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" 48 using namespace llvm; 49 using namespace PatternMatch; 50 51 STATISTIC(NumGVNInstr, "Number of instructions deleted"); 52 STATISTIC(NumGVNLoad, "Number of loads deleted"); 53 STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 54 STATISTIC(NumGVNBlocks, "Number of blocks merged"); 55 STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 56 STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 57 STATISTIC(NumPRELoad, "Number of loads PRE'd"); 58 59 static cl::opt<bool> EnablePRE("enable-pre", 60 cl::init(true), cl::Hidden); 61 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 62 63 // Maximum allowed recursion depth. 64 static cl::opt<uint32_t> 65 MaxRecurseDepth("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. 75 namespace { 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 } 134 135 namespace llvm { 136 template <> 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 //===----------------------------------------------------------------------===// 159 160 Expression 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 } 193 194 Expression 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 } 212 213 Expression 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. 270 void ValueTable::add(Value *V, uint32_t num) { 271 valueNumbering.insert(std::make_pair(V, num)); 272 } 273 274 uint32_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. 385 uint32_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. 456 uint32_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. 466 uint32_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. 476 void ValueTable::clear() { 477 valueNumbering.clear(); 478 expressionNumbering.clear(); 479 nextValueNumber = 1; 480 } 481 482 /// erase - Remove a value from the value numbering. 483 void ValueTable::erase(Value *V) { 484 valueNumbering.erase(V); 485 } 486 487 /// verifyRemoved - Verify that the value is removed from all internal data 488 /// structures. 489 void 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 //===----------------------------------------------------------------------===// 499 500 namespace { 501 502 class GVN : public FunctionPass { 503 bool NoLoads; 504 MemoryDependenceAnalysis *MD; 505 DominatorTree *DT; 506 const DataLayout *TD; 507 const TargetLibraryInfo *TLI; 508 509 ValueTable VN; 510 511 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 512 /// have that value number. Use findLeader to query it. 513 struct LeaderTableEntry { 514 Value *Val; 515 const BasicBlock *BB; 516 LeaderTableEntry *Next; 517 }; 518 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 519 BumpPtrAllocator TableAllocator; 520 521 SmallVector<Instruction*, 8> InstrsToErase; 522 public: 523 static char ID; // Pass identification, replacement for typeid 524 explicit GVN(bool noloads = false) 525 : FunctionPass(ID), NoLoads(noloads), MD(0) { 526 initializeGVNPass(*PassRegistry::getPassRegistry()); 527 } 528 529 bool runOnFunction(Function &F); 530 531 /// markInstructionForDeletion - This removes the specified instruction from 532 /// our various maps and marks it for deletion. 533 void markInstructionForDeletion(Instruction *I) { 534 VN.erase(I); 535 InstrsToErase.push_back(I); 536 } 537 538 const DataLayout *getDataLayout() const { return TD; } 539 DominatorTree &getDominatorTree() const { return *DT; } 540 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 541 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 542 private: 543 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 544 /// its value number. 545 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 546 LeaderTableEntry &Curr = LeaderTable[N]; 547 if (!Curr.Val) { 548 Curr.Val = V; 549 Curr.BB = BB; 550 return; 551 } 552 553 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 554 Node->Val = V; 555 Node->BB = BB; 556 Node->Next = Curr.Next; 557 Curr.Next = Node; 558 } 559 560 /// removeFromLeaderTable - Scan the list of values corresponding to a given 561 /// value number, and remove the given instruction if encountered. 562 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 563 LeaderTableEntry* Prev = 0; 564 LeaderTableEntry* Curr = &LeaderTable[N]; 565 566 while (Curr->Val != I || Curr->BB != BB) { 567 Prev = Curr; 568 Curr = Curr->Next; 569 } 570 571 if (Prev) { 572 Prev->Next = Curr->Next; 573 } else { 574 if (!Curr->Next) { 575 Curr->Val = 0; 576 Curr->BB = 0; 577 } else { 578 LeaderTableEntry* Next = Curr->Next; 579 Curr->Val = Next->Val; 580 Curr->BB = Next->BB; 581 Curr->Next = Next->Next; 582 } 583 } 584 } 585 586 // List of critical edges to be split between iterations. 587 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 588 589 // This transformation requires dominator postdominator info 590 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 591 AU.addRequired<DominatorTree>(); 592 AU.addRequired<TargetLibraryInfo>(); 593 if (!NoLoads) 594 AU.addRequired<MemoryDependenceAnalysis>(); 595 AU.addRequired<AliasAnalysis>(); 596 597 AU.addPreserved<DominatorTree>(); 598 AU.addPreserved<AliasAnalysis>(); 599 } 600 601 602 // Helper fuctions 603 // FIXME: eliminate or document these better 604 bool processLoad(LoadInst *L); 605 bool processInstruction(Instruction *I); 606 bool processNonLocalLoad(LoadInst *L); 607 bool processBlock(BasicBlock *BB); 608 void dump(DenseMap<uint32_t, Value*> &d); 609 bool iterateOnFunction(Function &F); 610 bool performPRE(Function &F); 611 Value *findLeader(const BasicBlock *BB, uint32_t num); 612 void cleanupGlobalSets(); 613 void verifyRemoved(const Instruction *I) const; 614 bool splitCriticalEdges(); 615 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 616 const BasicBlockEdge &Root); 617 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 618 }; 619 620 char GVN::ID = 0; 621 } 622 623 // createGVNPass - The public interface to this file... 624 FunctionPass *llvm::createGVNPass(bool NoLoads) { 625 return new GVN(NoLoads); 626 } 627 628 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 629 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 630 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 631 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 632 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 633 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 634 635 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 636 void GVN::dump(DenseMap<uint32_t, Value*>& d) { 637 errs() << "{\n"; 638 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 639 E = d.end(); I != E; ++I) { 640 errs() << I->first << "\n"; 641 I->second->dump(); 642 } 643 errs() << "}\n"; 644 } 645 #endif 646 647 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value 648 /// we're analyzing is fully available in the specified block. As we go, keep 649 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This 650 /// map is actually a tri-state map with the following values: 651 /// 0) we know the block *is not* fully available. 652 /// 1) we know the block *is* fully available. 653 /// 2) we do not know whether the block is fully available or not, but we are 654 /// currently speculating that it will be. 655 /// 3) we are speculating for this block and have used that to speculate for 656 /// other blocks. 657 static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 658 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 659 uint32_t RecurseDepth) { 660 if (RecurseDepth > MaxRecurseDepth) 661 return false; 662 663 // Optimistically assume that the block is fully available and check to see 664 // if we already know about this block in one lookup. 665 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 666 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 667 668 // If the entry already existed for this block, return the precomputed value. 669 if (!IV.second) { 670 // If this is a speculative "available" value, mark it as being used for 671 // speculation of other blocks. 672 if (IV.first->second == 2) 673 IV.first->second = 3; 674 return IV.first->second != 0; 675 } 676 677 // Otherwise, see if it is fully available in all predecessors. 678 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 679 680 // If this block has no predecessors, it isn't live-in here. 681 if (PI == PE) 682 goto SpeculationFailure; 683 684 for (; PI != PE; ++PI) 685 // If the value isn't fully available in one of our predecessors, then it 686 // isn't fully available in this block either. Undo our previous 687 // optimistic assumption and bail out. 688 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 689 goto SpeculationFailure; 690 691 return true; 692 693 // SpeculationFailure - If we get here, we found out that this is not, after 694 // all, a fully-available block. We have a problem if we speculated on this and 695 // used the speculation to mark other blocks as available. 696 SpeculationFailure: 697 char &BBVal = FullyAvailableBlocks[BB]; 698 699 // If we didn't speculate on this, just return with it set to false. 700 if (BBVal == 2) { 701 BBVal = 0; 702 return false; 703 } 704 705 // If we did speculate on this value, we could have blocks set to 1 that are 706 // incorrect. Walk the (transitive) successors of this block and mark them as 707 // 0 if set to one. 708 SmallVector<BasicBlock*, 32> BBWorklist; 709 BBWorklist.push_back(BB); 710 711 do { 712 BasicBlock *Entry = BBWorklist.pop_back_val(); 713 // Note that this sets blocks to 0 (unavailable) if they happen to not 714 // already be in FullyAvailableBlocks. This is safe. 715 char &EntryVal = FullyAvailableBlocks[Entry]; 716 if (EntryVal == 0) continue; // Already unavailable. 717 718 // Mark as unavailable. 