1//===- GVN.cpp - Eliminate redundant values and loads ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs global value numbering to eliminate fully redundant 11// instructions. It also performs simple dead load elimination. 12// 13// Note that this pass does the value numbering itself; it does not use the 14// ValueNumbering analysis passes. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "gvn" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/ADT/DenseMap.h" 21#include "llvm/ADT/DepthFirstIterator.h" 22#include "llvm/ADT/Hashing.h" 23#include "llvm/ADT/SmallPtrSet.h" 24#include "llvm/ADT/SetVector.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/Dominators.h" 30#include "llvm/Analysis/InstructionSimplify.h" 31#include "llvm/Analysis/Loads.h" 32#include "llvm/Analysis/MemoryBuiltins.h" 33#include "llvm/Analysis/MemoryDependenceAnalysis.h" 34#include "llvm/Analysis/PHITransAddr.h" 35#include "llvm/Analysis/ValueTracking.h" 36#include "llvm/Assembly/Writer.h" 37#include "llvm/IR/DataLayout.h" 38#include "llvm/IR/GlobalVariable.h" 39#include "llvm/IR/IRBuilder.h" 40#include "llvm/IR/IntrinsicInst.h" 41#include "llvm/IR/LLVMContext.h" 42#include "llvm/IR/Metadata.h" 43#include "llvm/Support/Allocator.h" 44#include "llvm/Support/CommandLine.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Support/PatternMatch.h" 47#include "llvm/Target/TargetLibraryInfo.h" 48#include "llvm/Transforms/Utils/BasicBlockUtils.h" 49#include "llvm/Transforms/Utils/SSAUpdater.h" 50#include <vector> 51using namespace llvm; 52using namespace PatternMatch; 53 54STATISTIC(NumGVNInstr, "Number of instructions deleted"); 55STATISTIC(NumGVNLoad, "Number of loads deleted"); 56STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 57STATISTIC(NumGVNBlocks, "Number of blocks merged"); 58STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 59STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 60STATISTIC(NumPRELoad, "Number of loads PRE'd"); 61 62static cl::opt<bool> EnablePRE("enable-pre", 63 cl::init(true), cl::Hidden); 64static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 65 66// Maximum allowed recursion depth. 67static cl::opt<uint32_t> 68MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 69 cl::desc("Max recurse depth (default = 1000)")); 70 71//===----------------------------------------------------------------------===// 72// ValueTable Class 73//===----------------------------------------------------------------------===// 74 75/// This class holds the mapping between values and value numbers. It is used 76/// as an efficient mechanism to determine the expression-wise equivalence of 77/// two values. 78namespace { 79 struct Expression { 80 uint32_t opcode; 81 Type *type; 82 SmallVector<uint32_t, 4> varargs; 83 84 Expression(uint32_t o = ~2U) : opcode(o) { } 85 86 bool operator==(const Expression &other) const { 87 if (opcode != other.opcode) 88 return false; 89 if (opcode == ~0U || opcode == ~1U) 90 return true; 91 if (type != other.type) 92 return false; 93 if (varargs != other.varargs) 94 return false; 95 return true; 96 } 97 98 friend hash_code hash_value(const Expression &Value) { 99 return hash_combine(Value.opcode, Value.type, 100 hash_combine_range(Value.varargs.begin(), 101 Value.varargs.end())); 102 } 103 }; 104 105 class ValueTable { 106 DenseMap<Value*, uint32_t> valueNumbering; 107 DenseMap<Expression, uint32_t> expressionNumbering; 108 AliasAnalysis *AA; 109 MemoryDependenceAnalysis *MD; 110 DominatorTree *DT; 111 112 uint32_t nextValueNumber; 113 114 Expression create_expression(Instruction* I); 115 Expression create_cmp_expression(unsigned Opcode, 116 CmpInst::Predicate Predicate, 117 Value *LHS, Value *RHS); 118 Expression create_extractvalue_expression(ExtractValueInst* EI); 119 uint32_t lookup_or_add_call(CallInst* C); 120 public: 121 ValueTable() : nextValueNumber(1) { } 122 uint32_t lookup_or_add(Value *V); 123 uint32_t lookup(Value *V) const; 124 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 125 Value *LHS, Value *RHS); 126 void add(Value *V, uint32_t num); 127 void clear(); 128 void erase(Value *v); 129 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 130 AliasAnalysis *getAliasAnalysis() const { return AA; } 131 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 132 void setDomTree(DominatorTree* D) { DT = D; } 133 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 134 void verifyRemoved(const Value *) const; 135 }; 136} 137 138namespace llvm { 139template <> struct DenseMapInfo<Expression> { 140 static inline Expression getEmptyKey() { 141 return ~0U; 142 } 143 144 static inline Expression getTombstoneKey() { 145 return ~1U; 146 } 147 148 static unsigned getHashValue(const Expression e) { 149 using llvm::hash_value; 150 return static_cast<unsigned>(hash_value(e)); 151 } 152 static bool isEqual(const Expression &LHS, const Expression &RHS) { 153 return LHS == RHS; 154 } 155}; 156 157} 158 159//===----------------------------------------------------------------------===// 160// ValueTable Internal Functions 161//===----------------------------------------------------------------------===// 162 163Expression ValueTable::create_expression(Instruction *I) { 164 Expression e; 165 e.type = I->getType(); 166 e.opcode = I->getOpcode(); 167 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 168 OI != OE; ++OI) 169 e.varargs.push_back(lookup_or_add(*OI)); 170 if (I->isCommutative()) { 171 // Ensure that commutative instructions that only differ by a permutation 172 // of their operands get the same value number by sorting the operand value 173 // numbers. Since all commutative instructions have two operands it is more 174 // efficient to sort by hand rather than using, say, std::sort. 175 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 176 if (e.varargs[0] > e.varargs[1]) 177 std::swap(e.varargs[0], e.varargs[1]); 178 } 179 180 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 181 // Sort the operand value numbers so x<y and y>x get the same value number. 182 CmpInst::Predicate Predicate = C->getPredicate(); 183 if (e.varargs[0] > e.varargs[1]) { 184 std::swap(e.varargs[0], e.varargs[1]); 185 Predicate = CmpInst::getSwappedPredicate(Predicate); 186 } 187 e.opcode = (C->getOpcode() << 8) | Predicate; 188 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 189 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 190 II != IE; ++II) 191 e.varargs.push_back(*II); 192 } 193 194 return e; 195} 196 197Expression ValueTable::create_cmp_expression(unsigned Opcode, 198 CmpInst::Predicate Predicate, 199 Value *LHS, Value *RHS) { 200 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 201 "Not a comparison!"); 202 Expression e; 203 e.type = CmpInst::makeCmpResultType(LHS->getType()); 204 e.varargs.push_back(lookup_or_add(LHS)); 205 e.varargs.push_back(lookup_or_add(RHS)); 206 207 // Sort the operand value numbers so x<y and y>x get the same value number. 208 if (e.varargs[0] > e.varargs[1]) { 209 std::swap(e.varargs[0], e.varargs[1]); 210 Predicate = CmpInst::getSwappedPredicate(Predicate); 211 } 212 e.opcode = (Opcode << 8) | Predicate; 213 return e; 214} 215 216Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 217 assert(EI != 0 && "Not an ExtractValueInst?"); 218 Expression e; 219 e.type = EI->getType(); 220 e.opcode = 0; 221 222 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 223 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 224 // EI might be an extract from one of our recognised intrinsics. If it 225 // is we'll synthesize a semantically equivalent expression instead on 226 // an extract value expression. 227 switch (I->getIntrinsicID()) { 228 case Intrinsic::sadd_with_overflow: 229 case Intrinsic::uadd_with_overflow: 230 e.opcode = Instruction::Add; 231 break; 232 case Intrinsic::ssub_with_overflow: 233 case Intrinsic::usub_with_overflow: 234 e.opcode = Instruction::Sub; 235 break; 236 case Intrinsic::smul_with_overflow: 237 case Intrinsic::umul_with_overflow: 238 e.opcode = Instruction::Mul; 239 break; 240 default: 241 break; 242 } 243 244 if (e.opcode != 0) { 245 // Intrinsic recognized. Grab its args to finish building the expression. 246 assert(I->getNumArgOperands() == 2 && 247 "Expect two args for recognised intrinsics."); 248 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 249 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 250 return e; 251 } 252 } 253 254 // Not a recognised intrinsic. Fall back to producing an extract value 255 // expression. 256 e.opcode = EI->getOpcode(); 257 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 258 OI != OE; ++OI) 259 e.varargs.push_back(lookup_or_add(*OI)); 260 261 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 262 II != IE; ++II) 263 e.varargs.push_back(*II); 264 265 return e; 266} 267 268//===----------------------------------------------------------------------===// 269// ValueTable External Functions 270//===----------------------------------------------------------------------===// 271 272/// add - Insert a value into the table with a specified value number. 273void ValueTable::add(Value *V, uint32_t num) { 274 valueNumbering.insert(std::make_pair(V, num)); 275} 276 277uint32_t ValueTable::lookup_or_add_call(CallInst *C) { 278 if (AA->doesNotAccessMemory(C)) { 279 Expression exp = create_expression(C); 280 uint32_t &e = expressionNumbering[exp]; 281 if (!e) e = nextValueNumber++; 282 valueNumbering[C] = e; 283 return e; 284 } else if (AA->onlyReadsMemory(C)) { 285 Expression exp = create_expression(C); 286 uint32_t &e = expressionNumbering[exp]; 287 if (!e) { 288 e = nextValueNumber++; 289 valueNumbering[C] = e; 290 return e; 291 } 292 if (!MD) { 293 e = nextValueNumber++; 294 valueNumbering[C] = e; 295 return e; 296 } 297 298 MemDepResult local_dep = MD->getDependency(C); 299 300 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 301 valueNumbering[C] = nextValueNumber; 302 return nextValueNumber++; 303 } 304 305 if (local_dep.isDef()) { 306 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 307 308 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 309 valueNumbering[C] = nextValueNumber; 310 return nextValueNumber++; 311 } 312 313 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 314 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 315 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 316 if (c_vn != cd_vn) { 317 valueNumbering[C] = nextValueNumber; 318 return nextValueNumber++; 319 } 320 } 321 322 uint32_t v = lookup_or_add(local_cdep); 323 valueNumbering[C] = v; 324 return v; 325 } 326 327 // Non-local case. 328 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 329 MD->getNonLocalCallDependency(CallSite(C)); 330 // FIXME: Move the checking logic to MemDep! 331 CallInst* cdep = 0; 332 333 // Check to see if we have a single dominating call instruction that is 334 // identical to C. 335 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 336 const NonLocalDepEntry *I = &deps[i]; 337 if (I->getResult().isNonLocal()) 338 continue; 339 340 // We don't handle non-definitions. If we already have a call, reject 341 // instruction dependencies. 342 if (!I->getResult().isDef() || cdep != 0) { 343 cdep = 0; 344 break; 345 } 346 347 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 348 // FIXME: All duplicated with non-local case. 349 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 350 cdep = NonLocalDepCall; 351 continue; 352 } 353 354 cdep = 0; 355 break; 356 } 357 358 if (!cdep) { 359 valueNumbering[C] = nextValueNumber; 360 return nextValueNumber++; 361 } 362 363 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 364 valueNumbering[C] = nextValueNumber; 365 return nextValueNumber++; 366 } 367 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 368 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 369 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 370 if (c_vn != cd_vn) { 371 valueNumbering[C] = nextValueNumber; 372 return nextValueNumber++; 373 } 374 } 375 376 uint32_t v = lookup_or_add(cdep); 377 valueNumbering[C] = v; 378 return v; 379 380 } else { 381 valueNumbering[C] = nextValueNumber; 382 return nextValueNumber++; 383 } 384} 385 386/// lookup_or_add - Returns the value number for the specified value, assigning 387/// it a new number if it did not have one before. 