GVN.cpp revision a990e071f2f29ba326b97a4288207a2c406c5b66
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: Move the checking logic to MemDep! 231 CallInst* cdep = 0; 232 233 // Check to see if we have a single dominating call instruction that is 234 // identical to C. 235 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 236 const NonLocalDepEntry *I = &deps[i]; 237 if (I->getResult().isNonLocal()) 238 continue; 239 240 // We don't handle non-definitions. If we already have a call, reject 241 // instruction dependencies. 242 if (!I->getResult().isDef() || cdep != 0) { 243 cdep = 0; 244 break; 245 } 246 247 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 248 // FIXME: All duplicated with non-local case. 249 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 250 cdep = NonLocalDepCall; 251 continue; 252 } 253 254 cdep = 0; 255 break; 256 } 257 258 if (!cdep) { 259 valueNumbering[C] = nextValueNumber; 260 return nextValueNumber++; 261 } 262 263 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 264 valueNumbering[C] = nextValueNumber; 265 return nextValueNumber++; 266 } 267 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 268 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 269 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 270 if (c_vn != cd_vn) { 271 valueNumbering[C] = nextValueNumber; 272 return nextValueNumber++; 273 } 274 } 275 276 uint32_t v = lookup_or_add(cdep); 277 valueNumbering[C] = v; 278 return v; 279 280 } else { 281 valueNumbering[C] = nextValueNumber; 282 return nextValueNumber++; 283 } 284} 285 286/// lookup_or_add - Returns the value number for the specified value, assigning 287/// it a new number if it did not have one before. 288uint32_t ValueTable::lookup_or_add(Value *V) { 289 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 290 if (VI != valueNumbering.end()) 291 return VI->second; 292 293 if (!isa<Instruction>(V)) { 294 valueNumbering[V] = nextValueNumber; 295 return nextValueNumber++; 296 } 297 298 Instruction* I = cast<Instruction>(V); 299 Expression exp; 300 switch (I->getOpcode()) { 301 case Instruction::Call: 302 return lookup_or_add_call(cast<CallInst>(I)); 303 case Instruction::Add: 304 case Instruction::FAdd: 305 case Instruction::Sub: 306 case Instruction::FSub: 307 case Instruction::Mul: 308 case Instruction::FMul: 309 case Instruction::UDiv: 310 case Instruction::SDiv: 311 case Instruction::FDiv: 312 case Instruction::URem: 313 case Instruction::SRem: 314 case Instruction::FRem: 315 case Instruction::Shl: 316 case Instruction::LShr: 317 case Instruction::AShr: 318 case Instruction::And: 319 case Instruction::Or : 320 case Instruction::Xor: 321 case Instruction::ICmp: 322 case Instruction::FCmp: 323 case Instruction::Trunc: 324 case Instruction::ZExt: 325 case Instruction::SExt: 326 case Instruction::FPToUI: 327 case Instruction::FPToSI: 328 case Instruction::UIToFP: 329 case Instruction::SIToFP: 330 case Instruction::FPTrunc: 331 case Instruction::FPExt: 332 case Instruction::PtrToInt: 333 case Instruction::IntToPtr: 334 case Instruction::BitCast: 335 case Instruction::Select: 336 case Instruction::ExtractElement: 337 case Instruction::InsertElement: 338 case Instruction::ShuffleVector: 339 case Instruction::ExtractValue: 340 case Instruction::InsertValue: 341 case Instruction::GetElementPtr: 342 exp = create_expression(I); 343 break; 344 default: 345 valueNumbering[V] = nextValueNumber; 346 return nextValueNumber++; 347 } 348 349 uint32_t& e = expressionNumbering[exp]; 350 if (!e) e = nextValueNumber++; 351 valueNumbering[V] = e; 352 return e; 353} 354 355/// lookup - Returns the value number of the specified value. Fails if 356/// the value has not yet been numbered. 357uint32_t ValueTable::lookup(Value *V) const { 358 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 359 assert(VI != valueNumbering.end() && "Value not numbered?"); 360 return VI->second; 361} 362 363/// clear - Remove all entries from the ValueTable. 364void ValueTable::clear() { 365 valueNumbering.clear(); 366 expressionNumbering.clear(); 367 nextValueNumber = 1; 368} 369 370/// erase - Remove a value from the value numbering. 371void ValueTable::erase(Value *V) { 372 valueNumbering.erase(V); 373} 374 375/// verifyRemoved - Verify that the value is removed from all internal data 376/// structures. 377void ValueTable::verifyRemoved(const Value *V) const { 378 for (DenseMap<Value*, uint32_t>::const_iterator 379 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 380 assert(I->first != V && "Inst still occurs in value numbering map!"); 381 } 382} 383 384//===----------------------------------------------------------------------===// 385// GVN Pass 386//===----------------------------------------------------------------------===// 387 388namespace { 389 390 class GVN : public FunctionPass { 391 bool NoLoads; 392 MemoryDependenceAnalysis *MD; 393 DominatorTree *DT; 394 const TargetData *TD; 395 396 ValueTable VN; 397 398 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 399 /// have that value number. Use findLeader to query it. 400 struct LeaderTableEntry { 401 Value *Val; 402 BasicBlock *BB; 403 LeaderTableEntry *Next; 404 }; 405 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 406 BumpPtrAllocator TableAllocator; 407 408 SmallVector<Instruction*, 8> InstrsToErase; 409 public: 410 static char ID; // Pass identification, replacement for typeid 411 explicit GVN(bool noloads = false) 412 : FunctionPass(ID), NoLoads(noloads), MD(0) { 413 initializeGVNPass(*PassRegistry::getPassRegistry()); 414 } 415 416 bool runOnFunction(Function &F); 417 418 /// markInstructionForDeletion - This removes the specified instruction from 419 /// our various maps and marks it for deletion. 420 void markInstructionForDeletion(Instruction *I) { 421 VN.erase(I); 422 InstrsToErase.push_back(I); 423 } 424 425 const TargetData *getTargetData() const { return TD; } 426 DominatorTree &getDominatorTree() const { return *DT; } 427 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 428 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 429 private: 430 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 431 /// its value number. 432 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 433 LeaderTableEntry &Curr = LeaderTable[N]; 434 if (!Curr.Val) { 435 Curr.Val = V; 436 Curr.BB = BB; 437 return; 438 } 439 440 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 441 Node->Val = V; 442 Node->BB = BB; 443 Node->Next = Curr.Next; 444 Curr.Next = Node; 445 } 446 447 /// removeFromLeaderTable - Scan the list of values corresponding to a given 448 /// value number, and remove the given value if encountered. 449 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 450 LeaderTableEntry* Prev = 0; 451 LeaderTableEntry* Curr = &LeaderTable[N]; 452 453 while (Curr->Val != V || Curr->BB != BB) { 454 Prev = Curr; 455 Curr = Curr->Next; 456 } 457 458 if (Prev) { 459 Prev->Next = Curr->Next; 460 } else { 461 if (!Curr->Next) { 462 Curr->Val = 0; 463 Curr->BB = 0; 464 } else { 465 LeaderTableEntry* Next = Curr->Next; 466 Curr->Val = Next->Val; 467 Curr->BB = Next->BB; 468 Curr->Next = Next->Next; 469 } 470 } 471 } 472 473 // List of critical edges to be split between iterations. 474 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 475 476 // This transformation requires dominator postdominator info 477 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 478 AU.addRequired<DominatorTree>(); 479 if (!NoLoads) 480 AU.addRequired<MemoryDependenceAnalysis>(); 481 AU.addRequired<AliasAnalysis>(); 482 483 AU.addPreserved<DominatorTree>(); 484 AU.addPreserved<AliasAnalysis>(); 485 } 486 487 488 // Helper fuctions 489 // FIXME: eliminate or document these better 490 bool processLoad(LoadInst *L); 491 bool processInstruction(Instruction *I); 492 bool processNonLocalLoad(LoadInst *L); 493 bool processBlock(BasicBlock *BB); 494 void dump(DenseMap<uint32_t, Value*> &d); 495 bool iterateOnFunction(Function &F); 496 bool performPRE(Function &F); 497 Value *findLeader(BasicBlock *BB, uint32_t num); 498 void cleanupGlobalSets(); 499 void verifyRemoved(const Instruction *I) const; 500 bool splitCriticalEdges(); 501 }; 502 503 char GVN::ID = 0; 504} 505 506// createGVNPass - The public interface to this file... 507FunctionPass *llvm::createGVNPass(bool NoLoads) { 508 return new GVN(NoLoads); 509} 510 511INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 512INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 513INITIALIZE_PASS_DEPENDENCY(DominatorTree) 514INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 515INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 516 517void GVN::dump(DenseMap<uint32_t, Value*>& d) { 518 errs() << "{\n"; 519 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 520 E = d.end(); I != E; ++I) { 521 errs() << I->first << "\n"; 522 I->second->dump(); 523 } 524 errs() << "}\n"; 525} 526 527/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 528/// we're analyzing is fully available in the specified block. As we go, keep 529/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 530/// map is actually a tri-state map with the following values: 531/// 0) we know the block *is not* fully available. 