GVN.cpp revision 968d689ec30a0df63d252b8193664e01944edb8b
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/ADT/DenseMap.h" 21#include "llvm/ADT/DepthFirstIterator.h" 22#include "llvm/ADT/Hashing.h" 23#include "llvm/ADT/SmallPtrSet.h" 24#include "llvm/ADT/Statistic.h" 25#include "llvm/Analysis/AliasAnalysis.h" 26#include "llvm/Analysis/ConstantFolding.h" 27#include "llvm/Analysis/Dominators.h" 28#include "llvm/Analysis/InstructionSimplify.h" 29#include "llvm/Analysis/Loads.h" 30#include "llvm/Analysis/MemoryBuiltins.h" 31#include "llvm/Analysis/MemoryDependenceAnalysis.h" 32#include "llvm/Analysis/PHITransAddr.h" 33#include "llvm/Analysis/ValueTracking.h" 34#include "llvm/Assembly/Writer.h" 35#include "llvm/IR/DataLayout.h" 36#include "llvm/IR/GlobalVariable.h" 37#include "llvm/IR/IRBuilder.h" 38#include "llvm/IR/IntrinsicInst.h" 39#include "llvm/IR/LLVMContext.h" 40#include "llvm/IR/Metadata.h" 41#include "llvm/Support/Allocator.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/PatternMatch.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 class GVN; 502 struct AvailableValueInBlock { 503 /// BB - The basic block in question. 504 BasicBlock *BB; 505 enum ValType { 506 SimpleVal, // A simple offsetted value that is accessed. 507 LoadVal, // A value produced by a load. 508 MemIntrin // A memory intrinsic which is loaded from. 509 }; 510 511 /// V - The value that is live out of the block. 512 PointerIntPair<Value *, 2, ValType> Val; 513 514 /// Offset - The byte offset in Val that is interesting for the load query. 515 unsigned Offset; 516 517 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 518 unsigned Offset = 0) { 519 AvailableValueInBlock Res; 520 Res.BB = BB; 521 Res.Val.setPointer(V); 522 Res.Val.setInt(SimpleVal); 523 Res.Offset = Offset; 524 return Res; 525 } 526 527 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 528 unsigned Offset = 0) { 529 AvailableValueInBlock Res; 530 Res.BB = BB; 531 Res.Val.setPointer(MI); 532 Res.Val.setInt(MemIntrin); 533 Res.Offset = Offset; 534 return Res; 535 } 536 537 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 538 unsigned Offset = 0) { 539 AvailableValueInBlock Res; 540 Res.BB = BB; 541 Res.Val.setPointer(LI); 542 Res.Val.setInt(LoadVal); 543 Res.Offset = Offset; 544 return Res; 545 } 546 547 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 548 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 549 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 550 551 Value *getSimpleValue() const { 552 assert(isSimpleValue() && "Wrong accessor"); 553 return Val.getPointer(); 554 } 555 556 LoadInst *getCoercedLoadValue() const { 557 assert(isCoercedLoadValue() && "Wrong accessor"); 558 return cast<LoadInst>(Val.getPointer()); 559 } 560 561 MemIntrinsic *getMemIntrinValue() const { 562 assert(isMemIntrinValue() && "Wrong accessor"); 563 return cast<MemIntrinsic>(Val.getPointer()); 564 } 565 566 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 567 /// defined here to the specified type. This handles various coercion cases. 568 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const; 569 }; 570 571 class GVN : public FunctionPass { 572 bool NoLoads; 573 MemoryDependenceAnalysis *MD; 574 DominatorTree *DT; 575 const DataLayout *TD; 576 const TargetLibraryInfo *TLI; 577 578 ValueTable VN; 579 580 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 581 /// have that value number. Use findLeader to query it. 582 struct LeaderTableEntry { 583 Value *Val; 584 const BasicBlock *BB; 585 LeaderTableEntry *Next; 586 }; 587 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 588 BumpPtrAllocator TableAllocator; 589 590 SmallVector<Instruction*, 8> InstrsToErase; 591 592 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect; 593 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect; 594 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect; 595 596 public: 597 static char ID; // Pass identification, replacement for typeid 598 explicit GVN(bool noloads = false) 599 : FunctionPass(ID), NoLoads(noloads), MD(0) { 600 initializeGVNPass(*PassRegistry::getPassRegistry()); 601 } 602 603 bool runOnFunction(Function &F); 604 605 /// markInstructionForDeletion - This removes the specified instruction from 606 /// our various maps and marks it for deletion. 607 void markInstructionForDeletion(Instruction *I) { 608 VN.erase(I); 609 InstrsToErase.push_back(I); 610 } 611 612 const DataLayout *getDataLayout() const { return TD; } 613 DominatorTree &getDominatorTree() const { return *DT; } 614 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 615 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 616 private: 617 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 618 /// its value number. 619 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 620 LeaderTableEntry &Curr = LeaderTable[N]; 621 if (!Curr.Val) { 622 Curr.Val = V; 623 Curr.BB = BB; 624 return; 625 } 626 627 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 628 Node->Val = V; 629 Node->BB = BB; 630 Node->Next = Curr.Next; 631 Curr.Next = Node; 632 } 633 634 /// removeFromLeaderTable - Scan the list of values corresponding to a given 635 /// value number, and remove the given instruction if encountered. 636 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 637 LeaderTableEntry* Prev = 0; 638 LeaderTableEntry* Curr = &LeaderTable[N]; 639 640 while (Curr->Val != I || Curr->BB != BB) { 641 Prev = Curr; 642 Curr = Curr->Next; 643 } 644 645 if (Prev) { 646 Prev->Next = Curr->Next; 647 } else { 648 if (!Curr->Next) { 649 Curr->Val = 0; 650 Curr->BB = 0; 651 } else { 652 LeaderTableEntry* Next = Curr->Next; 653 Curr->Val = Next->Val; 654 Curr->BB = Next->BB; 655 Curr->Next = Next->Next; 656 } 657 } 658 } 659 660 // List of critical edges to be split between iterations. 661 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 662 663 // This transformation requires dominator postdominator info 664 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 665 AU.addRequired<DominatorTree>(); 666 AU.addRequired<TargetLibraryInfo>(); 667 if (!NoLoads) 668 AU.addRequired<MemoryDependenceAnalysis>(); 669 AU.addRequired<AliasAnalysis>(); 670 671 AU.addPreserved<DominatorTree>(); 672 AU.addPreserved<AliasAnalysis>(); 673 } 674 675 676 // Helper fuctions of redundant load elimination 677 bool processLoad(LoadInst *L); 678 bool processNonLocalLoad(LoadInst *L); 679 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 680 AvailValInBlkVect &ValuesPerBlock, 681 UnavailBlkVect &UnavailableBlocks); 682 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 683 UnavailBlkVect &UnavailableBlocks); 684 685 // Other helper routines 686 bool processInstruction(Instruction *I); 687 bool processBlock(BasicBlock *BB); 688 void dump(DenseMap<uint32_t, Value*> &d); 689 bool iterateOnFunction(Function &F); 690 bool performPRE(Function &F); 691 Value *findLeader(const BasicBlock *BB, uint32_t num); 692 void cleanupGlobalSets(); 693 void verifyRemoved(const Instruction *I) const; 694 bool splitCriticalEdges(); 695 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 696 const BasicBlockEdge &Root); 697 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 698 }; 699 700 char GVN::ID = 0; 701} 702 703// createGVNPass - The public interface to this file... 704FunctionPass *llvm::createGVNPass(bool NoLoads) { 705 return new GVN(NoLoads); 706} 707 708INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 709INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 710INITIALIZE_PASS_DEPENDENCY(DominatorTree) 711INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 712INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 713INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 714 715#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 716void GVN::dump(DenseMap<uint32_t, Value*>& d) { 717 errs() << "{\n"; 718 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 719 E = d.end(); I != E; ++I) { 720 errs() << I->first << "\n"; 721 I->second->dump(); 722 } 723 errs() << "}\n"; 724} 725#endif 726 727/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 728/// we're analyzing is fully available in the specified block. As we go, keep 729/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 730/// map is actually a tri-state map with the following values: 731/// 0) we know the block *is not* fully available. 732/// 1) we know the block *is* fully available. 733/// 2) we do not know whether the block is fully available or not, but we are 734/// currently speculating that it will be. 735/// 3) we are speculating for this block and have used that to speculate for 736/// other blocks. 737static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 738 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 739 uint32_t RecurseDepth) { 740 if (RecurseDepth > MaxRecurseDepth) 741 return false; 742 743 // Optimistically assume that the block is fully available and check to see 744 // if we already know about this block in one lookup. 745 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 746 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 747 748 // If the entry already existed for this block, return the precomputed value. 749 if (!IV.second) { 750 // If this is a speculative "available" value, mark it as being used for 751 // speculation of other blocks. 752 if (IV.first->second == 2) 753 IV.first->second = 3; 754 return IV.first->second != 0; 755 } 756 757 // Otherwise, see if it is fully available in all predecessors. 