GVN.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/SetVector.h" 24#include "llvm/ADT/SmallPtrSet.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/Analysis/AliasAnalysis.h" 27#include "llvm/Analysis/CFG.h" 28#include "llvm/Analysis/ConstantFolding.h" 29#include "llvm/Analysis/InstructionSimplify.h" 30#include "llvm/Analysis/Loads.h" 31#include "llvm/Analysis/MemoryBuiltins.h" 32#include "llvm/Analysis/MemoryDependenceAnalysis.h" 33#include "llvm/Analysis/PHITransAddr.h" 34#include "llvm/Analysis/ValueTracking.h" 35#include "llvm/IR/DataLayout.h" 36#include "llvm/IR/Dominators.h" 37#include "llvm/IR/GlobalVariable.h" 38#include "llvm/IR/IRBuilder.h" 39#include "llvm/IR/IntrinsicInst.h" 40#include "llvm/IR/LLVMContext.h" 41#include "llvm/IR/Metadata.h" 42#include "llvm/IR/PatternMatch.h" 43#include "llvm/Support/Allocator.h" 44#include "llvm/Support/CommandLine.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Target/TargetLibraryInfo.h" 47#include "llvm/Transforms/Utils/BasicBlockUtils.h" 48#include "llvm/Transforms/Utils/SSAUpdater.h" 49#include <vector> 50using namespace llvm; 51using namespace PatternMatch; 52 53STATISTIC(NumGVNInstr, "Number of instructions deleted"); 54STATISTIC(NumGVNLoad, "Number of loads deleted"); 55STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 56STATISTIC(NumGVNBlocks, "Number of blocks merged"); 57STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 58STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 59STATISTIC(NumPRELoad, "Number of loads PRE'd"); 60 61static cl::opt<bool> EnablePRE("enable-pre", 62 cl::init(true), cl::Hidden); 63static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 64 65// Maximum allowed recursion depth. 66static cl::opt<uint32_t> 67MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 68 cl::desc("Max recurse depth (default = 1000)")); 69 70//===----------------------------------------------------------------------===// 71// ValueTable Class 72//===----------------------------------------------------------------------===// 73 74/// This class holds the mapping between values and value numbers. It is used 75/// as an efficient mechanism to determine the expression-wise equivalence of 76/// two values. 77namespace { 78 struct Expression { 79 uint32_t opcode; 80 Type *type; 81 SmallVector<uint32_t, 4> varargs; 82 83 Expression(uint32_t o = ~2U) : opcode(o) { } 84 85 bool operator==(const Expression &other) const { 86 if (opcode != other.opcode) 87 return false; 88 if (opcode == ~0U || opcode == ~1U) 89 return true; 90 if (type != other.type) 91 return false; 92 if (varargs != other.varargs) 93 return false; 94 return true; 95 } 96 97 friend hash_code hash_value(const Expression &Value) { 98 return hash_combine(Value.opcode, Value.type, 99 hash_combine_range(Value.varargs.begin(), 100 Value.varargs.end())); 101 } 102 }; 103 104 class ValueTable { 105 DenseMap<Value*, uint32_t> valueNumbering; 106 DenseMap<Expression, uint32_t> expressionNumbering; 107 AliasAnalysis *AA; 108 MemoryDependenceAnalysis *MD; 109 DominatorTree *DT; 110 111 uint32_t nextValueNumber; 112 113 Expression create_expression(Instruction* I); 114 Expression create_cmp_expression(unsigned Opcode, 115 CmpInst::Predicate Predicate, 116 Value *LHS, Value *RHS); 117 Expression create_extractvalue_expression(ExtractValueInst* EI); 118 uint32_t lookup_or_add_call(CallInst* C); 119 public: 120 ValueTable() : nextValueNumber(1) { } 121 uint32_t lookup_or_add(Value *V); 122 uint32_t lookup(Value *V) const; 123 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 124 Value *LHS, Value *RHS); 125 void add(Value *V, uint32_t num); 126 void clear(); 127 void erase(Value *v); 128 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 129 AliasAnalysis *getAliasAnalysis() const { return AA; } 130 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 131 void setDomTree(DominatorTree* D) { DT = D; } 132 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 133 void verifyRemoved(const Value *) const; 134 }; 135} 136 137namespace llvm { 138template <> struct DenseMapInfo<Expression> { 139 static inline Expression getEmptyKey() { 140 return ~0U; 141 } 142 143 static inline Expression getTombstoneKey() { 144 return ~1U; 145 } 146 147 static unsigned getHashValue(const Expression e) { 148 using llvm::hash_value; 149 return static_cast<unsigned>(hash_value(e)); 150 } 151 static bool isEqual(const Expression &LHS, const Expression &RHS) { 152 return LHS == RHS; 153 } 154}; 155 156} 157 158//===----------------------------------------------------------------------===// 159// ValueTable Internal Functions 160//===----------------------------------------------------------------------===// 161 162Expression ValueTable::create_expression(Instruction *I) { 163 Expression e; 164 e.type = I->getType(); 165 e.opcode = I->getOpcode(); 166 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 167 OI != OE; ++OI) 168 e.varargs.push_back(lookup_or_add(*OI)); 169 if (I->isCommutative()) { 170 // Ensure that commutative instructions that only differ by a permutation 171 // of their operands get the same value number by sorting the operand value 172 // numbers. Since all commutative instructions have two operands it is more 173 // efficient to sort by hand rather than using, say, std::sort. 174 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 175 if (e.varargs[0] > e.varargs[1]) 176 std::swap(e.varargs[0], e.varargs[1]); 177 } 178 179 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 180 // Sort the operand value numbers so x<y and y>x get the same value number. 181 CmpInst::Predicate Predicate = C->getPredicate(); 182 if (e.varargs[0] > e.varargs[1]) { 183 std::swap(e.varargs[0], e.varargs[1]); 184 Predicate = CmpInst::getSwappedPredicate(Predicate); 185 } 186 e.opcode = (C->getOpcode() << 8) | Predicate; 187 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 188 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 189 II != IE; ++II) 190 e.varargs.push_back(*II); 191 } 192 193 return e; 194} 195 196Expression ValueTable::create_cmp_expression(unsigned Opcode, 197 CmpInst::Predicate Predicate, 198 Value *LHS, Value *RHS) { 199 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 200 "Not a comparison!"); 201 Expression e; 202 e.type = CmpInst::makeCmpResultType(LHS->getType()); 203 e.varargs.push_back(lookup_or_add(LHS)); 204 e.varargs.push_back(lookup_or_add(RHS)); 205 206 // Sort the operand value numbers so x<y and y>x get the same value number. 207 if (e.varargs[0] > e.varargs[1]) { 208 std::swap(e.varargs[0], e.varargs[1]); 209 Predicate = CmpInst::getSwappedPredicate(Predicate); 210 } 211 e.opcode = (Opcode << 8) | Predicate; 212 return e; 213} 214 215Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 216 assert(EI != 0 && "Not an ExtractValueInst?"); 217 Expression e; 218 e.type = EI->getType(); 219 e.opcode = 0; 220 221 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 222 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 223 // EI might be an extract from one of our recognised intrinsics. If it 224 // is we'll synthesize a semantically equivalent expression instead on 225 // an extract value expression. 226 switch (I->getIntrinsicID()) { 227 case Intrinsic::sadd_with_overflow: 228 case Intrinsic::uadd_with_overflow: 229 e.opcode = Instruction::Add; 230 break; 231 case Intrinsic::ssub_with_overflow: 232 case Intrinsic::usub_with_overflow: 233 e.opcode = Instruction::Sub; 234 break; 235 case Intrinsic::smul_with_overflow: 236 case Intrinsic::umul_with_overflow: 237 e.opcode = Instruction::Mul; 238 break; 239 default: 240 break; 241 } 242 243 if (e.opcode != 0) { 244 // Intrinsic recognized. Grab its args to finish building the expression. 245 assert(I->getNumArgOperands() == 2 && 246 "Expect two args for recognised intrinsics."); 247 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 248 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 249 return e; 250 } 251 } 252 253 // Not a recognised intrinsic. Fall back to producing an extract value 254 // expression. 255 e.opcode = EI->getOpcode(); 256 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 257 OI != OE; ++OI) 258 e.varargs.push_back(lookup_or_add(*OI)); 259 260 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 261 II != IE; ++II) 262 e.varargs.push_back(*II); 263 264 return e; 265} 266 267//===----------------------------------------------------------------------===// 268// ValueTable External Functions 269//===----------------------------------------------------------------------===// 270 271/// add - Insert a value into the table with a specified value number. 272void ValueTable::add(Value *V, uint32_t num) { 273 valueNumbering.insert(std::make_pair(V, num)); 274} 275 276uint32_t ValueTable::lookup_or_add_call(CallInst *C) { 277 if (AA->doesNotAccessMemory(C)) { 278 Expression exp = create_expression(C); 279 uint32_t &e = expressionNumbering[exp]; 280 if (!e) e = nextValueNumber++; 281 valueNumbering[C] = e; 282 return e; 283 } else if (AA->onlyReadsMemory(C)) { 284 Expression exp = create_expression(C); 285 uint32_t &e = expressionNumbering[exp]; 286 if (!e) { 287 e = nextValueNumber++; 288 valueNumbering[C] = e; 289 return e; 290 } 291 if (!MD) { 292 e = nextValueNumber++; 293 valueNumbering[C] = e; 294 return e; 295 } 296 297 MemDepResult local_dep = MD->getDependency(C); 298 299 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 300 valueNumbering[C] = nextValueNumber; 301 return nextValueNumber++; 302 } 303 304 if (local_dep.isDef()) { 305 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 306 307 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 308 valueNumbering[C] = nextValueNumber; 309 return nextValueNumber++; 310 } 311 312 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 313 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 314 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 315 if (c_vn != cd_vn) { 316 valueNumbering[C] = nextValueNumber; 317 return nextValueNumber++; 318 } 319 } 320 321 uint32_t v = lookup_or_add(local_cdep); 322 valueNumbering[C] = v; 323 return v; 324 } 325 326 // Non-local case. 327 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 328 MD->getNonLocalCallDependency(CallSite(C)); 329 // FIXME: Move the checking logic to MemDep! 330 CallInst* cdep = 0; 331 332 // Check to see if we have a single dominating call instruction that is 333 // identical to C. 334 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 335 const NonLocalDepEntry *I = &deps[i]; 336 if (I->getResult().isNonLocal()) 337 continue; 338 339 // We don't handle non-definitions. If we already have a call, reject 340 // instruction dependencies. 341 if (!I->getResult().isDef() || cdep != 0) { 342 cdep = 0; 343 break; 344 } 345 346 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 347 // FIXME: All duplicated with non-local case. 348 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 349 cdep = NonLocalDepCall; 350 continue; 351 } 352 353 cdep = 0; 354 break; 355 } 356 357 if (!cdep) { 358 valueNumbering[C] = nextValueNumber; 359 return nextValueNumber++; 360 } 361 362 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 363 valueNumbering[C] = nextValueNumber; 364 return nextValueNumber++; 365 } 366 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 367 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 368 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 369 if (c_vn != cd_vn) { 370 valueNumbering[C] = nextValueNumber; 371 return nextValueNumber++; 372 } 373 } 374 375 uint32_t v = lookup_or_add(cdep); 376 valueNumbering[C] = v; 377 return v; 378 379 } else { 380 valueNumbering[C] = nextValueNumber; 381 return nextValueNumber++; 382 } 383} 384 385/// lookup_or_add - Returns the value number for the specified value, assigning 386/// it a new number if it did not have one before. 