532/// 1) we know the block *is* fully available. 533/// 2) we do not know whether the block is fully available or not, but we are 534/// currently speculating that it will be. 535/// 3) we are speculating for this block and have used that to speculate for 536/// other blocks. 537static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 538 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 539 // Optimistically assume that the block is fully available and check to see 540 // if we already know about this block in one lookup. 541 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 542 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 543 544 // If the entry already existed for this block, return the precomputed value. 545 if (!IV.second) { 546 // If this is a speculative "available" value, mark it as being used for 547 // speculation of other blocks. 548 if (IV.first->second == 2) 549 IV.first->second = 3; 550 return IV.first->second != 0; 551 } 552 553 // Otherwise, see if it is fully available in all predecessors. 554 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 555 556 // If this block has no predecessors, it isn't live-in here. 557 if (PI == PE) 558 goto SpeculationFailure; 559 560 for (; PI != PE; ++PI) 561 // If the value isn't fully available in one of our predecessors, then it 562 // isn't fully available in this block either. Undo our previous 563 // optimistic assumption and bail out. 564 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 565 goto SpeculationFailure; 566 567 return true; 568 569// SpeculationFailure - If we get here, we found out that this is not, after 570// all, a fully-available block. We have a problem if we speculated on this and 571// used the speculation to mark other blocks as available. 572SpeculationFailure: 573 char &BBVal = FullyAvailableBlocks[BB]; 574 575 // If we didn't speculate on this, just return with it set to false. 576 if (BBVal == 2) { 577 BBVal = 0; 578 return false; 579 } 580 581 // If we did speculate on this value, we could have blocks set to 1 that are 582 // incorrect. Walk the (transitive) successors of this block and mark them as 583 // 0 if set to one. 584 SmallVector<BasicBlock*, 32> BBWorklist; 585 BBWorklist.push_back(BB); 586 587 do { 588 BasicBlock *Entry = BBWorklist.pop_back_val(); 589 // Note that this sets blocks to 0 (unavailable) if they happen to not 590 // already be in FullyAvailableBlocks. This is safe. 591 char &EntryVal = FullyAvailableBlocks[Entry]; 592 if (EntryVal == 0) continue; // Already unavailable. 593 594 // Mark as unavailable. 595 EntryVal = 0; 596 597 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 598 BBWorklist.push_back(*I); 599 } while (!BBWorklist.empty()); 600 601 return false; 602} 603 604 605/// CanCoerceMustAliasedValueToLoad - Return true if 606/// CoerceAvailableValueToLoadType will succeed. 607static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 608 const Type *LoadTy, 609 const TargetData &TD) { 610 // If the loaded or stored value is an first class array or struct, don't try 611 // to transform them. We need to be able to bitcast to integer. 612 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 613 StoredVal->getType()->isStructTy() || 614 StoredVal->getType()->isArrayTy()) 615 return false; 616 617 // The store has to be at least as big as the load. 618 if (TD.getTypeSizeInBits(StoredVal->getType()) < 619 TD.getTypeSizeInBits(LoadTy)) 620 return false; 621 622 return true; 623} 624 625 626/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 627/// then a load from a must-aliased pointer of a different type, try to coerce 628/// the stored value. LoadedTy is the type of the load we want to replace and 629/// InsertPt is the place to insert new instructions. 630/// 631/// If we can't do it, return null. 632static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 633 const Type *LoadedTy, 634 Instruction *InsertPt, 635 const TargetData &TD) { 636 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 637 return 0; 638 639 // If this is already the right type, just return it. 640 const Type *StoredValTy = StoredVal->getType(); 641 642 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 643 uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy); 644 645 // If the store and reload are the same size, we can always reuse it. 646 if (StoreSize == LoadSize) { 647 // Pointer to Pointer -> use bitcast. 648 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 649 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 650 651 // Convert source pointers to integers, which can be bitcast. 652 if (StoredValTy->isPointerTy()) { 653 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 654 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 655 } 656 657 const Type *TypeToCastTo = LoadedTy; 658 if (TypeToCastTo->isPointerTy()) 659 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 660 661 if (StoredValTy != TypeToCastTo) 662 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 663 664 // Cast to pointer if the load needs a pointer type. 665 if (LoadedTy->isPointerTy()) 666 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 667 668 return StoredVal; 669 } 670 671 // If the loaded value is smaller than the available value, then we can 672 // extract out a piece from it. If the available value is too small, then we 673 // can't do anything. 674 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 675 676 // Convert source pointers to integers, which can be manipulated. 677 if (StoredValTy->isPointerTy()) { 678 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 679 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 680 } 681 682 // Convert vectors and fp to integer, which can be manipulated. 683 if (!StoredValTy->isIntegerTy()) { 684 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 685 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 686 } 687 688 // If this is a big-endian system, we need to shift the value down to the low 689 // bits so that a truncate will work. 690 if (TD.isBigEndian()) { 691 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 692 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 693 } 694 695 // Truncate the integer to the right size now. 696 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 697 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 698 699 if (LoadedTy == NewIntTy) 700 return StoredVal; 701 702 // If the result is a pointer, inttoptr. 703 if (LoadedTy->isPointerTy()) 704 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 705 706 // Otherwise, bitcast. 707 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 708} 709 710/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 711/// memdep query of a load that ends up being a clobbering memory write (store, 712/// memset, memcpy, memmove). This means that the write *may* provide bits used 713/// by the load but we can't be sure because the pointers don't mustalias. 714/// 715/// Check this case to see if there is anything more we can do before we give 716/// up. This returns -1 if we have to give up, or a byte number in the stored 717/// value of the piece that feeds the load. 718static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 719 Value *WritePtr, 720 uint64_t WriteSizeInBits, 721 const TargetData &TD) { 722 // If the loaded or stored value is an first class array or struct, don't try 723 // to transform them. We need to be able to bitcast to integer. 724 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 725 return -1; 726 727 int64_t StoreOffset = 0, LoadOffset = 0; 728 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 729 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 730 if (StoreBase != LoadBase) 731 return -1; 732 733 // If the load and store are to the exact same address, they should have been 734 // a must alias. AA must have gotten confused. 735 // FIXME: Study to see if/when this happens. One case is forwarding a memset 736 // to a load from the base of the memset. 