GVN.cpp revision e170c76ccdcf9b0343d2d5a2805010ff77b8b56e
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/Target/TargetLibraryInfo.h" 35#include "llvm/Transforms/Utils/BasicBlockUtils.h" 36#include "llvm/Transforms/Utils/SSAUpdater.h" 37#include "llvm/ADT/DenseMap.h" 38#include "llvm/ADT/DepthFirstIterator.h" 39#include "llvm/ADT/SmallPtrSet.h" 40#include "llvm/ADT/Statistic.h" 41#include "llvm/Support/Allocator.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/IRBuilder.h" 45#include "llvm/Support/PatternMatch.h" 46using namespace llvm; 47using namespace PatternMatch; 48 49STATISTIC(NumGVNInstr, "Number of instructions deleted"); 50STATISTIC(NumGVNLoad, "Number of loads deleted"); 51STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 52STATISTIC(NumGVNBlocks, "Number of blocks merged"); 53STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 54STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 55STATISTIC(NumPRELoad, "Number of loads PRE'd"); 56 57static cl::opt<bool> EnablePRE("enable-pre", 58 cl::init(true), cl::Hidden); 59static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 60 61//===----------------------------------------------------------------------===// 62// ValueTable Class 63//===----------------------------------------------------------------------===// 64 65/// This class holds the mapping between values and value numbers. It is used 66/// as an efficient mechanism to determine the expression-wise equivalence of 67/// two values. 68namespace { 69 struct Expression { 70 uint32_t opcode; 71 Type *type; 72 SmallVector<uint32_t, 4> varargs; 73 74 Expression(uint32_t o = ~2U) : opcode(o) { } 75 76 bool operator==(const Expression &other) const { 77 if (opcode != other.opcode) 78 return false; 79 if (opcode == ~0U || opcode == ~1U) 80 return true; 81 if (type != other.type) 82 return false; 83 if (varargs != other.varargs) 84 return false; 85 return true; 86 } 87 }; 88 89 class ValueTable { 90 DenseMap<Value*, uint32_t> valueNumbering; 91 DenseMap<Expression, uint32_t> expressionNumbering; 92 AliasAnalysis *AA; 93 MemoryDependenceAnalysis *MD; 94 DominatorTree *DT; 95 96 uint32_t nextValueNumber; 97 98 Expression create_expression(Instruction* I); 99 Expression create_extractvalue_expression(ExtractValueInst* EI); 100 uint32_t lookup_or_add_call(CallInst* C); 101 public: 102 ValueTable() : nextValueNumber(1) { } 103 uint32_t lookup_or_add(Value *V); 104 uint32_t lookup(Value *V) const; 105 void add(Value *V, uint32_t num); 106 void clear(); 107 void erase(Value *v); 108 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 109 AliasAnalysis *getAliasAnalysis() const { return AA; } 110 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 111 void setDomTree(DominatorTree* D) { DT = D; } 112 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 113 void verifyRemoved(const Value *) const; 114 }; 115} 116 117namespace llvm { 118template <> struct DenseMapInfo<Expression> { 119 static inline Expression getEmptyKey() { 120 return ~0U; 121 } 122 123 static inline Expression getTombstoneKey() { 124 return ~1U; 125 } 126 127 static unsigned getHashValue(const Expression e) { 128 unsigned hash = e.opcode; 129 130 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 131 (unsigned)((uintptr_t)e.type >> 9)); 132 133 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 134 E = e.varargs.end(); I != E; ++I) 135 hash = *I + hash * 37; 136 137 return hash; 138 } 139 static bool isEqual(const Expression &LHS, const Expression &RHS) { 140 return LHS == RHS; 141 } 142}; 143 144} 145 146//===----------------------------------------------------------------------===// 147// ValueTable Internal Functions 148//===----------------------------------------------------------------------===// 149 150Expression ValueTable::create_expression(Instruction *I) { 151 Expression e; 152 e.type = I->getType(); 153 e.opcode = I->getOpcode(); 154 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 155 OI != OE; ++OI) 156 e.varargs.push_back(lookup_or_add(*OI)); 157 if (I->isCommutative()) { 158 // Ensure that commutative instructions that only differ by a permutation 159 // of their operands get the same value number by sorting the operand value 160 // numbers. Since all commutative instructions have two operands it is more 161 // efficient to sort by hand rather than using, say, std::sort. 162 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 163 if (e.varargs[0] > e.varargs[1]) 164 std::swap(e.varargs[0], e.varargs[1]); 165 } 166 167 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 168 // Sort the operand value numbers so x<y and y>x get the same value number. 169 CmpInst::Predicate Predicate = C->getPredicate(); 170 if (e.varargs[0] > e.varargs[1]) { 171 std::swap(e.varargs[0], e.varargs[1]); 172 Predicate = CmpInst::getSwappedPredicate(Predicate); 173 } 174 e.opcode = (C->getOpcode() << 8) | Predicate; 175 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 176 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 177 II != IE; ++II) 178 e.varargs.push_back(*II); 179 } 180 181 return e; 182} 183 184Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 185 assert(EI != 0 && "Not an ExtractValueInst?"); 186 Expression e; 187 e.type = EI->getType(); 188 e.opcode = 0; 189 190 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 191 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 192 // EI might be an extract from one of our recognised intrinsics. If it 193 // is we'll synthesize a semantically equivalent expression instead on 194 // an extract value expression. 195 switch (I->getIntrinsicID()) { 196 case Intrinsic::sadd_with_overflow: 197 case Intrinsic::uadd_with_overflow: 198 e.opcode = Instruction::Add; 199 break; 200 case Intrinsic::ssub_with_overflow: 201 case Intrinsic::usub_with_overflow: 202 e.opcode = Instruction::Sub; 203 break; 204 case Intrinsic::smul_with_overflow: 205 case Intrinsic::umul_with_overflow: 206 e.opcode = Instruction::Mul; 207 break; 208 default: 209 break; 210 } 211 212 if (e.opcode != 0) { 213 // Intrinsic recognized. Grab its args to finish building the expression. 214 assert(I->getNumArgOperands() == 2 && 215 "Expect two args for recognised intrinsics."); 216 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 217 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 218 return e; 219 } 220 } 221 222 // Not a recognised intrinsic. Fall back to producing an extract value 223 // expression. 224 e.opcode = EI->getOpcode(); 225 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 226 OI != OE; ++OI) 227 e.varargs.push_back(lookup_or_add(*OI)); 228 229 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 230 II != IE; ++II) 231 e.varargs.push_back(*II); 232 233 return e; 234} 235 236//===----------------------------------------------------------------------===// 237// ValueTable External Functions 238//===----------------------------------------------------------------------===// 239 240/// add - Insert a value into the table with a specified value number. 241void ValueTable::add(Value *V, uint32_t num) { 242 valueNumbering.insert(std::make_pair(V, num)); 243} 244 245uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 246 if (AA->doesNotAccessMemory(C)) { 247 Expression exp = create_expression(C); 248 uint32_t& e = expressionNumbering[exp]; 249 if (!e) e = nextValueNumber++; 250 valueNumbering[C] = e; 251 return e; 252 } else if (AA->onlyReadsMemory(C)) { 253 Expression exp = create_expression(C); 254 uint32_t& e = expressionNumbering[exp]; 255 if (!e) { 256 e = nextValueNumber++; 257 valueNumbering[C] = e; 258 return e; 259 } 260 if (!MD) { 261 e = nextValueNumber++; 262 valueNumbering[C] = e; 263 return e; 264 } 265 266 MemDepResult local_dep = MD->getDependency(C); 267 268 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 269 valueNumbering[C] = nextValueNumber; 270 return nextValueNumber++; 271 } 272 273 if (local_dep.isDef()) { 274 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 275 276 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 277 valueNumbering[C] = nextValueNumber; 278 return nextValueNumber++; 279 } 280 281 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 282 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 283 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 284 if (c_vn != cd_vn) { 285 valueNumbering[C] = nextValueNumber; 286 return nextValueNumber++; 287 } 288 } 289 290 uint32_t v = lookup_or_add(local_cdep); 291 valueNumbering[C] = v; 292 return v; 293 } 294 295 // Non-local case. 296 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 297 MD->getNonLocalCallDependency(CallSite(C)); 298 // FIXME: Move the checking logic to MemDep! 299 CallInst* cdep = 0; 300 301 // Check to see if we have a single dominating call instruction that is 302 // identical to C. 303 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 304 const NonLocalDepEntry *I = &deps[i]; 305 if (I->getResult().isNonLocal()) 306 continue; 307 308 // We don't handle non-definitions. If we already have a call, reject 309 // instruction dependencies. 310 if (!I->getResult().isDef() || cdep != 0) { 311 cdep = 0; 312 break; 313 } 314 315 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 316 // FIXME: All duplicated with non-local case. 317 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 318 cdep = NonLocalDepCall; 319 continue; 320 } 321 322 cdep = 0; 323 break; 324 } 325 326 if (!cdep) { 327 valueNumbering[C] = nextValueNumber; 328 return nextValueNumber++; 329 } 330 331 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 332 valueNumbering[C] = nextValueNumber; 333 return nextValueNumber++; 334 } 335 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 336 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 337 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 338 if (c_vn != cd_vn) { 339 valueNumbering[C] = nextValueNumber; 340 return nextValueNumber++; 341 } 342 } 343 344 uint32_t v = lookup_or_add(cdep); 345 valueNumbering[C] = v; 346 return v; 347 348 } else { 349 valueNumbering[C] = nextValueNumber; 350 return nextValueNumber++; 351 } 352} 353 354/// lookup_or_add - Returns the value number for the specified value, assigning 355/// it a new number if it did not have one before. 