719 EntryVal = 0; 720 721 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 722 BBWorklist.push_back(*I); 723 } while (!BBWorklist.empty()); 724 725 return false; 726 } 727 728 729 /// CanCoerceMustAliasedValueToLoad - Return true if 730 /// CoerceAvailableValueToLoadType will succeed. 731 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 732 Type *LoadTy, 733 const DataLayout &TD) { 734 // If the loaded or stored value is an first class array or struct, don't try 735 // to transform them. We need to be able to bitcast to integer. 736 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 737 StoredVal->getType()->isStructTy() || 738 StoredVal->getType()->isArrayTy()) 739 return false; 740 741 // The store has to be at least as big as the load. 742 if (TD.getTypeSizeInBits(StoredVal->getType()) < 743 TD.getTypeSizeInBits(LoadTy)) 744 return false; 745 746 return true; 747 } 748 749 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 750 /// then a load from a must-aliased pointer of a different type, try to coerce 751 /// the stored value. LoadedTy is the type of the load we want to replace and 752 /// InsertPt is the place to insert new instructions. 753 /// 754 /// If we can't do it, return null. 755 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 756 Type *LoadedTy, 757 Instruction *InsertPt, 758 const DataLayout &TD) { 759 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 760 return 0; 761 762 // If this is already the right type, just return it. 763 Type *StoredValTy = StoredVal->getType(); 764 765 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 766 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 767 768 // If the store and reload are the same size, we can always reuse it. 769 if (StoreSize == LoadSize) { 770 // Pointer to Pointer -> use bitcast. 771 if (StoredValTy->getScalarType()->isPointerTy() && 772 LoadedTy->getScalarType()->isPointerTy()) 773 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 774 775 // Convert source pointers to integers, which can be bitcast. 776 if (StoredValTy->getScalarType()->isPointerTy()) { 777 StoredValTy = TD.getIntPtrType(StoredValTy); 778 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 779 } 780 781 Type *TypeToCastTo = LoadedTy; 782 if (TypeToCastTo->getScalarType()->isPointerTy()) 783 TypeToCastTo = TD.getIntPtrType(TypeToCastTo); 784 785 if (StoredValTy != TypeToCastTo) 786 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 787 788 // Cast to pointer if the load needs a pointer type. 789 if (LoadedTy->getScalarType()->isPointerTy()) 790 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 791 792 return StoredVal; 793 } 794 795 // If the loaded value is smaller than the available value, then we can 796 // extract out a piece from it. If the available value is too small, then we 797 // can't do anything. 798 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 799 800 // Convert source pointers to integers, which can be manipulated. 801 if (StoredValTy->getScalarType()->isPointerTy()) { 802 StoredValTy = TD.getIntPtrType(StoredValTy); 803 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 804 } 805 806 // Convert vectors and fp to integer, which can be manipulated. 807 if (!StoredValTy->isIntegerTy()) { 808 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 809 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 810 } 811 812 // If this is a big-endian system, we need to shift the value down to the low 813 // bits so that a truncate will work. 814 if (TD.isBigEndian()) { 815 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 816 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 817 } 818 819 // Truncate the integer to the right size now. 820 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 821 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 822 823 if (LoadedTy == NewIntTy) 824 return StoredVal; 825 826 // If the result is a pointer, inttoptr. 827 if (LoadedTy->getScalarType()->isPointerTy()) 828 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 829 830 // Otherwise, bitcast. 831 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 832 } 833 834 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a 835 /// memdep query of a load that ends up being a clobbering memory write (store, 836 /// memset, memcpy, memmove). This means that the write *may* provide bits used 837 /// by the load but we can't be sure because the pointers don't mustalias. 838 /// 839 /// Check this case to see if there is anything more we can do before we give 840 /// up. This returns -1 if we have to give up, or a byte number in the stored 841 /// value of the piece that feeds the load. 842 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 843 Value *WritePtr, 844 uint64_t WriteSizeInBits, 845 const DataLayout &TD) { 846 // If the loaded or stored value is a first class array or struct, don't try 847 // to transform them. We need to be able to bitcast to integer. 848 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 849 return -1; 850 851 int64_t StoreOffset = 0, LoadOffset = 0; 852 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD); 853 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD); 854 if (StoreBase != LoadBase) 855 return -1; 856 857 // If the load and store are to the exact same address, they should have been 858 // a must alias. AA must have gotten confused. 859 // FIXME: Study to see if/when this happens. One case is forwarding a memset 860 // to a load from the base of the memset. 861 #if 0 862 if (LoadOffset == StoreOffset) { 863 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 864 << "Base = " << *StoreBase << "\n" 865 << "Store Ptr = " << *WritePtr << "\n" 866 << "Store Offs = " << StoreOffset << "\n" 867 << "Load Ptr = " << *LoadPtr << "\n"; 868 abort(); 869 } 870 #endif 871 872 // If the load and store don't overlap at all, the store doesn't provide 873 // anything to the load. In this case, they really don't alias at all, AA 874 // must have gotten confused. 875 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 876 877 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 878 return -1; 879 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 880 LoadSize >>= 3; 881 882 883 bool isAAFailure = false; 884 if (StoreOffset < LoadOffset) 885 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 886 else 887 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 888 889 if (isAAFailure) { 890 #if 0 891 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 892 << "Base = " << *StoreBase << "\n" 893 << "Store Ptr = " << *WritePtr << "\n" 894 << "Store Offs = " << StoreOffset << "\n" 895 << "Load Ptr = " << *LoadPtr << "\n"; 896 abort(); 897 #endif 898 return -1; 899 } 900 901 // If the Load isn't completely contained within the stored bits, we don't 902 // have all the bits to feed it. We could do something crazy in the future 903 // (issue a smaller load then merge the bits in) but this seems unlikely to be 904 // valuable. 905 if (StoreOffset > LoadOffset || 906 StoreOffset+StoreSize < LoadOffset+LoadSize) 907 return -1; 908 909 // Okay, we can do this transformation. Return the number of bytes into the 910 // store that the load is. 911 return LoadOffset-StoreOffset; 912 } 913 914 /// AnalyzeLoadFromClobberingStore - This function is called when we have a 915 /// memdep query of a load that ends up being a clobbering store. 916 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 917 StoreInst *DepSI, 918 const DataLayout &TD) { 919 // Cannot handle reading from store of first-class aggregate yet. 920 if (DepSI->getValueOperand()->getType()->isStructTy() || 921 DepSI->getValueOperand()->getType()->isArrayTy()) 922 return -1; 923 924 Value *StorePtr = DepSI->getPointerOperand(); 925 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 926 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 927 StorePtr, StoreSize, TD); 928 } 929 930 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a 931 /// memdep query of a load that ends up being clobbered by another load. See if 932 /// the other load can feed into the second load. 933 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 934 LoadInst *DepLI, const DataLayout &TD){ 935 // Cannot handle reading from store of first-class aggregate yet. 936 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 937 return -1; 938 939 Value *DepPtr = DepLI->getPointerOperand(); 940 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 941 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 942 if (R != -1) return R; 943 944 // If we have a load/load clobber an DepLI can be widened to cover this load, 945 // then we should widen it! 946 int64_t LoadOffs = 0; 947 const Value *LoadBase = 948 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD); 949 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 950 951 unsigned Size = MemoryDependenceAnalysis:: 952 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 953 if (Size == 0) return -1; 954 955 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 956 } 957 958 959 960 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 961 MemIntrinsic *MI, 962 const DataLayout &TD) { 963 // If the mem operation is a non-constant size, we can't handle it. 964 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 965 if (SizeCst == 0) return -1; 966 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 967 968 // If this is memset, we just need to see if the offset is valid in the size 969 // of the memset.. 970 if (MI->getIntrinsicID() == Intrinsic::memset) 971 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 972 MemSizeInBits, TD); 973 974 // If we have a memcpy/memmove, the only case we can handle is if this is a 975 // copy from constant memory. In that case, we can read directly from the 976 // constant memory. 