737#if 0 738 if (LoadOffset == StoreOffset) { 739 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 740 << "Base = " << *StoreBase << "\n" 741 << "Store Ptr = " << *WritePtr << "\n" 742 << "Store Offs = " << StoreOffset << "\n" 743 << "Load Ptr = " << *LoadPtr << "\n"; 744 abort(); 745 } 746#endif 747 748 // If the load and store don't overlap at all, the store doesn't provide 749 // anything to the load. In this case, they really don't alias at all, AA 750 // must have gotten confused. 751 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 752 753 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 754 return -1; 755 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 756 LoadSize >>= 3; 757 758 759 bool isAAFailure = false; 760 if (StoreOffset < LoadOffset) 761 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 762 else 763 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 764 765 if (isAAFailure) { 766#if 0 767 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 768 << "Base = " << *StoreBase << "\n" 769 << "Store Ptr = " << *WritePtr << "\n" 770 << "Store Offs = " << StoreOffset << "\n" 771 << "Load Ptr = " << *LoadPtr << "\n"; 772 abort(); 773#endif 774 return -1; 775 } 776 777 // If the Load isn't completely contained within the stored bits, we don't 778 // have all the bits to feed it. We could do something crazy in the future 779 // (issue a smaller load then merge the bits in) but this seems unlikely to be 780 // valuable. 781 if (StoreOffset > LoadOffset || 782 StoreOffset+StoreSize < LoadOffset+LoadSize) 783 return -1; 784 785 // Okay, we can do this transformation. Return the number of bytes into the 786 // store that the load is. 787 return LoadOffset-StoreOffset; 788} 789 790/// AnalyzeLoadFromClobberingStore - This function is called when we have a 791/// memdep query of a load that ends up being a clobbering store. 792static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 793 StoreInst *DepSI, 794 const TargetData &TD) { 795 // Cannot handle reading from store of first-class aggregate yet. 796 if (DepSI->getValueOperand()->getType()->isStructTy() || 797 DepSI->getValueOperand()->getType()->isArrayTy()) 798 return -1; 799 800 Value *StorePtr = DepSI->getPointerOperand(); 801 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 802 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 803 StorePtr, StoreSize, TD); 804} 805 806/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 807/// memdep query of a load that ends up being clobbered by another load. See if 808/// the other load can feed into the second load. 809static int AnalyzeLoadFromClobberingLoad(const Type *LoadTy, Value *LoadPtr, 810 LoadInst *DepLI, const TargetData &TD){ 811 // Cannot handle reading from store of first-class aggregate yet. 812 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 813 return -1; 814 815 Value *DepPtr = DepLI->getPointerOperand(); 816 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 817 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 818 if (R != -1) return R; 819 820 // If we have a load/load clobber an DepLI can be widened to cover this load, 821 // then we should widen it! 822 int64_t LoadOffs = 0; 823 const Value *LoadBase = 824 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 825 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 826 827 unsigned Size = MemoryDependenceAnalysis:: 828 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 829 if (Size == 0) return -1; 830 831 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 832} 833 834 835 836static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 837 MemIntrinsic *MI, 838 const TargetData &TD) { 839 // If the mem operation is a non-constant size, we can't handle it. 840 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 841 if (SizeCst == 0) return -1; 842 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 843 844 // If this is memset, we just need to see if the offset is valid in the size 845 // of the memset.. 846 if (MI->getIntrinsicID() == Intrinsic::memset) 847 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 848 MemSizeInBits, TD); 849 850 // If we have a memcpy/memmove, the only case we can handle is if this is a 851 // copy from constant memory. In that case, we can read directly from the 852 // constant memory. 853 MemTransferInst *MTI = cast<MemTransferInst>(MI); 854 855 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 856 if (Src == 0) return -1; 857 858 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 859 if (GV == 0 || !GV->isConstant()) return -1; 860 861 // See if the access is within the bounds of the transfer. 862 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 863 MI->getDest(), MemSizeInBits, TD); 864 if (Offset == -1) 865 return Offset; 866 867 // Otherwise, see if we can constant fold a load from the constant with the 868 // offset applied as appropriate. 869 Src = ConstantExpr::getBitCast(Src, 870 llvm::Type::getInt8PtrTy(Src->getContext())); 871 Constant *OffsetCst = 872 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 873 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 874 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 875 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 876 return Offset; 877 return -1; 878} 879 880 881/// GetStoreValueForLoad - This function is called when we have a 882/// memdep query of a load that ends up being a clobbering store. This means 883/// that the store provides bits used by the load but we the pointers don't 884/// mustalias. Check this case to see if there is anything more we can do 885/// before we give up. 886static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 887 const Type *LoadTy, 888 Instruction *InsertPt, const TargetData &TD){ 889 LLVMContext &Ctx = SrcVal->getType()->getContext(); 890 891 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 892 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 893 894 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 895 896 // Compute which bits of the stored value are being used by the load. Convert 897 // to an integer type to start with. 898 if (SrcVal->getType()->isPointerTy()) 899 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 900 if (!SrcVal->getType()->isIntegerTy()) 901 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 902 "tmp"); 903 904 // Shift the bits to the least significant depending on endianness. 905 unsigned ShiftAmt; 906 if (TD.isLittleEndian()) 907 ShiftAmt = Offset*8; 908 else 909 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 910 911 if (ShiftAmt) 912 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 913 914 if (LoadSize != StoreSize) 915 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 916 "tmp"); 917 918 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 919} 920 921/// GetStoreValueForLoad - This function is called when we have a 922/// memdep query of a load that ends up being a clobbering load. This means 923/// that the load *may* provide bits used by the load but we can't be sure 924/// because the pointers don't mustalias. Check this case to see if there is 925/// anything more we can do before we give up. 926static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 927 const Type *LoadTy, Instruction *InsertPt, 928 GVN &gvn) { 929 const TargetData &TD = *gvn.getTargetData(); 930 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 931 // widen SrcVal out to a larger load. 932 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 933 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 934 if (Offset+LoadSize > SrcValSize) { 935 assert(!SrcVal->isVolatile() && "Cannot widen volatile load!"); 936 assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load"); 937 // If we have a load/load clobber an DepLI can be widened to cover this 938 // load, then we should widen it to the next power of 2 size big enough! 939 unsigned NewLoadSize = Offset+LoadSize; 940 if (!isPowerOf2_32(NewLoadSize)) 941 NewLoadSize = NextPowerOf2(NewLoadSize); 942 943 Value *PtrVal = SrcVal->getPointerOperand(); 944 945 // Insert the new load after the old load. This ensures that subsequent 946 // memdep queries will find the new load. We can't easily remove the old 947 // load completely because it is already in the value numbering table. 