GVN.cpp revision 4b7b3a7c19db3d0da12a44a0849730296c26fbf9
1//===- GVN.cpp - Eliminate redundant values and loads ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs global value numbering to eliminate fully redundant 11// instructions. It also performs simple dead load elimination. 12// 13// Note that this pass does the value numbering itself; it does not use the 14// ValueNumbering analysis passes. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "gvn" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/ADT/DenseMap.h" 21#include "llvm/ADT/DepthFirstIterator.h" 22#include "llvm/ADT/Hashing.h" 23#include "llvm/ADT/SmallPtrSet.h" 24#include "llvm/ADT/Statistic.h" 25#include "llvm/Analysis/AliasAnalysis.h" 26#include "llvm/Analysis/ConstantFolding.h" 27#include "llvm/Analysis/Dominators.h" 28#include "llvm/Analysis/InstructionSimplify.h" 29#include "llvm/Analysis/Loads.h" 30#include "llvm/Analysis/MemoryBuiltins.h" 31#include "llvm/Analysis/MemoryDependenceAnalysis.h" 32#include "llvm/Analysis/PHITransAddr.h" 33#include "llvm/Analysis/ValueTracking.h" 34#include "llvm/Assembly/Writer.h" 35#include "llvm/IR/DataLayout.h" 36#include "llvm/IR/GlobalVariable.h" 37#include "llvm/IR/IRBuilder.h" 38#include "llvm/IR/IntrinsicInst.h" 39#include "llvm/IR/LLVMContext.h" 40#include "llvm/IR/Metadata.h" 41#include "llvm/Support/Allocator.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/PatternMatch.h" 45#include "llvm/Target/TargetLibraryInfo.h" 46#include "llvm/Transforms/Utils/BasicBlockUtils.h" 47#include "llvm/Transforms/Utils/SSAUpdater.h" 48#include <vector> 49using namespace llvm; 50using namespace PatternMatch; 51 52STATISTIC(NumGVNInstr, "Number of instructions deleted"); 53STATISTIC(NumGVNLoad, "Number of loads deleted"); 54STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 55STATISTIC(NumGVNBlocks, "Number of blocks merged"); 56STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 57STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 58STATISTIC(NumPRELoad, "Number of loads PRE'd"); 59 60static cl::opt<bool> EnablePRE("enable-pre", 61 cl::init(true), cl::Hidden); 62static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 63 64// Maximum allowed recursion depth. 65static cl::opt<uint32_t> 66MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 67 cl::desc("Max recurse depth (default = 1000)")); 68 69//===----------------------------------------------------------------------===// 70// ValueTable Class 71//===----------------------------------------------------------------------===// 72 73/// This class holds the mapping between values and value numbers. It is used 74/// as an efficient mechanism to determine the expression-wise equivalence of 75/// two values. 76namespace { 77 struct Expression { 78 uint32_t opcode; 79 Type *type; 80 SmallVector<uint32_t, 4> varargs; 81 82 Expression(uint32_t o = ~2U) : opcode(o) { } 83 84 bool operator==(const Expression &other) const { 85 if (opcode != other.opcode) 86 return false; 87 if (opcode == ~0U || opcode == ~1U) 88 return true; 89 if (type != other.type) 90 return false; 91 if (varargs != other.varargs) 92 return false; 93 return true; 94 } 95 96 friend hash_code hash_value(const Expression &Value) { 97 return hash_combine(Value.opcode, Value.type, 98 hash_combine_range(Value.varargs.begin(), 99 Value.varargs.end())); 100 } 101 }; 102 103 class ValueTable { 104 DenseMap<Value*, uint32_t> valueNumbering; 105 DenseMap<Expression, uint32_t> expressionNumbering; 106 AliasAnalysis *AA; 107 MemoryDependenceAnalysis *MD; 108 DominatorTree *DT; 109 110 uint32_t nextValueNumber; 111 112 Expression create_expression(Instruction* I); 113 Expression create_cmp_expression(unsigned Opcode, 114 CmpInst::Predicate Predicate, 115 Value *LHS, Value *RHS); 116 Expression create_extractvalue_expression(ExtractValueInst* EI); 117 uint32_t lookup_or_add_call(CallInst* C); 118 public: 119 ValueTable() : nextValueNumber(1) { } 120 uint32_t lookup_or_add(Value *V); 121 uint32_t lookup(Value *V) const; 122 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 123 Value *LHS, Value *RHS); 124 void add(Value *V, uint32_t num); 125 void clear(); 126 void erase(Value *v); 127 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 128 AliasAnalysis *getAliasAnalysis() const { return AA; } 129 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 130 void setDomTree(DominatorTree* D) { DT = D; } 131 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 132 void verifyRemoved(const Value *) const; 133 }; 134} 135 136namespace llvm { 137template <> struct DenseMapInfo<Expression> { 138 static inline Expression getEmptyKey() { 139 return ~0U; 140 } 141 142 static inline Expression getTombstoneKey() { 143 return ~1U; 144 } 145 146 static unsigned getHashValue(const Expression e) { 147 using llvm::hash_value; 148 return static_cast<unsigned>(hash_value(e)); 149 } 150 static bool isEqual(const Expression &LHS, const Expression &RHS) { 151 return LHS == RHS; 152 } 153}; 154 155} 156 157//===----------------------------------------------------------------------===// 158// ValueTable Internal Functions 159//===----------------------------------------------------------------------===// 160 161Expression ValueTable::create_expression(Instruction *I) { 162 Expression e; 163 e.type = I->getType(); 164 e.opcode = I->getOpcode(); 165 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 166 OI != OE; ++OI) 167 e.varargs.push_back(lookup_or_add(*OI)); 168 if (I->isCommutative()) { 169 // Ensure that commutative instructions that only differ by a permutation 170 // of their operands get the same value number by sorting the operand value 171 // numbers. Since all commutative instructions have two operands it is more 172 // efficient to sort by hand rather than using, say, std::sort. 173 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 174 if (e.varargs[0] > e.varargs[1]) 175 std::swap(e.varargs[0], e.varargs[1]); 176 } 177 178 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 179 // Sort the operand value numbers so x<y and y>x get the same value number. 180 CmpInst::Predicate Predicate = C->getPredicate(); 181 if (e.varargs[0] > e.varargs[1]) { 182 std::swap(e.varargs[0], e.varargs[1]); 183 Predicate = CmpInst::getSwappedPredicate(Predicate); 184 } 185 e.opcode = (C->getOpcode() << 8) | Predicate; 186 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 187 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 188 II != IE; ++II) 189 e.varargs.push_back(*II); 190 } 191 192 return e; 193} 194 195Expression ValueTable::create_cmp_expression(unsigned Opcode, 196 CmpInst::Predicate Predicate, 197 Value *LHS, Value *RHS) { 198 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 199 "Not a comparison!"); 200 Expression e; 201 e.type = CmpInst::makeCmpResultType(LHS->getType()); 202 e.varargs.push_back(lookup_or_add(LHS)); 203 e.varargs.push_back(lookup_or_add(RHS)); 204 205 // Sort the operand value numbers so x<y and y>x get the same value number. 206 if (e.varargs[0] > e.varargs[1]) { 207 std::swap(e.varargs[0], e.varargs[1]); 208 Predicate = CmpInst::getSwappedPredicate(Predicate); 209 } 210 e.opcode = (Opcode << 8) | Predicate; 211 return e; 212} 213 214Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 215 assert(EI != 0 && "Not an ExtractValueInst?"); 216 Expression e; 217 e.type = EI->getType(); 218 e.opcode = 0; 219 220 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 221 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 222 // EI might be an extract from one of our recognised intrinsics. If it 223 // is we'll synthesize a semantically equivalent expression instead on 224 // an extract value expression. 225 switch (I->getIntrinsicID()) { 226 case Intrinsic::sadd_with_overflow: 227 case Intrinsic::uadd_with_overflow: 228 e.opcode = Instruction::Add; 229 break; 230 case Intrinsic::ssub_with_overflow: 231 case Intrinsic::usub_with_overflow: 232 e.opcode = Instruction::Sub; 233 break; 234 case Intrinsic::smul_with_overflow: 235 case Intrinsic::umul_with_overflow: 236 e.opcode = Instruction::Mul; 237 break; 238 default: 239 break; 240 } 241 242 if (e.opcode != 0) { 243 // Intrinsic recognized. Grab its args to finish building the expression. 244 assert(I->getNumArgOperands() == 2 && 245 "Expect two args for recognised intrinsics."); 246 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 247 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 248 return e; 249 } 250 } 251 252 // Not a recognised intrinsic. Fall back to producing an extract value 253 // expression. 254 e.opcode = EI->getOpcode(); 255 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 256 OI != OE; ++OI) 257 e.varargs.push_back(lookup_or_add(*OI)); 258 259 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 260 II != IE; ++II) 261 e.varargs.push_back(*II); 262 263 return e; 264} 265 266//===----------------------------------------------------------------------===// 267// ValueTable External Functions 268//===----------------------------------------------------------------------===// 269 270/// add - Insert a value into the table with a specified value number. 271void ValueTable::add(Value *V, uint32_t num) { 272 valueNumbering.insert(std::make_pair(V, num)); 273} 274 275uint32_t ValueTable::lookup_or_add_call(CallInst *C) { 276 if (AA->doesNotAccessMemory(C)) { 277 Expression exp = create_expression(C); 278 uint32_t &e = expressionNumbering[exp]; 279 if (!e) e = nextValueNumber++; 280 valueNumbering[C] = e; 281 return e; 282 } else if (AA->onlyReadsMemory(C)) { 283 Expression exp = create_expression(C); 284 uint32_t &e = expressionNumbering[exp]; 285 if (!e) { 286 e = nextValueNumber++; 287 valueNumbering[C] = e; 288 return e; 289 } 290 if (!MD) { 291 e = nextValueNumber++; 292 valueNumbering[C] = e; 293 return e; 294 } 295 296 MemDepResult local_dep = MD->getDependency(C); 297 298 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 299 valueNumbering[C] = nextValueNumber; 300 return nextValueNumber++; 301 } 302 303 if (local_dep.isDef()) { 304 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 305 306 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 307 valueNumbering[C] = nextValueNumber; 308 return nextValueNumber++; 309 } 310 311 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 312 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 313 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 314 if (c_vn != cd_vn) { 315 valueNumbering[C] = nextValueNumber; 316 return nextValueNumber++; 317 } 318 } 319 320 uint32_t v = lookup_or_add(local_cdep); 321 valueNumbering[C] = v; 322 return v; 323 } 324 325 // Non-local case. 326 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 327 MD->getNonLocalCallDependency(CallSite(C)); 328 // FIXME: Move the checking logic to MemDep! 329 CallInst* cdep = 0; 330 331 // Check to see if we have a single dominating call instruction that is 332 // identical to C. 333 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 334 const NonLocalDepEntry *I = &deps[i]; 335 if (I->getResult().isNonLocal()) 336 continue; 337 338 // We don't handle non-definitions. If we already have a call, reject 339 // instruction dependencies. 