GVN.cpp revision ad3ba6a7de5eb24ea58ba0ec9dd1e41912f678a9
1//===- GVN.cpp - Eliminate redundant values and loads ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs global value numbering to eliminate fully redundant 11// instructions. It also performs simple dead load elimination. 12// 13// Note that this pass does the value numbering itself; it does not use the 14// ValueNumbering analysis passes. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "gvn" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/GlobalVariable.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/LLVMContext.h" 23#include "llvm/Analysis/AliasAnalysis.h" 24#include "llvm/Analysis/ConstantFolding.h" 25#include "llvm/Analysis/Dominators.h" 26#include "llvm/Analysis/InstructionSimplify.h" 27#include "llvm/Analysis/Loads.h" 28#include "llvm/Analysis/MemoryBuiltins.h" 29#include "llvm/Analysis/MemoryDependenceAnalysis.h" 30#include "llvm/Analysis/PHITransAddr.h" 31#include "llvm/Analysis/ValueTracking.h" 32#include "llvm/Assembly/Writer.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/BasicBlockUtils.h" 35#include "llvm/Transforms/Utils/SSAUpdater.h" 36#include "llvm/ADT/DenseMap.h" 37#include "llvm/ADT/DepthFirstIterator.h" 38#include "llvm/ADT/SmallPtrSet.h" 39#include "llvm/ADT/Statistic.h" 40#include "llvm/Support/Allocator.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/IRBuilder.h" 44using namespace llvm; 45 46STATISTIC(NumGVNInstr, "Number of instructions deleted"); 47STATISTIC(NumGVNLoad, "Number of loads deleted"); 48STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 49STATISTIC(NumGVNBlocks, "Number of blocks merged"); 50STATISTIC(NumPRELoad, "Number of loads PRE'd"); 51 52static cl::opt<bool> EnablePRE("enable-pre", 53 cl::init(true), cl::Hidden); 54static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 55 56//===----------------------------------------------------------------------===// 57// ValueTable Class 58//===----------------------------------------------------------------------===// 59 60/// This class holds the mapping between values and value numbers. It is used 61/// as an efficient mechanism to determine the expression-wise equivalence of 62/// two values. 63namespace { 64 struct Expression { 65 uint32_t opcode; 66 const Type *type; 67 SmallVector<uint32_t, 4> varargs; 68 69 Expression(uint32_t o = ~2U) : opcode(o) { } 70 71 bool operator==(const Expression &other) const { 72 if (opcode != other.opcode) 73 return false; 74 if (opcode == ~0U || opcode == ~1U) 75 return true; 76 if (type != other.type) 77 return false; 78 if (varargs != other.varargs) 79 return false; 80 return true; 81 } 82 }; 83 84 class ValueTable { 85 DenseMap<Value*, uint32_t> valueNumbering; 86 DenseMap<Expression, uint32_t> expressionNumbering; 87 AliasAnalysis *AA; 88 MemoryDependenceAnalysis *MD; 89 DominatorTree *DT; 90 91 uint32_t nextValueNumber; 92 93 Expression create_expression(Instruction* I); 94 uint32_t lookup_or_add_call(CallInst* C); 95 public: 96 ValueTable() : nextValueNumber(1) { } 97 uint32_t lookup_or_add(Value *V); 98 uint32_t lookup(Value *V) const; 99 void add(Value *V, uint32_t num); 100 void clear(); 101 void erase(Value *v); 102 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 103 AliasAnalysis *getAliasAnalysis() const { return AA; } 104 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 105 void setDomTree(DominatorTree* D) { DT = D; } 106 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 107 void verifyRemoved(const Value *) const; 108 }; 109} 110 111namespace llvm { 112template <> struct DenseMapInfo<Expression> { 113 static inline Expression getEmptyKey() { 114 return ~0U; 115 } 116 117 static inline Expression getTombstoneKey() { 118 return ~1U; 119 } 120 121 static unsigned getHashValue(const Expression e) { 122 unsigned hash = e.opcode; 123 124 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 125 (unsigned)((uintptr_t)e.type >> 9)); 126 127 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 128 E = e.varargs.end(); I != E; ++I) 129 hash = *I + hash * 37; 130 131 return hash; 132 } 133 static bool isEqual(const Expression &LHS, const Expression &RHS) { 134 return LHS == RHS; 135 } 136}; 137 138} 139 140//===----------------------------------------------------------------------===// 141// ValueTable Internal Functions 142//===----------------------------------------------------------------------===// 143 144 145Expression ValueTable::create_expression(Instruction *I) { 146 Expression e; 147 e.type = I->getType(); 148 e.opcode = I->getOpcode(); 149 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 150 OI != OE; ++OI) 151 e.varargs.push_back(lookup_or_add(*OI)); 152 153 if (CmpInst *C = dyn_cast<CmpInst>(I)) 154 e.opcode = (C->getOpcode() << 8) | C->getPredicate(); 155 else if (ExtractValueInst *E = dyn_cast<ExtractValueInst>(I)) { 156 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 157 II != IE; ++II) 158 e.varargs.push_back(*II); 159 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 160 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 161 II != IE; ++II) 162 e.varargs.push_back(*II); 163 } 164 165 return e; 166} 167 168//===----------------------------------------------------------------------===// 169// ValueTable External Functions 170//===----------------------------------------------------------------------===// 171 172/// add - Insert a value into the table with a specified value number. 173void ValueTable::add(Value *V, uint32_t num) { 174 valueNumbering.insert(std::make_pair(V, num)); 175} 176 177uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 178 if (AA->doesNotAccessMemory(C)) { 179 Expression exp = create_expression(C); 180 uint32_t& e = expressionNumbering[exp]; 181 if (!e) e = nextValueNumber++; 182 valueNumbering[C] = e; 183 return e; 184 } else if (AA->onlyReadsMemory(C)) { 185 Expression exp = create_expression(C); 186 uint32_t& e = expressionNumbering[exp]; 187 if (!e) { 188 e = nextValueNumber++; 189 valueNumbering[C] = e; 190 return e; 191 } 192 if (!MD) { 193 e = nextValueNumber++; 194 valueNumbering[C] = e; 195 return e; 196 } 197 198 MemDepResult local_dep = MD->getDependency(C); 199 200 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 201 valueNumbering[C] = nextValueNumber; 202 return nextValueNumber++; 203 } 204 205 if (local_dep.isDef()) { 206 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 207 208 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 209 valueNumbering[C] = nextValueNumber; 210 return nextValueNumber++; 211 } 212 213 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 214 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 215 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 216 if (c_vn != cd_vn) { 217 valueNumbering[C] = nextValueNumber; 218 return nextValueNumber++; 219 } 220 } 221 222 uint32_t v = lookup_or_add(local_cdep); 223 valueNumbering[C] = v; 224 return v; 225 } 226 227 // Non-local case. 228 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 229 MD->getNonLocalCallDependency(CallSite(C)); 230 // FIXME: call/call dependencies for readonly calls should return def, not 231 // clobber! Move the checking logic to MemDep! 232 CallInst* cdep = 0; 233 234 // Check to see if we have a single dominating call instruction that is 235 // identical to C. 236 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 237 const NonLocalDepEntry *I = &deps[i]; 238 // Ignore non-local dependencies. 239 if (I->getResult().isNonLocal()) 240 continue; 241 242 // We don't handle non-depedencies. If we already have a call, reject 243 // instruction dependencies. 244 if (I->getResult().isClobber() || cdep != 0) { 245 cdep = 0; 246 break; 247 } 248 249 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 250 // FIXME: All duplicated with non-local case. 251 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 252 cdep = NonLocalDepCall; 253 continue; 254 } 255 256 cdep = 0; 257 break; 258 } 259 260 if (!cdep) { 261 valueNumbering[C] = nextValueNumber; 262 return nextValueNumber++; 263 } 264 265 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 266 valueNumbering[C] = nextValueNumber; 267 return nextValueNumber++; 268 } 269 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 270 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 271 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 272 if (c_vn != cd_vn) { 273 valueNumbering[C] = nextValueNumber; 274 return nextValueNumber++; 275 } 276 } 277 278 uint32_t v = lookup_or_add(cdep); 279 valueNumbering[C] = v; 280 return v; 281 282 } else { 283 valueNumbering[C] = nextValueNumber; 284 return nextValueNumber++; 285 } 286} 287 288/// lookup_or_add - Returns the value number for the specified value, assigning 289/// it a new number if it did not have one before. 