356uint32_t ValueTable::lookup_or_add(Value *V) { 357 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 358 if (VI != valueNumbering.end()) 359 return VI->second; 360 361 if (!isa<Instruction>(V)) { 362 valueNumbering[V] = nextValueNumber; 363 return nextValueNumber++; 364 } 365 366 Instruction* I = cast<Instruction>(V); 367 Expression exp; 368 switch (I->getOpcode()) { 369 case Instruction::Call: 370 return lookup_or_add_call(cast<CallInst>(I)); 371 case Instruction::Add: 372 case Instruction::FAdd: 373 case Instruction::Sub: 374 case Instruction::FSub: 375 case Instruction::Mul: 376 case Instruction::FMul: 377 case Instruction::UDiv: 378 case Instruction::SDiv: 379 case Instruction::FDiv: 380 case Instruction::URem: 381 case Instruction::SRem: 382 case Instruction::FRem: 383 case Instruction::Shl: 384 case Instruction::LShr: 385 case Instruction::AShr: 386 case Instruction::And: 387 case Instruction::Or : 388 case Instruction::Xor: 389 case Instruction::ICmp: 390 case Instruction::FCmp: 391 case Instruction::Trunc: 392 case Instruction::ZExt: 393 case Instruction::SExt: 394 case Instruction::FPToUI: 395 case Instruction::FPToSI: 396 case Instruction::UIToFP: 397 case Instruction::SIToFP: 398 case Instruction::FPTrunc: 399 case Instruction::FPExt: 400 case Instruction::PtrToInt: 401 case Instruction::IntToPtr: 402 case Instruction::BitCast: 403 case Instruction::Select: 404 case Instruction::ExtractElement: 405 case Instruction::InsertElement: 406 case Instruction::ShuffleVector: 407 case Instruction::InsertValue: 408 case Instruction::GetElementPtr: 409 exp = create_expression(I); 410 break; 411 case Instruction::ExtractValue: 412 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 413 break; 414 default: 415 valueNumbering[V] = nextValueNumber; 416 return nextValueNumber++; 417 } 418 419 uint32_t& e = expressionNumbering[exp]; 420 if (!e) e = nextValueNumber++; 421 valueNumbering[V] = e; 422 return e; 423} 424 425/// lookup - Returns the value number of the specified value. Fails if 426/// the value has not yet been numbered. 427uint32_t ValueTable::lookup(Value *V) const { 428 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 429 assert(VI != valueNumbering.end() && "Value not numbered?"); 430 return VI->second; 431} 432 433/// clear - Remove all entries from the ValueTable. 434void ValueTable::clear() { 435 valueNumbering.clear(); 436 expressionNumbering.clear(); 437 nextValueNumber = 1; 438} 439 440/// erase - Remove a value from the value numbering. 441void ValueTable::erase(Value *V) { 442 valueNumbering.erase(V); 443} 444 445/// verifyRemoved - Verify that the value is removed from all internal data 446/// structures. 447void ValueTable::verifyRemoved(const Value *V) const { 448 for (DenseMap<Value*, uint32_t>::const_iterator 449 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 450 assert(I->first != V && "Inst still occurs in value numbering map!"); 451 } 452} 453 454//===----------------------------------------------------------------------===// 455// GVN Pass 456//===----------------------------------------------------------------------===// 457 458namespace { 459 460 class GVN : public FunctionPass { 461 bool NoLoads; 462 MemoryDependenceAnalysis *MD; 463 DominatorTree *DT; 464 const TargetData *TD; 465 const TargetLibraryInfo *TLI; 466 467 ValueTable VN; 468 469 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 470 /// have that value number. Use findLeader to query it. 471 struct LeaderTableEntry { 472 Value *Val; 473 BasicBlock *BB; 474 LeaderTableEntry *Next; 475 }; 476 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 477 BumpPtrAllocator TableAllocator; 478 479 SmallVector<Instruction*, 8> InstrsToErase; 480 public: 481 static char ID; // Pass identification, replacement for typeid 482 explicit GVN(bool noloads = false) 483 : FunctionPass(ID), NoLoads(noloads), MD(0) { 484 initializeGVNPass(*PassRegistry::getPassRegistry()); 485 } 486 487 bool runOnFunction(Function &F); 488 489 /// markInstructionForDeletion - This removes the specified instruction from 490 /// our various maps and marks it for deletion. 491 void markInstructionForDeletion(Instruction *I) { 492 VN.erase(I); 493 InstrsToErase.push_back(I); 494 } 495 496 const TargetData *getTargetData() const { return TD; } 497 DominatorTree &getDominatorTree() const { return *DT; } 498 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 499 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 500 private: 501 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 502 /// its value number. 503 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 504 LeaderTableEntry &Curr = LeaderTable[N]; 505 if (!Curr.Val) { 506 Curr.Val = V; 507 Curr.BB = BB; 508 return; 509 } 510 511 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 512 Node->Val = V; 513 Node->BB = BB; 514 Node->Next = Curr.Next; 515 Curr.Next = Node; 516 } 517 518 /// removeFromLeaderTable - Scan the list of values corresponding to a given 519 /// value number, and remove the given value if encountered. 520 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 521 LeaderTableEntry* Prev = 0; 522 LeaderTableEntry* Curr = &LeaderTable[N]; 523 524 while (Curr->Val != V || Curr->BB != BB) { 525 Prev = Curr; 526 Curr = Curr->Next; 527 } 528 529 if (Prev) { 530 Prev->Next = Curr->Next; 531 } else { 532 if (!Curr->Next) { 533 Curr->Val = 0; 534 Curr->BB = 0; 535 } else { 536 LeaderTableEntry* Next = Curr->Next; 537 Curr->Val = Next->Val; 538 Curr->BB = Next->BB; 539 Curr->Next = Next->Next; 540 } 541 } 542 } 543 544 // List of critical edges to be split between iterations. 545 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 546 547 // This transformation requires dominator postdominator info 548 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 549 AU.addRequired<DominatorTree>(); 550 AU.addRequired<TargetLibraryInfo>(); 551 if (!NoLoads) 552 AU.addRequired<MemoryDependenceAnalysis>(); 553 AU.addRequired<AliasAnalysis>(); 554 555 AU.addPreserved<DominatorTree>(); 556 AU.addPreserved<AliasAnalysis>(); 557 } 558 559 560 // Helper fuctions 561 // FIXME: eliminate or document these better 562 bool processLoad(LoadInst *L); 563 bool processInstruction(Instruction *I); 564 bool processNonLocalLoad(LoadInst *L); 565 bool processBlock(BasicBlock *BB); 566 void dump(DenseMap<uint32_t, Value*> &d); 567 bool iterateOnFunction(Function &F); 568 bool performPRE(Function &F); 569 Value *findLeader(BasicBlock *BB, uint32_t num); 570 void cleanupGlobalSets(); 571 void verifyRemoved(const Instruction *I) const; 572 bool splitCriticalEdges(); 573 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 574 BasicBlock *Root); 575 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root); 576 }; 577 578 char GVN::ID = 0; 579} 580 581// createGVNPass - The public interface to this file... 582FunctionPass *llvm::createGVNPass(bool NoLoads) { 583 return new GVN(NoLoads); 584} 585 586INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 587INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 588INITIALIZE_PASS_DEPENDENCY(DominatorTree) 589INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 590INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 591INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 592 593void GVN::dump(DenseMap<uint32_t, Value*>& d) { 594 errs() << "{\n"; 595 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 596 E = d.end(); I != E; ++I) { 597 errs() << I->first << "\n"; 598 I->second->dump(); 599 } 600 errs() << "}\n"; 601} 602 603/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 604/// we're analyzing is fully available in the specified block. As we go, keep 605/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 606/// map is actually a tri-state map with the following values: 607/// 0) we know the block *is not* fully available. 608/// 1) we know the block *is* fully available. 609/// 2) we do not know whether the block is fully available or not, but we are 610/// currently speculating that it will be. 611/// 3) we are speculating for this block and have used that to speculate for 612/// other blocks. 613static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 614 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 615 // Optimistically assume that the block is fully available and check to see 616 // if we already know about this block in one lookup. 617 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 618 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 619 620 // If the entry already existed for this block, return the precomputed value. 621 if (!IV.second) { 622 // If this is a speculative "available" value, mark it as being used for 623 // speculation of other blocks. 624 if (IV.first->second == 2) 625 IV.first->second = 3; 626 return IV.first->second != 0; 627 } 628 629 // Otherwise, see if it is fully available in all predecessors. 630 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 631 632 // If this block has no predecessors, it isn't live-in here. 633 if (PI == PE) 634 goto SpeculationFailure; 635 636 for (; PI != PE; ++PI) 637 // If the value isn't fully available in one of our predecessors, then it 638 // isn't fully available in this block either. Undo our previous 639 // optimistic assumption and bail out. 640 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 641 goto SpeculationFailure; 642 643 return true; 644 645// SpeculationFailure - If we get here, we found out that this is not, after 646// all, a fully-available block. We have a problem if we speculated on this and 647// used the speculation to mark other blocks as available. 648SpeculationFailure: 649 char &BBVal = FullyAvailableBlocks[BB]; 650 651 // If we didn't speculate on this, just return with it set to false. 652 if (BBVal == 2) { 653 BBVal = 0; 654 return false; 655 } 656 657 // If we did speculate on this value, we could have blocks set to 1 that are 658 // incorrect. Walk the (transitive) successors of this block and mark them as 659 // 0 if set to one. 660 SmallVector<BasicBlock*, 32> BBWorklist; 661 BBWorklist.push_back(BB); 662 663 do { 664 BasicBlock *Entry = BBWorklist.pop_back_val(); 665 // Note that this sets blocks to 0 (unavailable) if they happen to not 666 // already be in FullyAvailableBlocks. This is safe. 667 char &EntryVal = FullyAvailableBlocks[Entry]; 668 if (EntryVal == 0) continue; // Already unavailable. 669 670 // Mark as unavailable. 671 EntryVal = 0; 672 673 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 674 BBWorklist.push_back(*I); 675 } while (!BBWorklist.empty()); 676 677 return false; 678} 679 680 681/// CanCoerceMustAliasedValueToLoad - Return true if 682/// CoerceAvailableValueToLoadType will succeed. 