388uint32_t ValueTable::lookup_or_add(Value *V) { 389 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 390 if (VI != valueNumbering.end()) 391 return VI->second; 392 393 if (!isa<Instruction>(V)) { 394 valueNumbering[V] = nextValueNumber; 395 return nextValueNumber++; 396 } 397 398 Instruction* I = cast<Instruction>(V); 399 Expression exp; 400 switch (I->getOpcode()) { 401 case Instruction::Call: 402 return lookup_or_add_call(cast<CallInst>(I)); 403 case Instruction::Add: 404 case Instruction::FAdd: 405 case Instruction::Sub: 406 case Instruction::FSub: 407 case Instruction::Mul: 408 case Instruction::FMul: 409 case Instruction::UDiv: 410 case Instruction::SDiv: 411 case Instruction::FDiv: 412 case Instruction::URem: 413 case Instruction::SRem: 414 case Instruction::FRem: 415 case Instruction::Shl: 416 case Instruction::LShr: 417 case Instruction::AShr: 418 case Instruction::And: 419 case Instruction::Or: 420 case Instruction::Xor: 421 case Instruction::ICmp: 422 case Instruction::FCmp: 423 case Instruction::Trunc: 424 case Instruction::ZExt: 425 case Instruction::SExt: 426 case Instruction::FPToUI: 427 case Instruction::FPToSI: 428 case Instruction::UIToFP: 429 case Instruction::SIToFP: 430 case Instruction::FPTrunc: 431 case Instruction::FPExt: 432 case Instruction::PtrToInt: 433 case Instruction::IntToPtr: 434 case Instruction::BitCast: 435 case Instruction::Select: 436 case Instruction::ExtractElement: 437 case Instruction::InsertElement: 438 case Instruction::ShuffleVector: 439 case Instruction::InsertValue: 440 case Instruction::GetElementPtr: 441 exp = create_expression(I); 442 break; 443 case Instruction::ExtractValue: 444 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 445 break; 446 default: 447 valueNumbering[V] = nextValueNumber; 448 return nextValueNumber++; 449 } 450 451 uint32_t& e = expressionNumbering[exp]; 452 if (!e) e = nextValueNumber++; 453 valueNumbering[V] = e; 454 return e; 455} 456 457/// lookup - Returns the value number of the specified value. Fails if 458/// the value has not yet been numbered. 459uint32_t ValueTable::lookup(Value *V) const { 460 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 461 assert(VI != valueNumbering.end() && "Value not numbered?"); 462 return VI->second; 463} 464 465/// lookup_or_add_cmp - Returns the value number of the given comparison, 466/// assigning it a new number if it did not have one before. Useful when 467/// we deduced the result of a comparison, but don't immediately have an 468/// instruction realizing that comparison to hand. 469uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 470 CmpInst::Predicate Predicate, 471 Value *LHS, Value *RHS) { 472 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 473 uint32_t& e = expressionNumbering[exp]; 474 if (!e) e = nextValueNumber++; 475 return e; 476} 477 478/// clear - Remove all entries from the ValueTable. 479void ValueTable::clear() { 480 valueNumbering.clear(); 481 expressionNumbering.clear(); 482 nextValueNumber = 1; 483} 484 485/// erase - Remove a value from the value numbering. 486void ValueTable::erase(Value *V) { 487 valueNumbering.erase(V); 488} 489 490/// verifyRemoved - Verify that the value is removed from all internal data 491/// structures. 492void ValueTable::verifyRemoved(const Value *V) const { 493 for (DenseMap<Value*, uint32_t>::const_iterator 494 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 495 assert(I->first != V && "Inst still occurs in value numbering map!"); 496 } 497} 498 499//===----------------------------------------------------------------------===// 500// GVN Pass 501//===----------------------------------------------------------------------===// 502 503namespace { 504 class GVN; 505 struct AvailableValueInBlock { 506 /// BB - The basic block in question. 507 BasicBlock *BB; 508 enum ValType { 509 SimpleVal, // A simple offsetted value that is accessed. 510 LoadVal, // A value produced by a load. 511 MemIntrin, // A memory intrinsic which is loaded from. 512 UndefVal // A UndefValue representing a value from dead block (which 513 // is not yet physically removed from the CFG). 514 }; 515 516 /// V - The value that is live out of the block. 517 PointerIntPair<Value *, 2, ValType> Val; 518 519 /// Offset - The byte offset in Val that is interesting for the load query. 520 unsigned Offset; 521 522 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 523 unsigned Offset = 0) { 524 AvailableValueInBlock Res; 525 Res.BB = BB; 526 Res.Val.setPointer(V); 527 Res.Val.setInt(SimpleVal); 528 Res.Offset = Offset; 529 return Res; 530 } 531 532 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 533 unsigned Offset = 0) { 534 AvailableValueInBlock Res; 535 Res.BB = BB; 536 Res.Val.setPointer(MI); 537 Res.Val.setInt(MemIntrin); 538 Res.Offset = Offset; 539 return Res; 540 } 541 542 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 543 unsigned Offset = 0) { 544 AvailableValueInBlock Res; 545 Res.BB = BB; 546 Res.Val.setPointer(LI); 547 Res.Val.setInt(LoadVal); 548 Res.Offset = Offset; 549 return Res; 550 } 551 552 static AvailableValueInBlock getUndef(BasicBlock *BB) { 553 AvailableValueInBlock Res; 554 Res.BB = BB; 555 Res.Val.setPointer(0); 556 Res.Val.setInt(UndefVal); 557 Res.Offset = 0; 558 return Res; 559 } 560 561 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 562 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 563 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 564 bool isUndefValue() const { return Val.getInt() == UndefVal; } 565 566 Value *getSimpleValue() const { 567 assert(isSimpleValue() && "Wrong accessor"); 568 return Val.getPointer(); 569 } 570 571 LoadInst *getCoercedLoadValue() const { 572 assert(isCoercedLoadValue() && "Wrong accessor"); 573 return cast<LoadInst>(Val.getPointer()); 574 } 575 576 MemIntrinsic *getMemIntrinValue() const { 577 assert(isMemIntrinValue() && "Wrong accessor"); 578 return cast<MemIntrinsic>(Val.getPointer()); 579 } 580 581 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 582 /// defined here to the specified type. This handles various coercion cases. 583 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const; 584 }; 585 586 class GVN : public FunctionPass { 587 bool NoLoads; 588 MemoryDependenceAnalysis *MD; 589 DominatorTree *DT; 590 const DataLayout *TD; 591 const TargetLibraryInfo *TLI; 592 SetVector<BasicBlock *> DeadBlocks; 593 594 ValueTable VN; 595 596 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 597 /// have that value number. Use findLeader to query it. 598 struct LeaderTableEntry { 599 Value *Val; 600 const BasicBlock *BB; 601 LeaderTableEntry *Next; 602 }; 603 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 604 BumpPtrAllocator TableAllocator; 605 606 SmallVector<Instruction*, 8> InstrsToErase; 607 608 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect; 609 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect; 610 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect; 611 612 public: 613 static char ID; // Pass identification, replacement for typeid 614 explicit GVN(bool noloads = false) 615 : FunctionPass(ID), NoLoads(noloads), MD(0) { 616 initializeGVNPass(*PassRegistry::getPassRegistry()); 617 } 618 619 bool runOnFunction(Function &F); 620 621 /// markInstructionForDeletion - This removes the specified instruction from 622 /// our various maps and marks it for deletion. 623 void markInstructionForDeletion(Instruction *I) { 624 VN.erase(I); 625 InstrsToErase.push_back(I); 626 } 627 628 const DataLayout *getDataLayout() const { return TD; } 629 DominatorTree &getDominatorTree() const { return *DT; } 630 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 631 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 632 private: 633 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 634 /// its value number. 635 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 636 LeaderTableEntry &Curr = LeaderTable[N]; 637 if (!Curr.Val) { 638 Curr.Val = V; 639 Curr.BB = BB; 640 return; 641 } 642 643 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 644 Node->Val = V; 645 Node->BB = BB; 646 Node->Next = Curr.Next; 647 Curr.Next = Node; 648 } 649 650 /// removeFromLeaderTable - Scan the list of values corresponding to a given 651 /// value number, and remove the given instruction if encountered. 652 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 653 LeaderTableEntry* Prev = 0; 654 LeaderTableEntry* Curr = &LeaderTable[N]; 655 656 while (Curr->Val != I || Curr->BB != BB) { 657 Prev = Curr; 658 Curr = Curr->Next; 659 } 660 661 if (Prev) { 662 Prev->Next = Curr->Next; 663 } else { 664 if (!Curr->Next) { 665 Curr->Val = 0; 666 Curr->BB = 0; 667 } else { 668 LeaderTableEntry* Next = Curr->Next; 669 Curr->Val = Next->Val; 670 Curr->BB = Next->BB; 671 Curr->Next = Next->Next; 672 } 673 } 674 } 675 676 // List of critical edges to be split between iterations. 677 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 678 679 // This transformation requires dominator postdominator info 680 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 681 AU.addRequired<DominatorTree>(); 682 AU.addRequired<TargetLibraryInfo>(); 683 if (!NoLoads) 684 AU.addRequired<MemoryDependenceAnalysis>(); 685 AU.addRequired<AliasAnalysis>(); 686 687 AU.addPreserved<DominatorTree>(); 688 AU.addPreserved<AliasAnalysis>(); 689 } 690 691 692 // Helper fuctions of redundant load elimination 693 bool processLoad(LoadInst *L); 694 bool processNonLocalLoad(LoadInst *L); 695 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 696 AvailValInBlkVect &ValuesPerBlock, 697 UnavailBlkVect &UnavailableBlocks); 698 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 699 UnavailBlkVect &UnavailableBlocks); 700 701 // Other helper routines 702 bool processInstruction(Instruction *I); 703 bool processBlock(BasicBlock *BB); 704 void dump(DenseMap<uint32_t, Value*> &d); 705 bool iterateOnFunction(Function &F); 706 bool performPRE(Function &F); 707 Value *findLeader(const BasicBlock *BB, uint32_t num); 708 void cleanupGlobalSets(); 709 void verifyRemoved(const Instruction *I) const; 710 bool splitCriticalEdges(); 711 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ); 712 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 713 const BasicBlockEdge &Root); 714 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 715 bool processFoldableCondBr(BranchInst *BI); 716 void addDeadBlock(BasicBlock *BB); 717 void assignValNumForDeadCode(); 718 }; 719 720 char GVN::ID = 0; 721} 722 723// createGVNPass - The public interface to this file... 724FunctionPass *llvm::createGVNPass(bool NoLoads) { 725 return new GVN(NoLoads); 726} 727 728INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 729INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 730INITIALIZE_PASS_DEPENDENCY(DominatorTree) 731INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 732INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 733INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 734 735#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 736void GVN::dump(DenseMap<uint32_t, Value*>& d) { 737 errs() << "{\n"; 738 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 739 E = d.end(); I != E; ++I) { 740 errs() << I->first << "\n"; 741 I->second->dump(); 742 } 743 errs() << "}\n"; 744} 745#endif 746 747/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 748/// we're analyzing is fully available in the specified block. As we go, keep 749/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 750/// map is actually a tri-state map with the following values: 751/// 0) we know the block *is not* fully available. 752/// 1) we know the block *is* fully available. 753/// 2) we do not know whether the block is fully available or not, but we are 754/// currently speculating that it will be. 755/// 3) we are speculating for this block and have used that to speculate for 756/// other blocks. 757static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 758 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 759 uint32_t RecurseDepth) { 760 if (RecurseDepth > MaxRecurseDepth) 761 return false; 762 763 // Optimistically assume that the block is fully available and check to see 764 // if we already know about this block in one lookup. 765 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 766 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 767 768 // If the entry already existed for this block, return the precomputed value. 