GVN.cpp revision f2becca90b832cc02345fba063b9b439b2be33ad
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/Function.h" 24#include "llvm/IntrinsicInst.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Operator.h" 27#include "llvm/Value.h" 28#include "llvm/ADT/DenseMap.h" 29#include "llvm/ADT/DepthFirstIterator.h" 30#include "llvm/ADT/PostOrderIterator.h" 31#include "llvm/ADT/SmallPtrSet.h" 32#include "llvm/ADT/SmallVector.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/Analysis/Dominators.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/MallocFreeHelper.h" 37#include "llvm/Analysis/MemoryDependenceAnalysis.h" 38#include "llvm/Support/CFG.h" 39#include "llvm/Support/CommandLine.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/GetElementPtrTypeIterator.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/Target/TargetData.h" 45#include "llvm/Transforms/Utils/BasicBlockUtils.h" 46#include "llvm/Transforms/Utils/Local.h" 47#include "llvm/Transforms/Utils/SSAUpdater.h" 48#include <cstdio> 49using namespace llvm; 50 51STATISTIC(NumGVNInstr, "Number of instructions deleted"); 52STATISTIC(NumGVNLoad, "Number of loads deleted"); 53STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 54STATISTIC(NumGVNBlocks, "Number of blocks merged"); 55STATISTIC(NumPRELoad, "Number of loads PRE'd"); 56 57static cl::opt<bool> EnablePRE("enable-pre", 58 cl::init(true), cl::Hidden); 59static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 60 61//===----------------------------------------------------------------------===// 62// ValueTable Class 63//===----------------------------------------------------------------------===// 64 65/// This class holds the mapping between values and value numbers. It is used 66/// as an efficient mechanism to determine the expression-wise equivalence of 67/// two values. 68namespace { 69 struct Expression { 70 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 71 UDIV, SDIV, FDIV, UREM, SREM, 72 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 73 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 74 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 75 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 76 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 77 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 78 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 79 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 80 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 81 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE }; 82 83 ExpressionOpcode opcode; 84 const Type* type; 85 SmallVector<uint32_t, 4> varargs; 86 Value *function; 87 88 Expression() { } 89 Expression(ExpressionOpcode o) : opcode(o) { } 90 91 bool operator==(const Expression &other) const { 92 if (opcode != other.opcode) 93 return false; 94 else if (opcode == EMPTY || opcode == TOMBSTONE) 95 return true; 96 else if (type != other.type) 97 return false; 98 else if (function != other.function) 99 return false; 100 else { 101 if (varargs.size() != other.varargs.size()) 102 return false; 103 104 for (size_t i = 0; i < varargs.size(); ++i) 105 if (varargs[i] != other.varargs[i]) 106 return false; 107 108 return true; 109 } 110 } 111 112 bool operator!=(const Expression &other) const { 113 return !(*this == other); 114 } 115 }; 116 117 class ValueTable { 118 private: 119 DenseMap<Value*, uint32_t> valueNumbering; 120 DenseMap<Expression, uint32_t> expressionNumbering; 121 AliasAnalysis* AA; 122 MemoryDependenceAnalysis* MD; 123 DominatorTree* DT; 124 125 uint32_t nextValueNumber; 126 127 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 128 Expression::ExpressionOpcode getOpcode(CmpInst* C); 129 Expression::ExpressionOpcode getOpcode(CastInst* C); 130 Expression create_expression(BinaryOperator* BO); 131 Expression create_expression(CmpInst* C); 132 Expression create_expression(ShuffleVectorInst* V); 133 Expression create_expression(ExtractElementInst* C); 134 Expression create_expression(InsertElementInst* V); 135 Expression create_expression(SelectInst* V); 136 Expression create_expression(CastInst* C); 137 Expression create_expression(GetElementPtrInst* G); 138 Expression create_expression(CallInst* C); 139 Expression create_expression(Constant* C); 140 Expression create_expression(ExtractValueInst* C); 141 Expression create_expression(InsertValueInst* C); 142 143 uint32_t lookup_or_add_call(CallInst* C); 144 public: 145 ValueTable() : nextValueNumber(1) { } 146 uint32_t lookup_or_add(Value *V); 147 uint32_t lookup(Value *V) const; 148 void add(Value *V, uint32_t num); 149 void clear(); 150 void erase(Value *v); 151 unsigned size(); 152 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 153 AliasAnalysis *getAliasAnalysis() const { return AA; } 154 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 155 void setDomTree(DominatorTree* D) { DT = D; } 156 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 157 void verifyRemoved(const Value *) const; 158 }; 159} 160 161namespace llvm { 162template <> struct DenseMapInfo<Expression> { 163 static inline Expression getEmptyKey() { 164 return Expression(Expression::EMPTY); 165 } 166 167 static inline Expression getTombstoneKey() { 168 return Expression(Expression::TOMBSTONE); 169 } 170 171 static unsigned getHashValue(const Expression e) { 172 unsigned hash = e.opcode; 173 174 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 175 (unsigned)((uintptr_t)e.type >> 9)); 176 177 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 178 E = e.varargs.end(); I != E; ++I) 179 hash = *I + hash * 37; 180 181 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 182 (unsigned)((uintptr_t)e.function >> 9)) + 183 hash * 37; 184 185 return hash; 186 } 187 static bool isEqual(const Expression &LHS, const Expression &RHS) { 188 return LHS == RHS; 189 } 190 static bool isPod() { return true; } 191}; 192} 193 194//===----------------------------------------------------------------------===// 195// ValueTable Internal Functions 196//===----------------------------------------------------------------------===// 197Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 198 switch(BO->getOpcode()) { 199 default: // THIS SHOULD NEVER HAPPEN 200 llvm_unreachable("Binary operator with unknown opcode?"); 201 case Instruction::Add: return Expression::ADD; 202 case Instruction::FAdd: return Expression::FADD; 203 case Instruction::Sub: return Expression::SUB; 204 case Instruction::FSub: return Expression::FSUB; 205 case Instruction::Mul: return Expression::MUL; 206 case Instruction::FMul: return Expression::FMUL; 207 case Instruction::UDiv: return Expression::UDIV; 208 case Instruction::SDiv: return Expression::SDIV; 209 case Instruction::FDiv: return Expression::FDIV; 210 case Instruction::URem: return Expression::UREM; 211 case Instruction::SRem: return Expression::SREM; 212 case Instruction::FRem: return Expression::FREM; 213 case Instruction::Shl: return Expression::SHL; 214 case Instruction::LShr: return Expression::LSHR; 215 case Instruction::AShr: return Expression::ASHR; 216 case Instruction::And: return Expression::AND; 217 case Instruction::Or: return Expression::OR; 218 case Instruction::Xor: return Expression::XOR; 219 } 220} 221 222Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 223 if (isa<ICmpInst>(C)) { 224 switch (C->getPredicate()) { 225 default: // THIS SHOULD NEVER HAPPEN 226 llvm_unreachable("Comparison with unknown predicate?"); 227 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 228 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 229 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 230 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 231 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 232 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 233 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 234 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 235 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 236 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 237 } 238 } else { 239 switch (C->getPredicate()) { 240 default: // THIS SHOULD NEVER HAPPEN 241 llvm_unreachable("Comparison with unknown predicate?"); 242 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 243 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 244 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 245 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 