387uint32_t ValueTable::lookup_or_add(Value *V) { 388 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 389 if (VI != valueNumbering.end()) 390 return VI->second; 391 392 if (!isa<Instruction>(V)) { 393 valueNumbering[V] = nextValueNumber; 394 return nextValueNumber++; 395 } 396 397 Instruction* I = cast<Instruction>(V); 398 Expression exp; 399 switch (I->getOpcode()) { 400 case Instruction::Call: 401 return lookup_or_add_call(cast<CallInst>(I)); 402 case Instruction::Add: 403 case Instruction::FAdd: 404 case Instruction::Sub: 405 case Instruction::FSub: 406 case Instruction::Mul: 407 case Instruction::FMul: 408 case Instruction::UDiv: 409 case Instruction::SDiv: 410 case Instruction::FDiv: 411 case Instruction::URem: 412 case Instruction::SRem: 413 case Instruction::FRem: 414 case Instruction::Shl: 415 case Instruction::LShr: 416 case Instruction::AShr: 417 case Instruction::And: 418 case Instruction::Or: 419 case Instruction::Xor: 420 case Instruction::ICmp: 421 case Instruction::FCmp: 422 case Instruction::Trunc: 423 case Instruction::ZExt: 424 case Instruction::SExt: 425 case Instruction::FPToUI: 426 case Instruction::FPToSI: 427 case Instruction::UIToFP: 428 case Instruction::SIToFP: 429 case Instruction::FPTrunc: 430 case Instruction::FPExt: 431 case Instruction::PtrToInt: 432 case Instruction::IntToPtr: 433 case Instruction::BitCast: 434 case Instruction::Select: 435 case Instruction::ExtractElement: 436 case Instruction::InsertElement: 437 case Instruction::ShuffleVector: 438 case Instruction::InsertValue: 439 case Instruction::GetElementPtr: 440 exp = create_expression(I); 441 break; 442 case Instruction::ExtractValue: 443 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 444 break; 445 default: 446 valueNumbering[V] = nextValueNumber; 447 return nextValueNumber++; 448 } 449 450 uint32_t& e = expressionNumbering[exp]; 451 if (!e) e = nextValueNumber++; 452 valueNumbering[V] = e; 453 return e; 454} 455 456/// lookup - Returns the value number of the specified value. Fails if 457/// the value has not yet been numbered. 458uint32_t ValueTable::lookup(Value *V) const { 459 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 460 assert(VI != valueNumbering.end() && "Value not numbered?"); 461 return VI->second; 462} 463 464/// lookup_or_add_cmp - Returns the value number of the given comparison, 465/// assigning it a new number if it did not have one before. Useful when 466/// we deduced the result of a comparison, but don't immediately have an 467/// instruction realizing that comparison to hand. 468uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 469 CmpInst::Predicate Predicate, 470 Value *LHS, Value *RHS) { 471 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 472 uint32_t& e = expressionNumbering[exp]; 473 if (!e) e = nextValueNumber++; 474 return e; 475} 476 477/// clear - Remove all entries from the ValueTable. 478void ValueTable::clear() { 479 valueNumbering.clear(); 480 expressionNumbering.clear(); 481 nextValueNumber = 1; 482} 483 484/// erase - Remove a value from the value numbering. 485void ValueTable::erase(Value *V) { 486 valueNumbering.erase(V); 487} 488 489/// verifyRemoved - Verify that the value is removed from all internal data 490/// structures. 491void ValueTable::verifyRemoved(const Value *V) const { 492 for (DenseMap<Value*, uint32_t>::const_iterator 493 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 494 assert(I->first != V && "Inst still occurs in value numbering map!"); 495 } 496} 497 498//===----------------------------------------------------------------------===// 499// GVN Pass 500//===----------------------------------------------------------------------===// 501 502namespace { 503 class GVN; 504 struct AvailableValueInBlock { 505 /// BB - The basic block in question. 506 BasicBlock *BB; 507 enum ValType { 508 SimpleVal, // A simple offsetted value that is accessed. 509 LoadVal, // A value produced by a load. 510 MemIntrin, // A memory intrinsic which is loaded from. 511 UndefVal // A UndefValue representing a value from dead block (which 512 // is not yet physically removed from the CFG). 513 }; 514 515 /// V - The value that is live out of the block. 516 PointerIntPair<Value *, 2, ValType> Val; 517 518 /// Offset - The byte offset in Val that is interesting for the load query. 519 unsigned Offset; 520 521 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 522 unsigned Offset = 0) { 523 AvailableValueInBlock Res; 524 Res.BB = BB; 525 Res.Val.setPointer(V); 526 Res.Val.setInt(SimpleVal); 527 Res.Offset = Offset; 528 return Res; 529 } 530 531 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 532 unsigned Offset = 0) { 533 AvailableValueInBlock Res; 534 Res.BB = BB; 535 Res.Val.setPointer(MI); 536 Res.Val.setInt(MemIntrin); 537 Res.Offset = Offset; 538 return Res; 539 } 540 541 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 542 unsigned Offset = 0) { 543 AvailableValueInBlock Res; 544 Res.BB = BB; 545 Res.Val.setPointer(LI); 546 Res.Val.setInt(LoadVal); 547 Res.Offset = Offset; 548 return Res; 549 } 550 551 static AvailableValueInBlock getUndef(BasicBlock *BB) { 552 AvailableValueInBlock Res; 553 Res.BB = BB; 554 Res.Val.setPointer(0); 555 Res.Val.setInt(UndefVal); 556 Res.Offset = 0; 557 return Res; 558 } 559 560 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 561 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 562 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 563 bool isUndefValue() const { return Val.getInt() == UndefVal; } 564 565 Value *getSimpleValue() const { 566 assert(isSimpleValue() && "Wrong accessor"); 567 return Val.getPointer(); 568 } 569 570 LoadInst *getCoercedLoadValue() const { 571 assert(isCoercedLoadValue() && "Wrong accessor"); 572 return cast<LoadInst>(Val.getPointer()); 573 } 574 575 MemIntrinsic *getMemIntrinValue() const { 576 assert(isMemIntrinValue() && "Wrong accessor"); 577 return cast<MemIntrinsic>(Val.getPointer()); 578 } 579 580 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 581 /// defined here to the specified type. This handles various coercion cases. 582 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const; 583 }; 584 585 class GVN : public FunctionPass { 586 bool NoLoads; 587 MemoryDependenceAnalysis *MD; 588 DominatorTree *DT; 589 const DataLayout *DL; 590 const TargetLibraryInfo *TLI; 591 SetVector<BasicBlock *> DeadBlocks; 592 593 ValueTable VN; 594 595 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 596 /// have that value number. Use findLeader to query it. 597 struct LeaderTableEntry { 598 Value *Val; 599 const BasicBlock *BB; 600 LeaderTableEntry *Next; 601 }; 602 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 603 BumpPtrAllocator TableAllocator; 604 605 SmallVector<Instruction*, 8> InstrsToErase; 606 607 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect; 608 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect; 609 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect; 610 611 public: 612 static char ID; // Pass identification, replacement for typeid 613 explicit GVN(bool noloads = false) 614 : FunctionPass(ID), NoLoads(noloads), MD(0) { 615 initializeGVNPass(*PassRegistry::getPassRegistry()); 616 } 617 618 bool runOnFunction(Function &F) override; 619 620 /// markInstructionForDeletion - This removes the specified instruction from 621 /// our various maps and marks it for deletion. 622 void markInstructionForDeletion(Instruction *I) { 623 VN.erase(I); 624 InstrsToErase.push_back(I); 625 } 626 627 const DataLayout *getDataLayout() const { return DL; } 628 DominatorTree &getDominatorTree() const { return *DT; } 629 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 630 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 631 private: 632 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 633 /// its value number. 634 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 635 LeaderTableEntry &Curr = LeaderTable[N]; 636 if (!Curr.Val) { 637 Curr.Val = V; 638 Curr.BB = BB; 639 return; 640 } 641 642 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 643 Node->Val = V; 644 Node->BB = BB; 645 Node->Next = Curr.Next; 646 Curr.Next = Node; 647 } 648 649 /// removeFromLeaderTable - Scan the list of values corresponding to a given 650 /// value number, and remove the given instruction if encountered. 651 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 652 LeaderTableEntry* Prev = 0; 653 LeaderTableEntry* Curr = &LeaderTable[N]; 654 655 while (Curr->Val != I || Curr->BB != BB) { 656 Prev = Curr; 657 Curr = Curr->Next; 658 } 659 660 if (Prev) { 661 Prev->Next = Curr->Next; 662 } else { 663 if (!Curr->Next) { 664 Curr->Val = 0; 665 Curr->BB = 0; 666 } else { 667 LeaderTableEntry* Next = Curr->Next; 668 Curr->Val = Next->Val; 669 Curr->BB = Next->BB; 670 Curr->Next = Next->Next; 671 } 672 } 673 } 674 675 // List of critical edges to be split between iterations. 676 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 677 678 // This transformation requires dominator postdominator info 679 void getAnalysisUsage(AnalysisUsage &AU) const override { 680 AU.addRequired<DominatorTreeWrapperPass>(); 681 AU.addRequired<TargetLibraryInfo>(); 682 if (!NoLoads) 683 AU.addRequired<MemoryDependenceAnalysis>(); 684 AU.addRequired<AliasAnalysis>(); 685 686 AU.addPreserved<DominatorTreeWrapperPass>(); 687 AU.addPreserved<AliasAnalysis>(); 688 } 689 690 691 // Helper fuctions of redundant load elimination 692 bool processLoad(LoadInst *L); 693 bool processNonLocalLoad(LoadInst *L); 694 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 695 AvailValInBlkVect &ValuesPerBlock, 696 UnavailBlkVect &UnavailableBlocks); 697 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 698 UnavailBlkVect &UnavailableBlocks); 699 700 // Other helper routines 701 bool processInstruction(Instruction *I); 702 bool processBlock(BasicBlock *BB); 703 void dump(DenseMap<uint32_t, Value*> &d); 704 bool iterateOnFunction(Function &F); 705 bool performPRE(Function &F); 706 Value *findLeader(const BasicBlock *BB, uint32_t num); 707 void cleanupGlobalSets(); 708 void verifyRemoved(const Instruction *I) const; 709 bool splitCriticalEdges(); 710 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ); 711 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 712 const BasicBlockEdge &Root); 713 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 714 bool processFoldableCondBr(BranchInst *BI); 715 void addDeadBlock(BasicBlock *BB); 716 void assignValNumForDeadCode(); 717 }; 718 719 char GVN::ID = 0; 720} 721 722// createGVNPass - The public interface to this file... 723FunctionPass *llvm::createGVNPass(bool NoLoads) { 724 return new GVN(NoLoads); 725} 726 727INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 728INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 729INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 730INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 731INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 732INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 733 734#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 735void GVN::dump(DenseMap<uint32_t, Value*>& d) { 736 errs() << "{\n"; 737 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 738 E = d.end(); I != E; ++I) { 739 errs() << I->first << "\n"; 740 I->second->dump(); 741 } 742 errs() << "}\n"; 743} 744#endif 745 746/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 747/// we're analyzing is fully available in the specified block. As we go, keep 748/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 749/// map is actually a tri-state map with the following values: 750/// 0) we know the block *is not* fully available. 751/// 1) we know the block *is* fully available. 752/// 2) we do not know whether the block is fully available or not, but we are 753/// currently speculating that it will be. 754/// 3) we are speculating for this block and have used that to speculate for 755/// other blocks. 756static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 757 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 758 uint32_t RecurseDepth) { 759 if (RecurseDepth > MaxRecurseDepth) 760 return false; 761 762 // Optimistically assume that the block is fully available and check to see 763 // if we already know about this block in one lookup. 764 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 765 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 766 767 // If the entry already existed for this block, return the precomputed value. 768 if (!IV.second) { 769 // If this is a speculative "available" value, mark it as being used for 770 // speculation of other blocks. 771 if (IV.first->second == 2) 772 IV.first->second = 3; 773 return IV.first->second != 0; 774 } 775 776 // Otherwise, see if it is fully available in all predecessors. 