GVN.cpp revision 075fb5d68fcb55d26e44c48f07dfdbbfa21ccb2a
1//===- GVN.cpp - Eliminate redundant values and loads ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs global value numbering to eliminate fully redundant 11// instructions. It also performs simple dead load elimination. 12// 13// Note that this pass does the value numbering itself; it does not use the 14// ValueNumbering analysis passes. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "gvn" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/BasicBlock.h" 21#include "llvm/Constants.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/GlobalVariable.h" 24#include "llvm/Function.h" 25#include "llvm/IntrinsicInst.h" 26#include "llvm/LLVMContext.h" 27#include "llvm/Operator.h" 28#include "llvm/Value.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/DepthFirstIterator.h" 31#include "llvm/ADT/PostOrderIterator.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/SmallVector.h" 34#include "llvm/ADT/Statistic.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/ConstantFolding.h" 37#include "llvm/Analysis/Dominators.h" 38#include "llvm/Analysis/Loads.h" 39#include "llvm/Analysis/MemoryBuiltins.h" 40#include "llvm/Analysis/MemoryDependenceAnalysis.h" 41#include "llvm/Analysis/PHITransAddr.h" 42#include "llvm/Support/CFG.h" 43#include "llvm/Support/CommandLine.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/Support/ErrorHandling.h" 46#include "llvm/Support/GetElementPtrTypeIterator.h" 47#include "llvm/Support/IRBuilder.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/Target/TargetData.h" 50#include "llvm/Transforms/Utils/BasicBlockUtils.h" 51#include "llvm/Transforms/Utils/Local.h" 52#include "llvm/Transforms/Utils/SSAUpdater.h" 53using namespace llvm; 54 55STATISTIC(NumGVNInstr, "Number of instructions deleted"); 56STATISTIC(NumGVNLoad, "Number of loads deleted"); 57STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 58STATISTIC(NumGVNBlocks, "Number of blocks merged"); 59STATISTIC(NumPRELoad, "Number of loads PRE'd"); 60 61static cl::opt<bool> EnablePRE("enable-pre", 62 cl::init(true), cl::Hidden); 63static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 64 65//===----------------------------------------------------------------------===// 66// ValueTable Class 67//===----------------------------------------------------------------------===// 68 69/// This class holds the mapping between values and value numbers. It is used 70/// as an efficient mechanism to determine the expression-wise equivalence of 71/// two values. 72namespace { 73 struct Expression { 74 enum ExpressionOpcode { 75 ADD = Instruction::Add, 76 FADD = Instruction::FAdd, 77 SUB = Instruction::Sub, 78 FSUB = Instruction::FSub, 79 MUL = Instruction::Mul, 80 FMUL = Instruction::FMul, 81 UDIV = Instruction::UDiv, 82 SDIV = Instruction::SDiv, 83 FDIV = Instruction::FDiv, 84 UREM = Instruction::URem, 85 SREM = Instruction::SRem, 86 FREM = Instruction::FRem, 87 SHL = Instruction::Shl, 88 LSHR = Instruction::LShr, 89 ASHR = Instruction::AShr, 90 AND = Instruction::And, 91 OR = Instruction::Or, 92 XOR = Instruction::Xor, 93 TRUNC = Instruction::Trunc, 94 ZEXT = Instruction::ZExt, 95 SEXT = Instruction::SExt, 96 FPTOUI = Instruction::FPToUI, 97 FPTOSI = Instruction::FPToSI, 98 UITOFP = Instruction::UIToFP, 99 SITOFP = Instruction::SIToFP, 100 FPTRUNC = Instruction::FPTrunc, 101 FPEXT = Instruction::FPExt, 102 PTRTOINT = Instruction::PtrToInt, 103 INTTOPTR = Instruction::IntToPtr, 104 BITCAST = Instruction::BitCast, 105 ICMPEQ, ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 106 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 107 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 108 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 109 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 110 SHUFFLE, SELECT, GEP, CALL, CONSTANT, 111 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE }; 112 113 ExpressionOpcode opcode; 114 const Type* type; 115 SmallVector<uint32_t, 4> varargs; 116 Value *function; 117 118 Expression() { } 119 Expression(ExpressionOpcode o) : opcode(o) { } 120 121 bool operator==(const Expression &other) const { 122 if (opcode != other.opcode) 123 return false; 124 else if (opcode == EMPTY || opcode == TOMBSTONE) 125 return true; 126 else if (type != other.type) 127 return false; 128 else if (function != other.function) 129 return false; 130 else { 131 if (varargs.size() != other.varargs.size()) 132 return false; 133 134 for (size_t i = 0; i < varargs.size(); ++i) 135 if (varargs[i] != other.varargs[i]) 136 return false; 137 138 return true; 139 } 140 } 141 142 /*bool operator!=(const Expression &other) const { 143 return !(*this == other); 144 }*/ 145 }; 146 147 class ValueTable { 148 private: 149 DenseMap<Value*, uint32_t> valueNumbering; 150 DenseMap<Expression, uint32_t> expressionNumbering; 151 AliasAnalysis* AA; 152 MemoryDependenceAnalysis* MD; 153 DominatorTree* DT; 154 155 uint32_t nextValueNumber; 156 157 Expression::ExpressionOpcode getOpcode(CmpInst* C); 158 Expression create_expression(BinaryOperator* BO); 159 Expression create_expression(CmpInst* C); 160 Expression create_expression(ShuffleVectorInst* V); 161 Expression create_expression(ExtractElementInst* C); 162 Expression create_expression(InsertElementInst* V); 163 Expression create_expression(SelectInst* V); 164 Expression create_expression(CastInst* C); 165 Expression create_expression(GetElementPtrInst* G); 166 Expression create_expression(CallInst* C); 167 Expression create_expression(ExtractValueInst* C); 168 Expression create_expression(InsertValueInst* C); 169 170 uint32_t lookup_or_add_call(CallInst* C); 171 public: 172 ValueTable() : nextValueNumber(1) { } 173 uint32_t lookup_or_add(Value *V); 174 uint32_t lookup(Value *V) const; 175 void add(Value *V, uint32_t num); 176 void clear(); 177 void erase(Value *v); 178 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 179 AliasAnalysis *getAliasAnalysis() const { return AA; } 180 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 181 void setDomTree(DominatorTree* D) { DT = D; } 182 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 183 void verifyRemoved(const Value *) const; 184 }; 185} 186 187namespace llvm { 188template <> struct DenseMapInfo<Expression> { 189 static inline Expression getEmptyKey() { 190 return Expression(Expression::EMPTY); 191 } 192 193 static inline Expression getTombstoneKey() { 194 return Expression(Expression::TOMBSTONE); 195 } 196 197 static unsigned getHashValue(const Expression e) { 198 unsigned hash = e.opcode; 199 200 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 201 (unsigned)((uintptr_t)e.type >> 9)); 202 203 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 204 E = e.varargs.end(); I != E; ++I) 205 hash = *I + hash * 37; 206 207 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 208 (unsigned)((uintptr_t)e.function >> 9)) + 209 hash * 37; 210 211 return hash; 212 } 213 static bool isEqual(const Expression &LHS, const Expression &RHS) { 214 return LHS == RHS; 215 } 216}; 217 218template <> 219struct isPodLike<Expression> { static const bool value = true; }; 220 221} 222 223//===----------------------------------------------------------------------===// 224// ValueTable Internal Functions 225//===----------------------------------------------------------------------===// 226 227Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 228 if (isa<ICmpInst>(C)) { 229 switch (C->getPredicate()) { 230 default: // THIS SHOULD NEVER HAPPEN 231 llvm_unreachable("Comparison with unknown predicate?"); 232 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 233 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 234 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 235 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 236 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 237 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 238 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 239 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 240 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 241 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 242 } 243 } else { 244 switch (C->getPredicate()) { 245 default: // THIS SHOULD NEVER HAPPEN 246 llvm_unreachable("Comparison with unknown predicate?"); 247 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 248 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 249 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 250 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 251 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 252 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 253 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 254 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 255 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 256 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 257 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 258 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 259 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 260 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 261 } 262 } 263} 264 265Expression ValueTable::create_expression(CallInst* C) { 266 Expression e; 267 268 e.type = C->getType(); 269 e.function = C->getCalledFunction(); 270 e.opcode = Expression::CALL; 271 272 CallSite CS(C); 273 for (CallInst::op_iterator I = CS.arg_begin(), E = CS.arg_end(); 274 I != E; ++I) 275 e.varargs.push_back(lookup_or_add(*I)); 276 277 return e; 278} 279 280Expression ValueTable::create_expression(BinaryOperator* BO) { 281 Expression e; 282 e.varargs.push_back(lookup_or_add(BO->getOperand(0))); 283 e.varargs.push_back(lookup_or_add(BO->getOperand(1))); 284 e.function = 0; 285 e.type = BO->getType(); 