GVN.cpp revision 2ab36d350293c77fc8941ce1023e4899df7e3a82
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) { } 666 667 private: 668 bool NoLoads; 669 MemoryDependenceAnalysis *MD; 670 DominatorTree *DT; 671 672 ValueTable VN; 673 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 674 675 // List of critical edges to be split between iterations. 676 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 677 678 // This transformation requires dominator postdominator info 679 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 680 AU.addRequired<DominatorTree>(); 681 if (!NoLoads) 682 AU.addRequired<MemoryDependenceAnalysis>(); 683 AU.addRequired<AliasAnalysis>(); 684 685 AU.addPreserved<DominatorTree>(); 686 AU.addPreserved<AliasAnalysis>(); 687 } 688 689 // Helper fuctions 690 // FIXME: eliminate or document these better 691 bool processLoad(LoadInst* L, 692 SmallVectorImpl<Instruction*> &toErase); 693 bool processInstruction(Instruction *I, 694 SmallVectorImpl<Instruction*> &toErase); 695 bool processNonLocalLoad(LoadInst* L, 696 SmallVectorImpl<Instruction*> &toErase); 697 bool processBlock(BasicBlock *BB); 698 void dump(DenseMap<uint32_t, Value*>& d); 699 bool iterateOnFunction(Function &F); 700 Value *CollapsePhi(PHINode* p); 701 bool performPRE(Function& F); 702 Value *lookupNumber(BasicBlock *BB, uint32_t num); 703 void cleanupGlobalSets(); 704 void verifyRemoved(const Instruction *I) const; 705 bool splitCriticalEdges(); 706 }; 707 708 char GVN::ID = 0; 709} 710 711// createGVNPass - The public interface to this file... 712FunctionPass *llvm::createGVNPass(bool NoLoads) { 713 return new GVN(NoLoads); 714} 715 716INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 717INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 718INITIALIZE_PASS_DEPENDENCY(DominatorTree) 719INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 720INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 721 722void GVN::dump(DenseMap<uint32_t, Value*>& d) { 723 errs() << "{\n"; 724 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 725 E = d.end(); I != E; ++I) { 726 errs() << I->first << "\n"; 727 I->second->dump(); 728 } 729 errs() << "}\n"; 730} 731 732static bool isSafeReplacement(PHINode* p, Instruction *inst) { 733 if (!isa<PHINode>(inst)) 734 return true; 735 736 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 737 UI != E; ++UI) 738 if (PHINode* use_phi = dyn_cast<PHINode>(*UI)) 739 if (use_phi->getParent() == inst->getParent()) 740 return false; 741 742 return true; 743} 744 745Value *GVN::CollapsePhi(PHINode *PN) { 746 Value *ConstVal = PN->hasConstantValue(DT); 747 if (!ConstVal) return 0; 748 749 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 750 if (!Inst) 751 return ConstVal; 752 753 if (DT->dominates(Inst, PN)) 754 if (isSafeReplacement(PN, Inst)) 755 return Inst; 756 return 0; 757} 758 759/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 760/// we're analyzing is fully available in the specified block. As we go, keep 761/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 762/// map is actually a tri-state map with the following values: 763/// 0) we know the block *is not* fully available. 764/// 1) we know the block *is* fully available. 765/// 2) we do not know whether the block is fully available or not, but we are 766/// currently speculating that it will be. 767/// 3) we are speculating for this block and have used that to speculate for 768/// other blocks. 769static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 770 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 771 // Optimistically assume that the block is fully available and check to see 772 // if we already know about this block in one lookup. 773 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 774 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 775 776 // If the entry already existed for this block, return the precomputed value. 777 if (!IV.second) { 778 // If this is a speculative "available" value, mark it as being used for 779 // speculation of other blocks. 780 if (IV.first->second == 2) 781 IV.first->second = 3; 782 return IV.first->second != 0; 783 } 784 785 // Otherwise, see if it is fully available in all predecessors. 786 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 787 788 // If this block has no predecessors, it isn't live-in here. 789 if (PI == PE) 790 goto SpeculationFailure; 791 792 for (; PI != PE; ++PI) 793 // If the value isn't fully available in one of our predecessors, then it 794 // isn't fully available in this block either. Undo our previous 795 // optimistic assumption and bail out. 796 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 797 goto SpeculationFailure; 798 799 return true; 800 801// SpeculationFailure - If we get here, we found out that this is not, after 802// all, a fully-available block. We have a problem if we speculated on this and 803// used the speculation to mark other blocks as available. 804SpeculationFailure: 805 char &BBVal = FullyAvailableBlocks[BB]; 806 807 // If we didn't speculate on this, just return with it set to false. 808 if (BBVal == 2) { 809 BBVal = 0; 810 return false; 811 } 812 813 // If we did speculate on this value, we could have blocks set to 1 that are 814 // incorrect. Walk the (transitive) successors of this block and mark them as 815 // 0 if set to one. 816 SmallVector<BasicBlock*, 32> BBWorklist; 817 BBWorklist.push_back(BB); 818 819 do { 820 BasicBlock *Entry = BBWorklist.pop_back_val(); 821 // Note that this sets blocks to 0 (unavailable) if they happen to not 822 // already be in FullyAvailableBlocks. This is safe. 823 char &EntryVal = FullyAvailableBlocks[Entry]; 824 if (EntryVal == 0) continue; // Already unavailable. 825 826 // Mark as unavailable. 827 EntryVal = 0; 828 829 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 830 BBWorklist.push_back(*I); 831 } while (!BBWorklist.empty()); 832 833 return false; 834} 835 836 837/// CanCoerceMustAliasedValueToLoad - Return true if 838/// CoerceAvailableValueToLoadType will succeed. 839static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 840 const Type *LoadTy, 841 const TargetData &TD) { 842 // If the loaded or stored value is an first class array or struct, don't try 843 // to transform them. We need to be able to bitcast to integer. 844 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 845 StoredVal->getType()->isStructTy() || 846 StoredVal->getType()->isArrayTy()) 847 return false; 848 849 // The store has to be at least as big as the load. 850 if (TD.getTypeSizeInBits(StoredVal->getType()) < 851 TD.getTypeSizeInBits(LoadTy)) 852 return false; 853 854 return true; 855} 856 857 858/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 859/// then a load from a must-aliased pointer of a different type, try to coerce 860/// the stored value. LoadedTy is the type of the load we want to replace and 861/// InsertPt is the place to insert new instructions. 