977 MemTransferInst *MTI = cast<MemTransferInst>(MI); 978 979 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 980 if (Src == 0) return -1; 981 982 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 983 if (GV == 0 || !GV->isConstant()) return -1; 984 985 // See if the access is within the bounds of the transfer. 986 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 987 MI->getDest(), MemSizeInBits, TD); 988 if (Offset == -1) 989 return Offset; 990 991 // Otherwise, see if we can constant fold a load from the constant with the 992 // offset applied as appropriate. 993 Src = ConstantExpr::getBitCast(Src, 994 llvm::Type::getInt8PtrTy(Src->getContext())); 995 Constant *OffsetCst = 996 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 997 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 998 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 999 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1000 return Offset; 1001 return -1; 1002 } 1003 1004 1005 /// GetStoreValueForLoad - This function is called when we have a 1006 /// memdep query of a load that ends up being a clobbering store. This means 1007 /// that the store provides bits used by the load but we the pointers don't 1008 /// mustalias. Check this case to see if there is anything more we can do 1009 /// before we give up. 1010 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1011 Type *LoadTy, 1012 Instruction *InsertPt, const DataLayout &TD){ 1013 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1014 1015 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1016 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1017 1018 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1019 1020 // Compute which bits of the stored value are being used by the load. Convert 1021 // to an integer type to start with. 1022 if (SrcVal->getType()->getScalarType()->isPointerTy()) 1023 SrcVal = Builder.CreatePtrToInt(SrcVal, 1024 TD.getIntPtrType(SrcVal->getType())); 1025 if (!SrcVal->getType()->isIntegerTy()) 1026 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1027 1028 // Shift the bits to the least significant depending on endianness. 1029 unsigned ShiftAmt; 1030 if (TD.isLittleEndian()) 1031 ShiftAmt = Offset*8; 1032 else 1033 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1034 1035 if (ShiftAmt) 1036 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1037 1038 if (LoadSize != StoreSize) 1039 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1040 1041 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1042 } 1043 1044 /// GetLoadValueForLoad - This function is called when we have a 1045 /// memdep query of a load that ends up being a clobbering load. This means 1046 /// that the load *may* provide bits used by the load but we can't be sure 1047 /// because the pointers don't mustalias. Check this case to see if there is 1048 /// anything more we can do before we give up. 1049 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1050 Type *LoadTy, Instruction *InsertPt, 1051 GVN &gvn) { 1052 const DataLayout &TD = *gvn.getDataLayout(); 1053 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1054 // widen SrcVal out to a larger load. 1055 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1056 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1057 if (Offset+LoadSize > SrcValSize) { 1058 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1059 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1060 // If we have a load/load clobber an DepLI can be widened to cover this 1061 // load, then we should widen it to the next power of 2 size big enough! 1062 unsigned NewLoadSize = Offset+LoadSize; 1063 if (!isPowerOf2_32(NewLoadSize)) 1064 NewLoadSize = NextPowerOf2(NewLoadSize); 1065 1066 Value *PtrVal = SrcVal->getPointerOperand(); 1067 1068 // Insert the new load after the old load. This ensures that subsequent 1069 // memdep queries will find the new load. We can't easily remove the old 1070 // load completely because it is already in the value numbering table. 1071 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1072 Type *DestPTy = 1073 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1074 DestPTy = PointerType::get(DestPTy, 1075 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1076 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1077 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1078 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1079 NewLoad->takeName(SrcVal); 1080 NewLoad->setAlignment(SrcVal->getAlignment()); 1081 1082 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1083 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1084 1085 // Replace uses of the original load with the wider load. On a big endian 1086 // system, we need to shift down to get the relevant bits. 1087 Value *RV = NewLoad; 1088 if (TD.isBigEndian()) 1089 RV = Builder.CreateLShr(RV, 1090 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1091 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1092 SrcVal->replaceAllUsesWith(RV); 1093 1094 // We would like to use gvn.markInstructionForDeletion here, but we can't 1095 // because the load is already memoized into the leader map table that GVN 1096 // tracks. It is potentially possible to remove the load from the table, 1097 // but then there all of the operations based on it would need to be 1098 // rehashed. Just leave the dead load around. 1099 gvn.getMemDep().removeInstruction(SrcVal); 1100 SrcVal = NewLoad; 1101 } 1102 1103 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1104 } 1105 1106 1107 /// GetMemInstValueForLoad - This function is called when we have a 1108 /// memdep query of a load that ends up being a clobbering mem intrinsic. 1109 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1110 Type *LoadTy, Instruction *InsertPt, 1111 const DataLayout &TD){ 1112 LLVMContext &Ctx = LoadTy->getContext(); 1113 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1114 1115 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1116 1117 // We know that this method is only called when the mem transfer fully 1118 // provides the bits for the load. 1119 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1120 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1121 // independently of what the offset is. 1122 Value *Val = MSI->getValue(); 1123 if (LoadSize != 1) 1124 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1125 1126 Value *OneElt = Val; 1127 1128 // Splat the value out to the right number of bits. 1129 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1130 // If we can double the number of bytes set, do it. 1131 if (NumBytesSet*2 <= LoadSize) { 1132 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1133 Val = Builder.CreateOr(Val, ShVal); 1134 NumBytesSet <<= 1; 1135 continue; 1136 } 1137 1138 // Otherwise insert one byte at a time. 1139 Value *ShVal = Builder.CreateShl(Val, 1*8); 1140 Val = Builder.CreateOr(OneElt, ShVal); 1141 ++NumBytesSet; 1142 } 1143 1144 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1145 } 1146 1147 // Otherwise, this is a memcpy/memmove from a constant global. 1148 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1149 Constant *Src = cast<Constant>(MTI->getSource()); 1150 1151 // Otherwise, see if we can constant fold a load from the constant with the 1152 // offset applied as appropriate. 1153 Src = ConstantExpr::getBitCast(Src, 1154 llvm::Type::getInt8PtrTy(Src->getContext())); 1155 Constant *OffsetCst = 1156 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1157 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1158 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1159 return ConstantFoldLoadFromConstPtr(Src, &TD); 1160 } 1161 1162 namespace { 1163 1164 struct AvailableValueInBlock { 1165 /// BB - The basic block in question. 1166 BasicBlock *BB; 1167 enum ValType { 1168 SimpleVal, // A simple offsetted value that is accessed. 1169 LoadVal, // A value produced by a load. 1170 MemIntrin // A memory intrinsic which is loaded from. 1171 }; 1172 1173 /// V - The value that is live out of the block. 1174 PointerIntPair<Value *, 2, ValType> Val; 1175 1176 /// Offset - The byte offset in Val that is interesting for the load query. 1177 unsigned Offset; 1178 1179 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1180 unsigned Offset = 0) { 1181 AvailableValueInBlock Res; 1182 Res.BB = BB; 1183 Res.Val.setPointer(V); 1184 Res.Val.setInt(SimpleVal); 1185 Res.Offset = Offset; 1186 return Res; 1187 } 1188 1189 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1190 unsigned Offset = 0) { 1191 AvailableValueInBlock Res; 1192 Res.BB = BB; 1193 Res.Val.setPointer(MI); 1194 Res.Val.setInt(MemIntrin); 1195 Res.Offset = Offset; 1196 return Res; 1197 } 1198 1199 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1200 unsigned Offset = 0) { 1201 AvailableValueInBlock Res; 1202 Res.BB = BB; 1203 Res.Val.setPointer(LI); 1204 Res.Val.setInt(LoadVal); 1205 Res.Offset = Offset; 1206 return Res; 1207 } 1208 1209 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1210 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1211 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1212 1213 Value *getSimpleValue() const { 1214 assert(isSimpleValue() && "Wrong accessor"); 1215 return Val.getPointer(); 1216 } 1217 1218 LoadInst *getCoercedLoadValue() const { 1219 assert(isCoercedLoadValue() && "Wrong accessor"); 1220 return cast<LoadInst>(Val.getPointer()); 1221 } 1222 1223 MemIntrinsic *getMemIntrinValue() const { 1224 assert(isMemIntrinValue() && "Wrong accessor"); 1225 return cast<MemIntrinsic>(Val.getPointer()); 1226 } 1227 1228 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1229 /// defined here to the specified type. This handles various coercion cases. 