758 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 759 760 // If this block has no predecessors, it isn't live-in here. 761 if (PI == PE) 762 goto SpeculationFailure; 763 764 for (; PI != PE; ++PI) 765 // If the value isn't fully available in one of our predecessors, then it 766 // isn't fully available in this block either. Undo our previous 767 // optimistic assumption and bail out. 768 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 769 goto SpeculationFailure; 770 771 return true; 772 773// SpeculationFailure - If we get here, we found out that this is not, after 774// all, a fully-available block. We have a problem if we speculated on this and 775// used the speculation to mark other blocks as available. 776SpeculationFailure: 777 char &BBVal = FullyAvailableBlocks[BB]; 778 779 // If we didn't speculate on this, just return with it set to false. 780 if (BBVal == 2) { 781 BBVal = 0; 782 return false; 783 } 784 785 // If we did speculate on this value, we could have blocks set to 1 that are 786 // incorrect. Walk the (transitive) successors of this block and mark them as 787 // 0 if set to one. 788 SmallVector<BasicBlock*, 32> BBWorklist; 789 BBWorklist.push_back(BB); 790 791 do { 792 BasicBlock *Entry = BBWorklist.pop_back_val(); 793 // Note that this sets blocks to 0 (unavailable) if they happen to not 794 // already be in FullyAvailableBlocks. This is safe. 795 char &EntryVal = FullyAvailableBlocks[Entry]; 796 if (EntryVal == 0) continue; // Already unavailable. 797 798 // Mark as unavailable. 799 EntryVal = 0; 800 801 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 802 BBWorklist.push_back(*I); 803 } while (!BBWorklist.empty()); 804 805 return false; 806} 807 808 809/// CanCoerceMustAliasedValueToLoad - Return true if 810/// CoerceAvailableValueToLoadType will succeed. 811static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 812 Type *LoadTy, 813 const DataLayout &TD) { 814 // If the loaded or stored value is an first class array or struct, don't try 815 // to transform them. We need to be able to bitcast to integer. 816 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 817 StoredVal->getType()->isStructTy() || 818 StoredVal->getType()->isArrayTy()) 819 return false; 820 821 // The store has to be at least as big as the load. 822 if (TD.getTypeSizeInBits(StoredVal->getType()) < 823 TD.getTypeSizeInBits(LoadTy)) 824 return false; 825 826 return true; 827} 828 829/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 830/// then a load from a must-aliased pointer of a different type, try to coerce 831/// the stored value. LoadedTy is the type of the load we want to replace and 832/// InsertPt is the place to insert new instructions. 833/// 834/// If we can't do it, return null. 835static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 836 Type *LoadedTy, 837 Instruction *InsertPt, 838 const DataLayout &TD) { 839 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 840 return 0; 841 842 // If this is already the right type, just return it. 843 Type *StoredValTy = StoredVal->getType(); 844 845 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 846 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 847 848 // If the store and reload are the same size, we can always reuse it. 849 if (StoreSize == LoadSize) { 850 // Pointer to Pointer -> use bitcast. 851 if (StoredValTy->getScalarType()->isPointerTy() && 852 LoadedTy->getScalarType()->isPointerTy()) 853 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 854 855 // Convert source pointers to integers, which can be bitcast. 856 if (StoredValTy->getScalarType()->isPointerTy()) { 857 StoredValTy = TD.getIntPtrType(StoredValTy); 858 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 859 } 860 861 Type *TypeToCastTo = LoadedTy; 862 if (TypeToCastTo->getScalarType()->isPointerTy()) 863 TypeToCastTo = TD.getIntPtrType(TypeToCastTo); 864 865 if (StoredValTy != TypeToCastTo) 866 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 867 868 // Cast to pointer if the load needs a pointer type. 869 if (LoadedTy->getScalarType()->isPointerTy()) 870 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 871 872 return StoredVal; 873 } 874 875 // If the loaded value is smaller than the available value, then we can 876 // extract out a piece from it. If the available value is too small, then we 877 // can't do anything. 878 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 879 880 // Convert source pointers to integers, which can be manipulated. 881 if (StoredValTy->getScalarType()->isPointerTy()) { 882 StoredValTy = TD.getIntPtrType(StoredValTy); 883 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 884 } 885 886 // Convert vectors and fp to integer, which can be manipulated. 887 if (!StoredValTy->isIntegerTy()) { 888 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 889 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 890 } 891 892 // If this is a big-endian system, we need to shift the value down to the low 893 // bits so that a truncate will work. 894 if (TD.isBigEndian()) { 895 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 896 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 897 } 898 899 // Truncate the integer to the right size now. 900 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 901 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 902 903 if (LoadedTy == NewIntTy) 904 return StoredVal; 905 906 // If the result is a pointer, inttoptr. 907 if (LoadedTy->getScalarType()->isPointerTy()) 908 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 909 910 // Otherwise, bitcast. 911 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 912} 913 914/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 915/// memdep query of a load that ends up being a clobbering memory write (store, 916/// memset, memcpy, memmove). This means that the write *may* provide bits used 917/// by the load but we can't be sure because the pointers don't mustalias. 918/// 919/// Check this case to see if there is anything more we can do before we give 920/// up. This returns -1 if we have to give up, or a byte number in the stored 921/// value of the piece that feeds the load. 922static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 923 Value *WritePtr, 924 uint64_t WriteSizeInBits, 925 const DataLayout &TD) { 926 // If the loaded or stored value is a first class array or struct, don't try 927 // to transform them. We need to be able to bitcast to integer. 928 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 929 return -1; 930 931 int64_t StoreOffset = 0, LoadOffset = 0; 932 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD); 933 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD); 934 if (StoreBase != LoadBase) 935 return -1; 936 937 // If the load and store are to the exact same address, they should have been 938 // a must alias. AA must have gotten confused. 939 // FIXME: Study to see if/when this happens. One case is forwarding a memset 940 // to a load from the base of the memset. 941#if 0 942 if (LoadOffset == StoreOffset) { 943 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 944 << "Base = " << *StoreBase << "\n" 945 << "Store Ptr = " << *WritePtr << "\n" 946 << "Store Offs = " << StoreOffset << "\n" 947 << "Load Ptr = " << *LoadPtr << "\n"; 948 abort(); 949 } 950#endif 951 952 // If the load and store don't overlap at all, the store doesn't provide 953 // anything to the load. In this case, they really don't alias at all, AA 954 // must have gotten confused. 955 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 956 957 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 958 return -1; 959 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 960 LoadSize >>= 3; 961 962 963 bool isAAFailure = false; 964 if (StoreOffset < LoadOffset) 965 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 966 else 967 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 968 969 if (isAAFailure) { 970#if 0 971 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 972 << "Base = " << *StoreBase << "\n" 973 << "Store Ptr = " << *WritePtr << "\n" 974 << "Store Offs = " << StoreOffset << "\n" 975 << "Load Ptr = " << *LoadPtr << "\n"; 976 abort(); 977#endif 978 return -1; 979 } 980 981 // If the Load isn't completely contained within the stored bits, we don't 982 // have all the bits to feed it. We could do something crazy in the future 983 // (issue a smaller load then merge the bits in) but this seems unlikely to be 984 // valuable. 985 if (StoreOffset > LoadOffset || 986 StoreOffset+StoreSize < LoadOffset+LoadSize) 987 return -1; 988 989 // Okay, we can do this transformation. Return the number of bytes into the 990 // store that the load is. 991 return LoadOffset-StoreOffset; 992} 993 994/// AnalyzeLoadFromClobberingStore - This function is called when we have a 995/// memdep query of a load that ends up being a clobbering store. 