntrinsic *getMemIntrinValue() const { 1221 assert(isMemIntrinValue() && "Wrong accessor"); 1222 return cast<MemIntrinsic>(Val.getPointer()); 1223 } 1224 1225 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1226 /// defined here to the specified type. This handles various coercion cases. 1227 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1228 Value *Res; 1229 if (isSimpleValue()) { 1230 Res = getSimpleValue(); 1231 if (Res->getType() != LoadTy) { 1232 const TargetData *TD = gvn.getTargetData(); 1233 assert(TD && "Need target data to handle type mismatch case"); 1234 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1235 *TD); 1236 1237 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1238 << *getSimpleValue() << '\n' 1239 << *Res << '\n' << "\n\n\n"); 1240 } 1241 } else if (isCoercedLoadValue()) { 1242 LoadInst *Load = getCoercedLoadValue(); 1243 if (Load->getType() == LoadTy && Offset == 0) { 1244 Res = Load; 1245 } else { 1246 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1247 gvn); 1248 1249 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1250 << *getCoercedLoadValue() << '\n' 1251 << *Res << '\n' << "\n\n\n"); 1252 } 1253 } else { 1254 const TargetData *TD = gvn.getTargetData(); 1255 assert(TD && "Need target data to handle type mismatch case"); 1256 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1257 LoadTy, BB->getTerminator(), *TD); 1258 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1259 << " " << *getMemIntrinValue() << '\n' 1260 << *Res << '\n' << "\n\n\n"); 1261 } 1262 return Res; 1263 } 1264}; 1265 1266} // end anonymous namespace 1267 1268/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1269/// construct SSA form, allowing us to eliminate LI. This returns the value 1270/// that should be used at LI's definition site. 1271static Value *ConstructSSAForLoadSet(LoadInst *LI, 1272 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1273 GVN &gvn) { 1274 // Check for the fully redundant, dominating load case. In this case, we can 1275 // just use the dominating value directly. 1276 if (ValuesPerBlock.size() == 1 && 1277 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1278 LI->getParent())) 1279 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1280 1281 // Otherwise, we have to construct SSA form. 1282 SmallVector<PHINode*, 8> NewPHIs; 1283 SSAUpdater SSAUpdate(&NewPHIs); 1284 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1285 1286 Type *LoadTy = LI->getType(); 1287 1288 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1289 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1290 BasicBlock *BB = AV.BB; 1291 1292 if (SSAUpdate.HasValueForBlock(BB)) 1293 continue; 1294 1295 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1296 } 1297 1298 // Perform PHI construction. 1299 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1300 1301 // If new PHI nodes were created, notify alias analysis. 1302 if (V->getType()->isPointerTy()) { 1303 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1304 1305 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1306 AA->copyValue(LI, NewPHIs[i]); 1307 1308 // Now that we've copied information to the new PHIs, scan through 1309 // them again and inform alias analysis that we've added potentially 1310 // escaping uses to any values that are operands to these PHIs. 1311 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1312 PHINode *P = NewPHIs[i]; 1313 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1314 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1315 AA->addEscapingUse(P->getOperandUse(jj)); 1316 } 1317 } 1318 } 1319 1320 return V; 1321} 1322 1323static bool isLifetimeStart(const Instruction *Inst) { 1324 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1325 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1326 return false; 1327} 1328 1329/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1330/// non-local by performing PHI construction. 1331bool GVN::processNonLocalLoad(LoadInst *LI) { 1332 // Find the non-local dependencies of the load. 1333 SmallVector<NonLocalDepResult, 64> Deps; 1334 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1335 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1336 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1337 // << Deps.size() << *LI << '\n'); 1338 1339 // If we had to process more than one hundred blocks to find the 1340 // dependencies, this load isn't worth worrying about. Optimizing 1341 // it will be too expensive. 1342 unsigned NumDeps = Deps.size(); 1343 if (NumDeps > 100) 1344 return false; 1345 1346 // If we had a phi translation failure, we'll have a single entry which is a 1347 // clobber in the current block. Reject this early. 1348 if (NumDeps == 1 && 1349 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1350 DEBUG( 1351 dbgs() << "GVN: non-local load "; 1352 WriteAsOperand(dbgs(), LI); 1353 dbgs() << " has unknown dependencies\n"; 1354 ); 1355 return false; 1356 } 1357 1358 // Filter out useless results (non-locals, etc). Keep track of the blocks 1359 // where we have a value available in repl, also keep track of whether we see 1360 // dependencies that produce an unknown value for the load (such as a call 1361 // that could potentially clobber the load). 1362 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock; 1363 SmallVector<BasicBlock*, 64> UnavailableBlocks; 1364 1365 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1366 BasicBlock *DepBB = Deps[i].getBB(); 1367 MemDepResult DepInfo = Deps[i].getResult(); 1368 1369 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1370 UnavailableBlocks.push_back(DepBB); 1371 continue; 1372 } 1373 1374 if (DepInfo.isClobber()) { 1375 // The address being loaded in this non-local block may not be the same as 1376 // the pointer operand of the load if PHI translation occurs. Make sure 1377 // to consider the right address. 1378 Value *Address = Deps[i].getAddress(); 1379 1380 // If the dependence is to a store that writes to a superset of the bits 1381 // read by the load, we can extract the bits we need for the load from the 1382 // stored value. 