GVN.cpp revision 1ce0829e83290bdafc6dbe0427708b509d3c8c0e
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 <cstdio> 48using namespace llvm; 49 50STATISTIC(NumGVNInstr, "Number of instructions deleted"); 51STATISTIC(NumGVNLoad, "Number of loads deleted"); 52STATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 53STATISTIC(NumGVNBlocks, "Number of blocks merged"); 54STATISTIC(NumPRELoad, "Number of loads PRE'd"); 55 56static cl::opt<bool> EnablePRE("enable-pre", 57 cl::init(true), cl::Hidden); 58static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 59 60//===----------------------------------------------------------------------===// 61// ValueTable Class 62//===----------------------------------------------------------------------===// 63 64/// This class holds the mapping between values and value numbers. It is used 65/// as an efficient mechanism to determine the expression-wise equivalence of 66/// two values. 67namespace { 68 struct Expression { 69 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL, 70 UDIV, SDIV, FDIV, UREM, SREM, 71 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 72 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 73 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 74 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 75 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 76 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT, 77 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI, 78 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 79 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT, 80 EMPTY, TOMBSTONE }; 81 82 ExpressionOpcode opcode; 83 const Type* type; 84 uint32_t firstVN; 85 uint32_t secondVN; 86 uint32_t thirdVN; 87 SmallVector<uint32_t, 4> varargs; 88 Value *function; 89 90 Expression() { } 91 Expression(ExpressionOpcode o) : opcode(o) { } 92 93 bool operator==(const Expression &other) const { 94 if (opcode != other.opcode) 95 return false; 96 else if (opcode == EMPTY || opcode == TOMBSTONE) 97 return true; 98 else if (type != other.type) 99 return false; 100 else if (function != other.function) 101 return false; 102 else if (firstVN != other.firstVN) 103 return false; 104 else if (secondVN != other.secondVN) 105 return false; 106 else if (thirdVN != other.thirdVN) 107 return false; 108 else { 109 if (varargs.size() != other.varargs.size()) 110 return false; 111 112 for (size_t i = 0; i < varargs.size(); ++i) 113 if (varargs[i] != other.varargs[i]) 114 return false; 115 116 return true; 117 } 118 } 119 120 bool operator!=(const Expression &other) const { 121 return !(*this == other); 122 } 123 }; 124 125 class ValueTable { 126 private: 127 DenseMap<Value*, uint32_t> valueNumbering; 128 DenseMap<Expression, uint32_t> expressionNumbering; 129 AliasAnalysis* AA; 130 MemoryDependenceAnalysis* MD; 131 DominatorTree* DT; 132 133 uint32_t nextValueNumber; 134 135 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO); 136 Expression::ExpressionOpcode getOpcode(CmpInst* C); 137 Expression::ExpressionOpcode getOpcode(CastInst* C); 138 Expression create_expression(BinaryOperator* BO); 139 Expression create_expression(CmpInst* C); 140 Expression create_expression(ShuffleVectorInst* V); 141 Expression create_expression(ExtractElementInst* C); 142 Expression create_expression(InsertElementInst* V); 143 Expression create_expression(SelectInst* V); 144 Expression create_expression(CastInst* C); 145 Expression create_expression(GetElementPtrInst* G); 146 Expression create_expression(CallInst* C); 147 Expression create_expression(Constant* C); 148 public: 149 ValueTable() : nextValueNumber(1) { } 150 uint32_t lookup_or_add(Value *V); 151 uint32_t lookup(Value *V) const; 152 void add(Value *V, uint32_t num); 153 void clear(); 154 void erase(Value *v); 155 unsigned size(); 156 void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 157 AliasAnalysis *getAliasAnalysis() const { return AA; } 158 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 159 void setDomTree(DominatorTree* D) { DT = D; } 160 uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 161 void verifyRemoved(const Value *) const; 162 }; 163} 164 165namespace llvm { 166template <> struct DenseMapInfo<Expression> { 167 static inline Expression getEmptyKey() { 168 return Expression(Expression::EMPTY); 169 } 170 171 static inline Expression getTombstoneKey() { 172 return Expression(Expression::TOMBSTONE); 173 } 174 175 static unsigned getHashValue(const Expression e) { 