948 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 949 const Type *DestPTy = 950 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 951 DestPTy = PointerType::get(DestPTy, 952 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 953 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 954 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 955 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 956 NewLoad->takeName(SrcVal); 957 NewLoad->setAlignment(SrcVal->getAlignment()); 958 959 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 960 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 961 962 // Replace uses of the original load with the wider load. On a big endian 963 // system, we need to shift down to get the relevant bits. 964 Value *RV = NewLoad; 965 if (TD.isBigEndian()) 966 RV = Builder.CreateLShr(RV, 967 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 968 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 969 SrcVal->replaceAllUsesWith(RV); 970 971 // We would like to use gvn.markInstructionForDeletion here, but we can't 972 // because the load is already memoized into the leader map table that GVN 973 // tracks. It is potentially possible to remove the load from the table, 974 // but then there all of the operations based on it would need to be 975 // rehashed. Just leave the dead load around. 976 gvn.getMemDep().removeInstruction(SrcVal); 977 SrcVal = NewLoad; 978 } 979 980 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 981} 982 983 984/// GetMemInstValueForLoad - This function is called when we have a 985/// memdep query of a load that ends up being a clobbering mem intrinsic. 986static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 987 const Type *LoadTy, Instruction *InsertPt, 988 const TargetData &TD){ 989 LLVMContext &Ctx = LoadTy->getContext(); 990 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 991 992 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 993 994 // We know that this method is only called when the mem transfer fully 995 // provides the bits for the load. 996 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 997 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 998 // independently of what the offset is. 999 Value *Val = MSI->getValue(); 1000 if (LoadSize != 1) 1001 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1002 1003 Value *OneElt = Val; 1004 1005 // Splat the value out to the right number of bits. 1006 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1007 // If we can double the number of bytes set, do it. 1008 if (NumBytesSet*2 <= LoadSize) { 1009 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1010 Val = Builder.CreateOr(Val, ShVal); 1011 NumBytesSet <<= 1; 1012 continue; 1013 } 1014 1015 // Otherwise insert one byte at a time. 1016 Value *ShVal = Builder.CreateShl(Val, 1*8); 1017 Val = Builder.CreateOr(OneElt, ShVal); 1018 ++NumBytesSet; 1019 } 1020 1021 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1022 } 1023 1024 // Otherwise, this is a memcpy/memmove from a constant global. 1025 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1026 Constant *Src = cast<Constant>(MTI->getSource()); 1027 1028 // Otherwise, see if we can constant fold a load from the constant with the 1029 // offset applied as appropriate. 1030 Src = ConstantExpr::getBitCast(Src, 1031 llvm::Type::getInt8PtrTy(Src->getContext())); 1032 Constant *OffsetCst = 1033 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1034 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1035 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1036 return ConstantFoldLoadFromConstPtr(Src, &TD); 1037} 1038 1039namespace { 1040 1041struct AvailableValueInBlock { 1042 /// BB - The basic block in question. 1043 BasicBlock *BB; 1044 enum ValType { 1045 SimpleVal, // A simple offsetted value that is accessed. 1046 LoadVal, // A value produced by a load. 1047 MemIntrin // A memory intrinsic which is loaded from. 1048 }; 1049 1050 /// V - The value that is live out of the block. 1051 PointerIntPair<Value *, 2, ValType> Val; 1052 1053 /// Offset - The byte offset in Val that is interesting for the load query. 1054 unsigned Offset; 1055 1056 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1057 unsigned Offset = 0) { 1058 AvailableValueInBlock Res; 1059 Res.BB = BB; 1060 Res.Val.setPointer(V); 1061 Res.Val.setInt(SimpleVal); 1062 Res.Offset = Offset; 1063 return Res; 1064 } 1065 1066 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1067 unsigned Offset = 0) { 1068 AvailableValueInBlock Res; 1069 Res.BB = BB; 1070 Res.Val.setPointer(MI); 1071 Res.Val.setInt(MemIntrin); 1072 Res.Offset = Offset; 1073 return Res; 1074 } 1075 1076 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1077 unsigned Offset = 0) { 1078 AvailableValueInBlock Res; 1079 Res.BB = BB; 1080 Res.Val.setPointer(LI); 1081 Res.Val.setInt(LoadVal); 1082 Res.Offset = Offset; 1083 return Res; 1084 } 1085 1086 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1087 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1088 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1089 1090 Value *getSimpleValue() const { 1091 assert(isSimpleValue() && "Wrong accessor"); 1092 return Val.getPointer(); 1093 } 1094 1095 LoadInst *getCoercedLoadValue() const { 1096 assert(isCoercedLoadValue() && "Wrong accessor"); 1097 return cast<LoadInst>(Val.getPointer()); 1098 } 1099 1100 MemIntrinsic *getMemIntrinValue() const { 1101 assert(isMemIntrinValue() && "Wrong accessor"); 1102 return cast<MemIntrinsic>(Val.getPointer()); 1103 } 1104 1105 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1106 /// defined here to the specified type. This handles various coercion cases. 1107 Value *MaterializeAdjustedValue(const Type *LoadTy, GVN &gvn) const { 1108 Value *Res; 1109 if (isSimpleValue()) { 1110 Res = getSimpleValue(); 1111 if (Res->getType() != LoadTy) { 1112 const TargetData *TD = gvn.getTargetData(); 1113 assert(TD && "Need target data to handle type mismatch case"); 1114 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1115 *TD); 1116 1117 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1118 << *getSimpleValue() << '\n' 1119 << *Res << '\n' << "\n\n\n"); 1120 } 1121 } else if (isCoercedLoadValue()) { 1122 LoadInst *Load = getCoercedLoadValue(); 1123 if (Load->getType() == LoadTy && Offset == 0) { 1124 Res = Load; 1125 } else { 1126 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1127 gvn); 1128 1129 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1130 << *getCoercedLoadValue() << '\n' 1131 << *Res << '\n' << "\n\n\n"); 1132 } 1133 } else { 1134 const TargetData *TD = gvn.getTargetData(); 1135 assert(TD && "Need target data to handle type mismatch case"); 1136 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1137 LoadTy, BB->getTerminator(), *TD); 1138 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1139 << " " << *getMemIntrinValue() << '\n' 1140 << *Res << '\n' << "\n\n\n"); 1141 } 1142 return Res; 1143 } 1144}; 1145 1146} // end anonymous namespace 1147 1148/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1149/// construct SSA form, allowing us to eliminate LI. This returns the value 1150/// that should be used at LI's definition site. 1151static Value *ConstructSSAForLoadSet(LoadInst *LI, 1152 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1153 GVN &gvn) { 1154 // Check for the fully redundant, dominating load case. In this case, we can 1155 // just use the dominating value directly. 1156 if (ValuesPerBlock.size() == 1 && 1157 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1158 LI->getParent())) 1159 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1160 1161 // Otherwise, we have to construct SSA form. 1162 SmallVector<PHINode*, 8> NewPHIs; 1163 SSAUpdater SSAUpdate(&NewPHIs); 1164 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1165 1166 const Type *LoadTy = LI->getType(); 1167 1168 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1169 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1170 BasicBlock *BB = AV.BB; 1171 1172 if (SSAUpdate.HasValueForBlock(BB)) 1173 continue; 1174 1175 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1176 } 1177 1178 // Perform PHI construction. 1179 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1180 1181 // If new PHI nodes were created, notify alias analysis. 1182 if (V->getType()->isPointerTy()) { 1183 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1184 1185 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1186 AA->copyValue(LI, NewPHIs[i]); 1187 1188 // Now that we've copied information to the new PHIs, scan through 1189 // them again and inform alias analysis that we've added potentially 1190 // escaping uses to any values that are operands to these PHIs. 1191 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1192 PHINode *P = NewPHIs[i]; 1193 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) 1194 AA->addEscapingUse(P->getOperandUse(2*ii)); 1195 } 1196 } 1197 1198 return V; 1199} 1200 1201static bool isLifetimeStart(const Instruction *Inst) { 1202 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1203 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1204 return false; 1205} 1206 1207/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1208/// non-local by performing PHI construction. 