246 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 247 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 248 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 249 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 250 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 251 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 252 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 253 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 254 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 255 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 256 } 257 } 258} 259 260Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 261 switch(C->getOpcode()) { 262 default: // THIS SHOULD NEVER HAPPEN 263 llvm_unreachable("Cast operator with unknown opcode?"); 264 case Instruction::Trunc: return Expression::TRUNC; 265 case Instruction::ZExt: return Expression::ZEXT; 266 case Instruction::SExt: return Expression::SEXT; 267 case Instruction::FPToUI: return Expression::FPTOUI; 268 case Instruction::FPToSI: return Expression::FPTOSI; 269 case Instruction::UIToFP: return Expression::UITOFP; 270 case Instruction::SIToFP: return Expression::SITOFP; 271 case Instruction::FPTrunc: return Expression::FPTRUNC; 272 case Instruction::FPExt: return Expression::FPEXT; 273 case Instruction::PtrToInt: return Expression::PTRTOINT; 274 case Instruction::IntToPtr: return Expression::INTTOPTR; 275 case Instruction::BitCast: return Expression::BITCAST; 276 } 277} 278 279Expression ValueTable::create_expression(CallInst* C) { 280 Expression e; 281 282 e.type = C->getType(); 283 e.function = C->getCalledFunction(); 284 e.opcode = Expression::CALL; 285 286 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 287 I != E; ++I) 288 e.varargs.push_back(lookup_or_add(*I)); 289 290 return e; 291} 292 293Expression ValueTable::create_expression(BinaryOperator* BO) { 294 Expression e; 295 e.varargs.push_back(lookup_or_add(BO->getOperand(0))); 296 e.varargs.push_back(lookup_or_add(BO->getOperand(1))); 297 e.function = 0; 298 e.type = BO->getType(); 299 e.opcode = getOpcode(BO); 300 301 return e; 302} 303 304Expression ValueTable::create_expression(CmpInst* C) { 305 Expression e; 306 307 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 308 e.varargs.push_back(lookup_or_add(C->getOperand(1))); 309 e.function = 0; 310 e.type = C->getType(); 311 e.opcode = getOpcode(C); 312 313 return e; 314} 315 316Expression ValueTable::create_expression(CastInst* C) { 317 Expression e; 318 319 e.varargs.push_back(lookup_or_add(C->getOperand(0))); 320 e.function = 0; 321 e.type = C->getType(); 322 e.opcode = getOpcode(C); 323 324 return e; 325} 326 327Expression ValueTable::create_expression(ShuffleVectorInst* S) { 328 Expression e; 329 330 e.varargs.push_back(lookup_or_add(S->getOperand(0))); 331 e.varargs.push_back(lookup_or_add(S->getOperand(1))); 332 e.varargs.push_back(lookup_or_add(S->getOperand(2))); 333 e.function = 0; 334 e.type = S->getType(); 335 e.opcode = Expression::SHUFFLE; 336 337 return e; 338} 339 340Expression ValueTable::create_expression(ExtractElementInst* E) { 341 Expression e; 342 343 e.varargs.push_back(lookup_or_add(E->getOperand(0))); 344 e.varargs.push_back(lookup_or_add(E->getOperand(1))); 345 e.function = 0; 346 e.type = E->getType(); 347 e.opcode = Expression::EXTRACT; 348 349 return e; 350} 351 352Expression ValueTable::create_expression(InsertElementInst* I) { 353 Expression e; 354 355 e.varargs.push_back(lookup_or_add(I->getOperand(0))); 356 e.varargs.push_back(lookup_or_add(I->getOperand(1))); 357 e.varargs.push_back(lookup_or_add(I->getOperand(2))); 358 e.function = 0; 359 e.type = I->getType(); 360 e.opcode = Expression::INSERT; 361 362 return e; 363} 364 365Expression ValueTable::create_expression(SelectInst* I) { 366 Expression e; 367 368 e.varargs.push_back(lookup_or_add(I->getCondition())); 369 e.varargs.push_back(lookup_or_add(I->getTrueValue())); 370 e.varargs.push_back(lookup_or_add(I->getFalseValue())); 371 e.function = 0; 372 e.type = I->getType(); 373 e.opcode = Expression::SELECT; 374 375 return e; 376} 377 378Expression ValueTable::create_expression(GetElementPtrInst* G) { 379 Expression e; 380 381 e.varargs.push_back(lookup_or_add(G->getPointerOperand())); 382 e.function = 0; 383 e.type = G->getType(); 384 e.opcode = Expression::GEP; 385 386 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 387 I != E; ++I) 388 e.varargs.push_back(lookup_or_add(*I)); 389 390 return e; 391} 392 393Expression ValueTable::create_expression(ExtractValueInst* E) { 394 Expression e; 395 396 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 397 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 398 II != IE; ++II) 399 e.varargs.push_back(*II); 400 e.function = 0; 401 e.type = E->getType(); 402 e.opcode = Expression::EXTRACTVALUE; 403 404 return e; 405} 406 407Expression ValueTable::create_expression(InsertValueInst* E) { 408 Expression e; 409 410 e.varargs.push_back(lookup_or_add(E->getAggregateOperand())); 411 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand())); 412 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 413 II != IE; ++II) 414 e.varargs.push_back(*II); 415 e.function = 0; 416 e.type = E->getType(); 417 e.opcode = Expression::INSERTVALUE; 418 419 return e; 420} 421 422//===----------------------------------------------------------------------===// 423// ValueTable External Functions 424//===----------------------------------------------------------------------===// 425 426/// add - Insert a value into the table with a specified value number. 427void ValueTable::add(Value *V, uint32_t num) { 428 valueNumbering.insert(std::make_pair(V, num)); 429} 430 431uint32_t ValueTable::lookup_or_add_call(CallInst* C) { 432 if (AA->doesNotAccessMemory(C)) { 433 Expression exp = create_expression(C); 434 uint32_t& e = expressionNumbering[exp]; 435 if (!e) e = nextValueNumber++; 436 valueNumbering[C] = e; 437 return e; 438 } else if (AA->onlyReadsMemory(C)) { 439 Expression exp = create_expression(C); 440 uint32_t& e = expressionNumbering[exp]; 441 if (!e) { 442 e = nextValueNumber++; 443 valueNumbering[C] = e; 444 return e; 445 } 446 447 MemDepResult local_dep = MD->getDependency(C); 448 449 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 450 valueNumbering[C] = nextValueNumber; 451 return nextValueNumber++; 452 } 453 454 if (local_dep.isDef()) { 455 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 456 457 if (local_cdep->getNumOperands() != C->getNumOperands()) { 458 valueNumbering[C] = nextValueNumber; 459 return nextValueNumber++; 460 } 461 462 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 463 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 464 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 465 if (c_vn != cd_vn) { 466 valueNumbering[C] = nextValueNumber; 467 return nextValueNumber++; 468 } 469 } 470 471 uint32_t v = lookup_or_add(local_cdep); 472 valueNumbering[C] = v; 473 return v; 474 } 475 476 // Non-local case. 477 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 478 MD->getNonLocalCallDependency(CallSite(C)); 479 // FIXME: call/call dependencies for readonly calls should return def, not 480 // clobber! Move the checking logic to MemDep! 481 CallInst* cdep = 0; 482 483 // Check to see if we have a single dominating call instruction that is 484 // identical to C. 485 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 486 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 487 // Ignore non-local dependencies. 488 if (I->second.isNonLocal()) 489 continue; 490 491 // We don't handle non-depedencies. If we already have a call, reject 492 // instruction dependencies. 493 if (I->second.isClobber() || cdep != 0) { 494 cdep = 0; 495 break; 496 } 497 498 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 499 // FIXME: All duplicated with non-local case. 500 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 501 cdep = NonLocalDepCall; 502 continue; 503 } 504 505 cdep = 0; 506 break; 507 } 508 509 if (!cdep) { 510 valueNumbering[C] = nextValueNumber; 511 return nextValueNumber++; 512 } 513 514 if (cdep->getNumOperands() != C->getNumOperands()) { 515 valueNumbering[C] = nextValueNumber; 516 return nextValueNumber++; 517 } 518 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 519 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 520 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 521 if (c_vn != cd_vn) { 522 valueNumbering[C] = nextValueNumber; 523 return nextValueNumber++; 524 } 525 } 526 527 uint32_t v = lookup_or_add(cdep); 528 valueNumbering[C] = v; 529 return v; 530 531 } else { 532 valueNumbering[C] = nextValueNumber; 533 return nextValueNumber++; 534 } 535} 536 537/// lookup_or_add - Returns the value number for the specified value, assigning 538/// it a new number if it did not have one before. 