340 if (!I->getResult().isDef() || cdep != 0) { 341 cdep = 0; 342 break; 343 } 344 345 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 346 // FIXME: All duplicated with non-local case. 347 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 348 cdep = NonLocalDepCall; 349 continue; 350 } 351 352 cdep = 0; 353 break; 354 } 355 356 if (!cdep) { 357 valueNumbering[C] = nextValueNumber; 358 return nextValueNumber++; 359 } 360 361 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 362 valueNumbering[C] = nextValueNumber; 363 return nextValueNumber++; 364 } 365 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 366 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 367 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 368 if (c_vn != cd_vn) { 369 valueNumbering[C] = nextValueNumber; 370 return nextValueNumber++; 371 } 372 } 373 374 uint32_t v = lookup_or_add(cdep); 375 valueNumbering[C] = v; 376 return v; 377 378 } else { 379 valueNumbering[C] = nextValueNumber; 380 return nextValueNumber++; 381 } 382} 383 384/// lookup_or_add - Returns the value number for the specified value, assigning 385/// it a new number if it did not have one before. 386uint32_t ValueTable::lookup_or_add(Value *V) { 387 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 388 if (VI != valueNumbering.end()) 389 return VI->second; 390 391 if (!isa<Instruction>(V)) { 392 valueNumbering[V] = nextValueNumber; 393 return nextValueNumber++; 394 } 395 396 Instruction* I = cast<Instruction>(V); 397 Expression exp; 398 switch (I->getOpcode()) { 399 case Instruction::Call: 400 return lookup_or_add_call(cast<CallInst>(I)); 401 case Instruction::Add: 402 case Instruction::FAdd: 403 case Instruction::Sub: 404 case Instruction::FSub: 405 case Instruction::Mul: 406 case Instruction::FMul: 407 case Instruction::UDiv: 408 case Instruction::SDiv: 409 case Instruction::FDiv: 410 case Instruction::URem: 411 case Instruction::SRem: 412 case Instruction::FRem: 413 case Instruction::Shl: 414 case Instruction::LShr: 415 case Instruction::AShr: 416 case Instruction::And: 417 case Instruction::Or: 418 case Instruction::Xor: 419 case Instruction::ICmp: 420 case Instruction::FCmp: 421 case Instruction::Trunc: 422 case Instruction::ZExt: 423 case Instruction::SExt: 424 case Instruction::FPToUI: 425 case Instruction::FPToSI: 426 case Instruction::UIToFP: 427 case Instruction::SIToFP: 428 case Instruction::FPTrunc: 429 case Instruction::FPExt: 430 case Instruction::PtrToInt: 431 case Instruction::IntToPtr: 432 case Instruction::BitCast: 433 case Instruction::Select: 434 case Instruction::ExtractElement: 435 case Instruction::InsertElement: 436 case Instruction::ShuffleVector: 437 case Instruction::InsertValue: 438 case Instruction::GetElementPtr: 439 exp = create_expression(I); 440 break; 441 case Instruction::ExtractValue: 442 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 443 break; 444 default: 445 valueNumbering[V] = nextValueNumber; 446 return nextValueNumber++; 447 } 448 449 uint32_t& e = expressionNumbering[exp]; 450 if (!e) e = nextValueNumber++; 451 valueNumbering[V] = e; 452 return e; 453} 454 455/// lookup - Returns the value number of the specified value. Fails if 456/// the value has not yet been numbered. 457uint32_t ValueTable::lookup(Value *V) const { 458 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 459 assert(VI != valueNumbering.end() && "Value not numbered?"); 460 return VI->second; 461} 462 463/// lookup_or_add_cmp - Returns the value number of the given comparison, 464/// assigning it a new number if it did not have one before. Useful when 465/// we deduced the result of a comparison, but don't immediately have an 466/// instruction realizing that comparison to hand. 467uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 468 CmpInst::Predicate Predicate, 469 Value *LHS, Value *RHS) { 470 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 471 uint32_t& e = expressionNumbering[exp]; 472 if (!e) e = nextValueNumber++; 473 return e; 474} 475 476/// clear - Remove all entries from the ValueTable. 477void ValueTable::clear() { 478 valueNumbering.clear(); 479 expressionNumbering.clear(); 480 nextValueNumber = 1; 481} 482 483/// erase - Remove a value from the value numbering. 484void ValueTable::erase(Value *V) { 485 valueNumbering.erase(V); 486} 487 488/// verifyRemoved - Verify that the value is removed from all internal data 489/// structures. 490void ValueTable::verifyRemoved(const Value *V) const { 491 for (DenseMap<Value*, uint32_t>::const_iterator 492 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 493 assert(I->first != V && "Inst still occurs in value numbering map!"); 494 } 495} 496 497//===----------------------------------------------------------------------===// 498// GVN Pass 499//===----------------------------------------------------------------------===// 500 501namespace { 502 class GVN; 503 struct AvailableValueInBlock { 504 /// BB - The basic block in question. 505 BasicBlock *BB; 506 enum ValType { 507 SimpleVal, // A simple offsetted value that is accessed. 508 LoadVal, // A value produced by a load. 509 MemIntrin // A memory intrinsic which is loaded from. 510 }; 511 512 /// V - The value that is live out of the block. 513 PointerIntPair<Value *, 2, ValType> Val; 514 515 /// Offset - The byte offset in Val that is interesting for the load query. 516 unsigned Offset; 517 518 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 519 unsigned Offset = 0) { 520 AvailableValueInBlock Res; 521 Res.BB = BB; 522 Res.Val.setPointer(V); 523 Res.Val.setInt(SimpleVal); 524 Res.Offset = Offset; 525 return Res; 526 } 527 528 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 529 unsigned Offset = 0) { 530 AvailableValueInBlock Res; 531 Res.BB = BB; 532 Res.Val.setPointer(MI); 533 Res.Val.setInt(MemIntrin); 534 Res.Offset = Offset; 535 return Res; 536 } 537 538 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 539 unsigned Offset = 0) { 540 AvailableValueInBlock Res; 541 Res.BB = BB; 542 Res.Val.setPointer(LI); 543 Res.Val.setInt(LoadVal); 544 Res.Offset = Offset; 545 return Res; 546 } 547 548 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 549 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 550 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 551 552 Value *getSimpleValue() const { 553 assert(isSimpleValue() && "Wrong accessor"); 554 return Val.getPointer(); 555 } 556 557 LoadInst *getCoercedLoadValue() const { 558 assert(isCoercedLoadValue() && "Wrong accessor"); 559 return cast<LoadInst>(Val.getPointer()); 560 } 561 562 MemIntrinsic *getMemIntrinValue() const { 563 assert(isMemIntrinValue() && "Wrong accessor"); 564 return cast<MemIntrinsic>(Val.getPointer()); 565 } 566 567 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 568 /// defined here to the specified type. This handles various coercion cases. 569 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const; 570 }; 571 572 class GVN : public FunctionPass { 573 bool NoLoads; 574 MemoryDependenceAnalysis *MD; 575 DominatorTree *DT; 576 const DataLayout *TD; 577 const TargetLibraryInfo *TLI; 578 579 ValueTable VN; 580 581 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 582 /// have that value number. Use findLeader to query it. 583 struct LeaderTableEntry { 584 Value *Val; 585 const BasicBlock *BB; 586 LeaderTableEntry *Next; 587 }; 588 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 589 BumpPtrAllocator TableAllocator; 590 591 SmallVector<Instruction*, 8> InstrsToErase; 592 593 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect; 594 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect; 595 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect; 596 597 public: 598 static char ID; // Pass identification, replacement for typeid 599 explicit GVN(bool noloads = false) 600 : FunctionPass(ID), NoLoads(noloads), MD(0) { 601 initializeGVNPass(*PassRegistry::getPassRegistry()); 602 } 603 604 bool runOnFunction(Function &F); 605 606 /// markInstructionForDeletion - This removes the specified instruction from 607 /// our various maps and marks it for deletion. 608 void markInstructionForDeletion(Instruction *I) { 609 VN.erase(I); 610 InstrsToErase.push_back(I); 611 } 612 613 const DataLayout *getDataLayout() const { return TD; } 614 DominatorTree &getDominatorTree() const { return *DT; } 615 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 616 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 617 private: 618 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 619 /// its value number. 620 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 621 LeaderTableEntry &Curr = LeaderTable[N]; 622 if (!Curr.Val) { 623 Curr.Val = V; 624 Curr.BB = BB; 625 return; 626 } 627 628 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 629 Node->Val = V; 630 Node->BB = BB; 631 Node->Next = Curr.Next; 632 Curr.Next = Node; 633 } 634 635 /// removeFromLeaderTable - Scan the list of values corresponding to a given 636 /// value number, and remove the given instruction if encountered. 637 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 638 LeaderTableEntry* Prev = 0; 639 LeaderTableEntry* Curr = &LeaderTable[N]; 640 641 while (Curr->Val != I || Curr->BB != BB) { 642 Prev = Curr; 643 Curr = Curr->Next; 644 } 645 646 if (Prev) { 647 Prev->Next = Curr->Next; 648 } else { 649 if (!Curr->Next) { 650 Curr->Val = 0; 651 Curr->BB = 0; 652 } else { 653 LeaderTableEntry* Next = Curr->Next; 654 Curr->Val = Next->Val; 655 Curr->BB = Next->BB; 656 Curr->Next = Next->Next; 657 } 658 } 659 } 660 661 // List of critical edges to be split between iterations. 662 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 663 664 // This transformation requires dominator postdominator info 665 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 666 AU.addRequired<DominatorTree>(); 667 AU.addRequired<TargetLibraryInfo>(); 668 if (!NoLoads) 669 AU.addRequired<MemoryDependenceAnalysis>(); 670 AU.addRequired<AliasAnalysis>(); 671 672 AU.addPreserved<DominatorTree>(); 673 AU.addPreserved<AliasAnalysis>(); 674 } 675 676 677 // Helper fuctions of redundant load elimination 678 bool processLoad(LoadInst *L); 679 bool processNonLocalLoad(LoadInst *L); 680 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 681 AvailValInBlkVect &ValuesPerBlock, 682 UnavailBlkVect &UnavailableBlocks); 683 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 684 UnavailBlkVect &UnavailableBlocks); 685 686 // Other helper routines 687 bool processInstruction(Instruction *I); 688 bool processBlock(BasicBlock *BB); 689 void dump(DenseMap<uint32_t, Value*> &d); 690 bool iterateOnFunction(Function &F); 691 bool performPRE(Function &F); 692 Value *findLeader(const BasicBlock *BB, uint32_t num); 693 void cleanupGlobalSets(); 694 void verifyRemoved(const Instruction *I) const; 695 bool splitCriticalEdges(); 696 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ); 697 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 698 const BasicBlockEdge &Root); 699 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 700 }; 701 702 char GVN::ID = 0; 703} 704 705// createGVNPass - The public interface to this file... 706FunctionPass *llvm::createGVNPass(bool NoLoads) { 707 return new GVN(NoLoads); 708} 709 710INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 711INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 712INITIALIZE_PASS_DEPENDENCY(DominatorTree) 713INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 714INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 715INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 716 717#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 718void GVN::dump(DenseMap<uint32_t, Value*>& d) { 719 errs() << "{\n"; 720 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 721 E = d.end(); I != E; ++I) { 722 errs() << I->first << "\n"; 723 I->second->dump(); 724 } 725 errs() << "}\n"; 726} 727#endif 728 729/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 730/// we're analyzing is fully available in the specified block. As we go, keep 731/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 732/// map is actually a tri-state map with the following values: 733/// 0) we know the block *is not* fully available. 