996static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 997 StoreInst *DepSI, 998 const DataLayout &TD) { 999 // Cannot handle reading from store of first-class aggregate yet. 1000 if (DepSI->getValueOperand()->getType()->isStructTy() || 1001 DepSI->getValueOperand()->getType()->isArrayTy()) 1002 return -1; 1003 1004 Value *StorePtr = DepSI->getPointerOperand(); 1005 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1006 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1007 StorePtr, StoreSize, TD); 1008} 1009 1010/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 1011/// memdep query of a load that ends up being clobbered by another load. See if 1012/// the other load can feed into the second load. 1013static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 1014 LoadInst *DepLI, const DataLayout &TD){ 1015 // Cannot handle reading from store of first-class aggregate yet. 1016 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 1017 return -1; 1018 1019 Value *DepPtr = DepLI->getPointerOperand(); 1020 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 1021 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 1022 if (R != -1) return R; 1023 1024 // If we have a load/load clobber an DepLI can be widened to cover this load, 1025 // then we should widen it! 1026 int64_t LoadOffs = 0; 1027 const Value *LoadBase = 1028 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD); 1029 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1030 1031 unsigned Size = MemoryDependenceAnalysis:: 1032 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 1033 if (Size == 0) return -1; 1034 1035 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 1036} 1037 1038 1039 1040static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 1041 MemIntrinsic *MI, 1042 const DataLayout &TD) { 1043 // If the mem operation is a non-constant size, we can't handle it. 1044 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1045 if (SizeCst == 0) return -1; 1046 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1047 1048 // If this is memset, we just need to see if the offset is valid in the size 1049 // of the memset.. 1050 if (MI->getIntrinsicID() == Intrinsic::memset) 1051 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1052 MemSizeInBits, TD); 1053 1054 // If we have a memcpy/memmove, the only case we can handle is if this is a 1055 // copy from constant memory. In that case, we can read directly from the 1056 // constant memory. 1057 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1058 1059 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1060 if (Src == 0) return -1; 1061 1062 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 1063 if (GV == 0 || !GV->isConstant()) return -1; 1064 1065 // See if the access is within the bounds of the transfer. 1066 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1067 MI->getDest(), MemSizeInBits, TD); 1068 if (Offset == -1) 1069 return Offset; 1070 1071 // Otherwise, see if we can constant fold a load from the constant with the 1072 // offset applied as appropriate. 1073 Src = ConstantExpr::getBitCast(Src, 1074 llvm::Type::getInt8PtrTy(Src->getContext())); 1075 Constant *OffsetCst = 1076 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1077 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1078 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1079 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1080 return Offset; 1081 return -1; 1082} 1083 1084 1085/// GetStoreValueForLoad - This function is called when we have a 1086/// memdep query of a load that ends up being a clobbering store. This means 1087/// that the store provides bits used by the load but we the pointers don't 1088/// mustalias. Check this case to see if there is anything more we can do 1089/// before we give up. 1090static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1091 Type *LoadTy, 1092 Instruction *InsertPt, const DataLayout &TD){ 1093 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1094 1095 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1096 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1097 1098 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1099 1100 // Compute which bits of the stored value are being used by the load. Convert 1101 // to an integer type to start with. 1102 if (SrcVal->getType()->getScalarType()->isPointerTy()) 1103 SrcVal = Builder.CreatePtrToInt(SrcVal, 1104 TD.getIntPtrType(SrcVal->getType())); 1105 if (!SrcVal->getType()->isIntegerTy()) 1106 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1107 1108 // Shift the bits to the least significant depending on endianness. 1109 unsigned ShiftAmt; 1110 if (TD.isLittleEndian()) 1111 ShiftAmt = Offset*8; 1112 else 1113 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1114 1115 if (ShiftAmt) 1116 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1117 1118 if (LoadSize != StoreSize) 1119 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1120 1121 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1122} 1123 1124/// GetLoadValueForLoad - This function is called when we have a 1125/// memdep query of a load that ends up being a clobbering load. This means 1126/// that the load *may* provide bits used by the load but we can't be sure 1127/// because the pointers don't mustalias. Check this case to see if there is 1128/// anything more we can do before we give up. 1129static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1130 Type *LoadTy, Instruction *InsertPt, 1131 GVN &gvn) { 1132 const DataLayout &TD = *gvn.getDataLayout(); 1133 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1134 // widen SrcVal out to a larger load. 1135 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1136 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1137 if (Offset+LoadSize > SrcValSize) { 1138 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1139 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1140 // If we have a load/load clobber an DepLI can be widened to cover this 1141 // load, then we should widen it to the next power of 2 size big enough! 1142 unsigned NewLoadSize = Offset+LoadSize; 1143 if (!isPowerOf2_32(NewLoadSize)) 1144 NewLoadSize = NextPowerOf2(NewLoadSize); 1145 1146 Value *PtrVal = SrcVal->getPointerOperand(); 1147 1148 // Insert the new load after the old load. This ensures that subsequent 1149 // memdep queries will find the new load. We can't easily remove the old 1150 // load completely because it is already in the value numbering table. 1151 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1152 Type *DestPTy = 1153 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1154 DestPTy = PointerType::get(DestPTy, 1155 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1156 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1157 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1158 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1159 NewLoad->takeName(SrcVal); 1160 NewLoad->setAlignment(SrcVal->getAlignment()); 1161 1162 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1163 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1164 1165 // Replace uses of the original load with the wider load. On a big endian 1166 // system, we need to shift down to get the relevant bits. 1167 Value *RV = NewLoad; 1168 if (TD.isBigEndian()) 1169 RV = Builder.CreateLShr(RV, 1170 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1171 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1172 SrcVal->replaceAllUsesWith(RV); 1173 1174 // We would like to use gvn.markInstructionForDeletion here, but we can't 1175 // because the load is already memoized into the leader map table that GVN 1176 // tracks. It is potentially possible to remove the load from the table, 1177 // but then there all of the operations based on it would need to be 1178 // rehashed. Just leave the dead load around. 1179 gvn.getMemDep().removeInstruction(SrcVal); 1180 SrcVal = NewLoad; 1181 } 1182 1183 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1184} 1185 1186 1187/// GetMemInstValueForLoad - This function is called when we have a 1188/// memdep query of a load that ends up being a clobbering mem intrinsic. 1189static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1190 Type *LoadTy, Instruction *InsertPt, 1191 const DataLayout &TD){ 1192 LLVMContext &Ctx = LoadTy->getContext(); 1193 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1194 1195 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1196 1197 // We know that this method is only called when the mem transfer fully 1198 // provides the bits for the load. 1199 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1200 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1201 // independently of what the offset is. 1202 Value *Val = MSI->getValue(); 1203 if (LoadSize != 1) 1204 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1205 1206 Value *OneElt = Val; 1207 1208 // Splat the value out to the right number of bits. 1209 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1210 // If we can double the number of bytes set, do it. 1211 if (NumBytesSet*2 <= LoadSize) { 1212 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1213 Val = Builder.CreateOr(Val, ShVal); 1214 NumBytesSet <<= 1; 1215 continue; 1216 } 1217 1218 // Otherwise insert one byte at a time. 1219 Value *ShVal = Builder.CreateShl(Val, 1*8); 1220 Val = Builder.CreateOr(OneElt, ShVal); 1221 ++NumBytesSet; 1222 } 1223 1224 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1225 } 1226 1227 // Otherwise, this is a memcpy/memmove from a constant global. 1228 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1229 Constant *Src = cast<Constant>(MTI->getSource()); 1230 1231 // Otherwise, see if we can constant fold a load from the constant with the 1232 // offset applied as appropriate. 1233 Src = ConstantExpr::getBitCast(Src, 1234 llvm::Type::getInt8PtrTy(Src->getContext())); 1235 Constant *OffsetCst = 1236 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1237 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1238 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1239 return ConstantFoldLoadFromConstPtr(Src, &TD); 1240} 1241 1242 1243/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1244/// construct SSA form, allowing us to eliminate LI. This returns the value 1245/// that should be used at LI's definition site. 1246static Value *ConstructSSAForLoadSet(LoadInst *LI, 1247 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1248 GVN &gvn) { 1249 // Check for the fully redundant, dominating load case. In this case, we can 1250 // just use the dominating value directly. 