1230 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1231 Value *Res; 1232 if (isSimpleValue()) { 1233 Res = getSimpleValue(); 1234 if (Res->getType() != LoadTy) { 1235 const DataLayout *TD = gvn.getDataLayout(); 1236 assert(TD && "Need target data to handle type mismatch case"); 1237 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1238 *TD); 1239 1240 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1241 << *getSimpleValue() << '\n' 1242 << *Res << '\n' << "\n\n\n"); 1243 } 1244 } else if (isCoercedLoadValue()) { 1245 LoadInst *Load = getCoercedLoadValue(); 1246 if (Load->getType() == LoadTy && Offset == 0) { 1247 Res = Load; 1248 } else { 1249 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1250 gvn); 1251 1252 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1253 << *getCoercedLoadValue() << '\n' 1254 << *Res << '\n' << "\n\n\n"); 1255 } 1256 } else { 1257 const DataLayout *TD = gvn.getDataLayout(); 1258 assert(TD && "Need target data to handle type mismatch case"); 1259 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1260 LoadTy, BB->getTerminator(), *TD); 1261 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1262 << " " << *getMemIntrinValue() << '\n' 1263 << *Res << '\n' << "\n\n\n"); 1264 } 1265 return Res; 1266 } 1267 }; 1268 1269 } // end anonymous namespace 1270 1271 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1272 /// construct SSA form, allowing us to eliminate LI. This returns the value 1273 /// that should be used at LI's definition site. 1274 static Value *ConstructSSAForLoadSet(LoadInst *LI, 1275 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1276 GVN &gvn) { 1277 // Check for the fully redundant, dominating load case. In this case, we can 1278 // just use the dominating value directly. 1279 if (ValuesPerBlock.size() == 1 && 1280 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1281 LI->getParent())) 1282 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1283 1284 // Otherwise, we have to construct SSA form. 1285 SmallVector<PHINode*, 8> NewPHIs; 1286 SSAUpdater SSAUpdate(&NewPHIs); 1287 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1288 1289 Type *LoadTy = LI->getType(); 1290 1291 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1292 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1293 BasicBlock *BB = AV.BB; 1294 1295 if (SSAUpdate.HasValueForBlock(BB)) 1296 continue; 1297 1298 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1299 } 1300 1301 // Perform PHI construction. 1302 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1303 1304 // If new PHI nodes were created, notify alias analysis. 1305 if (V->getType()->getScalarType()->isPointerTy()) { 1306 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1307 1308 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1309 AA->copyValue(LI, NewPHIs[i]); 1310 1311 // Now that we've copied information to the new PHIs, scan through 1312 // them again and inform alias analysis that we've added potentially 1313 // escaping uses to any values that are operands to these PHIs. 1314 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1315 PHINode *P = NewPHIs[i]; 1316 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1317 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1318 AA->addEscapingUse(P->getOperandUse(jj)); 1319 } 1320 } 1321 } 1322 1323 return V; 1324 } 1325 1326 static bool isLifetimeStart(const Instruction *Inst) { 1327 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1328 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1329 return false; 1330 } 1331 1332 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1333 /// non-local by performing PHI construction. 1334 bool GVN::processNonLocalLoad(LoadInst *LI) { 1335 // Find the non-local dependencies of the load. 1336 SmallVector<NonLocalDepResult, 64> Deps; 1337 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1338 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1339 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1340 // << Deps.size() << *LI << '\n'); 1341 1342 // If we had to process more than one hundred blocks to find the 1343 // dependencies, this load isn't worth worrying about. Optimizing 1344 // it will be too expensive. 1345 unsigned NumDeps = Deps.size(); 1346 if (NumDeps > 100) 1347 return false; 1348 1349 // If we had a phi translation failure, we'll have a single entry which is a 1350 // clobber in the current block. Reject this early. 1351 if (NumDeps == 1 && 1352 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1353 DEBUG( 1354 dbgs() << "GVN: non-local load "; 1355 WriteAsOperand(dbgs(), LI); 1356 dbgs() << " has unknown dependencies\n"; 1357 ); 1358 return false; 1359 } 1360 1361 // Filter out useless results (non-locals, etc). Keep track of the blocks 1362 // where we have a value available in repl, also keep track of whether we see 1363 // dependencies that produce an unknown value for the load (such as a call 1364 // that could potentially clobber the load). 1365 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock; 1366 SmallVector<BasicBlock*, 64> UnavailableBlocks; 1367 1368 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1369 BasicBlock *DepBB = Deps[i].getBB(); 1370 MemDepResult DepInfo = Deps[i].getResult(); 1371 1372 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1373 UnavailableBlocks.push_back(DepBB); 1374 continue; 1375 } 1376 1377 if (DepInfo.isClobber()) { 1378 // The address being loaded in this non-local block may not be the same as 1379 // the pointer operand of the load if PHI translation occurs. Make sure 1380 // to consider the right address. 1381 Value *Address = Deps[i].getAddress(); 1382 1383 // If the dependence is to a store that writes to a superset of the bits 1384 // read by the load, we can extract the bits we need for the load from the 1385 // stored value. 1386 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1387 if (TD && Address) { 1388 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1389 DepSI, *TD); 1390 if (Offset != -1) { 1391 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1392 DepSI->getValueOperand(), 1393 Offset)); 1394 continue; 1395 } 1396 } 1397 } 1398 1399 // Check to see if we have something like this: 1400 // load i32* P 1401 // load i8* (P+1) 1402 // if we have this, replace the later with an extraction from the former. 1403 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1404 // If this is a clobber and L is the first instruction in its block, then 1405 // we have the first instruction in the entry block. 1406 if (DepLI != LI && Address && TD) { 1407 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1408 LI->getPointerOperand(), 1409 DepLI, *TD); 1410 1411 if (Offset != -1) { 1412 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1413 Offset)); 1414 continue; 1415 } 1416 } 1417 } 1418 1419 // If the clobbering value is a memset/memcpy/memmove, see if we can 1420 // forward a value on from it. 1421 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1422 if (TD && Address) { 1423 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1424 DepMI, *TD); 1425 if (Offset != -1) { 1426 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1427 Offset)); 1428 continue; 1429 } 1430 } 1431 } 1432 1433 UnavailableBlocks.push_back(DepBB); 1434 continue; 1435 } 1436 1437 // DepInfo.isDef() here 1438 1439 Instruction *DepInst = DepInfo.getInst(); 1440 1441 // Loading the allocation -> undef. 1442 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1443 // Loading immediately after lifetime begin -> undef. 1444 isLifetimeStart(DepInst)) { 1445 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1446 UndefValue::get(LI->getType()))); 1447 continue; 1448 } 1449 1450 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1451 // Reject loads and stores that are to the same address but are of 1452 // different types if we have to. 1453 if (S->getValueOperand()->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(S->getValueOperand(), 1457 LI->getType(), *TD)) { 1458 UnavailableBlocks.push_back(DepBB); 1459 continue; 1460 } 1461 } 1462 1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1464 S->getValueOperand())); 1465 continue; 1466 } 1467 1468 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1469 // If the types mismatch and we can't handle it, reject reuse of the load. 1470 if (LD->getType() != LI->getType()) { 1471 // If the stored value is larger or equal to the loaded value, we can 1472 // reuse it. 1473 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1474 UnavailableBlocks.push_back(DepBB); 1475 continue; 1476 } 1477 } 1478 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1479 continue; 1480 } 1481 1482 UnavailableBlocks.push_back(DepBB); 1483 continue; 1484 } 1485 1486 // If we have no predecessors that produce a known value for this load, exit 1487 // early. 1488 if (ValuesPerBlock.empty()) return false; 1489 1490 // If all of the instructions we depend on produce a known value for this 1491 // load, then it is fully redundant and we can use PHI insertion to compute 1492 // its value. Insert PHIs and remove the fully redundant value now. 1493 if (UnavailableBlocks.empty()) { 1494 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1495 1496 // Perform PHI construction. 1497 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1498 LI->replaceAllUsesWith(V); 1499 1500 if (isa<PHINode>(V)) 1501 V->takeName(LI); 1502 if (V->getType()->getScalarType()->isPointerTy()) 1503 MD->invalidateCachedPointerInfo(V); 1504 markInstructionForDeletion(LI); 1505 ++NumGVNLoad; 1506 return true; 1507 } 1508 1509 if (!EnablePRE || !EnableLoadPRE) 1510 return false; 1511 1512 // Okay, we have *some* definitions of the value. This means that the value 1513 // is available in some of our (transitive) predecessors. Lets think about 1514 // doing PRE of this load. This will involve inserting a new load into the 1515 // predecessor when it's not available. We could do this in general, but 1516 // prefer to not increase code size. As such, we only do this when we know 1517 // that we only have to insert *one* load (which means we're basically moving 1518 // the load, not inserting a new one). 