683static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 684 Type *LoadTy, 685 const TargetData &TD) { 686 // If the loaded or stored value is an first class array or struct, don't try 687 // to transform them. We need to be able to bitcast to integer. 688 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 689 StoredVal->getType()->isStructTy() || 690 StoredVal->getType()->isArrayTy()) 691 return false; 692 693 // The store has to be at least as big as the load. 694 if (TD.getTypeSizeInBits(StoredVal->getType()) < 695 TD.getTypeSizeInBits(LoadTy)) 696 return false; 697 698 return true; 699} 700 701 702/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 703/// then a load from a must-aliased pointer of a different type, try to coerce 704/// the stored value. LoadedTy is the type of the load we want to replace and 705/// InsertPt is the place to insert new instructions. 706/// 707/// If we can't do it, return null. 708static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 709 Type *LoadedTy, 710 Instruction *InsertPt, 711 const TargetData &TD) { 712 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 713 return 0; 714 715 // If this is already the right type, just return it. 716 Type *StoredValTy = StoredVal->getType(); 717 718 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 719 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 720 721 // If the store and reload are the same size, we can always reuse it. 722 if (StoreSize == LoadSize) { 723 // Pointer to Pointer -> use bitcast. 724 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 725 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 726 727 // Convert source pointers to integers, which can be bitcast. 728 if (StoredValTy->isPointerTy()) { 729 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 730 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 731 } 732 733 Type *TypeToCastTo = LoadedTy; 734 if (TypeToCastTo->isPointerTy()) 735 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 736 737 if (StoredValTy != TypeToCastTo) 738 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 739 740 // Cast to pointer if the load needs a pointer type. 741 if (LoadedTy->isPointerTy()) 742 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 743 744 return StoredVal; 745 } 746 747 // If the loaded value is smaller than the available value, then we can 748 // extract out a piece from it. If the available value is too small, then we 749 // can't do anything. 750 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 751 752 // Convert source pointers to integers, which can be manipulated. 753 if (StoredValTy->isPointerTy()) { 754 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 755 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 756 } 757 758 // Convert vectors and fp to integer, which can be manipulated. 759 if (!StoredValTy->isIntegerTy()) { 760 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 761 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 762 } 763 764 // If this is a big-endian system, we need to shift the value down to the low 765 // bits so that a truncate will work. 766 if (TD.isBigEndian()) { 767 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 768 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 769 } 770 771 // Truncate the integer to the right size now. 772 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 773 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 774 775 if (LoadedTy == NewIntTy) 776 return StoredVal; 777 778 // If the result is a pointer, inttoptr. 779 if (LoadedTy->isPointerTy()) 780 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 781 782 // Otherwise, bitcast. 783 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 784} 785 786/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 787/// memdep query of a load that ends up being a clobbering memory write (store, 788/// memset, memcpy, memmove). This means that the write *may* provide bits used 789/// by the load but we can't be sure because the pointers don't mustalias. 790/// 791/// Check this case to see if there is anything more we can do before we give 792/// up. This returns -1 if we have to give up, or a byte number in the stored 793/// value of the piece that feeds the load. 794static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 795 Value *WritePtr, 796 uint64_t WriteSizeInBits, 797 const TargetData &TD) { 798 // If the loaded or stored value is a first class array or struct, don't try 799 // to transform them. We need to be able to bitcast to integer. 800 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 801 return -1; 802 803 int64_t StoreOffset = 0, LoadOffset = 0; 804 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 805 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 806 if (StoreBase != LoadBase) 807 return -1; 808 809 // If the load and store are to the exact same address, they should have been 810 // a must alias. AA must have gotten confused. 811 // FIXME: Study to see if/when this happens. One case is forwarding a memset 812 // to a load from the base of the memset. 813#if 0 814 if (LoadOffset == StoreOffset) { 815 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 816 << "Base = " << *StoreBase << "\n" 817 << "Store Ptr = " << *WritePtr << "\n" 818 << "Store Offs = " << StoreOffset << "\n" 819 << "Load Ptr = " << *LoadPtr << "\n"; 820 abort(); 821 } 822#endif 823 824 // If the load and store don't overlap at all, the store doesn't provide 825 // anything to the load. In this case, they really don't alias at all, AA 826 // must have gotten confused. 827 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 828 829 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 830 return -1; 831 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 832 LoadSize >>= 3; 833 834 835 bool isAAFailure = false; 836 if (StoreOffset < LoadOffset) 837 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 838 else 839 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 840 841 if (isAAFailure) { 842#if 0 843 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 844 << "Base = " << *StoreBase << "\n" 845 << "Store Ptr = " << *WritePtr << "\n" 846 << "Store Offs = " << StoreOffset << "\n" 847 << "Load Ptr = " << *LoadPtr << "\n"; 848 abort(); 849#endif 850 return -1; 851 } 852 853 // If the Load isn't completely contained within the stored bits, we don't 854 // have all the bits to feed it. We could do something crazy in the future 855 // (issue a smaller load then merge the bits in) but this seems unlikely to be 856 // valuable. 857 if (StoreOffset > LoadOffset || 858 StoreOffset+StoreSize < LoadOffset+LoadSize) 859 return -1; 860 861 // Okay, we can do this transformation. Return the number of bytes into the 862 // store that the load is. 863 return LoadOffset-StoreOffset; 864} 865 866/// AnalyzeLoadFromClobberingStore - This function is called when we have a 867/// memdep query of a load that ends up being a clobbering store. 868static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 869 StoreInst *DepSI, 870 const TargetData &TD) { 871 // Cannot handle reading from store of first-class aggregate yet. 872 if (DepSI->getValueOperand()->getType()->isStructTy() || 873 DepSI->getValueOperand()->getType()->isArrayTy()) 874 return -1; 875 876 Value *StorePtr = DepSI->getPointerOperand(); 877 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 878 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 879 StorePtr, StoreSize, TD); 880} 881 882/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 883/// memdep query of a load that ends up being clobbered by another load. See if 884/// the other load can feed into the second load. 885static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 886 LoadInst *DepLI, const TargetData &TD){ 887 // Cannot handle reading from store of first-class aggregate yet. 888 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 889 return -1; 890 891 Value *DepPtr = DepLI->getPointerOperand(); 892 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 893 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 894 if (R != -1) return R; 895 896 // If we have a load/load clobber an DepLI can be widened to cover this load, 897 // then we should widen it! 898 int64_t LoadOffs = 0; 899 const Value *LoadBase = 900 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 901 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 902 903 unsigned Size = MemoryDependenceAnalysis:: 904 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 905 if (Size == 0) return -1; 906 907 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 908} 909 910 911 912static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 913 MemIntrinsic *MI, 914 const TargetData &TD) { 915 // If the mem operation is a non-constant size, we can't handle it. 916 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 917 if (SizeCst == 0) return -1; 918 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 919 920 // If this is memset, we just need to see if the offset is valid in the size 921 // of the memset.. 922 if (MI->getIntrinsicID() == Intrinsic::memset) 923 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 924 MemSizeInBits, TD); 925 926 // If we have a memcpy/memmove, the only case we can handle is if this is a 927 // copy from constant memory. In that case, we can read directly from the 928 // constant memory. 929 MemTransferInst *MTI = cast<MemTransferInst>(MI); 930 931 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 932 if (Src == 0) return -1; 933 934 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 935 if (GV == 0 || !GV->isConstant()) return -1; 936 937 // See if the access is within the bounds of the transfer. 938 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 939 MI->getDest(), MemSizeInBits, TD); 940 if (Offset == -1) 941 return Offset; 942 943 // Otherwise, see if we can constant fold a load from the constant with the 944 // offset applied as appropriate. 