1251 if (ValuesPerBlock.size() == 1 && 1252 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1253 LI->getParent())) 1254 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1255 1256 // Otherwise, we have to construct SSA form. 1257 SmallVector<PHINode*, 8> NewPHIs; 1258 SSAUpdater SSAUpdate(&NewPHIs); 1259 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1260 1261 Type *LoadTy = LI->getType(); 1262 1263 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1264 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1265 BasicBlock *BB = AV.BB; 1266 1267 if (SSAUpdate.HasValueForBlock(BB)) 1268 continue; 1269 1270 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1271 } 1272 1273 // Perform PHI construction. 1274 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1275 1276 // If new PHI nodes were created, notify alias analysis. 1277 if (V->getType()->getScalarType()->isPointerTy()) { 1278 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1279 1280 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1281 AA->copyValue(LI, NewPHIs[i]); 1282 1283 // Now that we've copied information to the new PHIs, scan through 1284 // them again and inform alias analysis that we've added potentially 1285 // escaping uses to any values that are operands to these PHIs. 1286 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1287 PHINode *P = NewPHIs[i]; 1288 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1289 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1290 AA->addEscapingUse(P->getOperandUse(jj)); 1291 } 1292 } 1293 } 1294 1295 return V; 1296} 1297 1298Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1299 Value *Res; 1300 if (isSimpleValue()) { 1301 Res = getSimpleValue(); 1302 if (Res->getType() != LoadTy) { 1303 const DataLayout *TD = gvn.getDataLayout(); 1304 assert(TD && "Need target data to handle type mismatch case"); 1305 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1306 *TD); 1307 1308 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1309 << *getSimpleValue() << '\n' 1310 << *Res << '\n' << "\n\n\n"); 1311 } 1312 } else if (isCoercedLoadValue()) { 1313 LoadInst *Load = getCoercedLoadValue(); 1314 if (Load->getType() == LoadTy && Offset == 0) { 1315 Res = Load; 1316 } else { 1317 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1318 gvn); 1319 1320 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1321 << *getCoercedLoadValue() << '\n' 1322 << *Res << '\n' << "\n\n\n"); 1323 } 1324 } else { 1325 const DataLayout *TD = gvn.getDataLayout(); 1326 assert(TD && "Need target data to handle type mismatch case"); 1327 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1328 LoadTy, BB->getTerminator(), *TD); 1329 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1330 << " " << *getMemIntrinValue() << '\n' 1331 << *Res << '\n' << "\n\n\n"); 1332 } 1333 return Res; 1334} 1335 1336static bool isLifetimeStart(const Instruction *Inst) { 1337 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1338 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1339 return false; 1340} 1341 1342void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 1343 AvailValInBlkVect &ValuesPerBlock, 1344 UnavailBlkVect &UnavailableBlocks) { 1345 1346 // Filter out useless results (non-locals, etc). Keep track of the blocks 1347 // where we have a value available in repl, also keep track of whether we see 1348 // dependencies that produce an unknown value for the load (such as a call 1349 // that could potentially clobber the load). 1350 unsigned NumDeps = Deps.size(); 1351 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1352 BasicBlock *DepBB = Deps[i].getBB(); 1353 MemDepResult DepInfo = Deps[i].getResult(); 1354 1355 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1356 UnavailableBlocks.push_back(DepBB); 1357 continue; 1358 } 1359 1360 if (DepInfo.isClobber()) { 1361 // The address being loaded in this non-local block may not be the same as 1362 // the pointer operand of the load if PHI translation occurs. Make sure 1363 // to consider the right address. 1364 Value *Address = Deps[i].getAddress(); 1365 1366 // If the dependence is to a store that writes to a superset of the bits 1367 // read by the load, we can extract the bits we need for the load from the 1368 // stored value. 1369 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1370 if (TD && Address) { 1371 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1372 DepSI, *TD); 1373 if (Offset != -1) { 1374 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1375 DepSI->getValueOperand(), 1376 Offset)); 1377 continue; 1378 } 1379 } 1380 } 1381 1382 // Check to see if we have something like this: 1383 // load i32* P 1384 // load i8* (P+1) 1385 // if we have this, replace the later with an extraction from the former. 1386 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1387 // If this is a clobber and L is the first instruction in its block, then 1388 // we have the first instruction in the entry block. 1389 if (DepLI != LI && Address && TD) { 1390 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1391 LI->getPointerOperand(), 1392 DepLI, *TD); 1393 1394 if (Offset != -1) { 1395 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1396 Offset)); 1397 continue; 1398 } 1399 } 1400 } 1401 1402 // If the clobbering value is a memset/memcpy/memmove, see if we can 1403 // forward a value on from it. 1404 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1405 if (TD && Address) { 1406 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1407 DepMI, *TD); 1408 if (Offset != -1) { 1409 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1410 Offset)); 1411 continue; 1412 } 1413 } 1414 } 1415 1416 UnavailableBlocks.push_back(DepBB); 1417 continue; 1418 } 1419 1420 // DepInfo.isDef() here 1421 1422 Instruction *DepInst = DepInfo.getInst(); 1423 1424 // Loading the allocation -> undef. 1425 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1426 // Loading immediately after lifetime begin -> undef. 1427 isLifetimeStart(DepInst)) { 1428 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1429 UndefValue::get(LI->getType()))); 1430 continue; 1431 } 1432 1433 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1434 // Reject loads and stores that are to the same address but are of 1435 // different types if we have to. 1436 if (S->getValueOperand()->getType() != LI->getType()) { 1437 // If the stored value is larger or equal to the loaded value, we can 1438 // reuse it. 1439 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1440 LI->getType(), *TD)) { 1441 UnavailableBlocks.push_back(DepBB); 1442 continue; 1443 } 1444 } 1445 1446 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1447 S->getValueOperand())); 1448 continue; 1449 } 1450 1451 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1452 // If the types mismatch and we can't handle it, reject reuse of the load. 1453 if (LD->getType() != LI->getType()) { 1454 // If the stored value is larger or equal to the loaded value, we can 1455 // reuse it. 1456 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1457 UnavailableBlocks.push_back(DepBB); 1458 continue; 1459 } 1460 } 1461 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1462 continue; 1463 } 1464 1465 UnavailableBlocks.push_back(DepBB); 1466 } 1467} 1468 1469bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 1470 UnavailBlkVect &UnavailableBlocks) { 1471 // Okay, we have *some* definitions of the value. This means that the value 1472 // is available in some of our (transitive) predecessors. Lets think about 1473 // doing PRE of this load. This will involve inserting a new load into the 1474 // predecessor when it's not available. We could do this in general, but 1475 // prefer to not increase code size. As such, we only do this when we know 1476 // that we only have to insert *one* load (which means we're basically moving 1477 // the load, not inserting a new one). 1478 1479 SmallPtrSet<BasicBlock *, 4> Blockers; 1480 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1481 Blockers.insert(UnavailableBlocks[i]); 1482 1483 // Let's find the first basic block with more than one predecessor. Walk 1484 // backwards through predecessors if needed. 1485 BasicBlock *LoadBB = LI->getParent(); 1486 BasicBlock *TmpBB = LoadBB; 1487 1488 while (TmpBB->getSinglePredecessor()) { 1489 TmpBB = TmpBB->getSinglePredecessor(); 1490 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1491 return false; 1492 if (Blockers.count(TmpBB)) 1493 return false; 1494 1495 // If any of these blocks has more than one successor (i.e. if the edge we 1496 // just traversed was critical), then there are other paths through this 1497 // block along which the load may not be anticipated. Hoisting the load 1498 // above this block would be adding the load to execution paths along 1499 // which it was not previously executed. 1500 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1501 return false; 1502 } 1503 1504 assert(TmpBB); 1505 LoadBB = TmpBB; 1506 1507 // Check to see how many predecessors have the loaded value fully 1508 // available. 