GVN.cpp revision 70704978af7d37d7951c576ccb7b0cd17053b5a5
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/MemoryBuiltins.h" 39#include "llvm/Analysis/MemoryDependenceAnalysis.h" 40#include "llvm/Analysis/PHITransAddr.h" 41#include "llvm/Support/CFG.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/ErrorHandling.h" 45#include "llvm/Support/GetElementPtrTypeIterator.h" 46#include "llvm/Support/IRBuilder.h" 47#include "llvm/Support/raw_ostream.h" 48#include "llvm/Target/TargetData.h" 49#include "llvm/Transforms/Utils/BasicBlockUtils.h" 50#include "llvm/Transforms/Utils/Local.h" 51#include "llvm/Transforms/Utils/SSAUpdater.h" 52using namespace llvm; 53 54STATISTIC(NumGVNInstr, "Number of instructions deleted"); 55STATISTIC(NumGVNLoad, "Number of loads deleted"); 56STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 57STATISTIC(NumGVNBlocks, "Number of blocks merged"); 58STATISTIC(NumPRELoad, "Number of loads PRE'd"); 59 60static cl::opt<bool> EnablePRE("enable-pre", 61 cl::init(true), cl::Hidden); 62static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 63static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false)); 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(Constant* C); 168 Expression create_expression(ExtractValueInst* C); 169 Expression create_expression(InsertValueInst* C); 170 171 uint32_t lookup_or_add_call(CallInst* C); 172 public: 173 ValueTable() : nextValueNumber(1) { } 174 uint32_t lookup_or_add(Value *V); 175 uint32_t lookup(Value *V) const; 176 void add(Value *V, uint32_t num); 177 void clear(); 178 void erase(Value *v); 179 unsigned size(); 180 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 181 AliasAnalysis *getAliasAnalysis() const { return AA; } 182 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 183 void setDomTree(DominatorTree* D) { DT = D; } 184 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 185 void verifyRemoved(const Value *) const; 186 }; 187} 188 189namespace llvm { 190template <> struct DenseMapInfo<Expression> { 191 static inline Expression getEmptyKey() { 192 return Expression(Expression::EMPTY); 193 } 194 195 static inline Expression getTombstoneKey() { 196 return Expression(Expression::TOMBSTONE); 197 } 198 199 static unsigned getHashValue(const Expression e) { 200 unsigned hash = e.opcode; 201 202 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 203 (unsigned)((uintptr_t)e.type >> 9)); 204 205 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 206 E = e.varargs.end(); I != E; ++I) 207 hash = *I + hash * 37; 208 209 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 210 (unsigned)((uintptr_t)e.function >> 9)) + 211 hash * 37; 212 213 return hash; 214 } 215 static bool isEqual(const Expression &LHS, const Expression &RHS) { 216 return LHS == RHS; 217 } 218}; 219 220template <> 221struct isPodLike<Expression> { static const bool value = true; }; 222 223} 224 225//===----------------------------------------------------------------------===// 226// ValueTable Internal Functions 227//===----------------------------------------------------------------------===// 228 229Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 230 if (isa<ICmpInst>(C)) { 231 switch (C->getPredicate()) { 232 default: // THIS SHOULD NEVER HAPPEN 233 llvm_unreachable("Comparison with unknown predicate?"); 234 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 235 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 236 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 237 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 238 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 239 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 240 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 241 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 242 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 243 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 244 } 245 } else { 246 switch (C->getPredicate()) { 247 default: // THIS SHOULD NEVER HAPPEN 248 llvm_unreachable("Comparison with unknown predicate?"); 249 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 250 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 251 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 252 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 253 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 254 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 255 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 256 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 257 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 258 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 259 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 260 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 261 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 262 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 263 } 264 } 265} 266 267Expression ValueTable::create_expression(CallInst* C) { 268 Expression e; 269 270 e.type = C->getType(); 271 e.function = C->getCalledFunction(); 272 e.opcode = Expression::CALL; 273 274 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 275 I != E; ++I) 276 e.varargs.push_back(lookup_or_add(*I)); 277 278 return e; 279} 280 281Expression ValueTable::create_expression(BinaryOperator* BO) { 282 Expression e; 283 e.varargs.push_back(lookup_or_add(BO->getOperand(0))); 284 e.varargs.push_back(lookup_or_add(BO->getOperand(1))); 285 e.function = 0; 286 e.type = BO->getType(); 287 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode()); 288 289 return e; 290} 291 292Expression ValueTable::create_expression(CmpInst* C) { 293 Expression e; 294 295 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 296 e.varargs.push_back(lookup_or_add(C->getOperand(1))); 297 e.function = 0; 298 e.type = C->getType(); 299 e.opcode = getOpcode(C); 300 301 return e; 302} 303 304Expression ValueTable::create_expression(CastInst* C) { 305 Expression e; 306 307 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 308 e.function = 0; 309 e.type = C->getType(); 310 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode()); 311 312 return e; 313} 314 315Expression ValueTable::create_expression(ShuffleVectorInst* S) { 316 Expression e; 317 318 e.varargs.push_back(lookup_or_add(S->getOperand(0))); 319 e.varargs.push_back(lookup_or_add(S->getOperand(1))); 320 e.varargs.push_back(lookup_or_add(S->getOperand(2))); 321 e.function = 0; 322 e.type = S->getType(); 323 e.opcode = Expression::SHUFFLE; 324 325 return e; 326} 327 328Expression ValueTable::create_expression(ExtractElementInst* E) { 329 Expression e; 330 331 e.varargs.push_back(lookup_or_add(E->getOperand(0))); 332 e.varargs.push_back(lookup_or_add(E->getOperand(1))); 333 e.function = 0; 334 e.type = E->getType(); 335 e.opcode = Expression::EXTRACT; 336 337 return e; 338} 339 340Expression ValueTable::create_expression(InsertElementInst* I) { 341 Expression e; 342 343 e.varargs.push_back(lookup_or_add(I->getOperand(0))); 344 e.varargs.push_back(lookup_or_add(I->getOperand(1))); 345 e.varargs.push_back(lookup_or_add(I->getOperand(2))); 346 e.function = 0; 347 e.type = I->getType(); 348 e.opcode = Expression::INSERT; 349 350 return e; 351} 352 353Expression ValueTable::create_expression(SelectInst* I) { 354 Expression e; 355 356 e.varargs.push_back(lookup_or_add(I->getCondition())); 357 e.varargs.push_back(lookup_or_add(I->getTrueValue())); 358 e.varargs.push_back(lookup_or_add(I->getFalseValue())); 359 e.function = 0; 360 e.type = I->getType(); 361 e.opcode = Expression::SELECT; 362 363 return e; 364} 365 366Expression ValueTable::create_expression(GetElementPtrInst* G) { 367 Expression e; 368 369 e.varargs.push_back(lookup_or_add(G->getPointerOperand())); 370 e.function = 0; 371 e.type = G->getType(); 372 e.opcode = Expression::GEP; 373 374 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 375 I != E; ++I) 376 e.varargs.push_back(lookup_or_add(*I)); 377 378 return e; 379} 380 381Expression ValueTable::create_expression(ExtractValueInst* E) { 382 Expression e; 383 384 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 385 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 386 II != IE; ++II) 387 e.varargs.push_back(*II); 388 e.function = 0; 389 e.type = E->getType(); 390 e.opcode = Expression::EXTRACTVALUE; 391 392 return e; 393} 394 395Expression ValueTable::create_expression(InsertValueInst* E) { 396 Expression e; 397 398 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 399 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand())); 400 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 401 II != IE; ++II) 402 e.varargs.push_back(*II); 403 e.function = 0; 404 e.type = E->getType(); 405 e.opcode = Expression::INSERTVALUE; 406 407 return e; 408} 409 410//===----------------------------------------------------------------------===// 411// ValueTable External Functions 412//===----------------------------------------------------------------------===// 413 414/// add - Insert a value into the table with a specified value number. 