539uint32_t ValueTable::lookup_or_add(Value *V) { 540 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 541 if (VI != valueNumbering.end()) 542 return VI->second; 543 544 if (!isa<Instruction>(V)) { 545 valueNumbering[V] = nextValueNumber; 546 return nextValueNumber++; 547 } 548 549 Instruction* I = cast<Instruction>(V); 550 Expression exp; 551 switch (I->getOpcode()) { 552 case Instruction::Call: 553 return lookup_or_add_call(cast<CallInst>(I)); 554 case Instruction::Add: 555 case Instruction::FAdd: 556 case Instruction::Sub: 557 case Instruction::FSub: 558 case Instruction::Mul: 559 case Instruction::FMul: 560 case Instruction::UDiv: 561 case Instruction::SDiv: 562 case Instruction::FDiv: 563 case Instruction::URem: 564 case Instruction::SRem: 565 case Instruction::FRem: 566 case Instruction::Shl: 567 case Instruction::LShr: 568 case Instruction::AShr: 569 case Instruction::And: 570 case Instruction::Or : 571 case Instruction::Xor: 572 exp = create_expression(cast<BinaryOperator>(I)); 573 break; 574 case Instruction::ICmp: 575 case Instruction::FCmp: 576 exp = create_expression(cast<CmpInst>(I)); 577 break; 578 case Instruction::Trunc: 579 case Instruction::ZExt: 580 case Instruction::SExt: 581 case Instruction::FPToUI: 582 case Instruction::FPToSI: 583 case Instruction::UIToFP: 584 case Instruction::SIToFP: 585 case Instruction::FPTrunc: 586 case Instruction::FPExt: 587 case Instruction::PtrToInt: 588 case Instruction::IntToPtr: 589 case Instruction::BitCast: 590 exp = create_expression(cast<CastInst>(I)); 591 break; 592 case Instruction::Select: 593 exp = create_expression(cast<SelectInst>(I)); 594 break; 595 case Instruction::ExtractElement: 596 exp = create_expression(cast<ExtractElementInst>(I)); 597 break; 598 case Instruction::InsertElement: 599 exp = create_expression(cast<InsertElementInst>(I)); 600 break; 601 case Instruction::ShuffleVector: 602 exp = create_expression(cast<ShuffleVectorInst>(I)); 603 break; 604 case Instruction::ExtractValue: 605 exp = create_expression(cast<ExtractValueInst>(I)); 606 break; 607 case Instruction::InsertValue: 608 exp = create_expression(cast<InsertValueInst>(I)); 609 break; 610 case Instruction::GetElementPtr: 611 exp = create_expression(cast<GetElementPtrInst>(I)); 612 break; 613 default: 614 valueNumbering[V] = nextValueNumber; 615 return nextValueNumber++; 616 } 617 618 uint32_t& e = expressionNumbering[exp]; 619 if (!e) e = nextValueNumber++; 620 valueNumbering[V] = e; 621 return e; 622} 623 624/// lookup - Returns the value number of the specified value. Fails if 625/// the value has not yet been numbered. 626uint32_t ValueTable::lookup(Value *V) const { 627 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 628 assert(VI != valueNumbering.end() && "Value not numbered?"); 629 return VI->second; 630} 631 632/// clear - Remove all entries from the ValueTable 633void ValueTable::clear() { 634 valueNumbering.clear(); 635 expressionNumbering.clear(); 636 nextValueNumber = 1; 637} 638 639/// erase - Remove a value from the value numbering 640void ValueTable::erase(Value *V) { 641 valueNumbering.erase(V); 642} 643 644/// verifyRemoved - Verify that the value is removed from all internal data 645/// structures. 646void ValueTable::verifyRemoved(const Value *V) const { 647 for (DenseMap<Value*, uint32_t>::iterator 648 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 649 assert(I->first != V && "Inst still occurs in value numbering map!"); 650 } 651} 652 653//===----------------------------------------------------------------------===// 654// GVN Pass 655//===----------------------------------------------------------------------===// 656 657namespace { 658 struct ValueNumberScope { 659 ValueNumberScope* parent; 660 DenseMap<uint32_t, Value*> table; 661 662 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 663 }; 664} 665 666namespace { 667 668 class GVN : public FunctionPass { 669 bool runOnFunction(Function &F); 670 public: 671 static char ID; // Pass identification, replacement for typeid 672 GVN() : FunctionPass(&ID) { } 673 674 private: 675 MemoryDependenceAnalysis *MD; 676 DominatorTree *DT; 677 678 ValueTable VN; 679 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 680 681 // This transformation requires dominator postdominator info 682 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 683 AU.addRequired<DominatorTree>(); 684 AU.addRequired<MemoryDependenceAnalysis>(); 685 AU.addRequired<AliasAnalysis>(); 686 687 AU.addPreserved<DominatorTree>(); 688 AU.addPreserved<AliasAnalysis>(); 689 } 690 691 // Helper fuctions 692 // FIXME: eliminate or document these better 693 bool processLoad(LoadInst* L, 694 SmallVectorImpl<Instruction*> &toErase); 695 bool processInstruction(Instruction *I, 696 SmallVectorImpl<Instruction*> &toErase); 697 bool processNonLocalLoad(LoadInst* L, 698 SmallVectorImpl<Instruction*> &toErase); 699 bool processBlock(BasicBlock *BB); 700 void dump(DenseMap<uint32_t, Value*>& d); 701 bool iterateOnFunction(Function &F); 702 Value *CollapsePhi(PHINode* p); 703 bool performPRE(Function& F); 704 Value *lookupNumber(BasicBlock *BB, uint32_t num); 705 void cleanupGlobalSets(); 706 void verifyRemoved(const Instruction *I) const; 707 }; 708 709 char GVN::ID = 0; 710} 711 712// createGVNPass - The public interface to this file... 713FunctionPass *llvm::createGVNPass() { return new GVN(); } 714 715static RegisterPass<GVN> X("gvn", 716 "Global Value Numbering"); 717 718void GVN::dump(DenseMap<uint32_t, Value*>& d) { 719 printf("{\n"); 720 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 721 E = d.end(); I != E; ++I) { 722 printf("%d\n", I->first); 723 I->second->dump(); 724 } 725 printf("}\n"); 726} 727 728static bool isSafeReplacement(PHINode* p, Instruction *inst) { 729 if (!isa<PHINode>(inst)) 730 return true; 731 732 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 733 UI != E; ++UI) 734 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 735 if (use_phi->getParent() == inst->getParent()) 736 return false; 737 738 return true; 739} 740 741Value *GVN::CollapsePhi(PHINode *PN) { 742 Value *ConstVal = PN->hasConstantValue(DT); 743 if (!ConstVal) return 0; 744 745 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 746 if (!Inst) 747 return ConstVal; 748 749 if (DT->dominates(Inst, PN)) 750 if (isSafeReplacement(PN, Inst)) 751 return Inst; 752 return 0; 753} 754 755/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 756/// we're analyzing is fully available in the specified block. As we go, keep 757/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 758/// map is actually a tri-state map with the following values: 759/// 0) we know the block *is not* fully available. 760/// 1) we know the block *is* fully available. 761/// 2) we do not know whether the block is fully available or not, but we are 762/// currently speculating that it will be. 763/// 3) we are speculating for this block and have used that to speculate for 764/// other blocks. 765static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 766 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 767 // Optimistically assume that the block is fully available and check to see 768 // if we already know about this block in one lookup. 769 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 770 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 771 772 // If the entry already existed for this block, return the precomputed value. 773 if (!IV.second) { 774 // If this is a speculative "available" value, mark it as being used for 775 // speculation of other blocks. 776 if (IV.first->second == 2) 777 IV.first->second = 3; 778 return IV.first->second != 0; 779 } 780 781 // Otherwise, see if it is fully available in all predecessors. 