290uint32_t ValueTable::lookup_or_add(Value *V) { 291 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 292 if (VI != valueNumbering.end()) 293 return VI->second; 294 295 if (!isa<Instruction>(V)) { 296 valueNumbering[V] = nextValueNumber; 297 return nextValueNumber++; 298 } 299 300 Instruction* I = cast<Instruction>(V); 301 Expression exp; 302 switch (I->getOpcode()) { 303 case Instruction::Call: 304 return lookup_or_add_call(cast<CallInst>(I)); 305 case Instruction::Add: 306 case Instruction::FAdd: 307 case Instruction::Sub: 308 case Instruction::FSub: 309 case Instruction::Mul: 310 case Instruction::FMul: 311 case Instruction::UDiv: 312 case Instruction::SDiv: 313 case Instruction::FDiv: 314 case Instruction::URem: 315 case Instruction::SRem: 316 case Instruction::FRem: 317 case Instruction::Shl: 318 case Instruction::LShr: 319 case Instruction::AShr: 320 case Instruction::And: 321 case Instruction::Or : 322 case Instruction::Xor: 323 case Instruction::ICmp: 324 case Instruction::FCmp: 325 case Instruction::Trunc: 326 case Instruction::ZExt: 327 case Instruction::SExt: 328 case Instruction::FPToUI: 329 case Instruction::FPToSI: 330 case Instruction::UIToFP: 331 case Instruction::SIToFP: 332 case Instruction::FPTrunc: 333 case Instruction::FPExt: 334 case Instruction::PtrToInt: 335 case Instruction::IntToPtr: 336 case Instruction::BitCast: 337 case Instruction::Select: 338 case Instruction::ExtractElement: 339 case Instruction::InsertElement: 340 case Instruction::ShuffleVector: 341 case Instruction::ExtractValue: 342 case Instruction::InsertValue: 343 case Instruction::GetElementPtr: 344 exp = create_expression(I); 345 break; 346 default: 347 valueNumbering[V] = nextValueNumber; 348 return nextValueNumber++; 349 } 350 351 uint32_t& e = expressionNumbering[exp]; 352 if (!e) e = nextValueNumber++; 353 valueNumbering[V] = e; 354 return e; 355} 356 357/// lookup - Returns the value number of the specified value. Fails if 358/// the value has not yet been numbered. 359uint32_t ValueTable::lookup(Value *V) const { 360 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 361 assert(VI != valueNumbering.end() && "Value not numbered?"); 362 return VI->second; 363} 364 365/// clear - Remove all entries from the ValueTable. 366void ValueTable::clear() { 367 valueNumbering.clear(); 368 expressionNumbering.clear(); 369 nextValueNumber = 1; 370} 371 372/// erase - Remove a value from the value numbering. 373void ValueTable::erase(Value *V) { 374 valueNumbering.erase(V); 375} 376 377/// verifyRemoved - Verify that the value is removed from all internal data 378/// structures. 379void ValueTable::verifyRemoved(const Value *V) const { 380 for (DenseMap<Value*, uint32_t>::const_iterator 381 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 382 assert(I->first != V && "Inst still occurs in value numbering map!"); 383 } 384} 385 386//===----------------------------------------------------------------------===// 387// GVN Pass 388//===----------------------------------------------------------------------===// 389 390namespace { 391 392 class GVN : public FunctionPass { 393 bool NoLoads; 394 MemoryDependenceAnalysis *MD; 395 DominatorTree *DT; 396 const TargetData *TD; 397 398 ValueTable VN; 399 400 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 401 /// have that value number. Use findLeader to query it. 402 struct LeaderTableEntry { 403 Value *Val; 404 BasicBlock *BB; 405 LeaderTableEntry *Next; 406 }; 407 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 408 BumpPtrAllocator TableAllocator; 409 410 SmallVector<Instruction*, 8> InstrsToErase; 411 public: 412 static char ID; // Pass identification, replacement for typeid 413 explicit GVN(bool noloads = false) 414 : FunctionPass(ID), NoLoads(noloads), MD(0) { 415 initializeGVNPass(*PassRegistry::getPassRegistry()); 416 } 417 418 bool runOnFunction(Function &F); 419 420 /// markInstructionForDeletion - This removes the specified instruction from 421 /// our various maps and marks it for deletion. 422 void markInstructionForDeletion(Instruction *I) { 423 VN.erase(I); 424 InstrsToErase.push_back(I); 425 } 426 427 const TargetData *getTargetData() const { return TD; } 428 DominatorTree &getDominatorTree() const { return *DT; } 429 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 430 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 431 private: 432 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 433 /// its value number. 434 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 435 LeaderTableEntry& Curr = LeaderTable[N]; 436 if (!Curr.Val) { 437 Curr.Val = V; 438 Curr.BB = BB; 439 return; 440 } 441 442 LeaderTableEntry* Node = TableAllocator.Allocate<LeaderTableEntry>(); 443 Node->Val = V; 444 Node->BB = BB; 445 Node->Next = Curr.Next; 446 Curr.Next = Node; 447 } 448 449 /// removeFromLeaderTable - Scan the list of values corresponding to a given 450 /// value number, and remove the given value if encountered. 451 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 452 LeaderTableEntry* Prev = 0; 453 LeaderTableEntry* Curr = &LeaderTable[N]; 454 455 while (Curr->Val != V || Curr->BB != BB) { 456 Prev = Curr; 457 Curr = Curr->Next; 458 } 459 460 if (Prev) { 461 Prev->Next = Curr->Next; 462 } else { 463 if (!Curr->Next) { 464 Curr->Val = 0; 465 Curr->BB = 0; 466 } else { 467 LeaderTableEntry* Next = Curr->Next; 468 Curr->Val = Next->Val; 469 Curr->BB = Next->BB; 470 Curr->Next = Next->Next; 471 } 472 } 473 } 474 475 // List of critical edges to be split between iterations. 476 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 477 478 // This transformation requires dominator postdominator info 479 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 480 AU.addRequired<DominatorTree>(); 481 if (!NoLoads) 482 AU.addRequired<MemoryDependenceAnalysis>(); 483 AU.addRequired<AliasAnalysis>(); 484 485 AU.addPreserved<DominatorTree>(); 486 AU.addPreserved<AliasAnalysis>(); 487 } 488 489 490 // Helper fuctions 491 // FIXME: eliminate or document these better 492 bool processLoad(LoadInst *L); 493 bool processInstruction(Instruction *I); 494 bool processNonLocalLoad(LoadInst *L); 495 bool processBlock(BasicBlock *BB); 496 void dump(DenseMap<uint32_t, Value*> &d); 497 bool iterateOnFunction(Function &F); 498 bool performPRE(Function &F); 499 Value *findLeader(BasicBlock *BB, uint32_t num); 500 void cleanupGlobalSets(); 501 void verifyRemoved(const Instruction *I) const; 502 bool splitCriticalEdges(); 503 }; 504 505 char GVN::ID = 0; 506} 507 508// createGVNPass - The public interface to this file... 509FunctionPass *llvm::createGVNPass(bool NoLoads) { 510 return new GVN(NoLoads); 511} 512 513INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 514INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 515INITIALIZE_PASS_DEPENDENCY(DominatorTree) 516INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 517INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 518 519void GVN::dump(DenseMap<uint32_t, Value*>& d) { 520 errs() << "{\n"; 521 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 522 E = d.end(); I != E; ++I) { 523 errs() << I->first << "\n"; 524 I->second->dump(); 525 } 526 errs() << "}\n"; 527} 528 529/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 530/// we're analyzing is fully available in the specified block. As we go, keep 531/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 532/// map is actually a tri-state map with the following values: 533/// 0) we know the block *is not* fully available. 534/// 1) we know the block *is* fully available. 535/// 2) we do not know whether the block is fully available or not, but we are 536/// currently speculating that it will be. 537/// 3) we are speculating for this block and have used that to speculate for 538/// other blocks. 539static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 540 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 541 // Optimistically assume that the block is fully available and check to see 542 // if we already know about this block in one lookup. 543 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 544 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 545 546 // If the entry already existed for this block, return the precomputed value. 547 if (!IV.second) { 548 // If this is a speculative "available" value, mark it as being used for 549 // speculation of other blocks. 550 if (IV.first->second == 2) 551 IV.first->second = 3; 552 return IV.first->second != 0; 553 } 554 555 // Otherwise, see if it is fully available in all predecessors. 556 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 557 558 // If this block has no predecessors, it isn't live-in here. 559 if (PI == PE) 560 goto SpeculationFailure; 561 562 for (; PI != PE; ++PI) 563 // If the value isn't fully available in one of our predecessors, then it 564 // isn't fully available in this block either. Undo our previous 565 // optimistic assumption and bail out. 566 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 567 goto SpeculationFailure; 568 569 return true; 570 571// SpeculationFailure - If we get here, we found out that this is not, after 572// all, a fully-available block. We have a problem if we speculated on this and 573// used the speculation to mark other blocks as available. 574SpeculationFailure: 575 char &BBVal = FullyAvailableBlocks[BB]; 576 577 // If we didn't speculate on this, just return with it set to false. 578 if (BBVal == 2) { 579 BBVal = 0; 580 return false; 581 } 582 583 // If we did speculate on this value, we could have blocks set to 1 that are 584 // incorrect. Walk the (transitive) successors of this block and mark them as 585 // 0 if set to one. 586 SmallVector<BasicBlock*, 32> BBWorklist; 587 BBWorklist.push_back(BB); 588 589 do { 590 BasicBlock *Entry = BBWorklist.pop_back_val(); 591 // Note that this sets blocks to 0 (unavailable) if they happen to not 592 // already be in FullyAvailableBlocks. This is safe. 593 char &EntryVal = FullyAvailableBlocks[Entry]; 594 if (EntryVal == 0) continue; // Already unavailable. 