286 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode()); 287 288 return e; 289} 290 291Expression ValueTable::create_expression(CmpInst* C) { 292 Expression e; 293 294 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 295 e.varargs.push_back(lookup_or_add(C->getOperand(1))); 296 e.function = 0; 297 e.type = C->getType(); 298 e.opcode = getOpcode(C); 299 300 return e; 301} 302 303Expression ValueTable::create_expression(CastInst* C) { 304 Expression e; 305 306 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 307 e.function = 0; 308 e.type = C->getType(); 309 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode()); 310 311 return e; 312} 313 314Expression ValueTable::create_expression(ShuffleVectorInst* S) { 315 Expression e; 316 317 e.varargs.push_back(lookup_or_add(S->getOperand(0))); 318 e.varargs.push_back(lookup_or_add(S->getOperand(1))); 319 e.varargs.push_back(lookup_or_add(S->getOperand(2))); 320 e.function = 0; 321 e.type = S->getType(); 322 e.opcode = Expression::SHUFFLE; 323 324 return e; 325} 326 327Expression ValueTable::create_expression(ExtractElementInst* E) { 328 Expression e; 329 330 e.varargs.push_back(lookup_or_add(E->getOperand(0))); 331 e.varargs.push_back(lookup_or_add(E->getOperand(1))); 332 e.function = 0; 333 e.type = E->getType(); 334 e.opcode = Expression::EXTRACT; 335 336 return e; 337} 338 339Expression ValueTable::create_expression(InsertElementInst* I) { 340 Expression e; 341 342 e.varargs.push_back(lookup_or_add(I->getOperand(0))); 343 e.varargs.push_back(lookup_or_add(I->getOperand(1))); 344 e.varargs.push_back(lookup_or_add(I->getOperand(2))); 345 e.function = 0; 346 e.type = I->getType(); 347 e.opcode = Expression::INSERT; 348 349 return e; 350} 351 352Expression ValueTable::create_expression(SelectInst* I) { 353 Expression e; 354 355 e.varargs.push_back(lookup_or_add(I->getCondition())); 356 e.varargs.push_back(lookup_or_add(I->getTrueValue())); 357 e.varargs.push_back(lookup_or_add(I->getFalseValue())); 358 e.function = 0; 359 e.type = I->getType(); 360 e.opcode = Expression::SELECT; 361 362 return e; 363} 364 365Expression ValueTable::create_expression(GetElementPtrInst* G) { 366 Expression e; 367 368 e.varargs.push_back(lookup_or_add(G->getPointerOperand())); 369 e.function = 0; 370 e.type = G->getType(); 371 e.opcode = Expression::GEP; 372 373 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 374 I != E; ++I) 375 e.varargs.push_back(lookup_or_add(*I)); 376 377 return e; 378} 379 380Expression ValueTable::create_expression(ExtractValueInst* E) { 381 Expression e; 382 383 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 384 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 385 II != IE; ++II) 386 e.varargs.push_back(*II); 387 e.function = 0; 388 e.type = E->getType(); 389 e.opcode = Expression::EXTRACTVALUE; 390 391 return e; 392} 393 394Expression ValueTable::create_expression(InsertValueInst* E) { 395 Expression e; 396 397 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 398 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand())); 399 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 400 II != IE; ++II) 401 e.varargs.push_back(*II); 402 e.function = 0; 403 e.type = E->getType(); 404 e.opcode = Expression::INSERTVALUE; 405 406 return e; 407} 408 409//===----------------------------------------------------------------------===// 410// ValueTable External Functions 411//===----------------------------------------------------------------------===// 412 413/// add - Insert a value into the table with a specified value number. 414void ValueTable::add(Value *V, uint32_t num) { 415 valueNumbering.insert(std::make_pair(V, num)); 416} 417 418uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 419 if (AA->doesNotAccessMemory(C)) { 420 Expression exp = create_expression(C); 421 uint32_t& e = expressionNumbering[exp]; 422 if (!e) e = nextValueNumber++; 423 valueNumbering[C] = e; 424 return e; 425 } else if (AA->onlyReadsMemory(C)) { 426 Expression exp = create_expression(C); 427 uint32_t& e = expressionNumbering[exp]; 428 if (!e) { 429 e = nextValueNumber++; 430 valueNumbering[C] = e; 431 return e; 432 } 433 if (!MD) { 434 e = nextValueNumber++; 435 valueNumbering[C] = e; 436 return e; 437 } 438 439 MemDepResult local_dep = MD->getDependency(C); 440 441 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 442 valueNumbering[C] = nextValueNumber; 443 return nextValueNumber++; 444 } 445 446 if (local_dep.isDef()) { 447 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 448 449 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 450 valueNumbering[C] = nextValueNumber; 451 return nextValueNumber++; 452 } 453 454 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 455 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 456 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 457 if (c_vn != cd_vn) { 458 valueNumbering[C] = nextValueNumber; 459 return nextValueNumber++; 460 } 461 } 462 463 uint32_t v = lookup_or_add(local_cdep); 464 valueNumbering[C] = v; 465 return v; 466 } 467 468 // Non-local case. 469 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 470 MD->getNonLocalCallDependency(CallSite(C)); 471 // FIXME: call/call dependencies for readonly calls should return def, not 472 // clobber! Move the checking logic to MemDep! 473 CallInst* cdep = 0; 474 475 // Check to see if we have a single dominating call instruction that is 476 // identical to C. 477 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 478 const NonLocalDepEntry *I = &deps[i]; 479 // Ignore non-local dependencies. 480 if (I->getResult().isNonLocal()) 481 continue; 482 483 // We don't handle non-depedencies. If we already have a call, reject 484 // instruction dependencies. 485 if (I->getResult().isClobber() || cdep != 0) { 486 cdep = 0; 487 break; 488 } 489 490 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 491 // FIXME: All duplicated with non-local case. 492 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 493 cdep = NonLocalDepCall; 494 continue; 495 } 496 497 cdep = 0; 498 break; 499 } 500 501 if (!cdep) { 502 valueNumbering[C] = nextValueNumber; 503 return nextValueNumber++; 504 } 505 506 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 507 valueNumbering[C] = nextValueNumber; 508 return nextValueNumber++; 509 } 510 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 511 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 512 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 513 if (c_vn != cd_vn) { 514 valueNumbering[C] = nextValueNumber; 515 return nextValueNumber++; 516 } 517 } 518 519 uint32_t v = lookup_or_add(cdep); 520 valueNumbering[C] = v; 521 return v; 522 523 } else { 524 valueNumbering[C] = nextValueNumber; 525 return nextValueNumber++; 526 } 527} 528 529/// lookup_or_add - Returns the value number for the specified value, assigning 530/// it a new number if it did not have one before. 531uint32_t ValueTable::lookup_or_add(Value *V) { 532 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 533 if (VI != valueNumbering.end()) 534 return VI->second; 535 536 if (!isa<Instruction>(V)) { 537 valueNumbering[V] = nextValueNumber; 538 return nextValueNumber++; 539 } 540 541 Instruction* I = cast<Instruction>(V); 542 Expression exp; 543 switch (I->getOpcode()) { 544 case Instruction::Call: 545 return lookup_or_add_call(cast<CallInst>(I)); 546 case Instruction::Add: 547 case Instruction::FAdd: 548 case Instruction::Sub: 549 case Instruction::FSub: 550 case Instruction::Mul: 551 case Instruction::FMul: 552 case Instruction::UDiv: 553 case Instruction::SDiv: 554 case Instruction::FDiv: 555 case Instruction::URem: 556 case Instruction::SRem: 557 case Instruction::FRem: 558 case Instruction::Shl: 559 case Instruction::LShr: 560 case Instruction::AShr: 561 case Instruction::And: 562 case Instruction::Or : 563 case Instruction::Xor: 564 exp = create_expression(cast<BinaryOperator>(I)); 565 break; 566 case Instruction::ICmp: 567 case Instruction::FCmp: 568 exp = create_expression(cast<CmpInst>(I)); 569 break; 570 case Instruction::Trunc: 571 case Instruction::ZExt: 572 case Instruction::SExt: 573 case Instruction::FPToUI: 574 case Instruction::FPToSI: 575 case Instruction::UIToFP: 576 case Instruction::SIToFP: 577 case Instruction::FPTrunc: 578 case Instruction::FPExt: 579 case Instruction::PtrToInt: 580 case Instruction::IntToPtr: 581 case Instruction::BitCast: 582 exp = create_expression(cast<CastInst>(I)); 583 break; 584 case Instruction::Select: 585 exp = create_expression(cast<SelectInst>(I)); 586 break; 587 case Instruction::ExtractElement: 588 exp = create_expression(cast<ExtractElementInst>(I)); 589 break; 590 case Instruction::InsertElement: 591 exp = create_expression(cast<InsertElementInst>(I)); 592 break; 593 case Instruction::ShuffleVector: 594 exp = create_expression(cast<ShuffleVectorInst>(I)); 595 break; 596 case Instruction::ExtractValue: 597 exp = create_expression(cast<ExtractValueInst>(I)); 598 break; 599 case Instruction::InsertValue: 600 exp = create_expression(cast<InsertValueInst>(I)); 601 break; 602 case Instruction::GetElementPtr: 603 exp = create_expression(cast<GetElementPtrInst>(I)); 604 break; 605 default: 606 valueNumbering[V] = nextValueNumber; 607 return nextValueNumber++; 608 } 609 610 uint32_t& e = expressionNumbering[exp]; 611 if (!e) e = nextValueNumber++; 612 valueNumbering[V] = e; 613 return e; 614} 615 616/// lookup - Returns the value number of the specified value. Fails if 617/// the value has not yet been numbered. 