415void ValueTable::add(Value *V, uint32_t num) { 416 valueNumbering.insert(std::make_pair(V, num)); 417} 418 419uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 420 if (AA->doesNotAccessMemory(C)) { 421 Expression exp = create_expression(C); 422 uint32_t& e = expressionNumbering[exp]; 423 if (!e) e = nextValueNumber++; 424 valueNumbering[C] = e; 425 return e; 426 } else if (AA->onlyReadsMemory(C)) { 427 Expression exp = create_expression(C); 428 uint32_t& e = expressionNumbering[exp]; 429 if (!e) { 430 e = nextValueNumber++; 431 valueNumbering[C] = e; 432 return e; 433 } 434 if (!MD) { 435 e = nextValueNumber++; 436 valueNumbering[C] = e; 437 return e; 438 } 439 440 MemDepResult local_dep = MD->getDependency(C); 441 442 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 443 valueNumbering[C] = nextValueNumber; 444 return nextValueNumber++; 445 } 446 447 if (local_dep.isDef()) { 448 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 449 450 if (local_cdep->getNumOperands() != C->getNumOperands()) { 451 valueNumbering[C] = nextValueNumber; 452 return nextValueNumber++; 453 } 454 455 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 456 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 457 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 458 if (c_vn != cd_vn) { 459 valueNumbering[C] = nextValueNumber; 460 return nextValueNumber++; 461 } 462 } 463 464 uint32_t v = lookup_or_add(local_cdep); 465 valueNumbering[C] = v; 466 return v; 467 } 468 469 // Non-local case. 470 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 471 MD->getNonLocalCallDependency(CallSite(C)); 472 // FIXME: call/call dependencies for readonly calls should return def, not 473 // clobber! Move the checking logic to MemDep! 474 CallInst* cdep = 0; 475 476 // Check to see if we have a single dominating call instruction that is 477 // identical to C. 478 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 479 const NonLocalDepEntry *I = &deps[i]; 480 // Ignore non-local dependencies. 481 if (I->getResult().isNonLocal()) 482 continue; 483 484 // We don't handle non-depedencies. If we already have a call, reject 485 // instruction dependencies. 486 if (I->getResult().isClobber() || cdep != 0) { 487 cdep = 0; 488 break; 489 } 490 491 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 492 // FIXME: All duplicated with non-local case. 493 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 494 cdep = NonLocalDepCall; 495 continue; 496 } 497 498 cdep = 0; 499 break; 500 } 501 502 if (!cdep) { 503 valueNumbering[C] = nextValueNumber; 504 return nextValueNumber++; 505 } 506 507 if (cdep->getNumOperands() != C->getNumOperands()) { 508 valueNumbering[C] = nextValueNumber; 509 return nextValueNumber++; 510 } 511 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 512 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 513 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 514 if (c_vn != cd_vn) { 515 valueNumbering[C] = nextValueNumber; 516 return nextValueNumber++; 517 } 518 } 519 520 uint32_t v = lookup_or_add(cdep); 521 valueNumbering[C] = v; 522 return v; 523 524 } else { 525 valueNumbering[C] = nextValueNumber; 526 return nextValueNumber++; 527 } 528} 529 530/// lookup_or_add - Returns the value number for the specified value, assigning 531/// it a new number if it did not have one before. 532uint32_t ValueTable::lookup_or_add(Value *V) { 533 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 534 if (VI != valueNumbering.end()) 535 return VI->second; 536 537 if (!isa<Instruction>(V)) { 538 valueNumbering[V] = nextValueNumber; 539 return nextValueNumber++; 540 } 541 542 Instruction* I = cast<Instruction>(V); 543 Expression exp; 544 switch (I->getOpcode()) { 545 case Instruction::Call: 546 return lookup_or_add_call(cast<CallInst>(I)); 547 case Instruction::Add: 548 case Instruction::FAdd: 549 case Instruction::Sub: 550 case Instruction::FSub: 551 case Instruction::Mul: 552 case Instruction::FMul: 553 case Instruction::UDiv: 554 case Instruction::SDiv: 555 case Instruction::FDiv: 556 case Instruction::URem: 557 case Instruction::SRem: 558 case Instruction::FRem: 559 case Instruction::Shl: 560 case Instruction::LShr: 561 case Instruction::AShr: 562 case Instruction::And: 563 case Instruction::Or : 564 case Instruction::Xor: 565 exp = create_expression(cast<BinaryOperator>(I)); 566 break; 567 case Instruction::ICmp: 568 case Instruction::FCmp: 569 exp = create_expression(cast<CmpInst>(I)); 570 break; 571 case Instruction::Trunc: 572 case Instruction::ZExt: 573 case Instruction::SExt: 574 case Instruction::FPToUI: 575 case Instruction::FPToSI: 576 case Instruction::UIToFP: 577 case Instruction::SIToFP: 578 case Instruction::FPTrunc: 579 case Instruction::FPExt: 580 case Instruction::PtrToInt: 581 case Instruction::IntToPtr: 582 case Instruction::BitCast: 583 exp = create_expression(cast<CastInst>(I)); 584 break; 585 case Instruction::Select: 586 exp = create_expression(cast<SelectInst>(I)); 587 break; 588 case Instruction::ExtractElement: 589 exp = create_expression(cast<ExtractElementInst>(I)); 590 break; 591 case Instruction::InsertElement: 592 exp = create_expression(cast<InsertElementInst>(I)); 593 break; 594 case Instruction::ShuffleVector: 595 exp = create_expression(cast<ShuffleVectorInst>(I)); 596 break; 597 case Instruction::ExtractValue: 598 exp = create_expression(cast<ExtractValueInst>(I)); 599 break; 600 case Instruction::InsertValue: 601 exp = create_expression(cast<InsertValueInst>(I)); 602 break; 603 case Instruction::GetElementPtr: 604 exp = create_expression(cast<GetElementPtrInst>(I)); 605 break; 606 default: 607 valueNumbering[V] = nextValueNumber; 608 return nextValueNumber++; 609 } 610 611 uint32_t& e = expressionNumbering[exp]; 612 if (!e) e = nextValueNumber++; 613 valueNumbering[V] = e; 614 return e; 615} 616 617/// lookup - Returns the value number of the specified value. Fails if 618/// the value has not yet been numbered. 619uint32_t ValueTable::lookup(Value *V) const { 620 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 621 assert(VI != valueNumbering.end() && "Value not numbered?"); 622 return VI->second; 623} 624 625/// clear - Remove all entries from the ValueTable 626void ValueTable::clear() { 627 valueNumbering.clear(); 628 expressionNumbering.clear(); 629 nextValueNumber = 1; 630} 631 632/// erase - Remove a value from the value numbering 633void ValueTable::erase(Value *V) { 634 valueNumbering.erase(V); 635} 636 637/// verifyRemoved - Verify that the value is removed from all internal data 638/// structures. 639void ValueTable::verifyRemoved(const Value *V) const { 640 for (DenseMap<Value*, uint32_t>::const_iterator 641 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 642 assert(I->first != V && "Inst still occurs in value numbering map!"); 643 } 644} 645 646//===----------------------------------------------------------------------===// 647// GVN Pass 648//===----------------------------------------------------------------------===// 649 650namespace { 651 struct ValueNumberScope { 652 ValueNumberScope* parent; 653 DenseMap<uint32_t, Value*> table; 654 655 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 656 }; 657} 658 659namespace { 660 661 class GVN : public FunctionPass { 662 bool runOnFunction(Function &F); 663 public: 664 static char ID; // Pass identification, replacement for typeid 665 explicit GVN(bool noloads = false) 666 : FunctionPass(&ID), NoLoads(noloads), MD(0) { } 667 668 private: 669 bool NoLoads; 670 MemoryDependenceAnalysis *MD; 671 DominatorTree *DT; 672 673 ValueTable VN; 674 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 675 676 // List of critical edges to be split between iterations. 677 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 678 679 // This transformation requires dominator postdominator info 680 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 681 AU.addRequired<DominatorTree>(); 682 if (!NoLoads) 683 AU.addRequired<MemoryDependenceAnalysis>(); 684 AU.addRequired<AliasAnalysis>(); 685 686 AU.addPreserved<DominatorTree>(); 687 AU.addPreserved<AliasAnalysis>(); 688 } 689 690 // Helper fuctions 691 // FIXME: eliminate or document these better 692 bool processLoad(LoadInst* L, 693 SmallVectorImpl<Instruction*> &toErase); 694 bool processInstruction(Instruction *I, 695 SmallVectorImpl<Instruction*> &toErase); 696 bool processNonLocalLoad(LoadInst* L, 697 SmallVectorImpl<Instruction*> &toErase); 698 bool processBlock(BasicBlock *BB); 699 void dump(DenseMap<uint32_t, Value*>& d); 700 bool iterateOnFunction(Function &F); 701 Value *CollapsePhi(PHINode* p); 702 bool performPRE(Function& F); 703 Value *lookupNumber(BasicBlock *BB, uint32_t num); 704 void cleanupGlobalSets(); 705 void verifyRemoved(const Instruction *I) const; 706 bool splitCriticalEdges(); 707 }; 708 709 char GVN::ID = 0; 710} 711 712// createGVNPass - The public interface to this file... 713FunctionPass *llvm::createGVNPass(bool NoLoads) { 714 return new GVN(NoLoads); 715} 716 717static RegisterPass<GVN> X("gvn", 718 "Global Value Numbering"); 719 720void GVN::dump(DenseMap<uint32_t, Value*>& d) { 721 errs() << "{\n"; 722 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 723 E = d.end(); I != E; ++I) { 724 errs() << I->first << "\n"; 725 I->second->dump(); 726 } 727 errs() << "}\n"; 728} 729 730static bool isSafeReplacement(PHINode* p, Instruction *inst) { 731 if (!isa<PHINode>(inst)) 732 return true; 733 734 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 735 UI != E; ++UI) 736 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 737 if (use_phi->getParent() == inst->getParent()) 738 return false; 739 740 return true; 741} 742 743Value *GVN::CollapsePhi(PHINode *PN) { 744 Value *ConstVal = PN->hasConstantValue(DT); 745 if (!ConstVal) return 0; 746 747 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 748 if (!Inst) 749 return ConstVal; 750 751 if (DT->dominates(Inst, PN)) 752 if (isSafeReplacement(PN, Inst)) 753 return Inst; 754 return 0; 755} 756 757/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 758/// we're analyzing is fully available in the specified block. As we go, keep 759/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 760/// map is actually a tri-state map with the following values: 761/// 0) we know the block *is not* fully available. 