1519 1520 SmallPtrSet<BasicBlock *, 4> Blockers; 1521 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1522 Blockers.insert(UnavailableBlocks[i]); 1523 1524 // Let's find the first basic block with more than one predecessor. Walk 1525 // backwards through predecessors if needed. 1526 BasicBlock *LoadBB = LI->getParent(); 1527 BasicBlock *TmpBB = LoadBB; 1528 1529 bool allSingleSucc = true; 1530 while (TmpBB->getSinglePredecessor()) { 1531 TmpBB = TmpBB->getSinglePredecessor(); 1532 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1533 return false; 1534 if (Blockers.count(TmpBB)) 1535 return false; 1536 1537 // If any of these blocks has more than one successor (i.e. if the edge we 1538 // just traversed was critical), then there are other paths through this 1539 // block along which the load may not be anticipated. Hoisting the load 1540 // above this block would be adding the load to execution paths along 1541 // which it was not previously executed. 1542 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1543 return false; 1544 } 1545 1546 assert(TmpBB); 1547 LoadBB = TmpBB; 1548 1549 // Check to see how many predecessors have the loaded value fully 1550 // available. 1551 DenseMap<BasicBlock*, Value*> PredLoads; 1552 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1553 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1554 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1555 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1556 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1557 1558 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1559 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1560 PI != E; ++PI) { 1561 BasicBlock *Pred = *PI; 1562 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1563 continue; 1564 } 1565 PredLoads[Pred] = 0; 1566 1567 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1568 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1569 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1570 << Pred->getName() << "': " << *LI << '\n'); 1571 return false; 1572 } 1573 1574 if (LoadBB->isLandingPad()) { 1575 DEBUG(dbgs() 1576 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1577 << Pred->getName() << "': " << *LI << '\n'); 1578 return false; 1579 } 1580 1581 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1582 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1583 } 1584 } 1585 1586 if (!NeedToSplit.empty()) { 1587 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1588 return false; 1589 } 1590 1591 // Decide whether PRE is profitable for this load. 1592 unsigned NumUnavailablePreds = PredLoads.size(); 1593 assert(NumUnavailablePreds != 0 && 1594 "Fully available value should be eliminated above!"); 1595 1596 // If this load is unavailable in multiple predecessors, reject it. 1597 // FIXME: If we could restructure the CFG, we could make a common pred with 1598 // all the preds that don't have an available LI and insert a new load into 1599 // that one block. 1600 if (NumUnavailablePreds != 1) 1601 return false; 1602 1603 // Check if the load can safely be moved to all the unavailable predecessors. 1604 bool CanDoPRE = true; 1605 SmallVector<Instruction*, 8> NewInsts; 1606 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1607 E = PredLoads.end(); I != E; ++I) { 1608 BasicBlock *UnavailablePred = I->first; 1609 1610 // Do PHI translation to get its value in the predecessor if necessary. The 1611 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1612 1613 // If all preds have a single successor, then we know it is safe to insert 1614 // the load on the pred (?!?), so we can insert code to materialize the 1615 // pointer if it is not available. 1616 PHITransAddr Address(LI->getPointerOperand(), TD); 1617 Value *LoadPtr = 0; 1618 if (allSingleSucc) { 1619 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1620 *DT, NewInsts); 1621 } else { 1622 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1623 LoadPtr = Address.getAddr(); 1624 } 1625 1626 // If we couldn't find or insert a computation of this phi translated value, 1627 // we fail PRE. 1628 if (LoadPtr == 0) { 1629 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1630 << *LI->getPointerOperand() << "\n"); 1631 CanDoPRE = false; 1632 break; 1633 } 1634 1635 // Make sure it is valid to move this load here. We have to watch out for: 1636 // @1 = getelementptr (i8* p, ... 1637 // test p and branch if == 0 1638 // load @1 1639 // It is valid to have the getelementptr before the test, even if p can 1640 // be 0, as getelementptr only does address arithmetic. 1641 // If we are not pushing the value through any multiple-successor blocks 1642 // we do not have this case. Otherwise, check that the load is safe to 1643 // put anywhere; this can be improved, but should be conservatively safe. 1644 if (!allSingleSucc && 1645 // FIXME: REEVALUTE THIS. 1646 !isSafeToLoadUnconditionally(LoadPtr, 1647 UnavailablePred->getTerminator(), 1648 LI->getAlignment(), TD)) { 1649 CanDoPRE = false; 1650 break; 1651 } 1652 1653 I->second = LoadPtr; 1654 } 1655 1656 if (!CanDoPRE) { 1657 while (!NewInsts.empty()) { 1658 Instruction *I = NewInsts.pop_back_val(); 1659 if (MD) MD->removeInstruction(I); 1660 I->eraseFromParent(); 1661 } 1662 return false; 1663 } 1664 1665 // Okay, we can eliminate this load by inserting a reload in the predecessor 1666 // and using PHI construction to get the value in the other predecessors, do 1667 // it. 1668 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1669 DEBUG(if (!NewInsts.empty()) 1670 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1671 << *NewInsts.back() << '\n'); 1672 1673 // Assign value numbers to the new instructions. 1674 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1675 // FIXME: We really _ought_ to insert these value numbers into their 1676 // parent's availability map. However, in doing so, we risk getting into 1677 // ordering issues. If a block hasn't been processed yet, we would be 1678 // marking a value as AVAIL-IN, which isn't what we intend. 1679 VN.lookup_or_add(NewInsts[i]); 1680 } 1681 1682 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1683 E = PredLoads.end(); I != E; ++I) { 1684 BasicBlock *UnavailablePred = I->first; 1685 Value *LoadPtr = I->second; 1686 1687 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1688 LI->getAlignment(), 1689 UnavailablePred->getTerminator()); 1690 1691 // Transfer the old load's TBAA tag to the new load. 1692 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1693 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1694 1695 // Transfer DebugLoc. 1696 NewLoad->setDebugLoc(LI->getDebugLoc()); 1697 1698 // Add the newly created load. 1699 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1700 NewLoad)); 1701 MD->invalidateCachedPointerInfo(LoadPtr); 1702 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1703 } 1704 1705 // Perform PHI construction. 1706 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1707 LI->replaceAllUsesWith(V); 1708 if (isa<PHINode>(V)) 1709 V->takeName(LI); 1710 if (V->getType()->getScalarType()->isPointerTy()) 1711 MD->invalidateCachedPointerInfo(V); 1712 markInstructionForDeletion(LI); 1713 ++NumPRELoad; 1714 return true; 1715 } 1716 1717 static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1718 // Patch the replacement so that it is not more restrictive than the value 1719 // being replaced. 1720 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1721 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1722 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1723 isa<OverflowingBinaryOperator>(ReplOp)) { 1724 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1725 ReplOp->setHasNoSignedWrap(false); 1726 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1727 ReplOp->setHasNoUnsignedWrap(false); 1728 } 1729 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1730 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1731 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1732 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1733 unsigned Kind = Metadata[i].first; 1734 MDNode *IMD = I->getMetadata(Kind); 1735 MDNode *ReplMD = Metadata[i].second; 1736 switch(Kind) { 1737 default: 1738 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1739 break; 1740 case LLVMContext::MD_dbg: 1741 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1742 case LLVMContext::MD_tbaa: 1743 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1744 break; 1745 case LLVMContext::MD_range: 1746 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1747 break; 1748 case LLVMContext::MD_prof: 1749 llvm_unreachable("MD_prof in a non terminator instruction"); 1750 break; 1751 case LLVMContext::MD_fpmath: 1752 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1753 break; 1754 } 1755 } 1756 } 1757 } 1758 1759 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1760 patchReplacementInstruction(I, Repl); 1761 I->replaceAllUsesWith(Repl); 1762 } 1763 1764 /// processLoad - Attempt to eliminate a load, first by eliminating it 1765 /// locally, and then attempting non-local elimination if that fails. 1766 bool GVN::processLoad(LoadInst *L) { 1767 if (!MD) 1768 return false; 1769 1770 if (!L->isSimple()) 1771 return false; 1772 1773 if (L->use_empty()) { 1774 markInstructionForDeletion(L); 1775 return true; 1776 } 1777 1778 // ... to a pointer that has been loaded from before... 