GVN.cpp revision 5015b3417fcf5b415fa9a1e5b111ea6f8e96fbfa
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)); 64static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false)); 65 66//===----------------------------------------------------------------------===// 67// ValueTable Class 68//===----------------------------------------------------------------------===// 69 70/// This class holds the mapping between values and value numbers. It is used 71/// as an efficient mechanism to determine the expression-wise equivalence of 72/// two values. 73namespace { 74 struct Expression { 75 enum ExpressionOpcode { 76 ADD = Instruction::Add, 77 FADD = Instruction::FAdd, 78 SUB = Instruction::Sub, 79 FSUB = Instruction::FSub, 80 MUL = Instruction::Mul, 81 FMUL = Instruction::FMul, 82 UDIV = Instruction::UDiv, 83 SDIV = Instruction::SDiv, 84 FDIV = Instruction::FDiv, 85 UREM = Instruction::URem, 86 SREM = Instruction::SRem, 87 FREM = Instruction::FRem, 88 SHL = Instruction::Shl, 89 LSHR = Instruction::LShr, 90 ASHR = Instruction::AShr, 91 AND = Instruction::And, 92 OR = Instruction::Or, 93 XOR = Instruction::Xor, 94 TRUNC = Instruction::Trunc, 95 ZEXT = Instruction::ZExt, 96 SEXT = Instruction::SExt, 97 FPTOUI = Instruction::FPToUI, 98 FPTOSI = Instruction::FPToSI, 99 UITOFP = Instruction::UIToFP, 100 SITOFP = Instruction::SIToFP, 101 FPTRUNC = Instruction::FPTrunc, 102 FPEXT = Instruction::FPExt, 103 PTRTOINT = Instruction::PtrToInt, 104 INTTOPTR = Instruction::IntToPtr, 105 BITCAST = Instruction::BitCast, 106 ICMPEQ, ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 107 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 108 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 109 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 110 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 111 SHUFFLE, SELECT, GEP, CALL, CONSTANT, 112 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE }; 113 114 ExpressionOpcode opcode; 115 const Type* type; 116 SmallVector<uint32_t, 4> varargs; 117 Value *function; 118 119 Expression() { } 120 Expression(ExpressionOpcode o) : opcode(o) { } 121 122 bool operator==(const Expression &other) const { 123 if (opcode != other.opcode) 124 return false; 125 else if (opcode == EMPTY || opcode == TOMBSTONE) 126 return true; 127 else if (type != other.type) 128 return false; 129 else if (function != other.function) 130 return false; 131 else { 132 if (varargs.size() != other.varargs.size()) 133 return false; 134 135 for (size_t i = 0; i < varargs.size(); ++i) 136 if (varargs[i] != other.varargs[i]) 137 return false; 138 139 return true; 140 } 141 } 142 143 bool operator!=(const Expression &other) const { 144 return !(*this == other); 145 } 146 }; 147 148 class ValueTable { 149 private: 150 DenseMap<Value*, uint32_t> valueNumbering; 151 DenseMap<Expression, uint32_t> expressionNumbering; 152 AliasAnalysis* AA; 153 MemoryDependenceAnalysis* MD; 154 DominatorTree* DT; 155 156 uint32_t nextValueNumber; 157 158 Expression::ExpressionOpcode getOpcode(CmpInst* C); 159 Expression create_expression(BinaryOperator* BO); 160 Expression create_expression(CmpInst* C); 161 Expression create_expression(ShuffleVectorInst* V); 162 Expression create_expression(ExtractElementInst* C); 163 Expression create_expression(InsertElementInst* V); 164 Expression create_expression(SelectInst* V); 165 Expression create_expression(CastInst* C); 166 Expression create_expression(GetElementPtrInst* G); 167 Expression create_expression(CallInst* C); 168 Expression create_expression(Constant* C); 169 Expression create_expression(ExtractValueInst* C); 170 Expression create_expression(InsertValueInst* C); 171 172 uint32_t lookup_or_add_call(CallInst* C); 173 public: 174 ValueTable() : nextValueNumber(1) { } 175 uint32_t lookup_or_add(Value *V); 176 uint32_t lookup(Value *V) const; 177 void add(Value *V, uint32_t num); 178 void clear(); 179 void erase(Value *v); 180 unsigned size(); 181 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 182 AliasAnalysis *getAliasAnalysis() const { return AA; } 183 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 184 void setDomTree(DominatorTree* D) { DT = D; } 185 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 186 void verifyRemoved(const Value *) const; 187 }; 188} 189 190namespace llvm { 191template <> struct DenseMapInfo<Expression> { 192 static inline Expression getEmptyKey() { 193 return Expression(Expression::EMPTY); 194 } 195 196 static inline Expression getTombstoneKey() { 197 return Expression(Expression::TOMBSTONE); 198 } 199 200 static unsigned getHashValue(const Expression e) { 201 unsigned hash = e.opcode; 202 203 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 204 (unsigned)((uintptr_t)e.type >> 9)); 205 206 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 207 E = e.varargs.end(); I != E; ++I) 208 hash = *I + hash * 37; 209 210 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 211 (unsigned)((uintptr_t)e.function >> 9)) + 212 hash * 37; 213 214 return hash; 215 } 216 static bool isEqual(const Expression &LHS, const Expression &RHS) { 217 return LHS == RHS; 218 } 219}; 220 221template <> 222struct isPodLike<Expression> { static const bool value = true; }; 223 224} 225 226//===----------------------------------------------------------------------===// 227// ValueTable Internal Functions 228//===----------------------------------------------------------------------===// 229 230Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 231 if (isa<ICmpInst>(C)) { 232 switch (C->getPredicate()) { 233 default: // THIS SHOULD NEVER HAPPEN 234 llvm_unreachable("Comparison with unknown predicate?"); 235 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 236 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 237 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 238 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 239 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 240 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 241 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 242 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 243 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 244 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 245 } 246 } else { 247 switch (C->getPredicate()) { 248 default: // THIS SHOULD NEVER HAPPEN 249 llvm_unreachable("Comparison with unknown predicate?"); 250 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 251 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 252 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 253 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 254 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 255 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 256 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 257 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 258 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 259 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 260 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 261 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 262 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 263 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 264 } 265 } 266} 267 268Expression ValueTable::create_expression(CallInst* C) { 269 Expression e; 270 271 e.type = C->getType(); 272 e.function = C->getCalledFunction(); 273 e.opcode = Expression::CALL; 274 275 CallSite CS(C); 276 for (CallInst::op_iterator I = CS.arg_begin(), E = CS.arg_end(); 277 I != E; ++I) 278 e.varargs.push_back(lookup_or_add(*I)); 279 280 return e; 281} 282 283Expression ValueTable::create_expression(BinaryOperator* BO) { 284 Expression e; 285 e.varargs.push_back(lookup_or_add(BO->getOperand(0))); 286 e.varargs.push_back(lookup_or_add(BO->getOperand(1))); 287 e.function = 0; 288 e.type = BO->getType(); 289 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode()); 290 291 return e; 292} 293 294Expression ValueTable::create_expression(CmpInst* C) { 295 Expression e; 296 297 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 298 e.varargs.push_back(lookup_or_add(C->getOperand(1))); 299 e.function = 0; 300 e.type = C->getType(); 301 e.opcode = getOpcode(C); 302 303 return e; 304} 305 306Expression ValueTable::create_expression(CastInst* C) { 307 Expression e; 308 309 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 310 e.function = 0; 311 e.type = C->getType(); 312 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode()); 313 314 return e; 315} 316 317Expression ValueTable::create_expression(ShuffleVectorInst* S) { 318 Expression e; 319 320 e.varargs.push_back(lookup_or_add(S->getOperand(0))); 321 e.varargs.push_back(lookup_or_add(S->getOperand(1))); 322 e.varargs.push_back(lookup_or_add(S->getOperand(2))); 323 e.function = 0; 324 e.type = S->getType(); 325 e.opcode = Expression::SHUFFLE; 326 327 return e; 328} 329 330Expression ValueTable::create_expression(ExtractElementInst* E) { 331 Expression e; 332 333 e.varargs.push_back(lookup_or_add(E->getOperand(0))); 334 e.varargs.push_back(lookup_or_add(E->getOperand(1))); 335 e.function = 0; 336 e.type = E->getType(); 337 e.opcode = Expression::EXTRACT; 338 339 return e; 340} 341 342Expression ValueTable::create_expression(InsertElementInst* I) { 343 Expression e; 344 345 e.varargs.push_back(lookup_or_add(I->getOperand(0))); 346 e.varargs.push_back(lookup_or_add(I->getOperand(1))); 347 e.varargs.push_back(lookup_or_add(I->getOperand(2))); 348 e.function = 0; 349 e.type = I->getType(); 350 e.opcode = Expression::INSERT; 351 352 return e; 353} 354 355Expression ValueTable::create_expression(SelectInst* I) { 356 Expression e; 357 358 e.varargs.push_back(lookup_or_add(I->getCondition())); 359 e.varargs.push_back(lookup_or_add(I->getTrueValue())); 360 e.varargs.push_back(lookup_or_add(I->getFalseValue())); 361 e.function = 0; 362 e.type = I->getType(); 363 e.opcode = Expression::SELECT; 364 365 return e; 366} 367 368Expression ValueTable::create_expression(GetElementPtrInst* G) { 369 Expression e; 370 371 e.varargs.push_back(lookup_or_add(G->getPointerOperand())); 372 e.function = 0; 373 e.type = G->getType(); 374 e.opcode = Expression::GEP; 375 376 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 377 I != E; ++I) 378 e.varargs.push_back(lookup_or_add(*I)); 379 380 return e; 381} 382 383Expression ValueTable::create_expression(ExtractValueInst* E) { 384 Expression e; 385 386 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 387 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 388 II != IE; ++II) 389 e.varargs.push_back(*II); 390 e.function = 0; 391 e.type = E->getType(); 392 e.opcode = Expression::EXTRACTVALUE; 393 394 return e; 395} 396 397Expression ValueTable::create_expression(InsertValueInst* E) { 398 Expression e; 399 400 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 401 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand())); 402 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 403 II != IE; ++II) 404 e.varargs.push_back(*II); 405 e.function = 0; 406 e.type = E->getType(); 407 e.opcode = Expression::INSERTVALUE; 408 409 return e; 410} 411 412//===----------------------------------------------------------------------===// 413// ValueTable External Functions 414//===----------------------------------------------------------------------===// 415 416/// add - Insert a value into the table with a specified value number. 417void ValueTable::add(Value *V, uint32_t num) { 418 valueNumbering.insert(std::make_pair(V, num)); 419} 420 421uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 422 if (AA->doesNotAccessMemory(C)) { 423 Expression exp = create_expression(C); 424 uint32_t& e = expressionNumbering[exp]; 425 if (!e) e = nextValueNumber++; 426 valueNumbering[C] = e; 427 return e; 428 } else if (AA->onlyReadsMemory(C)) { 429 Expression exp = create_expression(C); 430 uint32_t& e = expressionNumbering[exp]; 431 if (!e) { 432 e = nextValueNumber++; 433 valueNumbering[C] = e; 434 return e; 435 } 436 if (!MD) { 437 e = nextValueNumber++; 438 valueNumbering[C] = e; 439 return e; 440 } 441 442 MemDepResult local_dep = MD->getDependency(C); 443 444 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 445 valueNumbering[C] = nextValueNumber; 446 return nextValueNumber++; 447 } 448 449 if (local_dep.isDef()) { 450 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 451 452 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 453 valueNumbering[C] = nextValueNumber; 454 return nextValueNumber++; 455 } 456 457 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 458 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 459 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 460 if (c_vn != cd_vn) { 461 valueNumbering[C] = nextValueNumber; 462 return nextValueNumber++; 463 } 464 } 465 466 uint32_t v = lookup_or_add(local_cdep); 467 valueNumbering[C] = v; 468 return v; 469 } 470 471 // Non-local case. 472 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 473 MD->getNonLocalCallDependency(CallSite(C)); 474 // FIXME: call/call dependencies for readonly calls should return def, not 475 // clobber! Move the checking logic to MemDep! 476 CallInst* cdep = 0; 477 478 // Check to see if we have a single dominating call instruction that is 479 // identical to C. 480 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 481 const NonLocalDepEntry *I = &deps[i]; 482 // Ignore non-local dependencies. 483 if (I->getResult().isNonLocal()) 484 continue; 485 486 // We don't handle non-depedencies. If we already have a call, reject 487 // instruction dependencies. 488 if (I->getResult().isClobber() || cdep != 0) { 489 cdep = 0; 490 break; 491 } 492 493 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 494 // FIXME: All duplicated with non-local case. 495 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 496 cdep = NonLocalDepCall; 497 continue; 498 } 499 500 cdep = 0; 501 break; 502 } 503 504 if (!cdep) { 505 valueNumbering[C] = nextValueNumber; 506 return nextValueNumber++; 507 } 508 509 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 510 valueNumbering[C] = nextValueNumber; 511 return nextValueNumber++; 512 } 513 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 514 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 515 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 516 if (c_vn != cd_vn) { 517 valueNumbering[C] = nextValueNumber; 518 return nextValueNumber++; 519 } 520 } 521 522 uint32_t v = lookup_or_add(cdep); 523 valueNumbering[C] = v; 524 return v; 525 526 } else { 527 valueNumbering[C] = nextValueNumber; 528 return nextValueNumber++; 529 } 530} 531 532/// lookup_or_add - Returns the value number for the specified value, assigning 533/// it a new number if it did not have one before. 534uint32_t ValueTable::lookup_or_add(Value *V) { 535 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 536 if (VI != valueNumbering.end()) 537 return VI->second; 538 539 if (!isa<Instruction>(V)) { 540 valueNumbering[V] = nextValueNumber; 541 return nextValueNumber++; 542 } 543 544 Instruction* I = cast<Instruction>(V); 545 Expression exp; 546 switch (I->getOpcode()) { 547 case Instruction::Call: 548 return lookup_or_add_call(cast<CallInst>(I)); 549 case Instruction::Add: 550 case Instruction::FAdd: 551 case Instruction::Sub: 552 case Instruction::FSub: 553 case Instruction::Mul: 554 case Instruction::FMul: 555 case Instruction::UDiv: 556 case Instruction::SDiv: 557 case Instruction::FDiv: 558 case Instruction::URem: 559 case Instruction::SRem: 560 case Instruction::FRem: 561 case Instruction::Shl: 562 case Instruction::LShr: 563 case Instruction::AShr: 564 case Instruction::And: 565 case Instruction::Or : 566 case Instruction::Xor: 567 exp = create_expression(cast<BinaryOperator>(I)); 568 break; 569 case Instruction::ICmp: 570 case Instruction::FCmp: 571 exp = create_expression(cast<CmpInst>(I)); 572 break; 573 case Instruction::Trunc: 574 case Instruction::ZExt: 575 case Instruction::SExt: 576 case Instruction::FPToUI: 577 case Instruction::FPToSI: 578 case Instruction::UIToFP: 579 case Instruction::SIToFP: 580 case Instruction::FPTrunc: 581 case Instruction::FPExt: 582 case Instruction::PtrToInt: 583 case Instruction::IntToPtr: 584 case Instruction::BitCast: 585 exp = create_expression(cast<CastInst>(I)); 586 break; 587 case Instruction::Select: 588 exp = create_expression(cast<SelectInst>(I)); 589 break; 590 case Instruction::ExtractElement: 591 exp = create_expression(cast<ExtractElementInst>(I)); 592 break; 593 case Instruction::InsertElement: 594 exp = create_expression(cast<InsertElementInst>(I)); 595 break; 596 case Instruction::ShuffleVector: 597 exp = create_expression(cast<ShuffleVectorInst>(I)); 598 break; 599 case Instruction::ExtractValue: 600 exp = create_expression(cast<ExtractValueInst>(I)); 601 break; 602 case Instruction::InsertValue: 603 exp = create_expression(cast<InsertValueInst>(I)); 604 break; 605 case Instruction::GetElementPtr: 606 exp = create_expression(cast<GetElementPtrInst>(I)); 607 break; 608 default: 609 valueNumbering[V] = nextValueNumber; 610 return nextValueNumber++; 611 } 612 613 uint32_t& e = expressionNumbering[exp]; 614 if (!e) e = nextValueNumber++; 615 valueNumbering[V] = e; 616 return e; 617} 618 619/// lookup - Returns the value number of the specified value. Fails if 620/// the value has not yet been numbered. 621uint32_t ValueTable::lookup(Value *V) const { 622 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 623 assert(VI != valueNumbering.end() && "Value not numbered?"); 624 return VI->second; 625} 626 627/// clear - Remove all entries from the ValueTable 628void ValueTable::clear() { 629 valueNumbering.clear(); 630 expressionNumbering.clear(); 631 nextValueNumber = 1; 632} 633 634/// erase - Remove a value from the value numbering 635void ValueTable::erase(Value *V) { 636 valueNumbering.erase(V); 637} 638 639/// verifyRemoved - Verify that the value is removed from all internal data 640/// structures. 641void ValueTable::verifyRemoved(const Value *V) const { 642 for (DenseMap<Value*, uint32_t>::const_iterator 643 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 644 assert(I->first != V && "Inst still occurs in value numbering map!"); 645 } 646} 647 648//===----------------------------------------------------------------------===// 649// GVN Pass 650//===----------------------------------------------------------------------===// 651 652namespace { 653 struct ValueNumberScope { 654 ValueNumberScope* parent; 655 DenseMap<uint32_t, Value*> table; 656 657 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 658 }; 659} 660 661namespace { 662 663 class GVN : public FunctionPass { 664 bool runOnFunction(Function &F); 665 public: 666 static char ID; // Pass identification, replacement for typeid 667 explicit GVN(bool noloads = false) 668 : FunctionPass(ID), NoLoads(noloads), MD(0) { } 669 670 private: 671 bool NoLoads; 672 MemoryDependenceAnalysis *MD; 673 DominatorTree *DT; 674 675 ValueTable VN; 676 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 677 678 // List of critical edges to be split between iterations. 