595 596 // Mark as unavailable. 597 EntryVal = 0; 598 599 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 600 BBWorklist.push_back(*I); 601 } while (!BBWorklist.empty()); 602 603 return false; 604} 605 606 607/// CanCoerceMustAliasedValueToLoad - Return true if 608/// CoerceAvailableValueToLoadType will succeed. 609static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 610 const Type *LoadTy, 611 const TargetData &TD) { 612 // If the loaded or stored value is an first class array or struct, don't try 613 // to transform them. We need to be able to bitcast to integer. 614 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 615 StoredVal->getType()->isStructTy() || 616 StoredVal->getType()->isArrayTy()) 617 return false; 618 619 // The store has to be at least as big as the load. 620 if (TD.getTypeSizeInBits(StoredVal->getType()) < 621 TD.getTypeSizeInBits(LoadTy)) 622 return false; 623 624 return true; 625} 626 627 628/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 629/// then a load from a must-aliased pointer of a different type, try to coerce 630/// the stored value. LoadedTy is the type of the load we want to replace and 631/// InsertPt is the place to insert new instructions. 632/// 633/// If we can't do it, return null. 634static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 635 const Type *LoadedTy, 636 Instruction *InsertPt, 637 const TargetData &TD) { 638 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 639 return 0; 640 641 // If this is already the right type, just return it. 642 const Type *StoredValTy = StoredVal->getType(); 643 644 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 645 uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy); 646 647 // If the store and reload are the same size, we can always reuse it. 648 if (StoreSize == LoadSize) { 649 // Pointer to Pointer -> use bitcast. 650 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 651 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 652 653 // Convert source pointers to integers, which can be bitcast. 654 if (StoredValTy->isPointerTy()) { 655 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 656 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 657 } 658 659 const Type *TypeToCastTo = LoadedTy; 660 if (TypeToCastTo->isPointerTy()) 661 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 662 663 if (StoredValTy != TypeToCastTo) 664 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 665 666 // Cast to pointer if the load needs a pointer type. 667 if (LoadedTy->isPointerTy()) 668 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 669 670 return StoredVal; 671 } 672 673 // If the loaded value is smaller than the available value, then we can 674 // extract out a piece from it. If the available value is too small, then we 675 // can't do anything. 676 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 677 678 // Convert source pointers to integers, which can be manipulated. 679 if (StoredValTy->isPointerTy()) { 680 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 681 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 682 } 683 684 // Convert vectors and fp to integer, which can be manipulated. 685 if (!StoredValTy->isIntegerTy()) { 686 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 687 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 688 } 689 690 // If this is a big-endian system, we need to shift the value down to the low 691 // bits so that a truncate will work. 692 if (TD.isBigEndian()) { 693 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 694 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 695 } 696 697 // Truncate the integer to the right size now. 698 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 699 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 700 701 if (LoadedTy == NewIntTy) 702 return StoredVal; 703 704 // If the result is a pointer, inttoptr. 705 if (LoadedTy->isPointerTy()) 706 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 707 708 // Otherwise, bitcast. 709 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 710} 711 712/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 713/// memdep query of a load that ends up being a clobbering memory write (store, 714/// memset, memcpy, memmove). This means that the write *may* provide bits used 715/// by the load but we can't be sure because the pointers don't mustalias. 716/// 717/// Check this case to see if there is anything more we can do before we give 718/// up. This returns -1 if we have to give up, or a byte number in the stored 719/// value of the piece that feeds the load. 720static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 721 Value *WritePtr, 722 uint64_t WriteSizeInBits, 723 const TargetData &TD) { 724 // If the loaded or stored value is an first class array or struct, don't try 725 // to transform them. We need to be able to bitcast to integer. 726 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 727 return -1; 728 729 int64_t StoreOffset = 0, LoadOffset = 0; 730 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 731 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 732 if (StoreBase != LoadBase) 733 return -1; 734 735 // If the load and store are to the exact same address, they should have been 736 // a must alias. AA must have gotten confused. 737 // FIXME: Study to see if/when this happens. One case is forwarding a memset 738 // to a load from the base of the memset. 739#if 0 740 if (LoadOffset == StoreOffset) { 741 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 742 << "Base = " << *StoreBase << "\n" 743 << "Store Ptr = " << *WritePtr << "\n" 744 << "Store Offs = " << StoreOffset << "\n" 745 << "Load Ptr = " << *LoadPtr << "\n"; 746 abort(); 747 } 748#endif 749 750 // If the load and store don't overlap at all, the store doesn't provide 751 // anything to the load. In this case, they really don't alias at all, AA 752 // must have gotten confused. 753 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 754 755 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 756 return -1; 757 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 758 LoadSize >>= 3; 759 760 761 bool isAAFailure = false; 762 if (StoreOffset < LoadOffset) 763 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 764 else 765 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 766 767 if (isAAFailure) { 768#if 0 769 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 770 << "Base = " << *StoreBase << "\n" 771 << "Store Ptr = " << *WritePtr << "\n" 772 << "Store Offs = " << StoreOffset << "\n" 773 << "Load Ptr = " << *LoadPtr << "\n"; 774 abort(); 775#endif 776 return -1; 777 } 778 779 // If the Load isn't completely contained within the stored bits, we don't 780 // have all the bits to feed it. We could do something crazy in the future 781 // (issue a smaller load then merge the bits in) but this seems unlikely to be 782 // valuable. 783 if (StoreOffset > LoadOffset || 784 StoreOffset+StoreSize < LoadOffset+LoadSize) 785 return -1; 786 787 // Okay, we can do this transformation. Return the number of bytes into the 788 // store that the load is. 789 return LoadOffset-StoreOffset; 790} 791 792/// AnalyzeLoadFromClobberingStore - This function is called when we have a 793/// memdep query of a load that ends up being a clobbering store. 794static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 795 StoreInst *DepSI, 796 const TargetData &TD) { 797 // Cannot handle reading from store of first-class aggregate yet. 798 if (DepSI->getValueOperand()->getType()->isStructTy() || 799 DepSI->getValueOperand()->getType()->isArrayTy()) 800 return -1; 801 802 Value *StorePtr = DepSI->getPointerOperand(); 803 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 804 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 805 StorePtr, StoreSize, TD); 806} 807 808/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 809/// memdep query of a load that ends up being clobbered by another load. See if 810/// the other load can feed into the second load. 811static int AnalyzeLoadFromClobberingLoad(const Type *LoadTy, Value *LoadPtr, 812 LoadInst *DepLI, const TargetData &TD){ 813 // Cannot handle reading from store of first-class aggregate yet. 814 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 815 return -1; 816 817 Value *DepPtr = DepLI->getPointerOperand(); 818 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 819 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 820 if (R != -1) return R; 821 822 // If we have a load/load clobber an DepLI can be widened to cover this load, 823 // then we should widen it! 824 int64_t LoadOffs = 0; 825 const Value *LoadBase = 826 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 827 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 828 829 unsigned Size = MemoryDependenceAnalysis:: 830 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 831 if (Size == 0) return -1; 832 833 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 834} 835 836 837 838static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 839 MemIntrinsic *MI, 840 const TargetData &TD) { 841 // If the mem operation is a non-constant size, we can't handle it. 842 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 843 if (SizeCst == 0) return -1; 844 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 845 846 // If this is memset, we just need to see if the offset is valid in the size 847 // of the memset.. 