176 unsigned hash = e.opcode; 177 178 hash = e.firstVN + hash * 37; 179 hash = e.secondVN + hash * 37; 180 hash = e.thirdVN + hash * 37; 181 182 hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 183 (unsigned)((uintptr_t)e.type >> 9)) + 184 hash * 37; 185 186 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 187 E = e.varargs.end(); I != E; ++I) 188 hash = *I + hash * 37; 189 190 hash = ((unsigned)((uintptr_t)e.function >> 4) ^ 191 (unsigned)((uintptr_t)e.function >> 9)) + 192 hash * 37; 193 194 return hash; 195 } 196 static bool isEqual(const Expression &LHS, const Expression &RHS) { 197 return LHS == RHS; 198 } 199 static bool isPod() { return true; } 200}; 201} 202 203//===----------------------------------------------------------------------===// 204// ValueTable Internal Functions 205//===----------------------------------------------------------------------===// 206Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) { 207 switch(BO->getOpcode()) { 208 default: // THIS SHOULD NEVER HAPPEN 209 llvm_unreachable("Binary operator with unknown opcode?"); 210 case Instruction::Add: return Expression::ADD; 211 case Instruction::FAdd: return Expression::FADD; 212 case Instruction::Sub: return Expression::SUB; 213 case Instruction::FSub: return Expression::FSUB; 214 case Instruction::Mul: return Expression::MUL; 215 case Instruction::FMul: return Expression::FMUL; 216 case Instruction::UDiv: return Expression::UDIV; 217 case Instruction::SDiv: return Expression::SDIV; 218 case Instruction::FDiv: return Expression::FDIV; 219 case Instruction::URem: return Expression::UREM; 220 case Instruction::SRem: return Expression::SREM; 221 case Instruction::FRem: return Expression::FREM; 222 case Instruction::Shl: return Expression::SHL; 223 case Instruction::LShr: return Expression::LSHR; 224 case Instruction::AShr: return Expression::ASHR; 225 case Instruction::And: return Expression::AND; 226 case Instruction::Or: return Expression::OR; 227 case Instruction::Xor: return Expression::XOR; 228 } 229} 230 231Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) { 232 if (isa<ICmpInst>(C)) { 233 switch (C->getPredicate()) { 234 default: // THIS SHOULD NEVER HAPPEN 235 llvm_unreachable("Comparison with unknown predicate?"); 236 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ; 237 case ICmpInst::ICMP_NE: return Expression::ICMPNE; 238 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT; 239 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE; 240 case ICmpInst::ICMP_ULT: return Expression::ICMPULT; 241 case ICmpInst::ICMP_ULE: return Expression::ICMPULE; 242 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT; 243 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE; 244 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT; 245 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE; 246 } 247 } else { 248 switch (C->getPredicate()) { 249 default: // THIS SHOULD NEVER HAPPEN 250 llvm_unreachable("Comparison with unknown predicate?"); 251 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ; 252 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT; 253 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE; 254 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT; 255 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE; 256 case FCmpInst::FCMP_ONE: return Expression::FCMPONE; 257 case FCmpInst::FCMP_ORD: return Expression::FCMPORD; 258 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO; 259 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ; 260 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT; 261 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE; 262 case FCmpInst::FCMP_ULT: return Expression::FCMPULT; 263 case FCmpInst::FCMP_ULE: return Expression::FCMPULE; 264 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE; 265 } 266 } 267} 268 269Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) { 270 switch(C->getOpcode()) { 271 default: // THIS SHOULD NEVER HAPPEN 272 llvm_unreachable("Cast operator with unknown opcode?"); 273 case Instruction::Trunc: return Expression::TRUNC; 274 case Instruction::ZExt: return Expression::ZEXT; 275 case Instruction::SExt: return Expression::SEXT; 276 case Instruction::FPToUI: return Expression::FPTOUI; 277 case Instruction::FPToSI: return Expression::FPTOSI; 278 case Instruction::UIToFP: return Expression::UITOFP; 279 case Instruction::SIToFP: return Expression::SITOFP; 280 case Instruction::FPTrunc: return Expression::FPTRUNC; 281 