1383 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1384 if (TD && Address) { 1385 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1386 DepSI, *TD); 1387 if (Offset != -1) { 1388 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1389 DepSI->getValueOperand(), 1390 Offset)); 1391 continue; 1392 } 1393 } 1394 } 1395 1396 // Check to see if we have something like this: 1397 // load i32* P 1398 // load i8* (P+1) 1399 // if we have this, replace the later with an extraction from the former. 1400 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1401 // If this is a clobber and L is the first instruction in its block, then 1402 // we have the first instruction in the entry block. 1403 if (DepLI != LI && Address && TD) { 1404 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1405 LI->getPointerOperand(), 1406 DepLI, *TD); 1407 1408 if (Offset != -1) { 1409 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1410 Offset)); 1411 continue; 1412 } 1413 } 1414 } 1415 1416 // If the clobbering value is a memset/memcpy/memmove, see if we can 1417 // forward a value on from it. 1418 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1419 if (TD && Address) { 1420 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1421 DepMI, *TD); 1422 if (Offset != -1) { 1423 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1424 Offset)); 1425 continue; 1426 } 1427 } 1428 } 1429 1430 UnavailableBlocks.push_back(DepBB); 1431 continue; 1432 } 1433 1434 // DepInfo.isDef() here 1435 1436 Instruction *DepInst = DepInfo.getInst(); 1437 1438 // Loading the allocation -> undef. 1439 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst) || 1440 // Loading immediately after lifetime begin -> undef. 1441 isLifetimeStart(DepInst)) { 1442 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1443 UndefValue::get(LI->getType()))); 1444 continue; 1445 } 1446 1447 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1448 // Reject loads and stores that are to the same address but are of 1449 // different types if we have to. 1450 if (S->getValueOperand()->getType() != LI->getType()) { 1451 // If the stored value is larger or equal to the loaded value, we can 1452 // reuse it. 1453 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1454 LI->getType(), *TD)) { 1455 UnavailableBlocks.push_back(DepBB); 1456 continue; 1457 } 1458 } 1459 1460 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1461 S->getValueOperand())); 1462 continue; 1463 } 1464 1465 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1466 // If the types mismatch and we can't handle it, reject reuse of the load. 1467 if (LD->getType() != LI->getType()) { 1468 // If the stored value is larger or equal to the loaded value, we can 1469 // reuse it. 1470 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1471 UnavailableBlocks.push_back(DepBB); 1472 continue; 1473 } 1474 } 1475 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1476 continue; 1477 } 1478 1479 UnavailableBlocks.push_back(DepBB); 1480 continue; 1481 } 1482 1483 // If we have no predecessors that produce a known value for this load, exit 1484 // early. 1485 if (ValuesPerBlock.empty()) return false; 1486 1487 // If all of the instructions we depend on produce a known value for this 1488 // load, then it is fully redundant and we can use PHI insertion to compute 1489 // its value. Insert PHIs and remove the fully redundant value now. 1490 if (UnavailableBlocks.empty()) { 1491 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1492 1493 // Perform PHI construction. 1494 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1495 LI->replaceAllUsesWith(V); 1496 1497 if (isa<PHINode>(V)) 1498 V->takeName(LI); 1499 if (V->getType()->isPointerTy()) 1500 MD->invalidateCachedPointerInfo(V); 1501 markInstructionForDeletion(LI); 1502 ++NumGVNLoad; 1503 return true; 1504 } 1505 1506 if (!EnablePRE || !EnableLoadPRE) 1507 return false; 1508 1509 // Okay, we have *some* definitions of the value. This means that the value 1510 // is available in some of our (transitive) predecessors. Lets think about 1511 // doing PRE of this load. This will involve inserting a new load into the 1512 // predecessor when it's not available. We could do this in general, but 1513 // prefer to not increase code size. As such, we only do this when we know 1514 // that we only have to insert *one* load (which means we're basically moving 1515 // the load, not inserting a new one). 1516 1517 SmallPtrSet<BasicBlock *, 4> Blockers; 1518 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1519 Blockers.insert(UnavailableBlocks[i]); 1520 1521 // Let's find the first basic block with more than one predecessor. Walk 1522 // backwards through predecessors if needed. 1523 BasicBlock *LoadBB = LI->getParent(); 1524 BasicBlock *TmpBB = LoadBB; 1525 1526 bool isSinglePred = false; 1527 bool allSingleSucc = true; 1528 while (TmpBB->getSinglePredecessor()) { 1529 isSinglePred = true; 1530 TmpBB = TmpBB->getSinglePredecessor(); 1531 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1532 return false; 1533 if (Blockers.count(TmpBB)) 1534 return false; 1535 1536 // If any of these blocks has more than one successor (i.e. if the edge we 1537 // just traversed was critical), then there are other paths through this 1538 // block along which the load may not be anticipated. Hoisting the load 1539 // above this block would be adding the load to execution paths along 1540 // which it was not previously executed. 1541 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1542 return false; 1543 } 1544 1545 assert(TmpBB); 1546 LoadBB = TmpBB; 1547 1548 // FIXME: It is extremely unclear what this loop is doing, other than 1549 // artificially restricting loadpre. 1550 if (isSinglePred) { 1551 bool isHot = false; 1552 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1553 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1554 if (AV.isSimpleValue()) 1555 // "Hot" Instruction is in some loop (because it dominates its dep. 1556 // instruction). 1557 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1558 if (DT->dominates(LI, I)) { 1559 isHot = true; 1560 break; 1561 } 1562 } 1563 1564 // We are interested only in "hot" instructions. We don't want to do any 1565 // mis-optimizations here. 1566 if (!isHot) 1567 return false; 1568 } 1569 1570 // Check to see how many predecessors have the loaded value fully 1571 // available. 