762/// 1) we know the block *is* fully available. 763/// 2) we do not know whether the block is fully available or not, but we are 764/// currently speculating that it will be. 765/// 3) we are speculating for this block and have used that to speculate for 766/// other blocks. 767static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 768 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 769 // Optimistically assume that the block is fully available and check to see 770 // if we already know about this block in one lookup. 771 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 772 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 773 774 // If the entry already existed for this block, return the precomputed value. 775 if (!IV.second) { 776 // If this is a speculative "available" value, mark it as being used for 777 // speculation of other blocks. 778 if (IV.first->second == 2) 779 IV.first->second = 3; 780 return IV.first->second != 0; 781 } 782 783 // Otherwise, see if it is fully available in all predecessors. 784 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 785 786 // If this block has no predecessors, it isn't live-in here. 787 if (PI == PE) 788 goto SpeculationFailure; 789 790 for (; PI != PE; ++PI) 791 // If the value isn't fully available in one of our predecessors, then it 792 // isn't fully available in this block either. Undo our previous 793 // optimistic assumption and bail out. 794 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 795 goto SpeculationFailure; 796 797 return true; 798 799// SpeculationFailure - If we get here, we found out that this is not, after 800// all, a fully-available block. We have a problem if we speculated on this and 801// used the speculation to mark other blocks as available. 802SpeculationFailure: 803 char &BBVal = FullyAvailableBlocks[BB]; 804 805 // If we didn't speculate on this, just return with it set to false. 806 if (BBVal == 2) { 807 BBVal = 0; 808 return false; 809 } 810 811 // If we did speculate on this value, we could have blocks set to 1 that are 812 // incorrect. Walk the (transitive) successors of this block and mark them as 813 // 0 if set to one. 814 SmallVector<BasicBlock*, 32> BBWorklist; 815 BBWorklist.push_back(BB); 816 817 do { 818 BasicBlock *Entry = BBWorklist.pop_back_val(); 819 // Note that this sets blocks to 0 (unavailable) if they happen to not 820 // already be in FullyAvailableBlocks. This is safe. 821 char &EntryVal = FullyAvailableBlocks[Entry]; 822 if (EntryVal == 0) continue; // Already unavailable. 823 824 // Mark as unavailable. 825 EntryVal = 0; 826 827 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 828 BBWorklist.push_back(*I); 829 } while (!BBWorklist.empty()); 830 831 return false; 832} 833 834 835/// CanCoerceMustAliasedValueToLoad - Return true if 836/// CoerceAvailableValueToLoadType will succeed. 837static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 838 const Type *LoadTy, 839 const TargetData &TD) { 840 // If the loaded or stored value is an first class array or struct, don't try 841 // to transform them. We need to be able to bitcast to integer. 842 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 843 StoredVal->getType()->isStructTy() || 844 StoredVal->getType()->isArrayTy()) 845 return false; 846 847 // The store has to be at least as big as the load. 848 if (TD.getTypeSizeInBits(StoredVal->getType()) < 849 TD.getTypeSizeInBits(LoadTy)) 850 return false; 851 852 return true; 853} 854 855 856/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 857/// then a load from a must-aliased pointer of a different type, try to coerce 858/// the stored value. LoadedTy is the type of the load we want to replace and 859/// InsertPt is the place to insert new instructions. 860/// 861/// If we can't do it, return null. 862static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 863 const Type *LoadedTy, 864 Instruction *InsertPt, 865 const TargetData &TD) { 866 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 867 return 0; 868 869 const Type *StoredValTy = StoredVal->getType(); 870 871 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 872 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 873 874 // If the store and reload are the same size, we can always reuse it. 875 if (StoreSize == LoadSize) { 876 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) { 877 // Pointer to Pointer -> use bitcast. 878 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 879 } 880 881 // Convert source pointers to integers, which can be bitcast. 882 if (StoredValTy->isPointerTy()) { 883 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 884 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 885 } 886 887 const Type *TypeToCastTo = LoadedTy; 888 if (TypeToCastTo->isPointerTy()) 889 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 890 891 if (StoredValTy != TypeToCastTo) 892 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 893 894 // Cast to pointer if the load needs a pointer type. 895 if (LoadedTy->isPointerTy()) 896 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 897 898 return StoredVal; 899 } 900 901 // If the loaded value is smaller than the available value, then we can 902 // extract out a piece from it. If the available value is too small, then we 903 // can't do anything. 904 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 905 906 // Convert source pointers to integers, which can be manipulated. 907 if (StoredValTy->isPointerTy()) { 908 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 909 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 910 } 911 912 // Convert vectors and fp to integer, which can be manipulated. 913 if (!StoredValTy->isIntegerTy()) { 914 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 915 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 916 } 917 918 // If this is a big-endian system, we need to shift the value down to the low 919 // bits so that a truncate will work. 920 if (TD.isBigEndian()) { 921 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 922 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 923 } 924 925 // Truncate the integer to the right size now. 926 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 927 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 928 929 if (LoadedTy == NewIntTy) 930 return StoredVal; 931 932 // If the result is a pointer, inttoptr. 933 if (LoadedTy->isPointerTy()) 934 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 935 936 // Otherwise, bitcast. 937 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 938} 939 940/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 941/// be expressed as a base pointer plus a constant offset. Return the base and 942/// offset to the caller. 943static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 944 const TargetData &TD) { 945 Operator *PtrOp = dyn_cast<Operator>(Ptr); 946 if (PtrOp == 0) return Ptr; 947 948 // Just look through bitcasts. 949 if (PtrOp->getOpcode() == Instruction::BitCast) 950 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 951 952 // If this is a GEP with constant indices, we can look through it. 953 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 954 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 955 956 gep_type_iterator GTI = gep_type_begin(GEP); 957 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 958 ++I, ++GTI) { 959 ConstantInt *OpC = cast<ConstantInt>(*I); 960 if (OpC->isZero()) continue; 961 962 // Handle a struct and array indices which add their offset to the pointer. 963 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 964 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 965 } else { 966 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 967 Offset += OpC->getSExtValue()*Size; 968 } 969 } 970 971 // Re-sign extend from the pointer size if needed to get overflow edge cases 972 // right. 