769 if (!IV.second) { 770 // If this is a speculative "available" value, mark it as being used for 771 // speculation of other blocks. 772 if (IV.first->second == 2) 773 IV.first->second = 3; 774 return IV.first->second != 0; 775 } 776 777 // Otherwise, see if it is fully available in all predecessors. 778 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 779 780 // If this block has no predecessors, it isn't live-in here. 781 if (PI == PE) 782 goto SpeculationFailure; 783 784 for (; PI != PE; ++PI) 785 // If the value isn't fully available in one of our predecessors, then it 786 // isn't fully available in this block either. Undo our previous 787 // optimistic assumption and bail out. 788 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 789 goto SpeculationFailure; 790 791 return true; 792 793// SpeculationFailure - If we get here, we found out that this is not, after 794// all, a fully-available block. We have a problem if we speculated on this and 795// used the speculation to mark other blocks as available. 796SpeculationFailure: 797 char &BBVal = FullyAvailableBlocks[BB]; 798 799 // If we didn't speculate on this, just return with it set to false. 800 if (BBVal == 2) { 801 BBVal = 0; 802 return false; 803 } 804 805 // If we did speculate on this value, we could have blocks set to 1 that are 806 // incorrect. Walk the (transitive) successors of this block and mark them as 807 // 0 if set to one. 808 SmallVector<BasicBlock*, 32> BBWorklist; 809 BBWorklist.push_back(BB); 810 811 do { 812 BasicBlock *Entry = BBWorklist.pop_back_val(); 813 // Note that this sets blocks to 0 (unavailable) if they happen to not 814 // already be in FullyAvailableBlocks. This is safe. 815 char &EntryVal = FullyAvailableBlocks[Entry]; 816 if (EntryVal == 0) continue; // Already unavailable. 817 818 // Mark as unavailable. 819 EntryVal = 0; 820 821 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 822 BBWorklist.push_back(*I); 823 } while (!BBWorklist.empty()); 824 825 return false; 826} 827 828 829/// CanCoerceMustAliasedValueToLoad - Return true if 830/// CoerceAvailableValueToLoadType will succeed. 831static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 832 Type *LoadTy, 833 const DataLayout &TD) { 834 // If the loaded or stored value is an first class array or struct, don't try 835 // to transform them. We need to be able to bitcast to integer. 836 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 837 StoredVal->getType()->isStructTy() || 838 StoredVal->getType()->isArrayTy()) 839 return false; 840 841 // The store has to be at least as big as the load. 842 if (TD.getTypeSizeInBits(StoredVal->getType()) < 843 TD.getTypeSizeInBits(LoadTy)) 844 return false; 845 846 return true; 847} 848 849/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 850/// then a load from a must-aliased pointer of a different type, try to coerce 851/// the stored value. LoadedTy is the type of the load we want to replace and 852/// InsertPt is the place to insert new instructions. 853/// 854/// If we can't do it, return null. 855static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 856 Type *LoadedTy, 857 Instruction *InsertPt, 858 const DataLayout &TD) { 859 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 860 return 0; 861 862 // If this is already the right type, just return it. 863 Type *StoredValTy = StoredVal->getType(); 864 865 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 866 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 867 868 // If the store and reload are the same size, we can always reuse it. 869 if (StoreSize == LoadSize) { 870 // Pointer to Pointer -> use bitcast. 871 if (StoredValTy->getScalarType()->isPointerTy() && 872 LoadedTy->getScalarType()->isPointerTy()) 873 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 874 875 // Convert source pointers to integers, which can be bitcast. 876 if (StoredValTy->getScalarType()->isPointerTy()) { 877 StoredValTy = TD.getIntPtrType(StoredValTy); 878 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 879 } 880 881 Type *TypeToCastTo = LoadedTy; 882 if (TypeToCastTo->getScalarType()->isPointerTy()) 883 TypeToCastTo = TD.getIntPtrType(TypeToCastTo); 884 885 if (StoredValTy != TypeToCastTo) 886 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 887 888 // Cast to pointer if the load needs a pointer type. 889 if (LoadedTy->getScalarType()->isPointerTy()) 890 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 891 892 return StoredVal; 893 } 894 895 // If the loaded value is smaller than the available value, then we can 896 // extract out a piece from it. If the available value is too small, then we 897 // can't do anything. 898 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 899 900 // Convert source pointers to integers, which can be manipulated. 901 if (StoredValTy->getScalarType()->isPointerTy()) { 902 StoredValTy = TD.getIntPtrType(StoredValTy); 903 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 904 } 905 906 // Convert vectors and fp to integer, which can be manipulated. 907 if (!StoredValTy->isIntegerTy()) { 908 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 909 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 910 } 911 912 // If this is a big-endian system, we need to shift the value down to the low 913 // bits so that a truncate will work. 914 if (TD.isBigEndian()) { 915 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 916 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 917 } 918 919 // Truncate the integer to the right size now. 920 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 921 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 922 923 if (LoadedTy == NewIntTy) 924 return StoredVal; 925 926 // If the result is a pointer, inttoptr. 927 if (LoadedTy->getScalarType()->isPointerTy()) 928 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 929 930 // Otherwise, bitcast. 931 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 932} 933 934/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 935/// memdep query of a load that ends up being a clobbering memory write (store, 936/// memset, memcpy, memmove). This means that the write *may* provide bits used 937/// by the load but we can't be sure because the pointers don't mustalias. 938/// 939/// Check this case to see if there is anything more we can do before we give 940/// up. This returns -1 if we have to give up, or a byte number in the stored 941/// value of the piece that feeds the load. 942static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 943 Value *WritePtr, 944 uint64_t WriteSizeInBits, 945 const DataLayout &TD) { 946 // If the loaded or stored value is a first class array or struct, don't try 947 // to transform them. We need to be able to bitcast to integer. 948 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 949 return -1; 950 951 int64_t StoreOffset = 0, LoadOffset = 0; 952 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD); 953 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD); 954 if (StoreBase != LoadBase) 955 return -1; 956 957 // If the load and store are to the exact same address, they should have been 958 // a must alias. AA must have gotten confused. 959 // FIXME: Study to see if/when this happens. One case is forwarding a memset 960 // to a load from the base of the memset. 961#if 0 962 if (LoadOffset == StoreOffset) { 963 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 964 << "Base = " << *StoreBase << "\n" 965 << "Store Ptr = " << *WritePtr << "\n" 966 << "Store Offs = " << StoreOffset << "\n" 967 << "Load Ptr = " << *LoadPtr << "\n"; 968 abort(); 969 } 970#endif 971 972 // If the load and store don't overlap at all, the store doesn't provide 973 // anything to the load. In this case, they really don't alias at all, AA 974 // must have gotten confused. 975 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 976 977 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 978 return -1; 979 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 980 LoadSize >>= 3; 981 982 983 bool isAAFailure = false; 984 if (StoreOffset < LoadOffset) 985 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 986 else 987 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 988 989 if (isAAFailure) { 990#if 0 991 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 992 << "Base = " << *StoreBase << "\n" 993 << "Store Ptr = " << *WritePtr << "\n" 994 << "Store Offs = " << StoreOffset << "\n" 995 << "Load Ptr = " << *LoadPtr << "\n"; 996 abort(); 997#endif 998 return -1; 999 } 1000 1001 // If the Load isn't completely contained within the stored bits, we don't 1002 // have all the bits to feed it. We could do something crazy in the future 1003 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1004 // valuable. 1005 if (StoreOffset > LoadOffset || 1006 StoreOffset+StoreSize < LoadOffset+LoadSize) 1007 return -1; 1008 1009 // Okay, we can do this transformation. Return the number of bytes into the 1010 // store that the load is. 1011 return LoadOffset-StoreOffset; 1012} 1013 1014/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1015/// memdep query of a load that ends up being a clobbering store. 1016static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 1017 StoreInst *DepSI, 1018 const DataLayout &TD) { 1019 // Cannot handle reading from store of first-class aggregate yet. 1020 if (DepSI->getValueOperand()->getType()->isStructTy() || 1021 DepSI->getValueOperand()->getType()->isArrayTy()) 1022 return -1; 1023 1024 Value *StorePtr = DepSI->getPointerOperand(); 1025 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1026 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1027 StorePtr, StoreSize, TD); 1028} 1029 1030/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 1031/// memdep query of a load that ends up being clobbered by another load. See if 1032/// the other load can feed into the second load. 1033static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 1034 LoadInst *DepLI, const DataLayout &TD){ 1035 // Cannot handle reading from store of first-class aggregate yet. 1036 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 1037 return -1; 1038 1039 Value *DepPtr = DepLI->getPointerOperand(); 1040 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 1041 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 1042 if (R != -1) return R; 1043 1044 // If we have a load/load clobber an DepLI can be widened to cover this load, 1045 // then we should widen it! 1046 int64_t LoadOffs = 0; 1047 const Value *LoadBase = 1048 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD); 1049 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1050 1051 unsigned Size = MemoryDependenceAnalysis:: 1052 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 1053 if (Size == 0) return -1; 1054 1055 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 1056} 1057 1058 1059 1060static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 1061 MemIntrinsic *MI, 1062 const DataLayout &TD) { 1063 // If the mem operation is a non-constant size, we can't handle it. 1064 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1065 if (SizeCst == 0) return -1; 1066 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1067 1068 // If this is memset, we just need to see if the offset is valid in the size 1069 // of the memset.. 1070 if (MI->getIntrinsicID() == Intrinsic::memset) 1071 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1072 MemSizeInBits, TD); 1073 1074 // If we have a memcpy/memmove, the only case we can handle is if this is a 1075 // copy from constant memory. In that case, we can read directly from the 1076 // constant memory. 1077 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1078 1079 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1080 if (Src == 0) return -1; 1081 1082 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 1083 if (GV == 0 || !GV->isConstant()) return -1; 1084 1085 // See if the access is within the bounds of the transfer. 1086 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1087 MI->getDest(), MemSizeInBits, TD); 1088 if (Offset == -1) 1089 return Offset; 1090 1091 unsigned AS = Src->getType()->getPointerAddressSpace(); 1092 // Otherwise, see if we can constant fold a load from the constant with the 1093 // offset applied as appropriate. 