679 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 680 681 // This transformation requires dominator postdominator info 682 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 683 AU.addRequired<DominatorTree>(); 684 if (!NoLoads) 685 AU.addRequired<MemoryDependenceAnalysis>(); 686 AU.addRequired<AliasAnalysis>(); 687 688 AU.addPreserved<DominatorTree>(); 689 AU.addPreserved<AliasAnalysis>(); 690 } 691 692 // Helper fuctions 693 // FIXME: eliminate or document these better 694 bool processLoad(LoadInst* L, 695 SmallVectorImpl<Instruction*> &toErase); 696 bool processInstruction(Instruction *I, 697 SmallVectorImpl<Instruction*> &toErase); 698 bool processNonLocalLoad(LoadInst* L, 699 SmallVectorImpl<Instruction*> &toErase); 700 bool processBlock(BasicBlock *BB); 701 void dump(DenseMap<uint32_t, Value*>& d); 702 bool iterateOnFunction(Function &F); 703 Value *CollapsePhi(PHINode* p); 704 bool performPRE(Function& F); 705 Value *lookupNumber(BasicBlock *BB, uint32_t num); 706 void cleanupGlobalSets(); 707 void verifyRemoved(const Instruction *I) const; 708 bool splitCriticalEdges(); 709 }; 710 711 char GVN::ID = 0; 712} 713 714// createGVNPass - The public interface to this file... 715FunctionPass *llvm::createGVNPass(bool NoLoads) { 716 return new GVN(NoLoads); 717} 718 719INITIALIZE_PASS(GVN, "gvn", "Global Value Numbering", false, false); 720 721void GVN::dump(DenseMap<uint32_t, Value*>& d) { 722 errs() << "{\n"; 723 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 724 E = d.end(); I != E; ++I) { 725 errs() << I->first << "\n"; 726 I->second->dump(); 727 } 728 errs() << "}\n"; 729} 730 731static bool isSafeReplacement(PHINode* p, Instruction *inst) { 732 if (!isa<PHINode>(inst)) 733 return true; 734 735 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 736 UI != E; ++UI) 737 if (PHINode* use_phi = dyn_cast<PHINode>(*UI)) 738 if (use_phi->getParent() == inst->getParent()) 739 return false; 740 741 return true; 742} 743 744Value *GVN::CollapsePhi(PHINode *PN) { 745 Value *ConstVal = PN->hasConstantValue(DT); 746 if (!ConstVal) return 0; 747 748 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 749 if (!Inst) 750 return ConstVal; 751 752 if (DT->dominates(Inst, PN)) 753 if (isSafeReplacement(PN, Inst)) 754 return Inst; 755 return 0; 756} 757 758/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 759/// we're analyzing is fully available in the specified block. As we go, keep 760/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 761/// map is actually a tri-state map with the following values: 762/// 0) we know the block *is not* fully available. 763/// 1) we know the block *is* fully available. 764/// 2) we do not know whether the block is fully available or not, but we are 765/// currently speculating that it will be. 766/// 3) we are speculating for this block and have used that to speculate for 767/// other blocks. 768static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 769 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 770 // Optimistically assume that the block is fully available and check to see 771 // if we already know about this block in one lookup. 772 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 773 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 774 775 // If the entry already existed for this block, return the precomputed value. 776 if (!IV.second) { 777 // If this is a speculative "available" value, mark it as being used for 778 // speculation of other blocks. 779 if (IV.first->second == 2) 780 IV.first->second = 3; 781 return IV.first->second != 0; 782 } 783 784 // Otherwise, see if it is fully available in all predecessors. 785 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 786 787 // If this block has no predecessors, it isn't live-in here. 788 if (PI == PE) 789 goto SpeculationFailure; 790 791 for (; PI != PE; ++PI) 792 // If the value isn't fully available in one of our predecessors, then it 793 // isn't fully available in this block either. Undo our previous 794 // optimistic assumption and bail out. 795 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 796 goto SpeculationFailure; 797 798 return true; 799 800// SpeculationFailure - If we get here, we found out that this is not, after 801// all, a fully-available block. We have a problem if we speculated on this and 802// used the speculation to mark other blocks as available. 803SpeculationFailure: 804 char &BBVal = FullyAvailableBlocks[BB]; 805 806 // If we didn't speculate on this, just return with it set to false. 807 if (BBVal == 2) { 808 BBVal = 0; 809 return false; 810 } 811 812 // If we did speculate on this value, we could have blocks set to 1 that are 813 // incorrect. Walk the (transitive) successors of this block and mark them as 814 // 0 if set to one. 815 SmallVector<BasicBlock*, 32> BBWorklist; 816 BBWorklist.push_back(BB); 817 818 do { 819 BasicBlock *Entry = BBWorklist.pop_back_val(); 820 // Note that this sets blocks to 0 (unavailable) if they happen to not 821 // already be in FullyAvailableBlocks. This is safe. 822 char &EntryVal = FullyAvailableBlocks[Entry]; 823 if (EntryVal == 0) continue; // Already unavailable. 824 825 // Mark as unavailable. 826 EntryVal = 0; 827 828 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 829 BBWorklist.push_back(*I); 830 } while (!BBWorklist.empty()); 831 832 return false; 833} 834 835 836/// CanCoerceMustAliasedValueToLoad - Return true if 837/// CoerceAvailableValueToLoadType will succeed. 838static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 839 const Type *LoadTy, 840 const TargetData &TD) { 841 // If the loaded or stored value is an first class array or struct, don't try 842 // to transform them. We need to be able to bitcast to integer. 843 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 844 StoredVal->getType()->isStructTy() || 845 StoredVal->getType()->isArrayTy()) 846 return false; 847 848 // The store has to be at least as big as the load. 849 if (TD.getTypeSizeInBits(StoredVal->getType()) < 850 TD.getTypeSizeInBits(LoadTy)) 851 return false; 852 853 return true; 854} 855 856 857/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 858/// then a load from a must-aliased pointer of a different type, try to coerce 859/// the stored value. LoadedTy is the type of the load we want to replace and 860/// InsertPt is the place to insert new instructions. 861/// 862/// If we can't do it, return null. 863static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 864 const Type *LoadedTy, 865 Instruction *InsertPt, 866 const TargetData &TD) { 867 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 868 return 0; 869 870 const Type *StoredValTy = StoredVal->getType(); 871 872 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 873 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 874 875 // If the store and reload are the same size, we can always reuse it. 876 if (StoreSize == LoadSize) { 877 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) { 878 // Pointer to Pointer -> use bitcast. 879 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 880 } 881 882 // Convert source pointers to integers, which can be bitcast. 883 if (StoredValTy->isPointerTy()) { 884 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 885 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 886 } 887 888 const Type *TypeToCastTo = LoadedTy; 889 if (TypeToCastTo->isPointerTy()) 890 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 891 892 if (StoredValTy != TypeToCastTo) 893 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 894 895 // Cast to pointer if the load needs a pointer type. 896 if (LoadedTy->isPointerTy()) 897 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 898 899 return StoredVal; 900 } 901 902 // If the loaded value is smaller than the available value, then we can 903 // extract out a piece from it. If the available value is too small, then we 904 // can't do anything. 905 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 906 907 // Convert source pointers to integers, which can be manipulated. 908 if (StoredValTy->isPointerTy()) { 909 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 910 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 911 } 912 913 // Convert vectors and fp to integer, which can be manipulated. 914 if (!StoredValTy->isIntegerTy()) { 915 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 916 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 917 } 918 919 // If this is a big-endian system, we need to shift the value down to the low 920 // bits so that a truncate will work. 921 if (TD.isBigEndian()) { 922 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 923 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 924 } 925 926 // Truncate the integer to the right size now. 927 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 928 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 929 930 if (LoadedTy == NewIntTy) 931 return StoredVal; 932 933 // If the result is a pointer, inttoptr. 934 if (LoadedTy->isPointerTy()) 935 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 936 937 // Otherwise, bitcast. 938 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 939} 940 941/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 942/// be expressed as a base pointer plus a constant offset. Return the base and 943/// offset to the caller. 