782 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 783 784 // If this block has no predecessors, it isn't live-in here. 785 if (PI == PE) 786 goto SpeculationFailure; 787 788 for (; PI != PE; ++PI) 789 // If the value isn't fully available in one of our predecessors, then it 790 // isn't fully available in this block either. Undo our previous 791 // optimistic assumption and bail out. 792 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 793 goto SpeculationFailure; 794 795 return true; 796 797// SpeculationFailure - If we get here, we found out that this is not, after 798// all, a fully-available block. We have a problem if we speculated on this and 799// used the speculation to mark other blocks as available. 800SpeculationFailure: 801 char &BBVal = FullyAvailableBlocks[BB]; 802 803 // If we didn't speculate on this, just return with it set to false. 804 if (BBVal == 2) { 805 BBVal = 0; 806 return false; 807 } 808 809 // If we did speculate on this value, we could have blocks set to 1 that are 810 // incorrect. Walk the (transitive) successors of this block and mark them as 811 // 0 if set to one. 812 SmallVector<BasicBlock*, 32> BBWorklist; 813 BBWorklist.push_back(BB); 814 815 while (!BBWorklist.empty()) { 816 BasicBlock *Entry = BBWorklist.pop_back_val(); 817 // Note that this sets blocks to 0 (unavailable) if they happen to not 818 // already be in FullyAvailableBlocks. This is safe. 819 char &EntryVal = FullyAvailableBlocks[Entry]; 820 if (EntryVal == 0) continue; // Already unavailable. 821 822 // Mark as unavailable. 823 EntryVal = 0; 824 825 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 826 BBWorklist.push_back(*I); 827 } 828 829 return false; 830} 831 832 833/// CanCoerceMustAliasedValueToLoad - Return true if 834/// CoerceAvailableValueToLoadType will succeed. 835static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 836 const Type *LoadTy, 837 const TargetData &TD) { 838 // If the loaded or stored value is an first class array or struct, don't try 839 // to transform them. We need to be able to bitcast to integer. 840 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) || 841 isa<StructType>(StoredVal->getType()) || 842 isa<ArrayType>(StoredVal->getType())) 843 return false; 844 845 // The store has to be at least as big as the load. 846 if (TD.getTypeSizeInBits(StoredVal->getType()) < 847 TD.getTypeSizeInBits(LoadTy)) 848 return false; 849 850 return true; 851} 852 853 854/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 855/// then a load from a must-aliased pointer of a different type, try to coerce 856/// the stored value. LoadedTy is the type of the load we want to replace and 857/// InsertPt is the place to insert new instructions. 858/// 859/// If we can't do it, return null. 860static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 861 const Type *LoadedTy, 862 Instruction *InsertPt, 863 const TargetData &TD) { 864 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 865 return 0; 866 867 const Type *StoredValTy = StoredVal->getType(); 868 869 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 870 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 871 872 // If the store and reload are the same size, we can always reuse it. 873 if (StoreSize == LoadSize) { 874 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) { 875 // Pointer to Pointer -> use bitcast. 876 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 877 } 878 879 // Convert source pointers to integers, which can be bitcast. 880 if (isa<PointerType>(StoredValTy)) { 881 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 882 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 883 } 884 885 const Type *TypeToCastTo = LoadedTy; 886 if (isa<PointerType>(TypeToCastTo)) 887 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 888 889 if (StoredValTy != TypeToCastTo) 890 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 891 892 // Cast to pointer if the load needs a pointer type. 893 if (isa<PointerType>(LoadedTy)) 894 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 895 896 return StoredVal; 897 } 898 899 // If the loaded value is smaller than the available value, then we can 900 // extract out a piece from it. If the available value is too small, then we 901 // can't do anything. 902 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 903 904 // Convert source pointers to integers, which can be manipulated. 905 if (isa<PointerType>(StoredValTy)) { 906 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 907 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 908 } 909 910 // Convert vectors and fp to integer, which can be manipulated. 911 if (!isa<IntegerType>(StoredValTy)) { 912 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 913 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 914 } 915 916 // If this is a big-endian system, we need to shift the value down to the low 917 // bits so that a truncate will work. 918 if (TD.isBigEndian()) { 919 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 920 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 921 } 922 923 // Truncate the integer to the right size now. 924 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 925 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 926 927 if (LoadedTy == NewIntTy) 928 return StoredVal; 929 930 // If the result is a pointer, inttoptr. 931 if (isa<PointerType>(LoadedTy)) 932 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 933 934 // Otherwise, bitcast. 935 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 936} 937 938/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 939/// be expressed as a base pointer plus a constant offset. Return the base and 940/// offset to the caller. 941static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 942 const TargetData &TD) { 943 Operator *PtrOp = dyn_cast<Operator>(Ptr); 944 if (PtrOp == 0) return Ptr; 945 946 // Just look through bitcasts. 947 if (PtrOp->getOpcode() == Instruction::BitCast) 948 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 949 950 // If this is a GEP with constant indices, we can look through it. 951 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 952 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 953 954 gep_type_iterator GTI = gep_type_begin(GEP); 955 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 956 ++I, ++GTI) { 957 ConstantInt *OpC = cast<ConstantInt>(*I); 958 if (OpC->isZero()) continue; 959 960 // Handle a struct and array indices which add their offset to the pointer. 961 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 962 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 963 } else { 964 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 965 Offset += OpC->getSExtValue()*Size; 966 } 967 } 968 969 // Re-sign extend from the pointer size if needed to get overflow edge cases 970 // right. 971 unsigned PtrSize = TD.getPointerSizeInBits(); 972 if (PtrSize < 64) 973 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 974 975 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 976} 977 978 979/// AnalyzeLoadFromClobberingStore - This function is called when we have a 980/// memdep query of a load that ends up being a clobbering store. This means 981/// that the store *may* provide bits used by the load but we can't be sure 982/// because the pointers don't mustalias. Check this case to see if there is 983/// anything more we can do before we give up. This returns -1 if we have to 984/// give up, or a byte number in the stored value of the piece that feeds the 985/// load. 986static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI, 987 const TargetData &TD) { 988 // If the loaded or stored value is an first class array or struct, don't try 989 // to transform them. We need to be able to bitcast to integer. 990 if (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()) || 991 isa<StructType>(DepSI->getOperand(0)->getType()) || 992 isa<ArrayType>(DepSI->getOperand(0)->getType())) 993 return -1; 994 995 int64_t StoreOffset = 0, LoadOffset = 0; 996 Value *StoreBase = 997 GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD); 998 Value *LoadBase = 999 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD); 1000 if (StoreBase != LoadBase) 1001 return -1; 1002 1003 // If the load and store are to the exact same address, they should have been 1004 // a must alias. AA must have gotten confused. 1005 // FIXME: Study to see if/when this happens. 