1572 DenseMap<BasicBlock*, Value*> PredLoads; 1573 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1574 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1575 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1576 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1577 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1578 1579 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1580 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1581 PI != E; ++PI) { 1582 BasicBlock *Pred = *PI; 1583 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1584 continue; 1585 } 1586 PredLoads[Pred] = 0; 1587 1588 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1589 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1590 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1591 << Pred->getName() << "': " << *LI << '\n'); 1592 return false; 1593 } 1594 1595 if (LoadBB->isLandingPad()) { 1596 DEBUG(dbgs() 1597 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1598 << Pred->getName() << "': " << *LI << '\n'); 1599 return false; 1600 } 1601 1602 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1603 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1604 } 1605 } 1606 1607 if (!NeedToSplit.empty()) { 1608 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1609 return false; 1610 } 1611 1612 // Decide whether PRE is profitable for this load. 1613 unsigned NumUnavailablePreds = PredLoads.size(); 1614 assert(NumUnavailablePreds != 0 && 1615 "Fully available value should be eliminated above!"); 1616 1617 // If this load is unavailable in multiple predecessors, reject it. 1618 // FIXME: If we could restructure the CFG, we could make a common pred with 1619 // all the preds that don't have an available LI and insert a new load into 1620 // that one block. 1621 if (NumUnavailablePreds != 1) 1622 return false; 1623 1624 // Check if the load can safely be moved to all the unavailable predecessors. 1625 bool CanDoPRE = true; 1626 SmallVector<Instruction*, 8> NewInsts; 1627 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1628 E = PredLoads.end(); I != E; ++I) { 1629 BasicBlock *UnavailablePred = I->first; 1630 1631 // Do PHI translation to get its value in the predecessor if necessary. The 1632 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1633 1634 // If all preds have a single successor, then we know it is safe to insert 1635 // the load on the pred (?!?), so we can insert code to materialize the 1636 // pointer if it is not available. 1637 PHITransAddr Address(LI->getPointerOperand(), TD); 1638 Value *LoadPtr = 0; 1639 if (allSingleSucc) { 1640 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1641 *DT, NewInsts); 1642 } else { 1643 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1644 LoadPtr = Address.getAddr(); 1645 } 1646 1647 // If we couldn't find or insert a computation of this phi translated value, 1648 // we fail PRE. 1649 if (LoadPtr == 0) { 1650 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1651 << *LI->getPointerOperand() << "\n"); 1652 CanDoPRE = false; 1653 break; 1654 } 1655 1656 // Make sure it is valid to move this load here. We have to watch out for: 1657 // @1 = getelementptr (i8* p, ... 1658 // test p and branch if == 0 1659 // load @1 1660 // It is valid to have the getelementptr before the test, even if p can 1661 // be 0, as getelementptr only does address arithmetic. 1662 // If we are not pushing the value through any multiple-successor blocks 1663 // we do not have this case. Otherwise, check that the load is safe to 1664 // put anywhere; this can be improved, but should be conservatively safe. 1665 if (!allSingleSucc && 1666 // FIXME: REEVALUTE THIS. 1667 !isSafeToLoadUnconditionally(LoadPtr, 1668 UnavailablePred->getTerminator(), 1669 LI->getAlignment(), TD)) { 1670 CanDoPRE = false; 1671 break; 1672 } 1673 1674 I->second = LoadPtr; 1675 } 1676 1677 if (!CanDoPRE) { 1678 while (!NewInsts.empty()) { 1679 Instruction *I = NewInsts.pop_back_val(); 1680 if (MD) MD->removeInstruction(I); 1681 I->eraseFromParent(); 1682 } 1683 return false; 1684 } 1685 1686 // Okay, we can eliminate this load by inserting a reload in the predecessor 1687 // and using PHI construction to get the value in the other predecessors, do 1688 // it. 1689 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1690 DEBUG(if (!NewInsts.empty()) 1691 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1692 << *NewInsts.back() << '\n'); 1693 1694 // Assign value numbers to the new instructions. 1695 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1696 // FIXME: We really _ought_ to insert these value numbers into their 1697 // parent's availability map. However, in doing so, we risk getting into 1698 // ordering issues. If a block hasn't been processed yet, we would be 1699 // marking a value as AVAIL-IN, which isn't what we intend. 1700 VN.lookup_or_add(NewInsts[i]); 1701 } 1702 1703 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1704 E = PredLoads.end(); I != E; ++I) { 1705 BasicBlock *UnavailablePred = I->first; 1706 Value *LoadPtr = I->second; 1707 1708 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1709 LI->getAlignment(), 1710 UnavailablePred->getTerminator()); 1711 1712 // Transfer the old load's TBAA tag to the new load. 1713 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1714 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1715 1716 // Transfer DebugLoc. 1717 NewLoad->setDebugLoc(LI->getDebugLoc()); 1718 1719 // Add the newly created load. 1720 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1721 NewLoad)); 1722 MD->invalidateCachedPointerInfo(LoadPtr); 1723 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1724 } 1725 1726 // Perform PHI construction. 1727 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1728 LI->replaceAllUsesWith(V); 1729 if (isa<PHINode>(V)) 1730 V->takeName(LI); 1731 if (V->getType()->isPointerTy()) 1732 MD->invalidateCachedPointerInfo(V); 1733 markInstructionForDeletion(LI); 1734 ++NumPRELoad; 1735 return true; 1736} 1737 1738static void patchReplacementInstruction(Value *Repl, Instruction *I) { 1739 // Patch the replacement so that it is not more restrictive than the value 1740 // being replaced. 