case Instruction::FPExt: return Expression::FPEXT; 282 case Instruction::PtrToInt: return Expression::PTRTOINT; 283 case Instruction::IntToPtr: return Expression::INTTOPTR; 284 case Instruction::BitCast: return Expression::BITCAST; 285 } 286} 287 288Expression ValueTable::create_expression(CallInst* C) { 289 Expression e; 290 291 e.type = C->getType(); 292 e.firstVN = 0; 293 e.secondVN = 0; 294 e.thirdVN = 0; 295 e.function = C->getCalledFunction(); 296 e.opcode = Expression::CALL; 297 298 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end(); 299 I != E; ++I) 300 e.varargs.push_back(lookup_or_add(*I)); 301 302 return e; 303} 304 305Expression ValueTable::create_expression(BinaryOperator* BO) { 306 Expression e; 307 308 e.firstVN = lookup_or_add(BO->getOperand(0)); 309 e.secondVN = lookup_or_add(BO->getOperand(1)); 310 e.thirdVN = 0; 311 e.function = 0; 312 e.type = BO->getType(); 313 e.opcode = getOpcode(BO); 314 315 return e; 316} 317 318Expression ValueTable::create_expression(CmpInst* C) { 319 Expression e; 320 321 e.firstVN = lookup_or_add(C->getOperand(0)); 322 e.secondVN = lookup_or_add(C->getOperand(1)); 323 e.thirdVN = 0; 324 e.function = 0; 325 e.type = C->getType(); 326 e.opcode = getOpcode(C); 327 328 return e; 329} 330 331Expression ValueTable::create_expression(CastInst* C) { 332 Expression e; 333 334 e.firstVN = lookup_or_add(C->getOperand(0)); 335 e.secondVN = 0; 336 e.thirdVN = 0; 337 e.function = 0; 338 e.type = C->getType(); 339 e.opcode = getOpcode(C); 340 341 return e; 342} 343 344Expression ValueTable::create_expression(ShuffleVectorInst* S) { 345 Expression e; 346 347 e.firstVN = lookup_or_add(S->getOperand(0)); 348 e.secondVN = lookup_or_add(S->getOperand(1)); 349 e.thirdVN = lookup_or_add(S->getOperand(2)); 350 e.function = 0; 351 e.type = S->getType(); 352 e.opcode = Expression::SHUFFLE; 353 354 return e; 355} 356 357Expression ValueTable::create_expression(ExtractElementInst* E) { 358 Expression e; 359 360 e.firstVN = lookup_or_add(E->getOperand(0)); 361 e.secondVN = lookup_or_add(E->getOperand(1)); 362 e.thirdVN = 0; 363 e.function = 0; 364 e.type = E->getType(); 365 e.opcode = Expression::EXTRACT; 366 367 return e; 368} 369 370Expression ValueTable::create_expression(InsertElementInst* I) { 371 Expression e; 372 373 e.firstVN = lookup_or_add(I->getOperand(0)); 374 e.secondVN = lookup_or_add(I->getOperand(1)); 375 e.thirdVN = lookup_or_add(I->getOperand(2)); 376 e.function = 0; 377 e.type = I->getType(); 378 e.opcode = Expression::INSERT; 379 380 return e; 381} 382 383Expression ValueTable::create_expression(SelectInst* I) { 384 Expression e; 385 386 e.firstVN = lookup_or_add(I->getCondition()); 387 e.secondVN = lookup_or_add(I->getTrueValue()); 388 e.thirdVN = lookup_or_add(I->getFalseValue()); 389 e.function = 0; 390 e.type = I->getType(); 391 e.opcode = Expression::SELECT; 392 393 return e; 394} 395 396Expression ValueTable::create_expression(GetElementPtrInst* G) { 397 Expression e; 398 399 e.firstVN = lookup_or_add(G->getPointerOperand()); 400 e.secondVN = 0; 401 e.thirdVN = 0; 402 e.function = 0; 403 e.type = G->getType(); 404 e.opcode = Expression::GEP; 405 406 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end(); 407 I != E; ++I) 408 e.varargs.push_back(lookup_or_add(*I)); 409 410 return e; 411} 412 413//===----------------------------------------------------------------------===// 414// ValueTable External Functions 415//===----------------------------------------------------------------------===// 416 417/// add - Insert a value into the table with a specified value number. 418void ValueTable::add(Value *V, uint32_t num) { 419 valueNumbering.insert(std::make_pair(V, num)); 420} 421 422/// lookup_or_add - Returns the value number for the specified value, assigning 423/// it a new number if it did not have one before. 