1209bool GVN::processNonLocalLoad(LoadInst *LI) { 1210 // Find the non-local dependencies of the load. 1211 SmallVector<NonLocalDepResult, 64> Deps; 1212 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1213 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1214 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1215 // << Deps.size() << *LI << '\n'); 1216 1217 // If we had to process more than one hundred blocks to find the 1218 // dependencies, this load isn't worth worrying about. Optimizing 1219 // it will be too expensive. 1220 if (Deps.size() > 100) 1221 return false; 1222 1223 // If we had a phi translation failure, we'll have a single entry which is a 1224 // clobber in the current block. Reject this early. 1225 if (Deps.size() == 1 && Deps[0].getResult().isUnknown()) { 1226 DEBUG( 1227 dbgs() << "GVN: non-local load "; 1228 WriteAsOperand(dbgs(), LI); 1229 dbgs() << " has unknown dependencies\n"; 1230 ); 1231 return false; 1232 } 1233 1234 // Filter out useless results (non-locals, etc). Keep track of the blocks 1235 // where we have a value available in repl, also keep track of whether we see 1236 // dependencies that produce an unknown value for the load (such as a call 1237 // that could potentially clobber the load). 1238 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1239 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1240 1241 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1242 BasicBlock *DepBB = Deps[i].getBB(); 1243 MemDepResult DepInfo = Deps[i].getResult(); 1244 1245 if (DepInfo.isUnknown()) { 1246 UnavailableBlocks.push_back(DepBB); 1247 continue; 1248 } 1249 1250 if (DepInfo.isClobber()) { 1251 // The address being loaded in this non-local block may not be the same as 1252 // the pointer operand of the load if PHI translation occurs. Make sure 1253 // to consider the right address. 1254 Value *Address = Deps[i].getAddress(); 1255 1256 // If the dependence is to a store that writes to a superset of the bits 1257 // read by the load, we can extract the bits we need for the load from the 1258 // stored value. 1259 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1260 if (TD && Address) { 1261 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1262 DepSI, *TD); 1263 if (Offset != -1) { 1264 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1265 DepSI->getValueOperand(), 1266 Offset)); 1267 continue; 1268 } 1269 } 1270 } 1271 1272 // Check to see if we have something like this: 1273 // load i32* P 1274 // load i8* (P+1) 1275 // if we have this, replace the later with an extraction from the former. 1276 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1277 // If this is a clobber and L is the first instruction in its block, then 1278 // we have the first instruction in the entry block. 1279 if (DepLI != LI && Address && TD) { 1280 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1281 LI->getPointerOperand(), 1282 DepLI, *TD); 1283 1284 if (Offset != -1) { 1285 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1286 Offset)); 1287 continue; 1288 } 1289 } 1290 } 1291 1292 // If the clobbering value is a memset/memcpy/memmove, see if we can 1293 // forward a value on from it. 1294 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1295 if (TD && Address) { 1296 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1297 DepMI, *TD); 1298 if (Offset != -1) { 1299 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1300 Offset)); 1301 continue; 1302 } 1303 } 1304 } 1305 1306 UnavailableBlocks.push_back(DepBB); 1307 continue; 1308 } 1309 1310 assert(DepInfo.isDef() && "Expecting def here"); 1311 1312 Instruction *DepInst = DepInfo.getInst(); 1313 1314 // Loading the allocation -> undef. 1315 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1316 // Loading immediately after lifetime begin -> undef. 1317 isLifetimeStart(DepInst)) { 1318 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1319 UndefValue::get(LI->getType()))); 1320 continue; 1321 } 1322 1323 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1324 // Reject loads and stores that are to the same address but are of 1325 // different types if we have to. 1326 if (S->getValueOperand()->getType() != LI->getType()) { 1327 // If the stored value is larger or equal to the loaded value, we can 1328 // reuse it. 1329 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1330 LI->getType(), *TD)) { 1331 UnavailableBlocks.push_back(DepBB); 1332 continue; 1333 } 1334 } 1335 1336 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1337 S->getValueOperand())); 1338 continue; 1339 } 1340 1341 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1342 // If the types mismatch and we can't handle it, reject reuse of the load. 1343 if (LD->getType() != LI->getType()) { 1344 // If the stored value is larger or equal to the loaded value, we can 1345 // reuse it. 1346 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1347 UnavailableBlocks.push_back(DepBB); 1348 continue; 1349 } 1350 } 1351 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1352 continue; 1353 } 1354 1355 UnavailableBlocks.push_back(DepBB); 1356 continue; 1357 } 1358 1359 // If we have no predecessors that produce a known value for this load, exit 1360 // early. 1361 if (ValuesPerBlock.empty()) return false; 1362 1363 // If all of the instructions we depend on produce a known value for this 1364 // load, then it is fully redundant and we can use PHI insertion to compute 1365 // its value. Insert PHIs and remove the fully redundant value now. 1366 if (UnavailableBlocks.empty()) { 1367 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1368 1369 // Perform PHI construction. 1370 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1371 LI->replaceAllUsesWith(V); 1372 1373 if (isa<PHINode>(V)) 1374 V->takeName(LI); 1375 if (V->getType()->isPointerTy()) 1376 MD->invalidateCachedPointerInfo(V); 1377 markInstructionForDeletion(LI); 1378 ++NumGVNLoad; 1379 return true; 1380 } 1381 1382 if (!EnablePRE || !EnableLoadPRE) 1383 return false; 1384 1385 // Okay, we have *some* definitions of the value. This means that the value 1386 // is available in some of our (transitive) predecessors. Lets think about 1387 // doing PRE of this load. This will involve inserting a new load into the 1388 // predecessor when it's not available. We could do this in general, but 1389 // prefer to not increase code size. As such, we only do this when we know 1390 // that we only have to insert *one* load (which means we're basically moving 1391 // the load, not inserting a new one). 1392 1393 SmallPtrSet<BasicBlock *, 4> Blockers; 1394 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1395 Blockers.insert(UnavailableBlocks[i]); 1396 1397 // Lets find first basic block with more than one predecessor. Walk backwards 1398 // through predecessors if needed. 1399 BasicBlock *LoadBB = LI->getParent(); 1400 BasicBlock *TmpBB = LoadBB; 1401 1402 bool isSinglePred = false; 1403 bool allSingleSucc = true; 1404 while (TmpBB->getSinglePredecessor()) { 1405 isSinglePred = true; 1406 TmpBB = TmpBB->getSinglePredecessor(); 1407 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1408 return false; 1409 if (Blockers.count(TmpBB)) 1410 return false; 1411 1412 // If any of these blocks has more than one successor (i.e. if the edge we 1413 // just traversed was critical), then there are other paths through this 1414 // block along which the load may not be anticipated. Hoisting the load 1415 // above this block would be adding the load to execution paths along 1416 // which it was not previously executed. 1417 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1418 return false; 1419 } 1420 1421 assert(TmpBB); 1422 LoadBB = TmpBB; 1423 1424 // FIXME: It is extremely unclear what this loop is doing, other than 1425 // artificially restricting loadpre. 1426 if (isSinglePred) { 1427 bool isHot = false; 1428 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1429 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1430 if (AV.isSimpleValue()) 1431 // "Hot" Instruction is in some loop (because it dominates its dep. 1432 // instruction). 1433 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1434 if (DT->dominates(LI, I)) { 1435 isHot = true; 1436 break; 1437 } 1438 } 1439 1440 // We are interested only in "hot" instructions. We don't want to do any 1441 // mis-optimizations here. 1442 if (!isHot) 1443 return false; 1444 } 1445 1446 // Check to see how many predecessors have the loaded value fully 1447 // available. 