1741 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1742 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1743 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1744 isa<OverflowingBinaryOperator>(ReplOp)) { 1745 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1746 ReplOp->setHasNoSignedWrap(false); 1747 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1748 ReplOp->setHasNoUnsignedWrap(false); 1749 } 1750 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1751 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1752 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1753 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1754 unsigned Kind = Metadata[i].first; 1755 MDNode *IMD = I->getMetadata(Kind); 1756 MDNode *ReplMD = Metadata[i].second; 1757 switch(Kind) { 1758 default: 1759 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1760 break; 1761 case LLVMContext::MD_dbg: 1762 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1763 case LLVMContext::MD_tbaa: 1764 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1765 break; 1766 case LLVMContext::MD_range: 1767 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1768 break; 1769 case LLVMContext::MD_prof: 1770 llvm_unreachable("MD_prof in a non terminator instruction"); 1771 break; 1772 case LLVMContext::MD_fpmath: 1773 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1774 break; 1775 } 1776 } 1777 } 1778} 1779 1780static void patchAndReplaceAllUsesWith(Value *Repl, Instruction *I) { 1781 patchReplacementInstruction(Repl, I); 1782 I->replaceAllUsesWith(Repl); 1783} 1784 1785/// processLoad - Attempt to eliminate a load, first by eliminating it 1786/// locally, and then attempting non-local elimination if that fails. 1787bool GVN::processLoad(LoadInst *L) { 1788 if (!MD) 1789 return false; 1790 1791 if (!L->isSimple()) 1792 return false; 1793 1794 if (L->use_empty()) { 1795 markInstructionForDeletion(L); 1796 return true; 1797 } 1798 1799 // ... to a pointer that has been loaded from before... 1800 MemDepResult Dep = MD->getDependency(L); 1801 1802 // If we have a clobber and target data is around, see if this is a clobber 1803 // that we can fix up through code synthesis. 1804 if (Dep.isClobber() && TD) { 1805 // Check to see if we have something like this: 1806 // store i32 123, i32* %P 1807 // %A = bitcast i32* %P to i8* 1808 // %B = gep i8* %A, i32 1 1809 // %C = load i8* %B 1810 // 1811 // We could do that by recognizing if the clobber instructions are obviously 1812 // a common base + constant offset, and if the previous store (or memset) 1813 // completely covers this load. This sort of thing can happen in bitfield 1814 // access code. 1815 Value *AvailVal = 0; 1816 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1817 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1818 L->getPointerOperand(), 1819 DepSI, *TD); 1820 if (Offset != -1) 1821 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1822 L->getType(), L, *TD); 1823 } 1824 1825 // Check to see if we have something like this: 1826 // load i32* P 1827 // load i8* (P+1) 1828 // if we have this, replace the later with an extraction from the former. 1829 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1830 // If this is a clobber and L is the first instruction in its block, then 1831 // we have the first instruction in the entry block. 1832 if (DepLI == L) 1833 return false; 1834 1835 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1836 L->getPointerOperand(), 1837 DepLI, *TD); 1838 if (Offset != -1) 1839 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1840 } 1841 1842 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1843 // a value on from it. 1844 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1845 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1846 L->getPointerOperand(), 1847 DepMI, *TD); 1848 if (Offset != -1) 1849 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1850 } 1851 1852 if (AvailVal) { 1853 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1854 << *AvailVal << '\n' << *L << "\n\n\n"); 1855 1856 // Replace the load! 1857 L->replaceAllUsesWith(AvailVal); 1858 if (AvailVal->getType()->isPointerTy()) 1859 MD->invalidateCachedPointerInfo(AvailVal); 1860 markInstructionForDeletion(L); 1861 ++NumGVNLoad; 1862 return true; 1863 } 1864 } 1865 1866 // If the value isn't available, don't do anything! 1867 if (Dep.isClobber()) { 1868 DEBUG( 1869 // fast print dep, using operator<< on instruction is too slow. 1870 dbgs() << "GVN: load "; 1871 WriteAsOperand(dbgs(), L); 1872 Instruction *I = Dep.getInst(); 1873 dbgs() << " is clobbered by " << *I << '\n'; 1874 ); 1875 return false; 1876 } 1877 1878 // If it is defined in another block, try harder. 1879 if (Dep.isNonLocal()) 1880 return processNonLocalLoad(L); 1881 1882 if (!Dep.isDef()) { 1883 DEBUG( 1884 // fast print dep, using operator<< on instruction is too slow. 1885 dbgs() << "GVN: load "; 1886 WriteAsOperand(dbgs(), L); 1887 dbgs() << " has unknown dependence\n"; 1888 ); 1889 return false; 1890 } 1891 1892 Instruction *DepInst = Dep.getInst(); 1893 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1894 Value *StoredVal = DepSI->getValueOperand(); 1895 1896 // The store and load are to a must-aliased pointer, but they may not 1897 // actually have the same type. See if we know how to reuse the stored 1898 // value (depending on its type). 1899 if (StoredVal->getType() != L->getType()) { 1900 if (TD) { 1901 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1902 L, *TD); 1903 if (StoredVal == 0) 1904 return false; 1905 1906 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1907 << '\n' << *L << "\n\n\n"); 1908 } 1909 else 1910 return false; 1911 } 1912 1913 // Remove it! 1914 L->replaceAllUsesWith(StoredVal); 1915 if (StoredVal->getType()->isPointerTy()) 1916 MD->invalidateCachedPointerInfo(StoredVal); 1917 markInstructionForDeletion(L); 1918 ++NumGVNLoad; 1919 return true; 1920 } 1921 1922 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1923 Value *AvailableVal = DepLI; 1924 1925 // The loads are of a must-aliased pointer, but they may not actually have 1926 // the same type. See if we know how to reuse the previously loaded value 1927 // (depending on its type). 