1509 DenseMap<BasicBlock*, Value*> PredLoads; 1510 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1511 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1512 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1513 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1514 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1515 1516 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1517 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1518 PI != E; ++PI) { 1519 BasicBlock *Pred = *PI; 1520 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1521 continue; 1522 } 1523 PredLoads[Pred] = 0; 1524 1525 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1526 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1527 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1528 << Pred->getName() << "': " << *LI << '\n'); 1529 return false; 1530 } 1531 1532 if (LoadBB->isLandingPad()) { 1533 DEBUG(dbgs() 1534 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1535 << Pred->getName() << "': " << *LI << '\n'); 1536 return false; 1537 } 1538 1539 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1540 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1541 } 1542 } 1543 1544 if (!NeedToSplit.empty()) { 1545 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1546 return false; 1547 } 1548 1549 // Decide whether PRE is profitable for this load. 1550 unsigned NumUnavailablePreds = PredLoads.size(); 1551 assert(NumUnavailablePreds != 0 && 1552 "Fully available value should be eliminated above!"); 1553 1554 // If this load is unavailable in multiple predecessors, reject it. 1555 // FIXME: If we could restructure the CFG, we could make a common pred with 1556 // all the preds that don't have an available LI and insert a new load into 1557 // that one block. 1558 if (NumUnavailablePreds != 1) 1559 return false; 1560 1561 // Check if the load can safely be moved to all the unavailable predecessors. 1562 bool CanDoPRE = true; 1563 SmallVector<Instruction*, 8> NewInsts; 1564 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1565 E = PredLoads.end(); I != E; ++I) { 1566 BasicBlock *UnavailablePred = I->first; 1567 1568 // Do PHI translation to get its value in the predecessor if necessary. The 1569 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1570 1571 // If all preds have a single successor, then we know it is safe to insert 1572 // the load on the pred (?!?), so we can insert code to materialize the 1573 // pointer if it is not available. 1574 PHITransAddr Address(LI->getPointerOperand(), TD); 1575 Value *LoadPtr = 0; 1576 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1577 *DT, NewInsts); 1578 1579 // If we couldn't find or insert a computation of this phi translated value, 1580 // we fail PRE. 1581 if (LoadPtr == 0) { 1582 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1583 << *LI->getPointerOperand() << "\n"); 1584 CanDoPRE = false; 1585 break; 1586 } 1587 1588 I->second = LoadPtr; 1589 } 1590 1591 if (!CanDoPRE) { 1592 while (!NewInsts.empty()) { 1593 Instruction *I = NewInsts.pop_back_val(); 1594 if (MD) MD->removeInstruction(I); 1595 I->eraseFromParent(); 1596 } 1597 return false; 1598 } 1599 1600 // Okay, we can eliminate this load by inserting a reload in the predecessor 1601 // and using PHI construction to get the value in the other predecessors, do 1602 // it. 1603 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1604 DEBUG(if (!NewInsts.empty()) 1605 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1606 << *NewInsts.back() << '\n'); 1607 1608 // Assign value numbers to the new instructions. 1609 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1610 // FIXME: We really _ought_ to insert these value numbers into their 1611 // parent's availability map. However, in doing so, we risk getting into 1612 // ordering issues. If a block hasn't been processed yet, we would be 1613 // marking a value as AVAIL-IN, which isn't what we intend. 1614 VN.lookup_or_add(NewInsts[i]); 1615 } 1616 1617 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1618 E = PredLoads.end(); I != E; ++I) { 1619 BasicBlock *UnavailablePred = I->first; 1620 Value *LoadPtr = I->second; 1621 1622 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1623 LI->getAlignment(), 1624 UnavailablePred->getTerminator()); 1625 1626 // Transfer the old load's TBAA tag to the new load. 1627 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1628 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1629 1630 // Transfer DebugLoc. 1631 NewLoad->setDebugLoc(LI->getDebugLoc()); 1632 1633 // Add the newly created load. 1634 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1635 NewLoad)); 1636 MD->invalidateCachedPointerInfo(LoadPtr); 1637 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1638 } 1639 1640 // Perform PHI construction. 1641 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1642 LI->replaceAllUsesWith(V); 1643 if (isa<PHINode>(V)) 1644 V->takeName(LI); 1645 if (V->getType()->getScalarType()->isPointerTy()) 1646 MD->invalidateCachedPointerInfo(V); 1647 markInstructionForDeletion(LI); 1648 ++NumPRELoad; 1649 return true; 1650} 1651 1652/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1653/// non-local by performing PHI construction. 1654bool GVN::processNonLocalLoad(LoadInst *LI) { 1655 // Step 1: Find the non-local dependencies of the load. 1656 LoadDepVect Deps; 1657 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1658 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1659 1660 // If we had to process more than one hundred blocks to find the 1661 // dependencies, this load isn't worth worrying about. Optimizing 1662 // it will be too expensive. 1663 unsigned NumDeps = Deps.size(); 1664 if (NumDeps > 100) 1665 return false; 1666 1667 // If we had a phi translation failure, we'll have a single entry which is a 1668 // clobber in the current block. Reject this early. 1669 if (NumDeps == 1 && 1670 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1671 DEBUG( 1672 dbgs() << "GVN: non-local load "; 1673 WriteAsOperand(dbgs(), LI); 1674 dbgs() << " has unknown dependencies\n"; 1675 ); 1676 return false; 1677 } 1678 1679 // Step 2: Analyze the availability of the load 1680 AvailValInBlkVect ValuesPerBlock; 1681 UnavailBlkVect UnavailableBlocks; 1682 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks); 1683 1684 // If we have no predecessors that produce a known value for this load, exit 1685 // early. 1686 if (ValuesPerBlock.empty()) 1687 return false; 1688 1689 // Step 3: Eliminate fully redundancy. 1690 // 1691 // If all of the instructions we depend on produce a known value for this 1692 // load, then it is fully redundant and we can use PHI insertion to compute 1693 // its value. Insert PHIs and remove the fully redundant value now. 1694 if (UnavailableBlocks.empty()) { 1695 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1696 1697 // Perform PHI construction. 1698 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1699 LI->replaceAllUsesWith(V); 1700 1701 if (isa<PHINode>(V)) 1702 V->takeName(LI); 1703 if (V->getType()->getScalarType()->isPointerTy()) 1704 MD->invalidateCachedPointerInfo(V); 1705 markInstructionForDeletion(LI); 1706 ++NumGVNLoad; 1707 return true; 1708 } 1709 1710 // Step 4: Eliminate partial redundancy. 1711 if (!EnablePRE || !EnableLoadPRE) 1712 return false; 1713 1714 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks); 1715} 1716 1717 1718static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1719 // Patch the replacement so that it is not more restrictive than the value 1720 // being replaced. 1721 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1722 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1723 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1724 isa<OverflowingBinaryOperator>(ReplOp)) { 1725 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1726 ReplOp->setHasNoSignedWrap(false); 1727 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1728 ReplOp->setHasNoUnsignedWrap(false); 1729 } 1730 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1731 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1732 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1733 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1734 unsigned Kind = Metadata[i].first; 1735 MDNode *IMD = I->getMetadata(Kind); 1736 MDNode *ReplMD = Metadata[i].second; 1737 switch(Kind) { 1738 default: 1739 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1740 break; 1741 case LLVMContext::MD_dbg: 1742 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1743 case LLVMContext::MD_tbaa: 1744 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1745 break; 1746 case LLVMContext::MD_range: 1747 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1748 break; 1749 case LLVMContext::MD_prof: 1750 llvm_unreachable("MD_prof in a non terminator instruction"); 1751 break; 1752 case LLVMContext::MD_fpmath: 1753 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1754 break; 1755 } 1756 } 1757 } 1758} 1759 1760static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1761 patchReplacementInstruction(I, Repl); 1762 I->replaceAllUsesWith(Repl); 1763} 1764 1765/// processLoad - Attempt to eliminate a load, first by eliminating it 1766/// locally, and then attempting non-local elimination if that fails. 1767bool GVN::processLoad(LoadInst *L) { 1768 if (!MD) 1769 return false; 1770 1771 if (!L->isSimple()) 1772 return false; 1773 1774 if (L->use_empty()) { 1775 markInstructionForDeletion(L); 1776 return true; 1777 } 1778 1779 // ... to a pointer that has been loaded from before... 1780 MemDepResult Dep = MD->getDependency(L); 1781 1782 // If we have a clobber and target data is around, see if this is a clobber 1783 // that we can fix up through code synthesis. 