424uint32_t ValueTable::lookup_or_add(Value *V) { 425 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 426 if (VI != valueNumbering.end()) 427 return VI->second; 428 429 if (CallInst* C = dyn_cast<CallInst>(V)) { 430 if (AA->doesNotAccessMemory(C)) { 431 Expression e = create_expression(C); 432 433 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 434 if (EI != expressionNumbering.end()) { 435 valueNumbering.insert(std::make_pair(V, EI->second)); 436 return EI->second; 437 } else { 438 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 439 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 440 441 return nextValueNumber++; 442 } 443 } else if (AA->onlyReadsMemory(C)) { 444 Expression e = create_expression(C); 445 446 if (expressionNumbering.find(e) == expressionNumbering.end()) { 447 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 448 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 449 return nextValueNumber++; 450 } 451 452 MemDepResult local_dep = MD->getDependency(C); 453 454 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 455 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 456 return nextValueNumber++; 457 } 458 459 if (local_dep.isDef()) { 460 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 461 462 if (local_cdep->getNumOperands() != C->getNumOperands()) { 463 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 464 return nextValueNumber++; 465 } 466 467 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 468 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 469 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i)); 470 if (c_vn != cd_vn) { 471 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 472 return nextValueNumber++; 473 } 474 } 475 476 uint32_t v = lookup_or_add(local_cdep); 477 valueNumbering.insert(std::make_pair(V, v)); 478 return v; 479 } 480 481 // Non-local case. 482 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 483 MD->getNonLocalCallDependency(CallSite(C)); 484 // FIXME: call/call dependencies for readonly calls should return def, not 485 // clobber! Move the checking logic to MemDep! 486 CallInst* cdep = 0; 487 488 // Check to see if we have a single dominating call instruction that is 489 // identical to C. 490 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 491 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i]; 492 // Ignore non-local dependencies. 493 if (I->second.isNonLocal()) 494 continue; 495 496 // We don't handle non-depedencies. If we already have a call, reject 497 // instruction dependencies. 498 if (I->second.isClobber() || cdep != 0) { 499 cdep = 0; 500 break; 501 } 502 503 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst()); 504 // FIXME: All duplicated with non-local case. 505 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){ 506 cdep = NonLocalDepCall; 507 continue; 508 } 509 510 cdep = 0; 511 break; 512 } 513 514 if (!cdep) { 515 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 516 return nextValueNumber++; 517 } 518 519 if (cdep->getNumOperands() != C->getNumOperands()) { 520 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 521 return nextValueNumber++; 522 } 523 for (unsigned i = 1; i < C->getNumOperands(); ++i) { 524 uint32_t c_vn = lookup_or_add(C->getOperand(i)); 525 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i)); 526 if (c_vn != cd_vn) { 527 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 528 return nextValueNumber++; 529 } 530 } 531 532 uint32_t v = lookup_or_add(cdep); 533 valueNumbering.insert(std::make_pair(V, v)); 534 return v; 535 536 } else { 537 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 538 return nextValueNumber++; 539 } 540 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) { 541 Expression e = create_expression(BO); 542 543 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 544 if (EI != expressionNumbering.end()) { 545 valueNumbering.insert(std::make_pair(V, EI->second)); 546 return EI->second; 547 } else { 548 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 549 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 550 551 return nextValueNumber++; 552 } 553 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) { 554 Expression e = create_expression(C); 555 556 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 557 if (EI != expressionNumbering.end()) { 558 valueNumbering.insert(std::make_pair(V, EI->second)); 559 return EI->second; 560 } else { 561 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 562 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 563 564 return nextValueNumber++; 565 } 566 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) { 567 Expression e = create_expression(U); 568 569 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 570 if (EI != expressionNumbering.end()) { 571 valueNumbering.insert(std::make_pair(V, EI->second)); 572 return EI->second; 573 } else { 574 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 575 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 576 577 return nextValueNumber++; 578 } 579 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) { 580 Expression e = create_expression(U); 