848 if (MI->getIntrinsicID() == Intrinsic::memset) 849 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 850 MemSizeInBits, TD); 851 852 // If we have a memcpy/memmove, the only case we can handle is if this is a 853 // copy from constant memory. In that case, we can read directly from the 854 // constant memory. 855 MemTransferInst *MTI = cast<MemTransferInst>(MI); 856 857 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 858 if (Src == 0) return -1; 859 860 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 861 if (GV == 0 || !GV->isConstant()) return -1; 862 863 // See if the access is within the bounds of the transfer. 864 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 865 MI->getDest(), MemSizeInBits, TD); 866 if (Offset == -1) 867 return Offset; 868 869 // Otherwise, see if we can constant fold a load from the constant with the 870 // offset applied as appropriate. 871 Src = ConstantExpr::getBitCast(Src, 872 llvm::Type::getInt8PtrTy(Src->getContext())); 873 Constant *OffsetCst = 874 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 875 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 876 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 877 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 878 return Offset; 879 return -1; 880} 881 882 883/// GetStoreValueForLoad - This function is called when we have a 884/// memdep query of a load that ends up being a clobbering store. This means 885/// that the store provides bits used by the load but we the pointers don't 886/// mustalias. Check this case to see if there is anything more we can do 887/// before we give up. 888static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 889 const Type *LoadTy, 890 Instruction *InsertPt, const TargetData &TD){ 891 LLVMContext &Ctx = SrcVal->getType()->getContext(); 892 893 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 894 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 895 896 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 897 898 // Compute which bits of the stored value are being used by the load. Convert 899 // to an integer type to start with. 900 if (SrcVal->getType()->isPointerTy()) 901 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 902 if (!SrcVal->getType()->isIntegerTy()) 903 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 904 "tmp"); 905 906 // Shift the bits to the least significant depending on endianness. 907 unsigned ShiftAmt; 908 if (TD.isLittleEndian()) 909 ShiftAmt = Offset*8; 910 else 911 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 912 913 if (ShiftAmt) 914 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 915 916 if (LoadSize != StoreSize) 917 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 918 "tmp"); 919 920 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 921} 922 923/// GetStoreValueForLoad - This function is called when we have a 924/// memdep query of a load that ends up being a clobbering load. This means 925/// that the load *may* provide bits used by the load but we can't be sure 926/// because the pointers don't mustalias. Check this case to see if there is 927/// anything more we can do before we give up. 928static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 929 const Type *LoadTy, Instruction *InsertPt, 930 GVN &gvn) { 931 const TargetData &TD = *gvn.getTargetData(); 932 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 933 // widen SrcVal out to a larger load. 934 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 935 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 936 if (Offset+LoadSize > SrcValSize) { 937 assert(!SrcVal->isVolatile() && "Cannot widen volatile load!"); 938 assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load"); 939 // If we have a load/load clobber an DepLI can be widened to cover this 940 // load, then we should widen it to the next power of 2 size big enough! 941 unsigned NewLoadSize = Offset+LoadSize; 942 if (!isPowerOf2_32(NewLoadSize)) 943 NewLoadSize = NextPowerOf2(NewLoadSize); 944 945 Value *PtrVal = SrcVal->getPointerOperand(); 946 947 IRBuilder<> Builder(SrcVal->getParent(), SrcVal); 948 const Type *DestPTy = 949 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 950 DestPTy = PointerType::get(DestPTy, 951 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 952 953 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 954 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 955 NewLoad->takeName(SrcVal); 956 NewLoad->setAlignment(SrcVal->getAlignment()); 957 958 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 959 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 960 961 // Replace uses of the original load with the wider load. On a big endian 962 // system, we need to shift down to get the relevant bits. 963 Value *RV = NewLoad; 964 if (TD.isBigEndian()) 965 RV = Builder.CreateLShr(RV, 966 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 967 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 968 SrcVal->replaceAllUsesWith(RV); 969 gvn.getMemDep().removeInstruction(SrcVal); 970 SrcVal = NewLoad; 971 } 972 973 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 974} 975 976 977/// GetMemInstValueForLoad - This function is called when we have a 978/// memdep query of a load that ends up being a clobbering mem intrinsic. 979static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 980 const Type *LoadTy, Instruction *InsertPt, 981 const TargetData &TD){ 982 LLVMContext &Ctx = LoadTy->getContext(); 983 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 984 985 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 986 987 // We know that this method is only called when the mem transfer fully 988 // provides the bits for the load. 989 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 990 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 991 // independently of what the offset is. 992 Value *Val = MSI->getValue(); 993 if (LoadSize != 1) 994 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 995 996 Value *OneElt = Val; 997 998 // Splat the value out to the right number of bits. 999 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1000 // If we can double the number of bytes set, do it. 1001 if (NumBytesSet*2 <= LoadSize) { 1002 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1003 Val = Builder.CreateOr(Val, ShVal); 1004 NumBytesSet <<= 1; 1005 continue; 1006 } 1007 1008 // Otherwise insert one byte at a time. 1009 Value *ShVal = Builder.CreateShl(Val, 1*8); 1010 Val = Builder.CreateOr(OneElt, ShVal); 1011 ++NumBytesSet; 1012 } 1013 1014 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1015 } 1016 1017 // Otherwise, this is a memcpy/memmove from a constant global. 1018 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1019 Constant *Src = cast<Constant>(MTI->getSource()); 1020 1021 // Otherwise, see if we can constant fold a load from the constant with the 1022 // offset applied as appropriate. 1023 Src = ConstantExpr::getBitCast(Src, 1024 llvm::Type::getInt8PtrTy(Src->getContext())); 1025 Constant *OffsetCst = 1026 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1027 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1028 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1029 return ConstantFoldLoadFromConstPtr(Src, &TD); 1030} 1031 1032namespace { 1033 1034struct AvailableValueInBlock { 1035 /// BB - The basic block in question. 1036 BasicBlock *BB; 1037 enum ValType { 1038 SimpleVal, // A simple offsetted value that is accessed. 1039 LoadVal, // A value produced by a load. 1040 MemIntrin // A memory intrinsic which is loaded from. 1041 }; 1042 1043 /// V - The value that is live out of the block. 1044 PointerIntPair<Value *, 2, ValType> Val; 1045 1046 /// Offset - The byte offset in Val that is interesting for the load query. 1047 unsigned Offset; 1048 1049 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1050 unsigned Offset = 0) { 1051 AvailableValueInBlock Res; 1052 Res.BB = BB; 1053 Res.Val.setPointer(V); 1054 Res.Val.setInt(SimpleVal); 1055 Res.Offset = Offset; 1056 return Res; 1057 } 1058 1059 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1060 unsigned Offset = 0) { 1061 AvailableValueInBlock Res; 1062 Res.BB = BB; 1063 Res.Val.setPointer(MI); 1064 Res.Val.setInt(MemIntrin); 1065 Res.Offset = Offset; 1066 return Res; 1067 } 1068 1069 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1070 unsigned Offset = 0) { 1071 AvailableValueInBlock Res; 1072 Res.BB = BB; 1073 Res.Val.setPointer(LI); 1074 Res.Val.setInt(LoadVal); 1075 Res.Offset = Offset; 1076 return Res; 1077 } 1078 1079 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1080 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1081 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1082 1083 Value *getSimpleValue() const { 1084 assert(isSimpleValue() && "Wrong accessor"); 1085 return Val.getPointer(); 1086 } 1087 1088 LoadInst *getCoercedLoadValue() const { 1089 assert(isCoercedLoadValue() && "Wrong accessor"); 1090 return cast<LoadInst>(Val.getPointer()); 1091 } 1092 1093 MemIntrinsic *getMemIntrinValue() const { 1094 assert(isMemIntrinValue() && "Wrong accessor"); 1095 return cast<MemIntrinsic>(Val.getPointer()); 1096 } 1097 1098 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1099 /// defined here to the specified type. This handles various coercion cases. 