GVN.cpp revision 795cf5efba4255cadd0bfde0e9d3dec65e96dd50
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 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 Expression create_extractvalue_expression(ExtractValueInst* EI); 95 uint32_t lookup_or_add_call(CallInst* C); 96 public: 97 ValueTable() : nextValueNumber(1) { } 98 uint32_t lookup_or_add(Value *V); 99 uint32_t lookup(Value *V) const; 100 void add(Value *V, uint32_t num); 101 void clear(); 102 void erase(Value *v); 103 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 104 AliasAnalysis *getAliasAnalysis() const { return AA; } 105 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 106 void setDomTree(DominatorTree* D) { DT = D; } 107 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 108 void verifyRemoved(const Value *) const; 109 }; 110} 111 112namespace llvm { 113template <> struct DenseMapInfo<Expression> { 114 static inline Expression getEmptyKey() { 115 return ~0U; 116 } 117 118 static inline Expression getTombstoneKey() { 119 return ~1U; 120 } 121 122 static unsigned getHashValue(const Expression e) { 123 unsigned hash = e.opcode; 124 125 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 126 (unsigned)((uintptr_t)e.type >> 9)); 127 128 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 129 E = e.varargs.end(); I != E; ++I) 130 hash = *I + hash * 37; 131 132 return hash; 133 } 134 static bool isEqual(const Expression &LHS, const Expression &RHS) { 135 return LHS == RHS; 136 } 137}; 138 139} 140 141//===----------------------------------------------------------------------===// 142// ValueTable Internal Functions 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 (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 156 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 157 II != IE; ++II) 158 e.varargs.push_back(*II); 159 } 160 161 return e; 162} 163 164Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 165 assert(EI != 0 && "Not an ExtractValueInst?"); 166 Expression e; 167 e.type = EI->getType(); 168 e.opcode = 0; 169 170 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 171 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 172 // EI might be an extract from one of our recognised intrinsics. If it 173 // is we'll synthesize a semantically equivalent expression instead on 174 // an extract value expression. 175 switch (I->getIntrinsicID()) { 176 case Intrinsic::sadd_with_overflow: 177 case Intrinsic::uadd_with_overflow: 178 e.opcode = Instruction::Add; 179 break; 180 case Intrinsic::ssub_with_overflow: 181 case Intrinsic::usub_with_overflow: 182 e.opcode = Instruction::Sub; 183 break; 184 case Intrinsic::smul_with_overflow: 185 case Intrinsic::umul_with_overflow: 186 e.opcode = Instruction::Mul; 187 break; 188 default: 189 break; 190 } 191 192 if (e.opcode != 0) { 193 // Intrinsic recognized. Grab its args to finish building the expression. 194 assert(I->getNumArgOperands() == 2 && 195 "Expect two args for recognised intrinsics."); 196 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 197 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 198 return e; 199 } 200 } 201 202 // Not a recognised intrinsic. Fall back to producing an extract value 203 // expression. 204 e.opcode = EI->getOpcode(); 205 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 206 OI != OE; ++OI) 207 e.varargs.push_back(lookup_or_add(*OI)); 208 209 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 210 II != IE; ++II) 211 e.varargs.push_back(*II); 212 213 return e; 214} 215 216//===----------------------------------------------------------------------===// 217// ValueTable External Functions 218//===----------------------------------------------------------------------===// 219 220/// add - Insert a value into the table with a specified value number. 221void ValueTable::add(Value *V, uint32_t num) { 222 valueNumbering.insert(std::make_pair(V, num)); 223} 224 225uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 226 if (AA->doesNotAccessMemory(C)) { 227 Expression exp = create_expression(C); 228 uint32_t& e = expressionNumbering[exp]; 229 if (!e) e = nextValueNumber++; 230 valueNumbering[C] = e; 231 return e; 232 } else if (AA->onlyReadsMemory(C)) { 233 Expression exp = create_expression(C); 234 uint32_t& e = expressionNumbering[exp]; 235 if (!e) { 236 e = nextValueNumber++; 237 valueNumbering[C] = e; 238 return e; 239 } 240 if (!MD) { 241 e = nextValueNumber++; 242 valueNumbering[C] = e; 243 return e; 244 } 245 246 MemDepResult local_dep = MD->getDependency(C); 247 248 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 249 valueNumbering[C] = nextValueNumber; 250 return nextValueNumber++; 251 } 252 253 if (local_dep.isDef()) { 254 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 255 256 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 257 valueNumbering[C] = nextValueNumber; 258 return nextValueNumber++; 259 } 260 261 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 262 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 263 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 264 if (c_vn != cd_vn) { 265 valueNumbering[C] = nextValueNumber; 266 return nextValueNumber++; 267 } 268 } 269 270 uint32_t v = lookup_or_add(local_cdep); 271 valueNumbering[C] = v; 272 return v; 273 } 274 275 // Non-local case. 276 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 277 MD->getNonLocalCallDependency(CallSite(C)); 278 // FIXME: Move the checking logic to MemDep! 279 CallInst* cdep = 0; 280 281 // Check to see if we have a single dominating call instruction that is 282 // identical to C. 283 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 284 const NonLocalDepEntry *I = &deps[i]; 285 if (I->getResult().isNonLocal()) 286 continue; 287 288 // We don't handle non-definitions. If we already have a call, reject 289 // instruction dependencies. 290 if (!I->getResult().isDef() || cdep != 0) { 291 cdep = 0; 292 break; 293 } 294 295 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 296 // FIXME: All duplicated with non-local case. 297 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 298 cdep = NonLocalDepCall; 299 continue; 300 } 301 302 cdep = 0; 303 break; 304 } 305 306 if (!cdep) { 307 valueNumbering[C] = nextValueNumber; 308 return nextValueNumber++; 309 } 310 311 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 312 valueNumbering[C] = nextValueNumber; 313 return nextValueNumber++; 314 } 315 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 316 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 317 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 318 if (c_vn != cd_vn) { 319 valueNumbering[C] = nextValueNumber; 320 return nextValueNumber++; 321 } 322 } 323 324 uint32_t v = lookup_or_add(cdep); 325 valueNumbering[C] = v; 326 return v; 327 328 } else { 329 valueNumbering[C] = nextValueNumber; 330 return nextValueNumber++; 331 } 332} 333 334/// lookup_or_add - Returns the value number for the specified value, assigning 335/// it a new number if it did not have one before. 336uint32_t ValueTable::lookup_or_add(Value *V) { 337 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 338 if (VI != valueNumbering.end()) 339 return VI->second; 340 341 if (!isa<Instruction>(V)) { 342 valueNumbering[V] = nextValueNumber; 343 return nextValueNumber++; 344 } 345 346 Instruction* I = cast<Instruction>(V); 347 Expression exp; 348 switch (I->getOpcode()) { 349 case Instruction::Call: 350 return lookup_or_add_call(cast<CallInst>(I)); 351 case Instruction::Add: 352 case Instruction::FAdd: 353 case Instruction::Sub: 354 case Instruction::FSub: 355 case Instruction::Mul: 356 case Instruction::FMul: 357 case Instruction::UDiv: 358 case Instruction::SDiv: 359 case Instruction::FDiv: 360 case Instruction::URem: 361 case Instruction::SRem: 362 case Instruction::FRem: 363 case Instruction::Shl: 364 case Instruction::LShr: 365 case Instruction::AShr: 366 case Instruction::And: 367 case Instruction::Or : 368 case Instruction::Xor: 369 case Instruction::ICmp: 370 case Instruction::FCmp: 371 case Instruction::Trunc: 372 case Instruction::ZExt: 373 case Instruction::SExt: 374 case Instruction::FPToUI: 375 case Instruction::FPToSI: 376 case Instruction::UIToFP: 377 case Instruction::SIToFP: 378 case Instruction::FPTrunc: 379 case Instruction::FPExt: 380 case Instruction::PtrToInt: 381 case Instruction::IntToPtr: 382 case Instruction::BitCast: 383 case Instruction::Select: 384 case Instruction::ExtractElement: 385 case Instruction::InsertElement: 386 case Instruction::ShuffleVector: 387 case Instruction::InsertValue: 388 case Instruction::GetElementPtr: 389 exp = create_expression(I); 390 break; 391 case Instruction::ExtractValue: 392 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 393 break; 394 default: 395 valueNumbering[V] = nextValueNumber; 396 return nextValueNumber++; 397 } 398 399 uint32_t& e = expressionNumbering[exp]; 400 if (!e) e = nextValueNumber++; 401 valueNumbering[V] = e; 402 return e; 403} 404 405/// lookup - Returns the value number of the specified value. Fails if 406/// the value has not yet been numbered. 407uint32_t ValueTable::lookup(Value *V) const { 408 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 409 assert(VI != valueNumbering.end() && "Value not numbered?"); 410 return VI->second; 411} 412 413/// clear - Remove all entries from the ValueTable. 