618uint32_t ValueTable::lookup(Value *V) const { 619 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 620 assert(VI != valueNumbering.end() && "Value not numbered?"); 621 return VI->second; 622} 623 624/// clear - Remove all entries from the ValueTable 625void ValueTable::clear() { 626 valueNumbering.clear(); 627 expressionNumbering.clear(); 628 nextValueNumber = 1; 629} 630 631/// erase - Remove a value from the value numbering 632void ValueTable::erase(Value *V) { 633 valueNumbering.erase(V); 634} 635 636/// verifyRemoved - Verify that the value is removed from all internal data 637/// structures. 638void ValueTable::verifyRemoved(const Value *V) const { 639 for (DenseMap<Value*, uint32_t>::const_iterator 640 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 641 assert(I->first != V && "Inst still occurs in value numbering map!"); 642 } 643} 644 645//===----------------------------------------------------------------------===// 646// GVN Pass 647//===----------------------------------------------------------------------===// 648 649namespace { 650 struct ValueNumberScope { 651 ValueNumberScope* parent; 652 DenseMap<uint32_t, Value*> table; 653 654 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 655 }; 656} 657 658namespace { 659 660 class GVN : public FunctionPass { 661 bool runOnFunction(Function &F); 662 public: 663 static char ID; // Pass identification, replacement for typeid 664 explicit GVN(bool noloads = false) 665 : FunctionPass(ID), NoLoads(noloads), MD(0) { 666 initializeGVNPass(*PassRegistry::getPassRegistry()); 667 } 668 669 private: 670 bool NoLoads; 671 MemoryDependenceAnalysis *MD; 672 DominatorTree *DT; 673 674 ValueTable VN; 675 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 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 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 682 AU.addRequired<DominatorTree>(); 683 if (!NoLoads) 684 AU.addRequired<MemoryDependenceAnalysis>(); 685 AU.addRequired<AliasAnalysis>(); 686 687 AU.addPreserved<DominatorTree>(); 688 AU.addPreserved<AliasAnalysis>(); 689 } 690 691 // Helper fuctions 692 // FIXME: eliminate or document these better 693 bool processLoad(LoadInst* L, 694 SmallVectorImpl<Instruction*> &toErase); 695 bool processInstruction(Instruction *I, 696 SmallVectorImpl<Instruction*> &toErase); 697 bool processNonLocalLoad(LoadInst* L, 698 SmallVectorImpl<Instruction*> &toErase); 699 bool processBlock(BasicBlock *BB); 700 void dump(DenseMap<uint32_t, Value*>& d); 701 bool iterateOnFunction(Function &F); 702 Value *CollapsePhi(PHINode* p); 703 bool performPRE(Function& F); 704 Value *lookupNumber(BasicBlock *BB, uint32_t num); 705 void cleanupGlobalSets(); 706 void verifyRemoved(const Instruction *I) const; 707 bool splitCriticalEdges(); 708 }; 709 710 char GVN::ID = 0; 711} 712 713// createGVNPass - The public interface to this file... 714FunctionPass *llvm::createGVNPass(bool NoLoads) { 715 return new GVN(NoLoads); 716} 717 718INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 719INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 720INITIALIZE_PASS_DEPENDENCY(DominatorTree) 721INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 722INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 723 724void GVN::dump(DenseMap<uint32_t, Value*>& d) { 725 errs() << "{\n"; 726 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 727 E = d.end(); I != E; ++I) { 728 errs() << I->first << "\n"; 729 I->second->dump(); 730 } 731 errs() << "}\n"; 732} 733 734static bool isSafeReplacement(PHINode* p, Instruction *inst) { 735 if (!isa<PHINode>(inst)) 736 return true; 737 738 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 739 UI != E; ++UI) 740 if (PHINode* use_phi = dyn_cast<PHINode>(*UI)) 741 if (use_phi->getParent() == inst->getParent()) 742 return false; 743 744 return true; 745} 746 747Value *GVN::CollapsePhi(PHINode *PN) { 748 Value *ConstVal = PN->hasConstantValue(DT); 749 if (!ConstVal) return 0; 750 751 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 752 if (!Inst) 753 return ConstVal; 754 755 if (DT->dominates(Inst, PN)) 756 if (isSafeReplacement(PN, Inst)) 757 return Inst; 758 return 0; 759} 760 761/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 762/// we're analyzing is fully available in the specified block. As we go, keep 763/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 764/// map is actually a tri-state map with the following values: 765/// 0) we know the block *is not* fully available. 766/// 1) we know the block *is* fully available. 767/// 2) we do not know whether the block is fully available or not, but we are 768/// currently speculating that it will be. 769/// 3) we are speculating for this block and have used that to speculate for 770/// other blocks. 771static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 772 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 773 // Optimistically assume that the block is fully available and check to see 774 // if we already know about this block in one lookup. 775 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 776 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 777 778 // If the entry already existed for this block, return the precomputed value. 779 if (!IV.second) { 780 // If this is a speculative "available" value, mark it as being used for 781 // speculation of other blocks. 782 if (IV.first->second == 2) 783 IV.first->second = 3; 784 return IV.first->second != 0; 785 } 786 787 // Otherwise, see if it is fully available in all predecessors. 788 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 789 790 // If this block has no predecessors, it isn't live-in here. 791 if (PI == PE) 792 goto SpeculationFailure; 793 794 for (; PI != PE; ++PI) 795 // If the value isn't fully available in one of our predecessors, then it 796 // isn't fully available in this block either. Undo our previous 797 // optimistic assumption and bail out. 798 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 799 goto SpeculationFailure; 800 801 return true; 802 803// SpeculationFailure - If we get here, we found out that this is not, after 804// all, a fully-available block. We have a problem if we speculated on this and 805// used the speculation to mark other blocks as available. 806SpeculationFailure: 807 char &BBVal = FullyAvailableBlocks[BB]; 808 809 // If we didn't speculate on this, just return with it set to false. 810 if (BBVal == 2) { 811 BBVal = 0; 812 return false; 813 } 814 815 // If we did speculate on this value, we could have blocks set to 1 that are 816 // incorrect. Walk the (transitive) successors of this block and mark them as 817 // 0 if set to one. 818 SmallVector<BasicBlock*, 32> BBWorklist; 819 BBWorklist.push_back(BB); 820 821 do { 822 BasicBlock *Entry = BBWorklist.pop_back_val(); 823 // Note that this sets blocks to 0 (unavailable) if they happen to not 824 // already be in FullyAvailableBlocks. This is safe. 825 char &EntryVal = FullyAvailableBlocks[Entry]; 826 if (EntryVal == 0) continue; // Already unavailable. 827 828 // Mark as unavailable. 829 EntryVal = 0; 830 831 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 832 BBWorklist.push_back(*I); 833 } while (!BBWorklist.empty()); 834 835 return false; 836} 837 838 839/// CanCoerceMustAliasedValueToLoad - Return true if 840/// CoerceAvailableValueToLoadType will succeed. 841static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 842 const Type *LoadTy, 843 const TargetData &TD) { 844 // If the loaded or stored value is an first class array or struct, don't try 845 // to transform them. We need to be able to bitcast to integer. 846 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 847 StoredVal->getType()->isStructTy() || 848 StoredVal->getType()->isArrayTy()) 849 return false; 850 851 // The store has to be at least as big as the load. 852 if (TD.getTypeSizeInBits(StoredVal->getType()) < 853 TD.getTypeSizeInBits(LoadTy)) 854 return false; 855 856 return true; 857} 858 859 860/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 861/// then a load from a must-aliased pointer of a different type, try to coerce 862/// the stored value. LoadedTy is the type of the load we want to replace and 863/// InsertPt is the place to insert new instructions. 864/// 865/// If we can't do it, return null. 866static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 867 const Type *LoadedTy, 868 Instruction *InsertPt, 869 const TargetData &TD) { 870 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 871 return 0; 872 873 const Type *StoredValTy = StoredVal->getType(); 874 875 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 876 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 877 878 // If the store and reload are the same size, we can always reuse it. 879 if (StoreSize == LoadSize) { 880 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) { 881 // Pointer to Pointer -> use bitcast. 882 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 883 } 884 885 // Convert source pointers to integers, which can be bitcast. 886 if (StoredValTy->isPointerTy()) { 887 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 888 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 889 } 890 891 const Type *TypeToCastTo = LoadedTy; 892 if (TypeToCastTo->isPointerTy()) 893 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 894 895 if (StoredValTy != TypeToCastTo) 896 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 897 898 // Cast to pointer if the load needs a pointer type. 899 if (LoadedTy->isPointerTy()) 900 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 901 902 return StoredVal; 903 } 904 905 // If the loaded value is smaller than the available value, then we can 906 // extract out a piece from it. If the available value is too small, then we 907 // can't do anything. 908 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 909 910 // Convert source pointers to integers, which can be manipulated. 