GVN.cpp revision 158d86e5b0d0a11bfce694b16bafb60c7f9507ff
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/MallocHelper.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 EMPTY, TOMBSTONE }; 82 83 ExpressionOpcode opcode; 84 const Type* type; 85 uint32_t firstVN; 86 uint32_t secondVN; 87 uint32_t thirdVN; 88 SmallVector<uint32_t, 4> varargs; 89 Value *function; 90 91 Expression() { } 92 Expression(ExpressionOpcode o) : opcode(o) { } 93 94 bool operator==(const Expression &other) const { 95 if (opcode != other.opcode) 96 return false; 97 else if (opcode == EMPTY || opcode == TOMBSTONE) 98 return true; 99 else if (type != other.type) 100 return false; 101 else if (function != other.function) 102 return false; 103 else if (firstVN != other.firstVN) 104 return false; 105 else if (secondVN != other.secondVN) 106 return false; 107 else if (thirdVN != other.thirdVN) 108 return false; 109 else { 110 if (varargs.size() != other.varargs.size()) 111 return false; 112 113 for (size_t i = 0; i < varargs.size(); ++i) 114 if (varargs[i] != other.varargs[i]) 115 return false; 116 117 return true; 118 } 119 } 120 121 bool operator!=(const Expression &other) const { 122 return !(*this == other); 123 } 124 }; 125 126 class ValueTable { 127 private: 128 DenseMap<Value*, uint32_t> valueNumbering; 129 DenseMap<Expression, uint32_t> expressionNumbering; 130 AliasAnalysis* AA; 131 MemoryDependenceAnalysis* MD; 132 DominatorTree* DT; 133 134 uint32_t nextValueNumber; 135 136 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 137 Expression::ExpressionOpcode getOpcode(CmpInst* C); 138 Expression::ExpressionOpcode getOpcode(CastInst* C); 139 Expression create_expression(BinaryOperator* BO); 140 Expression create_expression(CmpInst* C); 141 Expression create_expression(ShuffleVectorInst* V); 142 Expression create_expression(ExtractElementInst* C); 143 Expression create_expression(InsertElementInst* V); 144 Expression create_expression(SelectInst* V); 145 Expression create_expression(CastInst* C); 146 Expression create_expression(GetElementPtrInst* G); 147 Expression create_expression(CallInst* C); 148 Expression create_expression(Constant* C); 149 public: 150 ValueTable() : nextValueNumber(1) { } 151 uint32_t lookup_or_add(Value *V); 152 uint32_t lookup(Value *V) const; 153 void add(Value *V, uint32_t num); 154 void clear(); 155 void erase(Value *v); 156 unsigned size(); 157 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 158 AliasAnalysis *getAliasAnalysis() const { return AA; } 159 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 160 void setDomTree(DominatorTree* D) { DT = D; } 161 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 162 void verifyRemoved(const Value *) const; 163 }; 164} 165 166namespace llvm { 167template <> struct DenseMapInfo<Expression> { 168 static inline Expression getEmptyKey() { 169 return Expression(Expression::EMPTY); 170 } 171 172 static inline Expression getTombstoneKey() { 173 return Expression(Expression::TOMBSTONE); 174 } 175 176 static unsigned getHashValue(const Expression e) { 177 unsigned hash = e.opcode; 178 179 hash = e.firstVN + hash * 37; 180 hash = e.secondVN + hash * 37; 181 hash = e.thirdVN + hash * 37; 182 183 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 184 (unsigned)((uintptr_t)e.type >> 9)) + 185 hash * 37; 186 187 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 188 E = e.varargs.end(); I != E; ++I) 189 hash = *I + hash * 37; 190 191 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 192 (unsigned)((uintptr_t)e.function >> 9)) + 193 hash * 37; 194 195 return hash; 196 } 197 static bool isEqual(const Expression &LHS, const Expression &RHS) { 198 return LHS == RHS; 199 } 200 static bool isPod() { return true; } 201}; 202} 203 204//===----------------------------------------------------------------------===// 205// ValueTable Internal Functions 206//===----------------------------------------------------------------------===// 207Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 208 switch(BO->getOpcode()) { 209 default: // THIS SHOULD NEVER HAPPEN 210 llvm_unreachable("Binary operator with unknown opcode?"); 211 case Instruction::Add: return Expression::ADD; 212 case Instruction::FAdd: return Expression::FADD; 213 case Instruction::Sub: return Expression::SUB; 214 case Instruction::FSub: return Expression::FSUB; 215 case Instruction::Mul: return Expression::MUL; 216 case Instruction::FMul: return Expression::FMUL; 217 case Instruction::UDiv: return Expression::UDIV; 218 case Instruction::SDiv: return Expression::SDIV; 219 case Instruction::FDiv: return Expression::FDIV; 220 case Instruction::URem: return Expression::UREM; 221 case Instruction::SRem: return Expression::SREM; 222 case Instruction::FRem: return Expression::FREM; 223 case Instruction::Shl: return Expression::SHL; 224 case Instruction::LShr: return Expression::LSHR; 225 case Instruction::AShr: return Expression::ASHR; 226 case Instruction::And: return Expression::AND; 227 case Instruction::Or: return Expression::OR; 228 case Instruction::Xor: return Expression::XOR; 229 } 230} 231 232Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 233 if (isa<ICmpInst>(C)) { 234 switch (C->getPredicate()) { 235 default: // THIS SHOULD NEVER HAPPEN 236 llvm_unreachable("Comparison with unknown predicate?"); 237 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 238 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 239 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 240 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 241 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 242 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 243 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 244 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 245 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 246 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 247 } 248 } else { 249 switch (C->getPredicate()) { 250 default: // THIS SHOULD NEVER HAPPEN 251 llvm_unreachable("Comparison with unknown predicate?"); 252 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 253 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 254 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 255 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 256 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 257 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 258 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 259 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 260 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 261 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 262 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 263 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 264 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 265 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 266 } 267 } 268} 269 270Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 271 switch(C->getOpcode()) { 272 default: // THIS SHOULD NEVER HAPPEN 273 llvm_unreachable("Cast operator with unknown opcode?"); 274 case Instruction::Trunc: return Expression::TRUNC; 275 case Instruction::ZExt: return Expression::ZEXT; 276 case Instruction::SExt: return Expression::SEXT; 277 case Instruction::FPToUI: return Expression::FPTOUI; 278 case Instruction::FPToSI: return Expression::FPTOSI; 279 case Instruction::UIToFP: return Expression::UITOFP; 280 case Instruction::SIToFP: return Expression::SITOFP; 281 case Instruction::FPTrunc: return Expression::FPTRUNC; 282 case Instruction::FPExt: return Expression::FPEXT; 283 case Instruction::PtrToInt: return Expression::PTRTOINT; 284 case Instruction::IntToPtr: return Expression::INTTOPTR; 285 case Instruction::BitCast: return Expression::BITCAST; 286 } 287} 288 289Expression ValueTable::create_expression(CallInst* C) { 290 Expression e; 291 292 e.type = C->getType(); 293 e.firstVN = 0; 294 e.secondVN = 0; 295 e.thirdVN = 0; 296 e.function = C->getCalledFunction(); 297 e.opcode = Expression::CALL; 298 299 