1006 if (LoadOffset == StoreOffset) { 1007#if 0 1008 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1009 << "Base = " << *StoreBase << "\n" 1010 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 1011 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 1012 << "Load Ptr = " << *L->getPointerOperand() << "\n" 1013 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 1014 errs() << "'" << L->getParent()->getParent()->getName() << "'" 1015 << *L->getParent(); 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 StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType()); 1026 uint64_t LoadSize = TD.getTypeSizeInBits(L->getType()); 1027 1028 if ((StoreSize & 7) | (LoadSize & 7)) 1029 return -1; 1030 StoreSize >>= 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 errs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1043 << "Base = " << *StoreBase << "\n" 1044 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 1045 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 1046 << "Load Ptr = " << *L->getPointerOperand() << "\n" 1047 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 1048 errs() << "'" << L->getParent()->getParent()->getName() << "'" 1049 << *L->getParent(); 1050#endif 1051 return -1; 1052 } 1053 1054 // If the Load isn't completely contained within the stored bits, we don't 1055 // have all the bits to feed it. We could do something crazy in the future 1056 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1057 // valuable. 1058 if (StoreOffset > LoadOffset || 1059 StoreOffset+StoreSize < LoadOffset+LoadSize) 1060 return -1; 1061 1062 // Okay, we can do this transformation. Return the number of bytes into the 1063 // store that the load is. 1064 return LoadOffset-StoreOffset; 1065} 1066 1067 1068/// GetStoreValueForLoad - This function is called when we have a 1069/// memdep query of a load that ends up being a clobbering store. This means 1070/// that the store *may* provide bits used by the load but we can't be sure 1071/// because the pointers don't mustalias. Check this case to see if there is 1072/// anything more we can do before we give up. 1073static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1074 const Type *LoadTy, 1075 Instruction *InsertPt, const TargetData &TD){ 1076 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1077 1078 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8; 1079 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1080 1081 1082 // Compute which bits of the stored value are being used by the load. Convert 1083 // to an integer type to start with. 1084 if (isa<PointerType>(SrcVal->getType())) 1085 SrcVal = new PtrToIntInst(SrcVal, TD.getIntPtrType(Ctx), "tmp", InsertPt); 1086 if (!isa<IntegerType>(SrcVal->getType())) 1087 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1088 "tmp", InsertPt); 1089 1090 // Shift the bits to the least significant depending on endianness. 1091 unsigned ShiftAmt; 1092 if (TD.isLittleEndian()) { 1093 ShiftAmt = Offset*8; 1094 } else { 1095 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1096 } 1097 1098 if (ShiftAmt) 1099 SrcVal = BinaryOperator::CreateLShr(SrcVal, 1100 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt); 1101 1102 if (LoadSize != StoreSize) 1103 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1104 "tmp", InsertPt); 1105 1106 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1107} 1108 1109struct AvailableValueInBlock { 1110 /// BB - The basic block in question. 1111 BasicBlock *BB; 1112 /// V - The value that is live out of the block. 1113 Value *V; 1114 /// Offset - The byte offset in V that is interesting for the load query. 1115 unsigned Offset; 1116 1117 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1118 unsigned Offset = 0) { 1119 AvailableValueInBlock Res; 1120 Res.BB = BB; 1121 Res.V = V; 1122 Res.Offset = Offset; 1123 return Res; 1124 } 1125}; 1126 1127/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1128/// construct SSA form, allowing us to eliminate LI. This returns the value 1129/// that should be used at LI's definition site. 1130static Value *ConstructSSAForLoadSet(LoadInst *LI, 1131 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1132 const TargetData *TD, 1133 AliasAnalysis *AA) { 1134 SmallVector<PHINode*, 8> NewPHIs; 1135 SSAUpdater SSAUpdate(&NewPHIs); 1136 SSAUpdate.Initialize(LI); 1137 1138 const Type *LoadTy = LI->getType(); 1139 1140 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1141 BasicBlock *BB = ValuesPerBlock[i].BB; 1142 Value *AvailableVal = ValuesPerBlock[i].V; 1143 unsigned Offset = ValuesPerBlock[i].Offset; 1144 1145 if (SSAUpdate.HasValueForBlock(BB)) 1146 continue; 1147 1148 if (AvailableVal->getType() != LoadTy) { 1149 assert(TD && "Need target data to handle type mismatch case"); 1150 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy, 1151 BB->getTerminator(), *TD); 1152 1153 if (Offset) { 1154 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1155 << *ValuesPerBlock[i].V << '\n' 1156 << *AvailableVal << '\n' << "\n\n\n"); 1157 } 1158 1159 1160 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1161 << *ValuesPerBlock[i].V << '\n' 1162 << *AvailableVal << '\n' << "\n\n\n"); 1163 } 1164 1165 SSAUpdate.AddAvailableValue(BB, AvailableVal); 1166 } 1167 1168 // Perform PHI construction. 1169 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1170 1171 // If new PHI nodes were created, notify alias analysis. 1172 if (isa<PointerType>(V->getType())) 1173 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1174 AA->copyValue(LI, NewPHIs[i]); 1175 1176 return V; 1177} 1178 1179/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1180/// non-local by performing PHI construction. 1181bool GVN::processNonLocalLoad(LoadInst *LI, 1182 SmallVectorImpl<Instruction*> &toErase) { 1183 // Find the non-local dependencies of the load. 1184 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 1185 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1186 Deps); 1187 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: " 1188 // << Deps.size() << *LI << '\n'); 1189 1190 // If we had to process more than one hundred blocks to find the 1191 // dependencies, this load isn't worth worrying about. Optimizing 1192 // it will be too expensive. 1193 if (Deps.size() > 100) 1194 return false; 1195 1196 // If we had a phi translation failure, we'll have a single entry which is a 1197 // clobber in the current block. Reject this early. 1198 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 1199 DEBUG( 1200 errs() << "GVN: non-local load "; 1201 WriteAsOperand(errs(), LI); 1202 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n'; 1203 ); 1204 return false; 1205 } 1206 1207 // Filter out useless results (non-locals, etc). Keep track of the blocks 1208 // where we have a value available in repl, also keep track of whether we see 1209 // dependencies that produce an unknown value for the load (such as a call 1210 // that could potentially clobber the load). 1211 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1212 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1213 1214 const TargetData *TD = 0; 1215 1216 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1217 BasicBlock *DepBB = Deps[i].first; 1218 MemDepResult DepInfo = Deps[i].second; 1219 1220 if (DepInfo.isClobber()) { 1221 // If the dependence is to a store that writes to a superset of the bits 1222 // read by the load, we can extract the bits we need for the load from the 1223 // stored value. 1224 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1225 if (TD == 0) 1226 TD = getAnalysisIfAvailable<TargetData>(); 1227 if (TD) { 1228 int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD); 1229 if (Offset != -1) { 1230 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1231 DepSI->getOperand(0), 1232 Offset)); 1233 continue; 1234 } 1235 } 1236 } 1237 1238 // FIXME: Handle memset/memcpy. 1239 UnavailableBlocks.push_back(DepBB); 1240 continue; 1241 } 1242 1243 Instruction *DepInst = DepInfo.getInst(); 1244 1245 // Loading the allocation -> undef. 