1928 if (DepLI->getType() != L->getType()) { 1929 if (TD) { 1930 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1931 L, *TD); 1932 if (AvailableVal == 0) 1933 return false; 1934 1935 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1936 << "\n" << *L << "\n\n\n"); 1937 } 1938 else 1939 return false; 1940 } 1941 1942 // Remove it! 1943 patchAndReplaceAllUsesWith(AvailableVal, L); 1944 if (DepLI->getType()->isPointerTy()) 1945 MD->invalidateCachedPointerInfo(DepLI); 1946 markInstructionForDeletion(L); 1947 ++NumGVNLoad; 1948 return true; 1949 } 1950 1951 // If this load really doesn't depend on anything, then we must be loading an 1952 // undef value. This can happen when loading for a fresh allocation with no 1953 // intervening stores, for example. 1954 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst)) { 1955 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1956 markInstructionForDeletion(L); 1957 ++NumGVNLoad; 1958 return true; 1959 } 1960 1961 // If this load occurs either right after a lifetime begin, 1962 // then the loaded value is undefined. 1963 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1964 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1965 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1966 markInstructionForDeletion(L); 1967 ++NumGVNLoad; 1968 return true; 1969 } 1970 } 1971 1972 return false; 1973} 1974 1975// findLeader - In order to find a leader for a given value number at a 1976// specific basic block, we first obtain the list of all Values for that number, 1977// and then scan the list to find one whose block dominates the block in 1978// question. This is fast because dominator tree queries consist of only 1979// a few comparisons of DFS numbers. 1980Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1981 LeaderTableEntry Vals = LeaderTable[num]; 1982 if (!Vals.Val) return 0; 1983 1984 Value *Val = 0; 1985 if (DT->dominates(Vals.BB, BB)) { 1986 Val = Vals.Val; 1987 if (isa<Constant>(Val)) return Val; 1988 } 1989 1990 LeaderTableEntry* Next = Vals.Next; 1991 while (Next) { 1992 if (DT->dominates(Next->BB, BB)) { 1993 if (isa<Constant>(Next->Val)) return Next->Val; 1994 if (!Val) Val = Next->Val; 1995 } 1996 1997 Next = Next->Next; 1998 } 1999 2000 return Val; 2001} 2002 2003/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 2004/// use is dominated by the given basic block. Returns the number of uses that 2005/// were replaced. 2006unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 2007 BasicBlock *Root) { 2008 unsigned Count = 0; 2009 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2010 UI != UE; ) { 2011 Use &U = (UI++).getUse(); 2012 2013 // If From occurs as a phi node operand then the use implicitly lives in the 2014 // corresponding incoming block. Otherwise it is the block containing the 2015 // user that must be dominated by Root. 2016 BasicBlock *UsingBlock; 2017 if (PHINode *PN = dyn_cast<PHINode>(U.getUser())) 2018 UsingBlock = PN->getIncomingBlock(U); 2019 else 2020 UsingBlock = cast<Instruction>(U.getUser())->getParent(); 2021 2022 if (DT->dominates(Root, UsingBlock)) { 2023 U.set(To); 2024 ++Count; 2025 } 2026 } 2027 return Count; 2028} 2029 2030/// propagateEquality - The given values are known to be equal in every block 2031/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2032/// 'RHS' everywhere in the scope. Returns whether a change was made. 2033bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) { 2034 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2035 Worklist.push_back(std::make_pair(LHS, RHS)); 2036 bool Changed = false; 2037 2038 while (!Worklist.empty()) { 2039 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2040 LHS = Item.first; RHS = Item.second; 2041 2042 if (LHS == RHS) continue; 2043 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2044 2045 // Don't try to propagate equalities between constants. 2046 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2047 2048 // Prefer a constant on the right-hand side, or an Argument if no constants. 2049 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2050 std::swap(LHS, RHS); 2051 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2052 2053 // If there is no obvious reason to prefer the left-hand side over the right- 2054 // hand side, ensure the longest lived term is on the right-hand side, so the 2055 // shortest lived term will be replaced by the longest lived. This tends to 2056 // expose more simplifications. 2057 uint32_t LVN = VN.lookup_or_add(LHS); 2058 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2059 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2060 // Move the 'oldest' value to the right-hand side, using the value number as 2061 // a proxy for age. 2062 uint32_t RVN = VN.lookup_or_add(RHS); 2063 if (LVN < RVN) { 2064 std::swap(LHS, RHS); 2065 LVN = RVN; 2066 } 2067 } 2068 assert((!isa<Instruction>(RHS) || 2069 DT->properlyDominates(cast<Instruction>(RHS)->getParent(), Root)) && 2070 "Instruction doesn't dominate scope!"); 2071 2072 // If value numbering later sees that an instruction in the scope is equal 2073 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2074 // the invariant that instructions only occur in the leader table for their 2075 // own value number (this is used by removeFromLeaderTable), do not do this 2076 // if RHS is an instruction (if an instruction in the scope is morphed into 2077 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2078 // using the leader table is about compiling faster, not optimizing better). 