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 300 I != E; ++I) 301 e.varargs.push_back(lookup_or_add(*I)); 302 303 return e; 304} 305 306Expression ValueTable::create_expression(BinaryOperator* BO) { 307 Expression e; 308 309 e.firstVN = lookup_or_add(BO->getOperand(0)); 310 e.secondVN = lookup_or_add(BO->getOperand(1)); 311 e.thirdVN = 0; 312 e.function = 0; 313 e.type = BO->getType(); 314 e.opcode = getOpcode(BO); 315 316 return e; 317} 318 319Expression ValueTable::create_expression(CmpInst* C) { 320 Expression e; 321 322 e.firstVN = lookup_or_add(C->getOperand(0)); 323 e.secondVN = lookup_or_add(C->getOperand(1)); 324 e.thirdVN = 0; 325 e.function = 0; 326 e.type = C->getType(); 327 e.opcode = getOpcode(C); 328 329 return e; 330} 331 332Expression ValueTable::create_expression(CastInst* C) { 333 Expression e; 334 335 e.firstVN = lookup_or_add(C->getOperand(0)); 336 e.secondVN = 0; 337 e.thirdVN = 0; 338 e.function = 0; 339 e.type = C->getType(); 340 e.opcode = getOpcode(C); 341 342 return e; 343} 344 345Expression ValueTable::create_expression(ShuffleVectorInst* S) { 346 Expression e; 347 348 e.firstVN = lookup_or_add(S->getOperand(0)); 349 e.secondVN = lookup_or_add(S->getOperand(1)); 350 e.thirdVN = lookup_or_add(S->getOperand(2)); 351 e.function = 0; 352 e.type = S->getType(); 353 e.opcode = Expression::SHUFFLE; 354 355 return e; 356} 357 358Expression ValueTable::create_expression(ExtractElementInst* E) { 359 Expression e; 360 361 e.firstVN = lookup_or_add(E->getOperand(0)); 362 e.secondVN = lookup_or_add(E->getOperand(1)); 363 e.thirdVN = 0; 364 e.function = 0; 365 e.type = E->getType(); 366 e.opcode = Expression::EXTRACT; 367 368 return e; 369} 370 371Expression ValueTable::create_expression(InsertElementInst* I) { 372 Expression e; 373 374 e.firstVN = lookup_or_add(I->getOperand(0)); 375 e.secondVN = lookup_or_add(I->getOperand(1)); 376 e.thirdVN = lookup_or_add(I->getOperand(2)); 377 e.function = 0; 378 e.type = I->getType(); 379 e.opcode = Expression::INSERT; 380 381 return e; 382} 383 384Expression ValueTable::create_expression(SelectInst* I) { 385 Expression e; 386 387 e.firstVN = lookup_or_add(I->getCondition()); 388 e.secondVN = lookup_or_add(I->getTrueValue()); 389 e.thirdVN = lookup_or_add(I->getFalseValue()); 390 e.function = 0; 391 e.type = I->getType(); 392 e.opcode = Expression::SELECT; 393 394 return e; 395} 396 397Expression ValueTable::create_expression(GetElementPtrInst* G) { 398 Expression e; 399 400 e.firstVN = lookup_or_add(G->getPointerOperand()); 401 e.secondVN = 0; 402 e.thirdVN = 0; 403 e.function = 0; 404 e.type = G->getType(); 405 e.opcode = Expression::GEP; 406 407 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 408 I != E; ++I) 409 e.varargs.push_back(lookup_or_add(*I)); 410 411 return e; 412} 413 414//===----------------------------------------------------------------------===// 415// ValueTable External Functions 416//===----------------------------------------------------------------------===// 417 418/// add - Insert a value into the table with a specified value number. 419void ValueTable::add(Value *V, uint32_t num) { 420 valueNumbering.insert(std::make_pair(V, num)); 421} 422 423/// lookup_or_add - Returns the value number for the specified value, assigning 424/// it a new number if it did not have one before. 425uint32_t ValueTable::lookup_or_add(Value *V) { 426 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 427 if (VI != valueNumbering.end()) 428 return VI->second; 429 430 if (CallInst* C = dyn_cast<CallInst>(V)) { 431 if (AA->doesNotAccessMemory(C)) { 432 Expression exp = create_expression(C); 433 uint32_t& e = expressionNumbering[exp]; 434 if (!e) e = nextValueNumber++; 435 valueNumbering[V] = e; 436 return e; 437 } else if (AA->onlyReadsMemory(C)) { 438 Expression exp = create_expression(C); 439 uint32_t& e = expressionNumbering[exp]; 440 if (!e) { 441 e = nextValueNumber++; 442 valueNumbering[V] = e; 443 return e; 444 } 445 446 MemDepResult local_dep = MD->getDependency(C); 447 448 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 449 valueNumbering[V] = nextValueNumber; 450 return nextValueNumber++; 451 } 452 453 if (local_dep.isDef()) { 454 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 455 456 if (local_cdep->getNumOperands() != C->getNumOperands()) { 457 valueNumbering[V] = nextValueNumber; 458 return nextValueNumber++; 459 } 460 461 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 462 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 463 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 464 if (c_vn != cd_vn) { 465 valueNumbering[V] = nextValueNumber; 466 return nextValueNumber++; 467 } 468 } 469 470 uint32_t v = lookup_or_add(local_cdep); 471 valueNumbering[V] = v; 472 return v; 473 } 474 475 // Non-local case. 476 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 477 MD->getNonLocalCallDependency(CallSite(C)); 478 // FIXME: call/call dependencies for readonly calls should return def, not 479 // clobber! Move the checking logic to MemDep! 480 CallInst* cdep = 0; 481 482 // Check to see if we have a single dominating call instruction that is 483 // identical to C. 484 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 485 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 486 // Ignore non-local dependencies. 487 if (I->second.isNonLocal()) 488 continue; 489 490 // We don't handle non-depedencies. If we already have a call, reject 491 // instruction dependencies. 492 if (I->second.isClobber() || cdep != 0) { 493 cdep = 0; 494 break; 495 } 496 497 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 498 // FIXME: All duplicated with non-local case. 499 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 500 cdep = NonLocalDepCall; 501 continue; 502 } 503 504 cdep = 0; 505 break; 506 } 507 508 if (!cdep) { 509 valueNumbering[V] = nextValueNumber; 510 return nextValueNumber++; 511 } 512 513 if (cdep->getNumOperands() != C->getNumOperands()) { 514 valueNumbering[V] = nextValueNumber; 515 return nextValueNumber++; 516 } 517 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 518 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 519 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 520 if (c_vn != cd_vn) { 521 valueNumbering[V] = nextValueNumber; 522 return nextValueNumber++; 523 } 524 } 525 526 uint32_t v = lookup_or_add(cdep); 527 valueNumbering[V] = v; 528 return v; 529 530 } else { 531 valueNumbering[V] = nextValueNumber; 532 return nextValueNumber++; 533 } 534 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 535 Expression exp = create_expression(BO); 536 uint32_t& e = expressionNumbering[exp]; 537 if (!e) e = nextValueNumber++; 538 valueNumbering[V] = e; 539 return e; 540 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 541 Expression exp = create_expression(C); 542 uint32_t& e = expressionNumbering[exp]; 543 if (!e) e = nextValueNumber++; 544 valueNumbering[V] = e; 545 return e; 546 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 547 Expression exp = create_expression(U); 548 uint32_t& e = expressionNumbering[exp]; 549 if (!e) e = nextValueNumber++; 550 valueNumbering[V] = e; 551 return e; 552 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 553 Expression exp = create_expression(U); 554 uint32_t& e = expressionNumbering[exp]; 555 if (!e) e = nextValueNumber++; 556 valueNumbering[V] = e; 557 return e; 558 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 559 Expression exp = create_expression(U); 560 uint32_t& e = expressionNumbering[exp]; 561 if (!e) e = nextValueNumber++; 562 valueNumbering[V] = e; 563 return e; 564 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 565 Expression exp = create_expression(U); 566 uint32_t& e = expressionNumbering[exp]; 567 if (!e) e = nextValueNumber++; 568 valueNumbering[V] = e; 569 return e; 570 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 571 Expression exp = create_expression(U); 572 uint32_t& e = expressionNumbering[exp]; 573 if (!e) e = nextValueNumber++; 574 valueNumbering[V] = e; 575 return e; 576 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 577 Expression exp = create_expression(U); 578 uint32_t& e = expressionNumbering[exp]; 579 if (!e) e = nextValueNumber++; 580 valueNumbering[V] = e; 581 return e; 582 } else { 583 valueNumbering[V] = nextValueNumber; 584 return nextValueNumber++; 585 } 586} 587 588/// lookup - Returns the value number of the specified value. Fails if 589/// the value has not yet been numbered. 