944static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 945 const TargetData &TD) { 946 Operator *PtrOp = dyn_cast<Operator>(Ptr); 947 if (PtrOp == 0) return Ptr; 948 949 // Just look through bitcasts. 950 if (PtrOp->getOpcode() == Instruction::BitCast) 951 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 952 953 // If this is a GEP with constant indices, we can look through it. 954 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 955 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 956 957 gep_type_iterator GTI = gep_type_begin(GEP); 958 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 959 ++I, ++GTI) { 960 ConstantInt *OpC = cast<ConstantInt>(*I); 961 if (OpC->isZero()) continue; 962 963 // Handle a struct and array indices which add their offset to the pointer. 964 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 965 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 966 } else { 967 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 968 Offset += OpC->getSExtValue()*Size; 969 } 970 } 971 972 // Re-sign extend from the pointer size if needed to get overflow edge cases 973 // right. 974 unsigned PtrSize = TD.getPointerSizeInBits(); 975 if (PtrSize < 64) 976 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 977 978 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 979} 980 981 982/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 983/// memdep query of a load that ends up being a clobbering memory write (store, 984/// memset, memcpy, memmove). This means that the write *may* provide bits used 985/// by the load but we can't be sure because the pointers don't mustalias. 986/// 987/// Check this case to see if there is anything more we can do before we give 988/// up. This returns -1 if we have to give up, or a byte number in the stored 989/// value of the piece that feeds the load. 990static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 991 Value *WritePtr, 992 uint64_t WriteSizeInBits, 993 const TargetData &TD) { 994 // If the loaded or stored value is an first class array or struct, don't try 995 // to transform them. We need to be able to bitcast to integer. 996 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 997 return -1; 998 999 int64_t StoreOffset = 0, LoadOffset = 0; 1000 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD); 1001 Value *LoadBase = 1002 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 1003 if (StoreBase != LoadBase) 1004 return -1; 1005 1006 // If the load and store are to the exact same address, they should have been 1007 // a must alias. AA must have gotten confused. 1008 // FIXME: Study to see if/when this happens. One case is forwarding a memset 1009 // to a load from the base of the memset. 1010#if 0 1011 if (LoadOffset == StoreOffset) { 1012 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1013 << "Base = " << *StoreBase << "\n" 1014 << "Store Ptr = " << *WritePtr << "\n" 1015 << "Store Offs = " << StoreOffset << "\n" 1016 << "Load Ptr = " << *LoadPtr << "\n"; 1017 abort(); 1018 } 1019#endif 1020 1021 // If the load and store don't overlap at all, the store doesn't provide 1022 // anything to the load. In this case, they really don't alias at all, AA 1023 // must have gotten confused. 1024 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 1025 // remove this check, as it is duplicated with what we have below. 1026 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 1027 1028 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 1029 return -1; 1030 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 1031 LoadSize >>= 3; 1032 1033 1034 bool isAAFailure = false; 1035 if (StoreOffset < LoadOffset) 1036 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1037 else 1038 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1039 1040 if (isAAFailure) { 1041#if 0 1042 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1043 << "Base = " << *StoreBase << "\n" 1044 << "Store Ptr = " << *WritePtr << "\n" 1045 << "Store Offs = " << StoreOffset << "\n" 1046 << "Load Ptr = " << *LoadPtr << "\n"; 1047 abort(); 1048#endif 1049 return -1; 1050 } 1051 1052 // If the Load isn't completely contained within the stored bits, we don't 1053 // have all the bits to feed it. We could do something crazy in the future 1054 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1055 // valuable. 1056 if (StoreOffset > LoadOffset || 1057 StoreOffset+StoreSize < LoadOffset+LoadSize) 1058 return -1; 1059 1060 // Okay, we can do this transformation. Return the number of bytes into the 1061 // store that the load is. 1062 return LoadOffset-StoreOffset; 1063} 1064 1065/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1066/// memdep query of a load that ends up being a clobbering store. 1067static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 1068 StoreInst *DepSI, 1069 const TargetData &TD) { 1070 // Cannot handle reading from store of first-class aggregate yet. 1071 if (DepSI->getOperand(0)->getType()->isStructTy() || 1072 DepSI->getOperand(0)->getType()->isArrayTy()) 1073 return -1; 1074 1075 Value *StorePtr = DepSI->getPointerOperand(); 1076 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType()); 1077 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1078 StorePtr, StoreSize, TD); 1079} 1080 1081static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 1082 MemIntrinsic *MI, 1083 const TargetData &TD) { 1084 // If the mem operation is a non-constant size, we can't handle it. 1085 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 1086 if (SizeCst == 0) return -1; 1087 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 1088 1089 // If this is memset, we just need to see if the offset is valid in the size 1090 // of the memset.. 1091 if (MI->getIntrinsicID() == Intrinsic::memset) 1092 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 1093 MemSizeInBits, TD); 1094 1095 // If we have a memcpy/memmove, the only case we can handle is if this is a 1096 // copy from constant memory. In that case, we can read directly from the 1097 // constant memory. 1098 MemTransferInst *MTI = cast<MemTransferInst>(MI); 1099 1100 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 1101 if (Src == 0) return -1; 1102 1103 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject()); 1104 if (GV == 0 || !GV->isConstant()) return -1; 1105 1106 // See if the access is within the bounds of the transfer. 1107 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 1108 MI->getDest(), MemSizeInBits, TD); 1109 if (Offset == -1) 1110 return Offset; 1111 1112 // Otherwise, see if we can constant fold a load from the constant with the 1113 // offset applied as appropriate. 1114 Src = ConstantExpr::getBitCast(Src, 1115 llvm::Type::getInt8PtrTy(Src->getContext())); 1116 Constant *OffsetCst = 1117 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1118 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1119 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1120 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 1121 return Offset; 1122 return -1; 1123} 1124 1125 1126/// GetStoreValueForLoad - This function is called when we have a 1127/// memdep query of a load that ends up being a clobbering store. This means 1128/// that the store *may* provide bits used by the load but we can't be sure 1129/// because the pointers don't mustalias. Check this case to see if there is 1130/// anything more we can do before we give up. 1131static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1132 const Type *LoadTy, 1133 Instruction *InsertPt, const TargetData &TD){ 1134 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1135 1136 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 1137 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 1138 1139 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1140 1141 // Compute which bits of the stored value are being used by the load. Convert 1142 // to an integer type to start with. 1143 if (SrcVal->getType()->isPointerTy()) 1144 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 1145 if (!SrcVal->getType()->isIntegerTy()) 1146 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1147 "tmp"); 1148 1149 // Shift the bits to the least significant depending on endianness. 1150 unsigned ShiftAmt; 1151 if (TD.isLittleEndian()) 1152 ShiftAmt = Offset*8; 1153 else 1154 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1155 1156 if (ShiftAmt) 1157 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 1158 1159 if (LoadSize != StoreSize) 1160 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1161 "tmp"); 1162 1163 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1164} 1165 1166/// GetMemInstValueForLoad - This function is called when we have a 1167/// memdep query of a load that ends up being a clobbering mem intrinsic. 1168static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 1169 const Type *LoadTy, Instruction *InsertPt, 1170 const TargetData &TD){ 1171 LLVMContext &Ctx = LoadTy->getContext(); 1172 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1173 1174 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 1175 1176 // We know that this method is only called when the mem transfer fully 1177 // provides the bits for the load. 1178 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 1179 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 1180 // independently of what the offset is. 1181 Value *Val = MSI->getValue(); 1182 if (LoadSize != 1) 1183 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 1184 1185 Value *OneElt = Val; 1186 1187 // Splat the value out to the right number of bits. 1188 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 1189 // If we can double the number of bytes set, do it. 1190 if (NumBytesSet*2 <= LoadSize) { 1191 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 1192 Val = Builder.CreateOr(Val, ShVal); 1193 NumBytesSet <<= 1; 1194 continue; 1195 } 1196 1197 // Otherwise insert one byte at a time. 1198 Value *ShVal = Builder.CreateShl(Val, 1*8); 1199 Val = Builder.CreateOr(OneElt, ShVal); 1200 ++NumBytesSet; 1201 } 1202 1203 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1204 } 1205 1206 // Otherwise, this is a memcpy/memmove from a constant global. 