2079 if (!isa<Instruction>(RHS)) 2080 addToLeaderTable(LVN, RHS, Root); 2081 2082 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2083 // LHS always has at least one use that is not dominated by Root, this will 2084 // never do anything if LHS has only one use. 2085 if (!LHS->hasOneUse()) { 2086 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2087 Changed |= NumReplacements > 0; 2088 NumGVNEqProp += NumReplacements; 2089 } 2090 2091 // Now try to deduce additional equalities from this one. For example, if the 2092 // known equality was "(A != B)" == "false" then it follows that A and B are 2093 // equal in the scope. Only boolean equalities with an explicit true or false 2094 // RHS are currently supported. 2095 if (!RHS->getType()->isIntegerTy(1)) 2096 // Not a boolean equality - bail out. 2097 continue; 2098 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2099 if (!CI) 2100 // RHS neither 'true' nor 'false' - bail out. 2101 continue; 2102 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2103 bool isKnownTrue = CI->isAllOnesValue(); 2104 bool isKnownFalse = !isKnownTrue; 2105 2106 // If "A && B" is known true then both A and B are known true. If "A || B" 2107 // is known false then both A and B are known false. 2108 Value *A, *B; 2109 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2110 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2111 Worklist.push_back(std::make_pair(A, RHS)); 2112 Worklist.push_back(std::make_pair(B, RHS)); 2113 continue; 2114 } 2115 2116 // If we are propagating an equality like "(A == B)" == "true" then also 2117 // propagate the equality A == B. When propagating a comparison such as 2118 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2119 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2120 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2121 2122 // If "A == B" is known true, or "A != B" is known false, then replace 2123 // A with B everywhere in the scope. 2124 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2125 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2126 Worklist.push_back(std::make_pair(Op0, Op1)); 2127 2128 // If "A >= B" is known true, replace "A < B" with false everywhere. 2129 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2130 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2131 // Since we don't have the instruction "A < B" immediately to hand, work out 2132 // the value number that it would have and use that to find an appropriate 2133 // instruction (if any). 2134 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2135 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2136 // If the number we were assigned was brand new then there is no point in 2137 // looking for an instruction realizing it: there cannot be one! 2138 if (Num < NextNum) { 2139 Value *NotCmp = findLeader(Root, Num); 2140 if (NotCmp && isa<Instruction>(NotCmp)) { 2141 unsigned NumReplacements = 2142 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2143 Changed |= NumReplacements > 0; 2144 NumGVNEqProp += NumReplacements; 2145 } 2146 } 2147 // Ensure that any instruction in scope that gets the "A < B" value number 2148 // is replaced with false. 2149 addToLeaderTable(Num, NotVal, Root); 2150 2151 continue; 2152 } 2153 } 2154 2155 return Changed; 2156} 2157 2158/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2159/// true if every path from the entry block to 'Dst' passes via this edge. In 2160/// particular 'Dst' must not be reachable via another edge from 'Src'. 2161static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst, 2162 DominatorTree *DT) { 2163 // While in theory it is interesting to consider the case in which Dst has 2164 // more than one predecessor, because Dst might be part of a loop which is 2165 // only reachable from Src, in practice it is pointless since at the time 2166 // GVN runs all such loops have preheaders, which means that Dst will have 2167 // been changed to have only one predecessor, namely Src. 2168 BasicBlock *Pred = Dst->getSinglePredecessor(); 2169 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2170 (void)Src; 2171 return Pred != 0; 2172} 2173 2174/// processInstruction - When calculating availability, handle an instruction 2175/// by inserting it into the appropriate sets 2176bool GVN::processInstruction(Instruction *I) { 2177 // Ignore dbg info intrinsics. 2178 if (isa<DbgInfoIntrinsic>(I)) 2179 return false; 2180 2181 // If the instruction can be easily simplified then do so now in preference 2182 // to value numbering it. Value numbering often exposes redundancies, for 2183 // example if it determines that %y is equal to %x then the instruction 2184 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2185 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2186 I->replaceAllUsesWith(V); 2187 if (MD && V->getType()->isPointerTy()) 2188 MD->invalidateCachedPointerInfo(V); 2189 markInstructionForDeletion(I); 2190 ++NumGVNSimpl; 2191 return true; 2192 } 2193 2194 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2195 if (processLoad(LI)) 2196 return true; 2197 2198 unsigned Num = VN.lookup_or_add(LI); 2199 addToLeaderTable(Num, LI, LI->getParent()); 2200 return false; 2201 } 2202 2203 // For conditional branches, we can perform simple conditional propagation on 2204 // the condition value itself. 2205 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2206 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2207 return false; 2208 2209 Value *BranchCond = BI->getCondition(); 2210 2211 BasicBlock *TrueSucc = BI->getSuccessor(0); 2212 BasicBlock *FalseSucc = BI->getSuccessor(1); 2213 BasicBlock *Parent = BI->getParent(); 2214 bool Changed = false; 2215 2216 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT)) 2217 Changed |= propagateEquality(BranchCond, 2218 ConstantInt::getTrue(TrueSucc->getContext()), 2219 TrueSucc); 2220 2221 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT)) 2222 Changed |= propagateEquality(BranchCond, 2223 ConstantInt::getFalse(FalseSucc->getContext()), 2224 FalseSucc); 2225 2226 return Changed; 2227 } 2228 2229 // For switches, propagate the case values into the case destinations. 