1779 MemDepResult Dep = MD->getDependency(L); 1780 1781 // If we have a clobber and target data is around, see if this is a clobber 1782 // that we can fix up through code synthesis. 1783 if (Dep.isClobber() && TD) { 1784 // Check to see if we have something like this: 1785 // store i32 123, i32* %P 1786 // %A = bitcast i32* %P to i8* 1787 // %B = gep i8* %A, i32 1 1788 // %C = load i8* %B 1789 // 1790 // We could do that by recognizing if the clobber instructions are obviously 1791 // a common base + constant offset, and if the previous store (or memset) 1792 // completely covers this load. This sort of thing can happen in bitfield 1793 // access code. 1794 Value *AvailVal = 0; 1795 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1796 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1797 L->getPointerOperand(), 1798 DepSI, *TD); 1799 if (Offset != -1) 1800 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1801 L->getType(), L, *TD); 1802 } 1803 1804 // Check to see if we have something like this: 1805 // load i32* P 1806 // load i8* (P+1) 1807 // if we have this, replace the later with an extraction from the former. 1808 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1809 // If this is a clobber and L is the first instruction in its block, then 1810 // we have the first instruction in the entry block. 1811 if (DepLI == L) 1812 return false; 1813 1814 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1815 L->getPointerOperand(), 1816 DepLI, *TD); 1817 if (Offset != -1) 1818 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1819 } 1820 1821 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1822 // a value on from it. 1823 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1824 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1825 L->getPointerOperand(), 1826 DepMI, *TD); 1827 if (Offset != -1) 1828 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1829 } 1830 1831 if (AvailVal) { 1832 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1833 << *AvailVal << '\n' << *L << "\n\n\n"); 1834 1835 // Replace the load! 1836 L->replaceAllUsesWith(AvailVal); 1837 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1838 MD->invalidateCachedPointerInfo(AvailVal); 1839 markInstructionForDeletion(L); 1840 ++NumGVNLoad; 1841 return true; 1842 } 1843 } 1844 1845 // If the value isn't available, don't do anything! 1846 if (Dep.isClobber()) { 1847 DEBUG( 1848 // fast print dep, using operator<< on instruction is too slow. 1849 dbgs() << "GVN: load "; 1850 WriteAsOperand(dbgs(), L); 1851 Instruction *I = Dep.getInst(); 1852 dbgs() << " is clobbered by " << *I << '\n'; 1853 ); 1854 return false; 1855 } 1856 1857 // If it is defined in another block, try harder. 1858 if (Dep.isNonLocal()) 1859 return processNonLocalLoad(L); 1860 1861 if (!Dep.isDef()) { 1862 DEBUG( 1863 // fast print dep, using operator<< on instruction is too slow. 1864 dbgs() << "GVN: load "; 1865 WriteAsOperand(dbgs(), L); 1866 dbgs() << " has unknown dependence\n"; 1867 ); 1868 return false; 1869 } 1870 1871 Instruction *DepInst = Dep.getInst(); 1872 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1873 Value *StoredVal = DepSI->getValueOperand(); 1874 1875 // The store and load are to a must-aliased pointer, but they may not 1876 // actually have the same type. See if we know how to reuse the stored 1877 // value (depending on its type). 1878 if (StoredVal->getType() != L->getType()) { 1879 if (TD) { 1880 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1881 L, *TD); 1882 if (StoredVal == 0) 1883 return false; 1884 1885 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1886 << '\n' << *L << "\n\n\n"); 1887 } 1888 else 1889 return false; 1890 } 1891 1892 // Remove it! 1893 L->replaceAllUsesWith(StoredVal); 1894 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1895 MD->invalidateCachedPointerInfo(StoredVal); 1896 markInstructionForDeletion(L); 1897 ++NumGVNLoad; 1898 return true; 1899 } 1900 1901 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1902 Value *AvailableVal = DepLI; 1903 1904 // The loads are of a must-aliased pointer, but they may not actually have 1905 // the same type. See if we know how to reuse the previously loaded value 1906 // (depending on its type). 1907 if (DepLI->getType() != L->getType()) { 1908 if (TD) { 1909 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1910 L, *TD); 1911 if (AvailableVal == 0) 1912 return false; 1913 1914 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1915 << "\n" << *L << "\n\n\n"); 1916 } 1917 else 1918 return false; 1919 } 1920 1921 // Remove it! 1922 patchAndReplaceAllUsesWith(L, AvailableVal); 1923 if (DepLI->getType()->getScalarType()->isPointerTy()) 1924 MD->invalidateCachedPointerInfo(DepLI); 1925 markInstructionForDeletion(L); 1926 ++NumGVNLoad; 1927 return true; 1928 } 1929 1930 // If this load really doesn't depend on anything, then we must be loading an 1931 // undef value. This can happen when loading for a fresh allocation with no 1932 // intervening stores, for example. 1933 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1934 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1935 markInstructionForDeletion(L); 1936 ++NumGVNLoad; 1937 return true; 1938 } 1939 1940 // If this load occurs either right after a lifetime begin, 1941 // then the loaded value is undefined. 1942 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1943 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1944 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1945 markInstructionForDeletion(L); 1946 ++NumGVNLoad; 1947 return true; 1948 } 1949 } 1950 1951 return false; 1952 } 1953 1954 // findLeader - In order to find a leader for a given value number at a 1955 // specific basic block, we first obtain the list of all Values for that number, 1956 // and then scan the list to find one whose block dominates the block in 1957 // question. This is fast because dominator tree queries consist of only 1958 // a few comparisons of DFS numbers. 1959 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 1960 LeaderTableEntry Vals = LeaderTable[num]; 1961 if (!Vals.Val) return 0; 1962 1963 Value *Val = 0; 1964 if (DT->dominates(Vals.BB, BB)) { 1965 Val = Vals.Val; 1966 if (isa<Constant>(Val)) return Val; 1967 } 1968 1969 LeaderTableEntry* Next = Vals.Next; 1970 while (Next) { 1971 if (DT->dominates(Next->BB, BB)) { 1972 if (isa<Constant>(Next->Val)) return Next->Val; 1973 if (!Val) Val = Next->Val; 1974 } 1975 1976 Next = Next->Next; 1977 } 1978 1979 return Val; 1980 } 1981 1982 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1983 /// use is dominated by the given basic block. Returns the number of uses that 1984 /// were replaced. 1985 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1986 const BasicBlockEdge &Root) { 1987 unsigned Count = 0; 1988 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1989 UI != UE; ) { 1990 Use &U = (UI++).getUse(); 1991 1992 if (DT->dominates(Root, U)) { 1993 U.set(To); 1994 ++Count; 1995 } 1996 } 1997 return Count; 1998 } 1999 2000 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2001 /// true if every path from the entry block to 'Dst' passes via this edge. In 2002 /// particular 'Dst' must not be reachable via another edge from 'Src'. 2003 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2004 DominatorTree *DT) { 2005 // While in theory it is interesting to consider the case in which Dst has 2006 // more than one predecessor, because Dst might be part of a loop which is 2007 // only reachable from Src, in practice it is pointless since at the time 2008 // GVN runs all such loops have preheaders, which means that Dst will have 2009 // been changed to have only one predecessor, namely Src. 2010 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2011 const BasicBlock *Src = E.getStart(); 2012 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2013 (void)Src; 2014 return Pred != 0; 2015 } 2016 2017 /// propagateEquality - The given values are known to be equal in every block 2018 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2019 /// 'RHS' everywhere in the scope. Returns whether a change was made. 2020 bool GVN::propagateEquality(Value *LHS, Value *RHS, 2021 const BasicBlockEdge &Root) { 2022 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2023 Worklist.push_back(std::make_pair(LHS, RHS)); 2024 bool Changed = false; 2025 // For speed, compute a conservative fast approximation to 2026 // DT->dominates(Root, Root.getEnd()); 2027 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2028 2029 while (!Worklist.empty()) { 2030 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2031 LHS = Item.first; RHS = Item.second; 2032 2033 if (LHS == RHS) continue; 2034 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2035 2036 // Don't try to propagate equalities between constants. 2037 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2038 2039 // Prefer a constant on the right-hand side, or an Argument if no constants. 2040 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2041 std::swap(LHS, RHS); 2042 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2043 2044 // If there is no obvious reason to prefer the left-hand side over the right- 2045 // hand side, ensure the longest lived term is on the right-hand side, so the 2046 // shortest lived term will be replaced by the longest lived. This tends to 2047 // expose more simplifications. 