973 unsigned PtrSize = TD.getPointerSizeInBits(); 974 if (PtrSize < 64) 975 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 976 977 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 978} 979 980 981/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 982/// memdep query of a load that ends up being a clobbering memory write (store, 983/// memset, memcpy, memmove). This means that the write *may* provide bits used 984/// by the load but we can't be sure because the pointers don't mustalias. 985/// 986/// Check this case to see if there is anything more we can do before we give 987/// up. This returns -1 if we have to give up, or a byte number in the stored 988/// value of the piece that feeds the load. 989static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 990 Value *WritePtr, 991 uint64_t WriteSizeInBits, 992 const TargetData &TD) { 993 // If the loaded or stored value is an first class array or struct, don't try 994 // to transform them. We need to be able to bitcast to integer. 995 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 996 return -1; 997 998 int64_t StoreOffset = 0, LoadOffset = 0; 999 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD); 1000 Value *LoadBase = 1001 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 1002 if (StoreBase != LoadBase) 1003 return -1; 1004 1005 // If the load and store are to the exact same address, they should have been 1006 // a must alias. AA must have gotten confused. 1007 // FIXME: Study to see if/when this happens. 1008 if (LoadOffset == StoreOffset) { 1009#if 0 1010 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1011 << "Base = " << *StoreBase << "\n" 1012 << "Store Ptr = " << *WritePtr << "\n" 1013 << "Store Offs = " << StoreOffset << "\n" 1014 << "Load Ptr = " << *LoadPtr << "\n"; 1015 abort(); 1016#endif 1017 return -1; 1018 } 1019 1020 // If the load and store don't overlap at all, the store doesn't provide 1021 // anything to the load. In this case, they really don't alias at all, AA 1022 // must have gotten confused. 1023 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 1024 // remove this check, as it is duplicated with what we have below. 1025 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 1026 1027 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 1028 return -1; 1029 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 1030 LoadSize >>= 3; 1031 1032 1033 bool isAAFailure = false; 1034 if (StoreOffset < LoadOffset) { 1035 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1036 } else { 1037 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1038 } 1039 if (isAAFailure) { 1040#if 0 1041 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1042 << "Base = " << *StoreBase << "\n" 1043 << "Store Ptr = " << *WritePtr << "\n" 1044 << "Store Offs = " << StoreOffset << "\n" 1045 << "Load Ptr = " << *LoadPtr << "\n"; 1046 abort(); 1047#endif 1048 return -1; 1049 } 1050 1051 // If the Load isn't completely contained within the stored bits, we don't 1052 // have all the bits to feed it. We could do something crazy in the future 1053 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1054 // valuable. 1055 if (StoreOffset > LoadOffset || 1056 StoreOffset+StoreSize < LoadOffset+LoadSize) 1057 return -1; 1058 1059 // Okay, we can do this transformation. Return the number of bytes into the 1060 // store that the load is. 1061 return LoadOffset-StoreOffset; 1062} 1063 1064/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1065/// memdep query of a load that ends up being a clobbering store. 1066static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 1067 StoreInst *DepSI, 1068 const TargetData &TD) { 1069 // Cannot handle reading from store of first-class aggregate yet. 1070 if (DepSI->getOperand(0)->getType()->isStructTy() || 1071 DepSI->getOperand(0)->getType()->isArrayTy()) 1072 return -1; 1073 1074 Value *StorePtr = DepSI->getPointerOperand(); 1075 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType()); 1076 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1077 StorePtr, StoreSize, TD); 1078} 1079 1080static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 1081 MemIntrinsic *MI, 1082 const TargetData &TD) { 1083 // If the mem operation is a non-constant size, we can't handle it. 1084 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1085 if (SizeCst == 0) return -1; 1086 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1087 1088 // If this is memset, we just need to see if the offset is valid in the size 1089 // of the memset.. 1090 if (MI->getIntrinsicID() == Intrinsic::memset) 1091 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1092 MemSizeInBits, TD); 1093 1094 // If we have a memcpy/memmove, the only case we can handle is if this is a 1095 // copy from constant memory. In that case, we can read directly from the 1096 // constant memory. 1097 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1098 1099 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1100 if (Src == 0) return -1; 1101 1102 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject()); 1103 if (GV == 0 || !GV->isConstant()) return -1; 1104 1105 // See if the access is within the bounds of the transfer. 1106 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1107 MI->getDest(), MemSizeInBits, TD); 1108 if (Offset == -1) 1109 return Offset; 1110 1111 // Otherwise, see if we can constant fold a load from the constant with the 1112 // offset applied as appropriate. 1113 Src = ConstantExpr::getBitCast(Src, 1114 llvm::Type::getInt8PtrTy(Src->getContext())); 1115 Constant *OffsetCst = 1116 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1117 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1118 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1119 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1120 return Offset; 1121 return -1; 1122} 1123 1124 1125/// GetStoreValueForLoad - This function is called when we have a 1126/// memdep query of a load that ends up being a clobbering store. This means 1127/// that the store *may* provide bits used by the load but we can't be sure 1128/// because the pointers don't mustalias. Check this case to see if there is 1129/// anything more we can do before we give up. 1130static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1131 const Type *LoadTy, 1132 Instruction *InsertPt, const TargetData &TD){ 1133 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1134 1135 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8; 1136 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1137 1138 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1139 1140 // Compute which bits of the stored value are being used by the load. Convert 1141 // to an integer type to start with. 1142 if (SrcVal->getType()->isPointerTy()) 1143 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 1144 if (!SrcVal->getType()->isIntegerTy()) 1145 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1146 "tmp"); 1147 1148 // Shift the bits to the least significant depending on endianness. 1149 unsigned ShiftAmt; 1150 if (TD.isLittleEndian()) 1151 ShiftAmt = Offset*8; 1152 else 1153 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1154 1155 if (ShiftAmt) 1156 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 1157 1158 if (LoadSize != StoreSize) 1159 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1160 "tmp"); 1161 1162 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1163} 1164 1165/// GetMemInstValueForLoad - This function is called when we have a 1166/// memdep query of a load that ends up being a clobbering mem intrinsic. 1167static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1168 const Type *LoadTy, Instruction *InsertPt, 1169 const TargetData &TD){ 1170 LLVMContext &Ctx = LoadTy->getContext(); 1171 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1172 1173 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1174 1175 // We know that this method is only called when the mem transfer fully 1176 // provides the bits for the load. 1177 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1178 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1179 // independently of what the offset is. 1180 Value *Val = MSI->getValue(); 1181 if (LoadSize != 1) 1182 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1183 1184 Value *OneElt = Val; 1185 1186 // Splat the value out to the right number of bits. 1187 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1188 // If we can double the number of bytes set, do it. 1189 if (NumBytesSet*2 <= LoadSize) { 1190 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1191 Val = Builder.CreateOr(Val, ShVal); 1192 NumBytesSet <<= 1; 1193 continue; 1194 } 1195 1196 // Otherwise insert one byte at a time. 1197 Value *ShVal = Builder.CreateShl(Val, 1*8); 1198 Val = Builder.CreateOr(OneElt, ShVal); 1199 ++NumBytesSet; 1200 } 1201 1202 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1203 } 1204 1205 // Otherwise, this is a memcpy/memmove from a constant global. 1206 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1207 Constant *Src = cast<Constant>(MTI->getSource()); 1208 1209 // Otherwise, see if we can constant fold a load from the constant with the 1210 // offset applied as appropriate. 