1094 Src = ConstantExpr::getBitCast(Src, 1095 Type::getInt8PtrTy(Src->getContext(), AS)); 1096 Constant *OffsetCst = 1097 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1098 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1099 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS)); 1100 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1101 return Offset; 1102 return -1; 1103} 1104 1105 1106/// GetStoreValueForLoad - This function is called when we have a 1107/// memdep query of a load that ends up being a clobbering store. This means 1108/// that the store provides bits used by the load but we the pointers don't 1109/// mustalias. Check this case to see if there is anything more we can do 1110/// before we give up. 1111static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1112 Type *LoadTy, 1113 Instruction *InsertPt, const DataLayout &TD){ 1114 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1115 1116 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1117 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1118 1119 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1120 1121 // Compute which bits of the stored value are being used by the load. Convert 1122 // to an integer type to start with. 1123 if (SrcVal->getType()->getScalarType()->isPointerTy()) 1124 SrcVal = Builder.CreatePtrToInt(SrcVal, 1125 TD.getIntPtrType(SrcVal->getType())); 1126 if (!SrcVal->getType()->isIntegerTy()) 1127 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); 1128 1129 // Shift the bits to the least significant depending on endianness. 1130 unsigned ShiftAmt; 1131 if (TD.isLittleEndian()) 1132 ShiftAmt = Offset*8; 1133 else 1134 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1135 1136 if (ShiftAmt) 1137 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt); 1138 1139 if (LoadSize != StoreSize) 1140 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8)); 1141 1142 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1143} 1144 1145/// GetLoadValueForLoad - This function is called when we have a 1146/// memdep query of a load that ends up being a clobbering load. This means 1147/// that the load *may* provide bits used by the load but we can't be sure 1148/// because the pointers don't mustalias. Check this case to see if there is 1149/// anything more we can do before we give up. 1150static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1151 Type *LoadTy, Instruction *InsertPt, 1152 GVN &gvn) { 1153 const DataLayout &TD = *gvn.getDataLayout(); 1154 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 1155 // widen SrcVal out to a larger load. 1156 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 1157 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 1158 if (Offset+LoadSize > SrcValSize) { 1159 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!"); 1160 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load"); 1161 // If we have a load/load clobber an DepLI can be widened to cover this 1162 // load, then we should widen it to the next power of 2 size big enough! 1163 unsigned NewLoadSize = Offset+LoadSize; 1164 if (!isPowerOf2_32(NewLoadSize)) 1165 NewLoadSize = NextPowerOf2(NewLoadSize); 1166 1167 Value *PtrVal = SrcVal->getPointerOperand(); 1168 1169 // Insert the new load after the old load. This ensures that subsequent 1170 // memdep queries will find the new load. We can't easily remove the old 1171 // load completely because it is already in the value numbering table. 1172 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal)); 1173 Type *DestPTy = 1174 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 1175 DestPTy = PointerType::get(DestPTy, 1176 PtrVal->getType()->getPointerAddressSpace()); 1177 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc()); 1178 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 1179 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 1180 NewLoad->takeName(SrcVal); 1181 NewLoad->setAlignment(SrcVal->getAlignment()); 1182 1183 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 1184 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 1185 1186 // Replace uses of the original load with the wider load. On a big endian 1187 // system, we need to shift down to get the relevant bits. 1188 Value *RV = NewLoad; 1189 if (TD.isBigEndian()) 1190 RV = Builder.CreateLShr(RV, 1191 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 1192 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 1193 SrcVal->replaceAllUsesWith(RV); 1194 1195 // We would like to use gvn.markInstructionForDeletion here, but we can't 1196 // because the load is already memoized into the leader map table that GVN 1197 // tracks. It is potentially possible to remove the load from the table, 1198 // but then there all of the operations based on it would need to be 1199 // rehashed. Just leave the dead load around. 1200 gvn.getMemDep().removeInstruction(SrcVal); 1201 SrcVal = NewLoad; 1202 } 1203 1204 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 1205} 1206 1207 1208/// GetMemInstValueForLoad - This function is called when we have a 1209/// memdep query of a load that ends up being a clobbering mem intrinsic. 1210static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1211 Type *LoadTy, Instruction *InsertPt, 1212 const DataLayout &TD){ 1213 LLVMContext &Ctx = LoadTy->getContext(); 1214 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1215 1216 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1217 1218 // We know that this method is only called when the mem transfer fully 1219 // provides the bits for the load. 1220 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1221 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1222 // independently of what the offset is. 1223 Value *Val = MSI->getValue(); 1224 if (LoadSize != 1) 1225 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1226 1227 Value *OneElt = Val; 1228 1229 // Splat the value out to the right number of bits. 1230 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1231 // If we can double the number of bytes set, do it. 1232 if (NumBytesSet*2 <= LoadSize) { 1233 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1234 Val = Builder.CreateOr(Val, ShVal); 1235 NumBytesSet <<= 1; 1236 continue; 1237 } 1238 1239 // Otherwise insert one byte at a time. 1240 Value *ShVal = Builder.CreateShl(Val, 1*8); 1241 Val = Builder.CreateOr(OneElt, ShVal); 1242 ++NumBytesSet; 1243 } 1244 1245 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1246 } 1247 1248 // Otherwise, this is a memcpy/memmove from a constant global. 1249 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1250 Constant *Src = cast<Constant>(MTI->getSource()); 1251 unsigned AS = Src->getType()->getPointerAddressSpace(); 1252 1253 // Otherwise, see if we can constant fold a load from the constant with the 1254 // offset applied as appropriate. 1255 Src = ConstantExpr::getBitCast(Src, 1256 Type::getInt8PtrTy(Src->getContext(), AS)); 1257 Constant *OffsetCst = 1258 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1259 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst); 1260 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS)); 1261 return ConstantFoldLoadFromConstPtr(Src, &TD); 1262} 1263 1264 1265/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1266/// construct SSA form, allowing us to eliminate LI. This returns the value 1267/// that should be used at LI's definition site. 1268static Value *ConstructSSAForLoadSet(LoadInst *LI, 1269 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1270 GVN &gvn) { 1271 // Check for the fully redundant, dominating load case. In this case, we can 1272 // just use the dominating value directly. 1273 if (ValuesPerBlock.size() == 1 && 1274 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, 1275 LI->getParent())) { 1276 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block"); 1277 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn); 1278 } 1279 1280 // Otherwise, we have to construct SSA form. 1281 SmallVector<PHINode*, 8> NewPHIs; 1282 SSAUpdater SSAUpdate(&NewPHIs); 1283 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1284 1285 Type *LoadTy = LI->getType(); 1286 1287 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1288 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1289 BasicBlock *BB = AV.BB; 1290 1291 if (SSAUpdate.HasValueForBlock(BB)) 1292 continue; 1293 1294 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn)); 1295 } 1296 1297 // Perform PHI construction. 1298 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1299 1300 // If new PHI nodes were created, notify alias analysis. 1301 if (V->getType()->getScalarType()->isPointerTy()) { 1302 AliasAnalysis *AA = gvn.getAliasAnalysis(); 1303 1304 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1305 AA->copyValue(LI, NewPHIs[i]); 1306 1307 // Now that we've copied information to the new PHIs, scan through 1308 // them again and inform alias analysis that we've added potentially 1309 // escaping uses to any values that are operands to these PHIs. 1310 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1311 PHINode *P = NewPHIs[i]; 1312 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) { 1313 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 1314 AA->addEscapingUse(P->getOperandUse(jj)); 1315 } 1316 } 1317 } 1318 1319 return V; 1320} 1321 1322Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1323 Value *Res; 1324 if (isSimpleValue()) { 1325 Res = getSimpleValue(); 1326 if (Res->getType() != LoadTy) { 1327 const DataLayout *TD = gvn.getDataLayout(); 1328 assert(TD && "Need target data to handle type mismatch case"); 1329 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1330 *TD); 1331 1332 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1333 << *getSimpleValue() << '\n' 1334 << *Res << '\n' << "\n\n\n"); 1335 } 1336 } else if (isCoercedLoadValue()) { 1337 LoadInst *Load = getCoercedLoadValue(); 1338 if (Load->getType() == LoadTy && Offset == 0) { 1339 Res = Load; 1340 } else { 1341 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1342 gvn); 1343 1344 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1345 << *getCoercedLoadValue() << '\n' 1346 << *Res << '\n' << "\n\n\n"); 1347 } 1348 } else if (isMemIntrinValue()) { 1349 const DataLayout *TD = gvn.getDataLayout(); 1350 assert(TD && "Need target data to handle type mismatch case"); 1351 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1352 LoadTy, BB->getTerminator(), *TD); 1353 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1354 << " " << *getMemIntrinValue() << '\n' 1355 << *Res << '\n' << "\n\n\n"); 1356 } else { 1357 assert(isUndefValue() && "Should be UndefVal"); 1358 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";); 1359 return UndefValue::get(LoadTy); 1360 } 1361 return Res; 1362} 1363 1364static bool isLifetimeStart(const Instruction *Inst) { 1365 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1366 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1367 return false; 1368} 1369 1370void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 1371 AvailValInBlkVect &ValuesPerBlock, 1372 UnavailBlkVect &UnavailableBlocks) { 1373 1374 // Filter out useless results (non-locals, etc). Keep track of the blocks 1375 // where we have a value available in repl, also keep track of whether we see 1376 // dependencies that produce an unknown value for the load (such as a call 1377 // that could potentially clobber the load). 1378 unsigned NumDeps = Deps.size(); 1379 for (unsigned i = 0, e = NumDeps; i != e; ++i) { 1380 BasicBlock *DepBB = Deps[i].getBB(); 1381 MemDepResult DepInfo = Deps[i].getResult(); 1382 1383 if (DeadBlocks.count(DepBB)) { 1384 // Dead dependent mem-op disguise as a load evaluating the same value 1385 // as the load in question. 1386 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB)); 1387 continue; 1388 } 1389 1390 if (!DepInfo.isDef() && !DepInfo.isClobber()) { 1391 UnavailableBlocks.push_back(DepBB); 1392 continue; 1393 } 1394 1395 if (DepInfo.isClobber()) { 1396 // The address being loaded in this non-local block may not be the same as 1397 // the pointer operand of the load if PHI translation occurs. Make sure 1398 // to consider the right address. 1399 Value *Address = Deps[i].getAddress(); 1400 1401 // If the dependence is to a store that writes to a superset of the bits 1402 // read by the load, we can extract the bits we need for the load from the 1403 // stored value. 