945 Src = ConstantExpr::getBitCast(Src, 946 llvm::Type::getInt8PtrTy(Src->getContext())); 947 Constant *OffsetCst = 948 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 949 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 950 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 951 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 952 return Offset; 953 return -1; 954} 955 956 957/// GetStoreValueForLoad - This function is called when we have a 958/// memdep query of a load that ends up being a clobbering store. This means 959/// that the store provides bits used by the load but we the pointers don't 960/// mustalias. Check this case to see if there is anything more we can do 961/// before we give up. 962static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 963 Type *LoadTy, 964 Instruction *InsertPt, const TargetData &TD){ 965 LLVMContext &Ctx = SrcVal->getType()->getContext(); 966 967 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 968 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 969 970 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 971 972 // Compute which bits of the stored value are being used by the load. Convert 973 // to an integer type to start with. 974 if (SrcVal->getType()->isPointerTy()) 975 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx)); 976 if (!SrcVal->getType()->isIntegerTy()) 977 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 978 979 // Shift the bits to the least significant depending on endianness. 980 unsigned ShiftAmt; 981 if (TD.isLittleEndian()) 982 ShiftAmt = Offset*8; 983 else 984 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 985 986 if (ShiftAmt) 987 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 988 989 if (LoadSize != StoreSize) 990 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 991 992 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 993} 994 995/// GetLoadValueForLoad - This function is called when we have a 996/// memdep query of a load that ends up being a clobbering load. This means 997/// that the load *may* provide bits used by the load but we can't be sure 998/// because the pointers don't mustalias. Check this case to see if there is 999/// anything more we can do before we give up. 1000static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1001 Type *LoadTy, Instruction *InsertPt, 1002 GVN &gvn) { 1003 const TargetData &TD = *gvn.getTargetData(); 1004 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1005 // widen SrcVal out to a larger load. 1006 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1007 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1008 if (Offset+LoadSize > SrcValSize) { 1009 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1010 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1011 // If we have a load/load clobber an DepLI can be widened to cover this 1012 // load, then we should widen it to the next power of 2 size big enough! 1013 unsigned NewLoadSize = Offset+LoadSize; 1014 if (!isPowerOf2_32(NewLoadSize)) 1015 NewLoadSize = NextPowerOf2(NewLoadSize); 1016 1017 Value *PtrVal = SrcVal->getPointerOperand(); 1018 1019 // Insert the new load after the old load. This ensures that subsequent 1020 // memdep queries will find the new load. We can't easily remove the old 1021 // load completely because it is already in the value numbering table. 1022 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1023 Type *DestPTy = 1024 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1025 DestPTy = PointerType::get(DestPTy, 1026 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1027 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1028 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1029 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1030 NewLoad->takeName(SrcVal); 1031 NewLoad->setAlignment(SrcVal->getAlignment()); 1032 1033 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1034 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1035 1036 // Replace uses of the original load with the wider load. On a big endian 1037 // system, we need to shift down to get the relevant bits. 1038 Value *RV = NewLoad; 1039 if (TD.isBigEndian()) 1040 RV = Builder.CreateLShr(RV, 1041 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1042 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1043 SrcVal->replaceAllUsesWith(RV); 1044 1045 // We would like to use gvn.markInstructionForDeletion here, but we can't 1046 // because the load is already memoized into the leader map table that GVN 1047 // tracks. It is potentially possible to remove the load from the table, 1048 // but then there all of the operations based on it would need to be 1049 // rehashed. Just leave the dead load around. 1050 gvn.getMemDep().removeInstruction(SrcVal); 1051 SrcVal = NewLoad; 1052 } 1053 1054 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1055} 1056 1057 1058/// GetMemInstValueForLoad - This function is called when we have a 1059/// memdep query of a load that ends up being a clobbering mem intrinsic. 1060static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1061 Type *LoadTy, Instruction *InsertPt, 1062 const TargetData &TD){ 1063 LLVMContext &Ctx = LoadTy->getContext(); 1064 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1065 1066 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1067 1068 // We know that this method is only called when the mem transfer fully 1069 // provides the bits for the load. 1070 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1071 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1072 // independently of what the offset is. 1073 Value *Val = MSI->getValue(); 1074 if (LoadSize != 1) 1075 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1076 1077 Value *OneElt = Val; 1078 1079 // Splat the value out to the right number of bits. 1080 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1081 // If we can double the number of bytes set, do it. 1082 if (NumBytesSet*2 <= LoadSize) { 1083 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1084 Val = Builder.CreateOr(Val, ShVal); 1085 NumBytesSet <<= 1; 1086 continue; 1087 } 1088 1089 // Otherwise insert one byte at a time. 1090 Value *ShVal = Builder.CreateShl(Val, 1*8); 1091 Val = Builder.CreateOr(OneElt, ShVal); 1092 ++NumBytesSet; 1093 } 1094 1095 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1096 } 1097 1098 // Otherwise, this is a memcpy/memmove from a constant global. 1099 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1100 Constant *Src = cast<Constant>(MTI->getSource()); 1101 1102 // Otherwise, see if we can constant fold a load from the constant with the 1103 // offset applied as appropriate. 1104 Src = ConstantExpr::getBitCast(Src, 1105 llvm::Type::getInt8PtrTy(Src->getContext())); 1106 Constant *OffsetCst = 1107 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1108 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1109 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1110 return ConstantFoldLoadFromConstPtr(Src, &TD); 1111} 1112 1113namespace { 1114 1115struct AvailableValueInBlock { 1116 /// BB - The basic block in question. 1117 BasicBlock *BB; 1118 enum ValType { 1119 SimpleVal, // A simple offsetted value that is accessed. 1120 LoadVal, // A value produced by a load. 1121 MemIntrin // A memory intrinsic which is loaded from. 1122 }; 1123 1124 /// V - The value that is live out of the block. 1125 PointerIntPair<Value *, 2, ValType> Val; 1126 1127 /// Offset - The byte offset in Val that is interesting for the load query. 1128 unsigned Offset; 1129 1130 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1131 unsigned Offset = 0) { 1132 AvailableValueInBlock Res; 1133 Res.BB = BB; 1134 Res.Val.setPointer(V); 1135 Res.Val.setInt(SimpleVal); 1136 Res.Offset = Offset; 1137 return Res; 1138 } 1139 1140 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1141 unsigned Offset = 0) { 1142 AvailableValueInBlock Res; 1143 Res.BB = BB; 1144 Res.Val.setPointer(MI); 1145 Res.Val.setInt(MemIntrin); 1146 Res.Offset = Offset; 1147 return Res; 1148 } 1149 1150 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1151 unsigned Offset = 0) { 1152 AvailableValueInBlock Res; 1153 Res.BB = BB; 1154 Res.Val.setPointer(LI); 1155 Res.Val.setInt(LoadVal); 1156 Res.Offset = Offset; 1157 return Res; 1158 } 1159 1160 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1161 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1162 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1163 1164 Value *getSimpleValue() const { 1165 assert(isSimpleValue() && "Wrong accessor"); 1166 return Val.getPointer(); 1167 } 1168 1169 LoadInst *getCoercedLoadValue() const { 1170 assert(isCoercedLoadValue() && "Wrong accessor"); 1171 return cast<LoadInst>(Val.getPointer()); 1172 } 1173 1174 MemIntrinsic *getMemIntrinValue() const { 1175 assert(isMemIntrinValue() && "Wrong accessor"); 1176 return cast<MemIntrinsic>(Val.getPointer()); 1177 } 1178 1179 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1180 /// defined here to the specified type. This handles various coercion cases. 1181 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1182 Value *Res; 1183 if (isSimpleValue()) { 1184 Res = getSimpleValue(); 1185 if (Res->getType() != LoadTy) { 1186 const TargetData *TD = gvn.getTargetData(); 1187 assert(TD && "Need target data to handle type mismatch case"); 1188 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1189 *TD); 1190 1191 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1192 << *getSimpleValue() << '\n' 1193 << *Res << '\n' << "\n\n\n"); 1194 } 1195 } else if (isCoercedLoadValue()) { 1196 LoadInst *Load = getCoercedLoadValue(); 1197 if (Load->getType() == LoadTy && Offset == 0) { 1198 Res = Load; 1199 } else { 1200 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1201 gvn); 1202 1203 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1204 << *getCoercedLoadValue() << '\n' 1205 << *Res << '\n' << "\n\n\n"); 1206 } 1207 } else { 1208 const TargetData *TD = gvn.getTargetData(); 1209 assert(TD && "Need target data to handle type mismatch case"); 1210 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1211 LoadTy, BB->getTerminator(), *TD); 1212 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1213 << " " << *getMemIntrinValue() << '\n' 1214 << *Res << '\n' << "\n\n\n"); 1215 } 1216 return Res; 1217 } 1218}; 1219 1220} // end anonymous namespace 1221 1222/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1223/// construct SSA form, allowing us to eliminate LI. This returns the value 1224/// that should be used at LI's definition site. 1225static Value *ConstructSSAForLoadSet(LoadInst *LI, 1226 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1227 GVN &gvn) { 1228 // Check for the fully redundant, dominating load case. In this case, we can 1229 // just use the dominating value directly. 