1784 if (Dep.isClobber() && TD) { 1785 // Check to see if we have something like this: 1786 // store i32 123, i32* %P 1787 // %A = bitcast i32* %P to i8* 1788 // %B = gep i8* %A, i32 1 1789 // %C = load i8* %B 1790 // 1791 // We could do that by recognizing if the clobber instructions are obviously 1792 // a common base + constant offset, and if the previous store (or memset) 1793 // completely covers this load. This sort of thing can happen in bitfield 1794 // access code. 1795 Value *AvailVal = 0; 1796 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1797 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1798 L->getPointerOperand(), 1799 DepSI, *TD); 1800 if (Offset != -1) 1801 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1802 L->getType(), L, *TD); 1803 } 1804 1805 // Check to see if we have something like this: 1806 // load i32* P 1807 // load i8* (P+1) 1808 // if we have this, replace the later with an extraction from the former. 1809 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1810 // If this is a clobber and L is the first instruction in its block, then 1811 // we have the first instruction in the entry block. 1812 if (DepLI == L) 1813 return false; 1814 1815 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1816 L->getPointerOperand(), 1817 DepLI, *TD); 1818 if (Offset != -1) 1819 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1820 } 1821 1822 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1823 // a value on from it. 1824 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1825 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1826 L->getPointerOperand(), 1827 DepMI, *TD); 1828 if (Offset != -1) 1829 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1830 } 1831 1832 if (AvailVal) { 1833 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1834 << *AvailVal << '\n' << *L << "\n\n\n"); 1835 1836 // Replace the load! 1837 L->replaceAllUsesWith(AvailVal); 1838 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1839 MD->invalidateCachedPointerInfo(AvailVal); 1840 markInstructionForDeletion(L); 1841 ++NumGVNLoad; 1842 return true; 1843 } 1844 } 1845 1846 // If the value isn't available, don't do anything! 1847 if (Dep.isClobber()) { 1848 DEBUG( 1849 // fast print dep, using operator<< on instruction is too slow. 1850 dbgs() << "GVN: load "; 1851 WriteAsOperand(dbgs(), L); 1852 Instruction *I = Dep.getInst(); 1853 dbgs() << " is clobbered by " << *I << '\n'; 1854 ); 1855 return false; 1856 } 1857 1858 // If it is defined in another block, try harder. 1859 if (Dep.isNonLocal()) 1860 return processNonLocalLoad(L); 1861 1862 if (!Dep.isDef()) { 1863 DEBUG( 1864 // fast print dep, using operator<< on instruction is too slow. 1865 dbgs() << "GVN: load "; 1866 WriteAsOperand(dbgs(), L); 1867 dbgs() << " has unknown dependence\n"; 1868 ); 1869 return false; 1870 } 1871 1872 Instruction *DepInst = Dep.getInst(); 1873 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1874 Value *StoredVal = DepSI->getValueOperand(); 1875 1876 // The store and load are to a must-aliased pointer, but they may not 1877 // actually have the same type. See if we know how to reuse the stored 1878 // value (depending on its type). 1879 if (StoredVal->getType() != L->getType()) { 1880 if (TD) { 1881 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1882 L, *TD); 1883 if (StoredVal == 0) 1884 return false; 1885 1886 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1887 << '\n' << *L << "\n\n\n"); 1888 } 1889 else 1890 return false; 1891 } 1892 1893 // Remove it! 1894 L->replaceAllUsesWith(StoredVal); 1895 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1896 MD->invalidateCachedPointerInfo(StoredVal); 1897 markInstructionForDeletion(L); 1898 ++NumGVNLoad; 1899 return true; 1900 } 1901 1902 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1903 Value *AvailableVal = DepLI; 1904 1905 // The loads are of a must-aliased pointer, but they may not actually have 1906 // the same type. See if we know how to reuse the previously loaded value 1907 // (depending on its type). 1908 if (DepLI->getType() != L->getType()) { 1909 if (TD) { 1910 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1911 L, *TD); 1912 if (AvailableVal == 0) 1913 return false; 1914 1915 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1916 << "\n" << *L << "\n\n\n"); 1917 } 1918 else 1919 return false; 1920 } 1921 1922 // Remove it! 1923 patchAndReplaceAllUsesWith(L, AvailableVal); 1924 if (DepLI->getType()->getScalarType()->isPointerTy()) 1925 MD->invalidateCachedPointerInfo(DepLI); 1926 markInstructionForDeletion(L); 1927 ++NumGVNLoad; 1928 return true; 1929 } 1930 1931 // If this load really doesn't depend on anything, then we must be loading an 1932 // undef value. This can happen when loading for a fresh allocation with no 1933 // intervening stores, for example. 1934 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1935 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1936 markInstructionForDeletion(L); 1937 ++NumGVNLoad; 1938 return true; 1939 } 1940 1941 // If this load occurs either right after a lifetime begin, 1942 // then the loaded value is undefined. 1943 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1944 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1945 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1946 markInstructionForDeletion(L); 1947 ++NumGVNLoad; 1948 return true; 1949 } 1950 } 1951 1952 return false; 1953} 1954 1955// findLeader - In order to find a leader for a given value number at a 1956// specific basic block, we first obtain the list of all Values for that number, 1957// and then scan the list to find one whose block dominates the block in 1958// question. This is fast because dominator tree queries consist of only 1959// a few comparisons of DFS numbers. 1960Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 1961 LeaderTableEntry Vals = LeaderTable[num]; 1962 if (!Vals.Val) return 0; 1963 1964 Value *Val = 0; 1965 if (DT->dominates(Vals.BB, BB)) { 1966 Val = Vals.Val; 1967 if (isa<Constant>(Val)) return Val; 1968 } 1969 1970 LeaderTableEntry* Next = Vals.Next; 1971 while (Next) { 1972 if (DT->dominates(Next->BB, BB)) { 1973 if (isa<Constant>(Next->Val)) return Next->Val; 1974 if (!Val) Val = Next->Val; 1975 } 1976 1977 Next = Next->Next; 1978 } 1979 1980 return Val; 1981} 1982 1983/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1984/// use is dominated by the given basic block. Returns the number of uses that 1985/// were replaced. 1986unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1987 const BasicBlockEdge &Root) { 1988 unsigned Count = 0; 1989 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1990 UI != UE; ) { 1991 Use &U = (UI++).getUse(); 1992 1993 if (DT->dominates(Root, U)) { 1994 U.set(To); 1995 ++Count; 1996 } 1997 } 1998 return Count; 1999} 2000 2001/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2002/// true if every path from the entry block to 'Dst' passes via this edge. In 2003/// particular 'Dst' must not be reachable via another edge from 'Src'. 2004static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2005 DominatorTree *DT) { 2006 // While in theory it is interesting to consider the case in which Dst has 2007 // more than one predecessor, because Dst might be part of a loop which is 2008 // only reachable from Src, in practice it is pointless since at the time 2009 // GVN runs all such loops have preheaders, which means that Dst will have 2010 // been changed to have only one predecessor, namely Src. 2011 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2012 const BasicBlock *Src = E.getStart(); 2013 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2014 (void)Src; 2015 return Pred != 0; 2016} 2017 2018/// propagateEquality - The given values are known to be equal in every block 2019/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2020/// 'RHS' everywhere in the scope. Returns whether a change was made. 2021bool GVN::propagateEquality(Value *LHS, Value *RHS, 2022 const BasicBlockEdge &Root) { 2023 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2024 Worklist.push_back(std::make_pair(LHS, RHS)); 2025 bool Changed = false; 2026 // For speed, compute a conservative fast approximation to 2027 // DT->dominates(Root, Root.getEnd()); 2028 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2029 2030 while (!Worklist.empty()) { 2031 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2032 LHS = Item.first; RHS = Item.second; 2033 2034 if (LHS == RHS) continue; 2035 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2036 2037 // Don't try to propagate equalities between constants. 2038 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2039 2040 // Prefer a constant on the right-hand side, or an Argument if no constants. 2041 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2042 std::swap(LHS, RHS); 2043 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2044 2045 // If there is no obvious reason to prefer the left-hand side over the right- 2046 // hand side, ensure the longest lived term is on the right-hand side, so the 2047 // shortest lived term will be replaced by the longest lived. This tends to 2048 // expose more simplifications. 2049 uint32_t LVN = VN.lookup_or_add(LHS); 2050 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2051 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2052 // Move the 'oldest' value to the right-hand side, using the value number as 2053 // a proxy for age. 2054 uint32_t RVN = VN.lookup_or_add(RHS); 2055 if (LVN < RVN) { 2056 std::swap(LHS, RHS); 2057 LVN = RVN; 2058 } 2059 } 2060 2061 // If value numbering later sees that an instruction in the scope is equal 2062 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2063 // the invariant that instructions only occur in the leader table for their 2064 // own value number (this is used by removeFromLeaderTable), do not do this 2065 // if RHS is an instruction (if an instruction in the scope is morphed into 2066 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2067 // using the leader table is about compiling faster, not optimizing better). 