1207 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1208 Constant *Src = cast<Constant>(MTI->getSource()); 1209 1210 // Otherwise, see if we can constant fold a load from the constant with the 1211 // offset applied as appropriate. 1212 Src = ConstantExpr::getBitCast(Src, 1213 llvm::Type::getInt8PtrTy(Src->getContext())); 1214 Constant *OffsetCst = 1215 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1216 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1217 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1218 return ConstantFoldLoadFromConstPtr(Src, &TD); 1219} 1220 1221namespace { 1222 1223struct AvailableValueInBlock { 1224 /// BB - The basic block in question. 1225 BasicBlock *BB; 1226 enum ValType { 1227 SimpleVal, // A simple offsetted value that is accessed. 1228 MemIntrin // A memory intrinsic which is loaded from. 1229 }; 1230 1231 /// V - The value that is live out of the block. 1232 PointerIntPair<Value *, 1, ValType> Val; 1233 1234 /// Offset - The byte offset in Val that is interesting for the load query. 1235 unsigned Offset; 1236 1237 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1238 unsigned Offset = 0) { 1239 AvailableValueInBlock Res; 1240 Res.BB = BB; 1241 Res.Val.setPointer(V); 1242 Res.Val.setInt(SimpleVal); 1243 Res.Offset = Offset; 1244 return Res; 1245 } 1246 1247 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1248 unsigned Offset = 0) { 1249 AvailableValueInBlock Res; 1250 Res.BB = BB; 1251 Res.Val.setPointer(MI); 1252 Res.Val.setInt(MemIntrin); 1253 Res.Offset = Offset; 1254 return Res; 1255 } 1256 1257 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1258 Value *getSimpleValue() const { 1259 assert(isSimpleValue() && "Wrong accessor"); 1260 return Val.getPointer(); 1261 } 1262 1263 MemIntrinsic *getMemIntrinValue() const { 1264 assert(!isSimpleValue() && "Wrong accessor"); 1265 return cast<MemIntrinsic>(Val.getPointer()); 1266 } 1267 1268 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1269 /// defined here to the specified type. This handles various coercion cases. 1270 Value *MaterializeAdjustedValue(const Type *LoadTy, 1271 const TargetData *TD) const { 1272 Value *Res; 1273 if (isSimpleValue()) { 1274 Res = getSimpleValue(); 1275 if (Res->getType() != LoadTy) { 1276 assert(TD && "Need target data to handle type mismatch case"); 1277 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1278 *TD); 1279 1280 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1281 << *getSimpleValue() << '\n' 1282 << *Res << '\n' << "\n\n\n"); 1283 } 1284 } else { 1285 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1286 LoadTy, BB->getTerminator(), *TD); 1287 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1288 << " " << *getMemIntrinValue() << '\n' 1289 << *Res << '\n' << "\n\n\n"); 1290 } 1291 return Res; 1292 } 1293}; 1294 1295} 1296 1297/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1298/// construct SSA form, allowing us to eliminate LI. This returns the value 1299/// that should be used at LI's definition site. 1300static Value *ConstructSSAForLoadSet(LoadInst *LI, 1301 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1302 const TargetData *TD, 1303 const DominatorTree &DT, 1304 AliasAnalysis *AA) { 1305 // Check for the fully redundant, dominating load case. In this case, we can 1306 // just use the dominating value directly. 1307 if (ValuesPerBlock.size() == 1 && 1308 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent())) 1309 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD); 1310 1311 // Otherwise, we have to construct SSA form. 1312 SmallVector<PHINode*, 8> NewPHIs; 1313 SSAUpdater SSAUpdate(&NewPHIs); 1314 SSAUpdate.Initialize(LI); 1315 1316 const Type *LoadTy = LI->getType(); 1317 1318 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1319 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1320 BasicBlock *BB = AV.BB; 1321 1322 if (SSAUpdate.HasValueForBlock(BB)) 1323 continue; 1324 1325 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD)); 1326 } 1327 1328 // Perform PHI construction. 1329 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1330 1331 // If new PHI nodes were created, notify alias analysis. 1332 if (V->getType()->isPointerTy()) 1333 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1334 AA->copyValue(LI, NewPHIs[i]); 1335 1336 return V; 1337} 1338 1339static bool isLifetimeStart(const Instruction *Inst) { 1340 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1341 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1342 return false; 1343} 1344 1345/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1346/// non-local by performing PHI construction. 1347bool GVN::processNonLocalLoad(LoadInst *LI, 1348 SmallVectorImpl<Instruction*> &toErase) { 1349 // Find the non-local dependencies of the load. 1350 SmallVector<NonLocalDepResult, 64> Deps; 1351 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1352 Deps); 1353 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1354 // << Deps.size() << *LI << '\n'); 1355 1356 // If we had to process more than one hundred blocks to find the 1357 // dependencies, this load isn't worth worrying about. Optimizing 1358 // it will be too expensive. 1359 if (Deps.size() > 100) 1360 return false; 1361 1362 // If we had a phi translation failure, we'll have a single entry which is a 1363 // clobber in the current block. Reject this early. 1364 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) { 1365 DEBUG( 1366 dbgs() << "GVN: non-local load "; 1367 WriteAsOperand(dbgs(), LI); 1368 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1369 ); 1370 return false; 1371 } 1372 1373 // Filter out useless results (non-locals, etc). Keep track of the blocks 1374 // where we have a value available in repl, also keep track of whether we see 1375 // dependencies that produce an unknown value for the load (such as a call 1376 // that could potentially clobber the load). 1377 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1378 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1379 1380 const TargetData *TD = 0; 1381 1382 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1383 BasicBlock *DepBB = Deps[i].getBB(); 1384 MemDepResult DepInfo = Deps[i].getResult(); 1385 1386 if (DepInfo.isClobber()) { 1387 // The address being loaded in this non-local block may not be the same as 1388 // the pointer operand of the load if PHI translation occurs. Make sure 1389 // to consider the right address. 1390 Value *Address = Deps[i].getAddress(); 1391 1392 // If the dependence is to a store that writes to a superset of the bits 1393 // read by the load, we can extract the bits we need for the load from the 1394 // stored value. 1395 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1396 if (TD == 0) 1397 TD = getAnalysisIfAvailable<TargetData>(); 1398 if (TD && Address) { 1399 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1400 DepSI, *TD); 1401 if (Offset != -1) { 1402 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1403 DepSI->getOperand(0), 1404 Offset)); 1405 continue; 1406 } 1407 } 1408 } 1409 1410 // If the clobbering value is a memset/memcpy/memmove, see if we can 1411 // forward a value on from it. 1412 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1413 if (TD == 0) 1414 TD = getAnalysisIfAvailable<TargetData>(); 1415 if (TD && Address) { 1416 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1417 DepMI, *TD); 1418 if (Offset != -1) { 1419 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1420 Offset)); 1421 continue; 1422 } 1423 } 1424 } 1425 1426 UnavailableBlocks.push_back(DepBB); 1427 continue; 1428 } 1429 1430 Instruction *DepInst = DepInfo.getInst(); 1431 1432 // Loading the allocation -> undef. 1433 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1434 // Loading immediately after lifetime begin -> undef. 1435 isLifetimeStart(DepInst)) { 1436 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1437 UndefValue::get(LI->getType()))); 1438 continue; 1439 } 1440 1441 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1442 // Reject loads and stores that are to the same address but are of 1443 // different types if we have to. 1444 if (S->getOperand(0)->getType() != LI->getType()) { 1445 if (TD == 0) 1446 TD = getAnalysisIfAvailable<TargetData>(); 1447 1448 // If the stored value is larger or equal to the loaded value, we can 1449 // reuse it. 1450 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0), 1451 LI->getType(), *TD)) { 1452 UnavailableBlocks.push_back(DepBB); 1453 continue; 1454 } 1455 } 1456 1457 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1458 S->getOperand(0))); 1459 continue; 1460 } 1461 1462 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1463 // If the types mismatch and we can't handle it, reject reuse of the load. 1464 if (LD->getType() != LI->getType()) { 1465 if (TD == 0) 1466 TD = getAnalysisIfAvailable<TargetData>(); 1467 1468 // If the stored value is larger or equal to the loaded value, we can 1469 // reuse it. 1470 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1471 UnavailableBlocks.push_back(DepBB); 1472 continue; 1473 } 1474 } 1475 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1476 continue; 1477 } 1478 1479 UnavailableBlocks.push_back(DepBB); 1480 continue; 1481 } 1482 1483 // If we have no predecessors that produce a known value for this load, exit 1484 // early. 