1211 Src = ConstantExpr::getBitCast(Src, 1212 llvm::Type::getInt8PtrTy(Src->getContext())); 1213 Constant *OffsetCst = 1214 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1215 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1216 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1217 return ConstantFoldLoadFromConstPtr(Src, &TD); 1218} 1219 1220 1221 1222struct AvailableValueInBlock { 1223 /// BB - The basic block in question. 1224 BasicBlock *BB; 1225 enum ValType { 1226 SimpleVal, // A simple offsetted value that is accessed. 1227 MemIntrin // A memory intrinsic which is loaded from. 1228 }; 1229 1230 /// V - The value that is live out of the block. 1231 PointerIntPair<Value *, 1, ValType> Val; 1232 1233 /// Offset - The byte offset in Val that is interesting for the load query. 1234 unsigned Offset; 1235 1236 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1237 unsigned Offset = 0) { 1238 AvailableValueInBlock Res; 1239 Res.BB = BB; 1240 Res.Val.setPointer(V); 1241 Res.Val.setInt(SimpleVal); 1242 Res.Offset = Offset; 1243 return Res; 1244 } 1245 1246 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1247 unsigned Offset = 0) { 1248 AvailableValueInBlock Res; 1249 Res.BB = BB; 1250 Res.Val.setPointer(MI); 1251 Res.Val.setInt(MemIntrin); 1252 Res.Offset = Offset; 1253 return Res; 1254 } 1255 1256 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1257 Value *getSimpleValue() const { 1258 assert(isSimpleValue() && "Wrong accessor"); 1259 return Val.getPointer(); 1260 } 1261 1262 MemIntrinsic *getMemIntrinValue() const { 1263 assert(!isSimpleValue() && "Wrong accessor"); 1264 return cast<MemIntrinsic>(Val.getPointer()); 1265 } 1266 1267 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1268 /// defined here to the specified type. This handles various coercion cases. 1269 Value *MaterializeAdjustedValue(const Type *LoadTy, 1270 const TargetData *TD) const { 1271 Value *Res; 1272 if (isSimpleValue()) { 1273 Res = getSimpleValue(); 1274 if (Res->getType() != LoadTy) { 1275 assert(TD && "Need target data to handle type mismatch case"); 1276 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1277 *TD); 1278 1279 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1280 << *getSimpleValue() << '\n' 1281 << *Res << '\n' << "\n\n\n"); 1282 } 1283 } else { 1284 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1285 LoadTy, BB->getTerminator(), *TD); 1286 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1287 << " " << *getMemIntrinValue() << '\n' 1288 << *Res << '\n' << "\n\n\n"); 1289 } 1290 return Res; 1291 } 1292}; 1293 1294/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1295/// construct SSA form, allowing us to eliminate LI. This returns the value 1296/// that should be used at LI's definition site. 1297static Value *ConstructSSAForLoadSet(LoadInst *LI, 1298 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1299 const TargetData *TD, 1300 const DominatorTree &DT, 1301 AliasAnalysis *AA) { 1302 // Check for the fully redundant, dominating load case. In this case, we can 1303 // just use the dominating value directly. 1304 if (ValuesPerBlock.size() == 1 && 1305 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent())) 1306 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD); 1307 1308 // Otherwise, we have to construct SSA form. 1309 SmallVector<PHINode*, 8> NewPHIs; 1310 SSAUpdater SSAUpdate(&NewPHIs); 1311 SSAUpdate.Initialize(LI); 1312 1313 const Type *LoadTy = LI->getType(); 1314 1315 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1316 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1317 BasicBlock *BB = AV.BB; 1318 1319 if (SSAUpdate.HasValueForBlock(BB)) 1320 continue; 1321 1322 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD)); 1323 } 1324 1325 // Perform PHI construction. 1326 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1327 1328 // If new PHI nodes were created, notify alias analysis. 1329 if (V->getType()->isPointerTy()) 1330 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1331 AA->copyValue(LI, NewPHIs[i]); 1332 1333 return V; 1334} 1335 1336static bool isLifetimeStart(Instruction *Inst) { 1337 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1338 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1339 return false; 1340} 1341 1342/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1343/// non-local by performing PHI construction. 1344bool GVN::processNonLocalLoad(LoadInst *LI, 1345 SmallVectorImpl<Instruction*> &toErase) { 1346 // Find the non-local dependencies of the load. 1347 SmallVector<NonLocalDepResult, 64> Deps; 1348 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1349 Deps); 1350 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1351 // << Deps.size() << *LI << '\n'); 1352 1353 // If we had to process more than one hundred blocks to find the 1354 // dependencies, this load isn't worth worrying about. Optimizing 1355 // it will be too expensive. 1356 if (Deps.size() > 100) 1357 return false; 1358 1359 // If we had a phi translation failure, we'll have a single entry which is a 1360 // clobber in the current block. Reject this early. 1361 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) { 1362 DEBUG( 1363 dbgs() << "GVN: non-local load "; 1364 WriteAsOperand(dbgs(), LI); 1365 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1366 ); 1367 return false; 1368 } 1369 1370 // Filter out useless results (non-locals, etc). Keep track of the blocks 1371 // where we have a value available in repl, also keep track of whether we see 1372 // dependencies that produce an unknown value for the load (such as a call 1373 // that could potentially clobber the load). 1374 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1375 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1376 1377 const TargetData *TD = 0; 1378 1379 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1380 BasicBlock *DepBB = Deps[i].getBB(); 1381 MemDepResult DepInfo = Deps[i].getResult(); 1382 1383 if (DepInfo.isClobber()) { 1384 // The address being loaded in this non-local block may not be the same as 1385 // the pointer operand of the load if PHI translation occurs. Make sure 1386 // to consider the right address. 1387 Value *Address = Deps[i].getAddress(); 1388 1389 // If the dependence is to a store that writes to a superset of the bits 1390 // read by the load, we can extract the bits we need for the load from the 1391 // stored value. 1392 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1393 if (TD == 0) 1394 TD = getAnalysisIfAvailable<TargetData>(); 1395 if (TD && Address) { 1396 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1397 DepSI, *TD); 1398 if (Offset != -1) { 1399 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1400 DepSI->getOperand(0), 1401 Offset)); 1402 continue; 1403 } 1404 } 1405 } 1406 1407 // If the clobbering value is a memset/memcpy/memmove, see if we can 1408 // forward a value on from it. 1409 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1410 if (TD == 0) 1411 TD = getAnalysisIfAvailable<TargetData>(); 1412 if (TD && Address) { 1413 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1414 DepMI, *TD); 1415 if (Offset != -1) { 1416 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1417 Offset)); 1418 continue; 1419 } 1420 } 1421 } 1422 1423 UnavailableBlocks.push_back(DepBB); 1424 continue; 1425 } 1426 1427 Instruction *DepInst = DepInfo.getInst(); 1428 1429 // Loading the allocation -> undef. 1430 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1431 // Loading immediately after lifetime begin -> undef. 1432 isLifetimeStart(DepInst)) { 1433 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1434 UndefValue::get(LI->getType()))); 1435 continue; 1436 } 1437 1438 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1439 // Reject loads and stores that are to the same address but are of 1440 // different types if we have to. 1441 if (S->getOperand(0)->getType() != LI->getType()) { 1442 if (TD == 0) 1443 TD = getAnalysisIfAvailable<TargetData>(); 1444 1445 // If the stored value is larger or equal to the loaded value, we can 1446 // reuse it. 1447 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0), 1448 LI->getType(), *TD)) { 1449 UnavailableBlocks.push_back(DepBB); 1450 continue; 1451 } 1452 } 1453 1454 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1455 S->getOperand(0))); 1456 continue; 1457 } 1458 1459 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1460 // If the types mismatch and we can't handle it, reject reuse of the load. 1461 if (LD->getType() != LI->getType()) { 1462 if (TD == 0) 1463 TD = getAnalysisIfAvailable<TargetData>(); 1464 1465 // If the stored value is larger or equal to the loaded value, we can 1466 // reuse it. 1467 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1468 UnavailableBlocks.push_back(DepBB); 1469 continue; 1470 } 1471 } 1472 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1473 continue; 1474 } 1475 1476 UnavailableBlocks.push_back(DepBB); 1477 continue; 1478 } 1479 1480 // If we have no predecessors that produce a known value for this load, exit 1481 // early. 