777 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 778 779 // If this block has no predecessors, it isn't live-in here. 780 if (PI == PE) 781 goto SpeculationFailure; 782 783 for (; PI != PE; ++PI) 784 // If the value isn't fully available in one of our predecessors, then it 785 // isn't fully available in this block either. Undo our previous 786 // optimistic assumption and bail out. 787 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 788 goto SpeculationFailure; 789 790 return true; 791 792// SpeculationFailure - If we get here, we found out that this is not, after 793// all, a fully-available block. We have a problem if we speculated on this and 794// used the speculation to mark other blocks as available. 795SpeculationFailure: 796 char &BBVal = FullyAvailableBlocks[BB]; 797 798 // If we didn't speculate on this, just return with it set to false. 799 if (BBVal == 2) { 800 BBVal = 0; 801 return false; 802 } 803 804 // If we did speculate on this value, we could have blocks set to 1 that are 805 // incorrect. Walk the (transitive) successors of this block and mark them as 806 // 0 if set to one. 807 SmallVector<BasicBlock*, 32> BBWorklist; 808 BBWorklist.push_back(BB); 809 810 do { 811 BasicBlock *Entry = BBWorklist.pop_back_val(); 812 // Note that this sets blocks to 0 (unavailable) if they happen to not 813 // already be in FullyAvailableBlocks. This is safe. 814 char &EntryVal = FullyAvailableBlocks[Entry]; 815 if (EntryVal == 0) continue; // Already unavailable. 816 817 // Mark as unavailable. 818 EntryVal = 0; 819 820 BBWorklist.append(succ_begin(Entry), succ_end(Entry)); 821 } while (!BBWorklist.empty()); 822 823 return false; 824} 825 826 827/// CanCoerceMustAliasedValueToLoad - Return true if 828/// CoerceAvailableValueToLoadType will succeed. 829static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 830 Type *LoadTy, 831 const DataLayout &DL) { 832 // If the loaded or stored value is an first class array or struct, don't try 833 // to transform them. We need to be able to bitcast to integer. 834 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 835 StoredVal->getType()->isStructTy() || 836 StoredVal->getType()->isArrayTy()) 837 return false; 838 839 // The store has to be at least as big as the load. 840 if (DL.getTypeSizeInBits(StoredVal->getType()) < 841 DL.getTypeSizeInBits(LoadTy)) 842 return false; 843 844 return true; 845} 846 847/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 848/// then a load from a must-aliased pointer of a different type, try to coerce 849/// the stored value. LoadedTy is the type of the load we want to replace and 850/// InsertPt is the place to insert new instructions. 851/// 852/// If we can't do it, return null. 853static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 854 Type *LoadedTy, 855 Instruction *InsertPt, 856 const DataLayout &DL) { 857 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL)) 858 return 0; 859 860 // If this is already the right type, just return it. 861 Type *StoredValTy = StoredVal->getType(); 862 863 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy); 864 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy); 865 866 // If the store and reload are the same size, we can always reuse it. 867 if (StoreSize == LoadSize) { 868 // Pointer to Pointer -> use bitcast. 869 if (StoredValTy->getScalarType()->isPointerTy() && 870 LoadedTy->getScalarType()->isPointerTy()) 871 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 872 873 // Convert source pointers to integers, which can be bitcast. 874 if (StoredValTy->getScalarType()->isPointerTy()) { 875 StoredValTy = DL.getIntPtrType(StoredValTy); 876 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 877 } 878 879 Type *TypeToCastTo = LoadedTy; 880 if (TypeToCastTo->getScalarType()->isPointerTy()) 881 TypeToCastTo = DL.getIntPtrType(TypeToCastTo); 882 883 if (StoredValTy != TypeToCastTo) 884 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 885 886 // Cast to pointer if the load needs a pointer type. 887 if (LoadedTy->getScalarType()->isPointerTy()) 888 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 889 890 return StoredVal; 891 } 892 893 // If the loaded value is smaller than the available value, then we can 894 // extract out a piece from it. If the available value is too small, then we 895 // can't do anything. 896 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 897 898 // Convert source pointers to integers, which can be manipulated. 899 if (StoredValTy->getScalarType()->isPointerTy()) { 900 StoredValTy = DL.getIntPtrType(StoredValTy); 901 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 902 } 903 904 // Convert vectors and fp to integer, which can be manipulated. 905 if (!StoredValTy->isIntegerTy()) { 906 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 907 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 908 } 909 910 // If this is a big-endian system, we need to shift the value down to the low 911 // bits so that a truncate will work. 912 if (DL.isBigEndian()) { 913 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 914 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 915 } 916 917 // Truncate the integer to the right size now. 918 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 919 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 920 921 if (LoadedTy == NewIntTy) 922 return StoredVal; 923 924 // If the result is a pointer, inttoptr. 925 if (LoadedTy->getScalarType()->isPointerTy()) 926 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 927 928 // Otherwise, bitcast. 929 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 930} 931 932/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 933/// memdep query of a load that ends up being a clobbering memory write (store, 934/// memset, memcpy, memmove). This means that the write *may* provide bits used 935/// by the load but we can't be sure because the pointers don't mustalias. 936/// 937/// Check this case to see if there is anything more we can do before we give 938/// up. This returns -1 if we have to give up, or a byte number in the stored 939/// value of the piece that feeds the load. 940static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 941 Value *WritePtr, 942 uint64_t WriteSizeInBits, 943 const DataLayout &DL) { 944 // If the loaded or stored value is a first class array or struct, don't try 945 // to transform them. We need to be able to bitcast to integer. 946 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 947 return -1; 948 949 int64_t StoreOffset = 0, LoadOffset = 0; 950 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL); 951 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL); 952 if (StoreBase != LoadBase) 953 return -1; 954 955 // If the load and store are to the exact same address, they should have been 956 // a must alias. AA must have gotten confused. 957 // FIXME: Study to see if/when this happens. One case is forwarding a memset 958 // to a load from the base of the memset. 959#if 0 960 if (LoadOffset == StoreOffset) { 961 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 962 << "Base = " << *StoreBase << "\n" 963 << "Store Ptr = " << *WritePtr << "\n" 964 << "Store Offs = " << StoreOffset << "\n" 965 << "Load Ptr = " << *LoadPtr << "\n"; 966 abort(); 967 } 968#endif 969 970 // If the load and store don't overlap at all, the store doesn't provide 971 // anything to the load. In this case, they really don't alias at all, AA 972 // must have gotten confused. 973 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy); 974 975 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 976 return -1; 977 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 978 LoadSize >>= 3; 979 980 981 bool isAAFailure = false; 982 if (StoreOffset < LoadOffset) 983 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 984 else 985 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 986 987 if (isAAFailure) { 988#if 0 989 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 990 << "Base = " << *StoreBase << "\n" 991 << "Store Ptr = " << *WritePtr << "\n" 992 << "Store Offs = " << StoreOffset << "\n" 993 << "Load Ptr = " << *LoadPtr << "\n"; 994 abort(); 995#endif 996 return -1; 997 } 998 999 // If the Load isn't completely contained within the stored bits, we don't 1000 // have all the bits to feed it. We could do something crazy in the future 1001 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1002 // valuable. 1003 if (StoreOffset > LoadOffset || 1004 StoreOffset+StoreSize < LoadOffset+LoadSize) 1005 return -1; 1006 1007 // Okay, we can do this transformation. Return the number of bytes into the 1008 // store that the load is. 1009 return LoadOffset-StoreOffset; 1010} 1011 1012/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1013/// memdep query of a load that ends up being a clobbering store. 1014static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 1015 StoreInst *DepSI, 1016 const DataLayout &DL) { 1017 // Cannot handle reading from store of first-class aggregate yet. 1018 if (DepSI->getValueOperand()->getType()->isStructTy() || 1019 DepSI->getValueOperand()->getType()->isArrayTy()) 1020 return -1; 1021 1022 Value *StorePtr = DepSI->getPointerOperand(); 1023 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1024 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1025 StorePtr, StoreSize, DL); 1026} 1027 1028/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 1029/// memdep query of a load that ends up being clobbered by another load. See if 1030/// the other load can feed into the second load. 1031static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 1032 LoadInst *DepLI, const DataLayout &DL){ 1033 // Cannot handle reading from store of first-class aggregate yet. 1034 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 1035 return -1; 1036 1037 Value *DepPtr = DepLI->getPointerOperand(); 1038 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType()); 1039 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL); 1040 if (R != -1) return R; 1041 1042 // If we have a load/load clobber an DepLI can be widened to cover this load, 1043 // then we should widen it! 1044 int64_t LoadOffs = 0; 1045 const Value *LoadBase = 1046 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL); 1047 unsigned LoadSize = DL.getTypeStoreSize(LoadTy); 1048 1049 unsigned Size = MemoryDependenceAnalysis:: 1050 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL); 1051 if (Size == 0) return -1; 1052 1053 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL); 1054} 1055 1056 1057 1058static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 1059 MemIntrinsic *MI, 1060 const DataLayout &DL) { 1061 // If the mem operation is a non-constant size, we can't handle it. 1062 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1063 if (SizeCst == 0) return -1; 1064 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1065 1066 // If this is memset, we just need to see if the offset is valid in the size 1067 // of the memset.. 1068 if (MI->getIntrinsicID() == Intrinsic::memset) 1069 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1070 MemSizeInBits, DL); 1071 1072 // If we have a memcpy/memmove, the only case we can handle is if this is a 1073 // copy from constant memory. In that case, we can read directly from the 1074 // constant memory. 1075 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1076 1077 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1078 if (Src == 0) return -1; 1079 1080 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL)); 1081 if (GV == 0 || !GV->isConstant()) return -1; 1082 1083 // See if the access is within the bounds of the transfer. 1084 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1085 MI->getDest(), MemSizeInBits, DL); 1086 if (Offset == -1) 1087 return Offset; 1088 1089 unsigned AS = Src->getType()->getPointerAddressSpace(); 1090 // Otherwise, see if we can constant fold a load from the constant with the 1091 // offset applied as appropriate. 