414void ValueTable::clear() { 415 valueNumbering.clear(); 416 expressionNumbering.clear(); 417 nextValueNumber = 1; 418} 419 420/// erase - Remove a value from the value numbering. 421void ValueTable::erase(Value *V) { 422 valueNumbering.erase(V); 423} 424 425/// verifyRemoved - Verify that the value is removed from all internal data 426/// structures. 427void ValueTable::verifyRemoved(const Value *V) const { 428 for (DenseMap<Value*, uint32_t>::const_iterator 429 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 430 assert(I->first != V && "Inst still occurs in value numbering map!"); 431 } 432} 433 434//===----------------------------------------------------------------------===// 435// GVN Pass 436//===----------------------------------------------------------------------===// 437 438namespace { 439 440 class GVN : public FunctionPass { 441 bool NoLoads; 442 MemoryDependenceAnalysis *MD; 443 DominatorTree *DT; 444 const TargetData *TD; 445 446 ValueTable VN; 447 448 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 449 /// have that value number. Use findLeader to query it. 450 struct LeaderTableEntry { 451 Value *Val; 452 BasicBlock *BB; 453 LeaderTableEntry *Next; 454 }; 455 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 456 BumpPtrAllocator TableAllocator; 457 458 SmallVector<Instruction*, 8> InstrsToErase; 459 public: 460 static char ID; // Pass identification, replacement for typeid 461 explicit GVN(bool noloads = false) 462 : FunctionPass(ID), NoLoads(noloads), MD(0) { 463 initializeGVNPass(*PassRegistry::getPassRegistry()); 464 } 465 466 bool runOnFunction(Function &F); 467 468 /// markInstructionForDeletion - This removes the specified instruction from 469 /// our various maps and marks it for deletion. 470 void markInstructionForDeletion(Instruction *I) { 471 VN.erase(I); 472 InstrsToErase.push_back(I); 473 } 474 475 const TargetData *getTargetData() const { return TD; } 476 DominatorTree &getDominatorTree() const { return *DT; } 477 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 478 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 479 private: 480 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 481 /// its value number. 482 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 483 LeaderTableEntry &Curr = LeaderTable[N]; 484 if (!Curr.Val) { 485 Curr.Val = V; 486 Curr.BB = BB; 487 return; 488 } 489 490 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 491 Node->Val = V; 492 Node->BB = BB; 493 Node->Next = Curr.Next; 494 Curr.Next = Node; 495 } 496 497 /// removeFromLeaderTable - Scan the list of values corresponding to a given 498 /// value number, and remove the given value if encountered. 499 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 500 LeaderTableEntry* Prev = 0; 501 LeaderTableEntry* Curr = &LeaderTable[N]; 502 503 while (Curr->Val != V || Curr->BB != BB) { 504 Prev = Curr; 505 Curr = Curr->Next; 506 } 507 508 if (Prev) { 509 Prev->Next = Curr->Next; 510 } else { 511 if (!Curr->Next) { 512 Curr->Val = 0; 513 Curr->BB = 0; 514 } else { 515 LeaderTableEntry* Next = Curr->Next; 516 Curr->Val = Next->Val; 517 Curr->BB = Next->BB; 518 Curr->Next = Next->Next; 519 } 520 } 521 } 522 523 // List of critical edges to be split between iterations. 524 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 525 526 // This transformation requires dominator postdominator info 527 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 528 AU.addRequired<DominatorTree>(); 529 if (!NoLoads) 530 AU.addRequired<MemoryDependenceAnalysis>(); 531 AU.addRequired<AliasAnalysis>(); 532 533 AU.addPreserved<DominatorTree>(); 534 AU.addPreserved<AliasAnalysis>(); 535 } 536 537 538 // Helper fuctions 539 // FIXME: eliminate or document these better 540 bool processLoad(LoadInst *L); 541 bool processInstruction(Instruction *I); 542 bool processNonLocalLoad(LoadInst *L); 543 bool processBlock(BasicBlock *BB); 544 void dump(DenseMap<uint32_t, Value*> &d); 545 bool iterateOnFunction(Function &F); 546 bool performPRE(Function &F); 547 Value *findLeader(BasicBlock *BB, uint32_t num); 548 void cleanupGlobalSets(); 549 void verifyRemoved(const Instruction *I) const; 550 bool splitCriticalEdges(); 551 }; 552 553 char GVN::ID = 0; 554} 555 556// createGVNPass - The public interface to this file... 557FunctionPass *llvm::createGVNPass(bool NoLoads) { 558 return new GVN(NoLoads); 559} 560 561INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 562INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 563INITIALIZE_PASS_DEPENDENCY(DominatorTree) 564INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 565INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 566 567void GVN::dump(DenseMap<uint32_t, Value*>& d) { 568 errs() << "{\n"; 569 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 570 E = d.end(); I != E; ++I) { 571 errs() << I->first << "\n"; 572 I->second->dump(); 573 } 574 errs() << "}\n"; 575} 576 577/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 578/// we're analyzing is fully available in the specified block. As we go, keep 579/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 580/// map is actually a tri-state map with the following values: 581/// 0) we know the block *is not* fully available. 582/// 1) we know the block *is* fully available. 583/// 2) we do not know whether the block is fully available or not, but we are 584/// currently speculating that it will be. 585/// 3) we are speculating for this block and have used that to speculate for 586/// other blocks. 587static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 588 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 589 // Optimistically assume that the block is fully available and check to see 590 // if we already know about this block in one lookup. 591 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 592 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 593 594 // If the entry already existed for this block, return the precomputed value. 595 if (!IV.second) { 596 // If this is a speculative "available" value, mark it as being used for 597 // speculation of other blocks. 598 if (IV.first->second == 2) 599 IV.first->second = 3; 600 return IV.first->second != 0; 601 } 602 603 // Otherwise, see if it is fully available in all predecessors. 604 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 605 606 // If this block has no predecessors, it isn't live-in here. 607 if (PI == PE) 608 goto SpeculationFailure; 609 610 for (; PI != PE; ++PI) 611 // If the value isn't fully available in one of our predecessors, then it 612 // isn't fully available in this block either. Undo our previous 613 // optimistic assumption and bail out. 614 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 615 goto SpeculationFailure; 616 617 return true; 618 619// SpeculationFailure - If we get here, we found out that this is not, after 620// all, a fully-available block. We have a problem if we speculated on this and 621// used the speculation to mark other blocks as available. 622SpeculationFailure: 623 char &BBVal = FullyAvailableBlocks[BB]; 624 625 // If we didn't speculate on this, just return with it set to false. 626 if (BBVal == 2) { 627 BBVal = 0; 628 return false; 629 } 630 631 // If we did speculate on this value, we could have blocks set to 1 that are 632 // incorrect. Walk the (transitive) successors of this block and mark them as 633 // 0 if set to one. 634 SmallVector<BasicBlock*, 32> BBWorklist; 635 BBWorklist.push_back(BB); 636 637 do { 638 BasicBlock *Entry = BBWorklist.pop_back_val(); 639 // Note that this sets blocks to 0 (unavailable) if they happen to not 640 // already be in FullyAvailableBlocks. This is safe. 641 char &EntryVal = FullyAvailableBlocks[Entry]; 642 if (EntryVal == 0) continue; // Already unavailable. 643 644 // Mark as unavailable. 645 EntryVal = 0; 646 647 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 648 BBWorklist.push_back(*I); 649 } while (!BBWorklist.empty()); 650 651 return false; 652} 653 654 655/// CanCoerceMustAliasedValueToLoad - Return true if 656/// CoerceAvailableValueToLoadType will succeed. 657static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 658 Type *LoadTy, 659 const TargetData &TD) { 660 // If the loaded or stored value is an first class array or struct, don't try 661 // to transform them. We need to be able to bitcast to integer. 662 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 663 StoredVal->getType()->isStructTy() || 664 StoredVal->getType()->isArrayTy()) 665 return false; 666 667 // The store has to be at least as big as the load. 668 if (TD.getTypeSizeInBits(StoredVal->getType()) < 669 TD.getTypeSizeInBits(LoadTy)) 670 return false; 671 672 return true; 673} 674 675 676/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 677/// then a load from a must-aliased pointer of a different type, try to coerce 678/// the stored value. LoadedTy is the type of the load we want to replace and 679/// InsertPt is the place to insert new instructions. 680/// 681/// If we can't do it, return null. 682static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 683 Type *LoadedTy, 684 Instruction *InsertPt, 685 const TargetData &TD) { 686 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 687 return 0; 688 689 // If this is already the right type, just return it. 690 Type *StoredValTy = StoredVal->getType(); 691 692 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 693 uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy); 694 695 // If the store and reload are the same size, we can always reuse it. 696 if (StoreSize == LoadSize) { 697 // Pointer to Pointer -> use bitcast. 