862/// 863/// If we can't do it, return null. 864static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 865 const Type *LoadedTy, 866 Instruction *InsertPt, 867 const TargetData &TD) { 868 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 869 return 0; 870 871 const Type *StoredValTy = StoredVal->getType(); 872 873 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 874 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 875 876 // If the store and reload are the same size, we can always reuse it. 877 if (StoreSize == LoadSize) { 878 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) { 879 // Pointer to Pointer -> use bitcast. 880 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 881 } 882 883 // Convert source pointers to integers, which can be bitcast. 884 if (StoredValTy->isPointerTy()) { 885 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 886 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 887 } 888 889 const Type *TypeToCastTo = LoadedTy; 890 if (TypeToCastTo->isPointerTy()) 891 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 892 893 if (StoredValTy != TypeToCastTo) 894 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 895 896 // Cast to pointer if the load needs a pointer type. 897 if (LoadedTy->isPointerTy()) 898 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 899 900 return StoredVal; 901 } 902 903 // If the loaded value is smaller than the available value, then we can 904 // extract out a piece from it. If the available value is too small, then we 905 // can't do anything. 906 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 907 908 // Convert source pointers to integers, which can be manipulated. 909 if (StoredValTy->isPointerTy()) { 910 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 911 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 912 } 913 914 // Convert vectors and fp to integer, which can be manipulated. 915 if (!StoredValTy->isIntegerTy()) { 916 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 917 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 918 } 919 920 // If this is a big-endian system, we need to shift the value down to the low 921 // bits so that a truncate will work. 922 if (TD.isBigEndian()) { 923 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 924 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 925 } 926 927 // Truncate the integer to the right size now. 928 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 929 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 930 931 if (LoadedTy == NewIntTy) 932 return StoredVal; 933 934 // If the result is a pointer, inttoptr. 935 if (LoadedTy->isPointerTy()) 936 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 937 938 // Otherwise, bitcast. 939 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 940} 941 942/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 943/// be expressed as a base pointer plus a constant offset. Return the base and 944/// offset to the caller. 945static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 946 const TargetData &TD) { 947 Operator *PtrOp = dyn_cast<Operator>(Ptr); 948 if (PtrOp == 0) return Ptr; 949 950 // Just look through bitcasts. 951 if (PtrOp->getOpcode() == Instruction::BitCast) 952 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 953 954 // If this is a GEP with constant indices, we can look through it. 955 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 956 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 957 958 gep_type_iterator GTI = gep_type_begin(GEP); 959 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 960 ++I, ++GTI) { 961 ConstantInt *OpC = cast<ConstantInt>(*I); 962 if (OpC->isZero()) continue; 963 964 // Handle a struct and array indices which add their offset to the pointer. 965 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 966 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 967 } else { 968 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 969 Offset += OpC->getSExtValue()*Size; 970 } 971 } 972 973 // Re-sign extend from the pointer size if needed to get overflow edge cases 974 // right. 975 unsigned PtrSize = TD.getPointerSizeInBits(); 976 if (PtrSize < 64) 977 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 978 979 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 980} 981 982 983/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 984/// memdep query of a load that ends up being a clobbering memory write (store, 985/// memset, memcpy, memmove). This means that the write *may* provide bits used 986/// by the load but we can't be sure because the pointers don't mustalias. 987/// 988/// Check this case to see if there is anything more we can do before we give 989/// up. This returns -1 if we have to give up, or a byte number in the stored 990/// value of the piece that feeds the load. 991static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 992 Value *WritePtr, 993 uint64_t WriteSizeInBits, 994 const TargetData &TD) { 995 // If the loaded or stored value is an first class array or struct, don't try 996 // to transform them. We need to be able to bitcast to integer. 997 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 998 return -1; 999 1000 int64_t StoreOffset = 0, LoadOffset = 0; 1001 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD); 1002 Value *LoadBase = 1003 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 1004 if (StoreBase != LoadBase) 1005 return -1; 1006 1007 // If the load and store are to the exact same address, they should have been 1008 // a must alias. AA must have gotten confused. 1009 // FIXME: Study to see if/when this happens. One case is forwarding a memset 1010 // to a load from the base of the memset. 1011#if 0 1012 if (LoadOffset == StoreOffset) { 1013 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1014 << "Base = " << *StoreBase << "\n" 1015 << "Store Ptr = " << *WritePtr << "\n" 1016 << "Store Offs = " << StoreOffset << "\n" 1017 << "Load Ptr = " << *LoadPtr << "\n"; 1018 abort(); 1019 } 1020#endif 1021 1022 // If the load and store don't overlap at all, the store doesn't provide 1023 // anything to the load. In this case, they really don't alias at all, AA 1024 // must have gotten confused. 1025 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 1026 // remove this check, as it is duplicated with what we have below. 1027 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 1028 1029 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 1030 return -1; 1031 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 1032 LoadSize >>= 3; 1033 1034 1035 bool isAAFailure = false; 1036 if (StoreOffset < LoadOffset) 1037 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1038 else 1039 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1040 1041 if (isAAFailure) { 1042#if 0 1043 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1044 << "Base = " << *StoreBase << "\n" 1045 << "Store Ptr = " << *WritePtr << "\n" 1046 << "Store Offs = " << StoreOffset << "\n" 1047 << "Load Ptr = " << *LoadPtr << "\n"; 1048 abort(); 1049#endif 1050 return -1; 1051 } 1052 1053 // If the Load isn't completely contained within the stored bits, we don't 1054 // have all the bits to feed it. We could do something crazy in the future 1055 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1056 // valuable. 