1092 Src = ConstantExpr::getBitCast(Src, 1093 Type::getInt8PtrTy(Src->getContext(), AS)); 1094 Constant *OffsetCst = 1095 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1096 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1097 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS)); 1098 if (ConstantFoldLoadFromConstPtr(Src, &DL)) 1099 return Offset; 1100 return -1; 1101} 1102 1103 1104/// GetStoreValueForLoad - This function is called when we have a 1105/// memdep query of a load that ends up being a clobbering store. This means 1106/// that the store provides bits used by the load but we the pointers don't 1107/// mustalias. Check this case to see if there is anything more we can do 1108/// before we give up. 1109static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1110 Type *LoadTy, 1111 Instruction *InsertPt, const DataLayout &DL){ 1112 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1113 1114 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1115 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8; 1116 1117 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1118 1119 // Compute which bits of the stored value are being used by the load. Convert 1120 // to an integer type to start with. 1121 if (SrcVal->getType()->getScalarType()->isPointerTy()) 1122 SrcVal = Builder.CreatePtrToInt(SrcVal, 1123 DL.getIntPtrType(SrcVal->getType())); 1124 if (!SrcVal->getType()->isIntegerTy()) 1125 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1126 1127 // Shift the bits to the least significant depending on endianness. 1128 unsigned ShiftAmt; 1129 if (DL.isLittleEndian()) 1130 ShiftAmt = Offset*8; 1131 else 1132 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1133 1134 if (ShiftAmt) 1135 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1136 1137 if (LoadSize != StoreSize) 1138 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1139 1140 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL); 1141} 1142 1143/// GetLoadValueForLoad - This function is called when we have a 1144/// memdep query of a load that ends up being a clobbering load. This means 1145/// that the load *may* provide bits used by the load but we can't be sure 1146/// because the pointers don't mustalias. Check this case to see if there is 1147/// anything more we can do before we give up. 1148static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1149 Type *LoadTy, Instruction *InsertPt, 1150 GVN &gvn) { 1151 const DataLayout &DL = *gvn.getDataLayout(); 1152 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1153 // widen SrcVal out to a larger load. 1154 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType()); 1155 unsigned LoadSize = DL.getTypeStoreSize(LoadTy); 1156 if (Offset+LoadSize > SrcValSize) { 1157 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1158 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1159 // If we have a load/load clobber an DepLI can be widened to cover this 1160 // load, then we should widen it to the next power of 2 size big enough! 1161 unsigned NewLoadSize = Offset+LoadSize; 1162 if (!isPowerOf2_32(NewLoadSize)) 1163 NewLoadSize = NextPowerOf2(NewLoadSize); 1164 1165 Value *PtrVal = SrcVal->getPointerOperand(); 1166 1167 // Insert the new load after the old load. This ensures that subsequent 1168 // memdep queries will find the new load. We can't easily remove the old 1169 // load completely because it is already in the value numbering table. 1170 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1171 Type *DestPTy = 1172 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1173 DestPTy = PointerType::get(DestPTy, 1174 PtrVal->getType()->getPointerAddressSpace()); 1175 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1176 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1177 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1178 NewLoad->takeName(SrcVal); 1179 NewLoad->setAlignment(SrcVal->getAlignment()); 1180 1181 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1182 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1183 1184 // Replace uses of the original load with the wider load. On a big endian 1185 // system, we need to shift down to get the relevant bits. 1186 Value *RV = NewLoad; 1187 if (DL.isBigEndian()) 1188 RV = Builder.CreateLShr(RV, 1189 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1190 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1191 SrcVal->replaceAllUsesWith(RV); 1192 1193 // We would like to use gvn.markInstructionForDeletion here, but we can't 1194 // because the load is already memoized into the leader map table that GVN 1195 // tracks. It is potentially possible to remove the load from the table, 1196 // but then there all of the operations based on it would need to be 1197 // rehashed. Just leave the dead load around. 1198 gvn.getMemDep().removeInstruction(SrcVal); 1199 SrcVal = NewLoad; 1200 } 1201 1202 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL); 1203} 1204 1205 1206/// GetMemInstValueForLoad - This function is called when we have a 1207/// memdep query of a load that ends up being a clobbering mem intrinsic. 1208static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1209 Type *LoadTy, Instruction *InsertPt, 1210 const DataLayout &DL){ 1211 LLVMContext &Ctx = LoadTy->getContext(); 1212 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8; 1213 1214 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1215 1216 // We know that this method is only called when the mem transfer fully 1217 // provides the bits for the load. 1218 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1219 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1220 // independently of what the offset is. 1221 Value *Val = MSI->getValue(); 1222 if (LoadSize != 1) 1223 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1224 1225 Value *OneElt = Val; 1226 1227 // Splat the value out to the right number of bits. 1228 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1229 // If we can double the number of bytes set, do it. 1230 if (NumBytesSet*2 <= LoadSize) { 1231 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1232 Val = Builder.CreateOr(Val, ShVal); 1233 NumBytesSet <<= 1; 1234 continue; 1235 } 1236 1237 // Otherwise insert one byte at a time. 1238 Value *ShVal = Builder.CreateShl(Val, 1*8); 1239 Val = Builder.CreateOr(OneElt, ShVal); 1240 ++NumBytesSet; 1241 } 1242 1243 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL); 1244 } 1245 1246 // Otherwise, this is a memcpy/memmove from a constant global. 1247 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1248 Constant *Src = cast<Constant>(MTI->getSource()); 1249 unsigned AS = Src->getType()->getPointerAddressSpace(); 1250 1251 // Otherwise, see if we can constant fold a load from the constant with the 1252 // offset applied as appropriate. 1253 Src = ConstantExpr::getBitCast(Src, 1254 Type::getInt8PtrTy(Src->getContext(), AS)); 1255 Constant *OffsetCst = 1256 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1257 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1258 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS)); 1259 return ConstantFoldLoadFromConstPtr(Src, &DL); 1260} 1261 1262 1263/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1264/// construct SSA form, allowing us to eliminate LI. This returns the value 1265/// that should be used at LI's definition site. 1266static Value *ConstructSSAForLoadSet(LoadInst *LI, 1267 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1268 GVN &gvn) { 1269 // Check for the fully redundant, dominating load case. In this case, we can 1270 // just use the dominating value directly. 1271 if (ValuesPerBlock.size() == 1 && 1272 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1273 LI->getParent())) { 1274 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block"); 1275 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1276 } 1277 1278 // Otherwise, we have to construct SSA form. 1279 SmallVector<PHINode*, 8> NewPHIs; 1280 SSAUpdater SSAUpdate(&NewPHIs); 1281 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1282 1283 Type *LoadTy = LI->getType(); 1284 1285 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1286 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1287 BasicBlock *BB = AV.BB; 1288 1289 if (SSAUpdate.HasValueForBlock(BB)) 1290 continue; 1291 1292 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1293 } 1294 1295 // Perform PHI construction. 1296 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1297 1298 // If new PHI nodes were created, notify alias analysis. 1299 if (V->getType()->getScalarType()->isPointerTy()) { 1300 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1301 1302 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1303 AA->copyValue(LI, NewPHIs[i]); 1304 1305 // Now that we've copied information to the new PHIs, scan through 1306 // them again and inform alias analysis that we've added potentially 1307 // escaping uses to any values that are operands to these PHIs. 1308 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1309 PHINode *P = NewPHIs[i]; 1310 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1311 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1312 AA->addEscapingUse(P->getOperandUse(jj)); 1313 } 1314 } 1315 } 1316 1317 return V; 1318} 1319 1320Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1321 Value *Res; 1322 if (isSimpleValue()) { 1323 Res = getSimpleValue(); 1324 if (Res->getType() != LoadTy) { 1325 const DataLayout *DL = gvn.getDataLayout(); 1326 assert(DL && "Need target data to handle type mismatch case"); 1327 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1328 *DL); 1329 1330 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1331 << *getSimpleValue() << '\n' 1332 << *Res << '\n' << "\n\n\n"); 1333 } 1334 } else if (isCoercedLoadValue()) { 1335 LoadInst *Load = getCoercedLoadValue(); 1336 if (Load->getType() == LoadTy && Offset == 0) { 1337 Res = Load; 1338 } else { 1339 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1340 gvn); 1341 1342 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1343 << *getCoercedLoadValue() << '\n' 1344 << *Res << '\n' << "\n\n\n"); 1345 } 1346 } else if (isMemIntrinValue()) { 1347 const DataLayout *DL = gvn.getDataLayout(); 1348 assert(DL && "Need target data to handle type mismatch case"); 1349 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1350 LoadTy, BB->getTerminator(), *DL); 1351 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1352 << " " << *getMemIntrinValue() << '\n' 1353 << *Res << '\n' << "\n\n\n"); 1354 } else { 1355 assert(isUndefValue() && "Should be UndefVal"); 1356 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";); 1357 return UndefValue::get(LoadTy); 1358 } 1359 return Res; 1360} 1361 1362static bool isLifetimeStart(const Instruction *Inst) { 1363 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1364 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1365 return false; 1366} 1367 1368void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 1369 AvailValInBlkVect &ValuesPerBlock, 1370 UnavailBlkVect &UnavailableBlocks) { 1371 1372 // Filter out useless results (non-locals, etc). Keep track of the blocks 1373 // where we have a value available in repl, also keep track of whether we see 1374 // dependencies that produce an unknown value for the load (such as a call 1375 // that could potentially clobber the load). 1376 unsigned NumDeps = Deps.size(); 1377 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1378 BasicBlock *DepBB = Deps[i].getBB(); 1379 MemDepResult DepInfo = Deps[i].getResult(); 1380 1381 if (DeadBlocks.count(DepBB)) { 1382 // Dead dependent mem-op disguise as a load evaluating the same value 1383 // as the load in question. 1384 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB)); 1385 continue; 1386 } 1387 1388 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1389 UnavailableBlocks.push_back(DepBB); 1390 continue; 1391 } 1392 1393 if (DepInfo.isClobber()) { 1394 // The address being loaded in this non-local block may not be the same as 1395 // the pointer operand of the load if PHI translation occurs. Make sure 1396 // to consider the right address. 1397 Value *Address = Deps[i].getAddress(); 1398 1399 // If the dependence is to a store that writes to a superset of the bits 1400 // read by the load, we can extract the bits we need for the load from the 1401 // stored value. 1402 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1403 if (DL && Address) { 1404 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1405 DepSI, *DL); 1406 if (Offset != -1) { 1407 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1408 DepSI->getValueOperand(), 1409 Offset)); 1410 continue; 1411 } 1412 } 1413 } 1414 1415 // Check to see if we have something like this: 1416 // load i32* P 1417 // load i8* (P+1) 1418 // if we have this, replace the later with an extraction from the former. 