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 #include "llvm/Transforms/Scalar.h" 19 #include "llvm/ADT/DenseMap.h" 20 #include "llvm/ADT/DepthFirstIterator.h" 21 #include "llvm/ADT/Hashing.h" 22 #include "llvm/ADT/MapVector.h" 23 #include "llvm/ADT/SetVector.h" 24 #include "llvm/ADT/SmallPtrSet.h" 25 #include "llvm/ADT/Statistic.h" 26 #include "llvm/Analysis/AliasAnalysis.h" 27 #include "llvm/Analysis/CFG.h" 28 #include "llvm/Analysis/ConstantFolding.h" 29 #include "llvm/Analysis/InstructionSimplify.h" 30 #include "llvm/Analysis/Loads.h" 31 #include "llvm/Analysis/MemoryBuiltins.h" 32 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 33 #include "llvm/Analysis/PHITransAddr.h" 34 #include "llvm/Analysis/ValueTracking.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/Dominators.h" 37 #include "llvm/IR/GlobalVariable.h" 38 #include "llvm/IR/IRBuilder.h" 39 #include "llvm/IR/IntrinsicInst.h" 40 #include "llvm/IR/LLVMContext.h" 41 #include "llvm/IR/Metadata.h" 42 #include "llvm/IR/PatternMatch.h" 43 #include "llvm/Support/Allocator.h" 44 #include "llvm/Support/CommandLine.h" 45 #include "llvm/Support/Debug.h" 46 #include "llvm/Target/TargetLibraryInfo.h" 47 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 48 #include "llvm/Transforms/Utils/SSAUpdater.h" 49 #include <vector> 50 using namespace llvm; 51 using namespace PatternMatch; 52 53 #define DEBUG_TYPE "gvn" 54 55 STATISTIC(NumGVNInstr, "Number of instructions deleted"); 56 STATISTIC(NumGVNLoad, "Number of loads deleted"); 57 STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 58 STATISTIC(NumGVNBlocks, "Number of blocks merged"); 59 STATISTIC(NumGVNSimpl, "Number of instructions simplified"); 60 STATISTIC(NumGVNEqProp, "Number of equalities propagated"); 61 STATISTIC(NumPRELoad, "Number of loads PRE'd"); 62 63 static cl::opt<bool> EnablePRE("enable-pre", 64 cl::init(true), cl::Hidden); 65 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 66 67 // Maximum allowed recursion depth. 68 static cl::opt<uint32_t> 69 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, 70 cl::desc("Max recurse depth (default = 1000)")); 71 72 //===----------------------------------------------------------------------===// 73 // ValueTable Class 74 //===----------------------------------------------------------------------===// 75 76 /// This class holds the mapping between values and value numbers. It is used 77 /// as an efficient mechanism to determine the expression-wise equivalence of 78 /// two values. 79 namespace { 80 struct Expression { 81 uint32_t opcode; 82 Type *type; 83 SmallVector<uint32_t, 4> varargs; 84 85 Expression(uint32_t o = ~2U) : opcode(o) { } 86 87 bool operator==(const Expression &other) const { 88 if (opcode != other.opcode) 89 return false; 90 if (opcode == ~0U || opcode == ~1U) 91 return true; 92 if (type != other.type) 93 return false; 94 if (varargs != other.varargs) 95 return false; 96 return true; 97 } 98 99 friend hash_code hash_value(const Expression &Value) { 100 return hash_combine(Value.opcode, Value.type, 101 hash_combine_range(Value.varargs.begin(), 102 Value.varargs.end())); 103 } 104 }; 105 106 class ValueTable { 107 DenseMap<Value*, uint32_t> valueNumbering; 108 DenseMap<Expression, uint32_t> expressionNumbering; 109 AliasAnalysis *AA; 110 MemoryDependenceAnalysis *MD; 111 DominatorTree *DT; 112 113 uint32_t nextValueNumber; 114 115 Expression create_expression(Instruction* I); 116 Expression create_cmp_expression(unsigned Opcode, 117 CmpInst::Predicate Predicate, 118 Value *LHS, Value *RHS); 119 Expression create_extractvalue_expression(ExtractValueInst* EI); 120 uint32_t lookup_or_add_call(CallInst* C); 121 public: 122 ValueTable() : nextValueNumber(1) { } 123 uint32_t lookup_or_add(Value *V); 124 uint32_t lookup(Value *V) const; 125 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred, 126 Value *LHS, Value *RHS); 127 void add(Value *V, uint32_t num); 128 void clear(); 129 void erase(Value *v); 130 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 131 AliasAnalysis *getAliasAnalysis() const { return AA; } 132 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 133 void setDomTree(DominatorTree* D) { DT = D; } 134 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 135 void verifyRemoved(const Value *) const; 136 }; 137 } 138 139 namespace llvm { 140 template <> struct DenseMapInfo<Expression> { 141 static inline Expression getEmptyKey() { 142 return ~0U; 143 } 144 145 static inline Expression getTombstoneKey() { 146 return ~1U; 147 } 148 149 static unsigned getHashValue(const Expression e) { 150 using llvm::hash_value; 151 return static_cast<unsigned>(hash_value(e)); 152 } 153 static bool isEqual(const Expression &LHS, const Expression &RHS) { 154 return LHS == RHS; 155 } 156 }; 157 158 } 159 160 //===----------------------------------------------------------------------===// 161 // ValueTable Internal Functions 162 //===----------------------------------------------------------------------===// 163 164 Expression ValueTable::create_expression(Instruction *I) { 165 Expression e; 166 e.type = I->getType(); 167 e.opcode = I->getOpcode(); 168 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 169 OI != OE; ++OI) 170 e.varargs.push_back(lookup_or_add(*OI)); 171 if (I->isCommutative()) { 172 // Ensure that commutative instructions that only differ by a permutation 173 // of their operands get the same value number by sorting the operand value 174 // numbers. Since all commutative instructions have two operands it is more 175 // efficient to sort by hand rather than using, say, std::sort. 176 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!"); 177 if (e.varargs[0] > e.varargs[1]) 178 std::swap(e.varargs[0], e.varargs[1]); 179 } 180 181 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 182 // Sort the operand value numbers so x<y and y>x get the same value number. 183 CmpInst::Predicate Predicate = C->getPredicate(); 184 if (e.varargs[0] > e.varargs[1]) { 185 std::swap(e.varargs[0], e.varargs[1]); 186 Predicate = CmpInst::getSwappedPredicate(Predicate); 187 } 188 e.opcode = (C->getOpcode() << 8) | Predicate; 189 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 190 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 191 II != IE; ++II) 192 e.varargs.push_back(*II); 193 } 194 195 return e; 196 } 197 198 Expression ValueTable::create_cmp_expression(unsigned Opcode, 199 CmpInst::Predicate Predicate, 200 Value *LHS, Value *RHS) { 201 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && 202 "Not a comparison!"); 203 Expression e; 204 e.type = CmpInst::makeCmpResultType(LHS->getType()); 205 e.varargs.push_back(lookup_or_add(LHS)); 206 e.varargs.push_back(lookup_or_add(RHS)); 207 208 // Sort the operand value numbers so x<y and y>x get the same value number. 209 if (e.varargs[0] > e.varargs[1]) { 210 std::swap(e.varargs[0], e.varargs[1]); 211 Predicate = CmpInst::getSwappedPredicate(Predicate); 212 } 213 e.opcode = (Opcode << 8) | Predicate; 214 return e; 215 } 216 217 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) { 218 assert(EI && "Not an ExtractValueInst?"); 219 Expression e; 220 e.type = EI->getType(); 221 e.opcode = 0; 222 223 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand()); 224 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) { 225 // EI might be an extract from one of our recognised intrinsics. If it 226 // is we'll synthesize a semantically equivalent expression instead on 227 // an extract value expression. 228 switch (I->getIntrinsicID()) { 229 case Intrinsic::sadd_with_overflow: 230 case Intrinsic::uadd_with_overflow: 231 e.opcode = Instruction::Add; 232 break; 233 case Intrinsic::ssub_with_overflow: 234 case Intrinsic::usub_with_overflow: 235 e.opcode = Instruction::Sub; 236 break; 237 case Intrinsic::smul_with_overflow: 238 case Intrinsic::umul_with_overflow: 239 e.opcode = Instruction::Mul; 240 break; 241 default: 242 break; 243 } 244 245 if (e.opcode != 0) { 246 // Intrinsic recognized. Grab its args to finish building the expression. 247 assert(I->getNumArgOperands() == 2 && 248 "Expect two args for recognised intrinsics."); 249 e.varargs.push_back(lookup_or_add(I->getArgOperand(0))); 250 e.varargs.push_back(lookup_or_add(I->getArgOperand(1))); 251 return e; 252 } 253 } 254 255 // Not a recognised intrinsic. Fall back to producing an extract value 256 // expression. 257 e.opcode = EI->getOpcode(); 258 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end(); 259 OI != OE; ++OI) 260 e.varargs.push_back(lookup_or_add(*OI)); 261 262 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end(); 263 II != IE; ++II) 264 e.varargs.push_back(*II); 265 266 return e; 267 } 268 269 //===----------------------------------------------------------------------===// 270 // ValueTable External Functions 271 //===----------------------------------------------------------------------===// 272 273 /// add - Insert a value into the table with a specified value number. 274 void ValueTable::add(Value *V, uint32_t num) { 275 valueNumbering.insert(std::make_pair(V, num)); 276 } 277 278 uint32_t ValueTable::lookup_or_add_call(CallInst *C) { 279 if (AA->doesNotAccessMemory(C)) { 280 Expression exp = create_expression(C); 281 uint32_t &e = expressionNumbering[exp]; 282 if (!e) e = nextValueNumber++; 283 valueNumbering[C] = e; 284 return e; 285 } else if (AA->onlyReadsMemory(C)) { 286 Expression exp = create_expression(C); 287 uint32_t &e = expressionNumbering[exp]; 288 if (!e) { 289 e = nextValueNumber++; 290 valueNumbering[C] = e; 291 return e; 292 } 293 if (!MD) { 294 e = nextValueNumber++; 295 valueNumbering[C] = e; 296 return e; 297 } 298 299 MemDepResult local_dep = MD->getDependency(C); 300 301 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 302 valueNumbering[C] = nextValueNumber; 303 return nextValueNumber++; 304 } 305 306 if (local_dep.isDef()) { 307 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 308 309 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 310 valueNumbering[C] = nextValueNumber; 311 return nextValueNumber++; 312 } 313 314 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 315 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 316 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 317 if (c_vn != cd_vn) { 318 valueNumbering[C] = nextValueNumber; 319 return nextValueNumber++; 320 } 321 } 322 323 uint32_t v = lookup_or_add(local_cdep); 324 valueNumbering[C] = v; 325 return v; 326 } 327 328 // Non-local case. 329 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 330 MD->getNonLocalCallDependency(CallSite(C)); 331 // FIXME: Move the checking logic to MemDep! 