1057 if (StoreOffset > LoadOffset || 1058 StoreOffset+StoreSize < LoadOffset+LoadSize) 1059 return -1; 1060 1061 // Okay, we can do this transformation. Return the number of bytes into the 1062 // store that the load is. 1063 return LoadOffset-StoreOffset; 1064} 1065 1066/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1067/// memdep query of a load that ends up being a clobbering store. 1068static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 1069 StoreInst *DepSI, 1070 const TargetData &TD) { 1071 // Cannot handle reading from store of first-class aggregate yet. 1072 if (DepSI->getOperand(0)->getType()->isStructTy() || 1073 DepSI->getOperand(0)->getType()->isArrayTy()) 1074 return -1; 1075 1076 Value *StorePtr = DepSI->getPointerOperand(); 1077 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType()); 1078 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1079 StorePtr, StoreSize, TD); 1080} 1081 1082static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 1083 MemIntrinsic *MI, 1084 const TargetData &TD) { 1085 // If the mem operation is a non-constant size, we can't handle it. 1086 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1087 if (SizeCst == 0) return -1; 1088 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1089 1090 // If this is memset, we just need to see if the offset is valid in the size 1091 // of the memset.. 1092 if (MI->getIntrinsicID() == Intrinsic::memset) 1093 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1094 MemSizeInBits, TD); 1095 1096 // If we have a memcpy/memmove, the only case we can handle is if this is a 1097 // copy from constant memory. In that case, we can read directly from the 1098 // constant memory. 1099 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1100 1101 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1102 if (Src == 0) return -1; 1103 1104 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject()); 1105 if (GV == 0 || !GV->isConstant()) return -1; 1106 1107 // See if the access is within the bounds of the transfer. 1108 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1109 MI->getDest(), MemSizeInBits, TD); 1110 if (Offset == -1) 1111 return Offset; 1112 1113 // Otherwise, see if we can constant fold a load from the constant with the 1114 // offset applied as appropriate. 1115 Src = ConstantExpr::getBitCast(Src, 1116 llvm::Type::getInt8PtrTy(Src->getContext())); 1117 Constant *OffsetCst = 1118 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1119 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1120 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1121 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1122 return Offset; 1123 return -1; 1124} 1125 1126 1127/// GetStoreValueForLoad - This function is called when we have a 1128/// memdep query of a load that ends up being a clobbering store. This means 1129/// that the store *may* provide bits used by the load but we can't be sure 1130/// because the pointers don't mustalias. Check this case to see if there is 1131/// anything more we can do before we give up. 1132static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1133 const Type *LoadTy, 1134 Instruction *InsertPt, const TargetData &TD){ 1135 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1136 1137 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1138 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1139 1140 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1141 1142 // Compute which bits of the stored value are being used by the load. Convert 1143 // to an integer type to start with. 1144 if (SrcVal->getType()->isPointerTy()) 1145 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 1146 if (!SrcVal->getType()->isIntegerTy()) 1147 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1148 "tmp"); 1149 1150 // Shift the bits to the least significant depending on endianness. 1151 unsigned ShiftAmt; 1152 if (TD.isLittleEndian()) 1153 ShiftAmt = Offset*8; 1154 else 1155 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1156 1157 if (ShiftAmt) 1158 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 1159 1160 if (LoadSize != StoreSize) 1161 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1162 "tmp"); 1163 1164 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1165} 1166 1167/// GetMemInstValueForLoad - This function is called when we have a 1168/// memdep query of a load that ends up being a clobbering mem intrinsic. 1169static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1170 const Type *LoadTy, Instruction *InsertPt, 1171 const TargetData &TD){ 1172 LLVMContext &Ctx = LoadTy->getContext(); 1173 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1174 1175 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1176 1177 // We know that this method is only called when the mem transfer fully 1178 // provides the bits for the load. 1179 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1180 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1181 // independently of what the offset is. 1182 Value *Val = MSI->getValue(); 1183 if (LoadSize != 1) 1184 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1185 1186 Value *OneElt = Val; 1187 1188 // Splat the value out to the right number of bits. 1189 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1190 // If we can double the number of bytes set, do it. 1191 if (NumBytesSet*2 <= LoadSize) { 1192 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1193 Val = Builder.CreateOr(Val, ShVal); 1194 NumBytesSet <<= 1; 1195 continue; 1196 } 1197 1198 // Otherwise insert one byte at a time. 1199 Value *ShVal = Builder.CreateShl(Val, 1*8); 1200 Val = Builder.CreateOr(OneElt, ShVal); 1201 ++NumBytesSet; 1202 } 1203 1204 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1205 } 1206 1207 // Otherwise, this is a memcpy/memmove from a constant global. 1208 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1209 Constant *Src = cast<Constant>(MTI->getSource()); 1210 1211 // Otherwise, see if we can constant fold a load from the constant with the 1212 // offset applied as appropriate. 1213 Src = ConstantExpr::getBitCast(Src, 1214 llvm::Type::getInt8PtrTy(Src->getContext())); 1215 Constant *OffsetCst = 1216 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1217 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1218 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1219 return ConstantFoldLoadFromConstPtr(Src, &TD); 1220} 1221 1222namespace { 1223 1224struct AvailableValueInBlock { 1225 /// BB - The basic block in question. 1226 BasicBlock *BB; 1227 enum ValType { 1228 SimpleVal, // A simple offsetted value that is accessed. 1229 MemIntrin // A memory intrinsic which is loaded from. 1230 }; 1231 1232 /// V - The value that is live out of the block. 1233 PointerIntPair<Value *, 1, ValType> Val; 1234 1235 /// Offset - The byte offset in Val that is interesting for the load query. 