1230 if (ValuesPerBlock.size() == 1 && 1231 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1232 LI->getParent())) 1233 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1234 1235 // Otherwise, we have to construct SSA form. 1236 SmallVector<PHINode*, 8> NewPHIs; 1237 SSAUpdater SSAUpdate(&NewPHIs); 1238 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1239 1240 Type *LoadTy = LI->getType(); 1241 1242 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1243 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1244 BasicBlock *BB = AV.BB; 1245 1246 if (SSAUpdate.HasValueForBlock(BB)) 1247 continue; 1248 1249 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1250 } 1251 1252 // Perform PHI construction. 1253 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1254 1255 // If new PHI nodes were created, notify alias analysis. 1256 if (V->getType()->isPointerTy()) { 1257 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1258 1259 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1260 AA->copyValue(LI, NewPHIs[i]); 1261 1262 // Now that we've copied information to the new PHIs, scan through 1263 // them again and inform alias analysis that we've added potentially 1264 // escaping uses to any values that are operands to these PHIs. 1265 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1266 PHINode *P = NewPHIs[i]; 1267 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1268 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1269 AA->addEscapingUse(P->getOperandUse(jj)); 1270 } 1271 } 1272 } 1273 1274 return V; 1275} 1276 1277static bool isLifetimeStart(const Instruction *Inst) { 1278 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1279 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1280 return false; 1281} 1282 1283/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1284/// non-local by performing PHI construction. 1285bool GVN::processNonLocalLoad(LoadInst *LI) { 1286 // Find the non-local dependencies of the load. 1287 SmallVector<NonLocalDepResult, 64> Deps; 1288 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1289 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1290 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1291 // << Deps.size() << *LI << '\n'); 1292 1293 // If we had to process more than one hundred blocks to find the 1294 // dependencies, this load isn't worth worrying about. Optimizing 1295 // it will be too expensive. 1296 unsigned NumDeps = Deps.size(); 1297 if (NumDeps > 100) 1298 return false; 1299 1300 // If we had a phi translation failure, we'll have a single entry which is a 1301 // clobber in the current block. Reject this early. 1302 if (NumDeps == 1 && 1303 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1304 DEBUG( 1305 dbgs() << "GVN: non-local load "; 1306 WriteAsOperand(dbgs(), LI); 1307 dbgs() << " has unknown dependencies\n"; 1308 ); 1309 return false; 1310 } 1311 1312 // Filter out useless results (non-locals, etc). Keep track of the blocks 1313 // where we have a value available in repl, also keep track of whether we see 1314 // dependencies that produce an unknown value for the load (such as a call 1315 // that could potentially clobber the load). 1316 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock; 1317 SmallVector<BasicBlock*, 64> UnavailableBlocks; 1318 1319 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1320 BasicBlock *DepBB = Deps[i].getBB(); 1321 MemDepResult DepInfo = Deps[i].getResult(); 1322 1323 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1324 UnavailableBlocks.push_back(DepBB); 1325 continue; 1326 } 1327 1328 if (DepInfo.isClobber()) { 1329 // The address being loaded in this non-local block may not be the same as 1330 // the pointer operand of the load if PHI translation occurs. Make sure 1331 // to consider the right address. 1332 Value *Address = Deps[i].getAddress(); 1333 1334 // If the dependence is to a store that writes to a superset of the bits 1335 // read by the load, we can extract the bits we need for the load from the 1336 // stored value. 1337 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1338 if (TD && Address) { 1339 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1340 DepSI, *TD); 1341 if (Offset != -1) { 1342 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1343 DepSI->getValueOperand(), 1344 Offset)); 1345 continue; 1346 } 1347 } 1348 } 1349 1350 // Check to see if we have something like this: 1351 // load i32* P 1352 // load i8* (P+1) 1353 // if we have this, replace the later with an extraction from the former. 1354 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1355 // If this is a clobber and L is the first instruction in its block, then 1356 // we have the first instruction in the entry block. 1357 if (DepLI != LI && Address && TD) { 1358 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1359 LI->getPointerOperand(), 1360 DepLI, *TD); 1361 1362 if (Offset != -1) { 1363 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1364 Offset)); 1365 continue; 1366 } 1367 } 1368 } 1369 1370 // If the clobbering value is a memset/memcpy/memmove, see if we can 1371 // forward a value on from it. 1372 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1373 if (TD && Address) { 1374 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1375 DepMI, *TD); 1376 if (Offset != -1) { 1377 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1378 Offset)); 1379 continue; 1380 } 1381 } 1382 } 1383 1384 UnavailableBlocks.push_back(DepBB); 1385 continue; 1386 } 1387 1388 // DepInfo.isDef() here 1389 1390 Instruction *DepInst = DepInfo.getInst(); 1391 1392 // Loading the allocation -> undef. 1393 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1394 // Loading immediately after lifetime begin -> undef. 1395 isLifetimeStart(DepInst)) { 1396 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1397 UndefValue::get(LI->getType()))); 1398 continue; 1399 } 1400 1401 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1402 // Reject loads and stores that are to the same address but are of 1403 // different types if we have to. 1404 if (S->getValueOperand()->getType() != LI->getType()) { 1405 // If the stored value is larger or equal to the loaded value, we can 1406 // reuse it. 1407 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1408 LI->getType(), *TD)) { 1409 UnavailableBlocks.push_back(DepBB); 1410 continue; 1411 } 1412 } 1413 1414 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1415 S->getValueOperand())); 1416 continue; 1417 } 1418 1419 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1420 // If the types mismatch and we can't handle it, reject reuse of the load. 1421 if (LD->getType() != LI->getType()) { 1422 // If the stored value is larger or equal to the loaded value, we can 1423 // reuse it. 1424 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1425 UnavailableBlocks.push_back(DepBB); 1426 continue; 1427 } 1428 } 1429 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1430 continue; 1431 } 1432 1433 UnavailableBlocks.push_back(DepBB); 1434 continue; 1435 } 1436 1437 // If we have no predecessors that produce a known value for this load, exit 1438 // early. 1439 if (ValuesPerBlock.empty()) return false; 1440 1441 // If all of the instructions we depend on produce a known value for this 1442 // load, then it is fully redundant and we can use PHI insertion to compute 1443 // its value. Insert PHIs and remove the fully redundant value now. 1444 if (UnavailableBlocks.empty()) { 1445 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1446 1447 // Perform PHI construction. 1448 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1449 LI->replaceAllUsesWith(V); 1450 1451 if (isa<PHINode>(V)) 1452 V->takeName(LI); 1453 if (V->getType()->isPointerTy()) 1454 MD->invalidateCachedPointerInfo(V); 1455 markInstructionForDeletion(LI); 1456 ++NumGVNLoad; 1457 return true; 1458 } 1459 1460 if (!EnablePRE || !EnableLoadPRE) 1461 return false; 1462 1463 // Okay, we have *some* definitions of the value. This means that the value 1464 // is available in some of our (transitive) predecessors. Lets think about 1465 // doing PRE of this load. This will involve inserting a new load into the 1466 // predecessor when it's not available. We could do this in general, but 1467 // prefer to not increase code size. As such, we only do this when we know 1468 // that we only have to insert *one* load (which means we're basically moving 1469 // the load, not inserting a new one). 1470 1471 SmallPtrSet<BasicBlock *, 4> Blockers; 1472 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1473 Blockers.insert(UnavailableBlocks[i]); 1474 1475 // Let's find the first basic block with more than one predecessor. Walk 1476 // backwards through predecessors if needed. 1477 BasicBlock *LoadBB = LI->getParent(); 1478 BasicBlock *TmpBB = LoadBB; 1479 1480 bool isSinglePred = false; 1481 bool allSingleSucc = true; 1482 while (TmpBB->getSinglePredecessor()) { 1483 isSinglePred = true; 1484 TmpBB = TmpBB->getSinglePredecessor(); 1485 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1486 return false; 1487 if (Blockers.count(TmpBB)) 1488 return false; 1489 1490 // If any of these blocks has more than one successor (i.e. if the edge we 1491 // just traversed was critical), then there are other paths through this 1492 // block along which the load may not be anticipated. Hoisting the load 1493 // above this block would be adding the load to execution paths along 1494 // which it was not previously executed. 1495 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1496 return false; 1497 } 1498 1499 assert(TmpBB); 1500 LoadBB = TmpBB; 1501 1502 // FIXME: It is extremely unclear what this loop is doing, other than 1503 // artificially restricting loadpre. 1504 if (isSinglePred) { 1505 bool isHot = false; 1506 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1507 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1508 if (AV.isSimpleValue()) 1509 // "Hot" Instruction is in some loop (because it dominates its dep. 1510 // instruction). 1511 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1512 if (DT->dominates(LI, I)) { 1513 isHot = true; 1514 break; 1515 } 1516 } 1517 1518 // We are interested only in "hot" instructions. We don't want to do any 1519 // mis-optimizations here. 1520 if (!isHot) 1521 return false; 1522 } 1523 1524 // Check to see how many predecessors have the loaded value fully 1525 // available. 