1485 if (ValuesPerBlock.empty()) return false; 1486 1487 // If all of the instructions we depend on produce a known value for this 1488 // load, then it is fully redundant and we can use PHI insertion to compute 1489 // its value. Insert PHIs and remove the fully redundant value now. 1490 if (UnavailableBlocks.empty()) { 1491 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1492 1493 // Perform PHI construction. 1494 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1495 VN.getAliasAnalysis()); 1496 LI->replaceAllUsesWith(V); 1497 1498 if (isa<PHINode>(V)) 1499 V->takeName(LI); 1500 if (V->getType()->isPointerTy()) 1501 MD->invalidateCachedPointerInfo(V); 1502 VN.erase(LI); 1503 toErase.push_back(LI); 1504 ++NumGVNLoad; 1505 return true; 1506 } 1507 1508 if (!EnablePRE || !EnableLoadPRE) 1509 return false; 1510 1511 // Okay, we have *some* definitions of the value. This means that the value 1512 // is available in some of our (transitive) predecessors. Lets think about 1513 // doing PRE of this load. This will involve inserting a new load into the 1514 // predecessor when it's not available. We could do this in general, but 1515 // prefer to not increase code size. As such, we only do this when we know 1516 // that we only have to insert *one* load (which means we're basically moving 1517 // the load, not inserting a new one). 1518 1519 SmallPtrSet<BasicBlock *, 4> Blockers; 1520 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1521 Blockers.insert(UnavailableBlocks[i]); 1522 1523 // Lets find first basic block with more than one predecessor. Walk backwards 1524 // through predecessors if needed. 1525 BasicBlock *LoadBB = LI->getParent(); 1526 BasicBlock *TmpBB = LoadBB; 1527 1528 bool isSinglePred = false; 1529 bool allSingleSucc = true; 1530 while (TmpBB->getSinglePredecessor()) { 1531 isSinglePred = true; 1532 TmpBB = TmpBB->getSinglePredecessor(); 1533 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1534 return false; 1535 if (Blockers.count(TmpBB)) 1536 return false; 1537 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1538 allSingleSucc = false; 1539 } 1540 1541 assert(TmpBB); 1542 LoadBB = TmpBB; 1543 1544 // If we have a repl set with LI itself in it, this means we have a loop where 1545 // at least one of the values is LI. Since this means that we won't be able 1546 // to eliminate LI even if we insert uses in the other predecessors, we will 1547 // end up increasing code size. Reject this by scanning for LI. 1548 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1549 if (ValuesPerBlock[i].isSimpleValue() && 1550 ValuesPerBlock[i].getSimpleValue() == LI) { 1551 // Skip cases where LI is the only definition, even for EnableFullLoadPRE. 1552 if (!EnableFullLoadPRE || e == 1) 1553 return false; 1554 } 1555 } 1556 1557 // FIXME: It is extremely unclear what this loop is doing, other than 1558 // artificially restricting loadpre. 1559 if (isSinglePred) { 1560 bool isHot = false; 1561 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1562 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1563 if (AV.isSimpleValue()) 1564 // "Hot" Instruction is in some loop (because it dominates its dep. 1565 // instruction). 1566 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1567 if (DT->dominates(LI, I)) { 1568 isHot = true; 1569 break; 1570 } 1571 } 1572 1573 // We are interested only in "hot" instructions. We don't want to do any 1574 // mis-optimizations here. 1575 if (!isHot) 1576 return false; 1577 } 1578 1579 // Check to see how many predecessors have the loaded value fully 1580 // available. 1581 DenseMap<BasicBlock*, Value*> PredLoads; 1582 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1583 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1584 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1585 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1586 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1587 1588 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1589 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1590 PI != E; ++PI) { 1591 BasicBlock *Pred = *PI; 1592 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1593 continue; 1594 } 1595 PredLoads[Pred] = 0; 1596 1597 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1598 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1599 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1600 << Pred->getName() << "': " << *LI << '\n'); 1601 return false; 1602 } 1603 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1604 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1605 } 1606 } 1607 if (!NeedToSplit.empty()) { 1608 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1609 return false; 1610 } 1611 1612 // Decide whether PRE is profitable for this load. 1613 unsigned NumUnavailablePreds = PredLoads.size(); 1614 assert(NumUnavailablePreds != 0 && 1615 "Fully available value should be eliminated above!"); 1616 if (!EnableFullLoadPRE) { 1617 // If this load is unavailable in multiple predecessors, reject it. 1618 // FIXME: If we could restructure the CFG, we could make a common pred with 1619 // all the preds that don't have an available LI and insert a new load into 1620 // that one block. 1621 if (NumUnavailablePreds != 1) 1622 return false; 1623 } 1624 1625 // Check if the load can safely be moved to all the unavailable predecessors. 1626 bool CanDoPRE = true; 1627 SmallVector<Instruction*, 8> NewInsts; 1628 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1629 E = PredLoads.end(); I != E; ++I) { 1630 BasicBlock *UnavailablePred = I->first; 1631 1632 // Do PHI translation to get its value in the predecessor if necessary. The 1633 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1634 1635 // If all preds have a single successor, then we know it is safe to insert 1636 // the load on the pred (?!?), so we can insert code to materialize the 1637 // pointer if it is not available. 1638 PHITransAddr Address(LI->getOperand(0), TD); 1639 Value *LoadPtr = 0; 1640 if (allSingleSucc) { 1641 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1642 *DT, NewInsts); 1643 } else { 1644 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1645 LoadPtr = Address.getAddr(); 1646 } 1647 1648 // If we couldn't find or insert a computation of this phi translated value, 1649 // we fail PRE. 1650 if (LoadPtr == 0) { 1651 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1652 << *LI->getOperand(0) << "\n"); 1653 CanDoPRE = false; 1654 break; 1655 } 1656 1657 // Make sure it is valid to move this load here. We have to watch out for: 1658 // @1 = getelementptr (i8* p, ... 1659 // test p and branch if == 0 1660 // load @1 1661 // It is valid to have the getelementptr before the test, even if p can be 0, 1662 // as getelementptr only does address arithmetic. 1663 // If we are not pushing the value through any multiple-successor blocks 1664 // we do not have this case. Otherwise, check that the load is safe to 1665 // put anywhere; this can be improved, but should be conservatively safe. 1666 if (!allSingleSucc && 1667 // FIXME: REEVALUTE THIS. 1668 !isSafeToLoadUnconditionally(LoadPtr, 1669 UnavailablePred->getTerminator(), 1670 LI->getAlignment(), TD)) { 1671 CanDoPRE = false; 1672 break; 1673 } 1674 1675 I->second = LoadPtr; 1676 } 1677 1678 if (!CanDoPRE) { 1679 while (!NewInsts.empty()) 1680 NewInsts.pop_back_val()->eraseFromParent(); 1681 return false; 1682 } 1683 1684 // Okay, we can eliminate this load by inserting a reload in the predecessor 1685 // and using PHI construction to get the value in the other predecessors, do 1686 // it. 1687 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1688 DEBUG(if (!NewInsts.empty()) 1689 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1690 << *NewInsts.back() << '\n'); 1691 1692 // Assign value numbers to the new instructions. 1693 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1694 // FIXME: We really _ought_ to insert these value numbers into their 1695 // parent's availability map. However, in doing so, we risk getting into 1696 // ordering issues. If a block hasn't been processed yet, we would be 1697 // marking a value as AVAIL-IN, which isn't what we intend. 1698 VN.lookup_or_add(NewInsts[i]); 1699 } 1700 1701 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1702 E = PredLoads.end(); I != E; ++I) { 1703 BasicBlock *UnavailablePred = I->first; 1704 Value *LoadPtr = I->second; 1705 1706 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1707 LI->getAlignment(), 1708 UnavailablePred->getTerminator()); 1709 1710 // Add the newly created load. 1711 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1712 NewLoad)); 1713 MD->invalidateCachedPointerInfo(LoadPtr); 1714 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1715 } 1716 1717 // Perform PHI construction. 