590uint32_t ValueTable::lookup(Value *V) const { 591 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 592 assert(VI != valueNumbering.end() && "Value not numbered?"); 593 return VI->second; 594} 595 596/// clear - Remove all entries from the ValueTable 597void ValueTable::clear() { 598 valueNumbering.clear(); 599 expressionNumbering.clear(); 600 nextValueNumber = 1; 601} 602 603/// erase - Remove a value from the value numbering 604void ValueTable::erase(Value *V) { 605 valueNumbering.erase(V); 606} 607 608/// verifyRemoved - Verify that the value is removed from all internal data 609/// structures. 610void ValueTable::verifyRemoved(const Value *V) const { 611 for (DenseMap<Value*, uint32_t>::iterator 612 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 613 assert(I->first != V && "Inst still occurs in value numbering map!"); 614 } 615} 616 617//===----------------------------------------------------------------------===// 618// GVN Pass 619//===----------------------------------------------------------------------===// 620 621namespace { 622 struct ValueNumberScope { 623 ValueNumberScope* parent; 624 DenseMap<uint32_t, Value*> table; 625 626 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 627 }; 628} 629 630namespace { 631 632 class GVN : public FunctionPass { 633 bool runOnFunction(Function &F); 634 public: 635 static char ID; // Pass identification, replacement for typeid 636 GVN() : FunctionPass(&ID) { } 637 638 private: 639 MemoryDependenceAnalysis *MD; 640 DominatorTree *DT; 641 642 ValueTable VN; 643 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 644 645 // This transformation requires dominator postdominator info 646 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 647 AU.addRequired<DominatorTree>(); 648 AU.addRequired<MemoryDependenceAnalysis>(); 649 AU.addRequired<AliasAnalysis>(); 650 651 AU.addPreserved<DominatorTree>(); 652 AU.addPreserved<AliasAnalysis>(); 653 } 654 655 // Helper fuctions 656 // FIXME: eliminate or document these better 657 bool processLoad(LoadInst* L, 658 SmallVectorImpl<Instruction*> &toErase); 659 bool processInstruction(Instruction *I, 660 SmallVectorImpl<Instruction*> &toErase); 661 bool processNonLocalLoad(LoadInst* L, 662 SmallVectorImpl<Instruction*> &toErase); 663 bool processBlock(BasicBlock *BB); 664 void dump(DenseMap<uint32_t, Value*>& d); 665 bool iterateOnFunction(Function &F); 666 Value *CollapsePhi(PHINode* p); 667 bool performPRE(Function& F); 668 Value *lookupNumber(BasicBlock *BB, uint32_t num); 669 void cleanupGlobalSets(); 670 void verifyRemoved(const Instruction *I) const; 671 }; 672 673 char GVN::ID = 0; 674} 675 676// createGVNPass - The public interface to this file... 677FunctionPass *llvm::createGVNPass() { return new GVN(); } 678 679static RegisterPass<GVN> X("gvn", 680 "Global Value Numbering"); 681 682void GVN::dump(DenseMap<uint32_t, Value*>& d) { 683 printf("{\n"); 684 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 685 E = d.end(); I != E; ++I) { 686 printf("%d\n", I->first); 687 I->second->dump(); 688 } 689 printf("}\n"); 690} 691 692static bool isSafeReplacement(PHINode* p, Instruction *inst) { 693 if (!isa<PHINode>(inst)) 694 return true; 695 696 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 697 UI != E; ++UI) 698 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 699 if (use_phi->getParent() == inst->getParent()) 700 return false; 701 702 return true; 703} 704 705Value *GVN::CollapsePhi(PHINode *PN) { 706 Value *ConstVal = PN->hasConstantValue(DT); 707 if (!ConstVal) return 0; 708 709 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 710 if (!Inst) 711 return ConstVal; 712 713 if (DT->dominates(Inst, PN)) 714 if (isSafeReplacement(PN, Inst)) 715 return Inst; 716 return 0; 717} 718 719/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 720/// we're analyzing is fully available in the specified block. As we go, keep 721/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 722/// map is actually a tri-state map with the following values: 723/// 0) we know the block *is not* fully available. 724/// 1) we know the block *is* fully available. 725/// 2) we do not know whether the block is fully available or not, but we are 726/// currently speculating that it will be. 727/// 3) we are speculating for this block and have used that to speculate for 728/// other blocks. 729static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 730 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 731 // Optimistically assume that the block is fully available and check to see 732 // if we already know about this block in one lookup. 733 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 734 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 735 736 // If the entry already existed for this block, return the precomputed value. 737 if (!IV.second) { 738 // If this is a speculative "available" value, mark it as being used for 739 // speculation of other blocks. 740 if (IV.first->second == 2) 741 IV.first->second = 3; 742 return IV.first->second != 0; 743 } 744 745 // Otherwise, see if it is fully available in all predecessors. 746 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 747 748 // If this block has no predecessors, it isn't live-in here. 749 if (PI == PE) 750 goto SpeculationFailure; 751 752 for (; PI != PE; ++PI) 753 // If the value isn't fully available in one of our predecessors, then it 754 // isn't fully available in this block either. Undo our previous 755 // optimistic assumption and bail out. 756 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 757 goto SpeculationFailure; 758 759 return true; 760 761// SpeculationFailure - If we get here, we found out that this is not, after 762// all, a fully-available block. We have a problem if we speculated on this and 763// used the speculation to mark other blocks as available. 764SpeculationFailure: 765 char &BBVal = FullyAvailableBlocks[BB]; 766 767 // If we didn't speculate on this, just return with it set to false. 768 if (BBVal == 2) { 769 BBVal = 0; 770 return false; 771 } 772 773 // If we did speculate on this value, we could have blocks set to 1 that are 774 // incorrect. Walk the (transitive) successors of this block and mark them as 775 // 0 if set to one. 776 SmallVector<BasicBlock*, 32> BBWorklist; 777 BBWorklist.push_back(BB); 778 779 while (!BBWorklist.empty()) { 780 BasicBlock *Entry = BBWorklist.pop_back_val(); 781 // Note that this sets blocks to 0 (unavailable) if they happen to not 782 // already be in FullyAvailableBlocks. This is safe. 783 char &EntryVal = FullyAvailableBlocks[Entry]; 784 if (EntryVal == 0) continue; // Already unavailable. 785 786 // Mark as unavailable. 787 EntryVal = 0; 788 789 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 790 BBWorklist.push_back(*I); 791 } 792 793 return false; 794} 795 796 797/// CanCoerceMustAliasedValueToLoad - Return true if 798/// CoerceAvailableValueToLoadType will succeed. 799static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 800 const Type *LoadTy, 801 const TargetData &TD) { 802 // If the loaded or stored value is an first class array or struct, don't try 803 // to transform them. We need to be able to bitcast to integer. 804 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) || 805 isa<StructType>(StoredVal->getType()) || 806 isa<ArrayType>(StoredVal->getType())) 807 return false; 808 809 // The store has to be at least as big as the load. 810 if (TD.getTypeSizeInBits(StoredVal->getType()) < 811 TD.getTypeSizeInBits(LoadTy)) 812 return false; 813 814 return true; 815} 816 817 818/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 819/// then a load from a must-aliased pointer of a different type, try to coerce 820/// the stored value. LoadedTy is the type of the load we want to replace and 821/// InsertPt is the place to insert new instructions. 