1526 DenseMap<BasicBlock*, Value*> PredLoads; 1527 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1528 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1529 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1530 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1531 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1532 1533 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1534 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1535 PI != E; ++PI) { 1536 BasicBlock *Pred = *PI; 1537 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1538 continue; 1539 } 1540 PredLoads[Pred] = 0; 1541 1542 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1543 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1544 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1545 << Pred->getName() << "': " << *LI << '\n'); 1546 return false; 1547 } 1548 1549 if (LoadBB->isLandingPad()) { 1550 DEBUG(dbgs() 1551 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1552 << Pred->getName() << "': " << *LI << '\n'); 1553 return false; 1554 } 1555 1556 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1557 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1558 } 1559 } 1560 1561 if (!NeedToSplit.empty()) { 1562 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1563 return false; 1564 } 1565 1566 // Decide whether PRE is profitable for this load. 1567 unsigned NumUnavailablePreds = PredLoads.size(); 1568 assert(NumUnavailablePreds != 0 && 1569 "Fully available value should be eliminated above!"); 1570 1571 // If this load is unavailable in multiple predecessors, reject it. 1572 // FIXME: If we could restructure the CFG, we could make a common pred with 1573 // all the preds that don't have an available LI and insert a new load into 1574 // that one block. 1575 if (NumUnavailablePreds != 1) 1576 return false; 1577 1578 // Check if the load can safely be moved to all the unavailable predecessors. 1579 bool CanDoPRE = true; 1580 SmallVector<Instruction*, 8> NewInsts; 1581 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1582 E = PredLoads.end(); I != E; ++I) { 1583 BasicBlock *UnavailablePred = I->first; 1584 1585 // Do PHI translation to get its value in the predecessor if necessary. The 1586 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1587 1588 // If all preds have a single successor, then we know it is safe to insert 1589 // the load on the pred (?!?), so we can insert code to materialize the 1590 // pointer if it is not available. 1591 PHITransAddr Address(LI->getPointerOperand(), TD); 1592 Value *LoadPtr = 0; 1593 if (allSingleSucc) { 1594 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1595 *DT, NewInsts); 1596 } else { 1597 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1598 LoadPtr = Address.getAddr(); 1599 } 1600 1601 // If we couldn't find or insert a computation of this phi translated value, 1602 // we fail PRE. 1603 if (LoadPtr == 0) { 1604 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1605 << *LI->getPointerOperand() << "\n"); 1606 CanDoPRE = false; 1607 break; 1608 } 1609 1610 // Make sure it is valid to move this load here. We have to watch out for: 1611 // @1 = getelementptr (i8* p, ... 1612 // test p and branch if == 0 1613 // load @1 1614 // It is valid to have the getelementptr before the test, even if p can 1615 // be 0, as getelementptr only does address arithmetic. 1616 // If we are not pushing the value through any multiple-successor blocks 1617 // we do not have this case. Otherwise, check that the load is safe to 1618 // put anywhere; this can be improved, but should be conservatively safe. 1619 if (!allSingleSucc && 1620 // FIXME: REEVALUTE THIS. 1621 !isSafeToLoadUnconditionally(LoadPtr, 1622 UnavailablePred->getTerminator(), 1623 LI->getAlignment(), TD)) { 1624 CanDoPRE = false; 1625 break; 1626 } 1627 1628 I->second = LoadPtr; 1629 } 1630 1631 if (!CanDoPRE) { 1632 while (!NewInsts.empty()) { 1633 Instruction *I = NewInsts.pop_back_val(); 1634 if (MD) MD->removeInstruction(I); 1635 I->eraseFromParent(); 1636 } 1637 return false; 1638 } 1639 1640 // Okay, we can eliminate this load by inserting a reload in the predecessor 1641 // and using PHI construction to get the value in the other predecessors, do 1642 // it. 1643 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1644 DEBUG(if (!NewInsts.empty()) 1645 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1646 << *NewInsts.back() << '\n'); 1647 1648 // Assign value numbers to the new instructions. 1649 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1650 // FIXME: We really _ought_ to insert these value numbers into their 1651 // parent's availability map. However, in doing so, we risk getting into 1652 // ordering issues. If a block hasn't been processed yet, we would be 1653 // marking a value as AVAIL-IN, which isn't what we intend. 1654 VN.lookup_or_add(NewInsts[i]); 1655 } 1656 1657 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1658 E = PredLoads.end(); I != E; ++I) { 1659 BasicBlock *UnavailablePred = I->first; 1660 Value *LoadPtr = I->second; 1661 1662 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1663 LI->getAlignment(), 1664 UnavailablePred->getTerminator()); 1665 1666 // Transfer the old load's TBAA tag to the new load. 1667 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1668 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1669 1670 // Transfer DebugLoc. 1671 NewLoad->setDebugLoc(LI->getDebugLoc()); 1672 1673 // Add the newly created load. 1674 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1675 NewLoad)); 1676 MD->invalidateCachedPointerInfo(LoadPtr); 1677 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1678 } 1679 1680 // Perform PHI construction. 1681 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1682 LI->replaceAllUsesWith(V); 1683 if (isa<PHINode>(V)) 1684 V->takeName(LI); 1685 if (V->getType()->isPointerTy()) 1686 MD->invalidateCachedPointerInfo(V); 1687 markInstructionForDeletion(LI); 1688 ++NumPRELoad; 1689 return true; 1690} 1691 1692/// processLoad - Attempt to eliminate a load, first by eliminating it 1693/// locally, and then attempting non-local elimination if that fails. 1694bool GVN::processLoad(LoadInst *L) { 1695 if (!MD) 1696 return false; 1697 1698 if (!L->isSimple()) 1699 return false; 1700 1701 if (L->use_empty()) { 1702 markInstructionForDeletion(L); 1703 return true; 1704 } 1705 1706 // ... to a pointer that has been loaded from before... 1707 MemDepResult Dep = MD->getDependency(L); 1708 1709 // If we have a clobber and target data is around, see if this is a clobber 1710 // that we can fix up through code synthesis. 1711 if (Dep.isClobber() && TD) { 1712 // Check to see if we have something like this: 1713 // store i32 123, i32* %P 1714 // %A = bitcast i32* %P to i8* 1715 // %B = gep i8* %A, i32 1 1716 // %C = load i8* %B 1717 // 1718 // We could do that by recognizing if the clobber instructions are obviously 1719 // a common base + constant offset, and if the previous store (or memset) 1720 // completely covers this load. This sort of thing can happen in bitfield 1721 // access code. 1722 Value *AvailVal = 0; 1723 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1724 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1725 L->getPointerOperand(), 1726 DepSI, *TD); 1727 if (Offset != -1) 1728 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1729 L->getType(), L, *TD); 1730 } 1731 1732 // Check to see if we have something like this: 1733 // load i32* P 1734 // load i8* (P+1) 1735 // if we have this, replace the later with an extraction from the former. 1736 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1737 // If this is a clobber and L is the first instruction in its block, then 1738 // we have the first instruction in the entry block. 1739 if (DepLI == L) 1740 return false; 1741 1742 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1743 L->getPointerOperand(), 1744 DepLI, *TD); 1745 if (Offset != -1) 1746 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1747 } 1748 1749 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1750 // a value on from it. 1751 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1752 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1753 L->getPointerOperand(), 1754 DepMI, *TD); 1755 if (Offset != -1) 1756 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1757 } 1758 1759 if (AvailVal) { 1760 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1761 << *AvailVal << '\n' << *L << "\n\n\n"); 1762 1763 // Replace the load! 1764 L->replaceAllUsesWith(AvailVal); 1765 if (AvailVal->getType()->isPointerTy()) 1766 MD->invalidateCachedPointerInfo(AvailVal); 1767 markInstructionForDeletion(L); 1768 ++NumGVNLoad; 1769 return true; 1770 } 1771 } 1772 1773 // If the value isn't available, don't do anything! 1774 if (Dep.isClobber()) { 1775 DEBUG( 1776 // fast print dep, using operator<< on instruction is too slow. 1777 dbgs() << "GVN: load "; 1778 WriteAsOperand(dbgs(), L); 1779 Instruction *I = Dep.getInst(); 1780 dbgs() << " is clobbered by " << *I << '\n'; 1781 ); 1782 return false; 1783 } 1784 1785 // If it is defined in another block, try harder. 1786 if (Dep.isNonLocal()) 1787 return processNonLocalLoad(L); 1788 1789 if (!Dep.isDef()) { 1790 DEBUG( 1791 // fast print dep, using operator<< on instruction is too slow. 1792 dbgs() << "GVN: load "; 1793 WriteAsOperand(dbgs(), L); 1794 dbgs() << " has unknown dependence\n"; 1795 ); 1796 return false; 1797 } 1798 1799 Instruction *DepInst = Dep.getInst(); 1800 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1801 Value *StoredVal = DepSI->getValueOperand(); 1802 1803 // The store and load are to a must-aliased pointer, but they may not 1804 // actually have the same type. See if we know how to reuse the stored 1805 // value (depending on its type). 1806 if (StoredVal->getType() != L->getType()) { 1807 if (TD) { 1808 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1809 L, *TD); 1810 if (StoredVal == 0) 1811 return false; 1812 1813 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1814 << '\n' << *L << "\n\n\n"); 1815 } 1816 else 1817 return false; 1818 } 1819 1820 // Remove it! 1821 L->replaceAllUsesWith(StoredVal); 1822 if (StoredVal->getType()->isPointerTy()) 1823 MD->invalidateCachedPointerInfo(StoredVal); 1824 markInstructionForDeletion(L); 1825 ++NumGVNLoad; 1826 return true; 1827 } 1828 1829 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1830 Value *AvailableVal = DepLI; 1831 1832 // The loads are of a must-aliased pointer, but they may not actually have 1833 // the same type. See if we know how to reuse the previously loaded value 1834 // (depending on its type). 