332 CallInst* cdep = nullptr; 333 334 // Check to see if we have a single dominating call instruction that is 335 // identical to C. 336 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 337 const NonLocalDepEntry *I = &deps[i]; 338 if (I->getResult().isNonLocal()) 339 continue; 340 341 // We don't handle non-definitions. If we already have a call, reject 342 // instruction dependencies. 343 if (!I->getResult().isDef() || cdep != nullptr) { 344 cdep = nullptr; 345 break; 346 } 347 348 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 349 // FIXME: All duplicated with non-local case. 350 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 351 cdep = NonLocalDepCall; 352 continue; 353 } 354 355 cdep = nullptr; 356 break; 357 } 358 359 if (!cdep) { 360 valueNumbering[C] = nextValueNumber; 361 return nextValueNumber++; 362 } 363 364 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 365 valueNumbering[C] = nextValueNumber; 366 return nextValueNumber++; 367 } 368 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 369 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 370 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 371 if (c_vn != cd_vn) { 372 valueNumbering[C] = nextValueNumber; 373 return nextValueNumber++; 374 } 375 } 376 377 uint32_t v = lookup_or_add(cdep); 378 valueNumbering[C] = v; 379 return v; 380 381 } else { 382 valueNumbering[C] = nextValueNumber; 383 return nextValueNumber++; 384 } 385 } 386 387 /// lookup_or_add - Returns the value number for the specified value, assigning 388 /// it a new number if it did not have one before. 389 uint32_t ValueTable::lookup_or_add(Value *V) { 390 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 391 if (VI != valueNumbering.end()) 392 return VI->second; 393 394 if (!isa<Instruction>(V)) { 395 valueNumbering[V] = nextValueNumber; 396 return nextValueNumber++; 397 } 398 399 Instruction* I = cast<Instruction>(V); 400 Expression exp; 401 switch (I->getOpcode()) { 402 case Instruction::Call: 403 return lookup_or_add_call(cast<CallInst>(I)); 404 case Instruction::Add: 405 case Instruction::FAdd: 406 case Instruction::Sub: 407 case Instruction::FSub: 408 case Instruction::Mul: 409 case Instruction::FMul: 410 case Instruction::UDiv: 411 case Instruction::SDiv: 412 case Instruction::FDiv: 413 case Instruction::URem: 414 case Instruction::SRem: 415 case Instruction::FRem: 416 case Instruction::Shl: 417 case Instruction::LShr: 418 case Instruction::AShr: 419 case Instruction::And: 420 case Instruction::Or: 421 case Instruction::Xor: 422 case Instruction::ICmp: 423 case Instruction::FCmp: 424 case Instruction::Trunc: 425 case Instruction::ZExt: 426 case Instruction::SExt: 427 case Instruction::FPToUI: 428 case Instruction::FPToSI: 429 case Instruction::UIToFP: 430 case Instruction::SIToFP: 431 case Instruction::FPTrunc: 432 case Instruction::FPExt: 433 case Instruction::PtrToInt: 434 case Instruction::IntToPtr: 435 case Instruction::BitCast: 436 case Instruction::Select: 437 case Instruction::ExtractElement: 438 case Instruction::InsertElement: 439 case Instruction::ShuffleVector: 440 case Instruction::InsertValue: 441 case Instruction::GetElementPtr: 442 exp = create_expression(I); 443 break; 444 case Instruction::ExtractValue: 445 exp = create_extractvalue_expression(cast<ExtractValueInst>(I)); 446 break; 447 default: 448 valueNumbering[V] = nextValueNumber; 449 return nextValueNumber++; 450 } 451 452 uint32_t& e = expressionNumbering[exp]; 453 if (!e) e = nextValueNumber++; 454 valueNumbering[V] = e; 455 return e; 456 } 457 458 /// lookup - Returns the value number of the specified value. Fails if 459 /// the value has not yet been numbered. 460 uint32_t ValueTable::lookup(Value *V) const { 461 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 462 assert(VI != valueNumbering.end() && "Value not numbered?"); 463 return VI->second; 464 } 465 466 /// lookup_or_add_cmp - Returns the value number of the given comparison, 467 /// assigning it a new number if it did not have one before. Useful when 468 /// we deduced the result of a comparison, but don't immediately have an 469 /// instruction realizing that comparison to hand. 470 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode, 471 CmpInst::Predicate Predicate, 472 Value *LHS, Value *RHS) { 473 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS); 474 uint32_t& e = expressionNumbering[exp]; 475 if (!e) e = nextValueNumber++; 476 return e; 477 } 478 479 /// clear - Remove all entries from the ValueTable. 480 void ValueTable::clear() { 481 valueNumbering.clear(); 482 expressionNumbering.clear(); 483 nextValueNumber = 1; 484 } 485 486 /// erase - Remove a value from the value numbering. 487 void ValueTable::erase(Value *V) { 488 valueNumbering.erase(V); 489 } 490 491 /// verifyRemoved - Verify that the value is removed from all internal data 492 /// structures. 493 void ValueTable::verifyRemoved(const Value *V) const { 494 for (DenseMap<Value*, uint32_t>::const_iterator 495 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 496 assert(I->first != V && "Inst still occurs in value numbering map!"); 497 } 498 } 499 500 //===----------------------------------------------------------------------===// 501 // GVN Pass 502 //===----------------------------------------------------------------------===// 503 504 namespace { 505 class GVN; 506 struct AvailableValueInBlock { 507 /// BB - The basic block in question. 508 BasicBlock *BB; 509 enum ValType { 510 SimpleVal, // A simple offsetted value that is accessed. 511 LoadVal, // A value produced by a load. 512 MemIntrin, // A memory intrinsic which is loaded from. 513 UndefVal // A UndefValue representing a value from dead block (which 514 // is not yet physically removed from the CFG). 515 }; 516 517 /// V - The value that is live out of the block. 518 PointerIntPair<Value *, 2, ValType> Val; 519 520 /// Offset - The byte offset in Val that is interesting for the load query. 521 unsigned Offset; 522 523 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 524 unsigned Offset = 0) { 525 AvailableValueInBlock Res; 526 Res.BB = BB; 527 Res.Val.setPointer(V); 528 Res.Val.setInt(SimpleVal); 529 Res.Offset = Offset; 530 return Res; 531 } 532 533 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 534 unsigned Offset = 0) { 535 AvailableValueInBlock Res; 536 Res.BB = BB; 537 Res.Val.setPointer(MI); 538 Res.Val.setInt(MemIntrin); 539 Res.Offset = Offset; 540 return Res; 541 } 542 543 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 544 unsigned Offset = 0) { 545 AvailableValueInBlock Res; 546 Res.BB = BB; 547 Res.Val.setPointer(LI); 548 Res.Val.setInt(LoadVal); 549 Res.Offset = Offset; 550 return Res; 551 } 552 553 static AvailableValueInBlock getUndef(BasicBlock *BB) { 554 AvailableValueInBlock Res; 555 Res.BB = BB; 556 Res.Val.setPointer(nullptr); 557 Res.Val.setInt(UndefVal); 558 Res.Offset = 0; 559 return Res; 560 } 561 562 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 563 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 564 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 565 bool isUndefValue() const { return Val.getInt() == UndefVal; } 566 567 Value *getSimpleValue() const { 568 assert(isSimpleValue() && "Wrong accessor"); 569 return Val.getPointer(); 570 } 571 572 LoadInst *getCoercedLoadValue() const { 573 assert(isCoercedLoadValue() && "Wrong accessor"); 574 return cast<LoadInst>(Val.getPointer()); 575 } 576 577 MemIntrinsic *getMemIntrinValue() const { 578 assert(isMemIntrinValue() && "Wrong accessor"); 579 return cast<MemIntrinsic>(Val.getPointer()); 580 } 581 582 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 583 /// defined here to the specified type. This handles various coercion cases. 584 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const; 585 }; 586 587 class GVN : public FunctionPass { 588 bool NoLoads; 589 MemoryDependenceAnalysis *MD; 590 DominatorTree *DT; 591 const DataLayout *DL; 592 const TargetLibraryInfo *TLI; 593 SetVector<BasicBlock *> DeadBlocks; 594 595 ValueTable VN; 596 597 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 598 /// have that value number. Use findLeader to query it. 599 struct LeaderTableEntry { 600 Value *Val; 601 const BasicBlock *BB; 602 LeaderTableEntry *Next; 603 }; 604 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 605 BumpPtrAllocator TableAllocator; 606 607 SmallVector<Instruction*, 8> InstrsToErase; 608 609 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect; 610 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect; 611 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect; 612 613 public: 614 static char ID; // Pass identification, replacement for typeid 615 explicit GVN(bool noloads = false) 616 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) { 617 initializeGVNPass(*PassRegistry::getPassRegistry()); 618 } 619 620 bool runOnFunction(Function &F) override; 621 622 /// markInstructionForDeletion - This removes the specified instruction from 623 /// our various maps and marks it for deletion. 624 void markInstructionForDeletion(Instruction *I) { 625 VN.erase(I); 626 InstrsToErase.push_back(I); 627 } 628 629 const DataLayout *getDataLayout() const { return DL; } 630 DominatorTree &getDominatorTree() const { return *DT; } 631 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } 632 MemoryDependenceAnalysis &getMemDep() const { return *MD; } 633 private: 634 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 635 /// its value number. 636 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) { 637 LeaderTableEntry &Curr = LeaderTable[N]; 638 if (!Curr.Val) { 639 Curr.Val = V; 640 Curr.BB = BB; 641 return; 642 } 643 644 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>(); 645 Node->Val = V; 646 Node->BB = BB; 647 Node->Next = Curr.Next; 648 Curr.Next = Node; 649 } 650 651 /// removeFromLeaderTable - Scan the list of values corresponding to a given 652 /// value number, and remove the given instruction if encountered. 