822/// 823/// If we can't do it, return null. 824static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 825 const Type *LoadedTy, 826 Instruction *InsertPt, 827 const TargetData &TD) { 828 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 829 return 0; 830 831 const Type *StoredValTy = StoredVal->getType(); 832 833 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 834 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 835 836 // If the store and reload are the same size, we can always reuse it. 837 if (StoreSize == LoadSize) { 838 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) { 839 // Pointer to Pointer -> use bitcast. 840 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 841 } 842 843 // Convert source pointers to integers, which can be bitcast. 844 if (isa<PointerType>(StoredValTy)) { 845 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 846 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 847 } 848 849 const Type *TypeToCastTo = LoadedTy; 850 if (isa<PointerType>(TypeToCastTo)) 851 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 852 853 if (StoredValTy != TypeToCastTo) 854 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 855 856 // Cast to pointer if the load needs a pointer type. 857 if (isa<PointerType>(LoadedTy)) 858 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 859 860 return StoredVal; 861 } 862 863 // If the loaded value is smaller than the available value, then we can 864 // extract out a piece from it. If the available value is too small, then we 865 // can't do anything. 866 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 867 868 // Convert source pointers to integers, which can be manipulated. 869 if (isa<PointerType>(StoredValTy)) { 870 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 871 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 872 } 873 874 // Convert vectors and fp to integer, which can be manipulated. 875 if (!isa<IntegerType>(StoredValTy)) { 876 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 877 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 878 } 879 880 // If this is a big-endian system, we need to shift the value down to the low 881 // bits so that a truncate will work. 882 if (TD.isBigEndian()) { 883 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 884 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 885 } 886 887 // Truncate the integer to the right size now. 888 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 889 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 890 891 if (LoadedTy == NewIntTy) 892 return StoredVal; 893 894 // If the result is a pointer, inttoptr. 895 if (isa<PointerType>(LoadedTy)) 896 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 897 898 // Otherwise, bitcast. 899 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 900} 901 902/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 903/// be expressed as a base pointer plus a constant offset. Return the base and 904/// offset to the caller. 905static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 906 const TargetData &TD) { 907 Operator *PtrOp = dyn_cast<Operator>(Ptr); 908 if (PtrOp == 0) return Ptr; 909 910 // Just look through bitcasts. 911 if (PtrOp->getOpcode() == Instruction::BitCast) 912 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 913 914 // If this is a GEP with constant indices, we can look through it. 915 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 916 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 917 918 gep_type_iterator GTI = gep_type_begin(GEP); 919 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 920 ++I, ++GTI) { 921 ConstantInt *OpC = cast<ConstantInt>(*I); 922 if (OpC->isZero()) continue; 923 924 // Handle a struct and array indices which add their offset to the pointer. 925 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 926 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 927 } else { 928 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 929 Offset += OpC->getSExtValue()*Size; 930 } 931 } 932 933 // Re-sign extend from the pointer size if needed to get overflow edge cases 934 // right. 935 unsigned PtrSize = TD.getPointerSizeInBits(); 936 if (PtrSize < 64) 937 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 938 939 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 940} 941 942 943/// AnalyzeLoadFromClobberingStore - This function is called when we have a 944/// memdep query of a load that ends up being a clobbering store. This means 945/// that the store *may* provide bits used by the load but we can't be sure 946/// because the pointers don't mustalias. Check this case to see if there is 947/// anything more we can do before we give up. This returns -1 if we have to 948/// give up, or a byte number in the stored value of the piece that feeds the 949/// load. 950static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI, 951 const TargetData &TD) { 952 // If the loaded or stored value is an first class array or struct, don't try 953 // to transform them. We need to be able to bitcast to integer. 954 if (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()) || 955 isa<StructType>(DepSI->getOperand(0)->getType()) || 956 isa<ArrayType>(DepSI->getOperand(0)->getType())) 957 return -1; 958 959 int64_t StoreOffset = 0, LoadOffset = 0; 960 Value *StoreBase = 961 GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD); 962 Value *LoadBase = 963 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD); 964 if (StoreBase != LoadBase) 965 return -1; 966 967 // If the load and store are to the exact same address, they should have been 968 // a must alias. AA must have gotten confused. 969 // FIXME: Study to see if/when this happens. 970 if (LoadOffset == StoreOffset) { 971#if 0 972 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 973 << "Base = " << *StoreBase << "\n" 974 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 975 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 976 << "Load Ptr = " << *L->getPointerOperand() << "\n" 977 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 978 errs() << "'" << L->getParent()->getParent()->getName() << "'" 979 << *L->getParent(); 980#endif 981 return -1; 982 } 983 984 // If the load and store don't overlap at all, the store doesn't provide 985 // anything to the load. In this case, they really don't alias at all, AA 986 // must have gotten confused. 987 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 988 // remove this check, as it is duplicated with what we have below. 989 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType()); 990 uint64_t LoadSize = TD.getTypeSizeInBits(L->getType()); 991 992 if ((StoreSize & 7) | (LoadSize & 7)) 993 return -1; 994 StoreSize >>= 3; // Convert to bytes. 995 LoadSize >>= 3; 996 997 998 bool isAAFailure = false; 999 if (StoreOffset < LoadOffset) { 1000 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1001 } else { 1002 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1003 } 1004 if (isAAFailure) { 1005#if 0 1006 errs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1007 << "Base = " << *StoreBase << "\n" 1008 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 1009 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 1010 << "Load Ptr = " << *L->getPointerOperand() << "\n" 1011 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 1012 errs() << "'" << L->getParent()->getParent()->getName() << "'" 1013 << *L->getParent(); 1014#endif 1015 return -1; 1016 } 1017 1018 // If the Load isn't completely contained within the stored bits, we don't 1019 // have all the bits to feed it. We could do something crazy in the future 1020 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1021 // valuable. 