2048 uint32_t LVN = VN.lookup_or_add(LHS); 2049 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2050 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2051 // Move the 'oldest' value to the right-hand side, using the value number as 2052 // a proxy for age. 2053 uint32_t RVN = VN.lookup_or_add(RHS); 2054 if (LVN < RVN) { 2055 std::swap(LHS, RHS); 2056 LVN = RVN; 2057 } 2058 } 2059 2060 // If value numbering later sees that an instruction in the scope is equal 2061 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2062 // the invariant that instructions only occur in the leader table for their 2063 // own value number (this is used by removeFromLeaderTable), do not do this 2064 // if RHS is an instruction (if an instruction in the scope is morphed into 2065 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2066 // using the leader table is about compiling faster, not optimizing better). 2067 // The leader table only tracks basic blocks, not edges. Only add to if we 2068 // have the simple case where the edge dominates the end. 2069 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2070 addToLeaderTable(LVN, RHS, Root.getEnd()); 2071 2072 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2073 // LHS always has at least one use that is not dominated by Root, this will 2074 // never do anything if LHS has only one use. 2075 if (!LHS->hasOneUse()) { 2076 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2077 Changed |= NumReplacements > 0; 2078 NumGVNEqProp += NumReplacements; 2079 } 2080 2081 // Now try to deduce additional equalities from this one. For example, if the 2082 // known equality was "(A != B)" == "false" then it follows that A and B are 2083 // equal in the scope. Only boolean equalities with an explicit true or false 2084 // RHS are currently supported. 2085 if (!RHS->getType()->isIntegerTy(1)) 2086 // Not a boolean equality - bail out. 2087 continue; 2088 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2089 if (!CI) 2090 // RHS neither 'true' nor 'false' - bail out. 2091 continue; 2092 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2093 bool isKnownTrue = CI->isAllOnesValue(); 2094 bool isKnownFalse = !isKnownTrue; 2095 2096 // If "A && B" is known true then both A and B are known true. If "A || B" 2097 // is known false then both A and B are known false. 2098 Value *A, *B; 2099 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2100 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2101 Worklist.push_back(std::make_pair(A, RHS)); 2102 Worklist.push_back(std::make_pair(B, RHS)); 2103 continue; 2104 } 2105 2106 // If we are propagating an equality like "(A == B)" == "true" then also 2107 // propagate the equality A == B. When propagating a comparison such as 2108 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2109 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2110 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2111 2112 // If "A == B" is known true, or "A != B" is known false, then replace 2113 // A with B everywhere in the scope. 2114 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2115 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2116 Worklist.push_back(std::make_pair(Op0, Op1)); 2117 2118 // If "A >= B" is known true, replace "A < B" with false everywhere. 2119 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2120 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2121 // Since we don't have the instruction "A < B" immediately to hand, work out 2122 // the value number that it would have and use that to find an appropriate 2123 // instruction (if any). 2124 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2125 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2126 // If the number we were assigned was brand new then there is no point in 2127 // looking for an instruction realizing it: there cannot be one! 2128 if (Num < NextNum) { 2129 Value *NotCmp = findLeader(Root.getEnd(), Num); 2130 if (NotCmp && isa<Instruction>(NotCmp)) { 2131 unsigned NumReplacements = 2132 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2133 Changed |= NumReplacements > 0; 2134 NumGVNEqProp += NumReplacements; 2135 } 2136 } 2137 // Ensure that any instruction in scope that gets the "A < B" value number 2138 // is replaced with false. 2139 // The leader table only tracks basic blocks, not edges. Only add to if we 2140 // have the simple case where the edge dominates the end. 2141 if (RootDominatesEnd) 2142 addToLeaderTable(Num, NotVal, Root.getEnd()); 2143 2144 continue; 2145 } 2146 } 2147 2148 return Changed; 2149 } 2150 2151 /// processInstruction - When calculating availability, handle an instruction 2152 /// by inserting it into the appropriate sets 2153 bool GVN::processInstruction(Instruction *I) { 2154 // Ignore dbg info intrinsics. 2155 if (isa<DbgInfoIntrinsic>(I)) 2156 return false; 2157 2158 // If the instruction can be easily simplified then do so now in preference 2159 // to value numbering it. Value numbering often exposes redundancies, for 2160 // example if it determines that %y is equal to %x then the instruction 2161 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2162 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2163 I->replaceAllUsesWith(V); 2164 if (MD && V->getType()->getScalarType()->isPointerTy()) 2165 MD->invalidateCachedPointerInfo(V); 2166 markInstructionForDeletion(I); 2167 ++NumGVNSimpl; 2168 return true; 2169 } 2170 2171 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2172 if (processLoad(LI)) 2173 return true; 2174 2175 unsigned Num = VN.lookup_or_add(LI); 2176 addToLeaderTable(Num, LI, LI->getParent()); 2177 return false; 2178 } 2179 2180 // For conditional branches, we can perform simple conditional propagation on 2181 // the condition value itself. 2182 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2183 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2184 return false; 2185 2186 Value *BranchCond = BI->getCondition(); 2187 2188 BasicBlock *TrueSucc = BI->getSuccessor(0); 2189 BasicBlock *FalseSucc = BI->getSuccessor(1); 2190 // Avoid multiple edges early. 2191 if (TrueSucc == FalseSucc) 2192 return false; 2193 2194 BasicBlock *Parent = BI->getParent(); 2195 bool Changed = false; 2196 2197 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2198 BasicBlockEdge TrueE(Parent, TrueSucc); 2199 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2200 2201 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2202 BasicBlockEdge FalseE(Parent, FalseSucc); 2203 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2204 2205 return Changed; 2206 } 2207 2208 // For switches, propagate the case values into the case destinations. 2209 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2210 Value *SwitchCond = SI->getCondition(); 2211 BasicBlock *Parent = SI->getParent(); 2212 bool Changed = false; 2213 2214 // Remember how many outgoing edges there are to every successor. 2215 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2216 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2217 ++SwitchEdges[SI->getSuccessor(i)]; 2218 2219 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2220 i != e; ++i) { 2221 BasicBlock *Dst = i.getCaseSuccessor(); 2222 // If there is only a single edge, propagate the case value into it. 2223 if (SwitchEdges.lookup(Dst) == 1) { 2224 BasicBlockEdge E(Parent, Dst); 2225 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2226 } 2227 } 2228 return Changed; 2229 } 2230 2231 // Instructions with void type don't return a value, so there's 2232 // no point in trying to find redundancies in them. 2233 if (I->getType()->isVoidTy()) return false; 2234 2235 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2236 unsigned Num = VN.lookup_or_add(I); 2237 2238 // Allocations are always uniquely numbered, so we can save time and memory 2239 // by fast failing them. 2240 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2241 addToLeaderTable(Num, I, I->getParent()); 2242 return false; 2243 } 2244 2245 // If the number we were assigned was a brand new VN, then we don't 2246 // need to do a lookup to see if the number already exists 2247 // somewhere in the domtree: it can't! 2248 if (Num >= NextNum) { 2249 addToLeaderTable(Num, I, I->getParent()); 2250 return false; 2251 } 2252 2253 // Perform fast-path value-number based elimination of values inherited from 2254 // dominators. 2255 Value *repl = findLeader(I->getParent(), Num); 2256 if (repl == 0) { 2257 // Failure, just remember this instance for future use. 2258 addToLeaderTable(Num, I, I->getParent()); 2259 return false; 2260 } 2261 2262 // Remove it! 2263 patchAndReplaceAllUsesWith(I, repl); 2264 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2265 MD->invalidateCachedPointerInfo(repl); 2266 markInstructionForDeletion(I); 2267 return true; 2268 } 2269 2270 /// runOnFunction - This is the main transformation entry point for a function. 2271 bool GVN::runOnFunction(Function& F) { 2272 if (!NoLoads) 2273 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2274 DT = &getAnalysis<DominatorTree>(); 2275 TD = getAnalysisIfAvailable<DataLayout>(); 2276 TLI = &getAnalysis<TargetLibraryInfo>(); 2277 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2278 VN.setMemDep(MD); 2279 VN.setDomTree(DT); 2280 2281 bool Changed = false; 2282 bool ShouldContinue = true; 2283 2284 // Merge unconditional branches, allowing PRE to catch more 2285 // optimization opportunities. 2286 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2287 BasicBlock *BB = FI++; 2288 2289 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2290 if (removedBlock) ++NumGVNBlocks; 2291 2292 Changed |= removedBlock; 2293 } 2294 2295 unsigned Iteration = 0; 2296 while (ShouldContinue) { 2297 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2298 ShouldContinue = iterateOnFunction(F); 2299 if (splitCriticalEdges()) 2300 ShouldContinue = true; 2301 Changed |= ShouldContinue; 2302 ++Iteration; 2303 } 2304 2305 if (EnablePRE) { 2306 bool PREChanged = true; 2307 while (PREChanged) { 2308 PREChanged = performPRE(F); 2309 Changed |= PREChanged; 2310 } 2311 } 2312 // FIXME: Should perform GVN again after PRE does something. PRE can move 2313 // computations into blocks where they become fully redundant. Note that 2314 // we can't do this until PRE's critical edge splitting updates memdep. 