1419 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1420 // If this is a clobber and L is the first instruction in its block, then 1421 // we have the first instruction in the entry block. 1422 if (DepLI != LI && Address && DL) { 1423 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1424 LI->getPointerOperand(), 1425 DepLI, *DL); 1426 1427 if (Offset != -1) { 1428 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1429 Offset)); 1430 continue; 1431 } 1432 } 1433 } 1434 1435 // If the clobbering value is a memset/memcpy/memmove, see if we can 1436 // forward a value on from it. 1437 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1438 if (DL && Address) { 1439 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1440 DepMI, *DL); 1441 if (Offset != -1) { 1442 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1443 Offset)); 1444 continue; 1445 } 1446 } 1447 } 1448 1449 UnavailableBlocks.push_back(DepBB); 1450 continue; 1451 } 1452 1453 // DepInfo.isDef() here 1454 1455 Instruction *DepInst = DepInfo.getInst(); 1456 1457 // Loading the allocation -> undef. 1458 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1459 // Loading immediately after lifetime begin -> undef. 1460 isLifetimeStart(DepInst)) { 1461 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1462 UndefValue::get(LI->getType()))); 1463 continue; 1464 } 1465 1466 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1467 // Reject loads and stores that are to the same address but are of 1468 // different types if we have to. 1469 if (S->getValueOperand()->getType() != LI->getType()) { 1470 // If the stored value is larger or equal to the loaded value, we can 1471 // reuse it. 1472 if (DL == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1473 LI->getType(), *DL)) { 1474 UnavailableBlocks.push_back(DepBB); 1475 continue; 1476 } 1477 } 1478 1479 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1480 S->getValueOperand())); 1481 continue; 1482 } 1483 1484 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1485 // If the types mismatch and we can't handle it, reject reuse of the load. 1486 if (LD->getType() != LI->getType()) { 1487 // If the stored value is larger or equal to the loaded value, we can 1488 // reuse it. 1489 if (DL == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)){ 1490 UnavailableBlocks.push_back(DepBB); 1491 continue; 1492 } 1493 } 1494 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1495 continue; 1496 } 1497 1498 UnavailableBlocks.push_back(DepBB); 1499 } 1500} 1501 1502bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 1503 UnavailBlkVect &UnavailableBlocks) { 1504 // Okay, we have *some* definitions of the value. This means that the value 1505 // is available in some of our (transitive) predecessors. Lets think about 1506 // doing PRE of this load. This will involve inserting a new load into the 1507 // predecessor when it's not available. We could do this in general, but 1508 // prefer to not increase code size. As such, we only do this when we know 1509 // that we only have to insert *one* load (which means we're basically moving 1510 // the load, not inserting a new one). 1511 1512 SmallPtrSet<BasicBlock *, 4> Blockers; 1513 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1514 Blockers.insert(UnavailableBlocks[i]); 1515 1516 // Let's find the first basic block with more than one predecessor. Walk 1517 // backwards through predecessors if needed. 1518 BasicBlock *LoadBB = LI->getParent(); 1519 BasicBlock *TmpBB = LoadBB; 1520 1521 while (TmpBB->getSinglePredecessor()) { 1522 TmpBB = TmpBB->getSinglePredecessor(); 1523 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1524 return false; 1525 if (Blockers.count(TmpBB)) 1526 return false; 1527 1528 // If any of these blocks has more than one successor (i.e. if the edge we 1529 // just traversed was critical), then there are other paths through this 1530 // block along which the load may not be anticipated. Hoisting the load 1531 // above this block would be adding the load to execution paths along 1532 // which it was not previously executed. 1533 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1534 return false; 1535 } 1536 1537 assert(TmpBB); 1538 LoadBB = TmpBB; 1539 1540 // Check to see how many predecessors have the loaded value fully 1541 // available. 1542 DenseMap<BasicBlock*, Value*> PredLoads; 1543 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1544 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1545 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1546 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1547 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1548 1549 SmallVector<BasicBlock *, 4> CriticalEdgePred; 1550 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1551 PI != E; ++PI) { 1552 BasicBlock *Pred = *PI; 1553 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1554 continue; 1555 } 1556 PredLoads[Pred] = 0; 1557 1558 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1559 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1560 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1561 << Pred->getName() << "': " << *LI << '\n'); 1562 return false; 1563 } 1564 1565 if (LoadBB->isLandingPad()) { 1566 DEBUG(dbgs() 1567 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1568 << Pred->getName() << "': " << *LI << '\n'); 1569 return false; 1570 } 1571 1572 CriticalEdgePred.push_back(Pred); 1573 } 1574 } 1575 1576 // Decide whether PRE is profitable for this load. 1577 unsigned NumUnavailablePreds = PredLoads.size(); 1578 assert(NumUnavailablePreds != 0 && 1579 "Fully available value should already be eliminated!"); 1580 1581 // If this load is unavailable in multiple predecessors, reject it. 1582 // FIXME: If we could restructure the CFG, we could make a common pred with 1583 // all the preds that don't have an available LI and insert a new load into 1584 // that one block. 1585 if (NumUnavailablePreds != 1) 1586 return false; 1587 1588 // Split critical edges, and update the unavailable predecessors accordingly. 1589 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(), 1590 E = CriticalEdgePred.end(); I != E; I++) { 1591 BasicBlock *OrigPred = *I; 1592 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); 1593 PredLoads.erase(OrigPred); 1594 PredLoads[NewPred] = 0; 1595 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" 1596 << LoadBB->getName() << '\n'); 1597 } 1598 1599 // Check if the load can safely be moved to all the unavailable predecessors. 1600 bool CanDoPRE = true; 1601 SmallVector<Instruction*, 8> NewInsts; 1602 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1603 E = PredLoads.end(); I != E; ++I) { 1604 BasicBlock *UnavailablePred = I->first; 1605 1606 // Do PHI translation to get its value in the predecessor if necessary. The 1607 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1608 1609 // If all preds have a single successor, then we know it is safe to insert 1610 // the load on the pred (?!?), so we can insert code to materialize the 1611 // pointer if it is not available. 1612 PHITransAddr Address(LI->getPointerOperand(), DL); 1613 Value *LoadPtr = 0; 1614 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1615 *DT, NewInsts); 1616 1617 // If we couldn't find or insert a computation of this phi translated value, 1618 // we fail PRE. 1619 if (LoadPtr == 0) { 1620 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1621 << *LI->getPointerOperand() << "\n"); 1622 CanDoPRE = false; 1623 break; 1624 } 1625 1626 I->second = LoadPtr; 1627 } 1628 1629 if (!CanDoPRE) { 1630 while (!NewInsts.empty()) { 1631 Instruction *I = NewInsts.pop_back_val(); 1632 if (MD) MD->removeInstruction(I); 1633 I->eraseFromParent(); 1634 } 1635 // HINT:Don't revert the edge-splitting as following transformation may 1636 // also need to split these critial edges. 1637 return !CriticalEdgePred.empty(); 1638 } 1639 1640 // Okay, we can eliminate this load by inserting a reload in the predecessor 1641 // and using PHI construction to get the value in the other predecessors, do 1642 // it. 1643 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1644 DEBUG(if (!NewInsts.empty()) 1645 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1646 << *NewInsts.back() << '\n'); 1647 1648 // Assign value numbers to the new instructions. 1649 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1650 // FIXME: We really _ought_ to insert these value numbers into their 1651 // parent's availability map. However, in doing so, we risk getting into 1652 // ordering issues. If a block hasn't been processed yet, we would be 1653 // marking a value as AVAIL-IN, which isn't what we intend. 1654 VN.lookup_or_add(NewInsts[i]); 1655 } 1656 1657 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1658 E = PredLoads.end(); I != E; ++I) { 1659 BasicBlock *UnavailablePred = I->first; 1660 Value *LoadPtr = I->second; 1661 1662 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1663 LI->getAlignment(), 1664 UnavailablePred->getTerminator()); 1665 1666 // Transfer the old load's TBAA tag to the new load. 1667 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1668 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1669 1670 // Transfer DebugLoc. 1671 NewLoad->setDebugLoc(LI->getDebugLoc()); 1672 1673 // Add the newly created load. 1674 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1675 NewLoad)); 1676 MD->invalidateCachedPointerInfo(LoadPtr); 1677 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1678 } 1679 1680 // Perform PHI construction. 1681 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1682 LI->replaceAllUsesWith(V); 1683 if (isa<PHINode>(V)) 1684 V->takeName(LI); 1685 if (V->getType()->getScalarType()->isPointerTy()) 1686 MD->invalidateCachedPointerInfo(V); 1687 markInstructionForDeletion(LI); 1688 ++NumPRELoad; 1689 return true; 1690} 1691 1692/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1693/// non-local by performing PHI construction. 1694bool GVN::processNonLocalLoad(LoadInst *LI) { 1695 // Step 1: Find the non-local dependencies of the load. 1696 LoadDepVect Deps; 1697 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1698 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1699 1700 // If we had to process more than one hundred blocks to find the 1701 // dependencies, this load isn't worth worrying about. Optimizing 1702 // it will be too expensive. 1703 unsigned NumDeps = Deps.size(); 1704 if (NumDeps > 100) 1705 return false; 1706 1707 // If we had a phi translation failure, we'll have a single entry which is a 1708 // clobber in the current block. Reject this early. 1709 if (NumDeps == 1 && 1710 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1711 DEBUG( 1712 dbgs() << "GVN: non-local load "; 1713 LI->printAsOperand(dbgs()); 1714 dbgs() << " has unknown dependencies\n"; 1715 ); 1716 return false; 1717 } 1718 1719 // Step 2: Analyze the availability of the load 1720 AvailValInBlkVect ValuesPerBlock; 1721 UnavailBlkVect UnavailableBlocks; 1722 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks); 1723 1724 // If we have no predecessors that produce a known value for this load, exit 1725 // early. 1726 if (ValuesPerBlock.empty()) 1727 return false; 1728 1729 // Step 3: Eliminate fully redundancy. 1730 // 1731 // If all of the instructions we depend on produce a known value for this 1732 // load, then it is fully redundant and we can use PHI insertion to compute 1733 // its value. Insert PHIs and remove the fully redundant value now. 1734 if (UnavailableBlocks.empty()) { 1735 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1736 1737 // Perform PHI construction. 1738 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1739 LI->replaceAllUsesWith(V); 1740 1741 if (isa<PHINode>(V)) 1742 V->takeName(LI); 1743 if (V->getType()->getScalarType()->isPointerTy()) 1744 MD->invalidateCachedPointerInfo(V); 1745 markInstructionForDeletion(LI); 1746 ++NumGVNLoad; 1747 return true; 1748 } 1749 1750 // Step 4: Eliminate partial redundancy. 1751 if (!EnablePRE || !EnableLoadPRE) 1752 return false; 1753 1754 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks); 1755} 1756 1757 1758static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1759 // Patch the replacement so that it is not more restrictive than the value 1760 // being replaced. 