653 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) { 654 LeaderTableEntry* Prev = nullptr; 655 LeaderTableEntry* Curr = &LeaderTable[N]; 656 657 while (Curr->Val != I || Curr->BB != BB) { 658 Prev = Curr; 659 Curr = Curr->Next; 660 } 661 662 if (Prev) { 663 Prev->Next = Curr->Next; 664 } else { 665 if (!Curr->Next) { 666 Curr->Val = nullptr; 667 Curr->BB = nullptr; 668 } else { 669 LeaderTableEntry* Next = Curr->Next; 670 Curr->Val = Next->Val; 671 Curr->BB = Next->BB; 672 Curr->Next = Next->Next; 673 } 674 } 675 } 676 677 // List of critical edges to be split between iterations. 678 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 679 680 // This transformation requires dominator postdominator info 681 void getAnalysisUsage(AnalysisUsage &AU) const override { 682 AU.addRequired<DominatorTreeWrapperPass>(); 683 AU.addRequired<TargetLibraryInfo>(); 684 if (!NoLoads) 685 AU.addRequired<MemoryDependenceAnalysis>(); 686 AU.addRequired<AliasAnalysis>(); 687 688 AU.addPreserved<DominatorTreeWrapperPass>(); 689 AU.addPreserved<AliasAnalysis>(); 690 } 691 692 693 // Helper fuctions of redundant load elimination 694 bool processLoad(LoadInst *L); 695 bool processNonLocalLoad(LoadInst *L); 696 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 697 AvailValInBlkVect &ValuesPerBlock, 698 UnavailBlkVect &UnavailableBlocks); 699 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 700 UnavailBlkVect &UnavailableBlocks); 701 702 // Other helper routines 703 bool processInstruction(Instruction *I); 704 bool processBlock(BasicBlock *BB); 705 void dump(DenseMap<uint32_t, Value*> &d); 706 bool iterateOnFunction(Function &F); 707 bool performPRE(Function &F); 708 Value *findLeader(const BasicBlock *BB, uint32_t num); 709 void cleanupGlobalSets(); 710 void verifyRemoved(const Instruction *I) const; 711 bool splitCriticalEdges(); 712 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ); 713 unsigned replaceAllDominatedUsesWith(Value *From, Value *To, 714 const BasicBlockEdge &Root); 715 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root); 716 bool processFoldableCondBr(BranchInst *BI); 717 void addDeadBlock(BasicBlock *BB); 718 void assignValNumForDeadCode(); 719 }; 720 721 char GVN::ID = 0; 722 } 723 724 // createGVNPass - The public interface to this file... 725 FunctionPass *llvm::createGVNPass(bool NoLoads) { 726 return new GVN(NoLoads); 727 } 728 729 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 730 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 731 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 732 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 733 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 734 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 735 736 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 737 void GVN::dump(DenseMap<uint32_t, Value*>& d) { 738 errs() << "{\n"; 739 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 740 E = d.end(); I != E; ++I) { 741 errs() << I->first << "\n"; 742 I->second->dump(); 743 } 744 errs() << "}\n"; 745 } 746 #endif 747 748 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value 749 /// we're analyzing is fully available in the specified block. As we go, keep 750 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This 751 /// map is actually a tri-state map with the following values: 752 /// 0) we know the block *is not* fully available. 753 /// 1) we know the block *is* fully available. 754 /// 2) we do not know whether the block is fully available or not, but we are 755 /// currently speculating that it will be. 756 /// 3) we are speculating for this block and have used that to speculate for 757 /// other blocks. 758 static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 759 DenseMap<BasicBlock*, char> &FullyAvailableBlocks, 760 uint32_t RecurseDepth) { 761 if (RecurseDepth > MaxRecurseDepth) 762 return false; 763 764 // Optimistically assume that the block is fully available and check to see 765 // if we already know about this block in one lookup. 766 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 767 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 768 769 // If the entry already existed for this block, return the precomputed value. 770 if (!IV.second) { 771 // If this is a speculative "available" value, mark it as being used for 772 // speculation of other blocks. 773 if (IV.first->second == 2) 774 IV.first->second = 3; 775 return IV.first->second != 0; 776 } 777 778 // Otherwise, see if it is fully available in all predecessors. 779 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 780 781 // If this block has no predecessors, it isn't live-in here. 782 if (PI == PE) 783 goto SpeculationFailure; 784 785 for (; PI != PE; ++PI) 786 // If the value isn't fully available in one of our predecessors, then it 787 // isn't fully available in this block either. Undo our previous 788 // optimistic assumption and bail out. 789 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1)) 790 goto SpeculationFailure; 791 792 return true; 793 794 // SpeculationFailure - If we get here, we found out that this is not, after 795 // all, a fully-available block. We have a problem if we speculated on this and 796 // used the speculation to mark other blocks as available. 797 SpeculationFailure: 798 char &BBVal = FullyAvailableBlocks[BB]; 799 800 // If we didn't speculate on this, just return with it set to false. 801 if (BBVal == 2) { 802 BBVal = 0; 803 return false; 804 } 805 806 // If we did speculate on this value, we could have blocks set to 1 that are 807 // incorrect. Walk the (transitive) successors of this block and mark them as 808 // 0 if set to one. 809 SmallVector<BasicBlock*, 32> BBWorklist; 810 BBWorklist.push_back(BB); 811 812 do { 813 BasicBlock *Entry = BBWorklist.pop_back_val(); 814 // Note that this sets blocks to 0 (unavailable) if they happen to not 815 // already be in FullyAvailableBlocks. This is safe. 816 char &EntryVal = FullyAvailableBlocks[Entry]; 817 if (EntryVal == 0) continue; // Already unavailable. 818 819 // Mark as unavailable. 820 EntryVal = 0; 821 822 BBWorklist.append(succ_begin(Entry), succ_end(Entry)); 823 } while (!BBWorklist.empty()); 824 825 return false; 826 } 827 828 829 /// CanCoerceMustAliasedValueToLoad - Return true if 830 /// CoerceAvailableValueToLoadType will succeed. 831 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 832 Type *LoadTy, 833 const DataLayout &DL) { 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 (DL.getTypeSizeInBits(StoredVal->getType()) < 843 DL.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. 855 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 856 Type *LoadedTy, 857 Instruction *InsertPt, 858 const DataLayout &DL) { 859 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL)) 860 return nullptr; 861 862 // If this is already the right type, just return it. 863 Type *StoredValTy = StoredVal->getType(); 864 865 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy); 866 uint64_t LoadSize = DL.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 = DL.getIntPtrType(StoredValTy); 878 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 879 } 880 881 Type *TypeToCastTo = LoadedTy; 882 if (TypeToCastTo->getScalarType()->isPointerTy()) 883 TypeToCastTo = DL.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 = DL.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 (DL.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. 942 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 943 Value *WritePtr, 944 uint64_t WriteSizeInBits, 945 const DataLayout &DL) { 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,&DL); 953 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL); 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 = DL.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. 1016 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 1017 StoreInst *DepSI, 1018 const DataLayout &DL) { 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 =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1026 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1027 StorePtr, StoreSize, DL); 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. 1033 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, 1034 LoadInst *DepLI, const DataLayout &DL){ 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 = DL.getTypeSizeInBits(DepLI->getType()); 1041 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL); 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, &DL); 1049 unsigned LoadSize = DL.getTypeStoreSize(LoadTy); 1050 1051 unsigned Size = MemoryDependenceAnalysis:: 1052 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL); 1053 if (Size == 0) return -1; 1054 1055 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL); 1056 } 1057 1058 1059 1060 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 1061 MemIntrinsic *MI, 1062 const DataLayout &DL) { 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) 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, DL); 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) return -1; 1081 1082 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL)); 1083 if (!GV || !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, DL); 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, &DL)) 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. 1111 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1112 Type *LoadTy, 1113 Instruction *InsertPt, const DataLayout &DL){ 1114 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1115 1116 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1117 uint64_t LoadSize = (DL.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 DL.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 (DL.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, DL); 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. 1150 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 1151 Type *LoadTy, Instruction *InsertPt, 1152 GVN &gvn) { 1153 const DataLayout &DL = *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 = DL.getTypeStoreSize(SrcVal->getType()); 1157 unsigned LoadSize = DL.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 (DL.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, DL); 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. 