1835 if (DepLI->getType() != L->getType()) { 1836 if (TD) { 1837 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1838 L, *TD); 1839 if (AvailableVal == 0) 1840 return false; 1841 1842 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1843 << "\n" << *L << "\n\n\n"); 1844 } 1845 else 1846 return false; 1847 } 1848 1849 // Remove it! 1850 L->replaceAllUsesWith(AvailableVal); 1851 if (DepLI->getType()->isPointerTy()) 1852 MD->invalidateCachedPointerInfo(DepLI); 1853 markInstructionForDeletion(L); 1854 ++NumGVNLoad; 1855 return true; 1856 } 1857 1858 // If this load really doesn't depend on anything, then we must be loading an 1859 // undef value. This can happen when loading for a fresh allocation with no 1860 // intervening stores, for example. 1861 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1862 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1863 markInstructionForDeletion(L); 1864 ++NumGVNLoad; 1865 return true; 1866 } 1867 1868 // If this load occurs either right after a lifetime begin, 1869 // then the loaded value is undefined. 1870 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1871 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1872 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1873 markInstructionForDeletion(L); 1874 ++NumGVNLoad; 1875 return true; 1876 } 1877 } 1878 1879 return false; 1880} 1881 1882// findLeader - In order to find a leader for a given value number at a 1883// specific basic block, we first obtain the list of all Values for that number, 1884// and then scan the list to find one whose block dominates the block in 1885// question. This is fast because dominator tree queries consist of only 1886// a few comparisons of DFS numbers. 1887Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1888 LeaderTableEntry Vals = LeaderTable[num]; 1889 if (!Vals.Val) return 0; 1890 1891 Value *Val = 0; 1892 if (DT->dominates(Vals.BB, BB)) { 1893 Val = Vals.Val; 1894 if (isa<Constant>(Val)) return Val; 1895 } 1896 1897 LeaderTableEntry* Next = Vals.Next; 1898 while (Next) { 1899 if (DT->dominates(Next->BB, BB)) { 1900 if (isa<Constant>(Next->Val)) return Next->Val; 1901 if (!Val) Val = Next->Val; 1902 } 1903 1904 Next = Next->Next; 1905 } 1906 1907 return Val; 1908} 1909 1910/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1911/// use is dominated by the given basic block. Returns the number of uses that 1912/// were replaced. 1913unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1914 BasicBlock *Root) { 1915 unsigned Count = 0; 1916 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1917 UI != UE; ) { 1918 Use &U = (UI++).getUse(); 1919 if (DT->dominates(Root, cast<Instruction>(U.getUser())->getParent())) { 1920 U.set(To); 1921 ++Count; 1922 } 1923 } 1924 return Count; 1925} 1926 1927/// propagateEquality - The given values are known to be equal in every block 1928/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 1929/// 'RHS' everywhere in the scope. Returns whether a change was made. 1930bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) { 1931 if (LHS == RHS) return false; 1932 assert(LHS->getType() == RHS->getType() && "Equal but types differ!"); 1933 1934 // Don't try to propagate equalities between constants. 1935 if (isa<Constant>(LHS) && isa<Constant>(RHS)) 1936 return false; 1937 1938 // Make sure that any constants are on the right-hand side. In general the 1939 // best results are obtained by placing the longest lived value on the RHS. 1940 if (isa<Constant>(LHS)) 1941 std::swap(LHS, RHS); 1942 1943 // If neither term is constant then bail out. This is not for correctness, 1944 // it's just that the non-constant case is much less useful: it occurs just 1945 // as often as the constant case but handling it hardly ever results in an 1946 // improvement. 1947 if (!isa<Constant>(RHS)) 1948 return false; 1949 1950 // If value numbering later deduces that an instruction in the scope is equal 1951 // to 'LHS' then ensure it will be turned into 'RHS'. 1952 addToLeaderTable(VN.lookup_or_add(LHS), RHS, Root); 1953 1954 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 1955 // LHS always has at least one use that is not dominated by Root, this will 1956 // never do anything if LHS has only one use. 1957 bool Changed = false; 1958 if (!LHS->hasOneUse()) { 1959 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 1960 Changed |= NumReplacements > 0; 1961 NumGVNEqProp += NumReplacements; 1962 } 1963 1964 // Now try to deduce additional equalities from this one. For example, if the 1965 // known equality was "(A != B)" == "false" then it follows that A and B are 1966 // equal in the scope. Only boolean equalities with an explicit true or false 1967 // RHS are currently supported. 1968 if (!RHS->getType()->isIntegerTy(1)) 1969 // Not a boolean equality - bail out. 1970 return Changed; 1971 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 1972 if (!CI) 1973 // RHS neither 'true' nor 'false' - bail out. 1974 return Changed; 1975 // Whether RHS equals 'true'. Otherwise it equals 'false'. 1976 bool isKnownTrue = CI->isAllOnesValue(); 1977 bool isKnownFalse = !isKnownTrue; 1978 1979 // If "A && B" is known true then both A and B are known true. If "A || B" 1980 // is known false then both A and B are known false. 1981 Value *A, *B; 1982 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 1983 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 1984 Changed |= propagateEquality(A, RHS, Root); 1985 Changed |= propagateEquality(B, RHS, Root); 1986 return Changed; 1987 } 1988 1989 // If we are propagating an equality like "(A == B)" == "true" then also 1990 // propagate the equality A == B. 1991 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 1992 // Only equality comparisons are supported. 1993 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 1994 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) { 1995 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 1996 Changed |= propagateEquality(Op0, Op1, Root); 1997 } 1998 return Changed; 1999 } 2000 2001 return Changed; 2002} 2003 2004/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2005/// true if every path from the entry block to 'Dst' passes via this edge. In 2006/// particular 'Dst' must not be reachable via another edge from 'Src'. 2007static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst, 2008 DominatorTree *DT) { 2009 // While in theory it is interesting to consider the case in which Dst has 2010 // more than one predecessor, because Dst might be part of a loop which is 2011 // only reachable from Src, in practice it is pointless since at the time 2012 // GVN runs all such loops have preheaders, which means that Dst will have 2013 // been changed to have only one predecessor, namely Src. 2014 BasicBlock *Pred = Dst->getSinglePredecessor(); 2015 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2016 (void)Src; 2017 return Pred != 0; 2018} 2019 2020/// processInstruction - When calculating availability, handle an instruction 2021/// by inserting it into the appropriate sets 2022bool GVN::processInstruction(Instruction *I) { 2023 // Ignore dbg info intrinsics. 2024 if (isa<DbgInfoIntrinsic>(I)) 2025 return false; 2026 2027 // If the instruction can be easily simplified then do so now in preference 2028 // to value numbering it. Value numbering often exposes redundancies, for 2029 // example if it determines that %y is equal to %x then the instruction 2030 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2031 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2032 I->replaceAllUsesWith(V); 2033 if (MD && V->getType()->isPointerTy()) 2034 MD->invalidateCachedPointerInfo(V); 2035 markInstructionForDeletion(I); 2036 ++NumGVNSimpl; 2037 return true; 2038 } 2039 2040 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2041 if (processLoad(LI)) 2042 return true; 2043 2044 unsigned Num = VN.lookup_or_add(LI); 2045 addToLeaderTable(Num, LI, LI->getParent()); 2046 return false; 2047 } 2048 2049 // For conditional branches, we can perform simple conditional propagation on 2050 // the condition value itself. 2051 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2052 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2053 return false; 2054 2055 Value *BranchCond = BI->getCondition(); 2056 2057 BasicBlock *TrueSucc = BI->getSuccessor(0); 2058 BasicBlock *FalseSucc = BI->getSuccessor(1); 2059 BasicBlock *Parent = BI->getParent(); 2060 bool Changed = false; 2061 2062 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT)) 2063 Changed |= propagateEquality(BranchCond, 2064 ConstantInt::getTrue(TrueSucc->getContext()), 2065 TrueSucc); 2066 2067 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT)) 2068 Changed |= propagateEquality(BranchCond, 2069 ConstantInt::getFalse(FalseSucc->getContext()), 2070 FalseSucc); 2071 2072 return Changed; 2073 } 2074 2075 // For switches, propagate the case values into the case destinations. 2076 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2077 Value *SwitchCond = SI->getCondition(); 2078 BasicBlock *Parent = SI->getParent(); 2079 bool Changed = false; 2080 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) { 2081 BasicBlock *Dst = SI->getCaseSuccessor(i); 2082 if (isOnlyReachableViaThisEdge(Parent, Dst, DT)) 2083 Changed |= propagateEquality(SwitchCond, SI->getCaseValue(i), Dst); 2084 } 2085 return Changed; 2086 } 2087 2088 // Instructions with void type don't return a value, so there's 2089 // no point in trying to find redudancies in them. 2090 if (I->getType()->isVoidTy()) return false; 2091 2092 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2093 unsigned Num = VN.lookup_or_add(I); 2094 2095 // Allocations are always uniquely numbered, so we can save time and memory 2096 // by fast failing them. 2097 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2098 addToLeaderTable(Num, I, I->getParent()); 2099 return false; 2100 } 2101 2102 // If the number we were assigned was a brand new VN, then we don't 2103 // need to do a lookup to see if the number already exists 2104 // somewhere in the domtree: it can't! 2105 if (Num == NextNum) { 2106 addToLeaderTable(Num, I, I->getParent()); 2107 return false; 2108 } 2109 2110 // Perform fast-path value-number based elimination of values inherited from 2111 // dominators. 2112 Value *repl = findLeader(I->getParent(), Num); 2113 if (repl == 0) { 2114 // Failure, just remember this instance for future use. 2115 addToLeaderTable(Num, I, I->getParent()); 2116 return false; 2117 } 2118 2119 // Remove it! 2120 I->replaceAllUsesWith(repl); 2121 if (MD && repl->getType()->isPointerTy()) 2122 MD->invalidateCachedPointerInfo(repl); 2123 markInstructionForDeletion(I); 2124 return true; 2125} 2126 2127/// runOnFunction - This is the main transformation entry point for a function. 