1100 Value *MaterializeAdjustedValue(const Type *LoadTy, GVN &gvn) const { 1101 Value *Res; 1102 if (isSimpleValue()) { 1103 Res = getSimpleValue(); 1104 if (Res->getType() != LoadTy) { 1105 const TargetData *TD = gvn.getTargetData(); 1106 assert(TD && "Need target data to handle type mismatch case"); 1107 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1108 *TD); 1109 1110 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1111 << *getSimpleValue() << '\n' 1112 << *Res << '\n' << "\n\n\n"); 1113 } 1114 } else if (isCoercedLoadValue()) { 1115 LoadInst *Load = getCoercedLoadValue(); 1116 if (Load->getType() == LoadTy && Offset == 0) { 1117 Res = Load; 1118 } else { 1119 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1120 gvn); 1121 1122 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1123 << *getCoercedLoadValue() << '\n' 1124 << *Res << '\n' << "\n\n\n"); 1125 } 1126 } else { 1127 const TargetData *TD = gvn.getTargetData(); 1128 assert(TD && "Need target data to handle type mismatch case"); 1129 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1130 LoadTy, BB->getTerminator(), *TD); 1131 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1132 << " " << *getMemIntrinValue() << '\n' 1133 << *Res << '\n' << "\n\n\n"); 1134 } 1135 return Res; 1136 } 1137}; 1138 1139} // end anonymous namespace 1140 1141/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1142/// construct SSA form, allowing us to eliminate LI. This returns the value 1143/// that should be used at LI's definition site. 1144static Value *ConstructSSAForLoadSet(LoadInst *LI, 1145 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1146 GVN &gvn) { 1147 // Check for the fully redundant, dominating load case. In this case, we can 1148 // just use the dominating value directly. 1149 if (ValuesPerBlock.size() == 1 && 1150 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1151 LI->getParent())) 1152 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1153 1154 // Otherwise, we have to construct SSA form. 1155 SmallVector<PHINode*, 8> NewPHIs; 1156 SSAUpdater SSAUpdate(&NewPHIs); 1157 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1158 1159 const Type *LoadTy = LI->getType(); 1160 1161 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1162 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1163 BasicBlock *BB = AV.BB; 1164 1165 if (SSAUpdate.HasValueForBlock(BB)) 1166 continue; 1167 1168 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1169 } 1170 1171 // Perform PHI construction. 1172 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1173 1174 // If new PHI nodes were created, notify alias analysis. 1175 if (V->getType()->isPointerTy()) { 1176 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1177 1178 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1179 AA->copyValue(LI, NewPHIs[i]); 1180 1181 // Now that we've copied information to the new PHIs, scan through 1182 // them again and inform alias analysis that we've added potentially 1183 // escaping uses to any values that are operands to these PHIs. 1184 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1185 PHINode *P = NewPHIs[i]; 1186 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) 1187 AA->addEscapingUse(P->getOperandUse(2*ii)); 1188 } 1189 } 1190 1191 return V; 1192} 1193 1194static bool isLifetimeStart(const Instruction *Inst) { 1195 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1196 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1197 return false; 1198} 1199 1200/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1201/// non-local by performing PHI construction. 1202bool GVN::processNonLocalLoad(LoadInst *LI) { 1203 // Find the non-local dependencies of the load. 1204 SmallVector<NonLocalDepResult, 64> Deps; 1205 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1206 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1207 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1208 // << Deps.size() << *LI << '\n'); 1209 1210 // If we had to process more than one hundred blocks to find the 1211 // dependencies, this load isn't worth worrying about. Optimizing 1212 // it will be too expensive. 1213 if (Deps.size() > 100) 1214 return false; 1215 1216 // If we had a phi translation failure, we'll have a single entry which is a 1217 // clobber in the current block. Reject this early. 1218 if (Deps.size() == 1 && Deps[0].getResult().isClobber() && 1219 Deps[0].getResult().getInst()->getParent() == LI->getParent()) { 1220 DEBUG( 1221 dbgs() << "GVN: non-local load "; 1222 WriteAsOperand(dbgs(), LI); 1223 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1224 ); 1225 return false; 1226 } 1227 1228 // Filter out useless results (non-locals, etc). Keep track of the blocks 1229 // where we have a value available in repl, also keep track of whether we see 1230 // dependencies that produce an unknown value for the load (such as a call 1231 // that could potentially clobber the load). 1232 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1233 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1234 1235 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1236 BasicBlock *DepBB = Deps[i].getBB(); 1237 MemDepResult DepInfo = Deps[i].getResult(); 1238 1239 if (DepInfo.isClobber()) { 1240 // The address being loaded in this non-local block may not be the same as 1241 // the pointer operand of the load if PHI translation occurs. Make sure 1242 // to consider the right address. 1243 Value *Address = Deps[i].getAddress(); 1244 1245 // If the dependence is to a store that writes to a superset of the bits 1246 // read by the load, we can extract the bits we need for the load from the 1247 // stored value. 1248 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1249 if (TD && Address) { 1250 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1251 DepSI, *TD); 1252 if (Offset != -1) { 1253 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1254 DepSI->getValueOperand(), 1255 Offset)); 1256 continue; 1257 } 1258 } 1259 } 1260 1261 // Check to see if we have something like this: 1262 // load i32* P 1263 // load i8* (P+1) 1264 // if we have this, replace the later with an extraction from the former. 1265 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1266 // If this is a clobber and L is the first instruction in its block, then 1267 // we have the first instruction in the entry block. 1268 if (DepLI != LI && Address && TD) { 1269 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1270 LI->getPointerOperand(), 1271 DepLI, *TD); 1272 1273 if (Offset != -1) { 1274 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1275 Offset)); 1276 continue; 1277 } 1278 } 1279 } 1280 1281 // If the clobbering value is a memset/memcpy/memmove, see if we can 1282 // forward a value on from it. 1283 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1284 if (TD && Address) { 1285 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1286 DepMI, *TD); 1287 if (Offset != -1) { 1288 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1289 Offset)); 1290 continue; 1291 } 1292 } 1293 } 1294 1295 UnavailableBlocks.push_back(DepBB); 1296 continue; 1297 } 1298 1299 Instruction *DepInst = DepInfo.getInst(); 1300 1301 // Loading the allocation -> undef. 1302 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1303 // Loading immediately after lifetime begin -> undef. 1304 isLifetimeStart(DepInst)) { 1305 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1306 UndefValue::get(LI->getType()))); 1307 continue; 1308 } 1309 1310 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1311 // Reject loads and stores that are to the same address but are of 1312 // different types if we have to. 1313 if (S->getValueOperand()->getType() != LI->getType()) { 1314 // If the stored value is larger or equal to the loaded value, we can 1315 // reuse it. 1316 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1317 LI->getType(), *TD)) { 1318 UnavailableBlocks.push_back(DepBB); 1319 continue; 1320 } 1321 } 1322 1323 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1324 S->getValueOperand())); 1325 continue; 1326 } 1327 1328 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1329 // If the types mismatch and we can't handle it, reject reuse of the load. 1330 if (LD->getType() != LI->getType()) { 1331 // If the stored value is larger or equal to the loaded value, we can 1332 // reuse it. 1333 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1334 UnavailableBlocks.push_back(DepBB); 1335 continue; 1336 } 1337 } 1338 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1339 continue; 1340 } 1341 1342 UnavailableBlocks.push_back(DepBB); 1343 continue; 1344 } 1345 1346 // If we have no predecessors that produce a known value for this load, exit 1347 // early. 1348 if (ValuesPerBlock.empty()) return false; 1349 1350 // If all of the instructions we depend on produce a known value for this 1351 // load, then it is fully redundant and we can use PHI insertion to compute 1352 // its value. Insert PHIs and remove the fully redundant value now. 1353 if (UnavailableBlocks.empty()) { 1354 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1355 1356 // Perform PHI construction. 