1210 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1211 Type *LoadTy, Instruction *InsertPt, 1212 const DataLayout &DL){ 1213 LLVMContext &Ctx = LoadTy->getContext(); 1214 uint64_t LoadSize = DL.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, DL); 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, &DL); 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. 1268 static 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 1322 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const { 1323 Value *Res; 1324 if (isSimpleValue()) { 1325 Res = getSimpleValue(); 1326 if (Res->getType() != LoadTy) { 1327 const DataLayout *DL = gvn.getDataLayout(); 1328 assert(DL && "Need target data to handle type mismatch case"); 1329 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1330 *DL); 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 *DL = gvn.getDataLayout(); 1350 assert(DL && "Need target data to handle type mismatch case"); 1351 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1352 LoadTy, BB->getTerminator(), *DL); 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 1364 static 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 1370 void 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 (DL && Address) { 1406 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1407 DepSI, *DL); 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 && DL) { 1425 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), Address, 1426 DepLI, *DL); 1427 1428 if (Offset != -1) { 1429 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1430 Offset)); 1431 continue; 1432 } 1433 } 1434 } 1435 1436 // If the clobbering value is a memset/memcpy/memmove, see if we can 1437 // forward a value on from it. 1438 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1439 if (DL && Address) { 1440 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1441 DepMI, *DL); 1442 if (Offset != -1) { 1443 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1444 Offset)); 1445 continue; 1446 } 1447 } 1448 } 1449 1450 UnavailableBlocks.push_back(DepBB); 1451 continue; 1452 } 1453 1454 // DepInfo.isDef() here 1455 1456 Instruction *DepInst = DepInfo.getInst(); 1457 1458 // Loading the allocation -> undef. 1459 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || 1460 // Loading immediately after lifetime begin -> undef. 1461 isLifetimeStart(DepInst)) { 1462 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1463 UndefValue::get(LI->getType()))); 1464 continue; 1465 } 1466 1467 // Loading from calloc (which zero initializes memory) -> zero 1468 if (isCallocLikeFn(DepInst, TLI)) { 1469 ValuesPerBlock.push_back(AvailableValueInBlock::get( 1470 DepBB, Constant::getNullValue(LI->getType()))); 1471 continue; 1472 } 1473 1474 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1475 // Reject loads and stores that are to the same address but are of 1476 // different types if we have to. 1477 if (S->getValueOperand()->getType() != LI->getType()) { 1478 // If the stored value is larger or equal to the loaded value, we can 1479 // reuse it. 1480 if (!DL || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1481 LI->getType(), *DL)) { 1482 UnavailableBlocks.push_back(DepBB); 1483 continue; 1484 } 1485 } 1486 1487 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1488 S->getValueOperand())); 1489 continue; 1490 } 1491 1492 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1493 // If the types mismatch and we can't handle it, reject reuse of the load. 1494 if (LD->getType() != LI->getType()) { 1495 // If the stored value is larger or equal to the loaded value, we can 1496 // reuse it. 1497 if (!DL || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)) { 1498 UnavailableBlocks.push_back(DepBB); 1499 continue; 1500 } 1501 } 1502 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1503 continue; 1504 } 1505 1506 UnavailableBlocks.push_back(DepBB); 1507 } 1508 } 1509 1510 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 1511 UnavailBlkVect &UnavailableBlocks) { 1512 // Okay, we have *some* definitions of the value. This means that the value 1513 // is available in some of our (transitive) predecessors. Lets think about 1514 // doing PRE of this load. This will involve inserting a new load into the 1515 // predecessor when it's not available. We could do this in general, but 1516 // prefer to not increase code size. As such, we only do this when we know 1517 // that we only have to insert *one* load (which means we're basically moving 1518 // the load, not inserting a new one). 1519 1520 SmallPtrSet<BasicBlock *, 4> Blockers; 1521 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1522 Blockers.insert(UnavailableBlocks[i]); 1523 1524 // Let's find the first basic block with more than one predecessor. Walk 1525 // backwards through predecessors if needed. 1526 BasicBlock *LoadBB = LI->getParent(); 1527 BasicBlock *TmpBB = LoadBB; 1528 1529 while (TmpBB->getSinglePredecessor()) { 1530 TmpBB = TmpBB->getSinglePredecessor(); 1531 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1532 return false; 1533 if (Blockers.count(TmpBB)) 1534 return false; 1535 1536 // If any of these blocks has more than one successor (i.e. if the edge we 1537 // just traversed was critical), then there are other paths through this 1538 // block along which the load may not be anticipated. Hoisting the load 1539 // above this block would be adding the load to execution paths along 1540 // which it was not previously executed. 1541 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1542 return false; 1543 } 1544 1545 assert(TmpBB); 1546 LoadBB = TmpBB; 1547 1548 // Check to see how many predecessors have the loaded value fully 1549 // available. 1550 MapVector<BasicBlock *, Value *> PredLoads; 1551 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1552 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1553 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1554 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1555 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1556 1557 SmallVector<BasicBlock *, 4> CriticalEdgePred; 1558 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1559 PI != E; ++PI) { 1560 BasicBlock *Pred = *PI; 1561 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) { 1562 continue; 1563 } 1564 1565 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1566 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1567 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1568 << Pred->getName() << "': " << *LI << '\n'); 1569 return false; 1570 } 1571 1572 if (LoadBB->isLandingPad()) { 1573 DEBUG(dbgs() 1574 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '" 1575 << Pred->getName() << "': " << *LI << '\n'); 1576 return false; 1577 } 1578 1579 CriticalEdgePred.push_back(Pred); 1580 } else { 1581 // Only add the predecessors that will not be split for now. 1582 PredLoads[Pred] = nullptr; 1583 } 1584 } 1585 1586 // Decide whether PRE is profitable for this load. 1587 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size(); 1588 assert(NumUnavailablePreds != 0 && 1589 "Fully available value should already be eliminated!"); 1590 1591 // If this load is unavailable in multiple predecessors, reject it. 1592 // FIXME: If we could restructure the CFG, we could make a common pred with 1593 // all the preds that don't have an available LI and insert a new load into 1594 // that one block. 1595 if (NumUnavailablePreds != 1) 1596 return false; 1597 1598 // Split critical edges, and update the unavailable predecessors accordingly. 1599 for (BasicBlock *OrigPred : CriticalEdgePred) { 1600 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); 1601 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!"); 1602 PredLoads[NewPred] = nullptr; 1603 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" 1604 << LoadBB->getName() << '\n'); 1605 } 1606 1607 // Check if the load can safely be moved to all the unavailable predecessors. 1608 bool CanDoPRE = true; 1609 SmallVector<Instruction*, 8> NewInsts; 1610 for (auto &PredLoad : PredLoads) { 1611 BasicBlock *UnavailablePred = PredLoad.first; 1612 1613 // Do PHI translation to get its value in the predecessor if necessary. The 1614 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1615 1616 // If all preds have a single successor, then we know it is safe to insert 1617 // the load on the pred (?!?), so we can insert code to materialize the 1618 // pointer if it is not available. 1619 PHITransAddr Address(LI->getPointerOperand(), DL); 1620 Value *LoadPtr = nullptr; 1621 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1622 *DT, NewInsts); 1623 1624 // If we couldn't find or insert a computation of this phi translated value, 1625 // we fail PRE. 1626 if (!LoadPtr) { 1627 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1628 << *LI->getPointerOperand() << "\n"); 1629 CanDoPRE = false; 1630 break; 1631 } 1632 1633 PredLoad.second = LoadPtr; 1634 } 1635 1636 if (!CanDoPRE) { 1637 while (!NewInsts.empty()) { 1638 Instruction *I = NewInsts.pop_back_val(); 1639 if (MD) MD->removeInstruction(I); 1640 I->eraseFromParent(); 1641 } 1642 // HINT: Don't revert the edge-splitting as following transformation may 1643 // also need to split these critical edges. 1644 return !CriticalEdgePred.empty(); 1645 } 1646 1647 // Okay, we can eliminate this load by inserting a reload in the predecessor 1648 // and using PHI construction to get the value in the other predecessors, do 1649 // it. 1650 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1651 DEBUG(if (!NewInsts.empty()) 1652 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1653 << *NewInsts.back() << '\n'); 1654 1655 // Assign value numbers to the new instructions. 1656 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1657 // FIXME: We really _ought_ to insert these value numbers into their 1658 // parent's availability map. However, in doing so, we risk getting into 1659 // ordering issues. If a block hasn't been processed yet, we would be 1660 // marking a value as AVAIL-IN, which isn't what we intend. 