734/// 1) we know the block *is* fully available. 735/// 2) we do not know whether the block is fully available or not, but we are 736/// currently speculating that it will be. 737/// 3) we are speculating for this block and have used that to speculate for 738/// other blocks. 739static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 740 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 741 uint32_t RecurseDepth) { 742 if (RecurseDepth > MaxRecurseDepth) 743 return false; 744 745 // Optimistically assume that the block is fully available and check to see 746 // if we already know about this block in one lookup. 747 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 748 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 749 750 // If the entry already existed for this block, return the precomputed value. 751 if (!IV.second) { 752 // If this is a speculative "available" value, mark it as being used for 753 // speculation of other blocks. 754 if (IV.first->second == 2) 755 IV.first->second = 3; 756 return IV.first->second != 0; 757 } 758 759 // Otherwise, see if it is fully available in all predecessors. 760 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 761 762 // If this block has no predecessors, it isn't live-in here. 763 if (PI == PE) 764 goto SpeculationFailure; 765 766 for (; PI != PE; ++PI) 767 // If the value isn't fully available in one of our predecessors, then it 768 // isn't fully available in this block either. Undo our previous 769 // optimistic assumption and bail out. 770 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 771 goto SpeculationFailure; 772 773 return true; 774 775// SpeculationFailure - If we get here, we found out that this is not, after 776// all, a fully-available block. We have a problem if we speculated on this and 777// used the speculation to mark other blocks as available. 778SpeculationFailure: 779 char &BBVal = FullyAvailableBlocks[BB]; 780 781 // If we didn't speculate on this, just return with it set to false. 782 if (BBVal == 2) { 783 BBVal = 0; 784 return false; 785 } 786 787 // If we did speculate on this value, we could have blocks set to 1 that are 788 // incorrect. Walk the (transitive) successors of this block and mark them as 789 // 0 if set to one. 790 SmallVector<BasicBlock*, 32> BBWorklist; 791 BBWorklist.push_back(BB); 792 793 do { 794 BasicBlock *Entry = BBWorklist.pop_back_val(); 795 // Note that this sets blocks to 0 (unavailable) if they happen to not 796 // already be in FullyAvailableBlocks. This is safe. 797 char &EntryVal = FullyAvailableBlocks[Entry]; 798 if (EntryVal == 0) continue; // Already unavailable. 799 800 // Mark as unavailable. 801 EntryVal = 0; 802 803 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 804 BBWorklist.push_back(*I); 805 } while (!BBWorklist.empty()); 806 807 return false; 808} 809 810 811/// CanCoerceMustAliasedValueToLoad - Return true if 812/// CoerceAvailableValueToLoadType will succeed. 813static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 814 Type *LoadTy, 815 const DataLayout &TD) { 816 // If the loaded or stored value is an first class array or struct, don't try 817 // to transform them. We need to be able to bitcast to integer. 818 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 819 StoredVal->getType()->isStructTy() || 820 StoredVal->getType()->isArrayTy()) 821 return false; 822 823 // The store has to be at least as big as the load. 824 if (TD.getTypeSizeInBits(StoredVal->getType()) < 825 TD.getTypeSizeInBits(LoadTy)) 826 return false; 827 828 return true; 829} 830 831/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 832/// then a load from a must-aliased pointer of a different type, try to coerce 833/// the stored value. LoadedTy is the type of the load we want to replace and 834/// InsertPt is the place to insert new instructions. 835/// 836/// If we can't do it, return null. 837static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 838 Type *LoadedTy, 839 Instruction *InsertPt, 840 const DataLayout &TD) { 841 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 842 return 0; 843 844 // If this is already the right type, just return it. 845 Type *StoredValTy = StoredVal->getType(); 846 847 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 848 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 849 850 // If the store and reload are the same size, we can always reuse it. 851 if (StoreSize == LoadSize) { 852 // Pointer to Pointer -> use bitcast. 853 if (StoredValTy->getScalarType()->isPointerTy() && 854 LoadedTy->getScalarType()->isPointerTy()) 855 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 856 857 // Convert source pointers to integers, which can be bitcast. 858 if (StoredValTy->getScalarType()->isPointerTy()) { 859 StoredValTy = TD.getIntPtrType(StoredValTy); 860 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 861 } 862 863 Type *TypeToCastTo = LoadedTy; 864 if (TypeToCastTo->getScalarType()->isPointerTy()) 865 TypeToCastTo = TD.getIntPtrType(TypeToCastTo); 866 867 if (StoredValTy != TypeToCastTo) 868 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 869 870 // Cast to pointer if the load needs a pointer type. 871 if (LoadedTy->getScalarType()->isPointerTy()) 872 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 873 874 return StoredVal; 875 } 876 877 // If the loaded value is smaller than the available value, then we can 878 // extract out a piece from it. If the available value is too small, then we 879 // can't do anything. 880 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 881 882 // Convert source pointers to integers, which can be manipulated. 883 if (StoredValTy->getScalarType()->isPointerTy()) { 884 StoredValTy = TD.getIntPtrType(StoredValTy); 885 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 886 } 887 888 // Convert vectors and fp to integer, which can be manipulated. 889 if (!StoredValTy->isIntegerTy()) { 890 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 891 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 892 } 893 894 // If this is a big-endian system, we need to shift the value down to the low 895 // bits so that a truncate will work. 896 if (TD.isBigEndian()) { 897 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 898 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 899 } 900 901 // Truncate the integer to the right size now. 902 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 903 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 904 905 if (LoadedTy == NewIntTy) 906 return StoredVal; 907 908 // If the result is a pointer, inttoptr. 909 if (LoadedTy->getScalarType()->isPointerTy()) 910 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 911 912 // Otherwise, bitcast. 913 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 914} 915 916/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 917/// memdep query of a load that ends up being a clobbering memory write (store, 918/// memset, memcpy, memmove). This means that the write *may* provide bits used 919/// by the load but we can't be sure because the pointers don't mustalias. 920/// 921/// Check this case to see if there is anything more we can do before we give 922/// up. This returns -1 if we have to give up, or a byte number in the stored 923/// value of the piece that feeds the load. 924static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 925 Value *WritePtr, 926 uint64_t WriteSizeInBits, 927 const DataLayout &TD) { 928 // If the loaded or stored value is a first class array or struct, don't try 929 // to transform them. We need to be able to bitcast to integer. 930 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 931 return -1; 932 933 int64_t StoreOffset = 0, LoadOffset = 0; 934 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD); 935 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD); 936 if (StoreBase != LoadBase) 937 return -1; 938 939 // If the load and store are to the exact same address, they should have been 940 // a must alias. AA must have gotten confused. 941 // FIXME: Study to see if/when this happens. One case is forwarding a memset 942 // to a load from the base of the memset. 943#if 0 944 if (LoadOffset == StoreOffset) { 945 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 946 << "Base = " << *StoreBase << "\n" 947 << "Store Ptr = " << *WritePtr << "\n" 948 << "Store Offs = " << StoreOffset << "\n" 949 << "Load Ptr = " << *LoadPtr << "\n"; 950 abort(); 951 } 952#endif 953 954 // If the load and store don't overlap at all, the store doesn't provide 955 // anything to the load. In this case, they really don't alias at all, AA 956 // must have gotten confused. 957 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 958 959 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 960 return -1; 961 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 962 LoadSize >>= 3; 963 964 965 bool isAAFailure = false; 966 if (StoreOffset < LoadOffset) 967 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 968 else 969 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 970 971 if (isAAFailure) { 972#if 0 973 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 974 << "Base = " << *StoreBase << "\n" 975 << "Store Ptr = " << *WritePtr << "\n" 976 << "Store Offs = " << StoreOffset << "\n" 977 << "Load Ptr = " << *LoadPtr << "\n"; 978 abort(); 979#endif 980 return -1; 981 } 982 983 // If the Load isn't completely contained within the stored bits, we don't 984 // have all the bits to feed it. We could do something crazy in the future 985 // (issue a smaller load then merge the bits in) but this seems unlikely to be 986 // valuable. 987 if (StoreOffset > LoadOffset || 988 StoreOffset+StoreSize < LoadOffset+LoadSize) 989 return -1; 990 991 // Okay, we can do this transformation. Return the number of bytes into the 992 // store that the load is. 993 return LoadOffset-StoreOffset; 994} 995 996/// AnalyzeLoadFromClobberingStore - This function is called when we have a 997/// memdep query of a load that ends up being a clobbering store. 998static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 999 StoreInst *DepSI, 1000 const DataLayout &TD) { 1001 // Cannot handle reading from store of first-class aggregate yet. 1002 if (DepSI->getValueOperand()->getType()->isStructTy() || 1003 DepSI->getValueOperand()->getType()->isArrayTy()) 1004 return -1; 1005 1006 Value *StorePtr = DepSI->getPointerOperand(); 1007 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1008 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1009 StorePtr, StoreSize, TD); 1010} 1011 1012/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 1013/// memdep query of a load that ends up being clobbered by another load. See if 1014/// the other load can feed into the second load. 1015static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 1016 LoadInst *DepLI, const DataLayout &TD){ 1017 // Cannot handle reading from store of first-class aggregate yet. 1018 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 1019 return -1; 1020 1021 Value *DepPtr = DepLI->getPointerOperand(); 1022 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 1023 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 1024 if (R != -1) return R; 1025 1026 // If we have a load/load clobber an DepLI can be widened to cover this load, 1027 // then we should widen it! 1028 int64_t LoadOffs = 0; 1029 const Value *LoadBase = 1030 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD); 1031 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1032 1033 unsigned Size = MemoryDependenceAnalysis:: 1034 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 1035 if (Size == 0) return -1; 1036 1037 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 1038} 1039 1040 1041 1042static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 1043 MemIntrinsic *MI, 1044 const DataLayout &TD) { 1045 // If the mem operation is a non-constant size, we can't handle it. 1046 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1047 if (SizeCst == 0) return -1; 1048 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1049 1050 // If this is memset, we just need to see if the offset is valid in the size 1051 // of the memset.. 