1482 if (ValuesPerBlock.empty()) return false; 1483 1484 // If all of the instructions we depend on produce a known value for this 1485 // load, then it is fully redundant and we can use PHI insertion to compute 1486 // its value. Insert PHIs and remove the fully redundant value now. 1487 if (UnavailableBlocks.empty()) { 1488 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1489 1490 // Perform PHI construction. 1491 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1492 VN.getAliasAnalysis()); 1493 LI->replaceAllUsesWith(V); 1494 1495 if (isa<PHINode>(V)) 1496 V->takeName(LI); 1497 if (V->getType()->isPointerTy()) 1498 MD->invalidateCachedPointerInfo(V); 1499 VN.erase(LI); 1500 toErase.push_back(LI); 1501 NumGVNLoad++; 1502 return true; 1503 } 1504 1505 if (!EnablePRE || !EnableLoadPRE) 1506 return false; 1507 1508 // Okay, we have *some* definitions of the value. This means that the value 1509 // is available in some of our (transitive) predecessors. Lets think about 1510 // doing PRE of this load. This will involve inserting a new load into the 1511 // predecessor when it's not available. We could do this in general, but 1512 // prefer to not increase code size. As such, we only do this when we know 1513 // that we only have to insert *one* load (which means we're basically moving 1514 // the load, not inserting a new one). 1515 1516 SmallPtrSet<BasicBlock *, 4> Blockers; 1517 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1518 Blockers.insert(UnavailableBlocks[i]); 1519 1520 // Lets find first basic block with more than one predecessor. Walk backwards 1521 // through predecessors if needed. 1522 BasicBlock *LoadBB = LI->getParent(); 1523 BasicBlock *TmpBB = LoadBB; 1524 1525 bool isSinglePred = false; 1526 bool allSingleSucc = true; 1527 while (TmpBB->getSinglePredecessor()) { 1528 isSinglePred = true; 1529 TmpBB = TmpBB->getSinglePredecessor(); 1530 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1531 return false; 1532 if (Blockers.count(TmpBB)) 1533 return false; 1534 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1535 allSingleSucc = false; 1536 } 1537 1538 assert(TmpBB); 1539 LoadBB = TmpBB; 1540 1541 // If we have a repl set with LI itself in it, this means we have a loop where 1542 // at least one of the values is LI. Since this means that we won't be able 1543 // to eliminate LI even if we insert uses in the other predecessors, we will 1544 // end up increasing code size. Reject this by scanning for LI. 1545 if (!EnableFullLoadPRE) { 1546 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1547 if (ValuesPerBlock[i].isSimpleValue() && 1548 ValuesPerBlock[i].getSimpleValue() == LI) 1549 return false; 1550 } 1551 1552 // FIXME: It is extremely unclear what this loop is doing, other than 1553 // artificially restricting loadpre. 1554 if (isSinglePred) { 1555 bool isHot = false; 1556 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1557 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1558 if (AV.isSimpleValue()) 1559 // "Hot" Instruction is in some loop (because it dominates its dep. 1560 // instruction). 1561 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1562 if (DT->dominates(LI, I)) { 1563 isHot = true; 1564 break; 1565 } 1566 } 1567 1568 // We are interested only in "hot" instructions. We don't want to do any 1569 // mis-optimizations here. 1570 if (!isHot) 1571 return false; 1572 } 1573 1574 // Check to see how many predecessors have the loaded value fully 1575 // available. 1576 DenseMap<BasicBlock*, Value*> PredLoads; 1577 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1578 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1579 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1580 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1581 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1582 1583 bool NeedToSplitEdges = false; 1584 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1585 PI != E; ++PI) { 1586 BasicBlock *Pred = *PI; 1587 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1588 continue; 1589 } 1590 PredLoads[Pred] = 0; 1591 1592 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1593 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1594 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1595 << Pred->getName() << "': " << *LI << '\n'); 1596 return false; 1597 } 1598 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1599 toSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1600 NeedToSplitEdges = true; 1601 } 1602 } 1603 if (NeedToSplitEdges) 1604 return false; 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 if (!EnableFullLoadPRE) { 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 1619 // Check if the load can safely be moved to all the unavailable predecessors. 1620 bool CanDoPRE = true; 1621 SmallVector<Instruction*, 8> NewInsts; 1622 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1623 E = PredLoads.end(); I != E; ++I) { 1624 BasicBlock *UnavailablePred = I->first; 1625 1626 // Do PHI translation to get its value in the predecessor if necessary. The 1627 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1628 1629 // If all preds have a single successor, then we know it is safe to insert 1630 // the load on the pred (?!?), so we can insert code to materialize the 1631 // pointer if it is not available. 1632 PHITransAddr Address(LI->getOperand(0), TD); 1633 Value *LoadPtr = 0; 1634 if (allSingleSucc) { 1635 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1636 *DT, NewInsts); 1637 } else { 1638 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1639 LoadPtr = Address.getAddr(); 1640 } 1641 1642 // If we couldn't find or insert a computation of this phi translated value, 1643 // we fail PRE. 1644 if (LoadPtr == 0) { 1645 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1646 << *LI->getOperand(0) << "\n"); 1647 CanDoPRE = false; 1648 break; 1649 } 1650 1651 // Make sure it is valid to move this load here. We have to watch out for: 1652 // @1 = getelementptr (i8* p, ... 1653 // test p and branch if == 0 1654 // load @1 1655 // It is valid to have the getelementptr before the test, even if p can be 0, 1656 // as getelementptr only does address arithmetic. 1657 // If we are not pushing the value through any multiple-successor blocks 1658 // we do not have this case. Otherwise, check that the load is safe to 1659 // put anywhere; this can be improved, but should be conservatively safe. 1660 if (!allSingleSucc && 1661 // FIXME: REEVALUTE THIS. 1662 !isSafeToLoadUnconditionally(LoadPtr, 1663 UnavailablePred->getTerminator(), 1664 LI->getAlignment(), TD)) { 1665 CanDoPRE = false; 1666 break; 1667 } 1668 1669 I->second = LoadPtr; 1670 } 1671 1672 if (!CanDoPRE) { 1673 while (!NewInsts.empty()) 1674 NewInsts.pop_back_val()->eraseFromParent(); 1675 return false; 1676 } 1677 1678 // Okay, we can eliminate this load by inserting a reload in the predecessor 1679 // and using PHI construction to get the value in the other predecessors, do 1680 // it. 1681 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1682 DEBUG(if (!NewInsts.empty()) 1683 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1684 << *NewInsts.back() << '\n'); 1685 1686 // Assign value numbers to the new instructions. 1687 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1688 // FIXME: We really _ought_ to insert these value numbers into their 1689 // parent's availability map. However, in doing so, we risk getting into 1690 // ordering issues. If a block hasn't been processed yet, we would be 1691 // marking a value as AVAIL-IN, which isn't what we intend. 1692 VN.lookup_or_add(NewInsts[i]); 1693 } 1694 1695 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1696 E = PredLoads.end(); I != E; ++I) { 1697 BasicBlock *UnavailablePred = I->first; 1698 Value *LoadPtr = I->second; 1699 1700 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1701 LI->getAlignment(), 1702 UnavailablePred->getTerminator()); 1703 1704 // Add the newly created load. 1705 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1706 NewLoad)); 1707 MD->invalidateCachedPointerInfo(LoadPtr); 1708 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1709 } 1710 1711 // Perform PHI construction. 1712 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1713 VN.getAliasAnalysis()); 1714 LI->replaceAllUsesWith(V); 1715 if (isa<PHINode>(V)) 1716 V->takeName(LI); 1717 if (V->getType()->isPointerTy()) 1718 MD->invalidateCachedPointerInfo(V); 1719 VN.erase(LI); 1720 toErase.push_back(LI); 1721 NumPRELoad++; 1722 return true; 1723} 1724 1725/// processLoad - Attempt to eliminate a load, first by eliminating it 1726/// locally, and then attempting non-local elimination if that fails. 1727bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1728 if (!MD) 1729 return false; 1730 1731 if (L->isVolatile()) 1732 return false; 1733 1734 // ... to a pointer that has been loaded from before... 1735 MemDepResult Dep = MD->getDependency(L); 1736 1737 // If the value isn't available, don't do anything! 