698 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 699 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 700 701 // Convert source pointers to integers, which can be bitcast. 702 if (StoredValTy->isPointerTy()) { 703 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 704 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 705 } 706 707 Type *TypeToCastTo = LoadedTy; 708 if (TypeToCastTo->isPointerTy()) 709 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 710 711 if (StoredValTy != TypeToCastTo) 712 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 713 714 // Cast to pointer if the load needs a pointer type. 715 if (LoadedTy->isPointerTy()) 716 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 717 718 return StoredVal; 719 } 720 721 // If the loaded value is smaller than the available value, then we can 722 // extract out a piece from it. If the available value is too small, then we 723 // can't do anything. 724 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 725 726 // Convert source pointers to integers, which can be manipulated. 727 if (StoredValTy->isPointerTy()) { 728 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 729 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 730 } 731 732 // Convert vectors and fp to integer, which can be manipulated. 733 if (!StoredValTy->isIntegerTy()) { 734 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 735 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 736 } 737 738 // If this is a big-endian system, we need to shift the value down to the low 739 // bits so that a truncate will work. 740 if (TD.isBigEndian()) { 741 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 742 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 743 } 744 745 // Truncate the integer to the right size now. 746 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 747 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 748 749 if (LoadedTy == NewIntTy) 750 return StoredVal; 751 752 // If the result is a pointer, inttoptr. 753 if (LoadedTy->isPointerTy()) 754 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 755 756 // Otherwise, bitcast. 757 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 758} 759 760/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 761/// memdep query of a load that ends up being a clobbering memory write (store, 762/// memset, memcpy, memmove). This means that the write *may* provide bits used 763/// by the load but we can't be sure because the pointers don't mustalias. 764/// 765/// Check this case to see if there is anything more we can do before we give 766/// up. This returns -1 if we have to give up, or a byte number in the stored 767/// value of the piece that feeds the load. 768static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 769 Value *WritePtr, 770 uint64_t WriteSizeInBits, 771 const TargetData &TD) { 772 // If the loaded or stored value is an first class array or struct, don't try 773 // to transform them. We need to be able to bitcast to integer. 774 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 775 return -1; 776 777 int64_t StoreOffset = 0, LoadOffset = 0; 778 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 779 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 780 if (StoreBase != LoadBase) 781 return -1; 782 783 // If the load and store are to the exact same address, they should have been 784 // a must alias. AA must have gotten confused. 785 // FIXME: Study to see if/when this happens. One case is forwarding a memset 786 // to a load from the base of the memset. 787#if 0 788 if (LoadOffset == StoreOffset) { 789 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 790 << "Base = " << *StoreBase << "\n" 791 << "Store Ptr = " << *WritePtr << "\n" 792 << "Store Offs = " << StoreOffset << "\n" 793 << "Load Ptr = " << *LoadPtr << "\n"; 794 abort(); 795 } 796#endif 797 798 // If the load and store don't overlap at all, the store doesn't provide 799 // anything to the load. In this case, they really don't alias at all, AA 800 // must have gotten confused. 801 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 802 803 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 804 return -1; 805 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 806 LoadSize >>= 3; 807 808 809 bool isAAFailure = false; 810 if (StoreOffset < LoadOffset) 811 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 812 else 813 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 814 815 if (isAAFailure) { 816#if 0 817 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 818 << "Base = " << *StoreBase << "\n" 819 << "Store Ptr = " << *WritePtr << "\n" 820 << "Store Offs = " << StoreOffset << "\n" 821 << "Load Ptr = " << *LoadPtr << "\n"; 822 abort(); 823#endif 824 return -1; 825 } 826 827 // If the Load isn't completely contained within the stored bits, we don't 828 // have all the bits to feed it. We could do something crazy in the future 829 // (issue a smaller load then merge the bits in) but this seems unlikely to be 830 // valuable. 831 if (StoreOffset > LoadOffset || 832 StoreOffset+StoreSize < LoadOffset+LoadSize) 833 return -1; 834 835 // Okay, we can do this transformation. Return the number of bytes into the 836 // store that the load is. 837 return LoadOffset-StoreOffset; 838} 839 840/// AnalyzeLoadFromClobberingStore - This function is called when we have a 841/// memdep query of a load that ends up being a clobbering store. 842static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 843 StoreInst *DepSI, 844 const TargetData &TD) { 845 // Cannot handle reading from store of first-class aggregate yet. 846 if (DepSI->getValueOperand()->getType()->isStructTy() || 847 DepSI->getValueOperand()->getType()->isArrayTy()) 848 return -1; 849 850 Value *StorePtr = DepSI->getPointerOperand(); 851 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 852 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 853 StorePtr, StoreSize, TD); 854} 855 856/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 857/// memdep query of a load that ends up being clobbered by another load. See if 858/// the other load can feed into the second load. 859static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 860 LoadInst *DepLI, const TargetData &TD){ 861 // Cannot handle reading from store of first-class aggregate yet. 862 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 863 return -1; 864 865 Value *DepPtr = DepLI->getPointerOperand(); 866 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 867 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 868 if (R != -1) return R; 869 870 // If we have a load/load clobber an DepLI can be widened to cover this load, 871 // then we should widen it! 872 int64_t LoadOffs = 0; 873 const Value *LoadBase = 874 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 875 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 876 877 unsigned Size = MemoryDependenceAnalysis:: 878 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 879 if (Size == 0) return -1; 880 881 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 882} 883 884 885 886static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 887 MemIntrinsic *MI, 888 const TargetData &TD) { 889 // If the mem operation is a non-constant size, we can't handle it. 890 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 891 if (SizeCst == 0) return -1; 892 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 893 894 // If this is memset, we just need to see if the offset is valid in the size 895 // of the memset.. 896 if (MI->getIntrinsicID() == Intrinsic::memset) 897 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 898 MemSizeInBits, TD); 899 900 // If we have a memcpy/memmove, the only case we can handle is if this is a 901 // copy from constant memory. In that case, we can read directly from the 902 // constant memory. 903 MemTransferInst *MTI = cast<MemTransferInst>(MI); 904 905 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 906 if (Src == 0) return -1; 907 908 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 909 if (GV == 0 || !GV->isConstant()) return -1; 910 911 // See if the access is within the bounds of the transfer. 912 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 913 MI->getDest(), MemSizeInBits, TD); 914 if (Offset == -1) 915 return Offset; 916 917 // Otherwise, see if we can constant fold a load from the constant with the 918 // offset applied as appropriate. 919 Src = ConstantExpr::getBitCast(Src, 920 llvm::Type::getInt8PtrTy(Src->getContext())); 921 Constant *OffsetCst = 922 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 923 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 924 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 925 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 926 return Offset; 927 return -1; 928} 929 930 931/// GetStoreValueForLoad - This function is called when we have a 932/// memdep query of a load that ends up being a clobbering store. This means 933/// that the store provides bits used by the load but we the pointers don't 934/// mustalias. Check this case to see if there is anything more we can do 935/// before we give up. 936static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 937 Type *LoadTy, 938 Instruction *InsertPt, const TargetData &TD){ 939 LLVMContext &Ctx = SrcVal->getType()->getContext(); 940 941 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 942 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 943 944 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 945 946 // Compute which bits of the stored value are being used by the load. Convert 947 // to an integer type to start with. 948 if (SrcVal->getType()->isPointerTy()) 949 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 950 if (!SrcVal->getType()->isIntegerTy()) 951 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 952 "tmp"); 953 954 // Shift the bits to the least significant depending on endianness. 