2230 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2231 Value *SwitchCond = SI->getCondition(); 2232 BasicBlock *Parent = SI->getParent(); 2233 bool Changed = false; 2234 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2235 i != e; ++i) { 2236 BasicBlock *Dst = i.getCaseSuccessor(); 2237 if (isOnlyReachableViaThisEdge(Parent, Dst, DT)) 2238 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), Dst); 2239 } 2240 return Changed; 2241 } 2242 2243 // Instructions with void type don't return a value, so there's 2244 // no point in trying to find redundancies in them. 2245 if (I->getType()->isVoidTy()) return false; 2246 2247 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2248 unsigned Num = VN.lookup_or_add(I); 2249 2250 // Allocations are always uniquely numbered, so we can save time and memory 2251 // by fast failing them. 2252 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2253 addToLeaderTable(Num, I, I->getParent()); 2254 return false; 2255 } 2256 2257 // If the number we were assigned was a brand new VN, then we don't 2258 // need to do a lookup to see if the number already exists 2259 // somewhere in the domtree: it can't! 2260 if (Num >= NextNum) { 2261 addToLeaderTable(Num, I, I->getParent()); 2262 return false; 2263 } 2264 2265 // Perform fast-path value-number based elimination of values inherited from 2266 // dominators. 2267 Value *repl = findLeader(I->getParent(), Num); 2268 if (repl == 0) { 2269 // Failure, just remember this instance for future use. 2270 addToLeaderTable(Num, I, I->getParent()); 2271 return false; 2272 } 2273 2274 // Remove it! 2275 patchAndReplaceAllUsesWith(repl, I); 2276 if (MD && repl->getType()->isPointerTy()) 2277 MD->invalidateCachedPointerInfo(repl); 2278 markInstructionForDeletion(I); 2279 return true; 2280} 2281 2282/// runOnFunction - This is the main transformation entry point for a function. 2283bool GVN::runOnFunction(Function& F) { 2284 if (!NoLoads) 2285 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2286 DT = &getAnalysis<DominatorTree>(); 2287 TD = getAnalysisIfAvailable<TargetData>(); 2288 TLI = &getAnalysis<TargetLibraryInfo>(); 2289 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2290 VN.setMemDep(MD); 2291 VN.setDomTree(DT); 2292 2293 bool Changed = false; 2294 bool ShouldContinue = true; 2295 2296 // Merge unconditional branches, allowing PRE to catch more 2297 // optimization opportunities. 2298 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2299 BasicBlock *BB = FI++; 2300 2301 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2302 if (removedBlock) ++NumGVNBlocks; 2303 2304 Changed |= removedBlock; 2305 } 2306 2307 unsigned Iteration = 0; 2308 while (ShouldContinue) { 2309 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2310 ShouldContinue = iterateOnFunction(F); 2311 if (splitCriticalEdges()) 2312 ShouldContinue = true; 2313 Changed |= ShouldContinue; 2314 ++Iteration; 2315 } 2316 2317 if (EnablePRE) { 2318 bool PREChanged = true; 2319 while (PREChanged) { 2320 PREChanged = performPRE(F); 2321 Changed |= PREChanged; 2322 } 2323 } 2324 // FIXME: Should perform GVN again after PRE does something. PRE can move 2325 // computations into blocks where they become fully redundant. Note that 2326 // we can't do this until PRE's critical edge splitting updates memdep. 2327 // Actually, when this happens, we should just fully integrate PRE into GVN. 2328 2329 cleanupGlobalSets(); 2330 2331 return Changed; 2332} 2333 2334 2335bool GVN::processBlock(BasicBlock *BB) { 2336 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2337 // (and incrementing BI before processing an instruction). 2338 assert(InstrsToErase.empty() && 2339 "We expect InstrsToErase to be empty across iterations"); 2340 bool ChangedFunction = false; 2341 2342 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2343 BI != BE;) { 2344 ChangedFunction |= processInstruction(BI); 2345 if (InstrsToErase.empty()) { 2346 ++BI; 2347 continue; 2348 } 2349 2350 // If we need some instructions deleted, do it now. 2351 NumGVNInstr += InstrsToErase.size(); 2352 2353 // Avoid iterator invalidation. 2354 bool AtStart = BI == BB->begin(); 2355 if (!AtStart) 2356 --BI; 2357 2358 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2359 E = InstrsToErase.end(); I != E; ++I) { 2360 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2361 if (MD) MD->removeInstruction(*I); 2362 (*I)->eraseFromParent(); 2363 DEBUG(verifyRemoved(*I)); 2364 } 2365 InstrsToErase.clear(); 2366 2367 if (AtStart) 2368 BI = BB->begin(); 2369 else 2370 ++BI; 2371 } 2372 2373 return ChangedFunction; 2374} 2375 2376/// performPRE - Perform a purely local form of PRE that looks for diamond 2377/// control flow patterns and attempts to perform simple PRE at the join point. 2378bool GVN::performPRE(Function &F) { 2379 bool Changed = false; 2380 DenseMap<BasicBlock*, Value*> predMap; 2381 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2382 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2383 BasicBlock *CurrentBlock = *DI; 2384 2385 // Nothing to PRE in the entry block. 2386 if (CurrentBlock == &F.getEntryBlock()) continue; 2387 2388 // Don't perform PRE on a landing pad. 2389 if (CurrentBlock->isLandingPad()) continue; 2390 2391 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2392 BE = CurrentBlock->end(); BI != BE; ) { 2393 Instruction *CurInst = BI++; 2394 2395 if (isa<AllocaInst>(CurInst) || 2396 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2397 CurInst->getType()->isVoidTy() || 2398 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2399 isa<DbgInfoIntrinsic>(CurInst)) 2400 continue; 2401 2402 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2403 // sinking the compare again, and it would force the code generator to 2404 // move the i1 from processor flags or predicate registers into a general 2405 // purpose register. 2406 if (isa<CmpInst>(CurInst)) 2407 continue; 2408 2409 // We don't currently value number ANY inline asm calls. 2410 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2411 if (CallI->isInlineAsm()) 2412 continue; 2413 2414 uint32_t ValNo = VN.lookup(CurInst); 2415 2416 // Look for the predecessors for PRE opportunities. We're 2417 // only trying to solve the basic diamond case, where 2418 // a value is computed in the successor and one predecessor, 2419 // but not the other. We also explicitly disallow cases 2420 // where the successor is its own predecessor, because they're 2421 // more complicated to get right. 