1738 if (Dep.isClobber()) { 1739 // Check to see if we have something like this: 1740 // store i32 123, i32* %P 1741 // %A = bitcast i32* %P to i8* 1742 // %B = gep i8* %A, i32 1 1743 // %C = load i8* %B 1744 // 1745 // We could do that by recognizing if the clobber instructions are obviously 1746 // a common base + constant offset, and if the previous store (or memset) 1747 // completely covers this load. This sort of thing can happen in bitfield 1748 // access code. 1749 Value *AvailVal = 0; 1750 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1751 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1752 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1753 L->getPointerOperand(), 1754 DepSI, *TD); 1755 if (Offset != -1) 1756 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1757 L->getType(), L, *TD); 1758 } 1759 1760 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1761 // a value on from it. 1762 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1763 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1764 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1765 L->getPointerOperand(), 1766 DepMI, *TD); 1767 if (Offset != -1) 1768 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD); 1769 } 1770 } 1771 1772 if (AvailVal) { 1773 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1774 << *AvailVal << '\n' << *L << "\n\n\n"); 1775 1776 // Replace the load! 1777 L->replaceAllUsesWith(AvailVal); 1778 if (AvailVal->getType()->isPointerTy()) 1779 MD->invalidateCachedPointerInfo(AvailVal); 1780 VN.erase(L); 1781 toErase.push_back(L); 1782 NumGVNLoad++; 1783 return true; 1784 } 1785 1786 DEBUG( 1787 // fast print dep, using operator<< on instruction would be too slow 1788 dbgs() << "GVN: load "; 1789 WriteAsOperand(dbgs(), L); 1790 Instruction *I = Dep.getInst(); 1791 dbgs() << " is clobbered by " << *I << '\n'; 1792 ); 1793 return false; 1794 } 1795 1796 // If it is defined in another block, try harder. 1797 if (Dep.isNonLocal()) 1798 return processNonLocalLoad(L, toErase); 1799 1800 Instruction *DepInst = Dep.getInst(); 1801 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1802 Value *StoredVal = DepSI->getOperand(0); 1803 1804 // The store and load are to a must-aliased pointer, but they may not 1805 // actually have the same type. See if we know how to reuse the stored 1806 // value (depending on its type). 1807 const TargetData *TD = 0; 1808 if (StoredVal->getType() != L->getType()) { 1809 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1810 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1811 L, *TD); 1812 if (StoredVal == 0) 1813 return false; 1814 1815 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1816 << '\n' << *L << "\n\n\n"); 1817 } 1818 else 1819 return false; 1820 } 1821 1822 // Remove it! 1823 L->replaceAllUsesWith(StoredVal); 1824 if (StoredVal->getType()->isPointerTy()) 1825 MD->invalidateCachedPointerInfo(StoredVal); 1826 VN.erase(L); 1827 toErase.push_back(L); 1828 NumGVNLoad++; 1829 return true; 1830 } 1831 1832 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1833 Value *AvailableVal = DepLI; 1834 1835 // The loads are of a must-aliased pointer, but they may not actually have 1836 // the same type. See if we know how to reuse the previously loaded value 1837 // (depending on its type). 1838 const TargetData *TD = 0; 1839 if (DepLI->getType() != L->getType()) { 1840 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1841 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1842 if (AvailableVal == 0) 1843 return false; 1844 1845 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1846 << "\n" << *L << "\n\n\n"); 1847 } 1848 else 1849 return false; 1850 } 1851 1852 // Remove it! 1853 L->replaceAllUsesWith(AvailableVal); 1854 if (DepLI->getType()->isPointerTy()) 1855 MD->invalidateCachedPointerInfo(DepLI); 1856 VN.erase(L); 1857 toErase.push_back(L); 1858 NumGVNLoad++; 1859 return true; 1860 } 1861 1862 // If this load really doesn't depend on anything, then we must be loading an 1863 // undef value. This can happen when loading for a fresh allocation with no 1864 // intervening stores, for example. 1865 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1866 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1867 VN.erase(L); 1868 toErase.push_back(L); 1869 NumGVNLoad++; 1870 return true; 1871 } 1872 1873 // If this load occurs either right after a lifetime begin, 1874 // then the loaded value is undefined. 1875 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) { 1876 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1877 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1878 VN.erase(L); 1879 toErase.push_back(L); 1880 NumGVNLoad++; 1881 return true; 1882 } 1883 } 1884 1885 return false; 1886} 1887 1888Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1889 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1890 if (I == localAvail.end()) 1891 return 0; 1892 1893 ValueNumberScope *Locals = I->second; 1894 while (Locals) { 1895 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1896 if (I != Locals->table.end()) 1897 return I->second; 1898 Locals = Locals->parent; 1899 } 1900 1901 return 0; 1902} 1903 1904 1905/// processInstruction - When calculating availability, handle an instruction 1906/// by inserting it into the appropriate sets 1907bool GVN::processInstruction(Instruction *I, 1908 SmallVectorImpl<Instruction*> &toErase) { 1909 // Ignore dbg info intrinsics. 1910 if (isa<DbgInfoIntrinsic>(I)) 1911 return false; 1912 1913 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1914 bool Changed = processLoad(LI, toErase); 1915 1916 if (!Changed) { 1917 unsigned Num = VN.lookup_or_add(LI); 1918 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1919 } 1920 1921 return Changed; 1922 } 1923 1924 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1925 unsigned Num = VN.lookup_or_add(I); 1926 1927 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1928 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1929 1930 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1931 return false; 1932 1933 Value *BranchCond = BI->getCondition(); 1934 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1935 1936 BasicBlock *TrueSucc = BI->getSuccessor(0); 1937 BasicBlock *FalseSucc = BI->getSuccessor(1); 1938 1939 if (TrueSucc->getSinglePredecessor()) 1940 localAvail[TrueSucc]->table[CondVN] = 1941 ConstantInt::getTrue(TrueSucc->getContext()); 1942 if (FalseSucc->getSinglePredecessor()) 1943 localAvail[FalseSucc]->table[CondVN] = 1944 ConstantInt::getFalse(TrueSucc->getContext()); 1945 1946 return false; 1947 1948 // Allocations are always uniquely numbered, so we can save time and memory 1949 // by fast failing them. 1950 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) { 1951 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1952 return false; 1953 } 1954 1955 // Collapse PHI nodes 1956 if (PHINode* p = dyn_cast<PHINode>(I)) { 1957 Value *constVal = CollapsePhi(p); 1958 1959 if (constVal) { 1960 p->replaceAllUsesWith(constVal); 1961 if (MD && constVal->getType()->isPointerTy()) 1962 MD->invalidateCachedPointerInfo(constVal); 1963 VN.erase(p); 1964 1965 toErase.push_back(p); 1966 } else { 1967 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1968 } 1969 1970 // If the number we were assigned was a brand new VN, then we don't 1971 // need to do a lookup to see if the number already exists 1972 // somewhere in the domtree: it can't! 1973 } else if (Num == NextNum) { 1974 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1975 1976 // Perform fast-path value-number based elimination of values inherited from 1977 // dominators. 1978 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1979 // Remove it! 1980 VN.erase(I); 1981 I->replaceAllUsesWith(repl); 1982 if (MD && repl->getType()->isPointerTy()) 1983 MD->invalidateCachedPointerInfo(repl); 1984 toErase.push_back(I); 1985 return true; 1986 1987 } else { 1988 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1989 } 1990 1991 return false; 1992} 1993 1994/// runOnFunction - This is the main transformation entry point for a function. 1995bool GVN::runOnFunction(Function& F) { 1996 if (!NoLoads) 1997 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1998 DT = &getAnalysis<DominatorTree>(); 1999 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2000 VN.setMemDep(MD); 2001 VN.setDomTree(DT); 2002 2003 bool Changed = false; 2004 bool ShouldContinue = true; 2005 2006 // Merge unconditional branches, allowing PRE to catch more 2007 // optimization opportunities. 2008 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2009 BasicBlock *BB = FI; 2010 ++FI; 2011 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2012 if (removedBlock) NumGVNBlocks++; 2013 2014 Changed |= removedBlock; 2015 } 2016 2017 unsigned Iteration = 0; 2018 2019 while (ShouldContinue) { 2020 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2021 ShouldContinue = iterateOnFunction(F); 2022 if (splitCriticalEdges()) 2023 ShouldContinue = true; 2024 Changed |= ShouldContinue; 2025 ++Iteration; 2026 } 2027 2028 if (EnablePRE) { 2029 bool PREChanged = true; 2030 while (PREChanged) { 2031 PREChanged = performPRE(F); 2032 Changed |= PREChanged; 2033 } 2034 } 2035 // FIXME: Should perform GVN again after PRE does something. PRE can move 2036 // computations into blocks where they become fully redundant. Note that 2037 // we can't do this until PRE's critical edge splitting updates memdep. 