1052 if (MI->getIntrinsicID() == Intrinsic::memset) 1053 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1054 MemSizeInBits, TD); 1055 1056 // If we have a memcpy/memmove, the only case we can handle is if this is a 1057 // copy from constant memory. In that case, we can read directly from the 1058 // constant memory. 1059 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1060 1061 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1062 if (Src == 0) return -1; 1063 1064 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 1065 if (GV == 0 || !GV->isConstant()) return -1; 1066 1067 // See if the access is within the bounds of the transfer. 1068 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1069 MI->getDest(), MemSizeInBits, TD); 1070 if (Offset == -1) 1071 return Offset; 1072 1073 // Otherwise, see if we can constant fold a load from the constant with the 1074 // offset applied as appropriate. 1075 Src = ConstantExpr::getBitCast(Src, 1076 llvm::Type::getInt8PtrTy(Src->getContext())); 1077 Constant *OffsetCst = 1078 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1079 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1080 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1081 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1082 return Offset; 1083 return -1; 1084} 1085 1086 1087/// GetStoreValueForLoad - This function is called when we have a 1088/// memdep query of a load that ends up being a clobbering store. This means 1089/// that the store provides bits used by the load but we the pointers don't 1090/// mustalias. Check this case to see if there is anything more we can do 1091/// before we give up. 1092static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1093 Type *LoadTy, 1094 Instruction *InsertPt, const DataLayout &TD){ 1095 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1096 1097 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1098 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1099 1100 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1101 1102 // Compute which bits of the stored value are being used by the load. Convert 1103 // to an integer type to start with. 1104 if (SrcVal->getType()->getScalarType()->isPointerTy()) 1105 SrcVal = Builder.CreatePtrToInt(SrcVal, 1106 TD.getIntPtrType(SrcVal->getType())); 1107 if (!SrcVal->getType()->isIntegerTy()) 1108 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1109 1110 // Shift the bits to the least significant depending on endianness. 1111 unsigned ShiftAmt; 1112 if (TD.isLittleEndian()) 1113 ShiftAmt = Offset*8; 1114 else 1115 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1116 1117 if (ShiftAmt) 1118 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1119 1120 if (LoadSize != StoreSize) 1121 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1122 1123 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1124} 1125 1126/// GetLoadValueForLoad - This function is called when we have a 1127/// memdep query of a load that ends up being a clobbering load. This means 1128/// that the load *may* provide bits used by the load but we can't be sure 1129/// because the pointers don't mustalias. Check this case to see if there is 1130/// anything more we can do before we give up. 1131static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1132 Type *LoadTy, Instruction *InsertPt, 1133 GVN &gvn) { 1134 const DataLayout &TD = *gvn.getDataLayout(); 1135 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1136 // widen SrcVal out to a larger load. 1137 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1138 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1139 if (Offset+LoadSize > SrcValSize) { 1140 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1141 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1142 // If we have a load/load clobber an DepLI can be widened to cover this 1143 // load, then we should widen it to the next power of 2 size big enough! 1144 unsigned NewLoadSize = Offset+LoadSize; 1145 if (!isPowerOf2_32(NewLoadSize)) 1146 NewLoadSize = NextPowerOf2(NewLoadSize); 1147 1148 Value *PtrVal = SrcVal->getPointerOperand(); 1149 1150 // Insert the new load after the old load. This ensures that subsequent 1151 // memdep queries will find the new load. We can't easily remove the old 1152 // load completely because it is already in the value numbering table. 1153 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1154 Type *DestPTy = 1155 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1156 DestPTy = PointerType::get(DestPTy, 1157 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 1158 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1159 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1160 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1161 NewLoad->takeName(SrcVal); 1162 NewLoad->setAlignment(SrcVal->getAlignment()); 1163 1164 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1165 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1166 1167 // Replace uses of the original load with the wider load. On a big endian 1168 // system, we need to shift down to get the relevant bits. 1169 Value *RV = NewLoad; 1170 if (TD.isBigEndian()) 1171 RV = Builder.CreateLShr(RV, 1172 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1173 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1174 SrcVal->replaceAllUsesWith(RV); 1175 1176 // We would like to use gvn.markInstructionForDeletion here, but we can't 1177 // because the load is already memoized into the leader map table that GVN 1178 // tracks. It is potentially possible to remove the load from the table, 1179 // but then there all of the operations based on it would need to be 1180 // rehashed. Just leave the dead load around. 1181 gvn.getMemDep().removeInstruction(SrcVal); 1182 SrcVal = NewLoad; 1183 } 1184 1185 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1186} 1187 1188 1189/// GetMemInstValueForLoad - This function is called when we have a 1190/// memdep query of a load that ends up being a clobbering mem intrinsic. 1191static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1192 Type *LoadTy, Instruction *InsertPt, 1193 const DataLayout &TD){ 1194 LLVMContext &Ctx = LoadTy->getContext(); 1195 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1196 1197 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1198 1199 // We know that this method is only called when the mem transfer fully 1200 // provides the bits for the load. 1201 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1202 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1203 // independently of what the offset is. 1204 Value *Val = MSI->getValue(); 1205 if (LoadSize != 1) 1206 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1207 1208 Value *OneElt = Val; 1209 1210 // Splat the value out to the right number of bits. 1211 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1212 // If we can double the number of bytes set, do it. 1213 if (NumBytesSet*2 <= LoadSize) { 1214 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1215 Val = Builder.CreateOr(Val, ShVal); 1216 NumBytesSet <<= 1; 1217 continue; 1218 } 1219 1220 // Otherwise insert one byte at a time. 1221 Value *ShVal = Builder.CreateShl(Val, 1*8); 1222 Val = Builder.CreateOr(OneElt, ShVal); 1223 ++NumBytesSet; 1224 } 1225 1226 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1227 } 1228 1229 // Otherwise, this is a memcpy/memmove from a constant global. 1230 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1231 Constant *Src = cast<Constant>(MTI->getSource()); 1232 1233 // Otherwise, see if we can constant fold a load from the constant with the 1234 // offset applied as appropriate. 1235 Src = ConstantExpr::getBitCast(Src, 1236 llvm::Type::getInt8PtrTy(Src->getContext())); 1237 Constant *OffsetCst = 1238 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1239 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1240 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1241 return ConstantFoldLoadFromConstPtr(Src, &TD); 1242} 1243 1244 1245/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1246/// construct SSA form, allowing us to eliminate LI. This returns the value 1247/// that should be used at LI's definition site. 1248static Value *ConstructSSAForLoadSet(LoadInst *LI, 1249 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1250 GVN &gvn) { 1251 // Check for the fully redundant, dominating load case. In this case, we can 1252 // just use the dominating value directly. 1253 if (ValuesPerBlock.size() == 1 && 1254 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1255 LI->getParent())) 1256 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1257 1258 // Otherwise, we have to construct SSA form. 1259 SmallVector<PHINode*, 8> NewPHIs; 1260 SSAUpdater SSAUpdate(&NewPHIs); 1261 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1262 1263 Type *LoadTy = LI->getType(); 1264 1265 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1266 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1267 BasicBlock *BB = AV.BB; 1268 1269 if (SSAUpdate.HasValueForBlock(BB)) 1270 continue; 1271 1272 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1273 } 1274 1275 // Perform PHI construction. 1276 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1277 1278 // If new PHI nodes were created, notify alias analysis. 1279 if (V->getType()->getScalarType()->isPointerTy()) { 1280 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1281 1282 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1283 AA->copyValue(LI, NewPHIs[i]); 1284 1285 // Now that we've copied information to the new PHIs, scan through 1286 // them again and inform alias analysis that we've added potentially 1287 // escaping uses to any values that are operands to these PHIs. 1288 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1289 PHINode *P = NewPHIs[i]; 1290 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1291 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1292 AA->addEscapingUse(P->getOperandUse(jj)); 1293 } 1294 } 1295 } 1296 1297 return V; 1298} 1299 1300Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1301 Value *Res; 1302 if (isSimpleValue()) { 1303 Res = getSimpleValue(); 1304 if (Res->getType() != LoadTy) { 1305 const DataLayout *TD = gvn.getDataLayout(); 1306 assert(TD && "Need target data to handle type mismatch case"); 1307 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1308 *TD); 1309 1310 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1311 << *getSimpleValue() << '\n' 1312 << *Res << '\n' << "\n\n\n"); 1313 } 1314 } else if (isCoercedLoadValue()) { 1315 LoadInst *Load = getCoercedLoadValue(); 1316 if (Load->getType() == LoadTy && Offset == 0) { 1317 Res = Load; 1318 } else { 1319 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1320 gvn); 1321 1322 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1323 << *getCoercedLoadValue() << '\n' 1324 << *Res << '\n' << "\n\n\n"); 1325 } 1326 } else { 1327 const DataLayout *TD = gvn.getDataLayout(); 1328 assert(TD && "Need target data to handle type mismatch case"); 1329 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1330 LoadTy, BB->getTerminator(), *TD); 1331 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1332 << " " << *getMemIntrinValue() << '\n' 1333 << *Res << '\n' << "\n\n\n"); 1334 } 1335 return Res; 1336} 1337 1338static bool isLifetimeStart(const Instruction *Inst) { 1339 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1340 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1341 return false; 1342} 1343 1344void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 1345 AvailValInBlkVect &ValuesPerBlock, 1346 UnavailBlkVect &UnavailableBlocks) { 1347 1348 // Filter out useless results (non-locals, etc). Keep track of the blocks 1349 // where we have a value available in repl, also keep track of whether we see 1350 // dependencies that produce an unknown value for the load (such as a call 1351 // that could potentially clobber the load). 