955 unsigned ShiftAmt; 956 if (TD.isLittleEndian()) 957 ShiftAmt = Offset*8; 958 else 959 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 960 961 if (ShiftAmt) 962 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 963 964 if (LoadSize != StoreSize) 965 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 966 "tmp"); 967 968 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 969} 970 971/// GetStoreValueForLoad - This function is called when we have a 972/// memdep query of a load that ends up being a clobbering load. This means 973/// that the load *may* provide bits used by the load but we can't be sure 974/// because the pointers don't mustalias. Check this case to see if there is 975/// anything more we can do before we give up. 976static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 977 Type *LoadTy, Instruction *InsertPt, 978 GVN &gvn) { 979 const TargetData &TD = *gvn.getTargetData(); 980 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 981 // widen SrcVal out to a larger load. 982 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 983 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 984 if (Offset+LoadSize > SrcValSize) { 985 assert(!SrcVal->isVolatile() && "Cannot widen volatile load!"); 986 assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load"); 987 // If we have a load/load clobber an DepLI can be widened to cover this 988 // load, then we should widen it to the next power of 2 size big enough! 989 unsigned NewLoadSize = Offset+LoadSize; 990 if (!isPowerOf2_32(NewLoadSize)) 991 NewLoadSize = NextPowerOf2(NewLoadSize); 992 993 Value *PtrVal = SrcVal->getPointerOperand(); 994 995 // Insert the new load after the old load. This ensures that subsequent 996 // memdep queries will find the new load. We can't easily remove the old 997 // load completely because it is already in the value numbering table. 998 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 999 Type *DestPTy = 1000 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1001 DestPTy = PointerType::get(DestPTy, 1002 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1003 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1004 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1005 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1006 NewLoad->takeName(SrcVal); 1007 NewLoad->setAlignment(SrcVal->getAlignment()); 1008 1009 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1010 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1011 1012 // Replace uses of the original load with the wider load. On a big endian 1013 // system, we need to shift down to get the relevant bits. 1014 Value *RV = NewLoad; 1015 if (TD.isBigEndian()) 1016 RV = Builder.CreateLShr(RV, 1017 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1018 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1019 SrcVal->replaceAllUsesWith(RV); 1020 1021 // We would like to use gvn.markInstructionForDeletion here, but we can't 1022 // because the load is already memoized into the leader map table that GVN 1023 // tracks. It is potentially possible to remove the load from the table, 1024 // but then there all of the operations based on it would need to be 1025 // rehashed. Just leave the dead load around. 1026 gvn.getMemDep().removeInstruction(SrcVal); 1027 SrcVal = NewLoad; 1028 } 1029 1030 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1031} 1032 1033 1034/// GetMemInstValueForLoad - This function is called when we have a 1035/// memdep query of a load that ends up being a clobbering mem intrinsic. 1036static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1037 Type *LoadTy, Instruction *InsertPt, 1038 const TargetData &TD){ 1039 LLVMContext &Ctx = LoadTy->getContext(); 1040 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1041 1042 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1043 1044 // We know that this method is only called when the mem transfer fully 1045 // provides the bits for the load. 1046 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1047 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1048 // independently of what the offset is. 1049 Value *Val = MSI->getValue(); 1050 if (LoadSize != 1) 1051 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1052 1053 Value *OneElt = Val; 1054 1055 // Splat the value out to the right number of bits. 1056 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1057 // If we can double the number of bytes set, do it. 1058 if (NumBytesSet*2 <= LoadSize) { 1059 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1060 Val = Builder.CreateOr(Val, ShVal); 1061 NumBytesSet <<= 1; 1062 continue; 1063 } 1064 1065 // Otherwise insert one byte at a time. 1066 Value *ShVal = Builder.CreateShl(Val, 1*8); 1067 Val = Builder.CreateOr(OneElt, ShVal); 1068 ++NumBytesSet; 1069 } 1070 1071 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1072 } 1073 1074 // Otherwise, this is a memcpy/memmove from a constant global. 1075 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1076 Constant *Src = cast<Constant>(MTI->getSource()); 1077 1078 // Otherwise, see if we can constant fold a load from the constant with the 1079 // offset applied as appropriate. 1080 Src = ConstantExpr::getBitCast(Src, 1081 llvm::Type::getInt8PtrTy(Src->getContext())); 1082 Constant *OffsetCst = 1083 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1084 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1085 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1086 return ConstantFoldLoadFromConstPtr(Src, &TD); 1087} 1088 1089namespace { 1090 1091struct AvailableValueInBlock { 1092 /// BB - The basic block in question. 1093 BasicBlock *BB; 1094 enum ValType { 1095 SimpleVal, // A simple offsetted value that is accessed. 1096 LoadVal, // A value produced by a load. 1097 MemIntrin // A memory intrinsic which is loaded from. 1098 }; 1099 1100 /// V - The value that is live out of the block. 1101 PointerIntPair<Value *, 2, ValType> Val; 1102 1103 /// Offset - The byte offset in Val that is interesting for the load query. 1104 unsigned Offset; 1105 1106 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1107 unsigned Offset = 0) { 1108 AvailableValueInBlock Res; 1109 Res.BB = BB; 1110 Res.Val.setPointer(V); 1111 Res.Val.setInt(SimpleVal); 1112 Res.Offset = Offset; 1113 return Res; 1114 } 1115 1116 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1117 unsigned Offset = 0) { 1118 AvailableValueInBlock Res; 1119 Res.BB = BB; 1120 Res.Val.setPointer(MI); 1121 Res.Val.setInt(MemIntrin); 1122 Res.Offset = Offset; 1123 return Res; 1124 } 1125 1126 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1127 unsigned Offset = 0) { 1128 AvailableValueInBlock Res; 1129 Res.BB = BB; 1130 Res.Val.setPointer(LI); 1131 Res.Val.setInt(LoadVal); 1132 Res.Offset = Offset; 1133 return Res; 1134 } 1135 1136 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1137 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1138 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1139 1140 Value *getSimpleValue() const { 1141 assert(isSimpleValue() && "Wrong accessor"); 1142 return Val.getPointer(); 1143 } 1144 1145 LoadInst *getCoercedLoadValue() const { 1146 assert(isCoercedLoadValue() && "Wrong accessor"); 1147 return cast<LoadInst>(Val.getPointer()); 1148 } 1149 1150 MemIntrinsic *getMemIntrinValue() const { 1151 assert(isMemIntrinValue() && "Wrong accessor"); 1152 return cast<MemIntrinsic>(Val.getPointer()); 1153 } 1154 1155 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1156 /// defined here to the specified type. This handles various coercion cases. 1157 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1158 Value *Res; 1159 if (isSimpleValue()) { 1160 Res = getSimpleValue(); 1161 if (Res->getType() != LoadTy) { 1162 const TargetData *TD = gvn.getTargetData(); 1163 assert(TD && "Need target data to handle type mismatch case"); 1164 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1165 *TD); 1166 1167 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1168 << *getSimpleValue() << '\n' 1169 << *Res << '\n' << "\n\n\n"); 1170 } 1171 } else if (isCoercedLoadValue()) { 1172 LoadInst *Load = getCoercedLoadValue(); 1173 if (Load->getType() == LoadTy && Offset == 0) { 1174 Res = Load; 1175 } else { 1176 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1177 gvn); 1178 1179 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1180 << *getCoercedLoadValue() << '\n' 1181 << *Res << '\n' << "\n\n\n"); 1182 } 1183 } else { 1184 const TargetData *TD = gvn.getTargetData(); 1185 assert(TD && "Need target data to handle type mismatch case"); 1186 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1187 LoadTy, BB->getTerminator(), *TD); 1188 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1189 << " " << *getMemIntrinValue() << '\n' 1190 << *Res << '\n' << "\n\n\n"); 1191 } 1192 return Res; 1193 } 1194}; 1195 1196} // end anonymous namespace 1197 1198/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1199/// construct SSA form, allowing us to eliminate LI. This returns the value 1200/// that should be used at LI's definition site. 1201static Value *ConstructSSAForLoadSet(LoadInst *LI, 1202 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1203 GVN &gvn) { 1204 // Check for the fully redundant, dominating load case. In this case, we can 1205 // just use the dominating value directly. 