1357 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1358 LI->replaceAllUsesWith(V); 1359 1360 if (isa<PHINode>(V)) 1361 V->takeName(LI); 1362 if (V->getType()->isPointerTy()) 1363 MD->invalidateCachedPointerInfo(V); 1364 markInstructionForDeletion(LI); 1365 ++NumGVNLoad; 1366 return true; 1367 } 1368 1369 if (!EnablePRE || !EnableLoadPRE) 1370 return false; 1371 1372 // Okay, we have *some* definitions of the value. This means that the value 1373 // is available in some of our (transitive) predecessors. Lets think about 1374 // doing PRE of this load. This will involve inserting a new load into the 1375 // predecessor when it's not available. We could do this in general, but 1376 // prefer to not increase code size. As such, we only do this when we know 1377 // that we only have to insert *one* load (which means we're basically moving 1378 // the load, not inserting a new one). 1379 1380 SmallPtrSet<BasicBlock *, 4> Blockers; 1381 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1382 Blockers.insert(UnavailableBlocks[i]); 1383 1384 // Lets find first basic block with more than one predecessor. Walk backwards 1385 // through predecessors if needed. 1386 BasicBlock *LoadBB = LI->getParent(); 1387 BasicBlock *TmpBB = LoadBB; 1388 1389 bool isSinglePred = false; 1390 bool allSingleSucc = true; 1391 while (TmpBB->getSinglePredecessor()) { 1392 isSinglePred = true; 1393 TmpBB = TmpBB->getSinglePredecessor(); 1394 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1395 return false; 1396 if (Blockers.count(TmpBB)) 1397 return false; 1398 1399 // If any of these blocks has more than one successor (i.e. if the edge we 1400 // just traversed was critical), then there are other paths through this 1401 // block along which the load may not be anticipated. Hoisting the load 1402 // above this block would be adding the load to execution paths along 1403 // which it was not previously executed. 1404 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1405 return false; 1406 } 1407 1408 assert(TmpBB); 1409 LoadBB = TmpBB; 1410 1411 // FIXME: It is extremely unclear what this loop is doing, other than 1412 // artificially restricting loadpre. 1413 if (isSinglePred) { 1414 bool isHot = false; 1415 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1416 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1417 if (AV.isSimpleValue()) 1418 // "Hot" Instruction is in some loop (because it dominates its dep. 1419 // instruction). 1420 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1421 if (DT->dominates(LI, I)) { 1422 isHot = true; 1423 break; 1424 } 1425 } 1426 1427 // We are interested only in "hot" instructions. We don't want to do any 1428 // mis-optimizations here. 1429 if (!isHot) 1430 return false; 1431 } 1432 1433 // Check to see how many predecessors have the loaded value fully 1434 // available. 1435 DenseMap<BasicBlock*, Value*> PredLoads; 1436 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1437 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1438 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1439 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1440 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1441 1442 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1443 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1444 PI != E; ++PI) { 1445 BasicBlock *Pred = *PI; 1446 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1447 continue; 1448 } 1449 PredLoads[Pred] = 0; 1450 1451 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1452 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1453 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1454 << Pred->getName() << "': " << *LI << '\n'); 1455 return false; 1456 } 1457 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1458 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1459 } 1460 } 1461 if (!NeedToSplit.empty()) { 1462 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1463 return false; 1464 } 1465 1466 // Decide whether PRE is profitable for this load. 1467 unsigned NumUnavailablePreds = PredLoads.size(); 1468 assert(NumUnavailablePreds != 0 && 1469 "Fully available value should be eliminated above!"); 1470 1471 // If this load is unavailable in multiple predecessors, reject it. 1472 // FIXME: If we could restructure the CFG, we could make a common pred with 1473 // all the preds that don't have an available LI and insert a new load into 1474 // that one block. 1475 if (NumUnavailablePreds != 1) 1476 return false; 1477 1478 // Check if the load can safely be moved to all the unavailable predecessors. 1479 bool CanDoPRE = true; 1480 SmallVector<Instruction*, 8> NewInsts; 1481 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1482 E = PredLoads.end(); I != E; ++I) { 1483 BasicBlock *UnavailablePred = I->first; 1484 1485 // Do PHI translation to get its value in the predecessor if necessary. The 1486 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1487 1488 // If all preds have a single successor, then we know it is safe to insert 1489 // the load on the pred (?!?), so we can insert code to materialize the 1490 // pointer if it is not available. 1491 PHITransAddr Address(LI->getPointerOperand(), TD); 1492 Value *LoadPtr = 0; 1493 if (allSingleSucc) { 1494 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1495 *DT, NewInsts); 1496 } else { 1497 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1498 LoadPtr = Address.getAddr(); 1499 } 1500 1501 // If we couldn't find or insert a computation of this phi translated value, 1502 // we fail PRE. 1503 if (LoadPtr == 0) { 1504 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1505 << *LI->getPointerOperand() << "\n"); 1506 CanDoPRE = false; 1507 break; 1508 } 1509 1510 // Make sure it is valid to move this load here. We have to watch out for: 1511 // @1 = getelementptr (i8* p, ... 1512 // test p and branch if == 0 1513 // load @1 1514 // It is valid to have the getelementptr before the test, even if p can 1515 // be 0, as getelementptr only does address arithmetic. 1516 // If we are not pushing the value through any multiple-successor blocks 1517 // we do not have this case. Otherwise, check that the load is safe to 1518 // put anywhere; this can be improved, but should be conservatively safe. 1519 if (!allSingleSucc && 1520 // FIXME: REEVALUTE THIS. 1521 !isSafeToLoadUnconditionally(LoadPtr, 1522 UnavailablePred->getTerminator(), 1523 LI->getAlignment(), TD)) { 1524 CanDoPRE = false; 1525 break; 1526 } 1527 1528 I->second = LoadPtr; 1529 } 1530 1531 if (!CanDoPRE) { 1532 while (!NewInsts.empty()) { 1533 Instruction *I = NewInsts.pop_back_val(); 1534 if (MD) MD->removeInstruction(I); 1535 I->eraseFromParent(); 1536 } 1537 return false; 1538 } 1539 1540 // Okay, we can eliminate this load by inserting a reload in the predecessor 1541 // and using PHI construction to get the value in the other predecessors, do 1542 // it. 1543 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1544 DEBUG(if (!NewInsts.empty()) 1545 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1546 << *NewInsts.back() << '\n'); 1547 1548 // Assign value numbers to the new instructions. 1549 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1550 // FIXME: We really _ought_ to insert these value numbers into their 1551 // parent's availability map. However, in doing so, we risk getting into 1552 // ordering issues. If a block hasn't been processed yet, we would be 1553 // marking a value as AVAIL-IN, which isn't what we intend. 1554 VN.lookup_or_add(NewInsts[i]); 1555 } 1556 1557 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1558 E = PredLoads.end(); I != E; ++I) { 1559 BasicBlock *UnavailablePred = I->first; 1560 Value *LoadPtr = I->second; 1561 1562 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1563 LI->getAlignment(), 1564 UnavailablePred->getTerminator()); 1565 1566 // Transfer the old load's TBAA tag to the new load. 1567 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1568 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1569 1570 // Add the newly created load. 1571 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1572 NewLoad)); 1573 MD->invalidateCachedPointerInfo(LoadPtr); 1574 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1575 } 1576 1577 // Perform PHI construction. 1578 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1579 LI->replaceAllUsesWith(V); 1580 if (isa<PHINode>(V)) 1581 V->takeName(LI); 1582 if (V->getType()->isPointerTy()) 1583 MD->invalidateCachedPointerInfo(V); 1584 markInstructionForDeletion(LI); 1585 ++NumPRELoad; 1586 return true; 1587} 1588 1589/// processLoad - Attempt to eliminate a load, first by eliminating it 1590/// locally, and then attempting non-local elimination if that fails. 1591bool GVN::processLoad(LoadInst *L) { 1592 if (!MD) 1593 return false; 1594 1595 if (L->isVolatile()) 1596 return false; 1597 1598 // ... to a pointer that has been loaded from before... 1599 MemDepResult Dep = MD->getDependency(L); 1600 1601 // If we have a clobber and target data is around, see if this is a clobber 1602 // that we can fix up through code synthesis. 1603 if (Dep.isClobber() && TD) { 1604 // Check to see if we have something like this: 1605 // store i32 123, i32* %P 1606 // %A = bitcast i32* %P to i8* 1607 // %B = gep i8* %A, i32 1 1608 // %C = load i8* %B 1609 // 1610 // We could do that by recognizing if the clobber instructions are obviously 1611 // a common base + constant offset, and if the previous store (or memset) 1612 // completely covers this load. This sort of thing can happen in bitfield 1613 // access code. 1614 Value *AvailVal = 0; 1615 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1616 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1617 L->getPointerOperand(), 1618 DepSI, *TD); 1619 if (Offset != -1) 1620 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1621 L->getType(), L, *TD); 1622 } 1623 1624 // Check to see if we have something like this: 1625 // load i32* P 1626 // load i8* (P+1) 1627 // if we have this, replace the later with an extraction from the former. 