1352 unsigned NumDeps = Deps.size(); 1353 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1354 BasicBlock *DepBB = Deps[i].getBB(); 1355 MemDepResult DepInfo = Deps[i].getResult(); 1356 1357 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1358 UnavailableBlocks.push_back(DepBB); 1359 continue; 1360 } 1361 1362 if (DepInfo.isClobber()) { 1363 // The address being loaded in this non-local block may not be the same as 1364 // the pointer operand of the load if PHI translation occurs. Make sure 1365 // to consider the right address. 1366 Value *Address = Deps[i].getAddress(); 1367 1368 // If the dependence is to a store that writes to a superset of the bits 1369 // read by the load, we can extract the bits we need for the load from the 1370 // stored value. 1371 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1372 if (TD && Address) { 1373 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1374 DepSI, *TD); 1375 if (Offset != -1) { 1376 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1377 DepSI->getValueOperand(), 1378 Offset)); 1379 continue; 1380 } 1381 } 1382 } 1383 1384 // Check to see if we have something like this: 1385 // load i32* P 1386 // load i8* (P+1) 1387 // if we have this, replace the later with an extraction from the former. 1388 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1389 // If this is a clobber and L is the first instruction in its block, then 1390 // we have the first instruction in the entry block. 1391 if (DepLI != LI && Address && TD) { 1392 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1393 LI->getPointerOperand(), 1394 DepLI, *TD); 1395 1396 if (Offset != -1) { 1397 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1398 Offset)); 1399 continue; 1400 } 1401 } 1402 } 1403 1404 // If the clobbering value is a memset/memcpy/memmove, see if we can 1405 // forward a value on from it. 1406 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1407 if (TD && Address) { 1408 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1409 DepMI, *TD); 1410 if (Offset != -1) { 1411 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1412 Offset)); 1413 continue; 1414 } 1415 } 1416 } 1417 1418 UnavailableBlocks.push_back(DepBB); 1419 continue; 1420 } 1421 1422 // DepInfo.isDef() here 1423 1424 Instruction *DepInst = DepInfo.getInst(); 1425 1426 // Loading the allocation -> undef. 1427 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1428 // Loading immediately after lifetime begin -> undef. 1429 isLifetimeStart(DepInst)) { 1430 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1431 UndefValue::get(LI->getType()))); 1432 continue; 1433 } 1434 1435 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1436 // Reject loads and stores that are to the same address but are of 1437 // different types if we have to. 1438 if (S->getValueOperand()->getType() != LI->getType()) { 1439 // If the stored value is larger or equal to the loaded value, we can 1440 // reuse it. 1441 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1442 LI->getType(), *TD)) { 1443 UnavailableBlocks.push_back(DepBB); 1444 continue; 1445 } 1446 } 1447 1448 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1449 S->getValueOperand())); 1450 continue; 1451 } 1452 1453 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1454 // If the types mismatch and we can't handle it, reject reuse of the load. 1455 if (LD->getType() != LI->getType()) { 1456 // If the stored value is larger or equal to the loaded value, we can 1457 // reuse it. 1458 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1459 UnavailableBlocks.push_back(DepBB); 1460 continue; 1461 } 1462 } 1463 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1464 continue; 1465 } 1466 1467 UnavailableBlocks.push_back(DepBB); 1468 } 1469} 1470 1471bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 1472 UnavailBlkVect &UnavailableBlocks) { 1473 // Okay, we have *some* definitions of the value. This means that the value 1474 // is available in some of our (transitive) predecessors. Lets think about 1475 // doing PRE of this load. This will involve inserting a new load into the 1476 // predecessor when it's not available. We could do this in general, but 1477 // prefer to not increase code size. As such, we only do this when we know 1478 // that we only have to insert *one* load (which means we're basically moving 1479 // the load, not inserting a new one). 1480 1481 SmallPtrSet<BasicBlock *, 4> Blockers; 1482 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1483 Blockers.insert(UnavailableBlocks[i]); 1484 1485 // Let's find the first basic block with more than one predecessor. Walk 1486 // backwards through predecessors if needed. 1487 BasicBlock *LoadBB = LI->getParent(); 1488 BasicBlock *TmpBB = LoadBB; 1489 1490 while (TmpBB->getSinglePredecessor()) { 1491 TmpBB = TmpBB->getSinglePredecessor(); 1492 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1493 return false; 1494 if (Blockers.count(TmpBB)) 1495 return false; 1496 1497 // If any of these blocks has more than one successor (i.e. if the edge we 1498 // just traversed was critical), then there are other paths through this 1499 // block along which the load may not be anticipated. Hoisting the load 1500 // above this block would be adding the load to execution paths along 1501 // which it was not previously executed. 1502 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1503 return false; 1504 } 1505 1506 assert(TmpBB); 1507 LoadBB = TmpBB; 1508 1509 // Check to see how many predecessors have the loaded value fully 1510 // available. 1511 DenseMap<BasicBlock*, Value*> PredLoads; 1512 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1513 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1514 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1515 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1516 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1517 1518 SmallVector<BasicBlock *, 4> CriticalEdgePred; 1519 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1520 PI != E; ++PI) { 1521 BasicBlock *Pred = *PI; 1522 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1523 continue; 1524 } 1525 PredLoads[Pred] = 0; 1526 1527 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1528 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1529 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1530 << Pred->getName() << "': " << *LI << '\n'); 1531 return false; 1532 } 1533 1534 if (LoadBB->isLandingPad()) { 1535 DEBUG(dbgs() 1536 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1537 << Pred->getName() << "': " << *LI << '\n'); 1538 return false; 1539 } 1540 1541 CriticalEdgePred.push_back(Pred); 1542 } 1543 } 1544 1545 // Decide whether PRE is profitable for this load. 1546 unsigned NumUnavailablePreds = PredLoads.size(); 1547 assert(NumUnavailablePreds != 0 && 1548 "Fully available value should already be eliminated!"); 1549 1550 // If this load is unavailable in multiple predecessors, reject it. 1551 // FIXME: If we could restructure the CFG, we could make a common pred with 1552 // all the preds that don't have an available LI and insert a new load into 1553 // that one block. 1554 if (NumUnavailablePreds != 1) 1555 return false; 1556 1557 // Split critical edges, and update the unavailable predecessors accordingly. 1558 for (SmallVector<BasicBlock *, 4>::iterator I = CriticalEdgePred.begin(), 1559 E = CriticalEdgePred.end(); I != E; I++) { 1560 BasicBlock *OrigPred = *I; 1561 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); 1562 PredLoads.erase(OrigPred); 1563 PredLoads[NewPred] = 0; 1564 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" 1565 << LoadBB->getName() << '\n'); 1566 } 1567 1568 // Check if the load can safely be moved to all the unavailable predecessors. 1569 bool CanDoPRE = true; 1570 SmallVector<Instruction*, 8> NewInsts; 1571 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1572 E = PredLoads.end(); I != E; ++I) { 1573 BasicBlock *UnavailablePred = I->first; 1574 1575 // Do PHI translation to get its value in the predecessor if necessary. The 1576 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1577 1578 // If all preds have a single successor, then we know it is safe to insert 1579 // the load on the pred (?!?), so we can insert code to materialize the 1580 // pointer if it is not available. 1581 PHITransAddr Address(LI->getPointerOperand(), TD); 1582 Value *LoadPtr = 0; 1583 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1584 *DT, NewInsts); 1585 1586 // If we couldn't find or insert a computation of this phi translated value, 1587 // we fail PRE. 1588 if (LoadPtr == 0) { 1589 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1590 << *LI->getPointerOperand() << "\n"); 1591 CanDoPRE = false; 1592 break; 1593 } 1594 1595 I->second = LoadPtr; 1596 } 1597 1598 if (!CanDoPRE) { 1599 while (!NewInsts.empty()) { 1600 Instruction *I = NewInsts.pop_back_val(); 1601 if (MD) MD->removeInstruction(I); 1602 I->eraseFromParent(); 1603 } 1604 // HINT:Don't revert the edge-splitting as following transformation may 1605 // also need to split these critial edges. 1606 return !CriticalEdgePred.empty(); 1607 } 1608 1609 // Okay, we can eliminate this load by inserting a reload in the predecessor 1610 // and using PHI construction to get the value in the other predecessors, do 1611 // it. 1612 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1613 DEBUG(if (!NewInsts.empty()) 1614 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1615 << *NewInsts.back() << '\n'); 1616 1617 // Assign value numbers to the new instructions. 1618 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1619 // FIXME: We really _ought_ to insert these value numbers into their 1620 // parent's availability map. However, in doing so, we risk getting into 1621 // ordering issues. If a block hasn't been processed yet, we would be 1622 // marking a value as AVAIL-IN, which isn't what we intend. 1623 VN.lookup_or_add(NewInsts[i]); 1624 } 1625 1626 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1627 E = PredLoads.end(); I != E; ++I) { 1628 BasicBlock *UnavailablePred = I->first; 1629 Value *LoadPtr = I->second; 1630 1631 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1632 LI->getAlignment(), 1633 UnavailablePred->getTerminator()); 1634 1635 // Transfer the old load's TBAA tag to the new load. 1636 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1637 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1638 1639 // Transfer DebugLoc. 