1022 if (StoreOffset > LoadOffset || 1023 StoreOffset+StoreSize < LoadOffset+LoadSize) 1024 return -1; 1025 1026 // Okay, we can do this transformation. Return the number of bytes into the 1027 // store that the load is. 1028 return LoadOffset-StoreOffset; 1029} 1030 1031 1032/// GetStoreValueForLoad - This function is called when we have a 1033/// memdep query of a load that ends up being a clobbering store. This means 1034/// that the store *may* provide bits used by the load but we can't be sure 1035/// because the pointers don't mustalias. Check this case to see if there is 1036/// anything more we can do before we give up. 1037static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 1038 const Type *LoadTy, 1039 Instruction *InsertPt, const TargetData &TD){ 1040 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1041 1042 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8; 1043 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 1044 1045 1046 // Compute which bits of the stored value are being used by the load. Convert 1047 // to an integer type to start with. 1048 if (isa<PointerType>(SrcVal->getType())) 1049 SrcVal = new PtrToIntInst(SrcVal, TD.getIntPtrType(Ctx), "tmp", InsertPt); 1050 if (!isa<IntegerType>(SrcVal->getType())) 1051 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1052 "tmp", InsertPt); 1053 1054 // Shift the bits to the least significant depending on endianness. 1055 unsigned ShiftAmt; 1056 if (TD.isLittleEndian()) { 1057 ShiftAmt = Offset*8; 1058 } else { 1059 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 1060 } 1061 1062 if (ShiftAmt) 1063 SrcVal = BinaryOperator::CreateLShr(SrcVal, 1064 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt); 1065 1066 if (LoadSize != StoreSize) 1067 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1068 "tmp", InsertPt); 1069 1070 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 1071} 1072 1073struct AvailableValueInBlock { 1074 /// BB - The basic block in question. 1075 BasicBlock *BB; 1076 /// V - The value that is live out of the block. 1077 Value *V; 1078 /// Offset - The byte offset in V that is interesting for the load query. 1079 unsigned Offset; 1080 1081 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1082 unsigned Offset = 0) { 1083 AvailableValueInBlock Res; 1084 Res.BB = BB; 1085 Res.V = V; 1086 Res.Offset = Offset; 1087 return Res; 1088 } 1089}; 1090 1091/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1092/// construct SSA form, allowing us to eliminate LI. This returns the value 1093/// that should be used at LI's definition site. 1094static Value *ConstructSSAForLoadSet(LoadInst *LI, 1095 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1096 const TargetData *TD, 1097 AliasAnalysis *AA) { 1098 SmallVector<PHINode*, 8> NewPHIs; 1099 SSAUpdater SSAUpdate(&NewPHIs); 1100 SSAUpdate.Initialize(LI); 1101 1102 const Type *LoadTy = LI->getType(); 1103 1104 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1105 BasicBlock *BB = ValuesPerBlock[i].BB; 1106 Value *AvailableVal = ValuesPerBlock[i].V; 1107 unsigned Offset = ValuesPerBlock[i].Offset; 1108 1109 if (SSAUpdate.HasValueForBlock(BB)) 1110 continue; 1111 1112 if (AvailableVal->getType() != LoadTy) { 1113 assert(TD && "Need target data to handle type mismatch case"); 1114 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy, 1115 BB->getTerminator(), *TD); 1116 1117 if (Offset) { 1118 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1119 << *ValuesPerBlock[i].V << '\n' 1120 << *AvailableVal << '\n' << "\n\n\n"); 1121 } 1122 1123 1124 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1125 << *ValuesPerBlock[i].V << '\n' 1126 << *AvailableVal << '\n' << "\n\n\n"); 1127 } 1128 1129 SSAUpdate.AddAvailableValue(BB, AvailableVal); 1130 } 1131 1132 // Perform PHI construction. 1133 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1134 1135 // If new PHI nodes were created, notify alias analysis. 1136 if (isa<PointerType>(V->getType())) 1137 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1138 AA->copyValue(LI, NewPHIs[i]); 1139 1140 return V; 1141} 1142 1143/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1144/// non-local by performing PHI construction. 1145bool GVN::processNonLocalLoad(LoadInst *LI, 1146 SmallVectorImpl<Instruction*> &toErase) { 1147 // Find the non-local dependencies of the load. 1148 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 1149 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1150 Deps); 1151 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: " 1152 // << Deps.size() << *LI << '\n'); 1153 1154 // If we had to process more than one hundred blocks to find the 1155 // dependencies, this load isn't worth worrying about. Optimizing 1156 // it will be too expensive. 1157 if (Deps.size() > 100) 1158 return false; 1159 1160 // If we had a phi translation failure, we'll have a single entry which is a 1161 // clobber in the current block. Reject this early. 1162 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 1163 DEBUG( 1164 errs() << "GVN: non-local load "; 1165 WriteAsOperand(errs(), LI); 1166 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n'; 1167 ); 1168 return false; 1169 } 1170 1171 // Filter out useless results (non-locals, etc). Keep track of the blocks 1172 // where we have a value available in repl, also keep track of whether we see 1173 // dependencies that produce an unknown value for the load (such as a call 1174 // that could potentially clobber the load). 1175 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1176 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1177 1178 const TargetData *TD = 0; 1179 1180 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1181 BasicBlock *DepBB = Deps[i].first; 1182 MemDepResult DepInfo = Deps[i].second; 1183 1184 if (DepInfo.isClobber()) { 1185 // If the dependence is to a store that writes to a superset of the bits 1186 // read by the load, we can extract the bits we need for the load from the 1187 // stored value. 1188 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1189 if (TD == 0) 1190 TD = getAnalysisIfAvailable<TargetData>(); 1191 if (TD) { 1192 int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD); 1193 if (Offset != -1) { 1194 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1195 DepSI->getOperand(0), 1196 Offset)); 1197 continue; 1198 } 1199 } 1200 } 1201 1202 // FIXME: Handle memset/memcpy. 1203 UnavailableBlocks.push_back(DepBB); 1204 continue; 1205 } 1206 1207 Instruction *DepInst = DepInfo.getInst(); 1208 1209 // Loading the allocation -> undef. 1210 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1211 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1212 UndefValue::get(LI->getType()))); 1213 continue; 1214 } 1215 1216 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1217 // Reject loads and stores that are to the same address but are of 1218 // different types if we have to. 1219 if (S->getOperand(0)->getType() != LI->getType()) { 1220 if (TD == 0) 1221 TD = getAnalysisIfAvailable<TargetData>(); 1222 1223 // If the stored value is larger or equal to the loaded value, we can 1224 // reuse it. 1225 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0), 1226 LI->getType(), *TD)) { 1227 UnavailableBlocks.push_back(DepBB); 1228 continue; 1229 } 1230 } 1231 1232 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1233 S->getOperand(0))); 1234 continue; 1235 } 1236 1237 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1238 // If the types mismatch and we can't handle it, reject reuse of the load. 1239 if (LD->getType() != LI->getType()) { 1240 if (TD == 0) 1241 TD = getAnalysisIfAvailable<TargetData>(); 1242 1243 // If the stored value is larger or equal to the loaded value, we can 1244 // reuse it. 1245 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1246 UnavailableBlocks.push_back(DepBB); 1247 continue; 1248 } 1249 } 1250 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD)); 1251 continue; 1252 } 1253 1254 UnavailableBlocks.push_back(DepBB); 1255 continue; 1256 } 1257 1258 // If we have no predecessors that produce a known value for this load, exit 1259 // early. 