1236 unsigned Offset; 1237 1238 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1239 unsigned Offset = 0) { 1240 AvailableValueInBlock Res; 1241 Res.BB = BB; 1242 Res.Val.setPointer(V); 1243 Res.Val.setInt(SimpleVal); 1244 Res.Offset = Offset; 1245 return Res; 1246 } 1247 1248 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1249 unsigned Offset = 0) { 1250 AvailableValueInBlock Res; 1251 Res.BB = BB; 1252 Res.Val.setPointer(MI); 1253 Res.Val.setInt(MemIntrin); 1254 Res.Offset = Offset; 1255 return Res; 1256 } 1257 1258 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1259 Value *getSimpleValue() const { 1260 assert(isSimpleValue() && "Wrong accessor"); 1261 return Val.getPointer(); 1262 } 1263 1264 MemIntrinsic *getMemIntrinValue() const { 1265 assert(!isSimpleValue() && "Wrong accessor"); 1266 return cast<MemIntrinsic>(Val.getPointer()); 1267 } 1268 1269 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1270 /// defined here to the specified type. This handles various coercion cases. 1271 Value *MaterializeAdjustedValue(const Type *LoadTy, 1272 const TargetData *TD) const { 1273 Value *Res; 1274 if (isSimpleValue()) { 1275 Res = getSimpleValue(); 1276 if (Res->getType() != LoadTy) { 1277 assert(TD && "Need target data to handle type mismatch case"); 1278 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1279 *TD); 1280 1281 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1282 << *getSimpleValue() << '\n' 1283 << *Res << '\n' << "\n\n\n"); 1284 } 1285 } else { 1286 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1287 LoadTy, BB->getTerminator(), *TD); 1288 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1289 << " " << *getMemIntrinValue() << '\n' 1290 << *Res << '\n' << "\n\n\n"); 1291 } 1292 return Res; 1293 } 1294}; 1295 1296} 1297 1298/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1299/// construct SSA form, allowing us to eliminate LI. This returns the value 1300/// that should be used at LI's definition site. 1301static Value *ConstructSSAForLoadSet(LoadInst *LI, 1302 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1303 const TargetData *TD, 1304 const DominatorTree &DT, 1305 AliasAnalysis *AA) { 1306 // Check for the fully redundant, dominating load case. In this case, we can 1307 // just use the dominating value directly. 1308 if (ValuesPerBlock.size() == 1 && 1309 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent())) 1310 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD); 1311 1312 // Otherwise, we have to construct SSA form. 1313 SmallVector<PHINode*, 8> NewPHIs; 1314 SSAUpdater SSAUpdate(&NewPHIs); 1315 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1316 1317 const Type *LoadTy = LI->getType(); 1318 1319 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1320 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1321 BasicBlock *BB = AV.BB; 1322 1323 if (SSAUpdate.HasValueForBlock(BB)) 1324 continue; 1325 1326 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD)); 1327 } 1328 1329 // Perform PHI construction. 1330 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1331 1332 // If new PHI nodes were created, notify alias analysis. 1333 if (V->getType()->isPointerTy()) 1334 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1335 AA->copyValue(LI, NewPHIs[i]); 1336 1337 return V; 1338} 1339 1340static bool isLifetimeStart(const Instruction *Inst) { 1341 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1342 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1343 return false; 1344} 1345 1346/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1347/// non-local by performing PHI construction. 1348bool GVN::processNonLocalLoad(LoadInst *LI, 1349 SmallVectorImpl<Instruction*> &toErase) { 1350 // Find the non-local dependencies of the load. 1351 SmallVector<NonLocalDepResult, 64> Deps; 1352 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1353 Deps); 1354 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1355 // << Deps.size() << *LI << '\n'); 1356 1357 // If we had to process more than one hundred blocks to find the 1358 // dependencies, this load isn't worth worrying about. Optimizing 1359 // it will be too expensive. 1360 if (Deps.size() > 100) 1361 return false; 1362 1363 // If we had a phi translation failure, we'll have a single entry which is a 1364 // clobber in the current block. Reject this early. 1365 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) { 1366 DEBUG( 1367 dbgs() << "GVN: non-local load "; 1368 WriteAsOperand(dbgs(), LI); 1369 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1370 ); 1371 return false; 1372 } 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 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1379 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1380 1381 const TargetData *TD = 0; 1382 1383 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1384 BasicBlock *DepBB = Deps[i].getBB(); 1385 MemDepResult DepInfo = Deps[i].getResult(); 1386 1387 if (DepInfo.isClobber()) { 1388 // The address being loaded in this non-local block may not be the same as 1389 // the pointer operand of the load if PHI translation occurs. Make sure 1390 // to consider the right address. 1391 Value *Address = Deps[i].getAddress(); 1392 1393 // If the dependence is to a store that writes to a superset of the bits 1394 // read by the load, we can extract the bits we need for the load from the 1395 // stored value. 1396 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1397 if (TD == 0) 1398 TD = getAnalysisIfAvailable<TargetData>(); 1399 if (TD && Address) { 1400 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1401 DepSI, *TD); 1402 if (Offset != -1) { 1403 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1404 DepSI->getOperand(0), 1405 Offset)); 1406 continue; 1407 } 1408 } 1409 } 1410 1411 // If the clobbering value is a memset/memcpy/memmove, see if we can 1412 // forward a value on from it. 1413 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1414 if (TD == 0) 1415 TD = getAnalysisIfAvailable<TargetData>(); 1416 if (TD && Address) { 1417 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1418 DepMI, *TD); 1419 if (Offset != -1) { 1420 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1421 Offset)); 1422 continue; 1423 } 1424 } 1425 } 1426 1427 UnavailableBlocks.push_back(DepBB); 1428 continue; 1429 } 1430 1431 Instruction *DepInst = DepInfo.getInst(); 1432 1433 // Loading the allocation -> undef. 1434 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1435 // Loading immediately after lifetime begin -> undef. 