1640 NewLoad->setDebugLoc(LI->getDebugLoc()); 1641 1642 // Add the newly created load. 1643 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1644 NewLoad)); 1645 MD->invalidateCachedPointerInfo(LoadPtr); 1646 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1647 } 1648 1649 // Perform PHI construction. 1650 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1651 LI->replaceAllUsesWith(V); 1652 if (isa<PHINode>(V)) 1653 V->takeName(LI); 1654 if (V->getType()->getScalarType()->isPointerTy()) 1655 MD->invalidateCachedPointerInfo(V); 1656 markInstructionForDeletion(LI); 1657 ++NumPRELoad; 1658 return true; 1659} 1660 1661/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1662/// non-local by performing PHI construction. 1663bool GVN::processNonLocalLoad(LoadInst *LI) { 1664 // Step 1: Find the non-local dependencies of the load. 1665 LoadDepVect Deps; 1666 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1667 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1668 1669 // If we had to process more than one hundred blocks to find the 1670 // dependencies, this load isn't worth worrying about. Optimizing 1671 // it will be too expensive. 1672 unsigned NumDeps = Deps.size(); 1673 if (NumDeps > 100) 1674 return false; 1675 1676 // If we had a phi translation failure, we'll have a single entry which is a 1677 // clobber in the current block. Reject this early. 1678 if (NumDeps == 1 && 1679 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1680 DEBUG( 1681 dbgs() << "GVN: non-local load "; 1682 WriteAsOperand(dbgs(), LI); 1683 dbgs() << " has unknown dependencies\n"; 1684 ); 1685 return false; 1686 } 1687 1688 // Step 2: Analyze the availability of the load 1689 AvailValInBlkVect ValuesPerBlock; 1690 UnavailBlkVect UnavailableBlocks; 1691 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks); 1692 1693 // If we have no predecessors that produce a known value for this load, exit 1694 // early. 1695 if (ValuesPerBlock.empty()) 1696 return false; 1697 1698 // Step 3: Eliminate fully redundancy. 1699 // 1700 // If all of the instructions we depend on produce a known value for this 1701 // load, then it is fully redundant and we can use PHI insertion to compute 1702 // its value. Insert PHIs and remove the fully redundant value now. 1703 if (UnavailableBlocks.empty()) { 1704 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1705 1706 // Perform PHI construction. 1707 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1708 LI->replaceAllUsesWith(V); 1709 1710 if (isa<PHINode>(V)) 1711 V->takeName(LI); 1712 if (V->getType()->getScalarType()->isPointerTy()) 1713 MD->invalidateCachedPointerInfo(V); 1714 markInstructionForDeletion(LI); 1715 ++NumGVNLoad; 1716 return true; 1717 } 1718 1719 // Step 4: Eliminate partial redundancy. 1720 if (!EnablePRE || !EnableLoadPRE) 1721 return false; 1722 1723 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks); 1724} 1725 1726 1727static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1728 // Patch the replacement so that it is not more restrictive than the value 1729 // being replaced. 1730 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1731 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1732 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1733 isa<OverflowingBinaryOperator>(ReplOp)) { 1734 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1735 ReplOp->setHasNoSignedWrap(false); 1736 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1737 ReplOp->setHasNoUnsignedWrap(false); 1738 } 1739 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1740 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1741 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1742 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1743 unsigned Kind = Metadata[i].first; 1744 MDNode *IMD = I->getMetadata(Kind); 1745 MDNode *ReplMD = Metadata[i].second; 1746 switch(Kind) { 1747 default: 1748 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1749 break; 1750 case LLVMContext::MD_dbg: 1751 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1752 case LLVMContext::MD_tbaa: 1753 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1754 break; 1755 case LLVMContext::MD_range: 1756 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1757 break; 1758 case LLVMContext::MD_prof: 1759 llvm_unreachable("MD_prof in a non terminator instruction"); 1760 break; 1761 case LLVMContext::MD_fpmath: 1762 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1763 break; 1764 } 1765 } 1766 } 1767} 1768 1769static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1770 patchReplacementInstruction(I, Repl); 1771 I->replaceAllUsesWith(Repl); 1772} 1773 1774/// processLoad - Attempt to eliminate a load, first by eliminating it 1775/// locally, and then attempting non-local elimination if that fails. 1776bool GVN::processLoad(LoadInst *L) { 1777 if (!MD) 1778 return false; 1779 1780 if (!L->isSimple()) 1781 return false; 1782 1783 if (L->use_empty()) { 1784 markInstructionForDeletion(L); 1785 return true; 1786 } 1787 1788 // ... to a pointer that has been loaded from before... 1789 MemDepResult Dep = MD->getDependency(L); 1790 1791 // If we have a clobber and target data is around, see if this is a clobber 1792 // that we can fix up through code synthesis. 1793 if (Dep.isClobber() && TD) { 1794 // Check to see if we have something like this: 1795 // store i32 123, i32* %P 1796 // %A = bitcast i32* %P to i8* 1797 // %B = gep i8* %A, i32 1 1798 // %C = load i8* %B 1799 // 1800 // We could do that by recognizing if the clobber instructions are obviously 1801 // a common base + constant offset, and if the previous store (or memset) 1802 // completely covers this load. This sort of thing can happen in bitfield 1803 // access code. 1804 Value *AvailVal = 0; 1805 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1806 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1807 L->getPointerOperand(), 1808 DepSI, *TD); 1809 if (Offset != -1) 1810 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1811 L->getType(), L, *TD); 1812 } 1813 1814 // Check to see if we have something like this: 1815 // load i32* P 1816 // load i8* (P+1) 1817 // if we have this, replace the later with an extraction from the former. 1818 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1819 // If this is a clobber and L is the first instruction in its block, then 1820 // we have the first instruction in the entry block. 1821 if (DepLI == L) 1822 return false; 1823 1824 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1825 L->getPointerOperand(), 1826 DepLI, *TD); 1827 if (Offset != -1) 1828 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1829 } 1830 1831 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1832 // a value on from it. 1833 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1834 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1835 L->getPointerOperand(), 1836 DepMI, *TD); 1837 if (Offset != -1) 1838 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1839 } 1840 1841 if (AvailVal) { 1842 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1843 << *AvailVal << '\n' << *L << "\n\n\n"); 1844 1845 // Replace the load! 1846 L->replaceAllUsesWith(AvailVal); 1847 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1848 MD->invalidateCachedPointerInfo(AvailVal); 1849 markInstructionForDeletion(L); 1850 ++NumGVNLoad; 1851 return true; 1852 } 1853 } 1854 1855 // If the value isn't available, don't do anything! 1856 if (Dep.isClobber()) { 1857 DEBUG( 1858 // fast print dep, using operator<< on instruction is too slow. 1859 dbgs() << "GVN: load "; 1860 WriteAsOperand(dbgs(), L); 1861 Instruction *I = Dep.getInst(); 1862 dbgs() << " is clobbered by " << *I << '\n'; 1863 ); 1864 return false; 1865 } 1866 1867 // If it is defined in another block, try harder. 1868 if (Dep.isNonLocal()) 1869 return processNonLocalLoad(L); 1870 1871 if (!Dep.isDef()) { 1872 DEBUG( 1873 // fast print dep, using operator<< on instruction is too slow. 1874 dbgs() << "GVN: load "; 1875 WriteAsOperand(dbgs(), L); 1876 dbgs() << " has unknown dependence\n"; 1877 ); 1878 return false; 1879 } 1880 1881 Instruction *DepInst = Dep.getInst(); 1882 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1883 Value *StoredVal = DepSI->getValueOperand(); 1884 1885 // The store and load are to a must-aliased pointer, but they may not 1886 // actually have the same type. See if we know how to reuse the stored 1887 // value (depending on its type). 1888 if (StoredVal->getType() != L->getType()) { 1889 if (TD) { 1890 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1891 L, *TD); 1892 if (StoredVal == 0) 1893 return false; 1894 1895 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1896 << '\n' << *L << "\n\n\n"); 1897 } 1898 else 1899 return false; 1900 } 1901 1902 // Remove it! 1903 L->replaceAllUsesWith(StoredVal); 1904 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1905 MD->invalidateCachedPointerInfo(StoredVal); 1906 markInstructionForDeletion(L); 1907 ++NumGVNLoad; 1908 return true; 1909 } 1910 1911 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1912 Value *AvailableVal = DepLI; 1913 1914 // The loads are of a must-aliased pointer, but they may not actually have 1915 // the same type. See if we know how to reuse the previously loaded value 1916 // (depending on its type). 1917 if (DepLI->getType() != L->getType()) { 1918 if (TD) { 1919 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1920 L, *TD); 1921 if (AvailableVal == 0) 1922 return false; 1923 1924 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1925 << "\n" << *L << "\n\n\n"); 1926 } 1927 else 1928 return false; 1929 } 1930 1931 // Remove it! 1932 patchAndReplaceAllUsesWith(L, AvailableVal); 1933 if (DepLI->getType()->getScalarType()->isPointerTy()) 1934 MD->invalidateCachedPointerInfo(DepLI); 1935 markInstructionForDeletion(L); 1936 ++NumGVNLoad; 1937 return true; 1938 } 1939 1940 // If this load really doesn't depend on anything, then we must be loading an 1941 // undef value. This can happen when loading for a fresh allocation with no 1942 // intervening stores, for example. 1943 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1944 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1945 markInstructionForDeletion(L); 1946 ++NumGVNLoad; 1947 return true; 1948 } 1949 1950 // If this load occurs either right after a lifetime begin, 1951 // then the loaded value is undefined. 1952 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1953 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1954 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1955 markInstructionForDeletion(L); 1956 ++NumGVNLoad; 1957 return true; 1958 } 1959 } 1960 1961 return false; 1962} 1963 1964// findLeader - In order to find a leader for a given value number at a 1965// specific basic block, we first obtain the list of all Values for that number, 1966// and then scan the list to find one whose block dominates the block in 1967// question. This is fast because dominator tree queries consist of only 1968// a few comparisons of DFS numbers. 1969Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 1970 LeaderTableEntry Vals = LeaderTable[num]; 1971 if (!Vals.Val) return 0; 1972 1973 Value *Val = 0; 1974 if (DT->dominates(Vals.BB, BB)) { 1975 Val = Vals.Val; 1976 if (isa<Constant>(Val)) return Val; 1977 } 1978 1979 LeaderTableEntry* Next = Vals.Next; 1980 while (Next) { 1981 if (DT->dominates(Next->BB, BB)) { 1982 if (isa<Constant>(Next->Val)) return Next->Val; 1983 if (!Val) Val = Next->Val; 1984 } 1985 1986 Next = Next->Next; 1987 } 1988 1989 return Val; 1990} 1991 1992/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 1993/// use is dominated by the given basic block. Returns the number of uses that 1994/// were replaced. 