1246 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1247 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1248 UndefValue::get(LI->getType()))); 1249 continue; 1250 } 1251 1252 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1253 // Reject loads and stores that are to the same address but are of 1254 // different types if we have to. 1255 if (S->getOperand(0)->getType() != LI->getType()) { 1256 if (TD == 0) 1257 TD = getAnalysisIfAvailable<TargetData>(); 1258 1259 // If the stored value is larger or equal to the loaded value, we can 1260 // reuse it. 1261 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0), 1262 LI->getType(), *TD)) { 1263 UnavailableBlocks.push_back(DepBB); 1264 continue; 1265 } 1266 } 1267 1268 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1269 S->getOperand(0))); 1270 continue; 1271 } 1272 1273 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1274 // If the types mismatch and we can't handle it, reject reuse of the load. 1275 if (LD->getType() != LI->getType()) { 1276 if (TD == 0) 1277 TD = getAnalysisIfAvailable<TargetData>(); 1278 1279 // If the stored value is larger or equal to the loaded value, we can 1280 // reuse it. 1281 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1282 UnavailableBlocks.push_back(DepBB); 1283 continue; 1284 } 1285 } 1286 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1287 continue; 1288 } 1289 1290 UnavailableBlocks.push_back(DepBB); 1291 continue; 1292 } 1293 1294 // If we have no predecessors that produce a known value for this load, exit 1295 // early. 1296 if (ValuesPerBlock.empty()) return false; 1297 1298 // If all of the instructions we depend on produce a known value for this 1299 // load, then it is fully redundant and we can use PHI insertion to compute 1300 // its value. Insert PHIs and remove the fully redundant value now. 1301 if (UnavailableBlocks.empty()) { 1302 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1303 1304 // Perform PHI construction. 1305 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, 1306 VN.getAliasAnalysis()); 1307 LI->replaceAllUsesWith(V); 1308 1309 if (isa<PHINode>(V)) 1310 V->takeName(LI); 1311 if (isa<PointerType>(V->getType())) 1312 MD->invalidateCachedPointerInfo(V); 1313 toErase.push_back(LI); 1314 NumGVNLoad++; 1315 return true; 1316 } 1317 1318 if (!EnablePRE || !EnableLoadPRE) 1319 return false; 1320 1321 // Okay, we have *some* definitions of the value. This means that the value 1322 // is available in some of our (transitive) predecessors. Lets think about 1323 // doing PRE of this load. This will involve inserting a new load into the 1324 // predecessor when it's not available. We could do this in general, but 1325 // prefer to not increase code size. As such, we only do this when we know 1326 // that we only have to insert *one* load (which means we're basically moving 1327 // the load, not inserting a new one). 1328 1329 SmallPtrSet<BasicBlock *, 4> Blockers; 1330 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1331 Blockers.insert(UnavailableBlocks[i]); 1332 1333 // Lets find first basic block with more than one predecessor. Walk backwards 1334 // through predecessors if needed. 1335 BasicBlock *LoadBB = LI->getParent(); 1336 BasicBlock *TmpBB = LoadBB; 1337 1338 bool isSinglePred = false; 1339 bool allSingleSucc = true; 1340 while (TmpBB->getSinglePredecessor()) { 1341 isSinglePred = true; 1342 TmpBB = TmpBB->getSinglePredecessor(); 1343 if (!TmpBB) // If haven't found any, bail now. 1344 return false; 1345 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1346 return false; 1347 if (Blockers.count(TmpBB)) 1348 return false; 1349 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1350 allSingleSucc = false; 1351 } 1352 1353 assert(TmpBB); 1354 LoadBB = TmpBB; 1355 1356 // If we have a repl set with LI itself in it, this means we have a loop where 1357 // at least one of the values is LI. Since this means that we won't be able 1358 // to eliminate LI even if we insert uses in the other predecessors, we will 1359 // end up increasing code size. Reject this by scanning for LI. 1360 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1361 if (ValuesPerBlock[i].V == LI) 1362 return false; 1363 1364 if (isSinglePred) { 1365 bool isHot = false; 1366 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1367 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].V)) 1368 // "Hot" Instruction is in some loop (because it dominates its dep. 1369 // instruction). 1370 if (DT->dominates(LI, I)) { 1371 isHot = true; 1372 break; 1373 } 1374 1375 // We are interested only in "hot" instructions. We don't want to do any 1376 // mis-optimizations here. 1377 if (!isHot) 1378 return false; 1379 } 1380 1381 // Okay, we have some hope :). Check to see if the loaded value is fully 1382 // available in all but one predecessor. 1383 // FIXME: If we could restructure the CFG, we could make a common pred with 1384 // all the preds that don't have an available LI and insert a new load into 1385 // that one block. 1386 BasicBlock *UnavailablePred = 0; 1387 1388 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1389 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1390 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1391 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1392 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1393 1394 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1395 PI != E; ++PI) { 1396 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1397 continue; 1398 1399 // If this load is not available in multiple predecessors, reject it. 1400 if (UnavailablePred && UnavailablePred != *PI) 1401 return false; 1402 UnavailablePred = *PI; 1403 } 1404 1405 assert(UnavailablePred != 0 && 1406 "Fully available value should be eliminated above!"); 1407 1408 // If the loaded pointer is PHI node defined in this block, do PHI translation 1409 // to get its value in the predecessor. 1410 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1411 1412 // Make sure the value is live in the predecessor. If it was defined by a 1413 // non-PHI instruction in this block, we don't know how to recompute it above. 1414 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1415 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1416 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1417 << *LPInst << '\n' << *LI << "\n"); 1418 return false; 1419 } 1420 1421 // We don't currently handle critical edges :( 1422 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1423 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1424 << UnavailablePred->getName() << "': " << *LI << '\n'); 1425 return false; 1426 } 1427 1428 // Make sure it is valid to move this load here. We have to watch out for: 1429 // @1 = getelementptr (i8* p, ... 1430 // test p and branch if == 0 1431 // load @1 1432 // It is valid to have the getelementptr before the test, even if p can be 0, 1433 // as getelementptr only does address arithmetic. 1434 // If we are not pushing the value through any multiple-successor blocks 1435 // we do not have this case. Otherwise, check that the load is safe to 1436 // put anywhere; this can be improved, but should be conservatively safe. 1437 if (!allSingleSucc && 1438 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1439 return false; 1440 1441 // Okay, we can eliminate this load by inserting a reload in the predecessor 1442 // and using PHI construction to get the value in the other predecessors, do 1443 // it. 1444 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1445 1446 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1447 LI->getAlignment(), 1448 UnavailablePred->getTerminator()); 1449 1450 // Add the newly created load. 1451 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad)); 1452 1453 // Perform PHI construction. 1454 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, 1455 VN.getAliasAnalysis()); 1456 LI->replaceAllUsesWith(V); 1457 if (isa<PHINode>(V)) 1458 V->takeName(LI); 1459 if (isa<PointerType>(V->getType())) 1460 MD->invalidateCachedPointerInfo(V); 1461 toErase.push_back(LI); 1462 NumPRELoad++; 1463 return true; 1464} 1465 1466/// processLoad - Attempt to eliminate a load, first by eliminating it 1467/// locally, and then attempting non-local elimination if that fails. 1468bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1469 if (L->isVolatile()) 1470 return false; 1471 1472 // ... to a pointer that has been loaded from before... 1473 MemDepResult Dep = MD->getDependency(L); 1474 1475 // If the value isn't available, don't do anything! 1476 if (Dep.isClobber()) { 1477 // FIXME: We should handle memset/memcpy/memmove as dependent instructions 1478 // to forward the value if available. 