1761 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1762 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1763 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1764 isa<OverflowingBinaryOperator>(ReplOp)) { 1765 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1766 ReplOp->setHasNoSignedWrap(false); 1767 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1768 ReplOp->setHasNoUnsignedWrap(false); 1769 } 1770 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1771 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1772 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1773 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1774 unsigned Kind = Metadata[i].first; 1775 MDNode *IMD = I->getMetadata(Kind); 1776 MDNode *ReplMD = Metadata[i].second; 1777 switch(Kind) { 1778 default: 1779 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1780 break; 1781 case LLVMContext::MD_dbg: 1782 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1783 case LLVMContext::MD_tbaa: 1784 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1785 break; 1786 case LLVMContext::MD_range: 1787 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1788 break; 1789 case LLVMContext::MD_prof: 1790 llvm_unreachable("MD_prof in a non-terminator instruction"); 1791 break; 1792 case LLVMContext::MD_fpmath: 1793 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1794 break; 1795 } 1796 } 1797 } 1798} 1799 1800static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1801 patchReplacementInstruction(I, Repl); 1802 I->replaceAllUsesWith(Repl); 1803} 1804 1805/// processLoad - Attempt to eliminate a load, first by eliminating it 1806/// locally, and then attempting non-local elimination if that fails. 1807bool GVN::processLoad(LoadInst *L) { 1808 if (!MD) 1809 return false; 1810 1811 if (!L->isSimple()) 1812 return false; 1813 1814 if (L->use_empty()) { 1815 markInstructionForDeletion(L); 1816 return true; 1817 } 1818 1819 // ... to a pointer that has been loaded from before... 1820 MemDepResult Dep = MD->getDependency(L); 1821 1822 // If we have a clobber and target data is around, see if this is a clobber 1823 // that we can fix up through code synthesis. 1824 if (Dep.isClobber() && DL) { 1825 // Check to see if we have something like this: 1826 // store i32 123, i32* %P 1827 // %A = bitcast i32* %P to i8* 1828 // %B = gep i8* %A, i32 1 1829 // %C = load i8* %B 1830 // 1831 // We could do that by recognizing if the clobber instructions are obviously 1832 // a common base + constant offset, and if the previous store (or memset) 1833 // completely covers this load. This sort of thing can happen in bitfield 1834 // access code. 1835 Value *AvailVal = 0; 1836 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1837 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1838 L->getPointerOperand(), 1839 DepSI, *DL); 1840 if (Offset != -1) 1841 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1842 L->getType(), L, *DL); 1843 } 1844 1845 // Check to see if we have something like this: 1846 // load i32* P 1847 // load i8* (P+1) 1848 // if we have this, replace the later with an extraction from the former. 1849 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1850 // If this is a clobber and L is the first instruction in its block, then 1851 // we have the first instruction in the entry block. 1852 if (DepLI == L) 1853 return false; 1854 1855 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1856 L->getPointerOperand(), 1857 DepLI, *DL); 1858 if (Offset != -1) 1859 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1860 } 1861 1862 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1863 // a value on from it. 1864 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1865 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1866 L->getPointerOperand(), 1867 DepMI, *DL); 1868 if (Offset != -1) 1869 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL); 1870 } 1871 1872 if (AvailVal) { 1873 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1874 << *AvailVal << '\n' << *L << "\n\n\n"); 1875 1876 // Replace the load! 1877 L->replaceAllUsesWith(AvailVal); 1878 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1879 MD->invalidateCachedPointerInfo(AvailVal); 1880 markInstructionForDeletion(L); 1881 ++NumGVNLoad; 1882 return true; 1883 } 1884 } 1885 1886 // If the value isn't available, don't do anything! 1887 if (Dep.isClobber()) { 1888 DEBUG( 1889 // fast print dep, using operator<< on instruction is too slow. 1890 dbgs() << "GVN: load "; 1891 L->printAsOperand(dbgs()); 1892 Instruction *I = Dep.getInst(); 1893 dbgs() << " is clobbered by " << *I << '\n'; 1894 ); 1895 return false; 1896 } 1897 1898 // If it is defined in another block, try harder. 1899 if (Dep.isNonLocal()) 1900 return processNonLocalLoad(L); 1901 1902 if (!Dep.isDef()) { 1903 DEBUG( 1904 // fast print dep, using operator<< on instruction is too slow. 1905 dbgs() << "GVN: load "; 1906 L->printAsOperand(dbgs()); 1907 dbgs() << " has unknown dependence\n"; 1908 ); 1909 return false; 1910 } 1911 1912 Instruction *DepInst = Dep.getInst(); 1913 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1914 Value *StoredVal = DepSI->getValueOperand(); 1915 1916 // The store and load are to a must-aliased pointer, but they may not 1917 // actually have the same type. See if we know how to reuse the stored 1918 // value (depending on its type). 1919 if (StoredVal->getType() != L->getType()) { 1920 if (DL) { 1921 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1922 L, *DL); 1923 if (StoredVal == 0) 1924 return false; 1925 1926 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1927 << '\n' << *L << "\n\n\n"); 1928 } 1929 else 1930 return false; 1931 } 1932 1933 // Remove it! 1934 L->replaceAllUsesWith(StoredVal); 1935 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1936 MD->invalidateCachedPointerInfo(StoredVal); 1937 markInstructionForDeletion(L); 1938 ++NumGVNLoad; 1939 return true; 1940 } 1941 1942 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1943 Value *AvailableVal = DepLI; 1944 1945 // The loads are of a must-aliased pointer, but they may not actually have 1946 // the same type. See if we know how to reuse the previously loaded value 1947 // (depending on its type). 1948 if (DepLI->getType() != L->getType()) { 1949 if (DL) { 1950 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1951 L, *DL); 1952 if (AvailableVal == 0) 1953 return false; 1954 1955 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1956 << "\n" << *L << "\n\n\n"); 1957 } 1958 else 1959 return false; 1960 } 1961 1962 // Remove it! 1963 patchAndReplaceAllUsesWith(L, AvailableVal); 1964 if (DepLI->getType()->getScalarType()->isPointerTy()) 1965 MD->invalidateCachedPointerInfo(DepLI); 1966 markInstructionForDeletion(L); 1967 ++NumGVNLoad; 1968 return true; 1969 } 1970 1971 // If this load really doesn't depend on anything, then we must be loading an 1972 // undef value. This can happen when loading for a fresh allocation with no 1973 // intervening stores, for example. 1974 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1975 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1976 markInstructionForDeletion(L); 1977 ++NumGVNLoad; 1978 return true; 1979 } 1980 1981 // If this load occurs either right after a lifetime begin, 1982 // then the loaded value is undefined. 1983 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1984 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1985 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1986 markInstructionForDeletion(L); 1987 ++NumGVNLoad; 1988 return true; 1989 } 1990 } 1991 1992 return false; 1993} 1994 1995// findLeader - In order to find a leader for a given value number at a 1996// specific basic block, we first obtain the list of all Values for that number, 1997// and then scan the list to find one whose block dominates the block in 1998// question. This is fast because dominator tree queries consist of only 1999// a few comparisons of DFS numbers. 2000Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 2001 LeaderTableEntry Vals = LeaderTable[num]; 2002 if (!Vals.Val) return 0; 2003 2004 Value *Val = 0; 2005 if (DT->dominates(Vals.BB, BB)) { 2006 Val = Vals.Val; 2007 if (isa<Constant>(Val)) return Val; 2008 } 2009 2010 LeaderTableEntry* Next = Vals.Next; 2011 while (Next) { 2012 if (DT->dominates(Next->BB, BB)) { 2013 if (isa<Constant>(Next->Val)) return Next->Val; 2014 if (!Val) Val = Next->Val; 2015 } 2016 2017 Next = Next->Next; 2018 } 2019 2020 return Val; 2021} 2022 2023/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 2024/// use is dominated by the given basic block. Returns the number of uses that 2025/// were replaced. 2026unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 2027 const BasicBlockEdge &Root) { 2028 unsigned Count = 0; 2029 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2030 UI != UE; ) { 2031 Use &U = *UI++; 2032 2033 if (DT->dominates(Root, U)) { 2034 U.set(To); 2035 ++Count; 2036 } 2037 } 2038 return Count; 2039} 2040 2041/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2042/// true if every path from the entry block to 'Dst' passes via this edge. In 2043/// particular 'Dst' must not be reachable via another edge from 'Src'. 2044static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2045 DominatorTree *DT) { 2046 // While in theory it is interesting to consider the case in which Dst has 2047 // more than one predecessor, because Dst might be part of a loop which is 2048 // only reachable from Src, in practice it is pointless since at the time 2049 // GVN runs all such loops have preheaders, which means that Dst will have 2050 // been changed to have only one predecessor, namely Src. 2051 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2052 const BasicBlock *Src = E.getStart(); 2053 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2054 (void)Src; 2055 return Pred != 0; 2056} 2057 2058/// propagateEquality - The given values are known to be equal in every block 2059/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2060/// 'RHS' everywhere in the scope. Returns whether a change was made. 2061bool GVN::propagateEquality(Value *LHS, Value *RHS, 2062 const BasicBlockEdge &Root) { 2063 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2064 Worklist.push_back(std::make_pair(LHS, RHS)); 2065 bool Changed = false; 2066 // For speed, compute a conservative fast approximation to 2067 // DT->dominates(Root, Root.getEnd()); 2068 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2069 2070 while (!Worklist.empty()) { 2071 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2072 LHS = Item.first; RHS = Item.second; 2073 2074 if (LHS == RHS) continue; 2075 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2076 2077 // Don't try to propagate equalities between constants. 2078 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2079 2080 // Prefer a constant on the right-hand side, or an Argument if no constants. 2081 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2082 std::swap(LHS, RHS); 2083 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2084 2085 // If there is no obvious reason to prefer the left-hand side over the right- 2086 // hand side, ensure the longest lived term is on the right-hand side, so the 2087 // shortest lived term will be replaced by the longest lived. This tends to 2088 // expose more simplifications. 2089 uint32_t LVN = VN.lookup_or_add(LHS); 2090 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2091 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2092 // Move the 'oldest' value to the right-hand side, using the value number as 2093 // a proxy for age. 2094 uint32_t RVN = VN.lookup_or_add(RHS); 2095 if (LVN < RVN) { 2096 std::swap(LHS, RHS); 2097 LVN = RVN; 2098 } 2099 } 2100 2101 // If value numbering later sees that an instruction in the scope is equal 2102 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2103 // the invariant that instructions only occur in the leader table for their 2104 // own value number (this is used by removeFromLeaderTable), do not do this 2105 // if RHS is an instruction (if an instruction in the scope is morphed into 2106 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2107 // using the leader table is about compiling faster, not optimizing better). 