2128bool GVN::runOnFunction(Function& F) { 2129 if (!NoLoads) 2130 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2131 DT = &getAnalysis<DominatorTree>(); 2132 TD = getAnalysisIfAvailable<TargetData>(); 2133 TLI = &getAnalysis<TargetLibraryInfo>(); 2134 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2135 VN.setMemDep(MD); 2136 VN.setDomTree(DT); 2137 2138 bool Changed = false; 2139 bool ShouldContinue = true; 2140 2141 // Merge unconditional branches, allowing PRE to catch more 2142 // optimization opportunities. 2143 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2144 BasicBlock *BB = FI++; 2145 2146 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2147 if (removedBlock) ++NumGVNBlocks; 2148 2149 Changed |= removedBlock; 2150 } 2151 2152 unsigned Iteration = 0; 2153 while (ShouldContinue) { 2154 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2155 ShouldContinue = iterateOnFunction(F); 2156 if (splitCriticalEdges()) 2157 ShouldContinue = true; 2158 Changed |= ShouldContinue; 2159 ++Iteration; 2160 } 2161 2162 if (EnablePRE) { 2163 bool PREChanged = true; 2164 while (PREChanged) { 2165 PREChanged = performPRE(F); 2166 Changed |= PREChanged; 2167 } 2168 } 2169 // FIXME: Should perform GVN again after PRE does something. PRE can move 2170 // computations into blocks where they become fully redundant. Note that 2171 // we can't do this until PRE's critical edge splitting updates memdep. 2172 // Actually, when this happens, we should just fully integrate PRE into GVN. 2173 2174 cleanupGlobalSets(); 2175 2176 return Changed; 2177} 2178 2179 2180bool GVN::processBlock(BasicBlock *BB) { 2181 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2182 // (and incrementing BI before processing an instruction). 2183 assert(InstrsToErase.empty() && 2184 "We expect InstrsToErase to be empty across iterations"); 2185 bool ChangedFunction = false; 2186 2187 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2188 BI != BE;) { 2189 ChangedFunction |= processInstruction(BI); 2190 if (InstrsToErase.empty()) { 2191 ++BI; 2192 continue; 2193 } 2194 2195 // If we need some instructions deleted, do it now. 2196 NumGVNInstr += InstrsToErase.size(); 2197 2198 // Avoid iterator invalidation. 2199 bool AtStart = BI == BB->begin(); 2200 if (!AtStart) 2201 --BI; 2202 2203 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2204 E = InstrsToErase.end(); I != E; ++I) { 2205 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2206 if (MD) MD->removeInstruction(*I); 2207 (*I)->eraseFromParent(); 2208 DEBUG(verifyRemoved(*I)); 2209 } 2210 InstrsToErase.clear(); 2211 2212 if (AtStart) 2213 BI = BB->begin(); 2214 else 2215 ++BI; 2216 } 2217 2218 return ChangedFunction; 2219} 2220 2221/// performPRE - Perform a purely local form of PRE that looks for diamond 2222/// control flow patterns and attempts to perform simple PRE at the join point. 2223bool GVN::performPRE(Function &F) { 2224 bool Changed = false; 2225 DenseMap<BasicBlock*, Value*> predMap; 2226 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2227 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2228 BasicBlock *CurrentBlock = *DI; 2229 2230 // Nothing to PRE in the entry block. 2231 if (CurrentBlock == &F.getEntryBlock()) continue; 2232 2233 // Don't perform PRE on a landing pad. 2234 if (CurrentBlock->isLandingPad()) continue; 2235 2236 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2237 BE = CurrentBlock->end(); BI != BE; ) { 2238 Instruction *CurInst = BI++; 2239 2240 if (isa<AllocaInst>(CurInst) || 2241 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2242 CurInst->getType()->isVoidTy() || 2243 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2244 isa<DbgInfoIntrinsic>(CurInst)) 2245 continue; 2246 2247 // We don't currently value number ANY inline asm calls. 2248 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2249 if (CallI->isInlineAsm()) 2250 continue; 2251 2252 uint32_t ValNo = VN.lookup(CurInst); 2253 2254 // Look for the predecessors for PRE opportunities. We're 2255 // only trying to solve the basic diamond case, where 2256 // a value is computed in the successor and one predecessor, 2257 // but not the other. We also explicitly disallow cases 2258 // where the successor is its own predecessor, because they're 2259 // more complicated to get right. 2260 unsigned NumWith = 0; 2261 unsigned NumWithout = 0; 2262 BasicBlock *PREPred = 0; 2263 predMap.clear(); 2264 2265 for (pred_iterator PI = pred_begin(CurrentBlock), 2266 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2267 BasicBlock *P = *PI; 2268 // We're not interested in PRE where the block is its 2269 // own predecessor, or in blocks with predecessors 2270 // that are not reachable. 2271 if (P == CurrentBlock) { 2272 NumWithout = 2; 2273 break; 2274 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2275 NumWithout = 2; 2276 break; 2277 } 2278 2279 Value* predV = findLeader(P, ValNo); 2280 if (predV == 0) { 2281 PREPred = P; 2282 ++NumWithout; 2283 } else if (predV == CurInst) { 2284 NumWithout = 2; 2285 } else { 2286 predMap[P] = predV; 2287 ++NumWith; 2288 } 2289 } 2290 2291 // Don't do PRE when it might increase code size, i.e. when 2292 // we would need to insert instructions in more than one pred. 2293 if (NumWithout != 1 || NumWith == 0) 2294 continue; 2295 2296 // Don't do PRE across indirect branch. 2297 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2298 continue; 2299 2300 // We can't do PRE safely on a critical edge, so instead we schedule 2301 // the edge to be split and perform the PRE the next time we iterate 2302 // on the function. 2303 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2304 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2305 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2306 continue; 2307 } 2308 2309 // Instantiate the expression in the predecessor that lacked it. 2310 // Because we are going top-down through the block, all value numbers 2311 // will be available in the predecessor by the time we need them. Any 2312 // that weren't originally present will have been instantiated earlier 2313 // in this loop. 2314 Instruction *PREInstr = CurInst->clone(); 2315 bool success = true; 2316 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2317 Value *Op = PREInstr->getOperand(i); 2318 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2319 continue; 2320 2321 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2322 PREInstr->setOperand(i, V); 2323 } else { 2324 success = false; 2325 break; 2326 } 2327 } 2328 2329 // Fail out if we encounter an operand that is not available in 2330 // the PRE predecessor. This is typically because of loads which 2331 // are not value numbered precisely. 2332 if (!success) { 2333 delete PREInstr; 2334 DEBUG(verifyRemoved(PREInstr)); 2335 continue; 2336 } 2337 2338 PREInstr->insertBefore(PREPred->getTerminator()); 2339 PREInstr->setName(CurInst->getName() + ".pre"); 2340 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2341 predMap[PREPred] = PREInstr; 2342 VN.add(PREInstr, ValNo); 2343 ++NumGVNPRE; 2344 2345 // Update the availability map to include the new instruction. 2346 addToLeaderTable(ValNo, PREInstr, PREPred); 2347 2348 // Create a PHI to make the value available in this block. 2349 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2350 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2351 CurInst->getName() + ".pre-phi", 2352 CurrentBlock->begin()); 2353 for (pred_iterator PI = PB; PI != PE; ++PI) { 2354 BasicBlock *P = *PI; 2355 Phi->addIncoming(predMap[P], P); 2356 } 2357 2358 VN.add(Phi, ValNo); 2359 addToLeaderTable(ValNo, Phi, CurrentBlock); 2360 Phi->setDebugLoc(CurInst->getDebugLoc()); 2361 CurInst->replaceAllUsesWith(Phi); 2362 if (Phi->getType()->isPointerTy()) { 2363 // Because we have added a PHI-use of the pointer value, it has now 2364 // "escaped" from alias analysis' perspective. We need to inform 2365 // AA of this. 2366 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2367 ++ii) { 2368 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2369 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2370 } 2371 2372 if (MD) 2373 MD->invalidateCachedPointerInfo(Phi); 2374 } 2375 VN.erase(CurInst); 2376 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2377 2378 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2379 if (MD) MD->removeInstruction(CurInst); 2380 CurInst->eraseFromParent(); 2381 DEBUG(verifyRemoved(CurInst)); 2382 Changed = true; 2383 } 2384 } 2385 2386 if (splitCriticalEdges()) 2387 Changed = true; 2388 2389 return Changed; 2390} 2391 2392/// splitCriticalEdges - Split critical edges found during the previous 2393/// iteration that may enable further optimization. 2394bool GVN::splitCriticalEdges() { 2395 if (toSplit.empty()) 2396 return false; 2397 do { 2398 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2399 SplitCriticalEdge(Edge.first, Edge.second, this); 2400 } while (!toSplit.empty()); 2401 if (MD) MD->invalidateCachedPredecessors(); 2402 return true; 2403} 2404 2405/// iterateOnFunction - Executes one iteration of GVN 2406bool GVN::iterateOnFunction(Function &F) { 2407 cleanupGlobalSets(); 2408 2409 // Top-down walk of the dominator tree 2410 bool Changed = false; 2411#if 0 2412 // Needed for value numbering with phi construction to work. 2413 ReversePostOrderTraversal<Function*> RPOT(&F); 2414 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2415 RE = RPOT.end(); RI != RE; ++RI) 2416 Changed |= processBlock(*RI); 2417#else 2418 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2419 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2420 Changed |= processBlock(DI->getBlock()); 2421#endif 2422 2423 return Changed; 2424} 2425 2426void GVN::cleanupGlobalSets() { 2427 VN.clear(); 2428 LeaderTable.clear(); 2429 TableAllocator.Reset(); 2430} 2431 2432/// verifyRemoved - Verify that the specified instruction does not occur in our 2433/// internal data structures. 2434void GVN::verifyRemoved(const Instruction *Inst) const { 2435 VN.verifyRemoved(Inst); 2436 2437 // Walk through the value number scope to make sure the instruction isn't 2438 // ferreted away in it. 2439 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2440 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2441 const LeaderTableEntry *Node = &I->second; 2442 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2443 2444 while (Node->Next) { 2445 Node = Node->Next; 2446 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2447 } 2448 } 2449} 2450