2038 // Actually, when this happens, we should just fully integrate PRE into GVN. 2039 2040 cleanupGlobalSets(); 2041 2042 return Changed; 2043} 2044 2045 2046bool GVN::processBlock(BasicBlock *BB) { 2047 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 2048 // incrementing BI before processing an instruction). 2049 SmallVector<Instruction*, 8> toErase; 2050 bool ChangedFunction = false; 2051 2052 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2053 BI != BE;) { 2054 ChangedFunction |= processInstruction(BI, toErase); 2055 if (toErase.empty()) { 2056 ++BI; 2057 continue; 2058 } 2059 2060 // If we need some instructions deleted, do it now. 2061 NumGVNInstr += toErase.size(); 2062 2063 // Avoid iterator invalidation. 2064 bool AtStart = BI == BB->begin(); 2065 if (!AtStart) 2066 --BI; 2067 2068 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 2069 E = toErase.end(); I != E; ++I) { 2070 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2071 if (MD) MD->removeInstruction(*I); 2072 (*I)->eraseFromParent(); 2073 DEBUG(verifyRemoved(*I)); 2074 } 2075 toErase.clear(); 2076 2077 if (AtStart) 2078 BI = BB->begin(); 2079 else 2080 ++BI; 2081 } 2082 2083 return ChangedFunction; 2084} 2085 2086/// performPRE - Perform a purely local form of PRE that looks for diamond 2087/// control flow patterns and attempts to perform simple PRE at the join point. 2088bool GVN::performPRE(Function &F) { 2089 bool Changed = false; 2090 DenseMap<BasicBlock*, Value*> predMap; 2091 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2092 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2093 BasicBlock *CurrentBlock = *DI; 2094 2095 // Nothing to PRE in the entry block. 2096 if (CurrentBlock == &F.getEntryBlock()) continue; 2097 2098 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2099 BE = CurrentBlock->end(); BI != BE; ) { 2100 Instruction *CurInst = BI++; 2101 2102 if (isa<AllocaInst>(CurInst) || 2103 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2104 CurInst->getType()->isVoidTy() || 2105 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2106 isa<DbgInfoIntrinsic>(CurInst)) 2107 continue; 2108 2109 uint32_t ValNo = VN.lookup(CurInst); 2110 2111 // Look for the predecessors for PRE opportunities. We're 2112 // only trying to solve the basic diamond case, where 2113 // a value is computed in the successor and one predecessor, 2114 // but not the other. We also explicitly disallow cases 2115 // where the successor is its own predecessor, because they're 2116 // more complicated to get right. 2117 unsigned NumWith = 0; 2118 unsigned NumWithout = 0; 2119 BasicBlock *PREPred = 0; 2120 predMap.clear(); 2121 2122 for (pred_iterator PI = pred_begin(CurrentBlock), 2123 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2124 // We're not interested in PRE where the block is its 2125 // own predecessor, or in blocks with predecessors 2126 // that are not reachable. 2127 if (*PI == CurrentBlock) { 2128 NumWithout = 2; 2129 break; 2130 } else if (!localAvail.count(*PI)) { 2131 NumWithout = 2; 2132 break; 2133 } 2134 2135 DenseMap<uint32_t, Value*>::iterator predV = 2136 localAvail[*PI]->table.find(ValNo); 2137 if (predV == localAvail[*PI]->table.end()) { 2138 PREPred = *PI; 2139 NumWithout++; 2140 } else if (predV->second == CurInst) { 2141 NumWithout = 2; 2142 } else { 2143 predMap[*PI] = predV->second; 2144 NumWith++; 2145 } 2146 } 2147 2148 // Don't do PRE when it might increase code size, i.e. when 2149 // we would need to insert instructions in more than one pred. 2150 if (NumWithout != 1 || NumWith == 0) 2151 continue; 2152 2153 // Don't do PRE across indirect branch. 2154 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2155 continue; 2156 2157 // We can't do PRE safely on a critical edge, so instead we schedule 2158 // the edge to be split and perform the PRE the next time we iterate 2159 // on the function. 2160 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2161 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2162 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2163 continue; 2164 } 2165 2166 // Instantiate the expression in the predecessor that lacked it. 2167 // Because we are going top-down through the block, all value numbers 2168 // will be available in the predecessor by the time we need them. Any 2169 // that weren't originally present will have been instantiated earlier 2170 // in this loop. 2171 Instruction *PREInstr = CurInst->clone(); 2172 bool success = true; 2173 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2174 Value *Op = PREInstr->getOperand(i); 2175 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2176 continue; 2177 2178 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 2179 PREInstr->setOperand(i, V); 2180 } else { 2181 success = false; 2182 break; 2183 } 2184 } 2185 2186 // Fail out if we encounter an operand that is not available in 2187 // the PRE predecessor. This is typically because of loads which 2188 // are not value numbered precisely. 2189 if (!success) { 2190 delete PREInstr; 2191 DEBUG(verifyRemoved(PREInstr)); 2192 continue; 2193 } 2194 2195 PREInstr->insertBefore(PREPred->getTerminator()); 2196 PREInstr->setName(CurInst->getName() + ".pre"); 2197 predMap[PREPred] = PREInstr; 2198 VN.add(PREInstr, ValNo); 2199 NumGVNPRE++; 2200 2201 // Update the availability map to include the new instruction. 2202 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 2203 2204 // Create a PHI to make the value available in this block. 2205 PHINode* Phi = PHINode::Create(CurInst->getType(), 2206 CurInst->getName() + ".pre-phi", 2207 CurrentBlock->begin()); 2208 for (pred_iterator PI = pred_begin(CurrentBlock), 2209 PE = pred_end(CurrentBlock); PI != PE; ++PI) 2210 Phi->addIncoming(predMap[*PI], *PI); 2211 2212 VN.add(Phi, ValNo); 2213 localAvail[CurrentBlock]->table[ValNo] = Phi; 2214 2215 CurInst->replaceAllUsesWith(Phi); 2216 if (MD && Phi->getType()->isPointerTy()) 2217 MD->invalidateCachedPointerInfo(Phi); 2218 VN.erase(CurInst); 2219 2220 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2221 if (MD) MD->removeInstruction(CurInst); 2222 CurInst->eraseFromParent(); 2223 DEBUG(verifyRemoved(CurInst)); 2224 Changed = true; 2225 } 2226 } 2227 2228 if (splitCriticalEdges()) 2229 Changed = true; 2230 2231 return Changed; 2232} 2233 2234/// splitCriticalEdges - Split critical edges found during the previous 2235/// iteration that may enable further optimization. 2236bool GVN::splitCriticalEdges() { 2237 if (toSplit.empty()) 2238 return false; 2239 do { 2240 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2241 SplitCriticalEdge(Edge.first, Edge.second, this); 2242 } while (!toSplit.empty()); 2243 if (MD) MD->invalidateCachedPredecessors(); 2244 return true; 2245} 2246 2247/// iterateOnFunction - Executes one iteration of GVN 2248bool GVN::iterateOnFunction(Function &F) { 2249 cleanupGlobalSets(); 2250 2251 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2252 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 2253 if (DI->getIDom()) 2254 localAvail[DI->getBlock()] = 2255 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 2256 else 2257 localAvail[DI->getBlock()] = new ValueNumberScope(0); 2258 } 2259 2260 // Top-down walk of the dominator tree 2261 bool Changed = false; 2262#if 0 2263 // Needed for value numbering with phi construction to work. 2264 ReversePostOrderTraversal<Function*> RPOT(&F); 2265 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2266 RE = RPOT.end(); RI != RE; ++RI) 2267 Changed |= processBlock(*RI); 2268#else 2269 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2270 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2271 Changed |= processBlock(DI->getBlock()); 2272#endif 2273 2274 return Changed; 2275} 2276 2277void GVN::cleanupGlobalSets() { 2278 VN.clear(); 2279 2280 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2281 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 2282 delete I->second; 2283 localAvail.clear(); 2284} 2285 2286/// verifyRemoved - Verify that the specified instruction does not occur in our 2287/// internal data structures. 2288void GVN::verifyRemoved(const Instruction *Inst) const { 2289 VN.verifyRemoved(Inst); 2290 2291 // Walk through the value number scope to make sure the instruction isn't 2292 // ferreted away in it. 2293 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator 2294 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 2295 const ValueNumberScope *VNS = I->second; 2296 2297 while (VNS) { 2298 for (DenseMap<uint32_t, Value*>::const_iterator 2299 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 2300 assert(II->second != Inst && "Inst still in value numbering scope!"); 2301 } 2302 2303 VNS = VNS->parent; 2304 } 2305 } 2306} 2307