1404 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1405 if (TD && Address) { 1406 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1407 DepSI, *TD); 1408 if (Offset != -1) { 1409 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1410 DepSI->getValueOperand(), 1411 Offset)); 1412 continue; 1413 } 1414 } 1415 } 1416 1417 // Check to see if we have something like this: 1418 // load i32* P 1419 // load i8* (P+1) 1420 // if we have this, replace the later with an extraction from the former. 1421 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1422 // If this is a clobber and L is the first instruction in its block, then 1423 // we have the first instruction in the entry block. 1424 if (DepLI != LI && Address && TD) { 1425 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1426 LI->getPointerOperand(), 1427 DepLI, *TD); 1428 1429 if (Offset != -1) { 1430 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1431 Offset)); 1432 continue; 1433 } 1434 } 1435 } 1436 1437 // If the clobbering value is a memset/memcpy/memmove, see if we can 1438 // forward a value on from it. 1439 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1440 if (TD && Address) { 1441 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1442 DepMI, *TD); 1443 if (Offset != -1) { 1444 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1445 Offset)); 1446 continue; 1447 } 1448 } 1449 } 1450 1451 UnavailableBlocks.push_back(DepBB); 1452 continue; 1453 } 1454 1455 // DepInfo.isDef() here 1456 1457 Instruction *DepInst = DepInfo.getInst(); 1458 1459 // Loading the allocation -> undef. 1460 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1461 // Loading immediately after lifetime begin -> undef. 1462 isLifetimeStart(DepInst)) { 1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1464 UndefValue::get(LI->getType()))); 1465 continue; 1466 } 1467 1468 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1469 // Reject loads and stores that are to the same address but are of 1470 // different types if we have to. 1471 if (S->getValueOperand()->getType() != LI->getType()) { 1472 // If the stored value is larger or equal to the loaded value, we can 1473 // reuse it. 1474 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1475 LI->getType(), *TD)) { 1476 UnavailableBlocks.push_back(DepBB); 1477 continue; 1478 } 1479 } 1480 1481 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1482 S->getValueOperand())); 1483 continue; 1484 } 1485 1486 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1487 // If the types mismatch and we can't handle it, reject reuse of the load. 1488 if (LD->getType() != LI->getType()) { 1489 // If the stored value is larger or equal to the loaded value, we can 1490 // reuse it. 1491 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1492 UnavailableBlocks.push_back(DepBB); 1493 continue; 1494 } 1495 } 1496 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1497 continue; 1498 } 1499 1500 UnavailableBlocks.push_back(DepBB); 1501 } 1502} 1503 1504bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 1505 UnavailBlkVect &UnavailableBlocks) { 1506 // Okay, we have *some* definitions of the value. This means that the value 1507 // is available in some of our (transitive) predecessors. Lets think about 1508 // doing PRE of this load. This will involve inserting a new load into the 1509 // predecessor when it's not available. We could do this in general, but 1510 // prefer to not increase code size. As such, we only do this when we know 1511 // that we only have to insert *one* load (which means we're basically moving 1512 // the load, not inserting a new one). 1513 1514 SmallPtrSet<BasicBlock *, 4> Blockers; 1515 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1516 Blockers.insert(UnavailableBlocks[i]); 1517 1518 // Let's find the first basic block with more than one predecessor. Walk 1519 // backwards through predecessors if needed. 1520 BasicBlock *LoadBB = LI->getParent(); 1521 BasicBlock *TmpBB = LoadBB; 1522 1523 while (TmpBB->getSinglePredecessor()) { 1524 TmpBB = TmpBB->getSinglePredecessor(); 1525 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1526 return false; 1527 if (Blockers.count(TmpBB)) 1528 return false; 1529 1530 // If any of these blocks has more than one successor (i.e. if the edge we 1531 // just traversed was critical), then there are other paths through this 1532 // block along which the load may not be anticipated. Hoisting the load 1533 // above this block would be adding the load to execution paths along 1534 // which it was not previously executed. 1535 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1536 return false; 1537 } 1538 1539 assert(TmpBB); 1540 LoadBB = TmpBB; 1541 1542 // Check to see how many predecessors have the loaded value fully 1543 // available. 1544 DenseMap<BasicBlock*, Value*> PredLoads; 1545 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1546 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1547 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1548 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1549 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1550 1551 SmallVector<BasicBlock *, 4> CriticalEdgePred; 1552 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1553 PI != E; ++PI) { 1554 BasicBlock *Pred = *PI; 1555 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1556 continue; 1557 } 1558 PredLoads[Pred] = 0; 1559 1560 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1561 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1562 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1563 << Pred->getName() << "': " << *LI << '\n'); 1564 return false; 1565 } 1566 1567 if (LoadBB->isLandingPad()) { 1568 DEBUG(dbgs() 1569 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1570 << Pred->getName() << "': " << *LI << '\n'); 1571 return false; 1572 } 1573 1574 CriticalEdgePred.push_back(Pred); 1575 } 1576 } 1577 1578 // Decide whether PRE is profitable for this load. 1579 unsigned NumUnavailablePreds = PredLoads.size(); 1580 assert(NumUnavailablePreds != 0 && 1581 "Fully available value should already be eliminated!"); 1582 1583 // If this load is unavailable in multiple predecessors, reject it. 1584 // FIXME: If we could restructure the CFG, we could make a common pred with 1585 // all the preds that don't have an available LI and insert a new load into 1586 // that one block. 1587 if (NumUnavailablePreds != 1) 1588 return false; 1589 1590 // Split critical edges, and update the unavailable predecessors accordingly. 1591 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(), 1592 E = CriticalEdgePred.end(); I != E; I++) { 1593 BasicBlock *OrigPred = *I; 1594 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); 1595 PredLoads.erase(OrigPred); 1596 PredLoads[NewPred] = 0; 1597 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" 1598 << LoadBB->getName() << '\n'); 1599 } 1600 1601 // Check if the load can safely be moved to all the unavailable predecessors. 1602 bool CanDoPRE = true; 1603 SmallVector<Instruction*, 8> NewInsts; 1604 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1605 E = PredLoads.end(); I != E; ++I) { 1606 BasicBlock *UnavailablePred = I->first; 1607 1608 // Do PHI translation to get its value in the predecessor if necessary. The 1609 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1610 1611 // If all preds have a single successor, then we know it is safe to insert 1612 // the load on the pred (?!?), so we can insert code to materialize the 1613 // pointer if it is not available. 1614 PHITransAddr Address(LI->getPointerOperand(), TD); 1615 Value *LoadPtr = 0; 1616 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1617 *DT, NewInsts); 1618 1619 // If we couldn't find or insert a computation of this phi translated value, 1620 // we fail PRE. 1621 if (LoadPtr == 0) { 1622 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1623 << *LI->getPointerOperand() << "\n"); 1624 CanDoPRE = false; 1625 break; 1626 } 1627 1628 I->second = LoadPtr; 1629 } 1630 1631 if (!CanDoPRE) { 1632 while (!NewInsts.empty()) { 1633 Instruction *I = NewInsts.pop_back_val(); 1634 if (MD) MD->removeInstruction(I); 1635 I->eraseFromParent(); 1636 } 1637 // HINT:Don't revert the edge-splitting as following transformation may 1638 // also need to split these critial edges. 1639 return !CriticalEdgePred.empty(); 1640 } 1641 1642 // Okay, we can eliminate this load by inserting a reload in the predecessor 1643 // and using PHI construction to get the value in the other predecessors, do 1644 // it. 1645 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1646 DEBUG(if (!NewInsts.empty()) 1647 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1648 << *NewInsts.back() << '\n'); 1649 1650 // Assign value numbers to the new instructions. 1651 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1652 // FIXME: We really _ought_ to insert these value numbers into their 1653 // parent's availability map. However, in doing so, we risk getting into 1654 // ordering issues. If a block hasn't been processed yet, we would be 1655 // marking a value as AVAIL-IN, which isn't what we intend. 1656 VN.lookup_or_add(NewInsts[i]); 1657 } 1658 1659 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1660 E = PredLoads.end(); I != E; ++I) { 1661 BasicBlock *UnavailablePred = I->first; 1662 Value *LoadPtr = I->second; 1663 1664 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1665 LI->getAlignment(), 1666 UnavailablePred->getTerminator()); 1667 1668 // Transfer the old load's TBAA tag to the new load. 1669 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1670 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1671 1672 // Transfer DebugLoc. 1673 NewLoad->setDebugLoc(LI->getDebugLoc()); 1674 1675 // Add the newly created load. 1676 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1677 NewLoad)); 1678 MD->invalidateCachedPointerInfo(LoadPtr); 1679 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1680 } 1681 1682 // Perform PHI construction. 1683 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1684 LI->replaceAllUsesWith(V); 1685 if (isa<PHINode>(V)) 1686 V->takeName(LI); 1687 if (V->getType()->getScalarType()->isPointerTy()) 1688 MD->invalidateCachedPointerInfo(V); 1689 markInstructionForDeletion(LI); 1690 ++NumPRELoad; 1691 return true; 1692} 1693 1694/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1695/// non-local by performing PHI construction. 1696bool GVN::processNonLocalLoad(LoadInst *LI) { 1697 // Step 1: Find the non-local dependencies of the load. 1698 LoadDepVect Deps; 1699 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1700 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1701 1702 // If we had to process more than one hundred blocks to find the 1703 // dependencies, this load isn't worth worrying about. Optimizing 1704 // it will be too expensive. 1705 unsigned NumDeps = Deps.size(); 1706 if (NumDeps > 100) 1707 return false; 1708 1709 // If we had a phi translation failure, we'll have a single entry which is a 1710 // clobber in the current block. Reject this early. 1711 if (NumDeps == 1 && 1712 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1713 DEBUG( 1714 dbgs() << "GVN: non-local load "; 1715 WriteAsOperand(dbgs(), LI); 1716 dbgs() << " has unknown dependencies\n"; 1717 ); 1718 return false; 1719 } 1720 1721 // Step 2: Analyze the availability of the load 1722 AvailValInBlkVect ValuesPerBlock; 1723 UnavailBlkVect UnavailableBlocks; 1724 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks); 1725 1726 // If we have no predecessors that produce a known value for this load, exit 1727 // early. 1728 if (ValuesPerBlock.empty()) 1729 return false; 1730 1731 // Step 3: Eliminate fully redundancy. 1732 // 1733 // If all of the instructions we depend on produce a known value for this 1734 // load, then it is fully redundant and we can use PHI insertion to compute 1735 // its value. Insert PHIs and remove the fully redundant value now. 1736 if (UnavailableBlocks.empty()) { 1737 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1738 1739 // Perform PHI construction. 1740 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1741 LI->replaceAllUsesWith(V); 1742 1743 if (isa<PHINode>(V)) 1744 V->takeName(LI); 1745 if (V->getType()->getScalarType()->isPointerTy()) 1746 MD->invalidateCachedPointerInfo(V); 1747 markInstructionForDeletion(LI); 1748 ++NumGVNLoad; 1749 return true; 1750 } 1751 1752 // Step 4: Eliminate partial redundancy. 