1260 if (ValuesPerBlock.empty()) return false; 1261 1262 // If all of the instructions we depend on produce a known value for this 1263 // load, then it is fully redundant and we can use PHI insertion to compute 1264 // its value. Insert PHIs and remove the fully redundant value now. 1265 if (UnavailableBlocks.empty()) { 1266 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1267 1268 // Perform PHI construction. 1269 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, 1270 VN.getAliasAnalysis()); 1271 LI->replaceAllUsesWith(V); 1272 1273 if (isa<PHINode>(V)) 1274 V->takeName(LI); 1275 if (isa<PointerType>(V->getType())) 1276 MD->invalidateCachedPointerInfo(V); 1277 toErase.push_back(LI); 1278 NumGVNLoad++; 1279 return true; 1280 } 1281 1282 if (!EnablePRE || !EnableLoadPRE) 1283 return false; 1284 1285 // Okay, we have *some* definitions of the value. This means that the value 1286 // is available in some of our (transitive) predecessors. Lets think about 1287 // doing PRE of this load. This will involve inserting a new load into the 1288 // predecessor when it's not available. We could do this in general, but 1289 // prefer to not increase code size. As such, we only do this when we know 1290 // that we only have to insert *one* load (which means we're basically moving 1291 // the load, not inserting a new one). 1292 1293 SmallPtrSet<BasicBlock *, 4> Blockers; 1294 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1295 Blockers.insert(UnavailableBlocks[i]); 1296 1297 // Lets find first basic block with more than one predecessor. Walk backwards 1298 // through predecessors if needed. 1299 BasicBlock *LoadBB = LI->getParent(); 1300 BasicBlock *TmpBB = LoadBB; 1301 1302 bool isSinglePred = false; 1303 bool allSingleSucc = true; 1304 while (TmpBB->getSinglePredecessor()) { 1305 isSinglePred = true; 1306 TmpBB = TmpBB->getSinglePredecessor(); 1307 if (!TmpBB) // If haven't found any, bail now. 1308 return false; 1309 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1310 return false; 1311 if (Blockers.count(TmpBB)) 1312 return false; 1313 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1314 allSingleSucc = false; 1315 } 1316 1317 assert(TmpBB); 1318 LoadBB = TmpBB; 1319 1320 // If we have a repl set with LI itself in it, this means we have a loop where 1321 // at least one of the values is LI. Since this means that we won't be able 1322 // to eliminate LI even if we insert uses in the other predecessors, we will 1323 // end up increasing code size. Reject this by scanning for LI. 1324 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1325 if (ValuesPerBlock[i].V == LI) 1326 return false; 1327 1328 if (isSinglePred) { 1329 bool isHot = false; 1330 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1331 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].V)) 1332 // "Hot" Instruction is in some loop (because it dominates its dep. 1333 // instruction). 1334 if (DT->dominates(LI, I)) { 1335 isHot = true; 1336 break; 1337 } 1338 1339 // We are interested only in "hot" instructions. We don't want to do any 1340 // mis-optimizations here. 1341 if (!isHot) 1342 return false; 1343 } 1344 1345 // Okay, we have some hope :). Check to see if the loaded value is fully 1346 // available in all but one predecessor. 1347 // FIXME: If we could restructure the CFG, we could make a common pred with 1348 // all the preds that don't have an available LI and insert a new load into 1349 // that one block. 1350 BasicBlock *UnavailablePred = 0; 1351 1352 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1353 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1354 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1355 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1356 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1357 1358 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1359 PI != E; ++PI) { 1360 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1361 continue; 1362 1363 // If this load is not available in multiple predecessors, reject it. 1364 if (UnavailablePred && UnavailablePred != *PI) 1365 return false; 1366 UnavailablePred = *PI; 1367 } 1368 1369 assert(UnavailablePred != 0 && 1370 "Fully available value should be eliminated above!"); 1371 1372 // If the loaded pointer is PHI node defined in this block, do PHI translation 1373 // to get its value in the predecessor. 1374 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1375 1376 // Make sure the value is live in the predecessor. If it was defined by a 1377 // non-PHI instruction in this block, we don't know how to recompute it above. 1378 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1379 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1380 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1381 << *LPInst << '\n' << *LI << "\n"); 1382 return false; 1383 } 1384 1385 // We don't currently handle critical edges :( 1386 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1387 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1388 << UnavailablePred->getName() << "': " << *LI << '\n'); 1389 return false; 1390 } 1391 1392 // Make sure it is valid to move this load here. We have to watch out for: 1393 // @1 = getelementptr (i8* p, ... 1394 // test p and branch if == 0 1395 // load @1 1396 // It is valid to have the getelementptr before the test, even if p can be 0, 1397 // as getelementptr only does address arithmetic. 1398 // If we are not pushing the value through any multiple-successor blocks 1399 // we do not have this case. Otherwise, check that the load is safe to 1400 // put anywhere; this can be improved, but should be conservatively safe. 1401 if (!allSingleSucc && 1402 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1403 return false; 1404 1405 // Okay, we can eliminate this load by inserting a reload in the predecessor 1406 // and using PHI construction to get the value in the other predecessors, do 1407 // it. 1408 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1409 1410 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1411 LI->getAlignment(), 1412 UnavailablePred->getTerminator()); 1413 1414 // Add the newly created load. 1415 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad)); 1416 1417 // Perform PHI construction. 1418 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, 1419 VN.getAliasAnalysis()); 1420 LI->replaceAllUsesWith(V); 1421 if (isa<PHINode>(V)) 1422 V->takeName(LI); 1423 if (isa<PointerType>(V->getType())) 1424 MD->invalidateCachedPointerInfo(V); 1425 toErase.push_back(LI); 1426 NumPRELoad++; 1427 return true; 1428} 1429 1430/// processLoad - Attempt to eliminate a load, first by eliminating it 1431/// locally, and then attempting non-local elimination if that fails. 1432bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1433 if (L->isVolatile()) 1434 return false; 1435 1436 // ... to a pointer that has been loaded from before... 1437 MemDepResult Dep = MD->getDependency(L); 1438 1439 // If the value isn't available, don't do anything! 1440 if (Dep.isClobber()) { 1441 // FIXME: We should handle memset/memcpy/memmove as dependent instructions 1442 // to forward the value if available. 1443 //if (isa<MemIntrinsic>(Dep.getInst())) 1444 //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n"; 1445 1446 // Check to see if we have something like this: 1447 // store i32 123, i32* %P 1448 // %A = bitcast i32* %P to i8* 1449 // %B = gep i8* %A, i32 1 1450 // %C = load i8* %B 1451 // 1452 // We could do that by recognizing if the clobber instructions are obviously 1453 // a common base + constant offset, and if the previous store (or memset) 1454 // completely covers this load. This sort of thing can happen in bitfield 1455 // access code. 1456 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1457 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1458 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD); 1459 if (Offset != -1) { 1460 Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1461 L->getType(), L, *TD); 1462 DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n' 1463 << *AvailVal << '\n' << *L << "\n\n\n"); 1464 1465 // Replace the load! 1466 L->replaceAllUsesWith(AvailVal); 1467 if (isa<PointerType>(AvailVal->getType())) 1468 MD->invalidateCachedPointerInfo(AvailVal); 1469 toErase.push_back(L); 1470 NumGVNLoad++; 1471 return true; 1472 } 1473 } 1474 1475 DEBUG( 1476 // fast print dep, using operator<< on instruction would be too slow 1477 errs() << "GVN: load "; 1478 WriteAsOperand(errs(), L); 1479 Instruction *I = Dep.getInst(); 1480 errs() << " is clobbered by " << *I << '\n'; 1481 ); 1482 return false; 1483 } 1484 1485 // If it is defined in another block, try harder. 