1479 //if (isa<MemIntrinsic>(Dep.getInst())) 1480 //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n"; 1481 1482 // Check to see if we have something like this: 1483 // store i32 123, i32* %P 1484 // %A = bitcast i32* %P to i8* 1485 // %B = gep i8* %A, i32 1 1486 // %C = load i8* %B 1487 // 1488 // We could do that by recognizing if the clobber instructions are obviously 1489 // a common base + constant offset, and if the previous store (or memset) 1490 // completely covers this load. This sort of thing can happen in bitfield 1491 // access code. 1492 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1493 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1494 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD); 1495 if (Offset != -1) { 1496 Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1497 L->getType(), L, *TD); 1498 DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n' 1499 << *AvailVal << '\n' << *L << "\n\n\n"); 1500 1501 // Replace the load! 1502 L->replaceAllUsesWith(AvailVal); 1503 if (isa<PointerType>(AvailVal->getType())) 1504 MD->invalidateCachedPointerInfo(AvailVal); 1505 toErase.push_back(L); 1506 NumGVNLoad++; 1507 return true; 1508 } 1509 } 1510 1511 DEBUG( 1512 // fast print dep, using operator<< on instruction would be too slow 1513 errs() << "GVN: load "; 1514 WriteAsOperand(errs(), L); 1515 Instruction *I = Dep.getInst(); 1516 errs() << " is clobbered by " << *I << '\n'; 1517 ); 1518 return false; 1519 } 1520 1521 // If it is defined in another block, try harder. 1522 if (Dep.isNonLocal()) 1523 return processNonLocalLoad(L, toErase); 1524 1525 Instruction *DepInst = Dep.getInst(); 1526 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1527 Value *StoredVal = DepSI->getOperand(0); 1528 1529 // The store and load are to a must-aliased pointer, but they may not 1530 // actually have the same type. See if we know how to reuse the stored 1531 // value (depending on its type). 1532 const TargetData *TD = 0; 1533 if (StoredVal->getType() != L->getType()) { 1534 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1535 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1536 L, *TD); 1537 if (StoredVal == 0) 1538 return false; 1539 1540 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1541 << '\n' << *L << "\n\n\n"); 1542 } 1543 else 1544 return false; 1545 } 1546 1547 // Remove it! 1548 L->replaceAllUsesWith(StoredVal); 1549 if (isa<PointerType>(StoredVal->getType())) 1550 MD->invalidateCachedPointerInfo(StoredVal); 1551 toErase.push_back(L); 1552 NumGVNLoad++; 1553 return true; 1554 } 1555 1556 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1557 Value *AvailableVal = DepLI; 1558 1559 // The loads are of a must-aliased pointer, but they may not actually have 1560 // the same type. See if we know how to reuse the previously loaded value 1561 // (depending on its type). 1562 const TargetData *TD = 0; 1563 if (DepLI->getType() != L->getType()) { 1564 if ((TD = getAnalysisIfAvailable<TargetData>())) { 1565 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1566 if (AvailableVal == 0) 1567 return false; 1568 1569 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1570 << "\n" << *L << "\n\n\n"); 1571 } 1572 else 1573 return false; 1574 } 1575 1576 // Remove it! 1577 L->replaceAllUsesWith(AvailableVal); 1578 if (isa<PointerType>(DepLI->getType())) 1579 MD->invalidateCachedPointerInfo(DepLI); 1580 toErase.push_back(L); 1581 NumGVNLoad++; 1582 return true; 1583 } 1584 1585 // If this load really doesn't depend on anything, then we must be loading an 1586 // undef value. This can happen when loading for a fresh allocation with no 1587 // intervening stores, for example. 1588 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1589 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1590 toErase.push_back(L); 1591 NumGVNLoad++; 1592 return true; 1593 } 1594 1595 return false; 1596} 1597 1598Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1599 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1600 if (I == localAvail.end()) 1601 return 0; 1602 1603 ValueNumberScope *Locals = I->second; 1604 while (Locals) { 1605 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1606 if (I != Locals->table.end()) 1607 return I->second; 1608 Locals = Locals->parent; 1609 } 1610 1611 return 0; 1612} 1613 1614 1615/// processInstruction - When calculating availability, handle an instruction 1616/// by inserting it into the appropriate sets 1617bool GVN::processInstruction(Instruction *I, 1618 SmallVectorImpl<Instruction*> &toErase) { 1619 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1620 bool Changed = processLoad(LI, toErase); 1621 1622 if (!Changed) { 1623 unsigned Num = VN.lookup_or_add(LI); 1624 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1625 } 1626 1627 return Changed; 1628 } 1629 1630 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1631 unsigned Num = VN.lookup_or_add(I); 1632 1633 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1634 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1635 1636 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1637 return false; 1638 1639 Value *BranchCond = BI->getCondition(); 1640 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1641 1642 BasicBlock *TrueSucc = BI->getSuccessor(0); 1643 BasicBlock *FalseSucc = BI->getSuccessor(1); 1644 1645 if (TrueSucc->getSinglePredecessor()) 1646 localAvail[TrueSucc]->table[CondVN] = 1647 ConstantInt::getTrue(TrueSucc->getContext()); 1648 if (FalseSucc->getSinglePredecessor()) 1649 localAvail[FalseSucc]->table[CondVN] = 1650 ConstantInt::getFalse(TrueSucc->getContext()); 1651 1652 return false; 1653 1654 // Allocations are always uniquely numbered, so we can save time and memory 1655 // by fast failing them. 1656 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) { 1657 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1658 return false; 1659 } 1660 1661 // Collapse PHI nodes 1662 if (PHINode* p = dyn_cast<PHINode>(I)) { 1663 Value *constVal = CollapsePhi(p); 1664 1665 if (constVal) { 1666 p->replaceAllUsesWith(constVal); 1667 if (isa<PointerType>(constVal->getType())) 1668 MD->invalidateCachedPointerInfo(constVal); 1669 VN.erase(p); 1670 1671 toErase.push_back(p); 1672 } else { 1673 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1674 } 1675 1676 // If the number we were assigned was a brand new VN, then we don't 1677 // need to do a lookup to see if the number already exists 1678 // somewhere in the domtree: it can't! 1679 } else if (Num == NextNum) { 1680 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1681 1682 // Perform fast-path value-number based elimination of values inherited from 1683 // dominators. 1684 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1685 // Remove it! 1686 VN.erase(I); 1687 I->replaceAllUsesWith(repl); 1688 if (isa<PointerType>(repl->getType())) 1689 MD->invalidateCachedPointerInfo(repl); 1690 toErase.push_back(I); 1691 return true; 1692 1693 } else { 1694 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1695 } 1696 1697 return false; 1698} 1699 1700/// runOnFunction - This is the main transformation entry point for a function. 1701bool GVN::runOnFunction(Function& F) { 1702 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1703 DT = &getAnalysis<DominatorTree>(); 1704 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1705 VN.setMemDep(MD); 1706 VN.setDomTree(DT); 1707 1708 bool Changed = false; 1709 bool ShouldContinue = true; 1710 1711 // Merge unconditional branches, allowing PRE to catch more 1712 // optimization opportunities. 1713 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1714 BasicBlock *BB = FI; 1715 ++FI; 1716 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1717 if (removedBlock) NumGVNBlocks++; 1718 1719 Changed |= removedBlock; 1720 } 1721 1722 unsigned Iteration = 0; 1723 1724 while (ShouldContinue) { 1725 DEBUG(errs() << "GVN iteration: " << Iteration << "\n"); 1726 ShouldContinue = iterateOnFunction(F); 1727 Changed |= ShouldContinue; 1728 ++Iteration; 1729 } 1730 1731 if (EnablePRE) { 1732 bool PREChanged = true; 1733 while (PREChanged) { 1734 PREChanged = performPRE(F); 1735 Changed |= PREChanged; 1736 } 1737 } 1738 // FIXME: Should perform GVN again after PRE does something. PRE can move 1739 // computations into blocks where they become fully redundant. Note that 1740 // we can't do this until PRE's critical edge splitting updates memdep. 