2315 // Actually, when this happens, we should just fully integrate PRE into GVN. 2316 2317 cleanupGlobalSets(); 2318 2319 return Changed; 2320 } 2321 2322 2323 bool GVN::processBlock(BasicBlock *BB) { 2324 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2325 // (and incrementing BI before processing an instruction). 2326 assert(InstrsToErase.empty() && 2327 "We expect InstrsToErase to be empty across iterations"); 2328 bool ChangedFunction = false; 2329 2330 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2331 BI != BE;) { 2332 ChangedFunction |= processInstruction(BI); 2333 if (InstrsToErase.empty()) { 2334 ++BI; 2335 continue; 2336 } 2337 2338 // If we need some instructions deleted, do it now. 2339 NumGVNInstr += InstrsToErase.size(); 2340 2341 // Avoid iterator invalidation. 2342 bool AtStart = BI == BB->begin(); 2343 if (!AtStart) 2344 --BI; 2345 2346 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2347 E = InstrsToErase.end(); I != E; ++I) { 2348 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2349 if (MD) MD->removeInstruction(*I); 2350 DEBUG(verifyRemoved(*I)); 2351 (*I)->eraseFromParent(); 2352 } 2353 InstrsToErase.clear(); 2354 2355 if (AtStart) 2356 BI = BB->begin(); 2357 else 2358 ++BI; 2359 } 2360 2361 return ChangedFunction; 2362 } 2363 2364 /// performPRE - Perform a purely local form of PRE that looks for diamond 2365 /// control flow patterns and attempts to perform simple PRE at the join point. 2366 bool GVN::performPRE(Function &F) { 2367 bool Changed = false; 2368 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2369 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2370 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2371 BasicBlock *CurrentBlock = *DI; 2372 2373 // Nothing to PRE in the entry block. 2374 if (CurrentBlock == &F.getEntryBlock()) continue; 2375 2376 // Don't perform PRE on a landing pad. 2377 if (CurrentBlock->isLandingPad()) continue; 2378 2379 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2380 BE = CurrentBlock->end(); BI != BE; ) { 2381 Instruction *CurInst = BI++; 2382 2383 if (isa<AllocaInst>(CurInst) || 2384 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2385 CurInst->getType()->isVoidTy() || 2386 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2387 isa<DbgInfoIntrinsic>(CurInst)) 2388 continue; 2389 2390 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2391 // sinking the compare again, and it would force the code generator to 2392 // move the i1 from processor flags or predicate registers into a general 2393 // purpose register. 2394 if (isa<CmpInst>(CurInst)) 2395 continue; 2396 2397 // We don't currently value number ANY inline asm calls. 2398 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2399 if (CallI->isInlineAsm()) 2400 continue; 2401 2402 uint32_t ValNo = VN.lookup(CurInst); 2403 2404 // Look for the predecessors for PRE opportunities. We're 2405 // only trying to solve the basic diamond case, where 2406 // a value is computed in the successor and one predecessor, 2407 // but not the other. We also explicitly disallow cases 2408 // where the successor is its own predecessor, because they're 2409 // more complicated to get right. 2410 unsigned NumWith = 0; 2411 unsigned NumWithout = 0; 2412 BasicBlock *PREPred = 0; 2413 predMap.clear(); 2414 2415 for (pred_iterator PI = pred_begin(CurrentBlock), 2416 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2417 BasicBlock *P = *PI; 2418 // We're not interested in PRE where the block is its 2419 // own predecessor, or in blocks with predecessors 2420 // that are not reachable. 2421 if (P == CurrentBlock) { 2422 NumWithout = 2; 2423 break; 2424 } else if (!DT->isReachableFromEntry(P)) { 2425 NumWithout = 2; 2426 break; 2427 } 2428 2429 Value* predV = findLeader(P, ValNo); 2430 if (predV == 0) { 2431 predMap.push_back(std::make_pair(static_cast<Value *>(0), P)); 2432 PREPred = P; 2433 ++NumWithout; 2434 } else if (predV == CurInst) { 2435 /* CurInst dominates this predecessor. */ 2436 NumWithout = 2; 2437 break; 2438 } else { 2439 predMap.push_back(std::make_pair(predV, P)); 2440 ++NumWith; 2441 } 2442 } 2443 2444 // Don't do PRE when it might increase code size, i.e. when 2445 // we would need to insert instructions in more than one pred. 2446 if (NumWithout != 1 || NumWith == 0) 2447 continue; 2448 2449 // Don't do PRE across indirect branch. 2450 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2451 continue; 2452 2453 // We can't do PRE safely on a critical edge, so instead we schedule 2454 // the edge to be split and perform the PRE the next time we iterate 2455 // on the function. 2456 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2457 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2458 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2459 continue; 2460 } 2461 2462 // Instantiate the expression in the predecessor that lacked it. 2463 // Because we are going top-down through the block, all value numbers 2464 // will be available in the predecessor by the time we need them. Any 2465 // that weren't originally present will have been instantiated earlier 2466 // in this loop. 2467 Instruction *PREInstr = CurInst->clone(); 2468 bool success = true; 2469 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2470 Value *Op = PREInstr->getOperand(i); 2471 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2472 continue; 2473 2474 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2475 PREInstr->setOperand(i, V); 2476 } else { 2477 success = false; 2478 break; 2479 } 2480 } 2481 2482 // Fail out if we encounter an operand that is not available in 2483 // the PRE predecessor. This is typically because of loads which 2484 // are not value numbered precisely. 2485 if (!success) { 2486 DEBUG(verifyRemoved(PREInstr)); 2487 delete PREInstr; 2488 continue; 2489 } 2490 2491 PREInstr->insertBefore(PREPred->getTerminator()); 2492 PREInstr->setName(CurInst->getName() + ".pre"); 2493 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2494 VN.add(PREInstr, ValNo); 2495 ++NumGVNPRE; 2496 2497 // Update the availability map to include the new instruction. 2498 addToLeaderTable(ValNo, PREInstr, PREPred); 2499 2500 // Create a PHI to make the value available in this block. 2501 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2502 CurInst->getName() + ".pre-phi", 2503 CurrentBlock->begin()); 2504 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2505 if (Value *V = predMap[i].first) 2506 Phi->addIncoming(V, predMap[i].second); 2507 else 2508 Phi->addIncoming(PREInstr, PREPred); 2509 } 2510 2511 VN.add(Phi, ValNo); 2512 addToLeaderTable(ValNo, Phi, CurrentBlock); 2513 Phi->setDebugLoc(CurInst->getDebugLoc()); 2514 CurInst->replaceAllUsesWith(Phi); 2515 if (Phi->getType()->getScalarType()->isPointerTy()) { 2516 // Because we have added a PHI-use of the pointer value, it has now 2517 // "escaped" from alias analysis' perspective. We need to inform 2518 // AA of this. 2519 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2520 ++ii) { 2521 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2522 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2523 } 2524 2525 if (MD) 2526 MD->invalidateCachedPointerInfo(Phi); 2527 } 2528 VN.erase(CurInst); 2529 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2530 2531 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2532 if (MD) MD->removeInstruction(CurInst); 2533 DEBUG(verifyRemoved(CurInst)); 2534 CurInst->eraseFromParent(); 2535 Changed = true; 2536 } 2537 } 2538 2539 if (splitCriticalEdges()) 2540 Changed = true; 2541 2542 return Changed; 2543 } 2544 2545 /// splitCriticalEdges - Split critical edges found during the previous 2546 /// iteration that may enable further optimization. 2547 bool GVN::splitCriticalEdges() { 2548 if (toSplit.empty()) 2549 return false; 2550 do { 2551 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2552 SplitCriticalEdge(Edge.first, Edge.second, this); 2553 } while (!toSplit.empty()); 2554 if (MD) MD->invalidateCachedPredecessors(); 2555 return true; 2556 } 2557 2558 /// iterateOnFunction - Executes one iteration of GVN 2559 bool GVN::iterateOnFunction(Function &F) { 2560 cleanupGlobalSets(); 2561 2562 // Top-down walk of the dominator tree 2563 bool Changed = false; 2564 #if 0 2565 // Needed for value numbering with phi construction to work. 2566 ReversePostOrderTraversal<Function*> RPOT(&F); 2567 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2568 RE = RPOT.end(); RI != RE; ++RI) 2569 Changed |= processBlock(*RI); 2570 #else 2571 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2572 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2573 Changed |= processBlock(DI->getBlock()); 2574 #endif 2575 2576 return Changed; 2577 } 2578 2579 void GVN::cleanupGlobalSets() { 2580 VN.clear(); 2581 LeaderTable.clear(); 2582 TableAllocator.Reset(); 2583 } 2584 2585 /// verifyRemoved - Verify that the specified instruction does not occur in our 2586 /// internal data structures. 2587 void GVN::verifyRemoved(const Instruction *Inst) const { 2588 VN.verifyRemoved(Inst); 2589 2590 // Walk through the value number scope to make sure the instruction isn't 2591 // ferreted away in it. 2592 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2593 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2594 const LeaderTableEntry *Node = &I->second; 2595 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2596 2597 while (Node->Next) { 2598 Node = Node->Next; 2599 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2600 } 2601 } 2602 } 2603