911 if (StoredValTy->isPointerTy()) { 912 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 913 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 914 } 915 916 // Convert vectors and fp to integer, which can be manipulated. 917 if (!StoredValTy->isIntegerTy()) { 918 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 919 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 920 } 921 922 // If this is a big-endian system, we need to shift the value down to the low 923 // bits so that a truncate will work. 924 if (TD.isBigEndian()) { 925 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 926 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 927 } 928 929 // Truncate the integer to the right size now. 930 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 931 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 932 933 if (LoadedTy == NewIntTy) 934 return StoredVal; 935 936 // If the result is a pointer, inttoptr. 937 if (LoadedTy->isPointerTy()) 938 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 939 940 // Otherwise, bitcast. 941 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 942} 943 944/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 945/// be expressed as a base pointer plus a constant offset. Return the base and 946/// offset to the caller. 947static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 948 const TargetData &TD) { 949 Operator *PtrOp = dyn_cast<Operator>(Ptr); 950 if (PtrOp == 0) return Ptr; 951 952 // Just look through bitcasts. 953 if (PtrOp->getOpcode() == Instruction::BitCast) 954 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 955 956 // If this is a GEP with constant indices, we can look through it. 957 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 958 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 959 960 gep_type_iterator GTI = gep_type_begin(GEP); 961 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 962 ++I, ++GTI) { 963 ConstantInt *OpC = cast<ConstantInt>(*I); 964 if (OpC->isZero()) continue; 965 966 // Handle a struct and array indices which add their offset to the pointer. 967 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 968 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 969 } else { 970 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 971 Offset += OpC->getSExtValue()*Size; 972 } 973 } 974 975 // Re-sign extend from the pointer size if needed to get overflow edge cases 976 // right. 977 unsigned PtrSize = TD.getPointerSizeInBits(); 978 if (PtrSize < 64) 979 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 980 981 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 982} 983 984 985/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 986/// memdep query of a load that ends up being a clobbering memory write (store, 987/// memset, memcpy, memmove). This means that the write *may* provide bits used 988/// by the load but we can't be sure because the pointers don't mustalias. 989/// 990/// Check this case to see if there is anything more we can do before we give 991/// up. This returns -1 if we have to give up, or a byte number in the stored 992/// value of the piece that feeds the load. 993static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 994 Value *WritePtr, 995 uint64_t WriteSizeInBits, 996 const TargetData &TD) { 997 // If the loaded or stored value is an first class array or struct, don't try 998 // to transform them. We need to be able to bitcast to integer. 999 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 1000 return -1; 1001 1002 int64_t StoreOffset = 0, LoadOffset = 0; 1003 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD); 1004 Value *LoadBase = 1005 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 1006 if (StoreBase != LoadBase) 1007 return -1; 1008 1009 // If the load and store are to the exact same address, they should have been 1010 // a must alias. AA must have gotten confused. 1011 // FIXME: Study to see if/when this happens. One case is forwarding a memset 1012 // to a load from the base of the memset. 1013#if 0 1014 if (LoadOffset == StoreOffset) { 1015 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1016 << "Base = " << *StoreBase << "\n" 1017 << "Store Ptr = " << *WritePtr << "\n" 1018 << "Store Offs = " << StoreOffset << "\n" 1019 << "Load Ptr = " << *LoadPtr << "\n"; 1020 abort(); 1021 } 1022#endif 1023 1024 // If the load and store don't overlap at all, the store doesn't provide 1025 // anything to the load. In this case, they really don't alias at all, AA 1026 // must have gotten confused. 1027 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 1028 // remove this check, as it is duplicated with what we have below. 1029 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 1030 1031 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 1032 return -1; 1033 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 1034 LoadSize >>= 3; 1035 1036 1037 bool isAAFailure = false; 1038 if (StoreOffset < LoadOffset) 1039 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1040 else 1041 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1042 1043 if (isAAFailure) { 1044#if 0 1045 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1046 << "Base = " << *StoreBase << "\n" 1047 << "Store Ptr = " << *WritePtr << "\n" 1048 << "Store Offs = " << StoreOffset << "\n" 1049 << "Load Ptr = " << *LoadPtr << "\n"; 1050 abort(); 1051#endif 1052 return -1; 1053 } 1054 1055 // If the Load isn't completely contained within the stored bits, we don't 1056 // have all the bits to feed it. We could do something crazy in the future 1057 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1058 // valuable. 1059 if (StoreOffset > LoadOffset || 1060 StoreOffset+StoreSize < LoadOffset+LoadSize) 1061 return -1; 1062 1063 // Okay, we can do this transformation. Return the number of bytes into the 1064 // store that the load is. 1065 return LoadOffset-StoreOffset; 1066} 1067 1068/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1069/// memdep query of a load that ends up being a clobbering store. 1070static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 1071 StoreInst *DepSI, 1072 const TargetData &TD) { 1073 // Cannot handle reading from store of first-class aggregate yet. 1074 if (DepSI->getValueOperand()->getType()->isStructTy() || 1075 DepSI->getValueOperand()->getType()->isArrayTy()) 1076 return -1; 1077 1078 Value *StorePtr = DepSI->getPointerOperand(); 1079 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 1080 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1081 StorePtr, StoreSize, TD); 1082} 1083 1084static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 1085 MemIntrinsic *MI, 1086 const TargetData &TD) { 1087 // If the mem operation is a non-constant size, we can't handle it. 1088 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1089 if (SizeCst == 0) return -1; 1090 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1091 1092 // If this is memset, we just need to see if the offset is valid in the size 1093 // of the memset.. 1094 if (MI->getIntrinsicID() == Intrinsic::memset) 1095 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1096 MemSizeInBits, TD); 1097 1098 // If we have a memcpy/memmove, the only case we can handle is if this is a 1099 // copy from constant memory. In that case, we can read directly from the 1100 // constant memory. 1101 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1102 1103 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1104 if (Src == 0) return -1; 1105 1106 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject()); 1107 if (GV == 0 || !GV->isConstant()) return -1; 1108 1109 // See if the access is within the bounds of the transfer. 1110 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1111 MI->getDest(), MemSizeInBits, TD); 1112 if (Offset == -1) 1113 return Offset; 1114 1115 // Otherwise, see if we can constant fold a load from the constant with the 1116 // offset applied as appropriate. 1117 Src = ConstantExpr::getBitCast(Src, 1118 llvm::Type::getInt8PtrTy(Src->getContext())); 1119 Constant *OffsetCst = 1120 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1121 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1122 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1123 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1124 return Offset; 1125 return -1; 1126} 1127 1128 1129/// GetStoreValueForLoad - This function is called when we have a 1130/// memdep query of a load that ends up being a clobbering store. This means 1131/// that the store *may* provide bits used by the load but we can't be sure 1132/// because the pointers don't mustalias. Check this case to see if there is 1133/// anything more we can do before we give up. 1134static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1135 const Type *LoadTy, 1136 Instruction *InsertPt, const TargetData &TD){ 1137 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1138 1139 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1140 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1141 1142 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1143 1144 // Compute which bits of the stored value are being used by the load. Convert 1145 // to an integer type to start with. 1146 if (SrcVal->getType()->isPointerTy()) 1147 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 1148 if (!SrcVal->getType()->isIntegerTy()) 1149 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1150 "tmp"); 1151 1152 // Shift the bits to the least significant depending on endianness. 