1628 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1629 // If this is a clobber and L is the first instruction in its block, then 1630 // we have the first instruction in the entry block. 1631 if (DepLI == L) 1632 return false; 1633 1634 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1635 L->getPointerOperand(), 1636 DepLI, *TD); 1637 if (Offset != -1) 1638 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1639 } 1640 1641 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1642 // a value on from it. 1643 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1644 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1645 L->getPointerOperand(), 1646 DepMI, *TD); 1647 if (Offset != -1) 1648 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1649 } 1650 1651 if (AvailVal) { 1652 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1653 << *AvailVal << '\n' << *L << "\n\n\n"); 1654 1655 // Replace the load! 1656 L->replaceAllUsesWith(AvailVal); 1657 if (AvailVal->getType()->isPointerTy()) 1658 MD->invalidateCachedPointerInfo(AvailVal); 1659 markInstructionForDeletion(L); 1660 ++NumGVNLoad; 1661 return true; 1662 } 1663 } 1664 1665 // If the value isn't available, don't do anything! 1666 if (Dep.isClobber()) { 1667 DEBUG( 1668 // fast print dep, using operator<< on instruction is too slow. 1669 dbgs() << "GVN: load "; 1670 WriteAsOperand(dbgs(), L); 1671 Instruction *I = Dep.getInst(); 1672 dbgs() << " is clobbered by " << *I << '\n'; 1673 ); 1674 return false; 1675 } 1676 1677 // If it is defined in another block, try harder. 1678 if (Dep.isNonLocal()) 1679 return processNonLocalLoad(L); 1680 1681 Instruction *DepInst = Dep.getInst(); 1682 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1683 Value *StoredVal = DepSI->getValueOperand(); 1684 1685 // The store and load are to a must-aliased pointer, but they may not 1686 // actually have the same type. See if we know how to reuse the stored 1687 // value (depending on its type). 1688 if (StoredVal->getType() != L->getType()) { 1689 if (TD) { 1690 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1691 L, *TD); 1692 if (StoredVal == 0) 1693 return false; 1694 1695 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1696 << '\n' << *L << "\n\n\n"); 1697 } 1698 else 1699 return false; 1700 } 1701 1702 // Remove it! 1703 L->replaceAllUsesWith(StoredVal); 1704 if (StoredVal->getType()->isPointerTy()) 1705 MD->invalidateCachedPointerInfo(StoredVal); 1706 markInstructionForDeletion(L); 1707 ++NumGVNLoad; 1708 return true; 1709 } 1710 1711 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1712 Value *AvailableVal = DepLI; 1713 1714 // The loads are of a must-aliased pointer, but they may not actually have 1715 // the same type. See if we know how to reuse the previously loaded value 1716 // (depending on its type). 1717 if (DepLI->getType() != L->getType()) { 1718 if (TD) { 1719 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1720 L, *TD); 1721 if (AvailableVal == 0) 1722 return false; 1723 1724 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1725 << "\n" << *L << "\n\n\n"); 1726 } 1727 else 1728 return false; 1729 } 1730 1731 // Remove it! 1732 L->replaceAllUsesWith(AvailableVal); 1733 if (DepLI->getType()->isPointerTy()) 1734 MD->invalidateCachedPointerInfo(DepLI); 1735 markInstructionForDeletion(L); 1736 ++NumGVNLoad; 1737 return true; 1738 } 1739 1740 // If this load really doesn't depend on anything, then we must be loading an 1741 // undef value. This can happen when loading for a fresh allocation with no 1742 // intervening stores, for example. 1743 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1744 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1745 markInstructionForDeletion(L); 1746 ++NumGVNLoad; 1747 return true; 1748 } 1749 1750 // If this load occurs either right after a lifetime begin, 1751 // then the loaded value is undefined. 1752 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1753 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1754 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1755 markInstructionForDeletion(L); 1756 ++NumGVNLoad; 1757 return true; 1758 } 1759 } 1760 1761 return false; 1762} 1763 1764// findLeader - In order to find a leader for a given value number at a 1765// specific basic block, we first obtain the list of all Values for that number, 1766// and then scan the list to find one whose block dominates the block in 1767// question. This is fast because dominator tree queries consist of only 1768// a few comparisons of DFS numbers. 1769Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1770 LeaderTableEntry Vals = LeaderTable[num]; 1771 if (!Vals.Val) return 0; 1772 1773 Value *Val = 0; 1774 if (DT->dominates(Vals.BB, BB)) { 1775 Val = Vals.Val; 1776 if (isa<Constant>(Val)) return Val; 1777 } 1778 1779 LeaderTableEntry* Next = Vals.Next; 1780 while (Next) { 1781 if (DT->dominates(Next->BB, BB)) { 1782 if (isa<Constant>(Next->Val)) return Next->Val; 1783 if (!Val) Val = Next->Val; 1784 } 1785 1786 Next = Next->Next; 1787 } 1788 1789 return Val; 1790} 1791 1792 1793/// processInstruction - When calculating availability, handle an instruction 1794/// by inserting it into the appropriate sets 1795bool GVN::processInstruction(Instruction *I) { 1796 // Ignore dbg info intrinsics. 1797 if (isa<DbgInfoIntrinsic>(I)) 1798 return false; 1799 1800 // If the instruction can be easily simplified then do so now in preference 1801 // to value numbering it. Value numbering often exposes redundancies, for 1802 // example if it determines that %y is equal to %x then the instruction 1803 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 1804 if (Value *V = SimplifyInstruction(I, TD, DT)) { 1805 I->replaceAllUsesWith(V); 1806 if (MD && V->getType()->isPointerTy()) 1807 MD->invalidateCachedPointerInfo(V); 1808 markInstructionForDeletion(I); 1809 return true; 1810 } 1811 1812 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1813 if (processLoad(LI)) 1814 return true; 1815 1816 unsigned Num = VN.lookup_or_add(LI); 1817 addToLeaderTable(Num, LI, LI->getParent()); 1818 return false; 1819 } 1820 1821 // For conditions branches, we can perform simple conditional propagation on 1822 // the condition value itself. 1823 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1824 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1825 return false; 1826 1827 Value *BranchCond = BI->getCondition(); 1828 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1829 1830 BasicBlock *TrueSucc = BI->getSuccessor(0); 1831 BasicBlock *FalseSucc = BI->getSuccessor(1); 1832 1833 if (TrueSucc->getSinglePredecessor()) 1834 addToLeaderTable(CondVN, 1835 ConstantInt::getTrue(TrueSucc->getContext()), 1836 TrueSucc); 1837 if (FalseSucc->getSinglePredecessor()) 1838 addToLeaderTable(CondVN, 1839 ConstantInt::getFalse(TrueSucc->getContext()), 1840 FalseSucc); 1841 1842 return false; 1843 } 1844 1845 // Instructions with void type don't return a value, so there's 1846 // no point in trying to find redudancies in them. 1847 if (I->getType()->isVoidTy()) return false; 1848 1849 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1850 unsigned Num = VN.lookup_or_add(I); 1851 1852 // Allocations are always uniquely numbered, so we can save time and memory 1853 // by fast failing them. 1854 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 1855 addToLeaderTable(Num, I, I->getParent()); 1856 return false; 1857 } 1858 1859 // If the number we were assigned was a brand new VN, then we don't 1860 // need to do a lookup to see if the number already exists 1861 // somewhere in the domtree: it can't! 1862 if (Num == NextNum) { 1863 addToLeaderTable(Num, I, I->getParent()); 1864 return false; 1865 } 1866 1867 // Perform fast-path value-number based elimination of values inherited from 1868 // dominators. 1869 Value *repl = findLeader(I->getParent(), Num); 1870 if (repl == 0) { 1871 // Failure, just remember this instance for future use. 1872 addToLeaderTable(Num, I, I->getParent()); 1873 return false; 1874 } 1875 1876 // Remove it! 1877 I->replaceAllUsesWith(repl); 1878 if (MD && repl->getType()->isPointerTy()) 1879 MD->invalidateCachedPointerInfo(repl); 1880 markInstructionForDeletion(I); 1881 return true; 1882} 1883 1884/// runOnFunction - This is the main transformation entry point for a function. 1885bool GVN::runOnFunction(Function& F) { 1886 if (!NoLoads) 1887 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1888 DT = &getAnalysis<DominatorTree>(); 1889 TD = getAnalysisIfAvailable<TargetData>(); 1890 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1891 VN.setMemDep(MD); 1892 VN.setDomTree(DT); 1893 1894 bool Changed = false; 1895 bool ShouldContinue = true; 1896 1897 // Merge unconditional branches, allowing PRE to catch more 1898 // optimization opportunities. 