1995unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 1996 const BasicBlockEdge &Root) { 1997 unsigned Count = 0; 1998 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1999 UI != UE; ) { 2000 Use &U = (UI++).getUse(); 2001 2002 if (DT->dominates(Root, U)) { 2003 U.set(To); 2004 ++Count; 2005 } 2006 } 2007 return Count; 2008} 2009 2010/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2011/// true if every path from the entry block to 'Dst' passes via this edge. In 2012/// particular 'Dst' must not be reachable via another edge from 'Src'. 2013static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2014 DominatorTree *DT) { 2015 // While in theory it is interesting to consider the case in which Dst has 2016 // more than one predecessor, because Dst might be part of a loop which is 2017 // only reachable from Src, in practice it is pointless since at the time 2018 // GVN runs all such loops have preheaders, which means that Dst will have 2019 // been changed to have only one predecessor, namely Src. 2020 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2021 const BasicBlock *Src = E.getStart(); 2022 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2023 (void)Src; 2024 return Pred != 0; 2025} 2026 2027/// propagateEquality - The given values are known to be equal in every block 2028/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2029/// 'RHS' everywhere in the scope. Returns whether a change was made. 2030bool GVN::propagateEquality(Value *LHS, Value *RHS, 2031 const BasicBlockEdge &Root) { 2032 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2033 Worklist.push_back(std::make_pair(LHS, RHS)); 2034 bool Changed = false; 2035 // For speed, compute a conservative fast approximation to 2036 // DT->dominates(Root, Root.getEnd()); 2037 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2038 2039 while (!Worklist.empty()) { 2040 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2041 LHS = Item.first; RHS = Item.second; 2042 2043 if (LHS == RHS) continue; 2044 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2045 2046 // Don't try to propagate equalities between constants. 2047 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2048 2049 // Prefer a constant on the right-hand side, or an Argument if no constants. 2050 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2051 std::swap(LHS, RHS); 2052 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2053 2054 // If there is no obvious reason to prefer the left-hand side over the right- 2055 // hand side, ensure the longest lived term is on the right-hand side, so the 2056 // shortest lived term will be replaced by the longest lived. This tends to 2057 // expose more simplifications. 2058 uint32_t LVN = VN.lookup_or_add(LHS); 2059 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2060 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2061 // Move the 'oldest' value to the right-hand side, using the value number as 2062 // a proxy for age. 2063 uint32_t RVN = VN.lookup_or_add(RHS); 2064 if (LVN < RVN) { 2065 std::swap(LHS, RHS); 2066 LVN = RVN; 2067 } 2068 } 2069 2070 // If value numbering later sees that an instruction in the scope is equal 2071 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2072 // the invariant that instructions only occur in the leader table for their 2073 // own value number (this is used by removeFromLeaderTable), do not do this 2074 // if RHS is an instruction (if an instruction in the scope is morphed into 2075 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2076 // using the leader table is about compiling faster, not optimizing better). 2077 // The leader table only tracks basic blocks, not edges. Only add to if we 2078 // have the simple case where the edge dominates the end. 2079 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2080 addToLeaderTable(LVN, RHS, Root.getEnd()); 2081 2082 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2083 // LHS always has at least one use that is not dominated by Root, this will 2084 // never do anything if LHS has only one use. 2085 if (!LHS->hasOneUse()) { 2086 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2087 Changed |= NumReplacements > 0; 2088 NumGVNEqProp += NumReplacements; 2089 } 2090 2091 // Now try to deduce additional equalities from this one. For example, if the 2092 // known equality was "(A != B)" == "false" then it follows that A and B are 2093 // equal in the scope. Only boolean equalities with an explicit true or false 2094 // RHS are currently supported. 2095 if (!RHS->getType()->isIntegerTy(1)) 2096 // Not a boolean equality - bail out. 2097 continue; 2098 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2099 if (!CI) 2100 // RHS neither 'true' nor 'false' - bail out. 2101 continue; 2102 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2103 bool isKnownTrue = CI->isAllOnesValue(); 2104 bool isKnownFalse = !isKnownTrue; 2105 2106 // If "A && B" is known true then both A and B are known true. If "A || B" 2107 // is known false then both A and B are known false. 2108 Value *A, *B; 2109 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2110 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2111 Worklist.push_back(std::make_pair(A, RHS)); 2112 Worklist.push_back(std::make_pair(B, RHS)); 2113 continue; 2114 } 2115 2116 // If we are propagating an equality like "(A == B)" == "true" then also 2117 // propagate the equality A == B. When propagating a comparison such as 2118 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2119 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2120 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2121 2122 // If "A == B" is known true, or "A != B" is known false, then replace 2123 // A with B everywhere in the scope. 2124 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2125 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2126 Worklist.push_back(std::make_pair(Op0, Op1)); 2127 2128 // If "A >= B" is known true, replace "A < B" with false everywhere. 2129 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2130 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2131 // Since we don't have the instruction "A < B" immediately to hand, work out 2132 // the value number that it would have and use that to find an appropriate 2133 // instruction (if any). 2134 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2135 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2136 // If the number we were assigned was brand new then there is no point in 2137 // looking for an instruction realizing it: there cannot be one! 2138 if (Num < NextNum) { 2139 Value *NotCmp = findLeader(Root.getEnd(), Num); 2140 if (NotCmp && isa<Instruction>(NotCmp)) { 2141 unsigned NumReplacements = 2142 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2143 Changed |= NumReplacements > 0; 2144 NumGVNEqProp += NumReplacements; 2145 } 2146 } 2147 // Ensure that any instruction in scope that gets the "A < B" value number 2148 // is replaced with false. 2149 // The leader table only tracks basic blocks, not edges. Only add to if we 2150 // have the simple case where the edge dominates the end. 2151 if (RootDominatesEnd) 2152 addToLeaderTable(Num, NotVal, Root.getEnd()); 2153 2154 continue; 2155 } 2156 } 2157 2158 return Changed; 2159} 2160 2161/// processInstruction - When calculating availability, handle an instruction 2162/// by inserting it into the appropriate sets 2163bool GVN::processInstruction(Instruction *I) { 2164 // Ignore dbg info intrinsics. 2165 if (isa<DbgInfoIntrinsic>(I)) 2166 return false; 2167 2168 // If the instruction can be easily simplified then do so now in preference 2169 // to value numbering it. Value numbering often exposes redundancies, for 2170 // example if it determines that %y is equal to %x then the instruction 2171 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2172 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2173 I->replaceAllUsesWith(V); 2174 if (MD && V->getType()->getScalarType()->isPointerTy()) 2175 MD->invalidateCachedPointerInfo(V); 2176 markInstructionForDeletion(I); 2177 ++NumGVNSimpl; 2178 return true; 2179 } 2180 2181 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2182 if (processLoad(LI)) 2183 return true; 2184 2185 unsigned Num = VN.lookup_or_add(LI); 2186 addToLeaderTable(Num, LI, LI->getParent()); 2187 return false; 2188 } 2189 2190 // For conditional branches, we can perform simple conditional propagation on 2191 // the condition value itself. 2192 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2193 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 2194 return false; 2195 2196 Value *BranchCond = BI->getCondition(); 2197 2198 BasicBlock *TrueSucc = BI->getSuccessor(0); 2199 BasicBlock *FalseSucc = BI->getSuccessor(1); 2200 // Avoid multiple edges early. 2201 if (TrueSucc == FalseSucc) 2202 return false; 2203 2204 BasicBlock *Parent = BI->getParent(); 2205 bool Changed = false; 2206 2207 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2208 BasicBlockEdge TrueE(Parent, TrueSucc); 2209 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2210 2211 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2212 BasicBlockEdge FalseE(Parent, FalseSucc); 2213 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2214 2215 return Changed; 2216 } 2217 2218 // For switches, propagate the case values into the case destinations. 2219 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2220 Value *SwitchCond = SI->getCondition(); 2221 BasicBlock *Parent = SI->getParent(); 2222 bool Changed = false; 2223 2224 // Remember how many outgoing edges there are to every successor. 2225 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2226 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2227 ++SwitchEdges[SI->getSuccessor(i)]; 2228 2229 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2230 i != e; ++i) { 2231 BasicBlock *Dst = i.getCaseSuccessor(); 2232 // If there is only a single edge, propagate the case value into it. 2233 if (SwitchEdges.lookup(Dst) == 1) { 2234 BasicBlockEdge E(Parent, Dst); 2235 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2236 } 2237 } 2238 return Changed; 2239 } 2240 2241 // Instructions with void type don't return a value, so there's 2242 // no point in trying to find redundancies in them. 2243 if (I->getType()->isVoidTy()) return false; 2244 2245 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2246 unsigned Num = VN.lookup_or_add(I); 2247 2248 // Allocations are always uniquely numbered, so we can save time and memory 2249 // by fast failing them. 2250 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2251 addToLeaderTable(Num, I, I->getParent()); 2252 return false; 2253 } 2254 2255 // If the number we were assigned was a brand new VN, then we don't 2256 // need to do a lookup to see if the number already exists 2257 // somewhere in the domtree: it can't! 2258 if (Num >= NextNum) { 2259 addToLeaderTable(Num, I, I->getParent()); 2260 return false; 2261 } 2262 2263 // Perform fast-path value-number based elimination of values inherited from 2264 // dominators. 2265 Value *repl = findLeader(I->getParent(), Num); 2266 if (repl == 0) { 2267 // Failure, just remember this instance for future use. 2268 addToLeaderTable(Num, I, I->getParent()); 2269 return false; 2270 } 2271 2272 // Remove it! 2273 patchAndReplaceAllUsesWith(I, repl); 2274 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2275 MD->invalidateCachedPointerInfo(repl); 2276 markInstructionForDeletion(I); 2277 return true; 2278} 2279 2280/// runOnFunction - This is the main transformation entry point for a function. 