1206 if (ValuesPerBlock.size() == 1 && 1207 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1208 LI->getParent())) 1209 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1210 1211 // Otherwise, we have to construct SSA form. 1212 SmallVector<PHINode*, 8> NewPHIs; 1213 SSAUpdater SSAUpdate(&NewPHIs); 1214 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1215 1216 Type *LoadTy = LI->getType(); 1217 1218 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1219 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1220 BasicBlock *BB = AV.BB; 1221 1222 if (SSAUpdate.HasValueForBlock(BB)) 1223 continue; 1224 1225 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1226 } 1227 1228 // Perform PHI construction. 1229 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1230 1231 // If new PHI nodes were created, notify alias analysis. 1232 if (V->getType()->isPointerTy()) { 1233 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1234 1235 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1236 AA->copyValue(LI, NewPHIs[i]); 1237 1238 // Now that we've copied information to the new PHIs, scan through 1239 // them again and inform alias analysis that we've added potentially 1240 // escaping uses to any values that are operands to these PHIs. 1241 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1242 PHINode *P = NewPHIs[i]; 1243 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1244 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1245 AA->addEscapingUse(P->getOperandUse(jj)); 1246 } 1247 } 1248 } 1249 1250 return V; 1251} 1252 1253static bool isLifetimeStart(const Instruction *Inst) { 1254 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1255 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1256 return false; 1257} 1258 1259/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1260/// non-local by performing PHI construction. 1261bool GVN::processNonLocalLoad(LoadInst *LI) { 1262 // Find the non-local dependencies of the load. 1263 SmallVector<NonLocalDepResult, 64> Deps; 1264 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1265 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1266 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1267 // << Deps.size() << *LI << '\n'); 1268 1269 // If we had to process more than one hundred blocks to find the 1270 // dependencies, this load isn't worth worrying about. Optimizing 1271 // it will be too expensive. 1272 if (Deps.size() > 100) 1273 return false; 1274 1275 // If we had a phi translation failure, we'll have a single entry which is a 1276 // clobber in the current block. Reject this early. 1277 if (Deps.size() == 1 && Deps[0].getResult().isUnknown()) { 1278 DEBUG( 1279 dbgs() << "GVN: non-local load "; 1280 WriteAsOperand(dbgs(), LI); 1281 dbgs() << " has unknown dependencies\n"; 1282 ); 1283 return false; 1284 } 1285 1286 // Filter out useless results (non-locals, etc). Keep track of the blocks 1287 // where we have a value available in repl, also keep track of whether we see 1288 // dependencies that produce an unknown value for the load (such as a call 1289 // that could potentially clobber the load). 1290 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1291 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1292 1293 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1294 BasicBlock *DepBB = Deps[i].getBB(); 1295 MemDepResult DepInfo = Deps[i].getResult(); 1296 1297 if (DepInfo.isUnknown()) { 1298 UnavailableBlocks.push_back(DepBB); 1299 continue; 1300 } 1301 1302 if (DepInfo.isClobber()) { 1303 // The address being loaded in this non-local block may not be the same as 1304 // the pointer operand of the load if PHI translation occurs. Make sure 1305 // to consider the right address. 1306 Value *Address = Deps[i].getAddress(); 1307 1308 // If the dependence is to a store that writes to a superset of the bits 1309 // read by the load, we can extract the bits we need for the load from the 1310 // stored value. 1311 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1312 if (TD && Address) { 1313 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1314 DepSI, *TD); 1315 if (Offset != -1) { 1316 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1317 DepSI->getValueOperand(), 1318 Offset)); 1319 continue; 1320 } 1321 } 1322 } 1323 1324 // Check to see if we have something like this: 1325 // load i32* P 1326 // load i8* (P+1) 1327 // if we have this, replace the later with an extraction from the former. 1328 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1329 // If this is a clobber and L is the first instruction in its block, then 1330 // we have the first instruction in the entry block. 1331 if (DepLI != LI && Address && TD) { 1332 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1333 LI->getPointerOperand(), 1334 DepLI, *TD); 1335 1336 if (Offset != -1) { 1337 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1338 Offset)); 1339 continue; 1340 } 1341 } 1342 } 1343 1344 // If the clobbering value is a memset/memcpy/memmove, see if we can 1345 // forward a value on from it. 1346 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1347 if (TD && Address) { 1348 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1349 DepMI, *TD); 1350 if (Offset != -1) { 1351 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1352 Offset)); 1353 continue; 1354 } 1355 } 1356 } 1357 1358 UnavailableBlocks.push_back(DepBB); 1359 continue; 1360 } 1361 1362 assert(DepInfo.isDef() && "Expecting def here"); 1363 1364 Instruction *DepInst = DepInfo.getInst(); 1365 1366 // Loading the allocation -> undef. 1367 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1368 // Loading immediately after lifetime begin -> undef. 1369 isLifetimeStart(DepInst)) { 1370 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1371 UndefValue::get(LI->getType()))); 1372 continue; 1373 } 1374 1375 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1376 // Reject loads and stores that are to the same address but are of 1377 // different types if we have to. 1378 if (S->getValueOperand()->getType() != LI->getType()) { 1379 // If the stored value is larger or equal to the loaded value, we can 1380 // reuse it. 1381 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1382 LI->getType(), *TD)) { 1383 UnavailableBlocks.push_back(DepBB); 1384 continue; 1385 } 1386 } 1387 1388 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1389 S->getValueOperand())); 1390 continue; 1391 } 1392 1393 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1394 // If the types mismatch and we can't handle it, reject reuse of the load. 1395 if (LD->getType() != LI->getType()) { 1396 // If the stored value is larger or equal to the loaded value, we can 1397 // reuse it. 1398 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1399 UnavailableBlocks.push_back(DepBB); 1400 continue; 1401 } 1402 } 1403 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1404 continue; 1405 } 1406 1407 UnavailableBlocks.push_back(DepBB); 1408 continue; 1409 } 1410 1411 // If we have no predecessors that produce a known value for this load, exit 1412 // early. 1413 if (ValuesPerBlock.empty()) return false; 1414 1415 // If all of the instructions we depend on produce a known value for this 1416 // load, then it is fully redundant and we can use PHI insertion to compute 1417 // its value. Insert PHIs and remove the fully redundant value now. 1418 if (UnavailableBlocks.empty()) { 1419 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1420 1421 // Perform PHI construction. 1422 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1423 LI->replaceAllUsesWith(V); 1424 1425 if (isa<PHINode>(V)) 1426 V->takeName(LI); 1427 if (V->getType()->isPointerTy()) 1428 MD->invalidateCachedPointerInfo(V); 1429 markInstructionForDeletion(LI); 1430 ++NumGVNLoad; 1431 return true; 1432 } 1433 1434 if (!EnablePRE || !EnableLoadPRE) 1435 return false; 1436 1437 // Okay, we have *some* definitions of the value. This means that the value 1438 // is available in some of our (transitive) predecessors. Lets think about 1439 // doing PRE of this load. This will involve inserting a new load into the 1440 // predecessor when it's not available. We could do this in general, but 1441 // prefer to not increase code size. As such, we only do this when we know 1442 // that we only have to insert *one* load (which means we're basically moving 1443 // the load, not inserting a new one). 1444 1445 SmallPtrSet<BasicBlock *, 4> Blockers; 1446 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1447 Blockers.insert(UnavailableBlocks[i]); 1448 1449 // Let's find the first basic block with more than one predecessor. Walk 1450 // backwards through predecessors if needed. 