1718 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1719 VN.getAliasAnalysis()); 1720 LI->replaceAllUsesWith(V); 1721 if (isa<PHINode>(V)) 1722 V->takeName(LI); 1723 if (V->getType()->isPointerTy()) 1724 MD->invalidateCachedPointerInfo(V); 1725 VN.erase(LI); 1726 toErase.push_back(LI); 1727 ++NumPRELoad; 1728 return true; 1729} 1730 1731/// processLoad - Attempt to eliminate a load, first by eliminating it 1732/// locally, and then attempting non-local elimination if that fails. 1733bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1734 if (!MD) 1735 return false; 1736 1737 if (L->isVolatile()) 1738 return false; 1739 1740 // ... to a pointer that has been loaded from before... 1741 MemDepResult Dep = MD->getDependency(L); 1742 1743 // If the value isn't available, don't do anything! 1744 if (Dep.isClobber()) { 1745 // Check to see if we have something like this: 1746 // store i32 123, i32* %P 1747 // %A = bitcast i32* %P to i8* 1748 // %B = gep i8* %A, i32 1 1749 // %C = load i8* %B 1750 // 1751 // We could do that by recognizing if the clobber instructions are obviously 1752 // a common base + constant offset, and if the previous store (or memset) 1753 // completely covers this load. This sort of thing can happen in bitfield 1754 // access code. 1755 Value *AvailVal = 0; 1756 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1757 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1758 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1759 L->getPointerOperand(), 1760 DepSI, *TD); 1761 if (Offset != -1) 1762 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1763 L->getType(), L, *TD); 1764 } 1765 1766 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1767 // a value on from it. 1768 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1769 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1770 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1771 L->getPointerOperand(), 1772 DepMI, *TD); 1773 if (Offset != -1) 1774 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD); 1775 } 1776 } 1777 1778 if (AvailVal) { 1779 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1780 << *AvailVal << '\n' << *L << "\n\n\n"); 1781 1782 // Replace the load! 1783 L->replaceAllUsesWith(AvailVal); 1784 if (AvailVal->getType()->isPointerTy()) 1785 MD->invalidateCachedPointerInfo(AvailVal); 1786 VN.erase(L); 1787 toErase.push_back(L); 1788 ++NumGVNLoad; 1789 return true; 1790 } 1791 1792 DEBUG( 1793 // fast print dep, using operator<< on instruction would be too slow 1794 dbgs() << "GVN: load "; 1795 WriteAsOperand(dbgs(), L); 1796 Instruction *I = Dep.getInst(); 1797 dbgs() << " is clobbered by " << *I << '\n'; 1798 ); 1799 return false; 1800 } 1801 1802 // If it is defined in another block, try harder. 1803 if (Dep.isNonLocal()) 1804 return processNonLocalLoad(L, toErase); 1805 1806 Instruction *DepInst = Dep.getInst(); 1807 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1808 Value *StoredVal = DepSI->getOperand(0); 1809 1810 // The store and load are to a must-aliased pointer, but they may not 1811 // actually have the same type. See if we know how to reuse the stored 1812 // value (depending on its type). 1813 const TargetData *TD = 0; 1814 if (StoredVal->getType() != L->getType()) { 1815 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1816 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1817 L, *TD); 1818 if (StoredVal == 0) 1819 return false; 1820 1821 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1822 << '\n' << *L << "\n\n\n"); 1823 } 1824 else 1825 return false; 1826 } 1827 1828 // Remove it! 1829 L->replaceAllUsesWith(StoredVal); 1830 if (StoredVal->getType()->isPointerTy()) 1831 MD->invalidateCachedPointerInfo(StoredVal); 1832 VN.erase(L); 1833 toErase.push_back(L); 1834 ++NumGVNLoad; 1835 return true; 1836 } 1837 1838 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1839 Value *AvailableVal = DepLI; 1840 1841 // The loads are of a must-aliased pointer, but they may not actually have 1842 // the same type. See if we know how to reuse the previously loaded value 1843 // (depending on its type). 1844 const TargetData *TD = 0; 1845 if (DepLI->getType() != L->getType()) { 1846 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1847 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1848 if (AvailableVal == 0) 1849 return false; 1850 1851 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1852 << "\n" << *L << "\n\n\n"); 1853 } 1854 else 1855 return false; 1856 } 1857 1858 // Remove it! 1859 L->replaceAllUsesWith(AvailableVal); 1860 if (DepLI->getType()->isPointerTy()) 1861 MD->invalidateCachedPointerInfo(DepLI); 1862 VN.erase(L); 1863 toErase.push_back(L); 1864 ++NumGVNLoad; 1865 return true; 1866 } 1867 1868 // If this load really doesn't depend on anything, then we must be loading an 1869 // undef value. This can happen when loading for a fresh allocation with no 1870 // intervening stores, for example. 1871 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1872 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1873 VN.erase(L); 1874 toErase.push_back(L); 1875 ++NumGVNLoad; 1876 return true; 1877 } 1878 1879 // If this load occurs either right after a lifetime begin, 1880 // then the loaded value is undefined. 1881 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) { 1882 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1883 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1884 VN.erase(L); 1885 toErase.push_back(L); 1886 ++NumGVNLoad; 1887 return true; 1888 } 1889 } 1890 1891 return false; 1892} 1893 1894Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1895 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1896 if (I == localAvail.end()) 1897 return 0; 1898 1899 ValueNumberScope *Locals = I->second; 1900 while (Locals) { 1901 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1902 if (I != Locals->table.end()) 1903 return I->second; 1904 Locals = Locals->parent; 1905 } 1906 1907 return 0; 1908} 1909 1910 1911/// processInstruction - When calculating availability, handle an instruction 1912/// by inserting it into the appropriate sets 1913bool GVN::processInstruction(Instruction *I, 1914 SmallVectorImpl<Instruction*> &toErase) { 1915 // Ignore dbg info intrinsics. 1916 if (isa<DbgInfoIntrinsic>(I)) 1917 return false; 1918 1919 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1920 bool Changed = processLoad(LI, toErase); 1921 1922 if (!Changed) { 1923 unsigned Num = VN.lookup_or_add(LI); 1924 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1925 } 1926 1927 return Changed; 1928 } 1929 1930 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1931 unsigned Num = VN.lookup_or_add(I); 1932 1933 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1934 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1935 1936 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1937 return false; 1938 1939 Value *BranchCond = BI->getCondition(); 1940 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1941 1942 BasicBlock *TrueSucc = BI->getSuccessor(0); 1943 BasicBlock *FalseSucc = BI->getSuccessor(1); 1944 1945 if (TrueSucc->getSinglePredecessor()) 1946 localAvail[TrueSucc]->table[CondVN] = 1947 ConstantInt::getTrue(TrueSucc->getContext()); 1948 if (FalseSucc->getSinglePredecessor()) 1949 localAvail[FalseSucc]->table[CondVN] = 1950 ConstantInt::getFalse(TrueSucc->getContext()); 1951 1952 return false; 1953 1954 // Allocations are always uniquely numbered, so we can save time and memory 1955 // by fast failing them. 1956 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) { 1957 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1958 return false; 1959 } 1960 1961 // Collapse PHI nodes 1962 if (PHINode* p = dyn_cast<PHINode>(I)) { 1963 Value *constVal = CollapsePhi(p); 1964 1965 if (constVal) { 1966 p->replaceAllUsesWith(constVal); 1967 if (MD && constVal->getType()->isPointerTy()) 1968 MD->invalidateCachedPointerInfo(constVal); 1969 VN.erase(p); 1970 1971 toErase.push_back(p); 1972 } else { 1973 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1974 } 1975 1976 // If the number we were assigned was a brand new VN, then we don't 1977 // need to do a lookup to see if the number already exists 1978 // somewhere in the domtree: it can't! 1979 } else if (Num == NextNum) { 1980 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1981 1982 // Perform fast-path value-number based elimination of values inherited from 1983 // dominators. 1984 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1985 // Remove it! 1986 VN.erase(I); 1987 I->replaceAllUsesWith(repl); 1988 if (MD && repl->getType()->isPointerTy()) 1989 MD->invalidateCachedPointerInfo(repl); 1990 toErase.push_back(I); 1991 return true; 1992 1993 } else { 1994 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1995 } 1996 1997 return false; 1998} 1999 2000/// runOnFunction - This is the main transformation entry point for a function. 2001bool GVN::runOnFunction(Function& F) { 2002 if (!NoLoads) 2003 MD = &getAnalysis<MemoryDependenceAnalysis>(); 2004 DT = &getAnalysis<DominatorTree>(); 2005 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 2006 VN.setMemDep(MD); 2007 VN.setDomTree(DT); 2008 2009 bool Changed = false; 2010 bool ShouldContinue = true; 2011 2012 // Merge unconditional branches, allowing PRE to catch more 2013 // optimization opportunities. 