1448 DenseMap<BasicBlock*, Value*> PredLoads; 1449 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1450 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1451 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1452 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1453 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1454 1455 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1456 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1457 PI != E; ++PI) { 1458 BasicBlock *Pred = *PI; 1459 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1460 continue; 1461 } 1462 PredLoads[Pred] = 0; 1463 1464 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1465 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1466 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1467 << Pred->getName() << "': " << *LI << '\n'); 1468 return false; 1469 } 1470 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1471 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1472 } 1473 } 1474 if (!NeedToSplit.empty()) { 1475 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1476 return false; 1477 } 1478 1479 // Decide whether PRE is profitable for this load. 1480 unsigned NumUnavailablePreds = PredLoads.size(); 1481 assert(NumUnavailablePreds != 0 && 1482 "Fully available value should be eliminated above!"); 1483 1484 // If this load is unavailable in multiple predecessors, reject it. 1485 // FIXME: If we could restructure the CFG, we could make a common pred with 1486 // all the preds that don't have an available LI and insert a new load into 1487 // that one block. 1488 if (NumUnavailablePreds != 1) 1489 return false; 1490 1491 // Check if the load can safely be moved to all the unavailable predecessors. 1492 bool CanDoPRE = true; 1493 SmallVector<Instruction*, 8> NewInsts; 1494 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1495 E = PredLoads.end(); I != E; ++I) { 1496 BasicBlock *UnavailablePred = I->first; 1497 1498 // Do PHI translation to get its value in the predecessor if necessary. The 1499 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1500 1501 // If all preds have a single successor, then we know it is safe to insert 1502 // the load on the pred (?!?), so we can insert code to materialize the 1503 // pointer if it is not available. 1504 PHITransAddr Address(LI->getPointerOperand(), TD); 1505 Value *LoadPtr = 0; 1506 if (allSingleSucc) { 1507 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1508 *DT, NewInsts); 1509 } else { 1510 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1511 LoadPtr = Address.getAddr(); 1512 } 1513 1514 // If we couldn't find or insert a computation of this phi translated value, 1515 // we fail PRE. 1516 if (LoadPtr == 0) { 1517 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1518 << *LI->getPointerOperand() << "\n"); 1519 CanDoPRE = false; 1520 break; 1521 } 1522 1523 // Make sure it is valid to move this load here. We have to watch out for: 1524 // @1 = getelementptr (i8* p, ... 1525 // test p and branch if == 0 1526 // load @1 1527 // It is valid to have the getelementptr before the test, even if p can 1528 // be 0, as getelementptr only does address arithmetic. 1529 // If we are not pushing the value through any multiple-successor blocks 1530 // we do not have this case. Otherwise, check that the load is safe to 1531 // put anywhere; this can be improved, but should be conservatively safe. 1532 if (!allSingleSucc && 1533 // FIXME: REEVALUTE THIS. 1534 !isSafeToLoadUnconditionally(LoadPtr, 1535 UnavailablePred->getTerminator(), 1536 LI->getAlignment(), TD)) { 1537 CanDoPRE = false; 1538 break; 1539 } 1540 1541 I->second = LoadPtr; 1542 } 1543 1544 if (!CanDoPRE) { 1545 while (!NewInsts.empty()) { 1546 Instruction *I = NewInsts.pop_back_val(); 1547 if (MD) MD->removeInstruction(I); 1548 I->eraseFromParent(); 1549 } 1550 return false; 1551 } 1552 1553 // Okay, we can eliminate this load by inserting a reload in the predecessor 1554 // and using PHI construction to get the value in the other predecessors, do 1555 // it. 1556 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1557 DEBUG(if (!NewInsts.empty()) 1558 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1559 << *NewInsts.back() << '\n'); 1560 1561 // Assign value numbers to the new instructions. 1562 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1563 // FIXME: We really _ought_ to insert these value numbers into their 1564 // parent's availability map. However, in doing so, we risk getting into 1565 // ordering issues. If a block hasn't been processed yet, we would be 1566 // marking a value as AVAIL-IN, which isn't what we intend. 1567 VN.lookup_or_add(NewInsts[i]); 1568 } 1569 1570 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1571 E = PredLoads.end(); I != E; ++I) { 1572 BasicBlock *UnavailablePred = I->first; 1573 Value *LoadPtr = I->second; 1574 1575 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1576 LI->getAlignment(), 1577 UnavailablePred->getTerminator()); 1578 1579 // Transfer the old load's TBAA tag to the new load. 1580 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1581 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1582 1583 // Transfer DebugLoc. 1584 NewLoad->setDebugLoc(LI->getDebugLoc()); 1585 1586 // Add the newly created load. 1587 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1588 NewLoad)); 1589 MD->invalidateCachedPointerInfo(LoadPtr); 1590 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1591 } 1592 1593 // Perform PHI construction. 1594 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1595 LI->replaceAllUsesWith(V); 1596 if (isa<PHINode>(V)) 1597 V->takeName(LI); 1598 if (V->getType()->isPointerTy()) 1599 MD->invalidateCachedPointerInfo(V); 1600 markInstructionForDeletion(LI); 1601 ++NumPRELoad; 1602 return true; 1603} 1604 1605/// processLoad - Attempt to eliminate a load, first by eliminating it 1606/// locally, and then attempting non-local elimination if that fails. 1607bool GVN::processLoad(LoadInst *L) { 1608 if (!MD) 1609 return false; 1610 1611 if (L->isVolatile()) 1612 return false; 1613 1614 if (L->use_empty()) { 1615 markInstructionForDeletion(L); 1616 return true; 1617 } 1618 1619 // ... to a pointer that has been loaded from before... 1620 MemDepResult Dep = MD->getDependency(L); 1621 1622 // If we have a clobber and target data is around, see if this is a clobber 1623 // that we can fix up through code synthesis. 1624 if (Dep.isClobber() && TD) { 1625 // Check to see if we have something like this: 1626 // store i32 123, i32* %P 1627 // %A = bitcast i32* %P to i8* 1628 // %B = gep i8* %A, i32 1 1629 // %C = load i8* %B 1630 // 1631 // We could do that by recognizing if the clobber instructions are obviously 1632 // a common base + constant offset, and if the previous store (or memset) 1633 // completely covers this load. This sort of thing can happen in bitfield 1634 // access code. 1635 Value *AvailVal = 0; 1636 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1637 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1638 L->getPointerOperand(), 1639 DepSI, *TD); 1640 if (Offset != -1) 1641 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1642 L->getType(), L, *TD); 1643 } 1644 1645 // Check to see if we have something like this: 1646 // load i32* P 1647 // load i8* (P+1) 1648 // if we have this, replace the later with an extraction from the former. 1649 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1650 // If this is a clobber and L is the first instruction in its block, then 1651 // we have the first instruction in the entry block. 1652 if (DepLI == L) 1653 return false; 1654 1655 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1656 L->getPointerOperand(), 1657 DepLI, *TD); 1658 if (Offset != -1) 1659 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1660 } 1661 1662 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1663 // a value on from it. 1664 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1665 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1666 L->getPointerOperand(), 1667 DepMI, *TD); 1668 if (Offset != -1) 1669 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1670 } 1671 1672 if (AvailVal) { 1673 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1674 << *AvailVal << '\n' << *L << "\n\n\n"); 1675 1676 // Replace the load! 1677 L->replaceAllUsesWith(AvailVal); 1678 if (AvailVal->getType()->isPointerTy()) 1679 MD->invalidateCachedPointerInfo(AvailVal); 1680 markInstructionForDeletion(L); 1681 ++NumGVNLoad; 1682 return true; 1683 } 1684 } 1685 1686 // If the value isn't available, don't do anything! 1687 if (Dep.isClobber()) { 1688 DEBUG( 1689 // fast print dep, using operator<< on instruction is too slow. 1690 dbgs() << "GVN: load "; 1691 WriteAsOperand(dbgs(), L); 1692 Instruction *I = Dep.getInst(); 1693 dbgs() << " is clobbered by " << *I << '\n'; 1694 ); 1695 return false; 1696 } 1697 1698 if (Dep.isUnknown()) { 1699 DEBUG( 1700 // fast print dep, using operator<< on instruction is too slow. 1701 dbgs() << "GVN: load "; 1702 WriteAsOperand(dbgs(), L); 1703 dbgs() << " has unknown dependence\n"; 1704 ); 1705 return false; 1706 } 1707 1708 // If it is defined in another block, try harder. 