2068 // The leader table only tracks basic blocks, not edges. Only add to if we 2069 // have the simple case where the edge dominates the end. 2070 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2071 addToLeaderTable(LVN, RHS, Root.getEnd()); 2072 2073 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2074 // LHS always has at least one use that is not dominated by Root, this will 2075 // never do anything if LHS has only one use. 2076 if (!LHS->hasOneUse()) { 2077 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2078 Changed |= NumReplacements > 0; 2079 NumGVNEqProp += NumReplacements; 2080 } 2081 2082 // Now try to deduce additional equalities from this one. For example, if the 2083 // known equality was "(A != B)" == "false" then it follows that A and B are 2084 // equal in the scope. Only boolean equalities with an explicit true or false 2085 // RHS are currently supported. 2086 if (!RHS->getType()->isIntegerTy(1)) 2087 // Not a boolean equality - bail out. 2088 continue; 2089 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2090 if (!CI) 2091 // RHS neither 'true' nor 'false' - bail out. 2092 continue; 2093 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2094 bool isKnownTrue = CI->isAllOnesValue(); 2095 bool isKnownFalse = !isKnownTrue; 2096 2097 // If "A && B" is known true then both A and B are known true. If "A || B" 2098 // is known false then both A and B are known false. 2099 Value *A, *B; 2100 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2101 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2102 Worklist.push_back(std::make_pair(A, RHS)); 2103 Worklist.push_back(std::make_pair(B, RHS)); 2104 continue; 2105 } 2106 2107 // If we are propagating an equality like "(A == B)" == "true" then also 2108 // propagate the equality A == B. When propagating a comparison such as 2109 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2110 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2111 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2112 2113 // If "A == B" is known true, or "A != B" is known false, then replace 2114 // A with B everywhere in the scope. 2115 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2116 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2117 Worklist.push_back(std::make_pair(Op0, Op1)); 2118 2119 // If "A >= B" is known true, replace "A < B" with false everywhere. 2120 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2121 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2122 // Since we don't have the instruction "A < B" immediately to hand, work out 2123 // the value number that it would have and use that to find an appropriate 2124 // instruction (if any). 2125 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2126 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2127 // If the number we were assigned was brand new then there is no point in 2128 // looking for an instruction realizing it: there cannot be one! 2129 if (Num < NextNum) { 2130 Value *NotCmp = findLeader(Root.getEnd(), Num); 2131 if (NotCmp && isa<Instruction>(NotCmp)) { 2132 unsigned NumReplacements = 2133 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2134 Changed |= NumReplacements > 0; 2135 NumGVNEqProp += NumReplacements; 2136 } 2137 } 2138 // Ensure that any instruction in scope that gets the "A < B" value number 2139 // is replaced with false. 2140 // The leader table only tracks basic blocks, not edges. Only add to if we 2141 // have the simple case where the edge dominates the end. 2142 if (RootDominatesEnd) 2143 addToLeaderTable(Num, NotVal, Root.getEnd()); 2144 2145 continue; 2146 } 2147 } 2148 2149 return Changed; 2150} 2151 2152/// processInstruction - When calculating availability, handle an instruction 2153/// by inserting it into the appropriate sets 2154bool GVN::processInstruction(Instruction *I) { 2155 // Ignore dbg info intrinsics. 2156 if (isa<DbgInfoIntrinsic>(I)) 2157 return false; 2158 2159 // If the instruction can be easily simplified then do so now in preference 2160 // to value numbering it. Value numbering often exposes redundancies, for 2161 // example if it determines that %y is equal to %x then the instruction 2162 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2163 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2164 I->replaceAllUsesWith(V); 2165 if (MD && V->getType()->getScalarType()->isPointerTy()) 2166 MD->invalidateCachedPointerInfo(V); 2167 markInstructionForDeletion(I); 2168 ++NumGVNSimpl; 2169 return true; 2170 } 2171 2172 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2173 if (processLoad(LI)) 2174 return true; 2175 2176 unsigned Num = VN.lookup_or_add(LI); 2177 addToLeaderTable(Num, LI, LI->getParent()); 2178 return false; 2179 } 2180 2181 // For conditional branches, we can perform simple conditional propagation on 2182 // the condition value itself. 2183 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2184 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2185 return false; 2186 2187 Value *BranchCond = BI->getCondition(); 2188 2189 BasicBlock *TrueSucc = BI->getSuccessor(0); 2190 BasicBlock *FalseSucc = BI->getSuccessor(1); 2191 // Avoid multiple edges early. 2192 if (TrueSucc == FalseSucc) 2193 return false; 2194 2195 BasicBlock *Parent = BI->getParent(); 2196 bool Changed = false; 2197 2198 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2199 BasicBlockEdge TrueE(Parent, TrueSucc); 2200 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2201 2202 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2203 BasicBlockEdge FalseE(Parent, FalseSucc); 2204 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2205 2206 return Changed; 2207 } 2208 2209 // For switches, propagate the case values into the case destinations. 2210 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2211 Value *SwitchCond = SI->getCondition(); 2212 BasicBlock *Parent = SI->getParent(); 2213 bool Changed = false; 2214 2215 // Remember how many outgoing edges there are to every successor. 2216 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2217 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2218 ++SwitchEdges[SI->getSuccessor(i)]; 2219 2220 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2221 i != e; ++i) { 2222 BasicBlock *Dst = i.getCaseSuccessor(); 2223 // If there is only a single edge, propagate the case value into it. 2224 if (SwitchEdges.lookup(Dst) == 1) { 2225 BasicBlockEdge E(Parent, Dst); 2226 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2227 } 2228 } 2229 return Changed; 2230 } 2231 2232 // Instructions with void type don't return a value, so there's 2233 // no point in trying to find redundancies in them. 2234 if (I->getType()->isVoidTy()) return false; 2235 2236 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2237 unsigned Num = VN.lookup_or_add(I); 2238 2239 // Allocations are always uniquely numbered, so we can save time and memory 2240 // by fast failing them. 2241 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2242 addToLeaderTable(Num, I, I->getParent()); 2243 return false; 2244 } 2245 2246 // If the number we were assigned was a brand new VN, then we don't 2247 // need to do a lookup to see if the number already exists 2248 // somewhere in the domtree: it can't! 2249 if (Num >= NextNum) { 2250 addToLeaderTable(Num, I, I->getParent()); 2251 return false; 2252 } 2253 2254 // Perform fast-path value-number based elimination of values inherited from 2255 // dominators. 2256 Value *repl = findLeader(I->getParent(), Num); 2257 if (repl == 0) { 2258 // Failure, just remember this instance for future use. 2259 addToLeaderTable(Num, I, I->getParent()); 2260 return false; 2261 } 2262 2263 // Remove it! 2264 patchAndReplaceAllUsesWith(I, repl); 2265 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2266 MD->invalidateCachedPointerInfo(repl); 2267 markInstructionForDeletion(I); 2268 return true; 2269} 2270 2271/// runOnFunction - This is the main transformation entry point for a function. 2272bool GVN::runOnFunction(Function& F) { 2273 if (!NoLoads) 2274 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2275 DT = &getAnalysis<DominatorTree>(); 2276 TD = getAnalysisIfAvailable<DataLayout>(); 2277 TLI = &getAnalysis<TargetLibraryInfo>(); 2278 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2279 VN.setMemDep(MD); 2280 VN.setDomTree(DT); 2281 2282 bool Changed = false; 2283 bool ShouldContinue = true; 2284 2285 // Merge unconditional branches, allowing PRE to catch more 2286 // optimization opportunities. 2287 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2288 BasicBlock *BB = FI++; 2289 2290 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2291 if (removedBlock) ++NumGVNBlocks; 2292 2293 Changed |= removedBlock; 2294 } 2295 2296 unsigned Iteration = 0; 2297 while (ShouldContinue) { 2298 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2299 ShouldContinue = iterateOnFunction(F); 2300 if (splitCriticalEdges()) 2301 ShouldContinue = true; 2302 Changed |= ShouldContinue; 2303 ++Iteration; 2304 } 2305 2306 if (EnablePRE) { 2307 bool PREChanged = true; 2308 while (PREChanged) { 2309 PREChanged = performPRE(F); 2310 Changed |= PREChanged; 2311 } 2312 } 2313 // FIXME: Should perform GVN again after PRE does something. PRE can move 2314 // computations into blocks where they become fully redundant. Note that 2315 // we can't do this until PRE's critical edge splitting updates memdep. 