2014 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 2015 BasicBlock *BB = FI; 2016 ++FI; 2017 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 2018 if (removedBlock) ++NumGVNBlocks; 2019 2020 Changed |= removedBlock; 2021 } 2022 2023 unsigned Iteration = 0; 2024 2025 while (ShouldContinue) { 2026 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 2027 ShouldContinue = iterateOnFunction(F); 2028 if (splitCriticalEdges()) 2029 ShouldContinue = true; 2030 Changed |= ShouldContinue; 2031 ++Iteration; 2032 } 2033 2034 if (EnablePRE) { 2035 bool PREChanged = true; 2036 while (PREChanged) { 2037 PREChanged = performPRE(F); 2038 Changed |= PREChanged; 2039 } 2040 } 2041 // FIXME: Should perform GVN again after PRE does something. PRE can move 2042 // computations into blocks where they become fully redundant. Note that 2043 // we can't do this until PRE's critical edge splitting updates memdep. 2044 // Actually, when this happens, we should just fully integrate PRE into GVN. 2045 2046 cleanupGlobalSets(); 2047 2048 return Changed; 2049} 2050 2051 2052bool GVN::processBlock(BasicBlock *BB) { 2053 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 2054 // incrementing BI before processing an instruction). 2055 SmallVector<Instruction*, 8> toErase; 2056 bool ChangedFunction = false; 2057 2058 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 2059 BI != BE;) { 2060 ChangedFunction |= processInstruction(BI, toErase); 2061 if (toErase.empty()) { 2062 ++BI; 2063 continue; 2064 } 2065 2066 // If we need some instructions deleted, do it now. 2067 NumGVNInstr += toErase.size(); 2068 2069 // Avoid iterator invalidation. 2070 bool AtStart = BI == BB->begin(); 2071 if (!AtStart) 2072 --BI; 2073 2074 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 2075 E = toErase.end(); I != E; ++I) { 2076 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 2077 if (MD) MD->removeInstruction(*I); 2078 (*I)->eraseFromParent(); 2079 DEBUG(verifyRemoved(*I)); 2080 } 2081 toErase.clear(); 2082 2083 if (AtStart) 2084 BI = BB->begin(); 2085 else 2086 ++BI; 2087 } 2088 2089 return ChangedFunction; 2090} 2091 2092/// performPRE - Perform a purely local form of PRE that looks for diamond 2093/// control flow patterns and attempts to perform simple PRE at the join point. 2094bool GVN::performPRE(Function &F) { 2095 bool Changed = false; 2096 DenseMap<BasicBlock*, Value*> predMap; 2097 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 2098 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 2099 BasicBlock *CurrentBlock = *DI; 2100 2101 // Nothing to PRE in the entry block. 2102 if (CurrentBlock == &F.getEntryBlock()) continue; 2103 2104 for (BasicBlock::iterator BI = CurrentBlock->begin(), 2105 BE = CurrentBlock->end(); BI != BE; ) { 2106 Instruction *CurInst = BI++; 2107 2108 if (isa<AllocaInst>(CurInst) || 2109 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 2110 CurInst->getType()->isVoidTy() || 2111 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 2112 isa<DbgInfoIntrinsic>(CurInst)) 2113 continue; 2114 2115 // We don't currently value number ANY inline asm calls. 2116 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2117 if (CallI->isInlineAsm()) 2118 continue; 2119 2120 uint32_t ValNo = VN.lookup(CurInst); 2121 2122 // Look for the predecessors for PRE opportunities. We're 2123 // only trying to solve the basic diamond case, where 2124 // a value is computed in the successor and one predecessor, 2125 // but not the other. We also explicitly disallow cases 2126 // where the successor is its own predecessor, because they're 2127 // more complicated to get right. 2128 unsigned NumWith = 0; 2129 unsigned NumWithout = 0; 2130 BasicBlock *PREPred = 0; 2131 predMap.clear(); 2132 2133 for (pred_iterator PI = pred_begin(CurrentBlock), 2134 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2135 BasicBlock *P = *PI; 2136 // We're not interested in PRE where the block is its 2137 // own predecessor, or in blocks with predecessors 2138 // that are not reachable. 2139 if (P == CurrentBlock) { 2140 NumWithout = 2; 2141 break; 2142 } else if (!localAvail.count(P)) { 2143 NumWithout = 2; 2144 break; 2145 } 2146 2147 DenseMap<uint32_t, Value*>::iterator predV = 2148 localAvail[P]->table.find(ValNo); 2149 if (predV == localAvail[P]->table.end()) { 2150 PREPred = P; 2151 ++NumWithout; 2152 } else if (predV->second == CurInst) { 2153 NumWithout = 2; 2154 } else { 2155 predMap[P] = predV->second; 2156 ++NumWith; 2157 } 2158 } 2159 2160 // Don't do PRE when it might increase code size, i.e. when 2161 // we would need to insert instructions in more than one pred. 2162 if (NumWithout != 1 || NumWith == 0) 2163 continue; 2164 2165 // Don't do PRE across indirect branch. 2166 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2167 continue; 2168 2169 // We can't do PRE safely on a critical edge, so instead we schedule 2170 // the edge to be split and perform the PRE the next time we iterate 2171 // on the function. 2172 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2173 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2174 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2175 continue; 2176 } 2177 2178 // Instantiate the expression in the predecessor that lacked it. 2179 // Because we are going top-down through the block, all value numbers 2180 // will be available in the predecessor by the time we need them. Any 2181 // that weren't originally present will have been instantiated earlier 2182 // in this loop. 2183 Instruction *PREInstr = CurInst->clone(); 2184 bool success = true; 2185 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2186 Value *Op = PREInstr->getOperand(i); 2187 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2188 continue; 2189 2190 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 2191 PREInstr->setOperand(i, V); 2192 } else { 2193 success = false; 2194 break; 2195 } 2196 } 2197 2198 // Fail out if we encounter an operand that is not available in 2199 // the PRE predecessor. This is typically because of loads which 2200 // are not value numbered precisely. 2201 if (!success) { 2202 delete PREInstr; 2203 DEBUG(verifyRemoved(PREInstr)); 2204 continue; 2205 } 2206 2207 PREInstr->insertBefore(PREPred->getTerminator()); 2208 PREInstr->setName(CurInst->getName() + ".pre"); 2209 predMap[PREPred] = PREInstr; 2210 VN.add(PREInstr, ValNo); 2211 ++NumGVNPRE; 2212 2213 // Update the availability map to include the new instruction. 2214 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 2215 2216 // Create a PHI to make the value available in this block. 2217 PHINode* Phi = PHINode::Create(CurInst->getType(), 2218 CurInst->getName() + ".pre-phi", 2219 CurrentBlock->begin()); 2220 for (pred_iterator PI = pred_begin(CurrentBlock), 2221 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2222 BasicBlock *P = *PI; 2223 Phi->addIncoming(predMap[P], P); 2224 } 2225 2226 VN.add(Phi, ValNo); 2227 localAvail[CurrentBlock]->table[ValNo] = Phi; 2228 2229 CurInst->replaceAllUsesWith(Phi); 2230 if (MD && Phi->getType()->isPointerTy()) 2231 MD->invalidateCachedPointerInfo(Phi); 2232 VN.erase(CurInst); 2233 2234 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2235 if (MD) MD->removeInstruction(CurInst); 2236 CurInst->eraseFromParent(); 2237 DEBUG(verifyRemoved(CurInst)); 2238 Changed = true; 2239 } 2240 } 2241 2242 if (splitCriticalEdges()) 2243 Changed = true; 2244 2245 return Changed; 2246} 2247 2248/// splitCriticalEdges - Split critical edges found during the previous 2249/// iteration that may enable further optimization. 2250bool GVN::splitCriticalEdges() { 2251 if (toSplit.empty()) 2252 return false; 2253 do { 2254 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2255 SplitCriticalEdge(Edge.first, Edge.second, this); 2256 } while (!toSplit.empty()); 2257 if (MD) MD->invalidateCachedPredecessors(); 2258 return true; 2259} 2260 2261/// iterateOnFunction - Executes one iteration of GVN 2262bool GVN::iterateOnFunction(Function &F) { 2263 cleanupGlobalSets(); 2264 2265 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2266 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 2267 if (DI->getIDom()) 2268 localAvail[DI->getBlock()] = 2269 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 2270 else 2271 localAvail[DI->getBlock()] = new ValueNumberScope(0); 2272 } 2273 2274 // Top-down walk of the dominator tree 2275 bool Changed = false; 2276#if 0 2277 // Needed for value numbering with phi construction to work. 2278 ReversePostOrderTraversal<Function*> RPOT(&F); 2279 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2280 RE = RPOT.end(); RI != RE; ++RI) 2281 Changed |= processBlock(*RI); 2282#else 2283 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2284 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2285 Changed |= processBlock(DI->getBlock()); 2286#endif 2287 2288 return Changed; 2289} 2290 2291void GVN::cleanupGlobalSets() { 2292 VN.clear(); 2293 2294 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2295 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 2296 delete I->second; 2297 localAvail.clear(); 2298} 2299 2300/// verifyRemoved - Verify that the specified instruction does not occur in our 2301/// internal data structures. 2302void GVN::verifyRemoved(const Instruction *Inst) const { 2303 VN.verifyRemoved(Inst); 2304 2305 // Walk through the value number scope to make sure the instruction isn't 2306 // ferreted away in it. 2307 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator 2308 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 2309 const ValueNumberScope *VNS = I->second; 2310 2311 while (VNS) { 2312 for (DenseMap<uint32_t, Value*>::const_iterator 2313 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 2314 assert(II->second != Inst && "Inst still in value numbering scope!"); 2315 } 2316 2317 VNS = VNS->parent; 2318 } 2319 } 2320} 2321