1451 BasicBlock *LoadBB = LI->getParent(); 1452 BasicBlock *TmpBB = LoadBB; 1453 1454 bool isSinglePred = false; 1455 bool allSingleSucc = true; 1456 while (TmpBB->getSinglePredecessor()) { 1457 isSinglePred = true; 1458 TmpBB = TmpBB->getSinglePredecessor(); 1459 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1460 return false; 1461 if (Blockers.count(TmpBB)) 1462 return false; 1463 1464 // If any of these blocks has more than one successor (i.e. if the edge we 1465 // just traversed was critical), then there are other paths through this 1466 // block along which the load may not be anticipated. Hoisting the load 1467 // above this block would be adding the load to execution paths along 1468 // which it was not previously executed. 1469 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1470 return false; 1471 } 1472 1473 assert(TmpBB); 1474 LoadBB = TmpBB; 1475 1476 // FIXME: It is extremely unclear what this loop is doing, other than 1477 // artificially restricting loadpre. 1478 if (isSinglePred) { 1479 bool isHot = false; 1480 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1481 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1482 if (AV.isSimpleValue()) 1483 // "Hot" Instruction is in some loop (because it dominates its dep. 1484 // instruction). 1485 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1486 if (DT->dominates(LI, I)) { 1487 isHot = true; 1488 break; 1489 } 1490 } 1491 1492 // We are interested only in "hot" instructions. We don't want to do any 1493 // mis-optimizations here. 1494 if (!isHot) 1495 return false; 1496 } 1497 1498 // Check to see how many predecessors have the loaded value fully 1499 // available. 1500 DenseMap<BasicBlock*, Value*> PredLoads; 1501 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1502 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1503 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1504 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1505 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1506 1507 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1508 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1509 PI != E; ++PI) { 1510 BasicBlock *Pred = *PI; 1511 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1512 continue; 1513 } 1514 PredLoads[Pred] = 0; 1515 1516 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1517 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1518 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1519 << Pred->getName() << "': " << *LI << '\n'); 1520 return false; 1521 } 1522 1523 if (LoadBB->isLandingPad()) { 1524 DEBUG(dbgs() 1525 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1526 << Pred->getName() << "': " << *LI << '\n'); 1527 return false; 1528 } 1529 1530 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1531 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1532 } 1533 } 1534 1535 if (!NeedToSplit.empty()) { 1536 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1537 return false; 1538 } 1539 1540 // Decide whether PRE is profitable for this load. 1541 unsigned NumUnavailablePreds = PredLoads.size(); 1542 assert(NumUnavailablePreds != 0 && 1543 "Fully available value should be eliminated above!"); 1544 1545 // If this load is unavailable in multiple predecessors, reject it. 1546 // FIXME: If we could restructure the CFG, we could make a common pred with 1547 // all the preds that don't have an available LI and insert a new load into 1548 // that one block. 1549 if (NumUnavailablePreds != 1) 1550 return false; 1551 1552 // Check if the load can safely be moved to all the unavailable predecessors. 1553 bool CanDoPRE = true; 1554 SmallVector<Instruction*, 8> NewInsts; 1555 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1556 E = PredLoads.end(); I != E; ++I) { 1557 BasicBlock *UnavailablePred = I->first; 1558 1559 // Do PHI translation to get its value in the predecessor if necessary. The 1560 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1561 1562 // If all preds have a single successor, then we know it is safe to insert 1563 // the load on the pred (?!?), so we can insert code to materialize the 1564 // pointer if it is not available. 1565 PHITransAddr Address(LI->getPointerOperand(), TD); 1566 Value *LoadPtr = 0; 1567 if (allSingleSucc) { 1568 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1569 *DT, NewInsts); 1570 } else { 1571 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1572 LoadPtr = Address.getAddr(); 1573 } 1574 1575 // If we couldn't find or insert a computation of this phi translated value, 1576 // we fail PRE. 1577 if (LoadPtr == 0) { 1578 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1579 << *LI->getPointerOperand() << "\n"); 1580 CanDoPRE = false; 1581 break; 1582 } 1583 1584 // Make sure it is valid to move this load here. We have to watch out for: 1585 // @1 = getelementptr (i8* p, ... 1586 // test p and branch if == 0 1587 // load @1 1588 // It is valid to have the getelementptr before the test, even if p can 1589 // be 0, as getelementptr only does address arithmetic. 1590 // If we are not pushing the value through any multiple-successor blocks 1591 // we do not have this case. Otherwise, check that the load is safe to 1592 // put anywhere; this can be improved, but should be conservatively safe. 1593 if (!allSingleSucc && 1594 // FIXME: REEVALUTE THIS. 1595 !isSafeToLoadUnconditionally(LoadPtr, 1596 UnavailablePred->getTerminator(), 1597 LI->getAlignment(), TD)) { 1598 CanDoPRE = false; 1599 break; 1600 } 1601 1602 I->second = LoadPtr; 1603 } 1604 1605 if (!CanDoPRE) { 1606 while (!NewInsts.empty()) { 1607 Instruction *I = NewInsts.pop_back_val(); 1608 if (MD) MD->removeInstruction(I); 1609 I->eraseFromParent(); 1610 } 1611 return false; 1612 } 1613 1614 // Okay, we can eliminate this load by inserting a reload in the predecessor 1615 // and using PHI construction to get the value in the other predecessors, do 1616 // it. 1617 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1618 DEBUG(if (!NewInsts.empty()) 1619 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1620 << *NewInsts.back() << '\n'); 1621 1622 // Assign value numbers to the new instructions. 1623 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1624 // FIXME: We really _ought_ to insert these value numbers into their 1625 // parent's availability map. However, in doing so, we risk getting into 1626 // ordering issues. If a block hasn't been processed yet, we would be 1627 // marking a value as AVAIL-IN, which isn't what we intend. 1628 VN.lookup_or_add(NewInsts[i]); 1629 } 1630 1631 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1632 E = PredLoads.end(); I != E; ++I) { 1633 BasicBlock *UnavailablePred = I->first; 1634 Value *LoadPtr = I->second; 1635 1636 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1637 LI->getAlignment(), 1638 UnavailablePred->getTerminator()); 1639 1640 // Transfer the old load's TBAA tag to the new load. 1641 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1642 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1643 1644 // Transfer DebugLoc. 1645 NewLoad->setDebugLoc(LI->getDebugLoc()); 1646 1647 // Add the newly created load. 1648 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1649 NewLoad)); 1650 MD->invalidateCachedPointerInfo(LoadPtr); 1651 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1652 } 1653 1654 // Perform PHI construction. 1655 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1656 LI->replaceAllUsesWith(V); 1657 if (isa<PHINode>(V)) 1658 V->takeName(LI); 1659 if (V->getType()->isPointerTy()) 1660 MD->invalidateCachedPointerInfo(V); 1661 markInstructionForDeletion(LI); 1662 ++NumPRELoad; 1663 return true; 1664} 1665 1666/// processLoad - Attempt to eliminate a load, first by eliminating it 1667/// locally, and then attempting non-local elimination if that fails. 1668bool GVN::processLoad(LoadInst *L) { 1669 if (!MD) 1670 return false; 1671 1672 if (L->isVolatile()) 1673 return false; 1674 1675 if (L->use_empty()) { 1676 markInstructionForDeletion(L); 1677 return true; 1678 } 1679 1680 // ... to a pointer that has been loaded from before... 1681 MemDepResult Dep = MD->getDependency(L); 1682 1683 // If we have a clobber and target data is around, see if this is a clobber 1684 // that we can fix up through code synthesis. 1685 if (Dep.isClobber() && TD) { 1686 // Check to see if we have something like this: 1687 // store i32 123, i32* %P 1688 // %A = bitcast i32* %P to i8* 1689 // %B = gep i8* %A, i32 1 1690 // %C = load i8* %B 1691 // 1692 // We could do that by recognizing if the clobber instructions are obviously 1693 // a common base + constant offset, and if the previous store (or memset) 1694 // completely covers this load. This sort of thing can happen in bitfield 1695 // access code. 1696 Value *AvailVal = 0; 1697 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1698 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1699 L->getPointerOperand(), 1700 DepSI, *TD); 1701 if (Offset != -1) 1702 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1703 L->getType(), L, *TD); 1704 } 1705 1706 // Check to see if we have something like this: 1707 // load i32* P 1708 // load i8* (P+1) 1709 // if we have this, replace the later with an extraction from the former. 1710 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1711 // If this is a clobber and L is the first instruction in its block, then 1712 // we have the first instruction in the entry block. 