1153 unsigned ShiftAmt; 1154 if (TD.isLittleEndian()) 1155 ShiftAmt = Offset*8; 1156 else 1157 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1158 1159 if (ShiftAmt) 1160 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 1161 1162 if (LoadSize != StoreSize) 1163 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1164 "tmp"); 1165 1166 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1167} 1168 1169/// GetMemInstValueForLoad - This function is called when we have a 1170/// memdep query of a load that ends up being a clobbering mem intrinsic. 1171static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1172 const Type *LoadTy, Instruction *InsertPt, 1173 const TargetData &TD){ 1174 LLVMContext &Ctx = LoadTy->getContext(); 1175 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1176 1177 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1178 1179 // We know that this method is only called when the mem transfer fully 1180 // provides the bits for the load. 1181 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1182 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1183 // independently of what the offset is. 1184 Value *Val = MSI->getValue(); 1185 if (LoadSize != 1) 1186 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1187 1188 Value *OneElt = Val; 1189 1190 // Splat the value out to the right number of bits. 1191 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1192 // If we can double the number of bytes set, do it. 1193 if (NumBytesSet*2 <= LoadSize) { 1194 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1195 Val = Builder.CreateOr(Val, ShVal); 1196 NumBytesSet <<= 1; 1197 continue; 1198 } 1199 1200 // Otherwise insert one byte at a time. 1201 Value *ShVal = Builder.CreateShl(Val, 1*8); 1202 Val = Builder.CreateOr(OneElt, ShVal); 1203 ++NumBytesSet; 1204 } 1205 1206 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1207 } 1208 1209 // Otherwise, this is a memcpy/memmove from a constant global. 1210 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1211 Constant *Src = cast<Constant>(MTI->getSource()); 1212 1213 // Otherwise, see if we can constant fold a load from the constant with the 1214 // offset applied as appropriate. 1215 Src = ConstantExpr::getBitCast(Src, 1216 llvm::Type::getInt8PtrTy(Src->getContext())); 1217 Constant *OffsetCst = 1218 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1219 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1220 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1221 return ConstantFoldLoadFromConstPtr(Src, &TD); 1222} 1223 1224namespace { 1225 1226struct AvailableValueInBlock { 1227 /// BB - The basic block in question. 1228 BasicBlock *BB; 1229 enum ValType { 1230 SimpleVal, // A simple offsetted value that is accessed. 1231 MemIntrin // A memory intrinsic which is loaded from. 1232 }; 1233 1234 /// V - The value that is live out of the block. 1235 PointerIntPair<Value *, 1, ValType> Val; 1236 1237 /// Offset - The byte offset in Val that is interesting for the load query. 1238 unsigned Offset; 1239 1240 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1241 unsigned Offset = 0) { 1242 AvailableValueInBlock Res; 1243 Res.BB = BB; 1244 Res.Val.setPointer(V); 1245 Res.Val.setInt(SimpleVal); 1246 Res.Offset = Offset; 1247 return Res; 1248 } 1249 1250 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1251 unsigned Offset = 0) { 1252 AvailableValueInBlock Res; 1253 Res.BB = BB; 1254 Res.Val.setPointer(MI); 1255 Res.Val.setInt(MemIntrin); 1256 Res.Offset = Offset; 1257 return Res; 1258 } 1259 1260 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1261 Value *getSimpleValue() const { 1262 assert(isSimpleValue() && "Wrong accessor"); 1263 return Val.getPointer(); 1264 } 1265 1266 MemIntrinsic *getMemIntrinValue() const { 1267 assert(!isSimpleValue() && "Wrong accessor"); 1268 return cast<MemIntrinsic>(Val.getPointer()); 1269 } 1270 1271 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1272 /// defined here to the specified type. This handles various coercion cases. 1273 Value *MaterializeAdjustedValue(const Type *LoadTy, 1274 const TargetData *TD) const { 1275 Value *Res; 1276 if (isSimpleValue()) { 1277 Res = getSimpleValue(); 1278 if (Res->getType() != LoadTy) { 1279 assert(TD && "Need target data to handle type mismatch case"); 1280 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1281 *TD); 1282 1283 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1284 << *getSimpleValue() << '\n' 1285 << *Res << '\n' << "\n\n\n"); 1286 } 1287 } else { 1288 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1289 LoadTy, BB->getTerminator(), *TD); 1290 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1291 << " " << *getMemIntrinValue() << '\n' 1292 << *Res << '\n' << "\n\n\n"); 1293 } 1294 return Res; 1295 } 1296}; 1297 1298} 1299 1300/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1301/// construct SSA form, allowing us to eliminate LI. This returns the value 1302/// that should be used at LI's definition site. 1303static Value *ConstructSSAForLoadSet(LoadInst *LI, 1304 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1305 const TargetData *TD, 1306 const DominatorTree &DT, 1307 AliasAnalysis *AA) { 1308 // Check for the fully redundant, dominating load case. In this case, we can 1309 // just use the dominating value directly. 1310 if (ValuesPerBlock.size() == 1 && 1311 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent())) 1312 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD); 1313 1314 // Otherwise, we have to construct SSA form. 1315 SmallVector<PHINode*, 8> NewPHIs; 1316 SSAUpdater SSAUpdate(&NewPHIs); 1317 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1318 1319 const Type *LoadTy = LI->getType(); 1320 1321 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1322 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1323 BasicBlock *BB = AV.BB; 1324 1325 if (SSAUpdate.HasValueForBlock(BB)) 1326 continue; 1327 1328 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD)); 1329 } 1330 1331 // Perform PHI construction. 1332 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1333 1334 // If new PHI nodes were created, notify alias analysis. 1335 if (V->getType()->isPointerTy()) 1336 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1337 AA->copyValue(LI, NewPHIs[i]); 1338 1339 return V; 1340} 1341 1342static bool isLifetimeStart(const Instruction *Inst) { 1343 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1344 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1345 return false; 1346} 1347 1348/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1349/// non-local by performing PHI construction. 1350bool GVN::processNonLocalLoad(LoadInst *LI, 1351 SmallVectorImpl<Instruction*> &toErase) { 1352 // Find the non-local dependencies of the load. 1353 SmallVector<NonLocalDepResult, 64> Deps; 1354 AliasAnalysis::Location Loc(LI->getPointerOperand(), 1355 VN.getAliasAnalysis()->getTypeStoreSize(LI->getType()), 1356 LI->getMetadata(LLVMContext::MD_tbaa)); 1357 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), 1358 Deps); 1359 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1360 // << Deps.size() << *LI << '\n'); 1361 1362 // If we had to process more than one hundred blocks to find the 1363 // dependencies, this load isn't worth worrying about. Optimizing 1364 // it will be too expensive. 1365 if (Deps.size() > 100) 1366 return false; 1367 1368 // If we had a phi translation failure, we'll have a single entry which is a 1369 // clobber in the current block. Reject this early. 1370 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) { 1371 DEBUG( 1372 dbgs() << "GVN: non-local load "; 1373 WriteAsOperand(dbgs(), LI); 1374 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1375 ); 1376 return false; 1377 } 1378 1379 // Filter out useless results (non-locals, etc). Keep track of the blocks 1380 // where we have a value available in repl, also keep track of whether we see 1381 // dependencies that produce an unknown value for the load (such as a call 1382 // that could potentially clobber the load). 1383 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1384 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1385 1386 const TargetData *TD = 0; 1387 1388 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1389 BasicBlock *DepBB = Deps[i].getBB(); 1390 MemDepResult DepInfo = Deps[i].getResult(); 1391 1392 if (DepInfo.isClobber()) { 1393 // The address being loaded in this non-local block may not be the same as 1394 // the pointer operand of the load if PHI translation occurs. Make sure 1395 // to consider the right address. 1396 Value *Address = Deps[i].getAddress(); 1397 1398 // If the dependence is to a store that writes to a superset of the bits 1399 // read by the load, we can extract the bits we need for the load from the 1400 // stored value. 1401 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1402 if (TD == 0) 1403 TD = getAnalysisIfAvailable<TargetData>(); 1404 if (TD && Address) { 1405 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1406 DepSI, *TD); 1407 if (Offset != -1) { 1408 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1409 DepSI->getValueOperand(), 1410 Offset)); 1411 continue; 1412 } 1413 } 1414 } 1415 1416 // If the clobbering value is a memset/memcpy/memmove, see if we can 1417 // forward a value on from it. 1418 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1419 if (TD == 0) 1420 TD = getAnalysisIfAvailable<TargetData>(); 1421 if (TD && Address) { 1422 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1423 DepMI, *TD); 1424 if (Offset != -1) { 1425 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1426 Offset)); 1427 continue; 1428 } 1429 } 1430 } 1431 1432 UnavailableBlocks.push_back(DepBB); 1433 continue; 1434 } 1435 1436 Instruction *DepInst = DepInfo.getInst(); 1437 1438 // Loading the allocation -> undef. 1439 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1440 // Loading immediately after lifetime begin -> undef. 