1436 isLifetimeStart(DepInst)) { 1437 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1438 UndefValue::get(LI->getType()))); 1439 continue; 1440 } 1441 1442 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1443 // Reject loads and stores that are to the same address but are of 1444 // different types if we have to. 1445 if (S->getOperand(0)->getType() != LI->getType()) { 1446 if (TD == 0) 1447 TD = getAnalysisIfAvailable<TargetData>(); 1448 1449 // If the stored value is larger or equal to the loaded value, we can 1450 // reuse it. 1451 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0), 1452 LI->getType(), *TD)) { 1453 UnavailableBlocks.push_back(DepBB); 1454 continue; 1455 } 1456 } 1457 1458 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1459 S->getOperand(0))); 1460 continue; 1461 } 1462 1463 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1464 // If the types mismatch and we can't handle it, reject reuse of the load. 1465 if (LD->getType() != LI->getType()) { 1466 if (TD == 0) 1467 TD = getAnalysisIfAvailable<TargetData>(); 1468 1469 // If the stored value is larger or equal to the loaded value, we can 1470 // reuse it. 1471 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1472 UnavailableBlocks.push_back(DepBB); 1473 continue; 1474 } 1475 } 1476 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1477 continue; 1478 } 1479 1480 UnavailableBlocks.push_back(DepBB); 1481 continue; 1482 } 1483 1484 // If we have no predecessors that produce a known value for this load, exit 1485 // early. 1486 if (ValuesPerBlock.empty()) return false; 1487 1488 // If all of the instructions we depend on produce a known value for this 1489 // load, then it is fully redundant and we can use PHI insertion to compute 1490 // its value. Insert PHIs and remove the fully redundant value now. 1491 if (UnavailableBlocks.empty()) { 1492 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1493 1494 // Perform PHI construction. 1495 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1496 VN.getAliasAnalysis()); 1497 LI->replaceAllUsesWith(V); 1498 1499 if (isa<PHINode>(V)) 1500 V->takeName(LI); 1501 if (V->getType()->isPointerTy()) 1502 MD->invalidateCachedPointerInfo(V); 1503 VN.erase(LI); 1504 toErase.push_back(LI); 1505 ++NumGVNLoad; 1506 return true; 1507 } 1508 1509 if (!EnablePRE || !EnableLoadPRE) 1510 return false; 1511 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 // Lets find first basic block with more than one predecessor. Walk backwards 1525 // through predecessors if needed. 1526 BasicBlock *LoadBB = LI->getParent(); 1527 BasicBlock *TmpBB = LoadBB; 1528 1529 bool isSinglePred = false; 1530 bool allSingleSucc = true; 1531 while (TmpBB->getSinglePredecessor()) { 1532 isSinglePred = true; 1533 TmpBB = TmpBB->getSinglePredecessor(); 1534 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1535 return false; 1536 if (Blockers.count(TmpBB)) 1537 return false; 1538 1539 // If any of these blocks has more than one successor (i.e. if the edge we 1540 // just traversed was critical), then there are other paths through this 1541 // block along which the load may not be anticipated. Hoisting the load 1542 // above this block would be adding the load to execution paths along 1543 // which it was not previously executed. 1544 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1545 return false; 1546 } 1547 1548 assert(TmpBB); 1549 LoadBB = TmpBB; 1550 1551 // FIXME: It is extremely unclear what this loop is doing, other than 1552 // artificially restricting loadpre. 1553 if (isSinglePred) { 1554 bool isHot = false; 1555 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1556 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1557 if (AV.isSimpleValue()) 1558 // "Hot" Instruction is in some loop (because it dominates its dep. 1559 // instruction). 1560 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1561 if (DT->dominates(LI, I)) { 1562 isHot = true; 1563 break; 1564 } 1565 } 1566 1567 // We are interested only in "hot" instructions. We don't want to do any 1568 // mis-optimizations here. 1569 if (!isHot) 1570 return false; 1571 } 1572 1573 // Check to see how many predecessors have the loaded value fully 1574 // available. 1575 DenseMap<BasicBlock*, Value*> PredLoads; 1576 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1577 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1578 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1579 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1580 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1581 1582 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1583 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1584 PI != E; ++PI) { 1585 BasicBlock *Pred = *PI; 1586 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1587 continue; 1588 } 1589 PredLoads[Pred] = 0; 1590 1591 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1592 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1593 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1594 << Pred->getName() << "': " << *LI << '\n'); 1595 return false; 1596 } 1597 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1598 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1599 } 1600 } 1601 if (!NeedToSplit.empty()) { 1602 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1603 return false; 1604 } 1605 1606 // Decide whether PRE is profitable for this load. 1607 unsigned NumUnavailablePreds = PredLoads.size(); 1608 assert(NumUnavailablePreds != 0 && 1609 "Fully available value should be eliminated above!"); 1610 1611 // If this load is unavailable in multiple predecessors, reject it. 1612 // FIXME: If we could restructure the CFG, we could make a common pred with 1613 // all the preds that don't have an available LI and insert a new load into 1614 // that one block. 1615 if (NumUnavailablePreds != 1) 1616 return false; 1617 1618 // Check if the load can safely be moved to all the unavailable predecessors. 1619 bool CanDoPRE = true; 1620 SmallVector<Instruction*, 8> NewInsts; 1621 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1622 E = PredLoads.end(); I != E; ++I) { 1623 BasicBlock *UnavailablePred = I->first; 1624 1625 // Do PHI translation to get its value in the predecessor if necessary. The 1626 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1627 1628 // If all preds have a single successor, then we know it is safe to insert 1629 // the load on the pred (?!?), so we can insert code to materialize the 1630 // pointer if it is not available. 1631 PHITransAddr Address(LI->getOperand(0), TD); 1632 Value *LoadPtr = 0; 1633 if (allSingleSucc) { 1634 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1635 *DT, NewInsts); 1636 } else { 1637 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1638 LoadPtr = Address.getAddr(); 1639 } 1640 1641 // If we couldn't find or insert a computation of this phi translated value, 1642 // we fail PRE. 1643 if (LoadPtr == 0) { 1644 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1645 << *LI->getOperand(0) << "\n"); 1646 CanDoPRE = false; 1647 break; 1648 } 1649 1650 // Make sure it is valid to move this load here. We have to watch out for: 1651 // @1 = getelementptr (i8* p, ... 1652 // test p and branch if == 0 1653 // load @1 1654 // It is valid to have the getelementptr before the test, even if p can be 0, 1655 // as getelementptr only does address arithmetic. 