1661 VN.lookup_or_add(NewInsts[i]); 1662 } 1663 1664 for (const auto &PredLoad : PredLoads) { 1665 BasicBlock *UnavailablePred = PredLoad.first; 1666 Value *LoadPtr = PredLoad.second; 1667 1668 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1669 LI->getAlignment(), 1670 UnavailablePred->getTerminator()); 1671 1672 // Transfer the old load's TBAA tag to the new load. 1673 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1674 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1675 1676 // Transfer DebugLoc. 1677 NewLoad->setDebugLoc(LI->getDebugLoc()); 1678 1679 // Add the newly created load. 1680 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1681 NewLoad)); 1682 MD->invalidateCachedPointerInfo(LoadPtr); 1683 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1684 } 1685 1686 // Perform PHI construction. 1687 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1688 LI->replaceAllUsesWith(V); 1689 if (isa<PHINode>(V)) 1690 V->takeName(LI); 1691 if (V->getType()->getScalarType()->isPointerTy()) 1692 MD->invalidateCachedPointerInfo(V); 1693 markInstructionForDeletion(LI); 1694 ++NumPRELoad; 1695 return true; 1696 } 1697 1698 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1699 /// non-local by performing PHI construction. 1700 bool GVN::processNonLocalLoad(LoadInst *LI) { 1701 // Step 1: Find the non-local dependencies of the load. 1702 LoadDepVect Deps; 1703 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1704 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1705 1706 // If we had to process more than one hundred blocks to find the 1707 // dependencies, this load isn't worth worrying about. Optimizing 1708 // it will be too expensive. 1709 unsigned NumDeps = Deps.size(); 1710 if (NumDeps > 100) 1711 return false; 1712 1713 // If we had a phi translation failure, we'll have a single entry which is a 1714 // clobber in the current block. Reject this early. 1715 if (NumDeps == 1 && 1716 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { 1717 DEBUG( 1718 dbgs() << "GVN: non-local load "; 1719 LI->printAsOperand(dbgs()); 1720 dbgs() << " has unknown dependencies\n"; 1721 ); 1722 return false; 1723 } 1724 1725 // Step 2: Analyze the availability of the load 1726 AvailValInBlkVect ValuesPerBlock; 1727 UnavailBlkVect UnavailableBlocks; 1728 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks); 1729 1730 // If we have no predecessors that produce a known value for this load, exit 1731 // early. 1732 if (ValuesPerBlock.empty()) 1733 return false; 1734 1735 // Step 3: Eliminate fully redundancy. 1736 // 1737 // If all of the instructions we depend on produce a known value for this 1738 // load, then it is fully redundant and we can use PHI insertion to compute 1739 // its value. Insert PHIs and remove the fully redundant value now. 1740 if (UnavailableBlocks.empty()) { 1741 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1742 1743 // Perform PHI construction. 1744 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this); 1745 LI->replaceAllUsesWith(V); 1746 1747 if (isa<PHINode>(V)) 1748 V->takeName(LI); 1749 if (V->getType()->getScalarType()->isPointerTy()) 1750 MD->invalidateCachedPointerInfo(V); 1751 markInstructionForDeletion(LI); 1752 ++NumGVNLoad; 1753 return true; 1754 } 1755 1756 // Step 4: Eliminate partial redundancy. 1757 if (!EnablePRE || !EnableLoadPRE) 1758 return false; 1759 1760 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks); 1761 } 1762 1763 1764 static void patchReplacementInstruction(Instruction *I, Value *Repl) { 1765 // Patch the replacement so that it is not more restrictive than the value 1766 // being replaced. 1767 BinaryOperator *Op = dyn_cast<BinaryOperator>(I); 1768 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl); 1769 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) && 1770 isa<OverflowingBinaryOperator>(ReplOp)) { 1771 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap()) 1772 ReplOp->setHasNoSignedWrap(false); 1773 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap()) 1774 ReplOp->setHasNoUnsignedWrap(false); 1775 } 1776 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) { 1777 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1778 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata); 1779 for (int i = 0, n = Metadata.size(); i < n; ++i) { 1780 unsigned Kind = Metadata[i].first; 1781 MDNode *IMD = I->getMetadata(Kind); 1782 MDNode *ReplMD = Metadata[i].second; 1783 switch(Kind) { 1784 default: 1785 ReplInst->setMetadata(Kind, nullptr); // Remove unknown metadata 1786 break; 1787 case LLVMContext::MD_dbg: 1788 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1789 case LLVMContext::MD_tbaa: 1790 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD)); 1791 break; 1792 case LLVMContext::MD_range: 1793 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD)); 1794 break; 1795 case LLVMContext::MD_prof: 1796 llvm_unreachable("MD_prof in a non-terminator instruction"); 1797 break; 1798 case LLVMContext::MD_fpmath: 1799 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD)); 1800 break; 1801 case LLVMContext::MD_invariant_load: 1802 // Only set the !invariant.load if it is present in both instructions. 1803 ReplInst->setMetadata(Kind, IMD); 1804 break; 1805 } 1806 } 1807 } 1808 } 1809 1810 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { 1811 patchReplacementInstruction(I, Repl); 1812 I->replaceAllUsesWith(Repl); 1813 } 1814 1815 /// processLoad - Attempt to eliminate a load, first by eliminating it 1816 /// locally, and then attempting non-local elimination if that fails. 1817 bool GVN::processLoad(LoadInst *L) { 1818 if (!MD) 1819 return false; 1820 1821 if (!L->isSimple()) 1822 return false; 1823 1824 if (L->use_empty()) { 1825 markInstructionForDeletion(L); 1826 return true; 1827 } 1828 1829 // ... to a pointer that has been loaded from before... 1830 MemDepResult Dep = MD->getDependency(L); 1831 1832 // If we have a clobber and target data is around, see if this is a clobber 1833 // that we can fix up through code synthesis. 1834 if (Dep.isClobber() && DL) { 1835 // Check to see if we have something like this: 1836 // store i32 123, i32* %P 1837 // %A = bitcast i32* %P to i8* 1838 // %B = gep i8* %A, i32 1 1839 // %C = load i8* %B 1840 // 1841 // We could do that by recognizing if the clobber instructions are obviously 1842 // a common base + constant offset, and if the previous store (or memset) 1843 // completely covers this load. This sort of thing can happen in bitfield 1844 // access code. 1845 Value *AvailVal = nullptr; 1846 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1847 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1848 L->getPointerOperand(), 1849 DepSI, *DL); 1850 if (Offset != -1) 1851 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1852 L->getType(), L, *DL); 1853 } 1854 1855 // Check to see if we have something like this: 1856 // load i32* P 1857 // load i8* (P+1) 1858 // if we have this, replace the later with an extraction from the former. 1859 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1860 // If this is a clobber and L is the first instruction in its block, then 1861 // we have the first instruction in the entry block. 1862 if (DepLI == L) 1863 return false; 1864 1865 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1866 L->getPointerOperand(), 1867 DepLI, *DL); 1868 if (Offset != -1) 1869 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this); 1870 } 1871 1872 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1873 // a value on from it. 1874 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1875 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1876 L->getPointerOperand(), 1877 DepMI, *DL); 1878 if (Offset != -1) 1879 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL); 1880 } 1881 1882 if (AvailVal) { 1883 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1884 << *AvailVal << '\n' << *L << "\n\n\n"); 1885 1886 // Replace the load! 1887 L->replaceAllUsesWith(AvailVal); 1888 if (AvailVal->getType()->getScalarType()->isPointerTy()) 1889 MD->invalidateCachedPointerInfo(AvailVal); 1890 markInstructionForDeletion(L); 1891 ++NumGVNLoad; 1892 return true; 1893 } 1894 } 1895 1896 // If the value isn't available, don't do anything! 1897 if (Dep.isClobber()) { 1898 DEBUG( 1899 // fast print dep, using operator<< on instruction is too slow. 1900 dbgs() << "GVN: load "; 1901 L->printAsOperand(dbgs()); 1902 Instruction *I = Dep.getInst(); 1903 dbgs() << " is clobbered by " << *I << '\n'; 1904 ); 1905 return false; 1906 } 1907 1908 // If it is defined in another block, try harder. 1909 if (Dep.isNonLocal()) 1910 return processNonLocalLoad(L); 1911 1912 if (!Dep.isDef()) { 1913 DEBUG( 1914 // fast print dep, using operator<< on instruction is too slow. 1915 dbgs() << "GVN: load "; 1916 L->printAsOperand(dbgs()); 1917 dbgs() << " has unknown dependence\n"; 1918 ); 1919 return false; 1920 } 1921 1922 Instruction *DepInst = Dep.getInst(); 1923 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1924 Value *StoredVal = DepSI->getValueOperand(); 1925 1926 // The store and load are to a must-aliased pointer, but they may not 1927 // actually have the same type. See if we know how to reuse the stored 1928 // value (depending on its type). 1929 if (StoredVal->getType() != L->getType()) { 1930 if (DL) { 1931 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1932 L, *DL); 1933 if (!StoredVal) 1934 return false; 1935 1936 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1937 << '\n' << *L << "\n\n\n"); 1938 } 1939 else 1940 return false; 1941 } 1942 1943 // Remove it! 1944 L->replaceAllUsesWith(StoredVal); 1945 if (StoredVal->getType()->getScalarType()->isPointerTy()) 1946 MD->invalidateCachedPointerInfo(StoredVal); 1947 markInstructionForDeletion(L); 1948 ++NumGVNLoad; 1949 return true; 1950 } 1951 1952 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1953 Value *AvailableVal = DepLI; 1954 1955 // The loads are of a must-aliased pointer, but they may not actually have 1956 // the same type. See if we know how to reuse the previously loaded value 1957 // (depending on its type). 