1441 isLifetimeStart(DepInst)) { 1442 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1443 UndefValue::get(LI->getType()))); 1444 continue; 1445 } 1446 1447 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1448 // Reject loads and stores that are to the same address but are of 1449 // different types if we have to. 1450 if (S->getValueOperand()->getType() != LI->getType()) { 1451 if (TD == 0) 1452 TD = getAnalysisIfAvailable<TargetData>(); 1453 1454 // If the stored value is larger or equal to the loaded value, we can 1455 // reuse it. 1456 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1457 LI->getType(), *TD)) { 1458 UnavailableBlocks.push_back(DepBB); 1459 continue; 1460 } 1461 } 1462 1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1464 S->getValueOperand())); 1465 continue; 1466 } 1467 1468 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1469 // If the types mismatch and we can't handle it, reject reuse of the load. 1470 if (LD->getType() != LI->getType()) { 1471 if (TD == 0) 1472 TD = getAnalysisIfAvailable<TargetData>(); 1473 1474 // If the stored value is larger or equal to the loaded value, we can 1475 // reuse it. 1476 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1477 UnavailableBlocks.push_back(DepBB); 1478 continue; 1479 } 1480 } 1481 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1482 continue; 1483 } 1484 1485 UnavailableBlocks.push_back(DepBB); 1486 continue; 1487 } 1488 1489 // If we have no predecessors that produce a known value for this load, exit 1490 // early. 1491 if (ValuesPerBlock.empty()) return false; 1492 1493 // If all of the instructions we depend on produce a known value for this 1494 // load, then it is fully redundant and we can use PHI insertion to compute 1495 // its value. Insert PHIs and remove the fully redundant value now. 1496 if (UnavailableBlocks.empty()) { 1497 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1498 1499 // Perform PHI construction. 1500 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1501 VN.getAliasAnalysis()); 1502 LI->replaceAllUsesWith(V); 1503 1504 if (isa<PHINode>(V)) 1505 V->takeName(LI); 1506 if (V->getType()->isPointerTy()) 1507 MD->invalidateCachedPointerInfo(V); 1508 VN.erase(LI); 1509 toErase.push_back(LI); 1510 ++NumGVNLoad; 1511 return true; 1512 } 1513 1514 if (!EnablePRE || !EnableLoadPRE) 1515 return false; 1516 1517 // Okay, we have *some* definitions of the value. This means that the value 1518 // is available in some of our (transitive) predecessors. Lets think about 1519 // doing PRE of this load. This will involve inserting a new load into the 1520 // predecessor when it's not available. We could do this in general, but 1521 // prefer to not increase code size. As such, we only do this when we know 1522 // that we only have to insert *one* load (which means we're basically moving 1523 // the load, not inserting a new one). 1524 1525 SmallPtrSet<BasicBlock *, 4> Blockers; 1526 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1527 Blockers.insert(UnavailableBlocks[i]); 1528 1529 // Lets find first basic block with more than one predecessor. Walk backwards 1530 // through predecessors if needed. 1531 BasicBlock *LoadBB = LI->getParent(); 1532 BasicBlock *TmpBB = LoadBB; 1533 1534 bool isSinglePred = false; 1535 bool allSingleSucc = true; 1536 while (TmpBB->getSinglePredecessor()) { 1537 isSinglePred = true; 1538 TmpBB = TmpBB->getSinglePredecessor(); 1539 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1540 return false; 1541 if (Blockers.count(TmpBB)) 1542 return false; 1543 1544 // If any of these blocks has more than one successor (i.e. if the edge we 1545 // just traversed was critical), then there are other paths through this 1546 // block along which the load may not be anticipated. Hoisting the load 1547 // above this block would be adding the load to execution paths along 1548 // which it was not previously executed. 1549 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1550 return false; 1551 } 1552 1553 assert(TmpBB); 1554 LoadBB = TmpBB; 1555 1556 // FIXME: It is extremely unclear what this loop is doing, other than 1557 // artificially restricting loadpre. 1558 if (isSinglePred) { 1559 bool isHot = false; 1560 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1561 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1562 if (AV.isSimpleValue()) 1563 // "Hot" Instruction is in some loop (because it dominates its dep. 1564 // instruction). 1565 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1566 if (DT->dominates(LI, I)) { 1567 isHot = true; 1568 break; 1569 } 1570 } 1571 1572 // We are interested only in "hot" instructions. We don't want to do any 1573 // mis-optimizations here. 1574 if (!isHot) 1575 return false; 1576 } 1577 1578 // Check to see how many predecessors have the loaded value fully 1579 // available. 1580 DenseMap<BasicBlock*, Value*> PredLoads; 1581 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1582 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1583 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1584 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1585 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1586 1587 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1588 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1589 PI != E; ++PI) { 1590 BasicBlock *Pred = *PI; 1591 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1592 continue; 1593 } 1594 PredLoads[Pred] = 0; 1595 1596 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1597 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1598 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1599 << Pred->getName() << "': " << *LI << '\n'); 1600 return false; 1601 } 1602 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1603 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1604 } 1605 } 1606 if (!NeedToSplit.empty()) { 1607 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1608 return false; 1609 } 1610 1611 // Decide whether PRE is profitable for this load. 1612 unsigned NumUnavailablePreds = PredLoads.size(); 1613 assert(NumUnavailablePreds != 0 && 1614 "Fully available value should be eliminated above!"); 1615 1616 // If this load is unavailable in multiple predecessors, reject it. 1617 // FIXME: If we could restructure the CFG, we could make a common pred with 1618 // all the preds that don't have an available LI and insert a new load into 1619 // that one block. 1620 if (NumUnavailablePreds != 1) 1621 return false; 1622 1623 // Check if the load can safely be moved to all the unavailable predecessors. 1624 bool CanDoPRE = true; 1625 SmallVector<Instruction*, 8> NewInsts; 1626 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1627 E = PredLoads.end(); I != E; ++I) { 1628 BasicBlock *UnavailablePred = I->first; 1629 1630 // Do PHI translation to get its value in the predecessor if necessary. The 1631 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1632 1633 // If all preds have a single successor, then we know it is safe to insert 1634 // the load on the pred (?!?), so we can insert code to materialize the 1635 // pointer if it is not available. 1636 PHITransAddr Address(LI->getPointerOperand(), TD); 1637 Value *LoadPtr = 0; 1638 if (allSingleSucc) { 1639 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1640 *DT, NewInsts); 1641 } else { 1642 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1643 LoadPtr = Address.getAddr(); 1644 } 1645 1646 // If we couldn't find or insert a computation of this phi translated value, 1647 // we fail PRE. 1648 if (LoadPtr == 0) { 1649 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1650 << *LI->getPointerOperand() << "\n"); 1651 CanDoPRE = false; 1652 break; 1653 } 1654 1655 // Make sure it is valid to move this load here. We have to watch out for: 1656 // @1 = getelementptr (i8* p, ... 1657 // test p and branch if == 0 1658 // load @1 1659 // It is valid to have the getelementptr before the test, even if p can be 0, 1660 // as getelementptr only does address arithmetic. 1661 // If we are not pushing the value through any multiple-successor blocks 1662 // we do not have this case. Otherwise, check that the load is safe to 1663 // put anywhere; this can be improved, but should be conservatively safe. 1664 if (!allSingleSucc && 1665 // FIXME: REEVALUTE THIS. 1666 !isSafeToLoadUnconditionally(LoadPtr, 1667 UnavailablePred->getTerminator(), 1668 LI->getAlignment(), TD)) { 1669 CanDoPRE = false; 1670 break; 1671 } 1672 1673 I->second = LoadPtr; 1674 } 1675 1676 if (!CanDoPRE) { 1677 while (!NewInsts.empty()) 1678 NewInsts.pop_back_val()->eraseFromParent(); 1679 return false; 1680 } 1681 1682 // Okay, we can eliminate this load by inserting a reload in the predecessor 1683 // and using PHI construction to get the value in the other predecessors, do 1684 // it. 1685 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1686 DEBUG(if (!NewInsts.empty()) 1687 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1688 << *NewInsts.back() << '\n'); 1689 1690 // Assign value numbers to the new instructions. 1691 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1692 // FIXME: We really _ought_ to insert these value numbers into their 1693 // parent's availability map. However, in doing so, we risk getting into 1694 // ordering issues. If a block hasn't been processed yet, we would be 1695 // marking a value as AVAIL-IN, which isn't what we intend. 1696 VN.lookup_or_add(NewInsts[i]); 1697 } 1698 1699 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1700 E = PredLoads.end(); I != E; ++I) { 1701 BasicBlock *UnavailablePred = I->first; 1702 Value *LoadPtr = I->second; 1703 1704 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1705 LI->getAlignment(), 1706 UnavailablePred->getTerminator()); 1707 1708 // Add the newly created load. 1709 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1710 NewLoad)); 1711 MD->invalidateCachedPointerInfo(LoadPtr); 1712 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1713 } 1714 1715 // Perform PHI construction. 