581 582 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 583 if (EI != expressionNumbering.end()) { 584 valueNumbering.insert(std::make_pair(V, EI->second)); 585 return EI->second; 586 } else { 587 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 588 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 589 590 return nextValueNumber++; 591 } 592 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) { 593 Expression e = create_expression(U); 594 595 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 596 if (EI != expressionNumbering.end()) { 597 valueNumbering.insert(std::make_pair(V, EI->second)); 598 return EI->second; 599 } else { 600 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 601 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 602 603 return nextValueNumber++; 604 } 605 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) { 606 Expression e = create_expression(U); 607 608 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 609 if (EI != expressionNumbering.end()) { 610 valueNumbering.insert(std::make_pair(V, EI->second)); 611 return EI->second; 612 } else { 613 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 614 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 615 616 return nextValueNumber++; 617 } 618 } else if (CastInst* U = dyn_cast<CastInst>(V)) { 619 Expression e = create_expression(U); 620 621 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 622 if (EI != expressionNumbering.end()) { 623 valueNumbering.insert(std::make_pair(V, EI->second)); 624 return EI->second; 625 } else { 626 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 627 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 628 629 return nextValueNumber++; 630 } 631 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) { 632 Expression e = create_expression(U); 633 634 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e); 635 if (EI != expressionNumbering.end()) { 636 valueNumbering.insert(std::make_pair(V, EI->second)); 637 return EI->second; 638 } else { 639 expressionNumbering.insert(std::make_pair(e, nextValueNumber)); 640 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 641 642 return nextValueNumber++; 643 } 644 } else { 645 valueNumbering.insert(std::make_pair(V, nextValueNumber)); 646 return nextValueNumber++; 647 } 648} 649 650/// lookup - Returns the value number of the specified value. Fails if 651/// the value has not yet been numbered. 652uint32_t ValueTable::lookup(Value *V) const { 653 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 654 assert(VI != valueNumbering.end() && "Value not numbered?"); 655 return VI->second; 656} 657 658/// clear - Remove all entries from the ValueTable 659void ValueTable::clear() { 660 valueNumbering.clear(); 661 expressionNumbering.clear(); 662 nextValueNumber = 1; 663} 664 665/// erase - Remove a value from the value numbering 666void ValueTable::erase(Value *V) { 667 valueNumbering.erase(V); 668} 669 670/// verifyRemoved - Verify that the value is removed from all internal data 671/// structures. 672void ValueTable::verifyRemoved(const Value *V) const { 673 for (DenseMap<Value*, uint32_t>::iterator 674 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 675 assert(I->first != V && "Inst still occurs in value numbering map!"); 676 } 677} 678 679//===----------------------------------------------------------------------===// 680// GVN Pass 681//===----------------------------------------------------------------------===// 682 683namespace { 684 struct ValueNumberScope { 685 ValueNumberScope* parent; 686 DenseMap<uint32_t, Value*> table; 687 688 ValueNumberScope(ValueNumberScope* p) : parent(p) { } 689 }; 690} 691 692namespace { 693 694 class GVN : public FunctionPass { 695 bool runOnFunction(Function &F); 696 public: 697 static char ID; // Pass identification, replacement for typeid 698 GVN() : FunctionPass(&ID) { } 699 700 private: 701 MemoryDependenceAnalysis *MD; 702 DominatorTree *DT; 703 704 ValueTable VN; 705 DenseMap<BasicBlock*, ValueNumberScope*> localAvail; 706 707 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType; 708 PhiMapType phiMap; 709 710 711 // This transformation requires dominator postdominator info 712 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 713 AU.addRequired<DominatorTree>(); 714 AU.addRequired<MemoryDependenceAnalysis>(); 715 AU.addRequired<AliasAnalysis>(); 716 717 AU.addPreserved<DominatorTree>(); 718 AU.addPreserved<AliasAnalysis>(); 719 } 720 721 // Helper