1656 // If we are not pushing the value through any multiple-successor blocks 1657 // we do not have this case. Otherwise, check that the load is safe to 1658 // put anywhere; this can be improved, but should be conservatively safe. 1659 if (!allSingleSucc && 1660 // FIXME: REEVALUTE THIS. 1661 !isSafeToLoadUnconditionally(LoadPtr, 1662 UnavailablePred->getTerminator(), 1663 LI->getAlignment(), TD)) { 1664 CanDoPRE = false; 1665 break; 1666 } 1667 1668 I->second = LoadPtr; 1669 } 1670 1671 if (!CanDoPRE) { 1672 while (!NewInsts.empty()) 1673 NewInsts.pop_back_val()->eraseFromParent(); 1674 return false; 1675 } 1676 1677 // Okay, we can eliminate this load by inserting a reload in the predecessor 1678 // and using PHI construction to get the value in the other predecessors, do 1679 // it. 1680 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1681 DEBUG(if (!NewInsts.empty()) 1682 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1683 << *NewInsts.back() << '\n'); 1684 1685 // Assign value numbers to the new instructions. 1686 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1687 // FIXME: We really _ought_ to insert these value numbers into their 1688 // parent's availability map. However, in doing so, we risk getting into 1689 // ordering issues. If a block hasn't been processed yet, we would be 1690 // marking a value as AVAIL-IN, which isn't what we intend. 1691 VN.lookup_or_add(NewInsts[i]); 1692 } 1693 1694 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1695 E = PredLoads.end(); I != E; ++I) { 1696 BasicBlock *UnavailablePred = I->first; 1697 Value *LoadPtr = I->second; 1698 1699 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1700 LI->getAlignment(), 1701 UnavailablePred->getTerminator()); 1702 1703 // Add the newly created load. 1704 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1705 NewLoad)); 1706 MD->invalidateCachedPointerInfo(LoadPtr); 1707 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1708 } 1709 1710 // Perform PHI construction. 1711 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1712 VN.getAliasAnalysis()); 1713 LI->replaceAllUsesWith(V); 1714 if (isa<PHINode>(V)) 1715 V->takeName(LI); 1716 if (V->getType()->isPointerTy()) 1717 MD->invalidateCachedPointerInfo(V); 1718 VN.erase(LI); 1719 toErase.push_back(LI); 1720 ++NumPRELoad; 1721 return true; 1722} 1723 1724/// processLoad - Attempt to eliminate a load, first by eliminating it 1725/// locally, and then attempting non-local elimination if that fails. 1726bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1727 if (!MD) 1728 return false; 1729 1730 if (L->isVolatile()) 1731 return false; 1732 1733 // ... to a pointer that has been loaded from before... 1734 MemDepResult Dep = MD->getDependency(L); 1735 1736 // If the value isn't available, don't do anything! 1737 if (Dep.isClobber()) { 1738 // Check to see if we have something like this: 1739 // store i32 123, i32* %P 1740 // %A = bitcast i32* %P to i8* 1741 // %B = gep i8* %A, i32 1 1742 // %C = load i8* %B 1743 // 1744 // We could do that by recognizing if the clobber instructions are obviously 1745 // a common base + constant offset, and if the previous store (or memset) 1746 // completely covers this load. This sort of thing can happen in bitfield 1747 // access code. 1748 Value *AvailVal = 0; 1749 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1750 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1751 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1752 L->getPointerOperand(), 1753 DepSI, *TD); 1754 if (Offset != -1) 1755 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1756 L->getType(), L, *TD); 1757 } 1758 1759 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1760 // a value on from it. 1761 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1762 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1763 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1764 L->getPointerOperand(), 1765 DepMI, *TD); 1766 if (Offset != -1) 1767 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD); 1768 } 1769 } 1770 1771 if (AvailVal) { 1772 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1773 << *AvailVal << '\n' << *L << "\n\n\n"); 1774 1775 // Replace the load! 1776 L->replaceAllUsesWith(AvailVal); 1777 if (AvailVal->getType()->isPointerTy()) 1778 MD->invalidateCachedPointerInfo(AvailVal); 1779 VN.erase(L); 1780 toErase.push_back(L); 1781 ++NumGVNLoad; 1782 return true; 1783 } 1784 1785 DEBUG( 1786 // fast print dep, using operator<< on instruction would be too slow 1787 dbgs() << "GVN: load "; 1788 WriteAsOperand(dbgs(), L); 1789 Instruction *I = Dep.getInst(); 1790 dbgs() << " is clobbered by " << *I << '\n'; 1791 ); 1792 return false; 1793 } 1794 1795 // If it is defined in another block, try harder. 1796 if (Dep.isNonLocal()) 1797 return processNonLocalLoad(L, toErase); 1798 1799 Instruction *DepInst = Dep.getInst(); 1800 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1801 Value *StoredVal = DepSI->getOperand(0); 1802 1803 // The store and load are to a must-aliased pointer, but they may not 1804 // actually have the same type. See if we know how to reuse the stored 1805 // value (depending on its type). 1806 const TargetData *TD = 0; 1807 if (StoredVal->getType() != L->getType()) { 1808 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1809 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1810 L, *TD); 1811 if (StoredVal == 0) 1812 return false; 1813 1814 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1815 << '\n' << *L << "\n\n\n"); 1816 } 1817 else 1818 return false; 1819 } 1820 1821 // Remove it! 1822 L->replaceAllUsesWith(StoredVal); 1823 if (StoredVal->getType()->isPointerTy()) 1824 MD->invalidateCachedPointerInfo(StoredVal); 1825 VN.erase(L); 1826 toErase.push_back(L); 1827 ++NumGVNLoad; 1828 return true; 1829 } 1830 1831 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1832 Value *AvailableVal = DepLI; 1833 1834 // The loads are of a must-aliased pointer, but they may not actually have 1835 // the same type. See if we know how to reuse the previously loaded value 1836 // (depending on its type). 1837 const TargetData *TD = 0; 1838 if (DepLI->getType() != L->getType()) { 1839 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1840 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1841 if (AvailableVal == 0) 1842 return false; 1843 1844 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1845 << "\n" << *L << "\n\n\n"); 1846 } 1847 else 1848 return false; 1849 } 1850 1851 // Remove it! 1852 L->replaceAllUsesWith(AvailableVal); 1853 if (DepLI->getType()->isPointerTy()) 1854 MD->invalidateCachedPointerInfo(DepLI); 1855 VN.erase(L); 1856 toErase.push_back(L); 1857 ++NumGVNLoad; 1858 return true; 1859 } 1860 1861 // If this load really doesn't depend on anything, then we must be loading an 1862 // undef value. This can happen when loading for a fresh allocation with no 1863 // intervening stores, for example. 1864 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1865 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1866 VN.erase(L); 1867 toErase.push_back(L); 1868 ++NumGVNLoad; 1869 return true; 1870 } 1871 1872 // If this load occurs either right after a lifetime begin, 1873 // then the loaded value is undefined. 