1716 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1717 VN.getAliasAnalysis()); 1718 LI->replaceAllUsesWith(V); 1719 if (isa<PHINode>(V)) 1720 V->takeName(LI); 1721 if (V->getType()->isPointerTy()) 1722 MD->invalidateCachedPointerInfo(V); 1723 VN.erase(LI); 1724 toErase.push_back(LI); 1725 ++NumPRELoad; 1726 return true; 1727} 1728 1729/// processLoad - Attempt to eliminate a load, first by eliminating it 1730/// locally, and then attempting non-local elimination if that fails. 1731bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1732 if (!MD) 1733 return false; 1734 1735 if (L->isVolatile()) 1736 return false; 1737 1738 // ... to a pointer that has been loaded from before... 1739 MemDepResult Dep = MD->getDependency(L); 1740 1741 // If the value isn't available, don't do anything! 1742 if (Dep.isClobber()) { 1743 // Check to see if we have something like this: 1744 // store i32 123, i32* %P 1745 // %A = bitcast i32* %P to i8* 1746 // %B = gep i8* %A, i32 1 1747 // %C = load i8* %B 1748 // 1749 // We could do that by recognizing if the clobber instructions are obviously 1750 // a common base + constant offset, and if the previous store (or memset) 1751 // completely covers this load. This sort of thing can happen in bitfield 1752 // access code. 1753 Value *AvailVal = 0; 1754 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1755 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1756 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1757 L->getPointerOperand(), 1758 DepSI, *TD); 1759 if (Offset != -1) 1760 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1761 L->getType(), L, *TD); 1762 } 1763 1764 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1765 // a value on from it. 1766 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1767 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1768 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1769 L->getPointerOperand(), 1770 DepMI, *TD); 1771 if (Offset != -1) 1772 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD); 1773 } 1774 } 1775 1776 if (AvailVal) { 1777 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1778 << *AvailVal << '\n' << *L << "\n\n\n"); 1779 1780 // Replace the load! 1781 L->replaceAllUsesWith(AvailVal); 1782 if (AvailVal->getType()->isPointerTy()) 1783 MD->invalidateCachedPointerInfo(AvailVal); 1784 VN.erase(L); 1785 toErase.push_back(L); 1786 ++NumGVNLoad; 1787 return true; 1788 } 1789 1790 DEBUG( 1791 // fast print dep, using operator<< on instruction would be too slow 1792 dbgs() << "GVN: load "; 1793 WriteAsOperand(dbgs(), L); 1794 Instruction *I = Dep.getInst(); 1795 dbgs() << " is clobbered by " << *I << '\n'; 1796 ); 1797 return false; 1798 } 1799 1800 // If it is defined in another block, try harder. 1801 if (Dep.isNonLocal()) 1802 return processNonLocalLoad(L, toErase); 1803 1804 Instruction *DepInst = Dep.getInst(); 1805 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1806 Value *StoredVal = DepSI->getValueOperand(); 1807 1808 // The store and load are to a must-aliased pointer, but they may not 1809 // actually have the same type. See if we know how to reuse the stored 1810 // value (depending on its type). 1811 const TargetData *TD = 0; 1812 if (StoredVal->getType() != L->getType()) { 1813 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1814 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1815 L, *TD); 1816 if (StoredVal == 0) 1817 return false; 1818 1819 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1820 << '\n' << *L << "\n\n\n"); 1821 } 1822 else 1823 return false; 1824 } 1825 1826 // Remove it! 1827 L->replaceAllUsesWith(StoredVal); 1828 if (StoredVal->getType()->isPointerTy()) 1829 MD->invalidateCachedPointerInfo(StoredVal); 1830 VN.erase(L); 1831 toErase.push_back(L); 1832 ++NumGVNLoad; 1833 return true; 1834 } 1835 1836 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1837 Value *AvailableVal = DepLI; 1838 1839 // The loads are of a must-aliased pointer, but they may not actually have 1840 // the same type. See if we know how to reuse the previously loaded value 1841 // (depending on its type). 1842 const TargetData *TD = 0; 1843 if (DepLI->getType() != L->getType()) { 1844 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1845 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1846 if (AvailableVal == 0) 1847 return false; 1848 1849 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1850 << "\n" << *L << "\n\n\n"); 1851 } 1852 else 1853 return false; 1854 } 1855 1856 // Remove it! 1857 L->replaceAllUsesWith(AvailableVal); 1858 if (DepLI->getType()->isPointerTy()) 1859 MD->invalidateCachedPointerInfo(DepLI); 1860 VN.erase(L); 1861 toErase.push_back(L); 1862 ++NumGVNLoad; 1863 return true; 1864 } 1865 1866 // If this load really doesn't depend on anything, then we must be loading an 1867 // undef value. This can happen when loading for a fresh allocation with no 1868 // intervening stores, for example. 1869 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1870 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1871 VN.erase(L); 1872 toErase.push_back(L); 1873 ++NumGVNLoad; 1874 return true; 1875 } 1876 1877 // If this load occurs either right after a lifetime begin, 1878 // then the loaded value is undefined. 1879 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) { 1880 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1881 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1882 VN.erase(L); 1883 toErase.push_back(L); 1884 ++NumGVNLoad; 1885 return true; 1886 } 1887 } 1888 1889 return false; 1890} 1891 1892Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1893 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1894 if (I == localAvail.end()) 1895 return 0; 1896 1897 ValueNumberScope *Locals = I->second; 1898 while (Locals) { 1899 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1900 if (I != Locals->table.end()) 1901 return I->second; 1902 Locals = Locals->parent; 1903 } 1904 1905 return 0; 1906} 1907 1908 1909/// processInstruction - When calculating availability, handle an instruction 1910/// by inserting it into the appropriate sets 1911bool GVN::processInstruction(Instruction *I, 1912 SmallVectorImpl<Instruction*> &toErase) { 1913 // Ignore dbg info intrinsics. 1914 if (isa<DbgInfoIntrinsic>(I)) 1915 return false; 1916 1917 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1918 bool Changed = processLoad(LI, toErase); 1919 1920 if (!Changed) { 1921 unsigned Num = VN.lookup_or_add(LI); 1922 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1923 } 1924 1925 return Changed; 1926 } 1927 1928 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1929 unsigned Num = VN.lookup_or_add(I); 1930 1931 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1932 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1933 1934 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1935 return false; 1936 1937 Value *BranchCond = BI->getCondition(); 1938 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1939 1940 BasicBlock *TrueSucc = BI->getSuccessor(0); 1941 BasicBlock *FalseSucc = BI->getSuccessor(1); 1942 1943 if (TrueSucc->getSinglePredecessor()) 1944 localAvail[TrueSucc]->table[CondVN] = 1945 ConstantInt::getTrue(TrueSucc->getContext()); 1946 if (FalseSucc->getSinglePredecessor()) 1947 localAvail[FalseSucc]->table[CondVN] = 1948 ConstantInt::getFalse(TrueSucc->getContext()); 1949 1950 return false; 1951 1952 // Allocations are always uniquely numbered, so we can save time and memory 1953 // by fast failing them. 1954 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) { 1955 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1956 return false; 1957 } 1958 1959 // Collapse PHI nodes 1960 if (PHINode* p = dyn_cast<PHINode>(I)) { 1961 Value *constVal = CollapsePhi(p); 1962 1963 if (constVal) { 1964 p->replaceAllUsesWith(constVal); 1965 if (MD && constVal->getType()->isPointerTy()) 1966 MD->invalidateCachedPointerInfo(constVal); 1967 VN.erase(p); 1968 1969 toErase.push_back(p); 1970 } else { 1971 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1972 } 1973 1974 // If the number we were assigned was a brand new VN, then we don't 1975 // need to do a lookup to see if the number already exists 1976 // somewhere in the domtree: it can't! 1977 } else if (Num == NextNum) { 1978 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1979 1980 // Perform fast-path value-number based elimination of values inherited from 1981 // dominators. 1982 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1983 // Remove it! 1984 VN.erase(I); 1985 I->replaceAllUsesWith(repl); 1986 if (MD && repl->getType()->isPointerTy()) 1987 MD->invalidateCachedPointerInfo(repl); 1988 toErase.push_back(I); 1989 return true; 1990 1991 } else { 1992 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1993 } 1994 1995 return false; 1996} 1997 1998/// runOnFunction - This is the main transformation entry point for a function. 1999bool GVN::runOnFunction(Function& F) { 2000 if (!NoLoads) 2001 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2002 DT = &getAnalysis<DominatorTree>(); 2003 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2004 VN.setMemDep(MD); 2005 VN.setDomTree(DT); 2006 2007 bool Changed = false; 2008 bool ShouldContinue = true; 2009 2010 // Merge unconditional branches, allowing PRE to catch more 2011 // optimization opportunities. 