1713 if (DepLI == L) 1714 return false; 1715 1716 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1717 L->getPointerOperand(), 1718 DepLI, *TD); 1719 if (Offset != -1) 1720 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1721 } 1722 1723 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1724 // a value on from it. 1725 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1726 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1727 L->getPointerOperand(), 1728 DepMI, *TD); 1729 if (Offset != -1) 1730 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1731 } 1732 1733 if (AvailVal) { 1734 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1735 << *AvailVal << '\n' << *L << "\n\n\n"); 1736 1737 // Replace the load! 1738 L->replaceAllUsesWith(AvailVal); 1739 if (AvailVal->getType()->isPointerTy()) 1740 MD->invalidateCachedPointerInfo(AvailVal); 1741 markInstructionForDeletion(L); 1742 ++NumGVNLoad; 1743 return true; 1744 } 1745 } 1746 1747 // If the value isn't available, don't do anything! 1748 if (Dep.isClobber()) { 1749 DEBUG( 1750 // fast print dep, using operator<< on instruction is too slow. 1751 dbgs() << "GVN: load "; 1752 WriteAsOperand(dbgs(), L); 1753 Instruction *I = Dep.getInst(); 1754 dbgs() << " is clobbered by " << *I << '\n'; 1755 ); 1756 return false; 1757 } 1758 1759 if (Dep.isUnknown()) { 1760 DEBUG( 1761 // fast print dep, using operator<< on instruction is too slow. 1762 dbgs() << "GVN: load "; 1763 WriteAsOperand(dbgs(), L); 1764 dbgs() << " has unknown dependence\n"; 1765 ); 1766 return false; 1767 } 1768 1769 // If it is defined in another block, try harder. 1770 if (Dep.isNonLocal()) 1771 return processNonLocalLoad(L); 1772 1773 assert(Dep.isDef() && "Expecting def here"); 1774 1775 Instruction *DepInst = Dep.getInst(); 1776 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1777 Value *StoredVal = DepSI->getValueOperand(); 1778 1779 // The store and load are to a must-aliased pointer, but they may not 1780 // actually have the same type. See if we know how to reuse the stored 1781 // value (depending on its type). 1782 if (StoredVal->getType() != L->getType()) { 1783 if (TD) { 1784 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1785 L, *TD); 1786 if (StoredVal == 0) 1787 return false; 1788 1789 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1790 << '\n' << *L << "\n\n\n"); 1791 } 1792 else 1793 return false; 1794 } 1795 1796 // Remove it! 1797 L->replaceAllUsesWith(StoredVal); 1798 if (StoredVal->getType()->isPointerTy()) 1799 MD->invalidateCachedPointerInfo(StoredVal); 1800 markInstructionForDeletion(L); 1801 ++NumGVNLoad; 1802 return true; 1803 } 1804 1805 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1806 Value *AvailableVal = DepLI; 1807 1808 // The loads are of a must-aliased pointer, but they may not actually have 1809 // the same type. See if we know how to reuse the previously loaded value 1810 // (depending on its type). 1811 if (DepLI->getType() != L->getType()) { 1812 if (TD) { 1813 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1814 L, *TD); 1815 if (AvailableVal == 0) 1816 return false; 1817 1818 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1819 << "\n" << *L << "\n\n\n"); 1820 } 1821 else 1822 return false; 1823 } 1824 1825 // Remove it! 1826 L->replaceAllUsesWith(AvailableVal); 1827 if (DepLI->getType()->isPointerTy()) 1828 MD->invalidateCachedPointerInfo(DepLI); 1829 markInstructionForDeletion(L); 1830 ++NumGVNLoad; 1831 return true; 1832 } 1833 1834 // If this load really doesn't depend on anything, then we must be loading an 1835 // undef value. This can happen when loading for a fresh allocation with no 1836 // intervening stores, for example. 1837 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1838 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1839 markInstructionForDeletion(L); 1840 ++NumGVNLoad; 1841 return true; 1842 } 1843 1844 // If this load occurs either right after a lifetime begin, 1845 // then the loaded value is undefined. 1846 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1847 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1848 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1849 markInstructionForDeletion(L); 1850 ++NumGVNLoad; 1851 return true; 1852 } 1853 } 1854 1855 return false; 1856} 1857 1858// findLeader - In order to find a leader for a given value number at a 1859// specific basic block, we first obtain the list of all Values for that number, 1860// and then scan the list to find one whose block dominates the block in 1861// question. This is fast because dominator tree queries consist of only 1862// a few comparisons of DFS numbers. 1863Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1864 LeaderTableEntry Vals = LeaderTable[num]; 1865 if (!Vals.Val) return 0; 1866 1867 Value *Val = 0; 1868 if (DT->dominates(Vals.BB, BB)) { 1869 Val = Vals.Val; 1870 if (isa<Constant>(Val)) return Val; 1871 } 1872 1873 LeaderTableEntry* Next = Vals.Next; 1874 while (Next) { 1875 if (DT->dominates(Next->BB, BB)) { 1876 if (isa<Constant>(Next->Val)) return Next->Val; 1877 if (!Val) Val = Next->Val; 1878 } 1879 1880 Next = Next->Next; 1881 } 1882 1883 return Val; 1884} 1885 1886 1887/// processInstruction - When calculating availability, handle an instruction 1888/// by inserting it into the appropriate sets 1889bool GVN::processInstruction(Instruction *I) { 1890 // Ignore dbg info intrinsics. 1891 if (isa<DbgInfoIntrinsic>(I)) 1892 return false; 1893 1894 // If the instruction can be easily simplified then do so now in preference 1895 // to value numbering it. Value numbering often exposes redundancies, for 1896 // example if it determines that %y is equal to %x then the instruction 1897 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 1898 if (Value *V = SimplifyInstruction(I, TD, DT)) { 1899 I->replaceAllUsesWith(V); 1900 if (MD && V->getType()->isPointerTy()) 1901 MD->invalidateCachedPointerInfo(V); 1902 markInstructionForDeletion(I); 1903 return true; 1904 } 1905 1906 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1907 if (processLoad(LI)) 1908 return true; 1909 1910 unsigned Num = VN.lookup_or_add(LI); 1911 addToLeaderTable(Num, LI, LI->getParent()); 1912 return false; 1913 } 1914 1915 // For conditions branches, we can perform simple conditional propagation on 1916 // the condition value itself. 1917 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1918 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1919 return false; 1920 1921 Value *BranchCond = BI->getCondition(); 1922 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1923 1924 BasicBlock *TrueSucc = BI->getSuccessor(0); 1925 BasicBlock *FalseSucc = BI->getSuccessor(1); 1926 1927 if (TrueSucc->getSinglePredecessor()) 1928 addToLeaderTable(CondVN, 1929 ConstantInt::getTrue(TrueSucc->getContext()), 1930 TrueSucc); 1931 if (FalseSucc->getSinglePredecessor()) 1932 addToLeaderTable(CondVN, 1933 ConstantInt::getFalse(TrueSucc->getContext()), 1934 FalseSucc); 1935 1936 return false; 1937 } 1938 1939 // Instructions with void type don't return a value, so there's 1940 // no point in trying to find redudancies in them. 1941 if (I->getType()->isVoidTy()) return false; 1942 1943 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1944 unsigned Num = VN.lookup_or_add(I); 1945 1946 // Allocations are always uniquely numbered, so we can save time and memory 1947 // by fast failing them. 1948 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 1949 addToLeaderTable(Num, I, I->getParent()); 1950 return false; 1951 } 1952 1953 // If the number we were assigned was a brand new VN, then we don't 1954 // need to do a lookup to see if the number already exists 1955 // somewhere in the domtree: it can't! 1956 if (Num == NextNum) { 1957 addToLeaderTable(Num, I, I->getParent()); 1958 return false; 1959 } 1960 1961 // Perform fast-path value-number based elimination of values inherited from 1962 // dominators. 1963 Value *repl = findLeader(I->getParent(), Num); 1964 if (repl == 0) { 1965 // Failure, just remember this instance for future use. 1966 addToLeaderTable(Num, I, I->getParent()); 1967 return false; 1968 } 1969 1970 // Remove it! 1971 I->replaceAllUsesWith(repl); 1972 if (MD && repl->getType()->isPointerTy()) 1973 MD->invalidateCachedPointerInfo(repl); 1974 markInstructionForDeletion(I); 1975 return true; 1976} 1977 1978/// runOnFunction - This is the main transformation entry point for a function. 1979bool GVN::runOnFunction(Function& F) { 1980 if (!NoLoads) 1981 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1982 DT = &getAnalysis<DominatorTree>(); 1983 TD = getAnalysisIfAvailable<TargetData>(); 1984 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1985 VN.setMemDep(MD); 1986 VN.setDomTree(DT); 1987 1988 bool Changed = false; 1989 bool ShouldContinue = true; 1990 1991 // Merge unconditional branches, allowing PRE to catch more 1992 // optimization opportunities. 1993 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1994 BasicBlock *BB = FI++; 1995 1996 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1997 if (removedBlock) ++NumGVNBlocks; 1998 1999 Changed |= removedBlock; 2000 } 2001 2002 unsigned Iteration = 0; 2003 while (ShouldContinue) { 2004 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2005 ShouldContinue = iterateOnFunction(F); 2006 if (splitCriticalEdges()) 2007 ShouldContinue = true; 2008 Changed |= ShouldContinue; 2009 ++Iteration; 2010 } 2011 2012 if (EnablePRE) { 2013 bool PREChanged = true; 2014 while (PREChanged) { 2015 PREChanged = performPRE(F); 2016 Changed |= PREChanged; 2017 } 2018 } 2019 // FIXME: Should perform GVN again after PRE does something. PRE can move 2020 // computations into blocks where they become fully redundant. Note that 2021 // we can't do this until PRE's critical edge splitting updates memdep. 