1486 if (Dep.isNonLocal()) 1487 return processNonLocalLoad(L, toErase); 1488 1489 Instruction *DepInst = Dep.getInst(); 1490 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1491 Value *StoredVal = DepSI->getOperand(0); 1492 1493 // The store and load are to a must-aliased pointer, but they may not 1494 // actually have the same type. See if we know how to reuse the stored 1495 // value (depending on its type). 1496 const TargetData *TD = 0; 1497 if (StoredVal->getType() != L->getType() && 1498 (TD = getAnalysisIfAvailable<TargetData>())) { 1499 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1500 L, *TD); 1501 if (StoredVal == 0) 1502 return false; 1503 1504 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1505 << '\n' << *L << "\n\n\n"); 1506 } 1507 1508 // Remove it! 1509 L->replaceAllUsesWith(StoredVal); 1510 if (isa<PointerType>(StoredVal->getType())) 1511 MD->invalidateCachedPointerInfo(StoredVal); 1512 toErase.push_back(L); 1513 NumGVNLoad++; 1514 return true; 1515 } 1516 1517 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1518 Value *AvailableVal = DepLI; 1519 1520 // The loads are of a must-aliased pointer, but they may not actually have 1521 // the same type. See if we know how to reuse the previously loaded value 1522 // (depending on its type). 1523 const TargetData *TD = 0; 1524 if (DepLI->getType() != L->getType() && 1525 (TD = getAnalysisIfAvailable<TargetData>())) { 1526 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD); 1527 if (AvailableVal == 0) 1528 return false; 1529 1530 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1531 << "\n" << *L << "\n\n\n"); 1532 } 1533 1534 // Remove it! 1535 L->replaceAllUsesWith(AvailableVal); 1536 if (isa<PointerType>(DepLI->getType())) 1537 MD->invalidateCachedPointerInfo(DepLI); 1538 toErase.push_back(L); 1539 NumGVNLoad++; 1540 return true; 1541 } 1542 1543 // If this load really doesn't depend on anything, then we must be loading an 1544 // undef value. This can happen when loading for a fresh allocation with no 1545 // intervening stores, for example. 1546 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1547 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1548 toErase.push_back(L); 1549 NumGVNLoad++; 1550 return true; 1551 } 1552 1553 return false; 1554} 1555 1556Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1557 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1558 if (I == localAvail.end()) 1559 return 0; 1560 1561 ValueNumberScope *Locals = I->second; 1562 while (Locals) { 1563 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1564 if (I != Locals->table.end()) 1565 return I->second; 1566 Locals = Locals->parent; 1567 } 1568 1569 return 0; 1570} 1571 1572 1573/// processInstruction - When calculating availability, handle an instruction 1574/// by inserting it into the appropriate sets 1575bool GVN::processInstruction(Instruction *I, 1576 SmallVectorImpl<Instruction*> &toErase) { 1577 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1578 bool Changed = processLoad(LI, toErase); 1579 1580 if (!Changed) { 1581 unsigned Num = VN.lookup_or_add(LI); 1582 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1583 } 1584 1585 return Changed; 1586 } 1587 1588 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1589 unsigned Num = VN.lookup_or_add(I); 1590 1591 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1592 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1593 1594 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1595 return false; 1596 1597 Value *BranchCond = BI->getCondition(); 1598 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1599 1600 BasicBlock *TrueSucc = BI->getSuccessor(0); 1601 BasicBlock *FalseSucc = BI->getSuccessor(1); 1602 1603 if (TrueSucc->getSinglePredecessor()) 1604 localAvail[TrueSucc]->table[CondVN] = 1605 ConstantInt::getTrue(TrueSucc->getContext()); 1606 if (FalseSucc->getSinglePredecessor()) 1607 localAvail[FalseSucc]->table[CondVN] = 1608 ConstantInt::getFalse(TrueSucc->getContext()); 1609 1610 return false; 1611 1612 // Allocations are always uniquely numbered, so we can save time and memory 1613 // by fast failing them. 1614 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) { 1615 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1616 return false; 1617 } 1618 1619 // Collapse PHI nodes 1620 if (PHINode* p = dyn_cast<PHINode>(I)) { 1621 Value *constVal = CollapsePhi(p); 1622 1623 if (constVal) { 1624 p->replaceAllUsesWith(constVal); 1625 if (isa<PointerType>(constVal->getType())) 1626 MD->invalidateCachedPointerInfo(constVal); 1627 VN.erase(p); 1628 1629 toErase.push_back(p); 1630 } else { 1631 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1632 } 1633 1634 // If the number we were assigned was a brand new VN, then we don't 1635 // need to do a lookup to see if the number already exists 1636 // somewhere in the domtree: it can't! 1637 } else if (Num == NextNum) { 1638 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1639 1640 // Perform fast-path value-number based elimination of values inherited from 1641 // dominators. 1642 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1643 // Remove it! 1644 VN.erase(I); 1645 I->replaceAllUsesWith(repl); 1646 if (isa<PointerType>(repl->getType())) 1647 MD->invalidateCachedPointerInfo(repl); 1648 toErase.push_back(I); 1649 return true; 1650 1651 } else { 1652 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1653 } 1654 1655 return false; 1656} 1657 1658/// runOnFunction - This is the main transformation entry point for a function. 1659bool GVN::runOnFunction(Function& F) { 1660 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1661 DT = &getAnalysis<DominatorTree>(); 1662 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1663 VN.setMemDep(MD); 1664 VN.setDomTree(DT); 1665 1666 bool Changed = false; 1667 bool ShouldContinue = true; 1668 1669 // Merge unconditional branches, allowing PRE to catch more 1670 // optimization opportunities. 1671 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1672 BasicBlock *BB = FI; 1673 ++FI; 1674 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1675 if (removedBlock) NumGVNBlocks++; 1676 1677 Changed |= removedBlock; 1678 } 1679 1680 unsigned Iteration = 0; 1681 1682 while (ShouldContinue) { 1683 DEBUG(errs() << "GVN iteration: " << Iteration << "\n"); 1684 ShouldContinue = iterateOnFunction(F); 1685 Changed |= ShouldContinue; 1686 ++Iteration; 1687 } 1688 1689 if (EnablePRE) { 1690 bool PREChanged = true; 1691 while (PREChanged) { 1692 PREChanged = performPRE(F); 1693 Changed |= PREChanged; 1694 } 1695 } 1696 // FIXME: Should perform GVN again after PRE does something. PRE can move 1697 // computations into blocks where they become fully redundant. Note that 1698 // we can't do this until PRE's critical edge splitting updates memdep. 1699 // Actually, when this happens, we should just fully integrate PRE into GVN. 1700 1701 cleanupGlobalSets(); 1702 1703 return Changed; 1704} 1705 1706 1707bool GVN::processBlock(BasicBlock *BB) { 1708 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1709 // incrementing BI before processing an instruction). 