1958 if (DepLI->getType() != L->getType()) { 1959 if (DL) { 1960 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1961 L, *DL); 1962 if (!AvailableVal) 1963 return false; 1964 1965 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1966 << "\n" << *L << "\n\n\n"); 1967 } 1968 else 1969 return false; 1970 } 1971 1972 // Remove it! 1973 patchAndReplaceAllUsesWith(L, AvailableVal); 1974 if (DepLI->getType()->getScalarType()->isPointerTy()) 1975 MD->invalidateCachedPointerInfo(DepLI); 1976 markInstructionForDeletion(L); 1977 ++NumGVNLoad; 1978 return true; 1979 } 1980 1981 // If this load really doesn't depend on anything, then we must be loading an 1982 // undef value. This can happen when loading for a fresh allocation with no 1983 // intervening stores, for example. 1984 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { 1985 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1986 markInstructionForDeletion(L); 1987 ++NumGVNLoad; 1988 return true; 1989 } 1990 1991 // If this load occurs either right after a lifetime begin, 1992 // then the loaded value is undefined. 1993 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) { 1994 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1995 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1996 markInstructionForDeletion(L); 1997 ++NumGVNLoad; 1998 return true; 1999 } 2000 } 2001 2002 // If this load follows a calloc (which zero initializes memory), 2003 // then the loaded value is zero 2004 if (isCallocLikeFn(DepInst, TLI)) { 2005 L->replaceAllUsesWith(Constant::getNullValue(L->getType())); 2006 markInstructionForDeletion(L); 2007 ++NumGVNLoad; 2008 return true; 2009 } 2010 2011 return false; 2012 } 2013 2014 // findLeader - In order to find a leader for a given value number at a 2015 // specific basic block, we first obtain the list of all Values for that number, 2016 // and then scan the list to find one whose block dominates the block in 2017 // question. This is fast because dominator tree queries consist of only 2018 // a few comparisons of DFS numbers. 2019 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { 2020 LeaderTableEntry Vals = LeaderTable[num]; 2021 if (!Vals.Val) return nullptr; 2022 2023 Value *Val = nullptr; 2024 if (DT->dominates(Vals.BB, BB)) { 2025 Val = Vals.Val; 2026 if (isa<Constant>(Val)) return Val; 2027 } 2028 2029 LeaderTableEntry* Next = Vals.Next; 2030 while (Next) { 2031 if (DT->dominates(Next->BB, BB)) { 2032 if (isa<Constant>(Next->Val)) return Next->Val; 2033 if (!Val) Val = Next->Val; 2034 } 2035 2036 Next = Next->Next; 2037 } 2038 2039 return Val; 2040 } 2041 2042 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the 2043 /// use is dominated by the given basic block. Returns the number of uses that 2044 /// were replaced. 2045 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To, 2046 const BasicBlockEdge &Root) { 2047 unsigned Count = 0; 2048 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2049 UI != UE; ) { 2050 Use &U = *UI++; 2051 2052 if (DT->dominates(Root, U)) { 2053 U.set(To); 2054 ++Count; 2055 } 2056 } 2057 return Count; 2058 } 2059 2060 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return 2061 /// true if every path from the entry block to 'Dst' passes via this edge. In 2062 /// particular 'Dst' must not be reachable via another edge from 'Src'. 2063 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, 2064 DominatorTree *DT) { 2065 // While in theory it is interesting to consider the case in which Dst has 2066 // more than one predecessor, because Dst might be part of a loop which is 2067 // only reachable from Src, in practice it is pointless since at the time 2068 // GVN runs all such loops have preheaders, which means that Dst will have 2069 // been changed to have only one predecessor, namely Src. 2070 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); 2071 const BasicBlock *Src = E.getStart(); 2072 assert((!Pred || Pred == Src) && "No edge between these basic blocks!"); 2073 (void)Src; 2074 return Pred != nullptr; 2075 } 2076 2077 /// propagateEquality - The given values are known to be equal in every block 2078 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with 2079 /// 'RHS' everywhere in the scope. Returns whether a change was made. 2080 bool GVN::propagateEquality(Value *LHS, Value *RHS, 2081 const BasicBlockEdge &Root) { 2082 SmallVector<std::pair<Value*, Value*>, 4> Worklist; 2083 Worklist.push_back(std::make_pair(LHS, RHS)); 2084 bool Changed = false; 2085 // For speed, compute a conservative fast approximation to 2086 // DT->dominates(Root, Root.getEnd()); 2087 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); 2088 2089 while (!Worklist.empty()) { 2090 std::pair<Value*, Value*> Item = Worklist.pop_back_val(); 2091 LHS = Item.first; RHS = Item.second; 2092 2093 if (LHS == RHS) continue; 2094 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!"); 2095 2096 // Don't try to propagate equalities between constants. 2097 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue; 2098 2099 // Prefer a constant on the right-hand side, or an Argument if no constants. 2100 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) 2101 std::swap(LHS, RHS); 2102 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!"); 2103 2104 // If there is no obvious reason to prefer the left-hand side over the right- 2105 // hand side, ensure the longest lived term is on the right-hand side, so the 2106 // shortest lived term will be replaced by the longest lived. This tends to 2107 // expose more simplifications. 2108 uint32_t LVN = VN.lookup_or_add(LHS); 2109 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || 2110 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { 2111 // Move the 'oldest' value to the right-hand side, using the value number as 2112 // a proxy for age. 2113 uint32_t RVN = VN.lookup_or_add(RHS); 2114 if (LVN < RVN) { 2115 std::swap(LHS, RHS); 2116 LVN = RVN; 2117 } 2118 } 2119 2120 // If value numbering later sees that an instruction in the scope is equal 2121 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve 2122 // the invariant that instructions only occur in the leader table for their 2123 // own value number (this is used by removeFromLeaderTable), do not do this 2124 // if RHS is an instruction (if an instruction in the scope is morphed into 2125 // LHS then it will be turned into RHS by the next GVN iteration anyway, so 2126 // using the leader table is about compiling faster, not optimizing better). 2127 // The leader table only tracks basic blocks, not edges. Only add to if we 2128 // have the simple case where the edge dominates the end. 2129 if (RootDominatesEnd && !isa<Instruction>(RHS)) 2130 addToLeaderTable(LVN, RHS, Root.getEnd()); 2131 2132 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As 2133 // LHS always has at least one use that is not dominated by Root, this will 2134 // never do anything if LHS has only one use. 2135 if (!LHS->hasOneUse()) { 2136 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root); 2137 Changed |= NumReplacements > 0; 2138 NumGVNEqProp += NumReplacements; 2139 } 2140 2141 // Now try to deduce additional equalities from this one. For example, if the 2142 // known equality was "(A != B)" == "false" then it follows that A and B are 2143 // equal in the scope. Only boolean equalities with an explicit true or false 2144 // RHS are currently supported. 2145 if (!RHS->getType()->isIntegerTy(1)) 2146 // Not a boolean equality - bail out. 2147 continue; 2148 ConstantInt *CI = dyn_cast<ConstantInt>(RHS); 2149 if (!CI) 2150 // RHS neither 'true' nor 'false' - bail out. 2151 continue; 2152 // Whether RHS equals 'true'. Otherwise it equals 'false'. 2153 bool isKnownTrue = CI->isAllOnesValue(); 2154 bool isKnownFalse = !isKnownTrue; 2155 2156 // If "A && B" is known true then both A and B are known true. If "A || B" 2157 // is known false then both A and B are known false. 2158 Value *A, *B; 2159 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) || 2160 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) { 2161 Worklist.push_back(std::make_pair(A, RHS)); 2162 Worklist.push_back(std::make_pair(B, RHS)); 2163 continue; 2164 } 2165 2166 // If we are propagating an equality like "(A == B)" == "true" then also 2167 // propagate the equality A == B. When propagating a comparison such as 2168 // "(A >= B)" == "true", replace all instances of "A < B" with "false". 2169 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) { 2170 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); 2171 2172 // If "A == B" is known true, or "A != B" is known false, then replace 2173 // A with B everywhere in the scope. 2174 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) || 2175 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) 2176 Worklist.push_back(std::make_pair(Op0, Op1)); 2177 2178 // If "A >= B" is known true, replace "A < B" with false everywhere. 2179 CmpInst::Predicate NotPred = Cmp->getInversePredicate(); 2180 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); 2181 // Since we don't have the instruction "A < B" immediately to hand, work out 2182 // the value number that it would have and use that to find an appropriate 2183 // instruction (if any). 2184 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2185 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1); 2186 // If the number we were assigned was brand new then there is no point in 2187 // looking for an instruction realizing it: there cannot be one! 2188 if (Num < NextNum) { 2189 Value *NotCmp = findLeader(Root.getEnd(), Num); 2190 if (NotCmp && isa<Instruction>(NotCmp)) { 2191 unsigned NumReplacements = 2192 replaceAllDominatedUsesWith(NotCmp, NotVal, Root); 2193 Changed |= NumReplacements > 0; 2194 NumGVNEqProp += NumReplacements; 2195 } 2196 } 2197 // Ensure that any instruction in scope that gets the "A < B" value number 2198 // is replaced with false. 2199 // The leader table only tracks basic blocks, not edges. Only add to if we 2200 // have the simple case where the edge dominates the end. 2201 if (RootDominatesEnd) 2202 addToLeaderTable(Num, NotVal, Root.getEnd()); 2203 2204 continue; 2205 } 2206 } 2207 2208 return Changed; 2209 } 2210 2211 /// processInstruction - When calculating availability, handle an instruction 2212 /// by inserting it into the appropriate sets 2213 bool GVN::processInstruction(Instruction *I) { 2214 // Ignore dbg info intrinsics. 