1753 if (!EnablePRE || !EnableLoadPRE) 1754 return false; 1755 1756 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks); 1757} 1758 1759 1760static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1761 // Patch the replacement so that it is not more restrictive than the value 1762 // being replaced. 1763 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1764 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1765 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1766 isa<OverflowingBinaryOperator>(ReplOp)) { 1767 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1768 ReplOp->setHasNoSignedWrap(false); 1769 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1770 ReplOp->setHasNoUnsignedWrap(false); 1771 } 1772 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1773 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1774 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1775 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1776 unsigned Kind = Metadata[i].first; 1777 MDNode *IMD = I->getMetadata(Kind); 1778 MDNode *ReplMD = Metadata[i].second; 1779 switch(Kind) { 1780 default: 1781 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata 1782 break; 1783 case LLVMContext::MD_dbg: 1784 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1785 case LLVMContext::MD_tbaa: 1786 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1787 break; 1788 case LLVMContext::MD_range: 1789 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1790 break; 1791 case LLVMContext::MD_prof: 1792 llvm_unreachable("MD_prof in a non terminator instruction"); 1793 break; 1794 case LLVMContext::MD_fpmath: 1795 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1796 break; 1797 } 1798 } 1799 } 1800} 1801 1802static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1803 patchReplacementInstruction(I, Repl); 1804 I->replaceAllUsesWith(Repl); 1805} 1806 1807/// processLoad - Attempt to eliminate a load, first by eliminating it 1808/// locally, and then attempting non-local elimination if that fails. 1809bool GVN::processLoad(LoadInst *L) { 1810 if (!MD) 1811 return false; 1812 1813 if (!L->isSimple()) 1814 return false; 1815 1816 if (L->use_empty()) { 1817 markInstructionForDeletion(L); 1818 return true; 1819 } 1820 1821 // ... to a pointer that has been loaded from before... 1822 MemDepResult Dep = MD->getDependency(L); 1823 1824 // If we have a clobber and target data is around, see if this is a clobber 1825 // that we can fix up through code synthesis. 1826 if (Dep.isClobber() && TD) { 1827 // Check to see if we have something like this: 1828 // store i32 123, i32* %P 1829 // %A = bitcast i32* %P to i8* 1830 // %B = gep i8* %A, i32 1 1831 // %C = load i8* %B 1832 // 1833 // We could do that by recognizing if the clobber instructions are obviously 1834 // a common base + constant offset, and if the previous store (or memset) 1835 // completely covers this load. This sort of thing can happen in bitfield 1836 // access code. 1837 Value *AvailVal = 0; 1838 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1839 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1840 L->getPointerOperand(), 1841 DepSI, *TD); 1842 if (Offset != -1) 1843 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1844 L->getType(), L, *TD); 1845 } 1846 1847 // Check to see if we have something like this: 1848 // load i32* P 1849 // load i8* (P+1) 1850 // if we have this, replace the later with an extraction from the former. 1851 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1852 // If this is a clobber and L is the first instruction in its block, then 1853 // we have the first instruction in the entry block. 1854 if (DepLI == L) 1855 return false; 1856 1857 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1858 L->getPointerOperand(), 1859 DepLI, *TD); 1860 if (Offset != -1) 1861 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1862 } 1863 1864 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1865 // a value on from it. 1866 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1867 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1868 L->getPointerOperand(), 1869 DepMI, *TD); 1870 if (Offset != -1) 1871 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1872 } 1873 1874 if (AvailVal) { 1875 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1876 << *AvailVal << '\n' << *L << "\n\n\n"); 1877 1878 // Replace the load! 1879 L->replaceAllUsesWith(AvailVal); 1880 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1881 MD->invalidateCachedPointerInfo(AvailVal); 1882 markInstructionForDeletion(L); 1883 ++NumGVNLoad; 1884 return true; 1885 } 1886 } 1887 1888 // If the value isn't available, don't do anything! 1889 if (Dep.isClobber()) { 1890 DEBUG( 1891 // fast print dep, using operator<< on instruction is too slow. 1892 dbgs() << "GVN: load "; 1893 WriteAsOperand(dbgs(), L); 1894 Instruction *I = Dep.getInst(); 1895 dbgs() << " is clobbered by " << *I << '\n'; 1896 ); 1897 return false; 1898 } 1899 1900 // If it is defined in another block, try harder. 1901 if (Dep.isNonLocal()) 1902 return processNonLocalLoad(L); 1903 1904 if (!Dep.isDef()) { 1905 DEBUG( 1906 // fast print dep, using operator<< on instruction is too slow. 1907 dbgs() << "GVN: load "; 1908 WriteAsOperand(dbgs(), L); 1909 dbgs() << " has unknown dependence\n"; 1910 ); 1911 return false; 1912 } 1913 1914 Instruction *DepInst = Dep.getInst(); 1915 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1916 Value *StoredVal = DepSI->getValueOperand(); 1917 1918 // The store and load are to a must-aliased pointer, but they may not 1919 // actually have the same type. See if we know how to reuse the stored 1920 // value (depending on its type). 1921 if (StoredVal->getType() != L->getType()) { 1922 if (TD) { 1923 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1924 L, *TD); 1925 if (StoredVal == 0) 1926 return false; 1927 1928 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1929 << '\n' << *L << "\n\n\n"); 1930 } 1931 else 1932 return false; 1933 } 1934 1935 // Remove it! 1936 L->replaceAllUsesWith(StoredVal); 1937 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1938 MD->invalidateCachedPointerInfo(StoredVal); 1939 markInstructionForDeletion(L); 1940 ++NumGVNLoad; 1941 return true; 1942 } 1943 1944 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1945 Value *AvailableVal = DepLI; 1946 1947 // The loads are of a must-aliased pointer, but they may not actually have 1948 // the same type. See if we know how to reuse the previously loaded value 1949 // (depending on its type). 1950 if (DepLI->getType() != L->getType()) { 1951 if (TD) { 1952 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1953 L, *TD); 1954 if (AvailableVal == 0) 1955 return false; 1956 1957 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1958 << "\n" << *L << "\n\n\n"); 1959 } 1960 else 1961 return false; 1962 } 1963 1964 // Remove it! 1965 patchAndReplaceAllUsesWith(L, AvailableVal); 1966 if (DepLI->getType()->getScalarType()->isPointerTy()) 1967 MD->invalidateCachedPointerInfo(DepLI); 1968 markInstructionForDeletion(L); 1969 ++NumGVNLoad; 1970 return true; 1971 } 1972 1973 // If this load really doesn't depend on anything, then we must be loading an 1974 // undef value. This can happen when loading for a fresh allocation with no 1975 // intervening stores, for example. 1976 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1977 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1978 markInstructionForDeletion(L); 1979 ++NumGVNLoad; 1980 return true; 1981 } 1982 1983 // If this load occurs either right after a lifetime begin, 1984 // then the loaded value is undefined. 1985 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1986 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1987 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1988 markInstructionForDeletion(L); 1989 ++NumGVNLoad; 1990 return true; 1991 } 1992 } 1993 1994 return false; 1995} 1996 1997// findLeader - In order to find a leader for a given value number at a 1998// specific basic block, we first obtain the list of all Values for that number, 1999// and then scan the list to find one whose block dominates the block in 2000// question. This is fast because dominator tree queries consist of only 2001// a few comparisons of DFS numbers. 2002Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 2003 LeaderTableEntry Vals = LeaderTable[num]; 2004 if (!Vals.Val) return 0; 2005 2006 Value *Val = 0; 2007 if (DT->dominates(Vals.BB, BB)) { 2008 Val = Vals.Val; 2009 if (isa<Constant>(Val)) return Val; 2010 } 2011 2012 LeaderTableEntry* Next = Vals.Next; 2013 while (Next) { 2014 if (DT->dominates(Next->BB, BB)) { 2015 if (isa<Constant>(Next->Val)) return Next->Val; 2016 if (!Val) Val = Next->Val; 2017 } 2018 2019 Next = Next->Next; 2020 } 2021 2022 return Val; 2023} 2024 2025/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 2026/// use is dominated by the given basic block. Returns the number of uses that 2027/// were replaced. 2028unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 2029 const BasicBlockEdge &Root) { 2030 unsigned Count = 0; 2031 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2032 UI != UE; ) { 2033 Use &U = (UI++).getUse(); 2034 2035 if (DT->dominates(Root, U)) { 2036 U.set(To); 2037 ++Count; 2038 } 2039 } 2040 return Count; 2041} 2042 2043/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2044/// true if every path from the entry block to 'Dst' passes via this edge. In 2045/// particular 'Dst' must not be reachable via another edge from 'Src'. 2046static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2047 DominatorTree *DT) { 2048 // While in theory it is interesting to consider the case in which Dst has 2049 // more than one predecessor, because Dst might be part of a loop which is 2050 // only reachable from Src, in practice it is pointless since at the time 2051 // GVN runs all such loops have preheaders, which means that Dst will have 2052 // been changed to have only one predecessor, namely Src. 2053 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2054 const BasicBlock *Src = E.getStart(); 2055 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2056 (void)Src; 2057 return Pred != 0; 2058} 2059 2060/// propagateEquality - The given values are known to be equal in every block 2061/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2062/// 'RHS' everywhere in the scope. Returns whether a change was made. 2063bool GVN::propagateEquality(Value *LHS, Value *RHS, 2064 const BasicBlockEdge &Root) { 2065 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2066 Worklist.push_back(std::make_pair(LHS, RHS)); 2067 bool Changed = false; 2068 // For speed, compute a conservative fast approximation to 2069 // DT->dominates(Root, Root.getEnd()); 2070 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2071 2072 while (!Worklist.empty()) { 2073 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2074 LHS = Item.first; RHS = Item.second; 2075 2076 if (LHS == RHS) continue; 2077 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2078 2079 // Don't try to propagate equalities between constants. 2080 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2081 2082 // Prefer a constant on the right-hand side, or an Argument if no constants. 2083 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2084 std::swap(LHS, RHS); 2085 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2086 2087 // If there is no obvious reason to prefer the left-hand side over the right- 2088 // hand side, ensure the longest lived term is on the right-hand side, so the 2089 // shortest lived term will be replaced by the longest lived. This tends to 2090 // expose more simplifications. 2091 uint32_t LVN = VN.lookup_or_add(LHS); 2092 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2093 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2094 // Move the 'oldest' value to the right-hand side, using the value number as 2095 // a proxy for age. 2096 uint32_t RVN = VN.lookup_or_add(RHS); 2097 if (LVN < RVN) { 2098 std::swap(LHS, RHS); 2099 LVN = RVN; 2100 } 2101 } 2102 2103 // If value numbering later sees that an instruction in the scope is equal 2104 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2105 // the invariant that instructions only occur in the leader table for their 2106 // own value number (this is used by removeFromLeaderTable), do not do this 2107 // if RHS is an instruction (if an instruction in the scope is morphed into 2108 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2109 // using the leader table is about compiling faster, not optimizing better). 