2022 // Actually, when this happens, we should just fully integrate PRE into GVN. 2023 2024 cleanupGlobalSets(); 2025 2026 return Changed; 2027} 2028 2029 2030bool GVN::processBlock(BasicBlock *BB) { 2031 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2032 // (and incrementing BI before processing an instruction). 2033 assert(InstrsToErase.empty() && 2034 "We expect InstrsToErase to be empty across iterations"); 2035 bool ChangedFunction = false; 2036 2037 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2038 BI != BE;) { 2039 ChangedFunction |= processInstruction(BI); 2040 if (InstrsToErase.empty()) { 2041 ++BI; 2042 continue; 2043 } 2044 2045 // If we need some instructions deleted, do it now. 2046 NumGVNInstr += InstrsToErase.size(); 2047 2048 // Avoid iterator invalidation. 2049 bool AtStart = BI == BB->begin(); 2050 if (!AtStart) 2051 --BI; 2052 2053 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2054 E = InstrsToErase.end(); I != E; ++I) { 2055 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2056 if (MD) MD->removeInstruction(*I); 2057 (*I)->eraseFromParent(); 2058 DEBUG(verifyRemoved(*I)); 2059 } 2060 InstrsToErase.clear(); 2061 2062 if (AtStart) 2063 BI = BB->begin(); 2064 else 2065 ++BI; 2066 } 2067 2068 return ChangedFunction; 2069} 2070 2071/// performPRE - Perform a purely local form of PRE that looks for diamond 2072/// control flow patterns and attempts to perform simple PRE at the join point. 2073bool GVN::performPRE(Function &F) { 2074 bool Changed = false; 2075 DenseMap<BasicBlock*, Value*> predMap; 2076 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2077 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2078 BasicBlock *CurrentBlock = *DI; 2079 2080 // Nothing to PRE in the entry block. 2081 if (CurrentBlock == &F.getEntryBlock()) continue; 2082 2083 // Don't perform PRE on a landing pad. 2084 if (CurrentBlock->isLandingPad()) continue; 2085 2086 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2087 BE = CurrentBlock->end(); BI != BE; ) { 2088 Instruction *CurInst = BI++; 2089 2090 if (isa<AllocaInst>(CurInst) || 2091 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2092 CurInst->getType()->isVoidTy() || 2093 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2094 isa<DbgInfoIntrinsic>(CurInst)) 2095 continue; 2096 2097 // We don't currently value number ANY inline asm calls. 2098 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2099 if (CallI->isInlineAsm()) 2100 continue; 2101 2102 uint32_t ValNo = VN.lookup(CurInst); 2103 2104 // Look for the predecessors for PRE opportunities. We're 2105 // only trying to solve the basic diamond case, where 2106 // a value is computed in the successor and one predecessor, 2107 // but not the other. We also explicitly disallow cases 2108 // where the successor is its own predecessor, because they're 2109 // more complicated to get right. 2110 unsigned NumWith = 0; 2111 unsigned NumWithout = 0; 2112 BasicBlock *PREPred = 0; 2113 predMap.clear(); 2114 2115 for (pred_iterator PI = pred_begin(CurrentBlock), 2116 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2117 BasicBlock *P = *PI; 2118 // We're not interested in PRE where the block is its 2119 // own predecessor, or in blocks with predecessors 2120 // that are not reachable. 2121 if (P == CurrentBlock) { 2122 NumWithout = 2; 2123 break; 2124 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2125 NumWithout = 2; 2126 break; 2127 } 2128 2129 Value* predV = findLeader(P, ValNo); 2130 if (predV == 0) { 2131 PREPred = P; 2132 ++NumWithout; 2133 } else if (predV == CurInst) { 2134 NumWithout = 2; 2135 } else { 2136 predMap[P] = predV; 2137 ++NumWith; 2138 } 2139 } 2140 2141 // Don't do PRE when it might increase code size, i.e. when 2142 // we would need to insert instructions in more than one pred. 2143 if (NumWithout != 1 || NumWith == 0) 2144 continue; 2145 2146 // Don't do PRE across indirect branch. 2147 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2148 continue; 2149 2150 // We can't do PRE safely on a critical edge, so instead we schedule 2151 // the edge to be split and perform the PRE the next time we iterate 2152 // on the function. 2153 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2154 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2155 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2156 continue; 2157 } 2158 2159 // Instantiate the expression in the predecessor that lacked it. 2160 // Because we are going top-down through the block, all value numbers 2161 // will be available in the predecessor by the time we need them. Any 2162 // that weren't originally present will have been instantiated earlier 2163 // in this loop. 2164 Instruction *PREInstr = CurInst->clone(); 2165 bool success = true; 2166 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2167 Value *Op = PREInstr->getOperand(i); 2168 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2169 continue; 2170 2171 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2172 PREInstr->setOperand(i, V); 2173 } else { 2174 success = false; 2175 break; 2176 } 2177 } 2178 2179 // Fail out if we encounter an operand that is not available in 2180 // the PRE predecessor. This is typically because of loads which 2181 // are not value numbered precisely. 2182 if (!success) { 2183 delete PREInstr; 2184 DEBUG(verifyRemoved(PREInstr)); 2185 continue; 2186 } 2187 2188 PREInstr->insertBefore(PREPred->getTerminator()); 2189 PREInstr->setName(CurInst->getName() + ".pre"); 2190 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2191 predMap[PREPred] = PREInstr; 2192 VN.add(PREInstr, ValNo); 2193 ++NumGVNPRE; 2194 2195 // Update the availability map to include the new instruction. 2196 addToLeaderTable(ValNo, PREInstr, PREPred); 2197 2198 // Create a PHI to make the value available in this block. 2199 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2200 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2201 CurInst->getName() + ".pre-phi", 2202 CurrentBlock->begin()); 2203 for (pred_iterator PI = PB; PI != PE; ++PI) { 2204 BasicBlock *P = *PI; 2205 Phi->addIncoming(predMap[P], P); 2206 } 2207 2208 VN.add(Phi, ValNo); 2209 addToLeaderTable(ValNo, Phi, CurrentBlock); 2210 Phi->setDebugLoc(CurInst->getDebugLoc()); 2211 CurInst->replaceAllUsesWith(Phi); 2212 if (Phi->getType()->isPointerTy()) { 2213 // Because we have added a PHI-use of the pointer value, it has now 2214 // "escaped" from alias analysis' perspective. We need to inform 2215 // AA of this. 2216 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2217 ++ii) { 2218 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2219 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2220 } 2221 2222 if (MD) 2223 MD->invalidateCachedPointerInfo(Phi); 2224 } 2225 VN.erase(CurInst); 2226 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2227 2228 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2229 if (MD) MD->removeInstruction(CurInst); 2230 CurInst->eraseFromParent(); 2231 DEBUG(verifyRemoved(CurInst)); 2232 Changed = true; 2233 } 2234 } 2235 2236 if (splitCriticalEdges()) 2237 Changed = true; 2238 2239 return Changed; 2240} 2241 2242/// splitCriticalEdges - Split critical edges found during the previous 2243/// iteration that may enable further optimization. 2244bool GVN::splitCriticalEdges() { 2245 if (toSplit.empty()) 2246 return false; 2247 do { 2248 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2249 SplitCriticalEdge(Edge.first, Edge.second, this); 2250 } while (!toSplit.empty()); 2251 if (MD) MD->invalidateCachedPredecessors(); 2252 return true; 2253} 2254 2255/// iterateOnFunction - Executes one iteration of GVN 2256bool GVN::iterateOnFunction(Function &F) { 2257 cleanupGlobalSets(); 2258 2259 // Top-down walk of the dominator tree 2260 bool Changed = false; 2261#if 0 2262 // Needed for value numbering with phi construction to work. 2263 ReversePostOrderTraversal<Function*> RPOT(&F); 2264 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2265 RE = RPOT.end(); RI != RE; ++RI) 2266 Changed |= processBlock(*RI); 2267#else 2268 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2269 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2270 Changed |= processBlock(DI->getBlock()); 2271#endif 2272 2273 return Changed; 2274} 2275 2276void GVN::cleanupGlobalSets() { 2277 VN.clear(); 2278 LeaderTable.clear(); 2279 TableAllocator.Reset(); 2280} 2281 2282/// verifyRemoved - Verify that the specified instruction does not occur in our 2283/// internal data structures. 2284void GVN::verifyRemoved(const Instruction *Inst) const { 2285 VN.verifyRemoved(Inst); 2286 2287 // Walk through the value number scope to make sure the instruction isn't 2288 // ferreted away in it. 2289 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2290 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2291 const LeaderTableEntry *Node = &I->second; 2292 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2293 2294 while (Node->Next) { 2295 Node = Node->Next; 2296 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2297 } 2298 } 2299} 2300