1710 SmallVector<Instruction*, 8> toErase; 1711 bool ChangedFunction = false; 1712 1713 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1714 BI != BE;) { 1715 ChangedFunction |= processInstruction(BI, toErase); 1716 if (toErase.empty()) { 1717 ++BI; 1718 continue; 1719 } 1720 1721 // If we need some instructions deleted, do it now. 1722 NumGVNInstr += toErase.size(); 1723 1724 // Avoid iterator invalidation. 1725 bool AtStart = BI == BB->begin(); 1726 if (!AtStart) 1727 --BI; 1728 1729 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1730 E = toErase.end(); I != E; ++I) { 1731 DEBUG(errs() << "GVN removed: " << **I << '\n'); 1732 MD->removeInstruction(*I); 1733 (*I)->eraseFromParent(); 1734 DEBUG(verifyRemoved(*I)); 1735 } 1736 toErase.clear(); 1737 1738 if (AtStart) 1739 BI = BB->begin(); 1740 else 1741 ++BI; 1742 } 1743 1744 return ChangedFunction; 1745} 1746 1747/// performPRE - Perform a purely local form of PRE that looks for diamond 1748/// control flow patterns and attempts to perform simple PRE at the join point. 1749bool GVN::performPRE(Function& F) { 1750 bool Changed = false; 1751 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1752 DenseMap<BasicBlock*, Value*> predMap; 1753 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1754 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1755 BasicBlock *CurrentBlock = *DI; 1756 1757 // Nothing to PRE in the entry block. 1758 if (CurrentBlock == &F.getEntryBlock()) continue; 1759 1760 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1761 BE = CurrentBlock->end(); BI != BE; ) { 1762 Instruction *CurInst = BI++; 1763 1764 if (isa<AllocationInst>(CurInst) || 1765 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1766 CurInst->getType()->isVoidTy() || 1767 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1768 isa<DbgInfoIntrinsic>(CurInst)) 1769 continue; 1770 1771 uint32_t ValNo = VN.lookup(CurInst); 1772 1773 // Look for the predecessors for PRE opportunities. We're 1774 // only trying to solve the basic diamond case, where 1775 // a value is computed in the successor and one predecessor, 1776 // but not the other. We also explicitly disallow cases 1777 // where the successor is its own predecessor, because they're 1778 // more complicated to get right. 1779 unsigned NumWith = 0; 1780 unsigned NumWithout = 0; 1781 BasicBlock *PREPred = 0; 1782 predMap.clear(); 1783 1784 for (pred_iterator PI = pred_begin(CurrentBlock), 1785 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1786 // We're not interested in PRE where the block is its 1787 // own predecessor, on in blocks with predecessors 1788 // that are not reachable. 1789 if (*PI == CurrentBlock) { 1790 NumWithout = 2; 1791 break; 1792 } else if (!localAvail.count(*PI)) { 1793 NumWithout = 2; 1794 break; 1795 } 1796 1797 DenseMap<uint32_t, Value*>::iterator predV = 1798 localAvail[*PI]->table.find(ValNo); 1799 if (predV == localAvail[*PI]->table.end()) { 1800 PREPred = *PI; 1801 NumWithout++; 1802 } else if (predV->second == CurInst) { 1803 NumWithout = 2; 1804 } else { 1805 predMap[*PI] = predV->second; 1806 NumWith++; 1807 } 1808 } 1809 1810 // Don't do PRE when it might increase code size, i.e. when 1811 // we would need to insert instructions in more than one pred. 1812 if (NumWithout != 1 || NumWith == 0) 1813 continue; 1814 1815 // We can't do PRE safely on a critical edge, so instead we schedule 1816 // the edge to be split and perform the PRE the next time we iterate 1817 // on the function. 1818 unsigned SuccNum = 0; 1819 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1820 i != e; ++i) 1821 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1822 SuccNum = i; 1823 break; 1824 } 1825 1826 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 1827 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 1828 continue; 1829 } 1830 1831 // Instantiate the expression the in predecessor that lacked it. 1832 // Because we are going top-down through the block, all value numbers 1833 // will be available in the predecessor by the time we need them. Any 1834 // that weren't original present will have been instantiated earlier 1835 // in this loop. 1836 Instruction *PREInstr = CurInst->clone(); 1837 bool success = true; 1838 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1839 Value *Op = PREInstr->getOperand(i); 1840 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1841 continue; 1842 1843 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1844 PREInstr->setOperand(i, V); 1845 } else { 1846 success = false; 1847 break; 1848 } 1849 } 1850 1851 // Fail out if we encounter an operand that is not available in 1852 // the PRE predecessor. This is typically because of loads which 1853 // are not value numbered precisely. 1854 if (!success) { 1855 delete PREInstr; 1856 DEBUG(verifyRemoved(PREInstr)); 1857 continue; 1858 } 1859 1860 PREInstr->insertBefore(PREPred->getTerminator()); 1861 PREInstr->setName(CurInst->getName() + ".pre"); 1862 predMap[PREPred] = PREInstr; 1863 VN.add(PREInstr, ValNo); 1864 NumGVNPRE++; 1865 1866 // Update the availability map to include the new instruction. 1867 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 1868 1869 // Create a PHI to make the value available in this block. 1870 PHINode* Phi = PHINode::Create(CurInst->getType(), 1871 CurInst->getName() + ".pre-phi", 1872 CurrentBlock->begin()); 1873 for (pred_iterator PI = pred_begin(CurrentBlock), 1874 PE = pred_end(CurrentBlock); PI != PE; ++PI) 1875 Phi->addIncoming(predMap[*PI], *PI); 1876 1877 VN.add(Phi, ValNo); 1878 localAvail[CurrentBlock]->table[ValNo] = Phi; 1879 1880 CurInst->replaceAllUsesWith(Phi); 1881 if (isa<PointerType>(Phi->getType())) 1882 MD->invalidateCachedPointerInfo(Phi); 1883 VN.erase(CurInst); 1884 1885 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n'); 1886 MD->removeInstruction(CurInst); 1887 CurInst->eraseFromParent(); 1888 DEBUG(verifyRemoved(CurInst)); 1889 Changed = true; 1890 } 1891 } 1892 1893 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 1894 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 1895 SplitCriticalEdge(I->first, I->second, this); 1896 1897 return Changed || toSplit.size(); 1898} 1899 1900/// iterateOnFunction - Executes one iteration of GVN 1901bool GVN::iterateOnFunction(Function &F) { 1902 cleanupGlobalSets(); 1903 1904 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1905 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 1906 if (DI->getIDom()) 1907 localAvail[DI->getBlock()] = 1908 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 1909 else 1910 localAvail[DI->getBlock()] = new ValueNumberScope(0); 1911 } 1912 1913 // Top-down walk of the dominator tree 1914 bool Changed = false; 1915#if 0 1916 // Needed for value numbering with phi construction to work. 1917 ReversePostOrderTraversal<Function*> RPOT(&F); 1918 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 1919 RE = RPOT.end(); RI != RE; ++RI) 1920 Changed |= processBlock(*RI); 1921#else 1922 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 1923 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 1924 Changed |= processBlock(DI->getBlock()); 1925#endif 1926 1927 return Changed; 1928} 1929 1930void GVN::cleanupGlobalSets() { 1931 VN.clear(); 1932 1933 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1934 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 1935 delete I->second; 1936 localAvail.clear(); 1937} 1938 1939/// verifyRemoved - Verify that the specified instruction does not occur in our 1940/// internal data structures. 1941void GVN::verifyRemoved(const Instruction *Inst) const { 1942 VN.verifyRemoved(Inst); 1943 1944 // Walk through the value number scope to make sure the instruction isn't 1945 // ferreted away in it. 1946 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 1947 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 1948 const ValueNumberScope *VNS = I->second; 1949 1950 while (VNS) { 1951 for (DenseMap<uint32_t, Value*>::iterator 1952 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 1953 assert(II->second != Inst && "Inst still in value numbering scope!"); 1954 } 1955 1956 VNS = VNS->parent; 1957 } 1958 } 1959} 1960