2110 // The leader table only tracks basic blocks, not edges. Only add to if we 2111 // have the simple case where the edge dominates the end. 2112 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2113 addToLeaderTable(LVN, RHS, Root.getEnd()); 2114 2115 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2116 // LHS always has at least one use that is not dominated by Root, this will 2117 // never do anything if LHS has only one use. 2118 if (!LHS->hasOneUse()) { 2119 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2120 Changed |= NumReplacements > 0; 2121 NumGVNEqProp += NumReplacements; 2122 } 2123 2124 // Now try to deduce additional equalities from this one. For example, if the 2125 // known equality was "(A != B)" == "false" then it follows that A and B are 2126 // equal in the scope. Only boolean equalities with an explicit true or false 2127 // RHS are currently supported. 2128 if (!RHS->getType()->isIntegerTy(1)) 2129 // Not a boolean equality - bail out. 2130 continue; 2131 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2132 if (!CI) 2133 // RHS neither 'true' nor 'false' - bail out. 2134 continue; 2135 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2136 bool isKnownTrue = CI->isAllOnesValue(); 2137 bool isKnownFalse = !isKnownTrue; 2138 2139 // If "A && B" is known true then both A and B are known true. If "A || B" 2140 // is known false then both A and B are known false. 2141 Value *A, *B; 2142 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2143 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2144 Worklist.push_back(std::make_pair(A, RHS)); 2145 Worklist.push_back(std::make_pair(B, RHS)); 2146 continue; 2147 } 2148 2149 // If we are propagating an equality like "(A == B)" == "true" then also 2150 // propagate the equality A == B. When propagating a comparison such as 2151 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2152 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2153 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2154 2155 // If "A == B" is known true, or "A != B" is known false, then replace 2156 // A with B everywhere in the scope. 2157 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2158 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2159 Worklist.push_back(std::make_pair(Op0, Op1)); 2160 2161 // If "A >= B" is known true, replace "A < B" with false everywhere. 2162 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2163 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2164 // Since we don't have the instruction "A < B" immediately to hand, work out 2165 // the value number that it would have and use that to find an appropriate 2166 // instruction (if any). 2167 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2168 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2169 // If the number we were assigned was brand new then there is no point in 2170 // looking for an instruction realizing it: there cannot be one! 2171 if (Num < NextNum) { 2172 Value *NotCmp = findLeader(Root.getEnd(), Num); 2173 if (NotCmp && isa<Instruction>(NotCmp)) { 2174 unsigned NumReplacements = 2175 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2176 Changed |= NumReplacements > 0; 2177 NumGVNEqProp += NumReplacements; 2178 } 2179 } 2180 // Ensure that any instruction in scope that gets the "A < B" value number 2181 // is replaced with false. 2182 // The leader table only tracks basic blocks, not edges. Only add to if we 2183 // have the simple case where the edge dominates the end. 2184 if (RootDominatesEnd) 2185 addToLeaderTable(Num, NotVal, Root.getEnd()); 2186 2187 continue; 2188 } 2189 } 2190 2191 return Changed; 2192} 2193 2194/// processInstruction - When calculating availability, handle an instruction 2195/// by inserting it into the appropriate sets 2196bool GVN::processInstruction(Instruction *I) { 2197 // Ignore dbg info intrinsics. 2198 if (isa<DbgInfoIntrinsic>(I)) 2199 return false; 2200 2201 // If the instruction can be easily simplified then do so now in preference 2202 // to value numbering it. Value numbering often exposes redundancies, for 2203 // example if it determines that %y is equal to %x then the instruction 2204 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2205 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { 2206 I->replaceAllUsesWith(V); 2207 if (MD && V->getType()->getScalarType()->isPointerTy()) 2208 MD->invalidateCachedPointerInfo(V); 2209 markInstructionForDeletion(I); 2210 ++NumGVNSimpl; 2211 return true; 2212 } 2213 2214 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2215 if (processLoad(LI)) 2216 return true; 2217 2218 unsigned Num = VN.lookup_or_add(LI); 2219 addToLeaderTable(Num, LI, LI->getParent()); 2220 return false; 2221 } 2222 2223 // For conditional branches, we can perform simple conditional propagation on 2224 // the condition value itself. 2225 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2226 if (!BI->isConditional()) 2227 return false; 2228 2229 if (isa<Constant>(BI->getCondition())) 2230 return processFoldableCondBr(BI); 2231 2232 Value *BranchCond = BI->getCondition(); 2233 BasicBlock *TrueSucc = BI->getSuccessor(0); 2234 BasicBlock *FalseSucc = BI->getSuccessor(1); 2235 // Avoid multiple edges early. 2236 if (TrueSucc == FalseSucc) 2237 return false; 2238 2239 BasicBlock *Parent = BI->getParent(); 2240 bool Changed = false; 2241 2242 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2243 BasicBlockEdge TrueE(Parent, TrueSucc); 2244 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2245 2246 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2247 BasicBlockEdge FalseE(Parent, FalseSucc); 2248 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2249 2250 return Changed; 2251 } 2252 2253 // For switches, propagate the case values into the case destinations. 2254 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2255 Value *SwitchCond = SI->getCondition(); 2256 BasicBlock *Parent = SI->getParent(); 2257 bool Changed = false; 2258 2259 // Remember how many outgoing edges there are to every successor. 2260 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2261 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2262 ++SwitchEdges[SI->getSuccessor(i)]; 2263 2264 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2265 i != e; ++i) { 2266 BasicBlock *Dst = i.getCaseSuccessor(); 2267 // If there is only a single edge, propagate the case value into it. 2268 if (SwitchEdges.lookup(Dst) == 1) { 2269 BasicBlockEdge E(Parent, Dst); 2270 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2271 } 2272 } 2273 return Changed; 2274 } 2275 2276 // Instructions with void type don't return a value, so there's 2277 // no point in trying to find redundancies in them. 2278 if (I->getType()->isVoidTy()) return false; 2279 2280 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2281 unsigned Num = VN.lookup_or_add(I); 2282 2283 // Allocations are always uniquely numbered, so we can save time and memory 2284 // by fast failing them. 2285 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2286 addToLeaderTable(Num, I, I->getParent()); 2287 return false; 2288 } 2289 2290 // If the number we were assigned was a brand new VN, then we don't 2291 // need to do a lookup to see if the number already exists 2292 // somewhere in the domtree: it can't! 2293 if (Num >= NextNum) { 2294 addToLeaderTable(Num, I, I->getParent()); 2295 return false; 2296 } 2297 2298 // Perform fast-path value-number based elimination of values inherited from 2299 // dominators. 2300 Value *repl = findLeader(I->getParent(), Num); 2301 if (repl == 0) { 2302 // Failure, just remember this instance for future use. 2303 addToLeaderTable(Num, I, I->getParent()); 2304 return false; 2305 } 2306 2307 // Remove it! 2308 patchAndReplaceAllUsesWith(I, repl); 2309 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2310 MD->invalidateCachedPointerInfo(repl); 2311 markInstructionForDeletion(I); 2312 return true; 2313} 2314 2315/// runOnFunction - This is the main transformation entry point for a function. 2316bool GVN::runOnFunction(Function& F) { 2317 if (!NoLoads) 2318 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2319 DT = &getAnalysis<DominatorTree>(); 2320 TD = getAnalysisIfAvailable<DataLayout>(); 2321 TLI = &getAnalysis<TargetLibraryInfo>(); 2322 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2323 VN.setMemDep(MD); 2324 VN.setDomTree(DT); 2325 2326 bool Changed = false; 2327 bool ShouldContinue = true; 2328 2329 // Merge unconditional branches, allowing PRE to catch more 2330 // optimization opportunities. 2331 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2332 BasicBlock *BB = FI++; 2333 2334 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2335 if (removedBlock) ++NumGVNBlocks; 2336 2337 Changed |= removedBlock; 2338 } 2339 2340 unsigned Iteration = 0; 2341 while (ShouldContinue) { 2342 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2343 ShouldContinue = iterateOnFunction(F); 2344 Changed |= ShouldContinue; 2345 ++Iteration; 2346 } 2347 2348 if (EnablePRE) { 2349 // Fabricate val-num for dead-code in order to suppress assertion in 2350 // performPRE(). 2351 assignValNumForDeadCode(); 2352 bool PREChanged = true; 2353 while (PREChanged) { 2354 PREChanged = performPRE(F); 2355 Changed |= PREChanged; 2356 } 2357 } 2358 2359 // FIXME: Should perform GVN again after PRE does something. PRE can move 2360 // computations into blocks where they become fully redundant. Note that 2361 // we can't do this until PRE's critical edge splitting updates memdep. 2362 // Actually, when this happens, we should just fully integrate PRE into GVN. 2363 2364 cleanupGlobalSets(); 2365 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each 2366 // iteration. 2367 DeadBlocks.clear(); 2368 2369 return Changed; 2370} 2371 2372 2373bool GVN::processBlock(BasicBlock *BB) { 2374 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2375 // (and incrementing BI before processing an instruction). 2376 assert(InstrsToErase.empty() && 2377 "We expect InstrsToErase to be empty across iterations"); 2378 if (DeadBlocks.count(BB)) 2379 return false; 2380 2381 bool ChangedFunction = false; 2382 2383 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2384 BI != BE;) { 2385 ChangedFunction |= processInstruction(BI); 2386 if (InstrsToErase.empty()) { 2387 ++BI; 2388 continue; 2389 } 2390 2391 // If we need some instructions deleted, do it now. 2392 NumGVNInstr += InstrsToErase.size(); 2393 2394 // Avoid iterator invalidation. 2395 bool AtStart = BI == BB->begin(); 2396 if (!AtStart) 2397 --BI; 2398 2399 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(), 2400 E = InstrsToErase.end(); I != E; ++I) { 2401 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2402 if (MD) MD->removeInstruction(*I); 2403 DEBUG(verifyRemoved(*I)); 2404 (*I)->eraseFromParent(); 2405 } 2406 InstrsToErase.clear(); 2407 2408 if (AtStart) 2409 BI = BB->begin(); 2410 else 2411 ++BI; 2412 } 2413 2414 return ChangedFunction; 2415} 2416 2417/// performPRE - Perform a purely local form of PRE that looks for diamond 2418/// control flow patterns and attempts to perform simple PRE at the join point. 2419bool GVN::performPRE(Function &F) { 2420 bool Changed = false; 2421 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2422 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2423 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2424 BasicBlock *CurrentBlock = *DI; 2425 2426 // Nothing to PRE in the entry block. 2427 if (CurrentBlock == &F.getEntryBlock()) continue; 2428 2429 // Don't perform PRE on a landing pad. 2430 if (CurrentBlock->isLandingPad()) continue; 2431 2432 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2433 BE = CurrentBlock->end(); BI != BE; ) { 2434 Instruction *CurInst = BI++; 2435 2436 if (isa<AllocaInst>(CurInst) || 2437 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2438 CurInst->getType()->isVoidTy() || 2439 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2440 isa<DbgInfoIntrinsic>(CurInst)) 2441 continue; 2442 2443 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2444 // sinking the compare again, and it would force the code generator to 2445 // move the i1 from processor flags or predicate registers into a general 2446 // purpose register. 