2281bool GVN::runOnFunction(Function& F) { 2282 if (!NoLoads) 2283 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2284 DT = &getAnalysis<DominatorTree>(); 2285 TD = getAnalysisIfAvailable<DataLayout>(); 2286 TLI = &getAnalysis<TargetLibraryInfo>(); 2287 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2288 VN.setMemDep(MD); 2289 VN.setDomTree(DT); 2290 2291 bool Changed = false; 2292 bool ShouldContinue = true; 2293 2294 // Merge unconditional branches, allowing PRE to catch more 2295 // optimization opportunities. 2296 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2297 BasicBlock *BB = FI++; 2298 2299 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2300 if (removedBlock) ++NumGVNBlocks; 2301 2302 Changed |= removedBlock; 2303 } 2304 2305 unsigned Iteration = 0; 2306 while (ShouldContinue) { 2307 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2308 ShouldContinue = iterateOnFunction(F); 2309 Changed |= ShouldContinue; 2310 ++Iteration; 2311 } 2312 2313 if (EnablePRE) { 2314 bool PREChanged = true; 2315 while (PREChanged) { 2316 PREChanged = performPRE(F); 2317 Changed |= PREChanged; 2318 } 2319 } 2320 2321 // FIXME: Should perform GVN again after PRE does something. PRE can move 2322 // computations into blocks where they become fully redundant. Note that 2323 // we can't do this until PRE's critical edge splitting updates memdep. 2324 // Actually, when this happens, we should just fully integrate PRE into GVN. 2325 2326 cleanupGlobalSets(); 2327 2328 return Changed; 2329} 2330 2331 2332bool GVN::processBlock(BasicBlock *BB) { 2333 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2334 // (and incrementing BI before processing an instruction). 2335 assert(InstrsToErase.empty() && 2336 "We expect InstrsToErase to be empty across iterations"); 2337 bool ChangedFunction = false; 2338 2339 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2340 BI != BE;) { 2341 ChangedFunction |= processInstruction(BI); 2342 if (InstrsToErase.empty()) { 2343 ++BI; 2344 continue; 2345 } 2346 2347 // If we need some instructions deleted, do it now. 2348 NumGVNInstr += InstrsToErase.size(); 2349 2350 // Avoid iterator invalidation. 2351 bool AtStart = BI == BB->begin(); 2352 if (!AtStart) 2353 --BI; 2354 2355 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 2356 E = InstrsToErase.end(); I != E; ++I) { 2357 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2358 if (MD) MD->removeInstruction(*I); 2359 DEBUG(verifyRemoved(*I)); 2360 (*I)->eraseFromParent(); 2361 } 2362 InstrsToErase.clear(); 2363 2364 if (AtStart) 2365 BI = BB->begin(); 2366 else 2367 ++BI; 2368 } 2369 2370 return ChangedFunction; 2371} 2372 2373/// performPRE - Perform a purely local form of PRE that looks for diamond 2374/// control flow patterns and attempts to perform simple PRE at the join point. 2375bool GVN::performPRE(Function &F) { 2376 bool Changed = false; 2377 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2378 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2379 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2380 BasicBlock *CurrentBlock = *DI; 2381 2382 // Nothing to PRE in the entry block. 2383 if (CurrentBlock == &F.getEntryBlock()) continue; 2384 2385 // Don't perform PRE on a landing pad. 2386 if (CurrentBlock->isLandingPad()) continue; 2387 2388 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2389 BE = CurrentBlock->end(); BI != BE; ) { 2390 Instruction *CurInst = BI++; 2391 2392 if (isa<AllocaInst>(CurInst) || 2393 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2394 CurInst->getType()->isVoidTy() || 2395 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2396 isa<DbgInfoIntrinsic>(CurInst)) 2397 continue; 2398 2399 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2400 // sinking the compare again, and it would force the code generator to 2401 // move the i1 from processor flags or predicate registers into a general 2402 // purpose register. 2403 if (isa<CmpInst>(CurInst)) 2404 continue; 2405 2406 // We don't currently value number ANY inline asm calls. 2407 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2408 if (CallI->isInlineAsm()) 2409 continue; 2410 2411 uint32_t ValNo = VN.lookup(CurInst); 2412 2413 // Look for the predecessors for PRE opportunities. We're 2414 // only trying to solve the basic diamond case, where 2415 // a value is computed in the successor and one predecessor, 2416 // but not the other. We also explicitly disallow cases 2417 // where the successor is its own predecessor, because they're 2418 // more complicated to get right. 2419 unsigned NumWith = 0; 2420 unsigned NumWithout = 0; 2421 BasicBlock *PREPred = 0; 2422 predMap.clear(); 2423 2424 for (pred_iterator PI = pred_begin(CurrentBlock), 2425 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2426 BasicBlock *P = *PI; 2427 // We're not interested in PRE where the block is its 2428 // own predecessor, or in blocks with predecessors 2429 // that are not reachable. 2430 if (P == CurrentBlock) { 2431 NumWithout = 2; 2432 break; 2433 } else if (!DT->isReachableFromEntry(P)) { 2434 NumWithout = 2; 2435 break; 2436 } 2437 2438 Value* predV = findLeader(P, ValNo); 2439 if (predV == 0) { 2440 predMap.push_back(std::make_pair(static_cast<Value *>(0), P)); 2441 PREPred = P; 2442 ++NumWithout; 2443 } else if (predV == CurInst) { 2444 /* CurInst dominates this predecessor. */ 2445 NumWithout = 2; 2446 break; 2447 } else { 2448 predMap.push_back(std::make_pair(predV, P)); 2449 ++NumWith; 2450 } 2451 } 2452 2453 // Don't do PRE when it might increase code size, i.e. when 2454 // we would need to insert instructions in more than one pred. 2455 if (NumWithout != 1 || NumWith == 0) 2456 continue; 2457 2458 // Don't do PRE across indirect branch. 2459 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2460 continue; 2461 2462 // We can't do PRE safely on a critical edge, so instead we schedule 2463 // the edge to be split and perform the PRE the next time we iterate 2464 // on the function. 2465 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2466 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2467 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2468 continue; 2469 } 2470 2471 // Instantiate the expression in the predecessor that lacked it. 2472 // Because we are going top-down through the block, all value numbers 2473 // will be available in the predecessor by the time we need them. Any 2474 // that weren't originally present will have been instantiated earlier 2475 // in this loop. 2476 Instruction *PREInstr = CurInst->clone(); 2477 bool success = true; 2478 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2479 Value *Op = PREInstr->getOperand(i); 2480 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2481 continue; 2482 2483 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2484 PREInstr->setOperand(i, V); 2485 } else { 2486 success = false; 2487 break; 2488 } 2489 } 2490 2491 // Fail out if we encounter an operand that is not available in 2492 // the PRE predecessor. This is typically because of loads which 2493 // are not value numbered precisely. 2494 if (!success) { 2495 DEBUG(verifyRemoved(PREInstr)); 2496 delete PREInstr; 2497 continue; 2498 } 2499 2500 PREInstr->insertBefore(PREPred->getTerminator()); 2501 PREInstr->setName(CurInst->getName() + ".pre"); 2502 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2503 VN.add(PREInstr, ValNo); 2504 ++NumGVNPRE; 2505 2506 // Update the availability map to include the new instruction. 2507 addToLeaderTable(ValNo, PREInstr, PREPred); 2508 2509 // Create a PHI to make the value available in this block. 2510 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2511 CurInst->getName() + ".pre-phi", 2512 CurrentBlock->begin()); 2513 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2514 if (Value *V = predMap[i].first) 2515 Phi->addIncoming(V, predMap[i].second); 2516 else 2517 Phi->addIncoming(PREInstr, PREPred); 2518 } 2519 2520 VN.add(Phi, ValNo); 2521 addToLeaderTable(ValNo, Phi, CurrentBlock); 2522 Phi->setDebugLoc(CurInst->getDebugLoc()); 2523 CurInst->replaceAllUsesWith(Phi); 2524 if (Phi->getType()->getScalarType()->isPointerTy()) { 2525 // Because we have added a PHI-use of the pointer value, it has now 2526 // "escaped" from alias analysis' perspective. We need to inform 2527 // AA of this. 2528 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2529 ++ii) { 2530 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2531 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2532 } 2533 2534 if (MD) 2535 MD->invalidateCachedPointerInfo(Phi); 2536 } 2537 VN.erase(CurInst); 2538 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2539 2540 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2541 if (MD) MD->removeInstruction(CurInst); 2542 DEBUG(verifyRemoved(CurInst)); 2543 CurInst->eraseFromParent(); 2544 Changed = true; 2545 } 2546 } 2547 2548 if (splitCriticalEdges()) 2549 Changed = true; 2550 2551 return Changed; 2552} 2553 2554/// Split the critical edge connecting the given two blocks, and return 2555/// the block inserted to the critical edge. 2556BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { 2557 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this); 2558 if (MD) 2559 MD->invalidateCachedPredecessors(); 2560 return BB; 2561} 2562 2563/// splitCriticalEdges - Split critical edges found during the previous 2564/// iteration that may enable further optimization. 2565bool GVN::splitCriticalEdges() { 2566 if (toSplit.empty()) 2567 return false; 2568 do { 2569 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2570 SplitCriticalEdge(Edge.first, Edge.second, this); 2571 } while (!toSplit.empty()); 2572 if (MD) MD->invalidateCachedPredecessors(); 2573 return true; 2574} 2575 2576/// iterateOnFunction - Executes one iteration of GVN 2577bool GVN::iterateOnFunction(Function &F) { 2578 cleanupGlobalSets(); 2579 2580 // Top-down walk of the dominator tree 2581 bool Changed = false; 2582#if 0 2583 // Needed for value numbering with phi construction to work. 2584 ReversePostOrderTraversal<Function*> RPOT(&F); 2585 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2586 RE = RPOT.end(); RI != RE; ++RI) 2587 Changed |= processBlock(*RI); 2588#else 2589 // Save the blocks this function have before transformation begins. GVN may 2590 // split critical edge, and hence may invalidate the RPO/DT iterator. 2591 // 2592 std::vector<BasicBlock *> BBVect; 2593 BBVect.reserve(256); 2594 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2595 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2596 BBVect.push_back(DI->getBlock()); 2597 2598 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end(); 2599 I != E; I++) 2600 Changed |= processBlock(*I); 2601#endif 2602 2603 return Changed; 2604} 2605 2606void GVN::cleanupGlobalSets() { 2607 VN.clear(); 2608 LeaderTable.clear(); 2609 TableAllocator.Reset(); 2610} 2611 2612/// verifyRemoved - Verify that the specified instruction does not occur in our 2613/// internal data structures. 2614void GVN::verifyRemoved(const Instruction *Inst) const { 2615 VN.verifyRemoved(Inst); 2616 2617 // Walk through the value number scope to make sure the instruction isn't 2618 // ferreted away in it. 2619 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2620 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2621 const LeaderTableEntry *Node = &I->second; 2622 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2623 2624 while (Node->Next) { 2625 Node = Node->Next; 2626 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2627 } 2628 } 2629} 2630