2108 // The leader table only tracks basic blocks, not edges. Only add to if we 2109 // have the simple case where the edge dominates the end. 2110 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2111 addToLeaderTable(LVN, RHS, Root.getEnd()); 2112 2113 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2114 // LHS always has at least one use that is not dominated by Root, this will 2115 // never do anything if LHS has only one use. 2116 if (!LHS->hasOneUse()) { 2117 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2118 Changed |= NumReplacements > 0; 2119 NumGVNEqProp += NumReplacements; 2120 } 2121 2122 // Now try to deduce additional equalities from this one. For example, if the 2123 // known equality was "(A != B)" == "false" then it follows that A and B are 2124 // equal in the scope. Only boolean equalities with an explicit true or false 2125 // RHS are currently supported. 2126 if (!RHS->getType()->isIntegerTy(1)) 2127 // Not a boolean equality - bail out. 2128 continue; 2129 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2130 if (!CI) 2131 // RHS neither 'true' nor 'false' - bail out. 2132 continue; 2133 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2134 bool isKnownTrue = CI->isAllOnesValue(); 2135 bool isKnownFalse = !isKnownTrue; 2136 2137 // If "A && B" is known true then both A and B are known true. If "A || B" 2138 // is known false then both A and B are known false. 2139 Value *A, *B; 2140 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2141 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2142 Worklist.push_back(std::make_pair(A, RHS)); 2143 Worklist.push_back(std::make_pair(B, RHS)); 2144 continue; 2145 } 2146 2147 // If we are propagating an equality like "(A == B)" == "true" then also 2148 // propagate the equality A == B. When propagating a comparison such as 2149 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2150 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2151 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2152 2153 // If "A == B" is known true, or "A != B" is known false, then replace 2154 // A with B everywhere in the scope. 2155 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2156 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2157 Worklist.push_back(std::make_pair(Op0, Op1)); 2158 2159 // If "A >= B" is known true, replace "A < B" with false everywhere. 2160 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2161 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2162 // Since we don't have the instruction "A < B" immediately to hand, work out 2163 // the value number that it would have and use that to find an appropriate 2164 // instruction (if any). 2165 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2166 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2167 // If the number we were assigned was brand new then there is no point in 2168 // looking for an instruction realizing it: there cannot be one! 2169 if (Num < NextNum) { 2170 Value *NotCmp = findLeader(Root.getEnd(), Num); 2171 if (NotCmp && isa<Instruction>(NotCmp)) { 2172 unsigned NumReplacements = 2173 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2174 Changed |= NumReplacements > 0; 2175 NumGVNEqProp += NumReplacements; 2176 } 2177 } 2178 // Ensure that any instruction in scope that gets the "A < B" value number 2179 // is replaced with false. 2180 // The leader table only tracks basic blocks, not edges. Only add to if we 2181 // have the simple case where the edge dominates the end. 2182 if (RootDominatesEnd) 2183 addToLeaderTable(Num, NotVal, Root.getEnd()); 2184 2185 continue; 2186 } 2187 } 2188 2189 return Changed; 2190} 2191 2192/// processInstruction - When calculating availability, handle an instruction 2193/// by inserting it into the appropriate sets 2194bool GVN::processInstruction(Instruction *I) { 2195 // Ignore dbg info intrinsics. 2196 if (isa<DbgInfoIntrinsic>(I)) 2197 return false; 2198 2199 // If the instruction can be easily simplified then do so now in preference 2200 // to value numbering it. Value numbering often exposes redundancies, for 2201 // example if it determines that %y is equal to %x then the instruction 2202 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2203 if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) { 2204 I->replaceAllUsesWith(V); 2205 if (MD && V->getType()->getScalarType()->isPointerTy()) 2206 MD->invalidateCachedPointerInfo(V); 2207 markInstructionForDeletion(I); 2208 ++NumGVNSimpl; 2209 return true; 2210 } 2211 2212 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2213 if (processLoad(LI)) 2214 return true; 2215 2216 unsigned Num = VN.lookup_or_add(LI); 2217 addToLeaderTable(Num, LI, LI->getParent()); 2218 return false; 2219 } 2220 2221 // For conditional branches, we can perform simple conditional propagation on 2222 // the condition value itself. 2223 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2224 if (!BI->isConditional()) 2225 return false; 2226 2227 if (isa<Constant>(BI->getCondition())) 2228 return processFoldableCondBr(BI); 2229 2230 Value *BranchCond = BI->getCondition(); 2231 BasicBlock *TrueSucc = BI->getSuccessor(0); 2232 BasicBlock *FalseSucc = BI->getSuccessor(1); 2233 // Avoid multiple edges early. 2234 if (TrueSucc == FalseSucc) 2235 return false; 2236 2237 BasicBlock *Parent = BI->getParent(); 2238 bool Changed = false; 2239 2240 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2241 BasicBlockEdge TrueE(Parent, TrueSucc); 2242 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2243 2244 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2245 BasicBlockEdge FalseE(Parent, FalseSucc); 2246 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2247 2248 return Changed; 2249 } 2250 2251 // For switches, propagate the case values into the case destinations. 2252 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2253 Value *SwitchCond = SI->getCondition(); 2254 BasicBlock *Parent = SI->getParent(); 2255 bool Changed = false; 2256 2257 // Remember how many outgoing edges there are to every successor. 2258 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2259 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2260 ++SwitchEdges[SI->getSuccessor(i)]; 2261 2262 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2263 i != e; ++i) { 2264 BasicBlock *Dst = i.getCaseSuccessor(); 2265 // If there is only a single edge, propagate the case value into it. 2266 if (SwitchEdges.lookup(Dst) == 1) { 2267 BasicBlockEdge E(Parent, Dst); 2268 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2269 } 2270 } 2271 return Changed; 2272 } 2273 2274 // Instructions with void type don't return a value, so there's 2275 // no point in trying to find redundancies in them. 2276 if (I->getType()->isVoidTy()) return false; 2277 2278 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2279 unsigned Num = VN.lookup_or_add(I); 2280 2281 // Allocations are always uniquely numbered, so we can save time and memory 2282 // by fast failing them. 2283 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2284 addToLeaderTable(Num, I, I->getParent()); 2285 return false; 2286 } 2287 2288 // If the number we were assigned was a brand new VN, then we don't 2289 // need to do a lookup to see if the number already exists 2290 // somewhere in the domtree: it can't! 2291 if (Num >= NextNum) { 2292 addToLeaderTable(Num, I, I->getParent()); 2293 return false; 2294 } 2295 2296 // Perform fast-path value-number based elimination of values inherited from 2297 // dominators. 2298 Value *repl = findLeader(I->getParent(), Num); 2299 if (repl == 0) { 2300 // Failure, just remember this instance for future use. 2301 addToLeaderTable(Num, I, I->getParent()); 2302 return false; 2303 } 2304 2305 // Remove it! 2306 patchAndReplaceAllUsesWith(I, repl); 2307 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2308 MD->invalidateCachedPointerInfo(repl); 2309 markInstructionForDeletion(I); 2310 return true; 2311} 2312 2313/// runOnFunction - This is the main transformation entry point for a function. 2314bool GVN::runOnFunction(Function& F) { 2315 if (skipOptnoneFunction(F)) 2316 return false; 2317 2318 if (!NoLoads) 2319 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2320 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2321 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 2322 DL = DLP ? &DLP->getDataLayout() : 0; 2323 TLI = &getAnalysis<TargetLibraryInfo>(); 2324 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2325 VN.setMemDep(MD); 2326 VN.setDomTree(DT); 2327 2328 bool Changed = false; 2329 bool ShouldContinue = true; 2330 2331 // Merge unconditional branches, allowing PRE to catch more 2332 // optimization opportunities. 2333 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2334 BasicBlock *BB = FI++; 2335 2336 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2337 if (removedBlock) ++NumGVNBlocks; 2338 2339 Changed |= removedBlock; 2340 } 2341 2342 unsigned Iteration = 0; 2343 while (ShouldContinue) { 2344 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2345 ShouldContinue = iterateOnFunction(F); 2346 Changed |= ShouldContinue; 2347 ++Iteration; 2348 } 2349 2350 if (EnablePRE) { 2351 // Fabricate val-num for dead-code in order to suppress assertion in 2352 // performPRE(). 2353 assignValNumForDeadCode(); 2354 bool PREChanged = true; 2355 while (PREChanged) { 2356 PREChanged = performPRE(F); 2357 Changed |= PREChanged; 2358 } 2359 } 2360 2361 // FIXME: Should perform GVN again after PRE does something. PRE can move 2362 // computations into blocks where they become fully redundant. Note that 2363 // we can't do this until PRE's critical edge splitting updates memdep. 2364 // Actually, when this happens, we should just fully integrate PRE into GVN. 2365 2366 cleanupGlobalSets(); 2367 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each 2368 // iteration. 2369 DeadBlocks.clear(); 2370 2371 return Changed; 2372} 2373 2374 2375bool GVN::processBlock(BasicBlock *BB) { 2376 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2377 // (and incrementing BI before processing an instruction). 2378 assert(InstrsToErase.empty() && 2379 "We expect InstrsToErase to be empty across iterations"); 2380 if (DeadBlocks.count(BB)) 2381 return false; 2382 2383 bool ChangedFunction = false; 2384 2385 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2386 BI != BE;) { 2387 ChangedFunction |= processInstruction(BI); 2388 if (InstrsToErase.empty()) { 2389 ++BI; 2390 continue; 2391 } 2392 2393 // If we need some instructions deleted, do it now. 2394 NumGVNInstr += InstrsToErase.size(); 2395 2396 // Avoid iterator invalidation. 2397 bool AtStart = BI == BB->begin(); 2398 if (!AtStart) 2399 --BI; 2400 2401 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(), 2402 E = InstrsToErase.end(); I != E; ++I) { 2403 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2404 if (MD) MD->removeInstruction(*I); 2405 DEBUG(verifyRemoved(*I)); 2406 (*I)->eraseFromParent(); 2407 } 2408 InstrsToErase.clear(); 2409 2410 if (AtStart) 2411 BI = BB->begin(); 2412 else 2413 ++BI; 2414 } 2415 2416 return ChangedFunction; 2417} 2418 2419/// performPRE - Perform a purely local form of PRE that looks for diamond 2420/// control flow patterns and attempts to perform simple PRE at the join point. 2421bool GVN::performPRE(Function &F) { 2422 bool Changed = false; 2423 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2424 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2425 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2426 BasicBlock *CurrentBlock = *DI; 2427 2428 // Nothing to PRE in the entry block. 2429 if (CurrentBlock == &F.getEntryBlock()) continue; 2430 2431 // Don't perform PRE on a landing pad. 2432 if (CurrentBlock->isLandingPad()) continue; 2433 2434 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2435 BE = CurrentBlock->end(); BI != BE; ) { 2436 Instruction *CurInst = BI++; 2437 2438 if (isa<AllocaInst>(CurInst) || 2439 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2440 CurInst->getType()->isVoidTy() || 2441 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2442 isa<DbgInfoIntrinsic>(CurInst)) 2443 continue; 2444 2445 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2446 // sinking the compare again, and it would force the code generator to 2447 // move the i1 from processor flags or predicate registers into a general 2448 // purpose register. 