2215 if (isa<DbgInfoIntrinsic>(I)) 2216 return false; 2217 2218 // If the instruction can be easily simplified then do so now in preference 2219 // to value numbering it. Value numbering often exposes redundancies, for 2220 // example if it determines that %y is equal to %x then the instruction 2221 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 2222 if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) { 2223 I->replaceAllUsesWith(V); 2224 if (MD && V->getType()->getScalarType()->isPointerTy()) 2225 MD->invalidateCachedPointerInfo(V); 2226 markInstructionForDeletion(I); 2227 ++NumGVNSimpl; 2228 return true; 2229 } 2230 2231 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 2232 if (processLoad(LI)) 2233 return true; 2234 2235 unsigned Num = VN.lookup_or_add(LI); 2236 addToLeaderTable(Num, LI, LI->getParent()); 2237 return false; 2238 } 2239 2240 // For conditional branches, we can perform simple conditional propagation on 2241 // the condition value itself. 2242 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 2243 if (!BI->isConditional()) 2244 return false; 2245 2246 if (isa<Constant>(BI->getCondition())) 2247 return processFoldableCondBr(BI); 2248 2249 Value *BranchCond = BI->getCondition(); 2250 BasicBlock *TrueSucc = BI->getSuccessor(0); 2251 BasicBlock *FalseSucc = BI->getSuccessor(1); 2252 // Avoid multiple edges early. 2253 if (TrueSucc == FalseSucc) 2254 return false; 2255 2256 BasicBlock *Parent = BI->getParent(); 2257 bool Changed = false; 2258 2259 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); 2260 BasicBlockEdge TrueE(Parent, TrueSucc); 2261 Changed |= propagateEquality(BranchCond, TrueVal, TrueE); 2262 2263 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); 2264 BasicBlockEdge FalseE(Parent, FalseSucc); 2265 Changed |= propagateEquality(BranchCond, FalseVal, FalseE); 2266 2267 return Changed; 2268 } 2269 2270 // For switches, propagate the case values into the case destinations. 2271 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 2272 Value *SwitchCond = SI->getCondition(); 2273 BasicBlock *Parent = SI->getParent(); 2274 bool Changed = false; 2275 2276 // Remember how many outgoing edges there are to every successor. 2277 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 2278 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) 2279 ++SwitchEdges[SI->getSuccessor(i)]; 2280 2281 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2282 i != e; ++i) { 2283 BasicBlock *Dst = i.getCaseSuccessor(); 2284 // If there is only a single edge, propagate the case value into it. 2285 if (SwitchEdges.lookup(Dst) == 1) { 2286 BasicBlockEdge E(Parent, Dst); 2287 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); 2288 } 2289 } 2290 return Changed; 2291 } 2292 2293 // Instructions with void type don't return a value, so there's 2294 // no point in trying to find redundancies in them. 2295 if (I->getType()->isVoidTy()) return false; 2296 2297 uint32_t NextNum = VN.getNextUnusedValueNumber(); 2298 unsigned Num = VN.lookup_or_add(I); 2299 2300 // Allocations are always uniquely numbered, so we can save time and memory 2301 // by fast failing them. 2302 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 2303 addToLeaderTable(Num, I, I->getParent()); 2304 return false; 2305 } 2306 2307 // If the number we were assigned was a brand new VN, then we don't 2308 // need to do a lookup to see if the number already exists 2309 // somewhere in the domtree: it can't! 2310 if (Num >= NextNum) { 2311 addToLeaderTable(Num, I, I->getParent()); 2312 return false; 2313 } 2314 2315 // Perform fast-path value-number based elimination of values inherited from 2316 // dominators. 2317 Value *repl = findLeader(I->getParent(), Num); 2318 if (!repl) { 2319 // Failure, just remember this instance for future use. 2320 addToLeaderTable(Num, I, I->getParent()); 2321 return false; 2322 } 2323 2324 // Remove it! 2325 patchAndReplaceAllUsesWith(I, repl); 2326 if (MD && repl->getType()->getScalarType()->isPointerTy()) 2327 MD->invalidateCachedPointerInfo(repl); 2328 markInstructionForDeletion(I); 2329 return true; 2330 } 2331 2332 /// runOnFunction - This is the main transformation entry point for a function. 2333 bool GVN::runOnFunction(Function& F) { 2334 if (skipOptnoneFunction(F)) 2335 return false; 2336 2337 if (!NoLoads) 2338 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2339 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2340 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 2341 DL = DLP ? &DLP->getDataLayout() : nullptr; 2342 TLI = &getAnalysis<TargetLibraryInfo>(); 2343 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2344 VN.setMemDep(MD); 2345 VN.setDomTree(DT); 2346 2347 bool Changed = false; 2348 bool ShouldContinue = true; 2349 2350 // Merge unconditional branches, allowing PRE to catch more 2351 // optimization opportunities. 2352 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2353 BasicBlock *BB = FI++; 2354 2355 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2356 if (removedBlock) ++NumGVNBlocks; 2357 2358 Changed |= removedBlock; 2359 } 2360 2361 unsigned Iteration = 0; 2362 while (ShouldContinue) { 2363 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2364 ShouldContinue = iterateOnFunction(F); 2365 Changed |= ShouldContinue; 2366 ++Iteration; 2367 } 2368 2369 if (EnablePRE) { 2370 // Fabricate val-num for dead-code in order to suppress assertion in 2371 // performPRE(). 2372 assignValNumForDeadCode(); 2373 bool PREChanged = true; 2374 while (PREChanged) { 2375 PREChanged = performPRE(F); 2376 Changed |= PREChanged; 2377 } 2378 } 2379 2380 // FIXME: Should perform GVN again after PRE does something. PRE can move 2381 // computations into blocks where they become fully redundant. Note that 2382 // we can't do this until PRE's critical edge splitting updates memdep. 2383 // Actually, when this happens, we should just fully integrate PRE into GVN. 2384 2385 cleanupGlobalSets(); 2386 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each 2387 // iteration. 2388 DeadBlocks.clear(); 2389 2390 return Changed; 2391 } 2392 2393 2394 bool GVN::processBlock(BasicBlock *BB) { 2395 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 2396 // (and incrementing BI before processing an instruction). 2397 assert(InstrsToErase.empty() && 2398 "We expect InstrsToErase to be empty across iterations"); 2399 if (DeadBlocks.count(BB)) 2400 return false; 2401 2402 bool ChangedFunction = false; 2403 2404 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2405 BI != BE;) { 2406 ChangedFunction |= processInstruction(BI); 2407 if (InstrsToErase.empty()) { 2408 ++BI; 2409 continue; 2410 } 2411 2412 // If we need some instructions deleted, do it now. 2413 NumGVNInstr += InstrsToErase.size(); 2414 2415 // Avoid iterator invalidation. 2416 bool AtStart = BI == BB->begin(); 2417 if (!AtStart) 2418 --BI; 2419 2420 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(), 2421 E = InstrsToErase.end(); I != E; ++I) { 2422 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2423 if (MD) MD->removeInstruction(*I); 2424 DEBUG(verifyRemoved(*I)); 2425 (*I)->eraseFromParent(); 2426 } 2427 InstrsToErase.clear(); 2428 2429 if (AtStart) 2430 BI = BB->begin(); 2431 else 2432 ++BI; 2433 } 2434 2435 return ChangedFunction; 2436 } 2437 2438 /// performPRE - Perform a purely local form of PRE that looks for diamond 2439 /// control flow patterns and attempts to perform simple PRE at the join point. 2440 bool GVN::performPRE(Function &F) { 2441 bool Changed = false; 2442 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap; 2443 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) { 2444 // Nothing to PRE in the entry block. 2445 if (CurrentBlock == &F.getEntryBlock()) continue; 2446 2447 // Don't perform PRE on a landing pad. 2448 if (CurrentBlock->isLandingPad()) continue; 2449 2450 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2451 BE = CurrentBlock->end(); BI != BE; ) { 2452 Instruction *CurInst = BI++; 2453 2454 if (isa<AllocaInst>(CurInst) || 2455 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2456 CurInst->getType()->isVoidTy() || 2457 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2458 isa<DbgInfoIntrinsic>(CurInst)) 2459 continue; 2460 2461 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from 2462 // sinking the compare again, and it would force the code generator to 2463 // move the i1 from processor flags or predicate registers into a general 2464 // purpose register. 2465 if (isa<CmpInst>(CurInst)) 2466 continue; 2467 2468 // We don't currently value number ANY inline asm calls. 2469 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2470 if (CallI->isInlineAsm()) 2471 continue; 2472 2473 uint32_t ValNo = VN.lookup(CurInst); 2474 2475 // Look for the predecessors for PRE opportunities. We're 2476 // only trying to solve the basic diamond case, where 2477 // a value is computed in the successor and one predecessor, 2478 // but not the other. We also explicitly disallow cases 2479 // where the successor is its own predecessor, because they're 2480 // more complicated to get right. 2481 unsigned NumWith = 0; 2482 unsigned NumWithout = 0; 2483 BasicBlock *PREPred = nullptr; 2484 predMap.clear(); 2485 2486 for (pred_iterator PI = pred_begin(CurrentBlock), 2487 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2488 BasicBlock *P = *PI; 2489 // We're not interested in PRE where the block is its 2490 // own predecessor, or in blocks with predecessors 2491 // that are not reachable. 2492 if (P == CurrentBlock) { 2493 NumWithout = 2; 2494 break; 2495 } else if (!DT->isReachableFromEntry(P)) { 2496 NumWithout = 2; 2497 break; 2498 } 2499 2500 Value* predV = findLeader(P, ValNo); 2501 if (!predV) { 2502 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P)); 2503 PREPred = P; 2504 ++NumWithout; 2505 } else if (predV == CurInst) { 2506 /* CurInst dominates this predecessor. */ 2507 NumWithout = 2; 2508 break; 2509 } else { 2510 predMap.push_back(std::make_pair(predV, P)); 2511 ++NumWith; 2512 } 2513 } 2514 2515 // Don't do PRE when it might increase code size, i.e. when 2516 // we would need to insert instructions in more than one pred. 2517 if (NumWithout != 1 || NumWith == 0) 2518 continue; 2519 2520 // Don't do PRE across indirect branch. 2521 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2522 continue; 2523 2524 // We can't do PRE safely on a critical edge, so instead we schedule 2525 // the edge to be split and perform the PRE the next time we iterate 2526 // on the function. 