1899 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1900 BasicBlock *BB = FI++; 1901 1902 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1903 if (removedBlock) ++NumGVNBlocks; 1904 1905 Changed |= removedBlock; 1906 } 1907 1908 unsigned Iteration = 0; 1909 while (ShouldContinue) { 1910 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 1911 ShouldContinue = iterateOnFunction(F); 1912 if (splitCriticalEdges()) 1913 ShouldContinue = true; 1914 Changed |= ShouldContinue; 1915 ++Iteration; 1916 } 1917 1918 if (EnablePRE) { 1919 bool PREChanged = true; 1920 while (PREChanged) { 1921 PREChanged = performPRE(F); 1922 Changed |= PREChanged; 1923 } 1924 } 1925 // FIXME: Should perform GVN again after PRE does something. PRE can move 1926 // computations into blocks where they become fully redundant. Note that 1927 // we can't do this until PRE's critical edge splitting updates memdep. 1928 // Actually, when this happens, we should just fully integrate PRE into GVN. 1929 1930 cleanupGlobalSets(); 1931 1932 return Changed; 1933} 1934 1935 1936bool GVN::processBlock(BasicBlock *BB) { 1937 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 1938 // (and incrementing BI before processing an instruction). 1939 assert(InstrsToErase.empty() && 1940 "We expect InstrsToErase to be empty across iterations"); 1941 bool ChangedFunction = false; 1942 1943 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1944 BI != BE;) { 1945 ChangedFunction |= processInstruction(BI); 1946 if (InstrsToErase.empty()) { 1947 ++BI; 1948 continue; 1949 } 1950 1951 // If we need some instructions deleted, do it now. 1952 NumGVNInstr += InstrsToErase.size(); 1953 1954 // Avoid iterator invalidation. 1955 bool AtStart = BI == BB->begin(); 1956 if (!AtStart) 1957 --BI; 1958 1959 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 1960 E = InstrsToErase.end(); I != E; ++I) { 1961 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 1962 if (MD) MD->removeInstruction(*I); 1963 (*I)->eraseFromParent(); 1964 DEBUG(verifyRemoved(*I)); 1965 } 1966 InstrsToErase.clear(); 1967 1968 if (AtStart) 1969 BI = BB->begin(); 1970 else 1971 ++BI; 1972 } 1973 1974 return ChangedFunction; 1975} 1976 1977/// performPRE - Perform a purely local form of PRE that looks for diamond 1978/// control flow patterns and attempts to perform simple PRE at the join point. 1979bool GVN::performPRE(Function &F) { 1980 bool Changed = false; 1981 DenseMap<BasicBlock*, Value*> predMap; 1982 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1983 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1984 BasicBlock *CurrentBlock = *DI; 1985 1986 // Nothing to PRE in the entry block. 1987 if (CurrentBlock == &F.getEntryBlock()) continue; 1988 1989 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1990 BE = CurrentBlock->end(); BI != BE; ) { 1991 Instruction *CurInst = BI++; 1992 1993 if (isa<AllocaInst>(CurInst) || 1994 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1995 CurInst->getType()->isVoidTy() || 1996 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1997 isa<DbgInfoIntrinsic>(CurInst)) 1998 continue; 1999 2000 // We don't currently value number ANY inline asm calls. 2001 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2002 if (CallI->isInlineAsm()) 2003 continue; 2004 2005 uint32_t ValNo = VN.lookup(CurInst); 2006 2007 // Look for the predecessors for PRE opportunities. We're 2008 // only trying to solve the basic diamond case, where 2009 // a value is computed in the successor and one predecessor, 2010 // but not the other. We also explicitly disallow cases 2011 // where the successor is its own predecessor, because they're 2012 // more complicated to get right. 2013 unsigned NumWith = 0; 2014 unsigned NumWithout = 0; 2015 BasicBlock *PREPred = 0; 2016 predMap.clear(); 2017 2018 for (pred_iterator PI = pred_begin(CurrentBlock), 2019 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2020 BasicBlock *P = *PI; 2021 // We're not interested in PRE where the block is its 2022 // own predecessor, or in blocks with predecessors 2023 // that are not reachable. 2024 if (P == CurrentBlock) { 2025 NumWithout = 2; 2026 break; 2027 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2028 NumWithout = 2; 2029 break; 2030 } 2031 2032 Value* predV = findLeader(P, ValNo); 2033 if (predV == 0) { 2034 PREPred = P; 2035 ++NumWithout; 2036 } else if (predV == CurInst) { 2037 NumWithout = 2; 2038 } else { 2039 predMap[P] = predV; 2040 ++NumWith; 2041 } 2042 } 2043 2044 // Don't do PRE when it might increase code size, i.e. when 2045 // we would need to insert instructions in more than one pred. 2046 if (NumWithout != 1 || NumWith == 0) 2047 continue; 2048 2049 // Don't do PRE across indirect branch. 2050 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2051 continue; 2052 2053 // We can't do PRE safely on a critical edge, so instead we schedule 2054 // the edge to be split and perform the PRE the next time we iterate 2055 // on the function. 2056 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2057 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2058 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2059 continue; 2060 } 2061 2062 // Instantiate the expression in the predecessor that lacked it. 2063 // Because we are going top-down through the block, all value numbers 2064 // will be available in the predecessor by the time we need them. Any 2065 // that weren't originally present will have been instantiated earlier 2066 // in this loop. 2067 Instruction *PREInstr = CurInst->clone(); 2068 bool success = true; 2069 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2070 Value *Op = PREInstr->getOperand(i); 2071 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2072 continue; 2073 2074 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2075 PREInstr->setOperand(i, V); 2076 } else { 2077 success = false; 2078 break; 2079 } 2080 } 2081 2082 // Fail out if we encounter an operand that is not available in 2083 // the PRE predecessor. This is typically because of loads which 2084 // are not value numbered precisely. 2085 if (!success) { 2086 delete PREInstr; 2087 DEBUG(verifyRemoved(PREInstr)); 2088 continue; 2089 } 2090 2091 PREInstr->insertBefore(PREPred->getTerminator()); 2092 PREInstr->setName(CurInst->getName() + ".pre"); 2093 predMap[PREPred] = PREInstr; 2094 VN.add(PREInstr, ValNo); 2095 ++NumGVNPRE; 2096 2097 // Update the availability map to include the new instruction. 2098 addToLeaderTable(ValNo, PREInstr, PREPred); 2099 2100 // Create a PHI to make the value available in this block. 2101 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2102 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2103 CurInst->getName() + ".pre-phi", 2104 CurrentBlock->begin()); 2105 for (pred_iterator PI = PB; PI != PE; ++PI) { 2106 BasicBlock *P = *PI; 2107 Phi->addIncoming(predMap[P], P); 2108 } 2109 2110 VN.add(Phi, ValNo); 2111 addToLeaderTable(ValNo, Phi, CurrentBlock); 2112 2113 CurInst->replaceAllUsesWith(Phi); 2114 if (Phi->getType()->isPointerTy()) { 2115 // Because we have added a PHI-use of the pointer value, it has now 2116 // "escaped" from alias analysis' perspective. We need to inform 2117 // AA of this. 2118 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) 2119 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii)); 2120 2121 if (MD) 2122 MD->invalidateCachedPointerInfo(Phi); 2123 } 2124 VN.erase(CurInst); 2125 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2126 2127 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2128 if (MD) MD->removeInstruction(CurInst); 2129 CurInst->eraseFromParent(); 2130 DEBUG(verifyRemoved(CurInst)); 2131 Changed = true; 2132 } 2133 } 2134 2135 if (splitCriticalEdges()) 2136 Changed = true; 2137 2138 return Changed; 2139} 2140 2141/// splitCriticalEdges - Split critical edges found during the previous 2142/// iteration that may enable further optimization. 2143bool GVN::splitCriticalEdges() { 2144 if (toSplit.empty()) 2145 return false; 2146 do { 2147 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2148 SplitCriticalEdge(Edge.first, Edge.second, this); 2149 } while (!toSplit.empty()); 2150 if (MD) MD->invalidateCachedPredecessors(); 2151 return true; 2152} 2153 2154/// iterateOnFunction - Executes one iteration of GVN 2155bool GVN::iterateOnFunction(Function &F) { 2156 cleanupGlobalSets(); 2157 2158 // Top-down walk of the dominator tree 2159 bool Changed = false; 2160#if 0 2161 // Needed for value numbering with phi construction to work. 2162 ReversePostOrderTraversal<Function*> RPOT(&F); 2163 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2164 RE = RPOT.end(); RI != RE; ++RI) 2165 Changed |= processBlock(*RI); 2166#else 2167 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2168 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2169 Changed |= processBlock(DI->getBlock()); 2170#endif 2171 2172 return Changed; 2173} 2174 2175void GVN::cleanupGlobalSets() { 2176 VN.clear(); 2177 LeaderTable.clear(); 2178 TableAllocator.Reset(); 2179} 2180 2181/// verifyRemoved - Verify that the specified instruction does not occur in our 2182/// internal data structures. 2183void GVN::verifyRemoved(const Instruction *Inst) const { 2184 VN.verifyRemoved(Inst); 2185 2186 // Walk through the value number scope to make sure the instruction isn't 2187 // ferreted away in it. 2188 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2189 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2190 const LeaderTableEntry *Node = &I->second; 2191 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2192 2193 while (Node->Next) { 2194 Node = Node->Next; 2195 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2196 } 2197 } 2198} 2199