1741 // Actually, when this happens, we should just fully integrate PRE into GVN. 1742 1743 cleanupGlobalSets(); 1744 1745 return Changed; 1746} 1747 1748 1749bool GVN::processBlock(BasicBlock *BB) { 1750 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1751 // incrementing BI before processing an instruction). 1752 SmallVector<Instruction*, 8> toErase; 1753 bool ChangedFunction = false; 1754 1755 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1756 BI != BE;) { 1757 ChangedFunction |= processInstruction(BI, toErase); 1758 if (toErase.empty()) { 1759 ++BI; 1760 continue; 1761 } 1762 1763 // If we need some instructions deleted, do it now. 1764 NumGVNInstr += toErase.size(); 1765 1766 // Avoid iterator invalidation. 1767 bool AtStart = BI == BB->begin(); 1768 if (!AtStart) 1769 --BI; 1770 1771 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1772 E = toErase.end(); I != E; ++I) { 1773 DEBUG(errs() << "GVN removed: " << **I << '\n'); 1774 MD->removeInstruction(*I); 1775 (*I)->eraseFromParent(); 1776 DEBUG(verifyRemoved(*I)); 1777 } 1778 toErase.clear(); 1779 1780 if (AtStart) 1781 BI = BB->begin(); 1782 else 1783 ++BI; 1784 } 1785 1786 return ChangedFunction; 1787} 1788 1789/// performPRE - Perform a purely local form of PRE that looks for diamond 1790/// control flow patterns and attempts to perform simple PRE at the join point. 1791bool GVN::performPRE(Function& F) { 1792 bool Changed = false; 1793 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1794 DenseMap<BasicBlock*, Value*> predMap; 1795 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1796 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1797 BasicBlock *CurrentBlock = *DI; 1798 1799 // Nothing to PRE in the entry block. 1800 if (CurrentBlock == &F.getEntryBlock()) continue; 1801 1802 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1803 BE = CurrentBlock->end(); BI != BE; ) { 1804 Instruction *CurInst = BI++; 1805 1806 if (isa<AllocaInst>(CurInst) || 1807 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1808 CurInst->getType()->isVoidTy() || 1809 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1810 isa<DbgInfoIntrinsic>(CurInst)) 1811 continue; 1812 1813 uint32_t ValNo = VN.lookup(CurInst); 1814 1815 // Look for the predecessors for PRE opportunities. We're 1816 // only trying to solve the basic diamond case, where 1817 // a value is computed in the successor and one predecessor, 1818 // but not the other. We also explicitly disallow cases 1819 // where the successor is its own predecessor, because they're 1820 // more complicated to get right. 1821 unsigned NumWith = 0; 1822 unsigned NumWithout = 0; 1823 BasicBlock *PREPred = 0; 1824 predMap.clear(); 1825 1826 for (pred_iterator PI = pred_begin(CurrentBlock), 1827 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1828 // We're not interested in PRE where the block is its 1829 // own predecessor, on in blocks with predecessors 1830 // that are not reachable. 1831 if (*PI == CurrentBlock) { 1832 NumWithout = 2; 1833 break; 1834 } else if (!localAvail.count(*PI)) { 1835 NumWithout = 2; 1836 break; 1837 } 1838 1839 DenseMap<uint32_t, Value*>::iterator predV = 1840 localAvail[*PI]->table.find(ValNo); 1841 if (predV == localAvail[*PI]->table.end()) { 1842 PREPred = *PI; 1843 NumWithout++; 1844 } else if (predV->second == CurInst) { 1845 NumWithout = 2; 1846 } else { 1847 predMap[*PI] = predV->second; 1848 NumWith++; 1849 } 1850 } 1851 1852 // Don't do PRE when it might increase code size, i.e. when 1853 // we would need to insert instructions in more than one pred. 1854 if (NumWithout != 1 || NumWith == 0) 1855 continue; 1856 1857 // We can't do PRE safely on a critical edge, so instead we schedule 1858 // the edge to be split and perform the PRE the next time we iterate 1859 // on the function. 1860 unsigned SuccNum = 0; 1861 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1862 i != e; ++i) 1863 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1864 SuccNum = i; 1865 break; 1866 } 1867 1868 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 1869 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 1870 continue; 1871 } 1872 1873 // Instantiate the expression the in predecessor that lacked it. 1874 // Because we are going top-down through the block, all value numbers 1875 // will be available in the predecessor by the time we need them. Any 1876 // that weren't original present will have been instantiated earlier 1877 // in this loop. 1878 Instruction *PREInstr = CurInst->clone(); 1879 bool success = true; 1880 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1881 Value *Op = PREInstr->getOperand(i); 1882 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1883 continue; 1884 1885 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1886 PREInstr->setOperand(i, V); 1887 } else { 1888 success = false; 1889 break; 1890 } 1891 } 1892 1893 // Fail out if we encounter an operand that is not available in 1894 // the PRE predecessor. This is typically because of loads which 1895 // are not value numbered precisely. 1896 if (!success) { 1897 delete PREInstr; 1898 DEBUG(verifyRemoved(PREInstr)); 1899 continue; 1900 } 1901 1902 PREInstr->insertBefore(PREPred->getTerminator()); 1903 PREInstr->setName(CurInst->getName() + ".pre"); 1904 predMap[PREPred] = PREInstr; 1905 VN.add(PREInstr, ValNo); 1906 NumGVNPRE++; 1907 1908 // Update the availability map to include the new instruction. 1909 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 1910 1911 // Create a PHI to make the value available in this block. 1912 PHINode* Phi = PHINode::Create(CurInst->getType(), 1913 CurInst->getName() + ".pre-phi", 1914 CurrentBlock->begin()); 1915 for (pred_iterator PI = pred_begin(CurrentBlock), 1916 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1917 Phi->addIncoming(predMap[*PI], *PI); 1918 1919 VN.add(Phi, ValNo); 1920 localAvail[CurrentBlock]->table[ValNo] = Phi; 1921 1922 CurInst->replaceAllUsesWith(Phi); 1923 if (isa<PointerType>(Phi->getType())) 1924 MD->invalidateCachedPointerInfo(Phi); 1925 VN.erase(CurInst); 1926 1927 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n'); 1928 MD->removeInstruction(CurInst); 1929 CurInst->eraseFromParent(); 1930 DEBUG(verifyRemoved(CurInst)); 1931 Changed = true; 1932 } 1933 } 1934 1935 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1936 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1937 SplitCriticalEdge(I->first, I->second, this); 1938 1939 return Changed || toSplit.size(); 1940} 1941 1942/// iterateOnFunction - Executes one iteration of GVN 1943bool GVN::iterateOnFunction(Function &F) { 1944 cleanupGlobalSets(); 1945 1946 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1947 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1948 if (DI->getIDom()) 1949 localAvail[DI->getBlock()] = 1950 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1951 else 1952 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1953 } 1954 1955 // Top-down walk of the dominator tree 1956 bool Changed = false; 1957#if 0 1958 // Needed for value numbering with phi construction to work. 1959 ReversePostOrderTraversal<Function*> RPOT(&F); 1960 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1961 RE = RPOT.end(); RI != RE; ++RI) 1962 Changed |= processBlock(*RI); 1963#else 1964 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1965 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1966 Changed |= processBlock(DI->getBlock()); 1967#endif 1968 1969 return Changed; 1970} 1971 1972void GVN::cleanupGlobalSets() { 1973 VN.clear(); 1974 1975 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1976 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1977 delete I->second; 1978 localAvail.clear(); 1979} 1980 1981/// verifyRemoved - Verify that the specified instruction does not occur in our 1982/// internal data structures. 1983void GVN::verifyRemoved(const Instruction *Inst) const { 1984 VN.verifyRemoved(Inst); 1985 1986 // Walk through the value number scope to make sure the instruction isn't 1987 // ferreted away in it. 1988 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1989 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1990 const ValueNumberScope *VNS = I->second; 1991 1992 while (VNS) { 1993 for (DenseMap<uint32_t, Value*>::iterator 1994 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1995 assert(II->second != Inst && "Inst still in value numbering scope!"); 1996 } 1997 1998 VNS = VNS->parent; 1999 } 2000 } 2001} 2002