2527 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2528 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2529 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2530 continue; 2531 } 2532 2533 // Instantiate the expression in the predecessor that lacked it. 2534 // Because we are going top-down through the block, all value numbers 2535 // will be available in the predecessor by the time we need them. Any 2536 // that weren't originally present will have been instantiated earlier 2537 // in this loop. 2538 Instruction *PREInstr = CurInst->clone(); 2539 bool success = true; 2540 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2541 Value *Op = PREInstr->getOperand(i); 2542 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2543 continue; 2544 2545 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2546 PREInstr->setOperand(i, V); 2547 } else { 2548 success = false; 2549 break; 2550 } 2551 } 2552 2553 // Fail out if we encounter an operand that is not available in 2554 // the PRE predecessor. This is typically because of loads which 2555 // are not value numbered precisely. 2556 if (!success) { 2557 DEBUG(verifyRemoved(PREInstr)); 2558 delete PREInstr; 2559 continue; 2560 } 2561 2562 PREInstr->insertBefore(PREPred->getTerminator()); 2563 PREInstr->setName(CurInst->getName() + ".pre"); 2564 PREInstr->setDebugLoc(CurInst->getDebugLoc()); 2565 VN.add(PREInstr, ValNo); 2566 ++NumGVNPRE; 2567 2568 // Update the availability map to include the new instruction. 2569 addToLeaderTable(ValNo, PREInstr, PREPred); 2570 2571 // Create a PHI to make the value available in this block. 2572 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(), 2573 CurInst->getName() + ".pre-phi", 2574 CurrentBlock->begin()); 2575 for (unsigned i = 0, e = predMap.size(); i != e; ++i) { 2576 if (Value *V = predMap[i].first) 2577 Phi->addIncoming(V, predMap[i].second); 2578 else 2579 Phi->addIncoming(PREInstr, PREPred); 2580 } 2581 2582 VN.add(Phi, ValNo); 2583 addToLeaderTable(ValNo, Phi, CurrentBlock); 2584 Phi->setDebugLoc(CurInst->getDebugLoc()); 2585 CurInst->replaceAllUsesWith(Phi); 2586 if (Phi->getType()->getScalarType()->isPointerTy()) { 2587 // Because we have added a PHI-use of the pointer value, it has now 2588 // "escaped" from alias analysis' perspective. We need to inform 2589 // AA of this. 2590 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; 2591 ++ii) { 2592 unsigned jj = PHINode::getOperandNumForIncomingValue(ii); 2593 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj)); 2594 } 2595 2596 if (MD) 2597 MD->invalidateCachedPointerInfo(Phi); 2598 } 2599 VN.erase(CurInst); 2600 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2601 2602 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2603 if (MD) MD->removeInstruction(CurInst); 2604 DEBUG(verifyRemoved(CurInst)); 2605 CurInst->eraseFromParent(); 2606 Changed = true; 2607 } 2608 } 2609 2610 if (splitCriticalEdges()) 2611 Changed = true; 2612 2613 return Changed; 2614 } 2615 2616 /// Split the critical edge connecting the given two blocks, and return 2617 /// the block inserted to the critical edge. 2618 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { 2619 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this); 2620 if (MD) 2621 MD->invalidateCachedPredecessors(); 2622 return BB; 2623 } 2624 2625 /// splitCriticalEdges - Split critical edges found during the previous 2626 /// iteration that may enable further optimization. 2627 bool GVN::splitCriticalEdges() { 2628 if (toSplit.empty()) 2629 return false; 2630 do { 2631 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2632 SplitCriticalEdge(Edge.first, Edge.second, this); 2633 } while (!toSplit.empty()); 2634 if (MD) MD->invalidateCachedPredecessors(); 2635 return true; 2636 } 2637 2638 /// iterateOnFunction - Executes one iteration of GVN 2639 bool GVN::iterateOnFunction(Function &F) { 2640 cleanupGlobalSets(); 2641 2642 // Top-down walk of the dominator tree 2643 bool Changed = false; 2644 #if 0 2645 // Needed for value numbering with phi construction to work. 2646 ReversePostOrderTraversal<Function*> RPOT(&F); 2647 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2648 RE = RPOT.end(); RI != RE; ++RI) 2649 Changed |= processBlock(*RI); 2650 #else 2651 // Save the blocks this function have before transformation begins. GVN may 2652 // split critical edge, and hence may invalidate the RPO/DT iterator. 2653 // 2654 std::vector<BasicBlock *> BBVect; 2655 BBVect.reserve(256); 2656 for (DomTreeNode *x : depth_first(DT->getRootNode())) 2657 BBVect.push_back(x->getBlock()); 2658 2659 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end(); 2660 I != E; I++) 2661 Changed |= processBlock(*I); 2662 #endif 2663 2664 return Changed; 2665 } 2666 2667 void GVN::cleanupGlobalSets() { 2668 VN.clear(); 2669 LeaderTable.clear(); 2670 TableAllocator.Reset(); 2671 } 2672 2673 /// verifyRemoved - Verify that the specified instruction does not occur in our 2674 /// internal data structures. 2675 void GVN::verifyRemoved(const Instruction *Inst) const { 2676 VN.verifyRemoved(Inst); 2677 2678 // Walk through the value number scope to make sure the instruction isn't 2679 // ferreted away in it. 2680 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2681 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2682 const LeaderTableEntry *Node = &I->second; 2683 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2684 2685 while (Node->Next) { 2686 Node = Node->Next; 2687 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2688 } 2689 } 2690 } 2691 2692 // BB is declared dead, which implied other blocks become dead as well. This 2693 // function is to add all these blocks to "DeadBlocks". For the dead blocks' 2694 // live successors, update their phi nodes by replacing the operands 2695 // corresponding to dead blocks with UndefVal. 2696 // 2697 void GVN::addDeadBlock(BasicBlock *BB) { 2698 SmallVector<BasicBlock *, 4> NewDead; 2699 SmallSetVector<BasicBlock *, 4> DF; 2700 2701 NewDead.push_back(BB); 2702 while (!NewDead.empty()) { 2703 BasicBlock *D = NewDead.pop_back_val(); 2704 if (DeadBlocks.count(D)) 2705 continue; 2706 2707 // All blocks dominated by D are dead. 2708 SmallVector<BasicBlock *, 8> Dom; 2709 DT->getDescendants(D, Dom); 2710 DeadBlocks.insert(Dom.begin(), Dom.end()); 2711 2712 // Figure out the dominance-frontier(D). 2713 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(), 2714 E = Dom.end(); I != E; I++) { 2715 BasicBlock *B = *I; 2716 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) { 2717 BasicBlock *S = *SI; 2718 if (DeadBlocks.count(S)) 2719 continue; 2720 2721 bool AllPredDead = true; 2722 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++) 2723 if (!DeadBlocks.count(*PI)) { 2724 AllPredDead = false; 2725 break; 2726 } 2727 2728 if (!AllPredDead) { 2729 // S could be proved dead later on. That is why we don't update phi 2730 // operands at this moment. 2731 DF.insert(S); 2732 } else { 2733 // While S is not dominated by D, it is dead by now. This could take 2734 // place if S already have a dead predecessor before D is declared 2735 // dead. 2736 NewDead.push_back(S); 2737 } 2738 } 2739 } 2740 } 2741 2742 // For the dead blocks' live successors, update their phi nodes by replacing 2743 // the operands corresponding to dead blocks with UndefVal. 2744 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end(); 2745 I != E; I++) { 2746 BasicBlock *B = *I; 2747 if (DeadBlocks.count(B)) 2748 continue; 2749 2750 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B)); 2751 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(), 2752 PE = Preds.end(); PI != PE; PI++) { 2753 BasicBlock *P = *PI; 2754 2755 if (!DeadBlocks.count(P)) 2756 continue; 2757 2758 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) { 2759 if (BasicBlock *S = splitCriticalEdges(P, B)) 2760 DeadBlocks.insert(P = S); 2761 } 2762 2763 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) { 2764 PHINode &Phi = cast<PHINode>(*II); 2765 Phi.setIncomingValue(Phi.getBasicBlockIndex(P), 2766 UndefValue::get(Phi.getType())); 2767 } 2768 } 2769 } 2770 } 2771 2772 // If the given branch is recognized as a foldable branch (i.e. conditional 2773 // branch with constant condition), it will perform following analyses and 2774 // transformation. 2775 // 1) If the dead out-coming edge is a critical-edge, split it. Let 2776 // R be the target of the dead out-coming edge. 2777 // 1) Identify the set of dead blocks implied by the branch's dead outcoming 2778 // edge. The result of this step will be {X| X is dominated by R} 2779 // 2) Identify those blocks which haves at least one dead prodecessor. The 2780 // result of this step will be dominance-frontier(R). 2781 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to 2782 // dead blocks with "UndefVal" in an hope these PHIs will optimized away. 2783 // 2784 // Return true iff *NEW* dead code are found. 2785 bool GVN::processFoldableCondBr(BranchInst *BI) { 2786 if (!BI || BI->isUnconditional()) 2787 return false; 2788 2789 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition()); 2790 if (!Cond) 2791 return false; 2792 2793 BasicBlock *DeadRoot = Cond->getZExtValue() ? 2794 BI->getSuccessor(1) : BI->getSuccessor(0); 2795 if (DeadBlocks.count(DeadRoot)) 2796 return false; 2797 2798 if (!DeadRoot->getSinglePredecessor()) 2799 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot); 2800 2801 addDeadBlock(DeadRoot); 2802 return true; 2803 } 2804 2805 // performPRE() will trigger assert if it come across an instruciton without 2806 // associated val-num. As it normally has far more live instructions than dead 2807 // instructions, it makes more sense just to "fabricate" a val-number for the 2808 // dead code than checking if instruction involved is dead or not. 2809 void GVN::assignValNumForDeadCode() { 2810 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(), 2811 E = DeadBlocks.end(); I != E; I++) { 2812 BasicBlock *BB = *I; 2813 for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); 2814 II != EE; II++) { 2815 Instruction *Inst = &*II; 2816 unsigned ValNum = VN.lookup_or_add(Inst); 2817 addToLeaderTable(ValNum, Inst, BB); 2818 } 2819 } 2820 } 2821