2012 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2013 BasicBlock *BB = FI; 2014 ++FI; 2015 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2016 if (removedBlock) ++NumGVNBlocks; 2017 2018 Changed |= removedBlock; 2019 } 2020 2021 unsigned Iteration = 0; 2022 2023 while (ShouldContinue) { 2024 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2025 ShouldContinue = iterateOnFunction(F); 2026 if (splitCriticalEdges()) 2027 ShouldContinue = true; 2028 Changed |= ShouldContinue; 2029 ++Iteration; 2030 } 2031 2032 if (EnablePRE) { 2033 bool PREChanged = true; 2034 while (PREChanged) { 2035 PREChanged = performPRE(F); 2036 Changed |= PREChanged; 2037 } 2038 } 2039 // FIXME: Should perform GVN again after PRE does something. PRE can move 2040 // computations into blocks where they become fully redundant. Note that 2041 // we can't do this until PRE's critical edge splitting updates memdep. 2042 // Actually, when this happens, we should just fully integrate PRE into GVN. 2043 2044 cleanupGlobalSets(); 2045 2046 return Changed; 2047} 2048 2049 2050bool GVN::processBlock(BasicBlock *BB) { 2051 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 2052 // incrementing BI before processing an instruction). 2053 SmallVector<Instruction*, 8> toErase; 2054 bool ChangedFunction = false; 2055 2056 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2057 BI != BE;) { 2058 ChangedFunction |= processInstruction(BI, toErase); 2059 if (toErase.empty()) { 2060 ++BI; 2061 continue; 2062 } 2063 2064 // If we need some instructions deleted, do it now. 2065 NumGVNInstr += toErase.size(); 2066 2067 // Avoid iterator invalidation. 2068 bool AtStart = BI == BB->begin(); 2069 if (!AtStart) 2070 --BI; 2071 2072 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 2073 E = toErase.end(); I != E; ++I) { 2074 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2075 if (MD) MD->removeInstruction(*I); 2076 (*I)->eraseFromParent(); 2077 DEBUG(verifyRemoved(*I)); 2078 } 2079 toErase.clear(); 2080 2081 if (AtStart) 2082 BI = BB->begin(); 2083 else 2084 ++BI; 2085 } 2086 2087 return ChangedFunction; 2088} 2089 2090/// performPRE - Perform a purely local form of PRE that looks for diamond 2091/// control flow patterns and attempts to perform simple PRE at the join point. 2092bool GVN::performPRE(Function &F) { 2093 bool Changed = false; 2094 DenseMap<BasicBlock*, Value*> predMap; 2095 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2096 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2097 BasicBlock *CurrentBlock = *DI; 2098 2099 // Nothing to PRE in the entry block. 2100 if (CurrentBlock == &F.getEntryBlock()) continue; 2101 2102 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2103 BE = CurrentBlock->end(); BI != BE; ) { 2104 Instruction *CurInst = BI++; 2105 2106 if (isa<AllocaInst>(CurInst) || 2107 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2108 CurInst->getType()->isVoidTy() || 2109 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2110 isa<DbgInfoIntrinsic>(CurInst)) 2111 continue; 2112 2113 // We don't currently value number ANY inline asm calls. 2114 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2115 if (CallI->isInlineAsm()) 2116 continue; 2117 2118 uint32_t ValNo = VN.lookup(CurInst); 2119 2120 // Look for the predecessors for PRE opportunities. We're 2121 // only trying to solve the basic diamond case, where 2122 // a value is computed in the successor and one predecessor, 2123 // but not the other. We also explicitly disallow cases 2124 // where the successor is its own predecessor, because they're 2125 // more complicated to get right. 2126 unsigned NumWith = 0; 2127 unsigned NumWithout = 0; 2128 BasicBlock *PREPred = 0; 2129 predMap.clear(); 2130 2131 for (pred_iterator PI = pred_begin(CurrentBlock), 2132 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2133 BasicBlock *P = *PI; 2134 // We're not interested in PRE where the block is its 2135 // own predecessor, or in blocks with predecessors 2136 // that are not reachable. 2137 if (P == CurrentBlock) { 2138 NumWithout = 2; 2139 break; 2140 } else if (!localAvail.count(P)) { 2141 NumWithout = 2; 2142 break; 2143 } 2144 2145 DenseMap<uint32_t, Value*>::iterator predV = 2146 localAvail[P]->table.find(ValNo); 2147 if (predV == localAvail[P]->table.end()) { 2148 PREPred = P; 2149 ++NumWithout; 2150 } else if (predV->second == CurInst) { 2151 NumWithout = 2; 2152 } else { 2153 predMap[P] = predV->second; 2154 ++NumWith; 2155 } 2156 } 2157 2158 // Don't do PRE when it might increase code size, i.e. when 2159 // we would need to insert instructions in more than one pred. 2160 if (NumWithout != 1 || NumWith == 0) 2161 continue; 2162 2163 // Don't do PRE across indirect branch. 2164 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2165 continue; 2166 2167 // We can't do PRE safely on a critical edge, so instead we schedule 2168 // the edge to be split and perform the PRE the next time we iterate 2169 // on the function. 2170 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2171 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2172 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2173 continue; 2174 } 2175 2176 // Instantiate the expression in the predecessor that lacked it. 2177 // Because we are going top-down through the block, all value numbers 2178 // will be available in the predecessor by the time we need them. Any 2179 // that weren't originally present will have been instantiated earlier 2180 // in this loop. 2181 Instruction *PREInstr = CurInst->clone(); 2182 bool success = true; 2183 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2184 Value *Op = PREInstr->getOperand(i); 2185 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2186 continue; 2187 2188 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 2189 PREInstr->setOperand(i, V); 2190 } else { 2191 success = false; 2192 break; 2193 } 2194 } 2195 2196 // Fail out if we encounter an operand that is not available in 2197 // the PRE predecessor. This is typically because of loads which 2198 // are not value numbered precisely. 2199 if (!success) { 2200 delete PREInstr; 2201 DEBUG(verifyRemoved(PREInstr)); 2202 continue; 2203 } 2204 2205 PREInstr->insertBefore(PREPred->getTerminator()); 2206 PREInstr->setName(CurInst->getName() + ".pre"); 2207 predMap[PREPred] = PREInstr; 2208 VN.add(PREInstr, ValNo); 2209 ++NumGVNPRE; 2210 2211 // Update the availability map to include the new instruction. 2212 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 2213 2214 // Create a PHI to make the value available in this block. 2215 PHINode* Phi = PHINode::Create(CurInst->getType(), 2216 CurInst->getName() + ".pre-phi", 2217 CurrentBlock->begin()); 2218 for (pred_iterator PI = pred_begin(CurrentBlock), 2219 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2220 BasicBlock *P = *PI; 2221 Phi->addIncoming(predMap[P], P); 2222 } 2223 2224 VN.add(Phi, ValNo); 2225 localAvail[CurrentBlock]->table[ValNo] = Phi; 2226 2227 CurInst->replaceAllUsesWith(Phi); 2228 if (MD && Phi->getType()->isPointerTy()) 2229 MD->invalidateCachedPointerInfo(Phi); 2230 VN.erase(CurInst); 2231 2232 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2233 if (MD) MD->removeInstruction(CurInst); 2234 CurInst->eraseFromParent(); 2235 DEBUG(verifyRemoved(CurInst)); 2236 Changed = true; 2237 } 2238 } 2239 2240 if (splitCriticalEdges()) 2241 Changed = true; 2242 2243 return Changed; 2244} 2245 2246/// splitCriticalEdges - Split critical edges found during the previous 2247/// iteration that may enable further optimization. 2248bool GVN::splitCriticalEdges() { 2249 if (toSplit.empty()) 2250 return false; 2251 do { 2252 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2253 SplitCriticalEdge(Edge.first, Edge.second, this); 2254 } while (!toSplit.empty()); 2255 if (MD) MD->invalidateCachedPredecessors(); 2256 return true; 2257} 2258 2259/// iterateOnFunction - Executes one iteration of GVN 2260bool GVN::iterateOnFunction(Function &F) { 2261 cleanupGlobalSets(); 2262 2263 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2264 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 2265 if (DI->getIDom()) 2266 localAvail[DI->getBlock()] = 2267 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 2268 else 2269 localAvail[DI->getBlock()] = new ValueNumberScope(0); 2270 } 2271 2272 // Top-down walk of the dominator tree 2273 bool Changed = false; 2274#if 0 2275 // Needed for value numbering with phi construction to work. 2276 ReversePostOrderTraversal<Function*> RPOT(&F); 2277 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2278 RE = RPOT.end(); RI != RE; ++RI) 2279 Changed |= processBlock(*RI); 2280#else 2281 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2282 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2283 Changed |= processBlock(DI->getBlock()); 2284#endif 2285 2286 return Changed; 2287} 2288 2289void GVN::cleanupGlobalSets() { 2290 VN.clear(); 2291 2292 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2293 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 2294 delete I->second; 2295 localAvail.clear(); 2296} 2297 2298/// verifyRemoved - Verify that the specified instruction does not occur in our 2299/// internal data structures. 2300void GVN::verifyRemoved(const Instruction *Inst) const { 2301 VN.verifyRemoved(Inst); 2302 2303 // Walk through the value number scope to make sure the instruction isn't 2304 // ferreted away in it. 2305 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator 2306 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 2307 const ValueNumberScope *VNS = I->second; 2308 2309 while (VNS) { 2310 for (DenseMap<uint32_t, Value*>::const_iterator 2311 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 2312 assert(II->second != Inst && "Inst still in value numbering scope!"); 2313 } 2314 2315 VNS = VNS->parent; 2316 } 2317 } 2318} 2319