1874 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) { 1875 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1876 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1877 VN.erase(L); 1878 toErase.push_back(L); 1879 ++NumGVNLoad; 1880 return true; 1881 } 1882 } 1883 1884 return false; 1885} 1886 1887Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1888 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1889 if (I == localAvail.end()) 1890 return 0; 1891 1892 ValueNumberScope *Locals = I->second; 1893 while (Locals) { 1894 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1895 if (I != Locals->table.end()) 1896 return I->second; 1897 Locals = Locals->parent; 1898 } 1899 1900 return 0; 1901} 1902 1903 1904/// processInstruction - When calculating availability, handle an instruction 1905/// by inserting it into the appropriate sets 1906bool GVN::processInstruction(Instruction *I, 1907 SmallVectorImpl<Instruction*> &toErase) { 1908 // Ignore dbg info intrinsics. 1909 if (isa<DbgInfoIntrinsic>(I)) 1910 return false; 1911 1912 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1913 bool Changed = processLoad(LI, toErase); 1914 1915 if (!Changed) { 1916 unsigned Num = VN.lookup_or_add(LI); 1917 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1918 } 1919 1920 return Changed; 1921 } 1922 1923 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1924 unsigned Num = VN.lookup_or_add(I); 1925 1926 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1927 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1928 1929 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1930 return false; 1931 1932 Value *BranchCond = BI->getCondition(); 1933 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1934 1935 BasicBlock *TrueSucc = BI->getSuccessor(0); 1936 BasicBlock *FalseSucc = BI->getSuccessor(1); 1937 1938 if (TrueSucc->getSinglePredecessor()) 1939 localAvail[TrueSucc]->table[CondVN] = 1940 ConstantInt::getTrue(TrueSucc->getContext()); 1941 if (FalseSucc->getSinglePredecessor()) 1942 localAvail[FalseSucc]->table[CondVN] = 1943 ConstantInt::getFalse(TrueSucc->getContext()); 1944 1945 return false; 1946 1947 // Allocations are always uniquely numbered, so we can save time and memory 1948 // by fast failing them. 1949 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) { 1950 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1951 return false; 1952 } 1953 1954 // Collapse PHI nodes 1955 if (PHINode* p = dyn_cast<PHINode>(I)) { 1956 Value *constVal = CollapsePhi(p); 1957 1958 if (constVal) { 1959 p->replaceAllUsesWith(constVal); 1960 if (MD && constVal->getType()->isPointerTy()) 1961 MD->invalidateCachedPointerInfo(constVal); 1962 VN.erase(p); 1963 1964 toErase.push_back(p); 1965 } else { 1966 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1967 } 1968 1969 // If the number we were assigned was a brand new VN, then we don't 1970 // need to do a lookup to see if the number already exists 1971 // somewhere in the domtree: it can't! 1972 } else if (Num == NextNum) { 1973 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1974 1975 // Perform fast-path value-number based elimination of values inherited from 1976 // dominators. 1977 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1978 // Remove it! 1979 VN.erase(I); 1980 I->replaceAllUsesWith(repl); 1981 if (MD && repl->getType()->isPointerTy()) 1982 MD->invalidateCachedPointerInfo(repl); 1983 toErase.push_back(I); 1984 return true; 1985 1986 } else { 1987 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1988 } 1989 1990 return false; 1991} 1992 1993/// runOnFunction - This is the main transformation entry point for a function. 1994bool GVN::runOnFunction(Function& F) { 1995 if (!NoLoads) 1996 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1997 DT = &getAnalysis<DominatorTree>(); 1998 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1999 VN.setMemDep(MD); 2000 VN.setDomTree(DT); 2001 2002 bool Changed = false; 2003 bool ShouldContinue = true; 2004 2005 // Merge unconditional branches, allowing PRE to catch more 2006 // optimization opportunities. 2007 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2008 BasicBlock *BB = FI; 2009 ++FI; 2010 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2011 if (removedBlock) ++NumGVNBlocks; 2012 2013 Changed |= removedBlock; 2014 } 2015 2016 unsigned Iteration = 0; 2017 2018 while (ShouldContinue) { 2019 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2020 ShouldContinue = iterateOnFunction(F); 2021 if (splitCriticalEdges()) 2022 ShouldContinue = true; 2023 Changed |= ShouldContinue; 2024 ++Iteration; 2025 } 2026 2027 if (EnablePRE) { 2028 bool PREChanged = true; 2029 while (PREChanged) { 2030 PREChanged = performPRE(F); 2031 Changed |= PREChanged; 2032 } 2033 } 2034 // FIXME: Should perform GVN again after PRE does something. PRE can move 2035 // computations into blocks where they become fully redundant. Note that 2036 // we can't do this until PRE's critical edge splitting updates memdep. 2037 // Actually, when this happens, we should just fully integrate PRE into GVN. 2038 2039 cleanupGlobalSets(); 2040 2041 return Changed; 2042} 2043 2044 2045bool GVN::processBlock(BasicBlock *BB) { 2046 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 2047 // incrementing BI before processing an instruction). 2048 SmallVector<Instruction*, 8> toErase; 2049 bool ChangedFunction = false; 2050 2051 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2052 BI != BE;) { 2053 ChangedFunction |= processInstruction(BI, toErase); 2054 if (toErase.empty()) { 2055 ++BI; 2056 continue; 2057 } 2058 2059 // If we need some instructions deleted, do it now. 2060 NumGVNInstr += toErase.size(); 2061 2062 // Avoid iterator invalidation. 2063 bool AtStart = BI == BB->begin(); 2064 if (!AtStart) 2065 --BI; 2066 2067 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 2068 E = toErase.end(); I != E; ++I) { 2069 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2070 if (MD) MD->removeInstruction(*I); 2071 (*I)->eraseFromParent(); 2072 DEBUG(verifyRemoved(*I)); 2073 } 2074 toErase.clear(); 2075 2076 if (AtStart) 2077 BI = BB->begin(); 2078 else 2079 ++BI; 2080 } 2081 2082 return ChangedFunction; 2083} 2084 2085/// performPRE - Perform a purely local form of PRE that looks for diamond 2086/// control flow patterns and attempts to perform simple PRE at the join point. 2087bool GVN::performPRE(Function &F) { 2088 bool Changed = false; 2089 DenseMap<BasicBlock*, Value*> predMap; 2090 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2091 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2092 BasicBlock *CurrentBlock = *DI; 2093 2094 // Nothing to PRE in the entry block. 