1709 if (Dep.isNonLocal()) 1710 return processNonLocalLoad(L); 1711 1712 assert(Dep.isDef() && "Expecting def here"); 1713 1714 Instruction *DepInst = Dep.getInst(); 1715 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1716 Value *StoredVal = DepSI->getValueOperand(); 1717 1718 // The store and load are to a must-aliased pointer, but they may not 1719 // actually have the same type. See if we know how to reuse the stored 1720 // value (depending on its type). 1721 if (StoredVal->getType() != L->getType()) { 1722 if (TD) { 1723 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1724 L, *TD); 1725 if (StoredVal == 0) 1726 return false; 1727 1728 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1729 << '\n' << *L << "\n\n\n"); 1730 } 1731 else 1732 return false; 1733 } 1734 1735 // Remove it! 1736 L->replaceAllUsesWith(StoredVal); 1737 if (StoredVal->getType()->isPointerTy()) 1738 MD->invalidateCachedPointerInfo(StoredVal); 1739 markInstructionForDeletion(L); 1740 ++NumGVNLoad; 1741 return true; 1742 } 1743 1744 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1745 Value *AvailableVal = DepLI; 1746 1747 // The loads are of a must-aliased pointer, but they may not actually have 1748 // the same type. See if we know how to reuse the previously loaded value 1749 // (depending on its type). 1750 if (DepLI->getType() != L->getType()) { 1751 if (TD) { 1752 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1753 L, *TD); 1754 if (AvailableVal == 0) 1755 return false; 1756 1757 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1758 << "\n" << *L << "\n\n\n"); 1759 } 1760 else 1761 return false; 1762 } 1763 1764 // Remove it! 1765 L->replaceAllUsesWith(AvailableVal); 1766 if (DepLI->getType()->isPointerTy()) 1767 MD->invalidateCachedPointerInfo(DepLI); 1768 markInstructionForDeletion(L); 1769 ++NumGVNLoad; 1770 return true; 1771 } 1772 1773 // If this load really doesn't depend on anything, then we must be loading an 1774 // undef value. This can happen when loading for a fresh allocation with no 1775 // intervening stores, for example. 1776 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1777 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1778 markInstructionForDeletion(L); 1779 ++NumGVNLoad; 1780 return true; 1781 } 1782 1783 // If this load occurs either right after a lifetime begin, 1784 // then the loaded value is undefined. 1785 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1786 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1787 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1788 markInstructionForDeletion(L); 1789 ++NumGVNLoad; 1790 return true; 1791 } 1792 } 1793 1794 return false; 1795} 1796 1797// findLeader - In order to find a leader for a given value number at a 1798// specific basic block, we first obtain the list of all Values for that number, 1799// and then scan the list to find one whose block dominates the block in 1800// question. This is fast because dominator tree queries consist of only 1801// a few comparisons of DFS numbers. 1802Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1803 LeaderTableEntry Vals = LeaderTable[num]; 1804 if (!Vals.Val) return 0; 1805 1806 Value *Val = 0; 1807 if (DT->dominates(Vals.BB, BB)) { 1808 Val = Vals.Val; 1809 if (isa<Constant>(Val)) return Val; 1810 } 1811 1812 LeaderTableEntry* Next = Vals.Next; 1813 while (Next) { 1814 if (DT->dominates(Next->BB, BB)) { 1815 if (isa<Constant>(Next->Val)) return Next->Val; 1816 if (!Val) Val = Next->Val; 1817 } 1818 1819 Next = Next->Next; 1820 } 1821 1822 return Val; 1823} 1824 1825 1826/// processInstruction - When calculating availability, handle an instruction 1827/// by inserting it into the appropriate sets 1828bool GVN::processInstruction(Instruction *I) { 1829 // Ignore dbg info intrinsics. 1830 if (isa<DbgInfoIntrinsic>(I)) 1831 return false; 1832 1833 // If the instruction can be easily simplified then do so now in preference 1834 // to value numbering it. Value numbering often exposes redundancies, for 1835 // example if it determines that %y is equal to %x then the instruction 1836 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 1837 if (Value *V = SimplifyInstruction(I, TD, DT)) { 1838 I->replaceAllUsesWith(V); 1839 if (MD && V->getType()->isPointerTy()) 1840 MD->invalidateCachedPointerInfo(V); 1841 markInstructionForDeletion(I); 1842 return true; 1843 } 1844 1845 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1846 if (processLoad(LI)) 1847 return true; 1848 1849 unsigned Num = VN.lookup_or_add(LI); 1850 addToLeaderTable(Num, LI, LI->getParent()); 1851 return false; 1852 } 1853 1854 // For conditions branches, we can perform simple conditional propagation on 1855 // the condition value itself. 1856 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1857 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1858 return false; 1859 1860 Value *BranchCond = BI->getCondition(); 1861 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1862 1863 BasicBlock *TrueSucc = BI->getSuccessor(0); 1864 BasicBlock *FalseSucc = BI->getSuccessor(1); 1865 1866 if (TrueSucc->getSinglePredecessor()) 1867 addToLeaderTable(CondVN, 1868 ConstantInt::getTrue(TrueSucc->getContext()), 1869 TrueSucc); 1870 if (FalseSucc->getSinglePredecessor()) 1871 addToLeaderTable(CondVN, 1872 ConstantInt::getFalse(TrueSucc->getContext()), 1873 FalseSucc); 1874 1875 return false; 1876 } 1877 1878 // Instructions with void type don't return a value, so there's 1879 // no point in trying to find redudancies in them. 1880 if (I->getType()->isVoidTy()) return false; 1881 1882 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1883 unsigned Num = VN.lookup_or_add(I); 1884 1885 // Allocations are always uniquely numbered, so we can save time and memory 1886 // by fast failing them. 1887 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 1888 addToLeaderTable(Num, I, I->getParent()); 1889 return false; 1890 } 1891 1892 // If the number we were assigned was a brand new VN, then we don't 1893 // need to do a lookup to see if the number already exists 1894 // somewhere in the domtree: it can't! 1895 if (Num == NextNum) { 1896 addToLeaderTable(Num, I, I->getParent()); 1897 return false; 1898 } 1899 1900 // Perform fast-path value-number based elimination of values inherited from 1901 // dominators. 1902 Value *repl = findLeader(I->getParent(), Num); 1903 if (repl == 0) { 1904 // Failure, just remember this instance for future use. 1905 addToLeaderTable(Num, I, I->getParent()); 1906 return false; 1907 } 1908 1909 // Remove it! 1910 I->replaceAllUsesWith(repl); 1911 if (MD && repl->getType()->isPointerTy()) 1912 MD->invalidateCachedPointerInfo(repl); 1913 markInstructionForDeletion(I); 1914 return true; 1915} 1916 1917/// runOnFunction - This is the main transformation entry point for a function. 1918bool GVN::runOnFunction(Function& F) { 1919 if (!NoLoads) 1920 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1921 DT = &getAnalysis<DominatorTree>(); 1922 TD = getAnalysisIfAvailable<TargetData>(); 1923 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1924 VN.setMemDep(MD); 1925 VN.setDomTree(DT); 1926 1927 bool Changed = false; 1928 bool ShouldContinue = true; 1929 1930 // Merge unconditional branches, allowing PRE to catch more 1931 // optimization opportunities. 1932 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1933 BasicBlock *BB = FI++; 1934 1935 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1936 if (removedBlock) ++NumGVNBlocks; 1937 1938 Changed |= removedBlock; 1939 } 1940 1941 unsigned Iteration = 0; 1942 while (ShouldContinue) { 1943 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 1944 ShouldContinue = iterateOnFunction(F); 1945 if (splitCriticalEdges()) 1946 ShouldContinue = true; 1947 Changed |= ShouldContinue; 1948 ++Iteration; 1949 } 1950 1951 if (EnablePRE) { 1952 bool PREChanged = true; 1953 while (PREChanged) { 1954 PREChanged = performPRE(F); 1955 Changed |= PREChanged; 1956 } 1957 } 1958 // FIXME: Should perform GVN again after PRE does something. PRE can move 1959 // computations into blocks where they become fully redundant. Note that 1960 // we can't do this until PRE's critical edge splitting updates memdep. 1961 // Actually, when this happens, we should just fully integrate PRE into GVN. 1962 1963 cleanupGlobalSets(); 1964 1965 return Changed; 1966} 1967 1968 1969bool GVN::processBlock(BasicBlock *BB) { 1970 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 1971 // (and incrementing BI before processing an instruction). 