2447 if (isa<CmpInst>(CurInst)) 2448 continue; 2449 2450 // We don't currently value number ANY inline asm calls. 2451 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2452 if (CallI->isInlineAsm()) 2453 continue; 2454 2455 uint32_t ValNo = VN.lookup(CurInst); 2456 2457 // Look for the predecessors for PRE opportunities. We're 2458 // only trying to solve the basic diamond case, where 2459 // a value is computed in the successor and one predecessor, 2460 // but not the other. We also explicitly disallow cases 2461 // where the successor is its own predecessor, because they're 2462 // more complicated to get right. 2463 unsigned NumWith = 0; 2464 unsigned NumWithout = 0; 2465 BasicBlock *PREPred = 0; 2466 predMap.clear(); 2467 2468 for (pred_iterator PI = pred_begin(CurrentBlock), 2469 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2470 BasicBlock *P = *PI; 2471 // We're not interested in PRE where the block is its 2472 // own predecessor, or in blocks with predecessors 2473 // that are not reachable. 2474 if (P == CurrentBlock) { 2475 NumWithout = 2; 2476 break; 2477 } else if (!DT->isReachableFromEntry(P)) { 2478 NumWithout = 2; 2479 break; 2480 } 2481 2482 Value* predV = findLeader(P, ValNo); 2483 if (predV == 0) { 2484 predMap.push_back(std::make_pair(static_cast<Value *>(0), P)); 2485 PREPred = P; 2486 ++NumWithout; 2487 } else if (predV == CurInst) { 2488 /* CurInst dominates this predecessor. */ 2489 NumWithout = 2; 2490 break; 2491 } else { 2492 predMap.push_back(std::make_pair(predV, P)); 2493 ++NumWith; 2494 } 2495 } 2496 2497 // Don't do PRE when it might increase code size, i.e. when 2498 // we would need to insert instructions in more than one pred. 2499 if (NumWithout != 1 || NumWith == 0) 2500 continue; 2501 2502 // Don't do PRE across indirect branch. 2503 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2504 continue; 2505 2506 // We can't do PRE safely on a critical edge, so instead we schedule 2507 // the edge to be split and perform the PRE the next time we iterate 2508 // on the function. 2509 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2510 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2511 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2512 continue; 2513 } 2514 2515 // Instantiate the expression in the predecessor that lacked it. 2516 // Because we are going top-down through the block, all value numbers 2517 // will be available in the predecessor by the time we need them. Any 2518 // that weren't originally present will have been instantiated earlier 2519 // in this loop. 2520 Instruction *PREInstr = CurInst->clone(); 2521 bool success = true; 2522 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2523 Value *Op = PREInstr->getOperand(i); 2524 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2525 continue; 2526 2527 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2528 PREInstr->setOperand(i, V); 2529 } else { 2530 success = false; 2531 break; 2532 } 2533 } 2534 2535 // Fail out if we encounter an operand that is not available in 2536 // the PRE predecessor. This is typically because of loads which 2537 // are not value numbered precisely. 2538 if (!success) { 2539 DEBUG(verifyRemoved(PREInstr)); 2540 delete PREInstr; 2541 continue; 2542 } 2543 2544 PREInstr->insertBefore(PREPred->getTerminator()); 2545 PREInstr->setName(CurInst->getName() + ".pre"); 2546 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2547 VN.add(PREInstr, ValNo); 2548 ++NumGVNPRE; 2549 2550 // Update the availability map to include the new instruction. 2551 addToLeaderTable(ValNo, PREInstr, PREPred); 2552 2553 // Create a PHI to make the value available in this block. 2554 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2555 CurInst->getName() + ".pre-phi", 2556 CurrentBlock->begin()); 2557 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2558 if (Value *V = predMap[i].first) 2559 Phi->addIncoming(V, predMap[i].second); 2560 else 2561 Phi->addIncoming(PREInstr, PREPred); 2562 } 2563 2564 VN.add(Phi, ValNo); 2565 addToLeaderTable(ValNo, Phi, CurrentBlock); 2566 Phi->setDebugLoc(CurInst->getDebugLoc()); 2567 CurInst->replaceAllUsesWith(Phi); 2568 if (Phi->getType()->getScalarType()->isPointerTy()) { 2569 // Because we have added a PHI-use of the pointer value, it has now 2570 // "escaped" from alias analysis' perspective. We need to inform 2571 // AA of this. 2572 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2573 ++ii) { 2574 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2575 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2576 } 2577 2578 if (MD) 2579 MD->invalidateCachedPointerInfo(Phi); 2580 } 2581 VN.erase(CurInst); 2582 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2583 2584 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2585 if (MD) MD->removeInstruction(CurInst); 2586 DEBUG(verifyRemoved(CurInst)); 2587 CurInst->eraseFromParent(); 2588 Changed = true; 2589 } 2590 } 2591 2592 if (splitCriticalEdges()) 2593 Changed = true; 2594 2595 return Changed; 2596} 2597 2598/// Split the critical edge connecting the given two blocks, and return 2599/// the block inserted to the critical edge. 2600BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { 2601 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this); 2602 if (MD) 2603 MD->invalidateCachedPredecessors(); 2604 return BB; 2605} 2606 2607/// splitCriticalEdges - Split critical edges found during the previous 2608/// iteration that may enable further optimization. 2609bool GVN::splitCriticalEdges() { 2610 if (toSplit.empty()) 2611 return false; 2612 do { 2613 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2614 SplitCriticalEdge(Edge.first, Edge.second, this); 2615 } while (!toSplit.empty()); 2616 if (MD) MD->invalidateCachedPredecessors(); 2617 return true; 2618} 2619 2620/// iterateOnFunction - Executes one iteration of GVN 2621bool GVN::iterateOnFunction(Function &F) { 2622 cleanupGlobalSets(); 2623 2624 // Top-down walk of the dominator tree 2625 bool Changed = false; 2626#if 0 2627 // Needed for value numbering with phi construction to work. 2628 ReversePostOrderTraversal<Function*> RPOT(&F); 2629 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2630 RE = RPOT.end(); RI != RE; ++RI) 2631 Changed |= processBlock(*RI); 2632#else 2633 // Save the blocks this function have before transformation begins. GVN may 2634 // split critical edge, and hence may invalidate the RPO/DT iterator. 2635 // 2636 std::vector<BasicBlock *> BBVect; 2637 BBVect.reserve(256); 2638 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2639 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2640 BBVect.push_back(DI->getBlock()); 2641 2642 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end(); 2643 I != E; I++) 2644 Changed |= processBlock(*I); 2645#endif 2646 2647 return Changed; 2648} 2649 2650void GVN::cleanupGlobalSets() { 2651 VN.clear(); 2652 LeaderTable.clear(); 2653 TableAllocator.Reset(); 2654} 2655 2656/// verifyRemoved - Verify that the specified instruction does not occur in our 2657/// internal data structures. 2658void GVN::verifyRemoved(const Instruction *Inst) const { 2659 VN.verifyRemoved(Inst); 2660 2661 // Walk through the value number scope to make sure the instruction isn't 2662 // ferreted away in it. 2663 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2664 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2665 const LeaderTableEntry *Node = &I->second; 2666 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2667 2668 while (Node->Next) { 2669 Node = Node->Next; 2670 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2671 } 2672 } 2673} 2674 2675// BB is declared dead, which implied other blocks become dead as well. This 2676// function is to add all these blocks to "DeadBlocks". For the dead blocks' 2677// live successors, update their phi nodes by replacing the operands 2678// corresponding to dead blocks with UndefVal. 2679// 2680void GVN::addDeadBlock(BasicBlock *BB) { 2681 SmallVector<BasicBlock *, 4> NewDead; 2682 SmallSetVector<BasicBlock *, 4> DF; 2683 2684 NewDead.push_back(BB); 2685 while (!NewDead.empty()) { 2686 BasicBlock *D = NewDead.pop_back_val(); 2687 if (DeadBlocks.count(D)) 2688 continue; 2689 2690 // All blocks dominated by D are dead. 2691 SmallVector<BasicBlock *, 8> Dom; 2692 DT->getDescendants(D, Dom); 2693 DeadBlocks.insert(Dom.begin(), Dom.end()); 2694 2695 // Figure out the dominance-frontier(D). 2696 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(), 2697 E = Dom.end(); I != E; I++) { 2698 BasicBlock *B = *I; 2699 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) { 2700 BasicBlock *S = *SI; 2701 if (DeadBlocks.count(S)) 2702 continue; 2703 2704 bool AllPredDead = true; 2705 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++) 2706 if (!DeadBlocks.count(*PI)) { 2707 AllPredDead = false; 2708 break; 2709 } 2710 2711 if (!AllPredDead) { 2712 // S could be proved dead later on. That is why we don't update phi 2713 // operands at this moment. 2714 DF.insert(S); 2715 } else { 2716 // While S is not dominated by D, it is dead by now. This could take 2717 // place if S already have a dead predecessor before D is declared 2718 // dead. 2719 NewDead.push_back(S); 2720 } 2721 } 2722 } 2723 } 2724 2725 // For the dead blocks' live successors, update their phi nodes by replacing 2726 // the operands corresponding to dead blocks with UndefVal. 2727 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end(); 2728 I != E; I++) { 2729 BasicBlock *B = *I; 2730 if (DeadBlocks.count(B)) 2731 continue; 2732 2733 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B)); 2734 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(), 2735 PE = Preds.end(); PI != PE; PI++) { 2736 BasicBlock *P = *PI; 2737 2738 if (!DeadBlocks.count(P)) 2739 continue; 2740 2741 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) { 2742 if (BasicBlock *S = splitCriticalEdges(P, B)) 2743 DeadBlocks.insert(P = S); 2744 } 2745 2746 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) { 2747 PHINode &Phi = cast<PHINode>(*II); 2748 Phi.setIncomingValue(Phi.getBasicBlockIndex(P), 2749 UndefValue::get(Phi.getType())); 2750 } 2751 } 2752 } 2753} 2754 2755// If the given branch is recognized as a foldable branch (i.e. conditional 2756// branch with constant condition), it will perform following analyses and 2757// transformation. 2758// 1) If the dead out-coming edge is a critical-edge, split it. Let 2759// R be the target of the dead out-coming edge. 2760// 1) Identify the set of dead blocks implied by the branch's dead outcoming 2761// edge. The result of this step will be {X| X is dominated by R} 2762// 2) Identify those blocks which haves at least one dead prodecessor. The 2763// result of this step will be dominance-frontier(R). 2764// 3) Update the PHIs in DF(R) by replacing the operands corresponding to 2765// dead blocks with "UndefVal" in an hope these PHIs will optimized away. 2766// 2767// Return true iff *NEW* dead code are found. 2768bool GVN::processFoldableCondBr(BranchInst *BI) { 2769 if (!BI || BI->isUnconditional()) 2770 return false; 2771 2772 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition()); 2773 if (!Cond) 2774 return false; 2775 2776 BasicBlock *DeadRoot = Cond->getZExtValue() ? 2777 BI->getSuccessor(1) : BI->getSuccessor(0); 2778 if (DeadBlocks.count(DeadRoot)) 2779 return false; 2780 2781 if (!DeadRoot->getSinglePredecessor()) 2782 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot); 2783 2784 addDeadBlock(DeadRoot); 2785 return true; 2786} 2787 2788// performPRE() will trigger assert if it come across an instruciton without 2789// associated val-num. As it normally has far more live instructions than dead 2790// instructions, it makes more sense just to "fabricate" a val-number for the 2791// dead code than checking if instruction involved is dead or not. 2792void GVN::assignValNumForDeadCode() { 2793 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(), 2794 E = DeadBlocks.end(); I != E; I++) { 2795 BasicBlock *BB = *I; 2796 for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); 2797 II != EE; II++) { 2798 Instruction *Inst = &*II; 2799 unsigned ValNum = VN.lookup_or_add(Inst); 2800 addToLeaderTable(ValNum, Inst, BB); 2801 } 2802 } 2803} 2804