2316 // Actually, when this happens, we should just fully integrate PRE into GVN. 2317 2318 cleanupGlobalSets(); 2319 2320 return Changed; 2321} 2322 2323 2324bool GVN::processBlock(BasicBlock *BB) { 2325 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2326 // (and incrementing BI before processing an instruction). 2327 assert(InstrsToErase.empty() && 2328 "We expect InstrsToErase to be empty across iterations"); 2329 bool ChangedFunction = false; 2330 2331 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2332 BI != BE;) { 2333 ChangedFunction |= processInstruction(BI); 2334 if (InstrsToErase.empty()) { 2335 ++BI; 2336 continue; 2337 } 2338 2339 // If we need some instructions deleted, do it now. 2340 NumGVNInstr += InstrsToErase.size(); 2341 2342 // Avoid iterator invalidation. 2343 bool AtStart = BI == BB->begin(); 2344 if (!AtStart) 2345 --BI; 2346 2347 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2348 E = InstrsToErase.end(); I != E; ++I) { 2349 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2350 if (MD) MD->removeInstruction(*I); 2351 DEBUG(verifyRemoved(*I)); 2352 (*I)->eraseFromParent(); 2353 } 2354 InstrsToErase.clear(); 2355 2356 if (AtStart) 2357 BI = BB->begin(); 2358 else 2359 ++BI; 2360 } 2361 2362 return ChangedFunction; 2363} 2364 2365/// performPRE - Perform a purely local form of PRE that looks for diamond 2366/// control flow patterns and attempts to perform simple PRE at the join point. 2367bool GVN::performPRE(Function &F) { 2368 bool Changed = false; 2369 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2370 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2371 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2372 BasicBlock *CurrentBlock = *DI; 2373 2374 // Nothing to PRE in the entry block. 2375 if (CurrentBlock == &F.getEntryBlock()) continue; 2376 2377 // Don't perform PRE on a landing pad. 2378 if (CurrentBlock->isLandingPad()) continue; 2379 2380 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2381 BE = CurrentBlock->end(); BI != BE; ) { 2382 Instruction *CurInst = BI++; 2383 2384 if (isa<AllocaInst>(CurInst) || 2385 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2386 CurInst->getType()->isVoidTy() || 2387 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2388 isa<DbgInfoIntrinsic>(CurInst)) 2389 continue; 2390 2391 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2392 // sinking the compare again, and it would force the code generator to 2393 // move the i1 from processor flags or predicate registers into a general 2394 // purpose register. 2395 if (isa<CmpInst>(CurInst)) 2396 continue; 2397 2398 // We don't currently value number ANY inline asm calls. 2399 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2400 if (CallI->isInlineAsm()) 2401 continue; 2402 2403 uint32_t ValNo = VN.lookup(CurInst); 2404 2405 // Look for the predecessors for PRE opportunities. We're 2406 // only trying to solve the basic diamond case, where 2407 // a value is computed in the successor and one predecessor, 2408 // but not the other. We also explicitly disallow cases 2409 // where the successor is its own predecessor, because they're 2410 // more complicated to get right. 2411 unsigned NumWith = 0; 2412 unsigned NumWithout = 0; 2413 BasicBlock *PREPred = 0; 2414 predMap.clear(); 2415 2416 for (pred_iterator PI = pred_begin(CurrentBlock), 2417 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2418 BasicBlock *P = *PI; 2419 // We're not interested in PRE where the block is its 2420 // own predecessor, or in blocks with predecessors 2421 // that are not reachable. 2422 if (P == CurrentBlock) { 2423 NumWithout = 2; 2424 break; 2425 } else if (!DT->isReachableFromEntry(P)) { 2426 NumWithout = 2; 2427 break; 2428 } 2429 2430 Value* predV = findLeader(P, ValNo); 2431 if (predV == 0) { 2432 predMap.push_back(std::make_pair(static_cast<Value *>(0), P)); 2433 PREPred = P; 2434 ++NumWithout; 2435 } else if (predV == CurInst) { 2436 /* CurInst dominates this predecessor. */ 2437 NumWithout = 2; 2438 break; 2439 } else { 2440 predMap.push_back(std::make_pair(predV, P)); 2441 ++NumWith; 2442 } 2443 } 2444 2445 // Don't do PRE when it might increase code size, i.e. when 2446 // we would need to insert instructions in more than one pred. 2447 if (NumWithout != 1 || NumWith == 0) 2448 continue; 2449 2450 // Don't do PRE across indirect branch. 2451 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2452 continue; 2453 2454 // We can't do PRE safely on a critical edge, so instead we schedule 2455 // the edge to be split and perform the PRE the next time we iterate 2456 // on the function. 2457 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2458 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2459 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2460 continue; 2461 } 2462 2463 // Instantiate the expression in the predecessor that lacked it. 2464 // Because we are going top-down through the block, all value numbers 2465 // will be available in the predecessor by the time we need them. Any 2466 // that weren't originally present will have been instantiated earlier 2467 // in this loop. 2468 Instruction *PREInstr = CurInst->clone(); 2469 bool success = true; 2470 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2471 Value *Op = PREInstr->getOperand(i); 2472 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2473 continue; 2474 2475 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2476 PREInstr->setOperand(i, V); 2477 } else { 2478 success = false; 2479 break; 2480 } 2481 } 2482 2483 // Fail out if we encounter an operand that is not available in 2484 // the PRE predecessor. This is typically because of loads which 2485 // are not value numbered precisely. 2486 if (!success) { 2487 DEBUG(verifyRemoved(PREInstr)); 2488 delete PREInstr; 2489 continue; 2490 } 2491 2492 PREInstr->insertBefore(PREPred->getTerminator()); 2493 PREInstr->setName(CurInst->getName() + ".pre"); 2494 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2495 VN.add(PREInstr, ValNo); 2496 ++NumGVNPRE; 2497 2498 // Update the availability map to include the new instruction. 2499 addToLeaderTable(ValNo, PREInstr, PREPred); 2500 2501 // Create a PHI to make the value available in this block. 2502 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2503 CurInst->getName() + ".pre-phi", 2504 CurrentBlock->begin()); 2505 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2506 if (Value *V = predMap[i].first) 2507 Phi->addIncoming(V, predMap[i].second); 2508 else 2509 Phi->addIncoming(PREInstr, PREPred); 2510 } 2511 2512 VN.add(Phi, ValNo); 2513 addToLeaderTable(ValNo, Phi, CurrentBlock); 2514 Phi->setDebugLoc(CurInst->getDebugLoc()); 2515 CurInst->replaceAllUsesWith(Phi); 2516 if (Phi->getType()->getScalarType()->isPointerTy()) { 2517 // Because we have added a PHI-use of the pointer value, it has now 2518 // "escaped" from alias analysis' perspective. We need to inform 2519 // AA of this. 2520 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2521 ++ii) { 2522 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2523 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2524 } 2525 2526 if (MD) 2527 MD->invalidateCachedPointerInfo(Phi); 2528 } 2529 VN.erase(CurInst); 2530 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2531 2532 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2533 if (MD) MD->removeInstruction(CurInst); 2534 DEBUG(verifyRemoved(CurInst)); 2535 CurInst->eraseFromParent(); 2536 Changed = true; 2537 } 2538 } 2539 2540 if (splitCriticalEdges()) 2541 Changed = true; 2542 2543 return Changed; 2544} 2545 2546/// splitCriticalEdges - Split critical edges found during the previous 2547/// iteration that may enable further optimization. 2548bool GVN::splitCriticalEdges() { 2549 if (toSplit.empty()) 2550 return false; 2551 do { 2552 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2553 SplitCriticalEdge(Edge.first, Edge.second, this); 2554 } while (!toSplit.empty()); 2555 if (MD) MD->invalidateCachedPredecessors(); 2556 return true; 2557} 2558 2559/// iterateOnFunction - Executes one iteration of GVN 2560bool GVN::iterateOnFunction(Function &F) { 2561 cleanupGlobalSets(); 2562 2563 // Top-down walk of the dominator tree 2564 bool Changed = false; 2565#if 0 2566 // Needed for value numbering with phi construction to work. 2567 ReversePostOrderTraversal<Function*> RPOT(&F); 2568 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2569 RE = RPOT.end(); RI != RE; ++RI) 2570 Changed |= processBlock(*RI); 2571#else 2572 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2573 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2574 Changed |= processBlock(DI->getBlock()); 2575#endif 2576 2577 return Changed; 2578} 2579 2580void GVN::cleanupGlobalSets() { 2581 VN.clear(); 2582 LeaderTable.clear(); 2583 TableAllocator.Reset(); 2584} 2585 2586/// verifyRemoved - Verify that the specified instruction does not occur in our 2587/// internal data structures. 2588void GVN::verifyRemoved(const Instruction *Inst) const { 2589 VN.verifyRemoved(Inst); 2590 2591 // Walk through the value number scope to make sure the instruction isn't 2592 // ferreted away in it. 2593 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2594 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2595 const LeaderTableEntry *Node = &I->second; 2596 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2597 2598 while (Node->Next) { 2599 Node = Node->Next; 2600 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2601 } 2602 } 2603} 2604