2095 if (CurrentBlock == &F.getEntryBlock()) continue; 2096 2097 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2098 BE = CurrentBlock->end(); BI != BE; ) { 2099 Instruction *CurInst = BI++; 2100 2101 if (isa<AllocaInst>(CurInst) || 2102 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2103 CurInst->getType()->isVoidTy() || 2104 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2105 isa<DbgInfoIntrinsic>(CurInst)) 2106 continue; 2107 2108 // We don't currently value number ANY inline asm calls. 2109 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2110 if (CallI->isInlineAsm()) 2111 continue; 2112 2113 uint32_t ValNo = VN.lookup(CurInst); 2114 2115 // Look for the predecessors for PRE opportunities. We're 2116 // only trying to solve the basic diamond case, where 2117 // a value is computed in the successor and one predecessor, 2118 // but not the other. We also explicitly disallow cases 2119 // where the successor is its own predecessor, because they're 2120 // more complicated to get right. 2121 unsigned NumWith = 0; 2122 unsigned NumWithout = 0; 2123 BasicBlock *PREPred = 0; 2124 predMap.clear(); 2125 2126 for (pred_iterator PI = pred_begin(CurrentBlock), 2127 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2128 BasicBlock *P = *PI; 2129 // We're not interested in PRE where the block is its 2130 // own predecessor, or in blocks with predecessors 2131 // that are not reachable. 2132 if (P == CurrentBlock) { 2133 NumWithout = 2; 2134 break; 2135 } else if (!localAvail.count(P)) { 2136 NumWithout = 2; 2137 break; 2138 } 2139 2140 DenseMap<uint32_t, Value*>::iterator predV = 2141 localAvail[P]->table.find(ValNo); 2142 if (predV == localAvail[P]->table.end()) { 2143 PREPred = P; 2144 ++NumWithout; 2145 } else if (predV->second == CurInst) { 2146 NumWithout = 2; 2147 } else { 2148 predMap[P] = predV->second; 2149 ++NumWith; 2150 } 2151 } 2152 2153 // Don't do PRE when it might increase code size, i.e. when 2154 // we would need to insert instructions in more than one pred. 2155 if (NumWithout != 1 || NumWith == 0) 2156 continue; 2157 2158 // Don't do PRE across indirect branch. 2159 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2160 continue; 2161 2162 // We can't do PRE safely on a critical edge, so instead we schedule 2163 // the edge to be split and perform the PRE the next time we iterate 2164 // on the function. 2165 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2166 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2167 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2168 continue; 2169 } 2170 2171 // Instantiate the expression in the predecessor that lacked it. 2172 // Because we are going top-down through the block, all value numbers 2173 // will be available in the predecessor by the time we need them. Any 2174 // that weren't originally present will have been instantiated earlier 2175 // in this loop. 2176 Instruction *PREInstr = CurInst->clone(); 2177 bool success = true; 2178 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2179 Value *Op = PREInstr->getOperand(i); 2180 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2181 continue; 2182 2183 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 2184 PREInstr->setOperand(i, V); 2185 } else { 2186 success = false; 2187 break; 2188 } 2189 } 2190 2191 // Fail out if we encounter an operand that is not available in 2192 // the PRE predecessor. This is typically because of loads which 2193 // are not value numbered precisely. 2194 if (!success) { 2195 delete PREInstr; 2196 DEBUG(verifyRemoved(PREInstr)); 2197 continue; 2198 } 2199 2200 PREInstr->insertBefore(PREPred->getTerminator()); 2201 PREInstr->setName(CurInst->getName() + ".pre"); 2202 predMap[PREPred] = PREInstr; 2203 VN.add(PREInstr, ValNo); 2204 ++NumGVNPRE; 2205 2206 // Update the availability map to include the new instruction. 2207 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 2208 2209 // Create a PHI to make the value available in this block. 2210 PHINode* Phi = PHINode::Create(CurInst->getType(), 2211 CurInst->getName() + ".pre-phi", 2212 CurrentBlock->begin()); 2213 for (pred_iterator PI = pred_begin(CurrentBlock), 2214 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2215 BasicBlock *P = *PI; 2216 Phi->addIncoming(predMap[P], P); 2217 } 2218 2219 VN.add(Phi, ValNo); 2220 localAvail[CurrentBlock]->table[ValNo] = Phi; 2221 2222 CurInst->replaceAllUsesWith(Phi); 2223 if (MD && Phi->getType()->isPointerTy()) 2224 MD->invalidateCachedPointerInfo(Phi); 2225 VN.erase(CurInst); 2226 2227 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2228 if (MD) MD->removeInstruction(CurInst); 2229 CurInst->eraseFromParent(); 2230 DEBUG(verifyRemoved(CurInst)); 2231 Changed = true; 2232 } 2233 } 2234 2235 if (splitCriticalEdges()) 2236 Changed = true; 2237 2238 return Changed; 2239} 2240 2241/// splitCriticalEdges - Split critical edges found during the previous 2242/// iteration that may enable further optimization. 2243bool GVN::splitCriticalEdges() { 2244 if (toSplit.empty()) 2245 return false; 2246 do { 2247 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2248 SplitCriticalEdge(Edge.first, Edge.second, this); 2249 } while (!toSplit.empty()); 2250 if (MD) MD->invalidateCachedPredecessors(); 2251 return true; 2252} 2253 2254/// iterateOnFunction - Executes one iteration of GVN 2255bool GVN::iterateOnFunction(Function &F) { 2256 cleanupGlobalSets(); 2257 2258 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2259 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 2260 if (DI->getIDom()) 2261 localAvail[DI->getBlock()] = 2262 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 2263 else 2264 localAvail[DI->getBlock()] = new ValueNumberScope(0); 2265 } 2266 2267 // Top-down walk of the dominator tree 2268 bool Changed = false; 2269#if 0 2270 // Needed for value numbering with phi construction to work. 2271 ReversePostOrderTraversal<Function*> RPOT(&F); 2272 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2273 RE = RPOT.end(); RI != RE; ++RI) 2274 Changed |= processBlock(*RI); 2275#else 2276 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2277 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2278 Changed |= processBlock(DI->getBlock()); 2279#endif 2280 2281 return Changed; 2282} 2283 2284void GVN::cleanupGlobalSets() { 2285 VN.clear(); 2286 2287 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2288 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 2289 delete I->second; 2290 localAvail.clear(); 2291} 2292 2293/// verifyRemoved - Verify that the specified instruction does not occur in our 2294/// internal data structures. 2295void GVN::verifyRemoved(const Instruction *Inst) const { 2296 VN.verifyRemoved(Inst); 2297 2298 // Walk through the value number scope to make sure the instruction isn't 2299 // ferreted away in it. 2300 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator 2301 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 2302 const ValueNumberScope *VNS = I->second; 2303 2304 while (VNS) { 2305 for (DenseMap<uint32_t, Value*>::const_iterator 2306 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 2307 assert(II->second != Inst && "Inst still in value numbering scope!"); 2308 } 2309 2310 VNS = VNS->parent; 2311 } 2312 } 2313} 2314