2422 unsigned NumWith = 0; 2423 unsigned NumWithout = 0; 2424 BasicBlock *PREPred = 0; 2425 predMap.clear(); 2426 2427 for (pred_iterator PI = pred_begin(CurrentBlock), 2428 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2429 BasicBlock *P = *PI; 2430 // We're not interested in PRE where the block is its 2431 // own predecessor, or in blocks with predecessors 2432 // that are not reachable. 2433 if (P == CurrentBlock) { 2434 NumWithout = 2; 2435 break; 2436 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2437 NumWithout = 2; 2438 break; 2439 } 2440 2441 Value* predV = findLeader(P, ValNo); 2442 if (predV == 0) { 2443 PREPred = P; 2444 ++NumWithout; 2445 } else if (predV == CurInst) { 2446 NumWithout = 2; 2447 } else { 2448 predMap[P] = predV; 2449 ++NumWith; 2450 } 2451 } 2452 2453 // Don't do PRE when it might increase code size, i.e. when 2454 // we would need to insert instructions in more than one pred. 2455 if (NumWithout != 1 || NumWith == 0) 2456 continue; 2457 2458 // Don't do PRE across indirect branch. 2459 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2460 continue; 2461 2462 // We can't do PRE safely on a critical edge, so instead we schedule 2463 // the edge to be split and perform the PRE the next time we iterate 2464 // on the function. 2465 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2466 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2467 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2468 continue; 2469 } 2470 2471 // Instantiate the expression in the predecessor that lacked it. 2472 // Because we are going top-down through the block, all value numbers 2473 // will be available in the predecessor by the time we need them. Any 2474 // that weren't originally present will have been instantiated earlier 2475 // in this loop. 2476 Instruction *PREInstr = CurInst->clone(); 2477 bool success = true; 2478 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2479 Value *Op = PREInstr->getOperand(i); 2480 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2481 continue; 2482 2483 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2484 PREInstr->setOperand(i, V); 2485 } else { 2486 success = false; 2487 break; 2488 } 2489 } 2490 2491 // Fail out if we encounter an operand that is not available in 2492 // the PRE predecessor. This is typically because of loads which 2493 // are not value numbered precisely. 2494 if (!success) { 2495 delete PREInstr; 2496 DEBUG(verifyRemoved(PREInstr)); 2497 continue; 2498 } 2499 2500 PREInstr->insertBefore(PREPred->getTerminator()); 2501 PREInstr->setName(CurInst->getName() + ".pre"); 2502 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2503 predMap[PREPred] = PREInstr; 2504 VN.add(PREInstr, ValNo); 2505 ++NumGVNPRE; 2506 2507 // Update the availability map to include the new instruction. 2508 addToLeaderTable(ValNo, PREInstr, PREPred); 2509 2510 // Create a PHI to make the value available in this block. 2511 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2512 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2513 CurInst->getName() + ".pre-phi", 2514 CurrentBlock->begin()); 2515 for (pred_iterator PI = PB; PI != PE; ++PI) { 2516 BasicBlock *P = *PI; 2517 Phi->addIncoming(predMap[P], P); 2518 } 2519 2520 VN.add(Phi, ValNo); 2521 addToLeaderTable(ValNo, Phi, CurrentBlock); 2522 Phi->setDebugLoc(CurInst->getDebugLoc()); 2523 CurInst->replaceAllUsesWith(Phi); 2524 if (Phi->getType()->isPointerTy()) { 2525 // Because we have added a PHI-use of the pointer value, it has now 2526 // "escaped" from alias analysis' perspective. We need to inform 2527 // AA of this. 2528 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2529 ++ii) { 2530 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2531 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2532 } 2533 2534 if (MD) 2535 MD->invalidateCachedPointerInfo(Phi); 2536 } 2537 VN.erase(CurInst); 2538 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2539 2540 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2541 if (MD) MD->removeInstruction(CurInst); 2542 CurInst->eraseFromParent(); 2543 DEBUG(verifyRemoved(CurInst)); 2544 Changed = true; 2545 } 2546 } 2547 2548 if (splitCriticalEdges()) 2549 Changed = true; 2550 2551 return Changed; 2552} 2553 2554/// splitCriticalEdges - Split critical edges found during the previous 2555/// iteration that may enable further optimization. 2556bool GVN::splitCriticalEdges() { 2557 if (toSplit.empty()) 2558 return false; 2559 do { 2560 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2561 SplitCriticalEdge(Edge.first, Edge.second, this); 2562 } while (!toSplit.empty()); 2563 if (MD) MD->invalidateCachedPredecessors(); 2564 return true; 2565} 2566 2567/// iterateOnFunction - Executes one iteration of GVN 2568bool GVN::iterateOnFunction(Function &F) { 2569 cleanupGlobalSets(); 2570 2571 // Top-down walk of the dominator tree 2572 bool Changed = false; 2573#if 0 2574 // Needed for value numbering with phi construction to work. 2575 ReversePostOrderTraversal<Function*> RPOT(&F); 2576 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2577 RE = RPOT.end(); RI != RE; ++RI) 2578 Changed |= processBlock(*RI); 2579#else 2580 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2581 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2582 Changed |= processBlock(DI->getBlock()); 2583#endif 2584 2585 return Changed; 2586} 2587 2588void GVN::cleanupGlobalSets() { 2589 VN.clear(); 2590 LeaderTable.clear(); 2591 TableAllocator.Reset(); 2592} 2593 2594/// verifyRemoved - Verify that the specified instruction does not occur in our 2595/// internal data structures. 2596void GVN::verifyRemoved(const Instruction *Inst) const { 2597 VN.verifyRemoved(Inst); 2598 2599 // Walk through the value number scope to make sure the instruction isn't 2600 // ferreted away in it. 2601 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2602 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2603 const LeaderTableEntry *Node = &I->second; 2604 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2605 2606 while (Node->Next) { 2607 Node = Node->Next; 2608 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2609 } 2610 } 2611} 2612