1972 assert(InstrsToErase.empty() && 1973 "We expect InstrsToErase to be empty across iterations"); 1974 bool ChangedFunction = false; 1975 1976 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1977 BI != BE;) { 1978 ChangedFunction |= processInstruction(BI); 1979 if (InstrsToErase.empty()) { 1980 ++BI; 1981 continue; 1982 } 1983 1984 // If we need some instructions deleted, do it now. 1985 NumGVNInstr += InstrsToErase.size(); 1986 1987 // Avoid iterator invalidation. 1988 bool AtStart = BI == BB->begin(); 1989 if (!AtStart) 1990 --BI; 1991 1992 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 1993 E = InstrsToErase.end(); I != E; ++I) { 1994 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 1995 if (MD) MD->removeInstruction(*I); 1996 (*I)->eraseFromParent(); 1997 DEBUG(verifyRemoved(*I)); 1998 } 1999 InstrsToErase.clear(); 2000 2001 if (AtStart) 2002 BI = BB->begin(); 2003 else 2004 ++BI; 2005 } 2006 2007 return ChangedFunction; 2008} 2009 2010/// performPRE - Perform a purely local form of PRE that looks for diamond 2011/// control flow patterns and attempts to perform simple PRE at the join point. 2012bool GVN::performPRE(Function &F) { 2013 bool Changed = false; 2014 DenseMap<BasicBlock*, Value*> predMap; 2015 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2016 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2017 BasicBlock *CurrentBlock = *DI; 2018 2019 // Nothing to PRE in the entry block. 2020 if (CurrentBlock == &F.getEntryBlock()) continue; 2021 2022 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2023 BE = CurrentBlock->end(); BI != BE; ) { 2024 Instruction *CurInst = BI++; 2025 2026 if (isa<AllocaInst>(CurInst) || 2027 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2028 CurInst->getType()->isVoidTy() || 2029 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2030 isa<DbgInfoIntrinsic>(CurInst)) 2031 continue; 2032 2033 // We don't currently value number ANY inline asm calls. 2034 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2035 if (CallI->isInlineAsm()) 2036 continue; 2037 2038 uint32_t ValNo = VN.lookup(CurInst); 2039 2040 // Look for the predecessors for PRE opportunities. We're 2041 // only trying to solve the basic diamond case, where 2042 // a value is computed in the successor and one predecessor, 2043 // but not the other. We also explicitly disallow cases 2044 // where the successor is its own predecessor, because they're 2045 // more complicated to get right. 2046 unsigned NumWith = 0; 2047 unsigned NumWithout = 0; 2048 BasicBlock *PREPred = 0; 2049 predMap.clear(); 2050 2051 for (pred_iterator PI = pred_begin(CurrentBlock), 2052 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2053 BasicBlock *P = *PI; 2054 // We're not interested in PRE where the block is its 2055 // own predecessor, or in blocks with predecessors 2056 // that are not reachable. 2057 if (P == CurrentBlock) { 2058 NumWithout = 2; 2059 break; 2060 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2061 NumWithout = 2; 2062 break; 2063 } 2064 2065 Value* predV = findLeader(P, ValNo); 2066 if (predV == 0) { 2067 PREPred = P; 2068 ++NumWithout; 2069 } else if (predV == CurInst) { 2070 NumWithout = 2; 2071 } else { 2072 predMap[P] = predV; 2073 ++NumWith; 2074 } 2075 } 2076 2077 // Don't do PRE when it might increase code size, i.e. when 2078 // we would need to insert instructions in more than one pred. 2079 if (NumWithout != 1 || NumWith == 0) 2080 continue; 2081 2082 // Don't do PRE across indirect branch. 2083 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2084 continue; 2085 2086 // We can't do PRE safely on a critical edge, so instead we schedule 2087 // the edge to be split and perform the PRE the next time we iterate 2088 // on the function. 2089 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2090 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2091 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2092 continue; 2093 } 2094 2095 // Instantiate the expression in the predecessor that lacked it. 2096 // Because we are going top-down through the block, all value numbers 2097 // will be available in the predecessor by the time we need them. Any 2098 // that weren't originally present will have been instantiated earlier 2099 // in this loop. 2100 Instruction *PREInstr = CurInst->clone(); 2101 bool success = true; 2102 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2103 Value *Op = PREInstr->getOperand(i); 2104 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2105 continue; 2106 2107 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2108 PREInstr->setOperand(i, V); 2109 } else { 2110 success = false; 2111 break; 2112 } 2113 } 2114 2115 // Fail out if we encounter an operand that is not available in 2116 // the PRE predecessor. This is typically because of loads which 2117 // are not value numbered precisely. 2118 if (!success) { 2119 delete PREInstr; 2120 DEBUG(verifyRemoved(PREInstr)); 2121 continue; 2122 } 2123 2124 PREInstr->insertBefore(PREPred->getTerminator()); 2125 PREInstr->setName(CurInst->getName() + ".pre"); 2126 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2127 predMap[PREPred] = PREInstr; 2128 VN.add(PREInstr, ValNo); 2129 ++NumGVNPRE; 2130 2131 // Update the availability map to include the new instruction. 2132 addToLeaderTable(ValNo, PREInstr, PREPred); 2133 2134 // Create a PHI to make the value available in this block. 2135 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2136 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2137 CurInst->getName() + ".pre-phi", 2138 CurrentBlock->begin()); 2139 for (pred_iterator PI = PB; PI != PE; ++PI) { 2140 BasicBlock *P = *PI; 2141 Phi->addIncoming(predMap[P], P); 2142 } 2143 2144 VN.add(Phi, ValNo); 2145 addToLeaderTable(ValNo, Phi, CurrentBlock); 2146 Phi->setDebugLoc(CurInst->getDebugLoc()); 2147 CurInst->replaceAllUsesWith(Phi); 2148 if (Phi->getType()->isPointerTy()) { 2149 // Because we have added a PHI-use of the pointer value, it has now 2150 // "escaped" from alias analysis' perspective. We need to inform 2151 // AA of this. 2152 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) 2153 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii)); 2154 2155 if (MD) 2156 MD->invalidateCachedPointerInfo(Phi); 2157 } 2158 VN.erase(CurInst); 2159 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2160 2161 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2162 if (MD) MD->removeInstruction(CurInst); 2163 CurInst->eraseFromParent(); 2164 DEBUG(verifyRemoved(CurInst)); 2165 Changed = true; 2166 } 2167 } 2168 2169 if (splitCriticalEdges()) 2170 Changed = true; 2171 2172 return Changed; 2173} 2174 2175/// splitCriticalEdges - Split critical edges found during the previous 2176/// iteration that may enable further optimization. 2177bool GVN::splitCriticalEdges() { 2178 if (toSplit.empty()) 2179 return false; 2180 do { 2181 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2182 SplitCriticalEdge(Edge.first, Edge.second, this); 2183 } while (!toSplit.empty()); 2184 if (MD) MD->invalidateCachedPredecessors(); 2185 return true; 2186} 2187 2188/// iterateOnFunction - Executes one iteration of GVN 2189bool GVN::iterateOnFunction(Function &F) { 2190 cleanupGlobalSets(); 2191 2192 // Top-down walk of the dominator tree 2193 bool Changed = false; 2194#if 0 2195 // Needed for value numbering with phi construction to work. 2196 ReversePostOrderTraversal<Function*> RPOT(&F); 2197 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2198 RE = RPOT.end(); RI != RE; ++RI) 2199 Changed |= processBlock(*RI); 2200#else 2201 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2202 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2203 Changed |= processBlock(DI->getBlock()); 2204#endif 2205 2206 return Changed; 2207} 2208 2209void GVN::cleanupGlobalSets() { 2210 VN.clear(); 2211 LeaderTable.clear(); 2212 TableAllocator.Reset(); 2213} 2214 2215/// verifyRemoved - Verify that the specified instruction does not occur in our 2216/// internal data structures. 2217void GVN::verifyRemoved(const Instruction *Inst) const { 2218 VN.verifyRemoved(Inst); 2219 2220 // Walk through the value number scope to make sure the instruction isn't 2221 // ferreted away in it. 2222 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2223 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2224 const LeaderTableEntry *Node = &I->second; 2225 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2226 2227 while (Node->Next) { 2228 Node = Node->Next; 2229 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2230 } 2231 } 2232} 2233