2449 if (isa<CmpInst>(CurInst)) 2450 continue; 2451 2452 // We don't currently value number ANY inline asm calls. 2453 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2454 if (CallI->isInlineAsm()) 2455 continue; 2456 2457 uint32_t ValNo = VN.lookup(CurInst); 2458 2459 // Look for the predecessors for PRE opportunities. We're 2460 // only trying to solve the basic diamond case, where 2461 // a value is computed in the successor and one predecessor, 2462 // but not the other. We also explicitly disallow cases 2463 // where the successor is its own predecessor, because they're 2464 // more complicated to get right. 2465 unsigned NumWith = 0; 2466 unsigned NumWithout = 0; 2467 BasicBlock *PREPred = 0; 2468 predMap.clear(); 2469 2470 for (pred_iterator PI = pred_begin(CurrentBlock), 2471 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2472 BasicBlock *P = *PI; 2473 // We're not interested in PRE where the block is its 2474 // own predecessor, or in blocks with predecessors 2475 // that are not reachable. 2476 if (P == CurrentBlock) { 2477 NumWithout = 2; 2478 break; 2479 } else if (!DT->isReachableFromEntry(P)) { 2480 NumWithout = 2; 2481 break; 2482 } 2483 2484 Value* predV = findLeader(P, ValNo); 2485 if (predV == 0) { 2486 predMap.push_back(std::make_pair(static_cast<Value *>(0), P)); 2487 PREPred = P; 2488 ++NumWithout; 2489 } else if (predV == CurInst) { 2490 /* CurInst dominates this predecessor. */ 2491 NumWithout = 2; 2492 break; 2493 } else { 2494 predMap.push_back(std::make_pair(predV, P)); 2495 ++NumWith; 2496 } 2497 } 2498 2499 // Don't do PRE when it might increase code size, i.e. when 2500 // we would need to insert instructions in more than one pred. 2501 if (NumWithout != 1 || NumWith == 0) 2502 continue; 2503 2504 // Don't do PRE across indirect branch. 2505 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2506 continue; 2507 2508 // We can't do PRE safely on a critical edge, so instead we schedule 2509 // the edge to be split and perform the PRE the next time we iterate 2510 // on the function. 2511 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2512 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2513 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2514 continue; 2515 } 2516 2517 // Instantiate the expression in the predecessor that lacked it. 2518 // Because we are going top-down through the block, all value numbers 2519 // will be available in the predecessor by the time we need them. Any 2520 // that weren't originally present will have been instantiated earlier 2521 // in this loop. 2522 Instruction *PREInstr = CurInst->clone(); 2523 bool success = true; 2524 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2525 Value *Op = PREInstr->getOperand(i); 2526 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2527 continue; 2528 2529 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2530 PREInstr->setOperand(i, V); 2531 } else { 2532 success = false; 2533 break; 2534 } 2535 } 2536 2537 // Fail out if we encounter an operand that is not available in 2538 // the PRE predecessor. This is typically because of loads which 2539 // are not value numbered precisely. 2540 if (!success) { 2541 DEBUG(verifyRemoved(PREInstr)); 2542 delete PREInstr; 2543 continue; 2544 } 2545 2546 PREInstr->insertBefore(PREPred->getTerminator()); 2547 PREInstr->setName(CurInst->getName() + ".pre"); 2548 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2549 VN.add(PREInstr, ValNo); 2550 ++NumGVNPRE; 2551 2552 // Update the availability map to include the new instruction. 2553 addToLeaderTable(ValNo, PREInstr, PREPred); 2554 2555 // Create a PHI to make the value available in this block. 2556 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2557 CurInst->getName() + ".pre-phi", 2558 CurrentBlock->begin()); 2559 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2560 if (Value *V = predMap[i].first) 2561 Phi->addIncoming(V, predMap[i].second); 2562 else 2563 Phi->addIncoming(PREInstr, PREPred); 2564 } 2565 2566 VN.add(Phi, ValNo); 2567 addToLeaderTable(ValNo, Phi, CurrentBlock); 2568 Phi->setDebugLoc(CurInst->getDebugLoc()); 2569 CurInst->replaceAllUsesWith(Phi); 2570 if (Phi->getType()->getScalarType()->isPointerTy()) { 2571 // Because we have added a PHI-use of the pointer value, it has now 2572 // "escaped" from alias analysis' perspective. We need to inform 2573 // AA of this. 2574 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2575 ++ii) { 2576 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2577 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2578 } 2579 2580 if (MD) 2581 MD->invalidateCachedPointerInfo(Phi); 2582 } 2583 VN.erase(CurInst); 2584 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2585 2586 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2587 if (MD) MD->removeInstruction(CurInst); 2588 DEBUG(verifyRemoved(CurInst)); 2589 CurInst->eraseFromParent(); 2590 Changed = true; 2591 } 2592 } 2593 2594 if (splitCriticalEdges()) 2595 Changed = true; 2596 2597 return Changed; 2598} 2599 2600/// Split the critical edge connecting the given two blocks, and return 2601/// the block inserted to the critical edge. 2602BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { 2603 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this); 2604 if (MD) 2605 MD->invalidateCachedPredecessors(); 2606 return BB; 2607} 2608 2609/// splitCriticalEdges - Split critical edges found during the previous 2610/// iteration that may enable further optimization. 2611bool GVN::splitCriticalEdges() { 2612 if (toSplit.empty()) 2613 return false; 2614 do { 2615 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2616 SplitCriticalEdge(Edge.first, Edge.second, this); 2617 } while (!toSplit.empty()); 2618 if (MD) MD->invalidateCachedPredecessors(); 2619 return true; 2620} 2621 2622/// iterateOnFunction - Executes one iteration of GVN 2623bool GVN::iterateOnFunction(Function &F) { 2624 cleanupGlobalSets(); 2625 2626 // Top-down walk of the dominator tree 2627 bool Changed = false; 2628#if 0 2629 // Needed for value numbering with phi construction to work. 2630 ReversePostOrderTraversal<Function*> RPOT(&F); 2631 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2632 RE = RPOT.end(); RI != RE; ++RI) 2633 Changed |= processBlock(*RI); 2634#else 2635 // Save the blocks this function have before transformation begins. GVN may 2636 // split critical edge, and hence may invalidate the RPO/DT iterator. 2637 // 2638 std::vector<BasicBlock *> BBVect; 2639 BBVect.reserve(256); 2640 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2641 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2642 BBVect.push_back(DI->getBlock()); 2643 2644 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end(); 2645 I != E; I++) 2646 Changed |= processBlock(*I); 2647#endif 2648 2649 return Changed; 2650} 2651 2652void GVN::cleanupGlobalSets() { 2653 VN.clear(); 2654 LeaderTable.clear(); 2655 TableAllocator.Reset(); 2656} 2657 2658/// verifyRemoved - Verify that the specified instruction does not occur in our 2659/// internal data structures. 2660void GVN::verifyRemoved(const Instruction *Inst) const { 2661 VN.verifyRemoved(Inst); 2662 2663 // Walk through the value number scope to make sure the instruction isn't 2664 // ferreted away in it. 2665 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2666 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2667 const LeaderTableEntry *Node = &I->second; 2668 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2669 2670 while (Node->Next) { 2671 Node = Node->Next; 2672 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2673 } 2674 } 2675} 2676 2677// BB is declared dead, which implied other blocks become dead as well. This 2678// function is to add all these blocks to "DeadBlocks". For the dead blocks' 2679// live successors, update their phi nodes by replacing the operands 2680// corresponding to dead blocks with UndefVal. 2681// 2682void GVN::addDeadBlock(BasicBlock *BB) { 2683 SmallVector<BasicBlock *, 4> NewDead; 2684 SmallSetVector<BasicBlock *, 4> DF; 2685 2686 NewDead.push_back(BB); 2687 while (!NewDead.empty()) { 2688 BasicBlock *D = NewDead.pop_back_val(); 2689 if (DeadBlocks.count(D)) 2690 continue; 2691 2692 // All blocks dominated by D are dead. 2693 SmallVector<BasicBlock *, 8> Dom; 2694 DT->getDescendants(D, Dom); 2695 DeadBlocks.insert(Dom.begin(), Dom.end()); 2696 2697 // Figure out the dominance-frontier(D). 2698 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(), 2699 E = Dom.end(); I != E; I++) { 2700 BasicBlock *B = *I; 2701 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) { 2702 BasicBlock *S = *SI; 2703 if (DeadBlocks.count(S)) 2704 continue; 2705 2706 bool AllPredDead = true; 2707 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++) 2708 if (!DeadBlocks.count(*PI)) { 2709 AllPredDead = false; 2710 break; 2711 } 2712 2713 if (!AllPredDead) { 2714 // S could be proved dead later on. That is why we don't update phi 2715 // operands at this moment. 2716 DF.insert(S); 2717 } else { 2718 // While S is not dominated by D, it is dead by now. This could take 2719 // place if S already have a dead predecessor before D is declared 2720 // dead. 2721 NewDead.push_back(S); 2722 } 2723 } 2724 } 2725 } 2726 2727 // For the dead blocks' live successors, update their phi nodes by replacing 2728 // the operands corresponding to dead blocks with UndefVal. 2729 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end(); 2730 I != E; I++) { 2731 BasicBlock *B = *I; 2732 if (DeadBlocks.count(B)) 2733 continue; 2734 2735 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B)); 2736 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(), 2737 PE = Preds.end(); PI != PE; PI++) { 2738 BasicBlock *P = *PI; 2739 2740 if (!DeadBlocks.count(P)) 2741 continue; 2742 2743 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) { 2744 if (BasicBlock *S = splitCriticalEdges(P, B)) 2745 DeadBlocks.insert(P = S); 2746 } 2747 2748 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) { 2749 PHINode &Phi = cast<PHINode>(*II); 2750 Phi.setIncomingValue(Phi.getBasicBlockIndex(P), 2751 UndefValue::get(Phi.getType())); 2752 } 2753 } 2754 } 2755} 2756 2757// If the given branch is recognized as a foldable branch (i.e. conditional 2758// branch with constant condition), it will perform following analyses and 2759// transformation. 2760// 1) If the dead out-coming edge is a critical-edge, split it. Let 2761// R be the target of the dead out-coming edge. 2762// 1) Identify the set of dead blocks implied by the branch's dead outcoming 2763// edge. The result of this step will be {X| X is dominated by R} 2764// 2) Identify those blocks which haves at least one dead prodecessor. The 2765// result of this step will be dominance-frontier(R). 2766// 3) Update the PHIs in DF(R) by replacing the operands corresponding to 2767// dead blocks with "UndefVal" in an hope these PHIs will optimized away. 2768// 2769// Return true iff *NEW* dead code are found. 2770bool GVN::processFoldableCondBr(BranchInst *BI) { 2771 if (!BI || BI->isUnconditional()) 2772 return false; 2773 2774 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition()); 2775 if (!Cond) 2776 return false; 2777 2778 BasicBlock *DeadRoot = Cond->getZExtValue() ? 2779 BI->getSuccessor(1) : BI->getSuccessor(0); 2780 if (DeadBlocks.count(DeadRoot)) 2781 return false; 2782 2783 if (!DeadRoot->getSinglePredecessor()) 2784 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot); 2785 2786 addDeadBlock(DeadRoot); 2787 return true; 2788} 2789 2790// performPRE() will trigger assert if it come across an instruciton without 2791// associated val-num. As it normally has far more live instructions than dead 2792// instructions, it makes more sense just to "fabricate" a val-number for the 2793// dead code than checking if instruction involved is dead or not. 2794void GVN::assignValNumForDeadCode() { 2795 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(), 2796 E = DeadBlocks.end(); I != E; I++) { 2797 BasicBlock *BB = *I; 2798 for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); 2799 II != EE; II++) { 2800 Instruction *Inst = &*II; 2801 unsigned ValNum = VN.lookup_or_add(Inst); 2802 addToLeaderTable(ValNum, Inst, BB); 2803 } 2804 } 2805} 2806