fuctions 722 // FIXME: eliminate or document these better 723 bool processLoad(LoadInst* L, 724 SmallVectorImpl<Instruction*> &toErase); 725 bool processInstruction(Instruction *I, 726 SmallVectorImpl<Instruction*> &toErase); 727 bool processNonLocalLoad(LoadInst* L, 728 SmallVectorImpl<Instruction*> &toErase); 729 bool processBlock(BasicBlock *BB); 730 Value *GetValueForBlock(BasicBlock *BB, Instruction *orig, 731 DenseMap<BasicBlock*, Value*> &Phis, 732 bool top_level = false); 733 void dump(DenseMap<uint32_t, Value*>& d); 734 bool iterateOnFunction(Function &F); 735 Value *CollapsePhi(PHINode* p); 736 bool performPRE(Function& F); 737 Value *lookupNumber(BasicBlock *BB, uint32_t num); 738 Value *AttemptRedundancyElimination(Instruction *orig, unsigned valno); 739 void cleanupGlobalSets(); 740 void verifyRemoved(const Instruction *I) const; 741 }; 742 743 char GVN::ID = 0; 744} 745 746// createGVNPass - The public interface to this file... 747FunctionPass *llvm::createGVNPass() { return new GVN(); } 748 749static RegisterPass<GVN> X("gvn", 750 "Global Value Numbering"); 751 752void GVN::dump(DenseMap<uint32_t, Value*>& d) { 753 printf("{\n"); 754 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 755 E = d.end(); I != E; ++I) { 756 printf("%d\n", I->first); 757 I->second->dump(); 758 } 759 printf("}\n"); 760} 761 762static bool isSafeReplacement(PHINode* p, Instruction *inst) { 763 if (!isa<PHINode>(inst)) 764 return true; 765 766 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end(); 767 UI != E; ++UI) 768 if (PHINode* use_phi = dyn_cast<PHINode>(UI)) 769 if (use_phi->getParent() == inst->getParent()) 770 return false; 771 772 return true; 773} 774 775Value *GVN::CollapsePhi(PHINode *PN) { 776 Value *ConstVal = PN->hasConstantValue(DT); 777 if (!ConstVal) return 0; 778 779 Instruction *Inst = dyn_cast<Instruction>(ConstVal); 780 if (!Inst) 781 return ConstVal; 782 783 if (DT->dominates(Inst, PN)) 784 if (isSafeReplacement(PN, Inst)) 785 return Inst; 786 return 0; 787} 788 789/// GetValueForBlock - Get the value to use within the specified basic block. 790/// available values are in Phis. 791Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction *Orig, 792 DenseMap<BasicBlock*, Value*> &Phis, 793 bool TopLevel) { 794 795 // If we have already computed this value, return the previously computed val. 796 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB); 797 if (V != Phis.end() && !TopLevel) return V->second; 798 799 // If the block is unreachable, just return undef, since this path 800 // can't actually occur at runtime. 801 if (!DT->isReachableFromEntry(BB)) 802 return Phis[BB] = UndefValue::get(Orig->getType()); 803 804 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 805 Value *ret = GetValueForBlock(Pred, Orig, Phis); 806 Phis[BB] = ret; 807 return ret; 808 } 809 810 // Get the number of predecessors of this block so we can reserve space later. 811 // If there is already a PHI in it, use the #preds from it, otherwise count. 812 // Getting it from the PHI is constant time. 813 unsigned NumPreds; 814 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin())) 815 NumPreds = ExistingPN->getNumIncomingValues(); 816 else 817 NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 818 819 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so 820 // now, then get values to fill in the incoming values for the PHI. 821 PHINode *PN = PHINode::Create(Orig->getType(), Orig->getName()+".rle", 822 BB->begin()); 823 PN->reserveOperandSpace(NumPreds); 824 825 Phis.insert(std::make_pair(BB, PN)); 826 827 // Fill in the incoming values for the block. 828 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 829 Value *val = GetValueForBlock(*PI, Orig, Phis); 830 PN->addIncoming(val, *PI); 831 } 832 833 VN.getAliasAnalysis()->copyValue(Orig, PN); 834 835 // Attempt to collapse PHI nodes that are trivially redundant 836 Value *v = CollapsePhi(PN); 837 if (!v) { 838 // Cache our phi construction results 839 if (LoadInst* L = dyn_cast<LoadInst>(Orig)) 840 phiMap[L->getPointerOperand()].insert(PN); 841 else 842 phiMap[Orig].insert(PN); 843 844 return PN; 845 } 846 847 PN->replaceAllUsesWith(v); 848 if (isa<PointerType>(v->getType())) 849 MD->invalidateCachedPointerInfo(v); 850 851 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(), 852 E = Phis.end(); I != E; ++I) 853 if (I->second == PN) 854 I->second = v; 855 856 DEBUG(errs() << "GVN removed: " << *PN << '\n'); 857 MD->removeInstruction(PN); 858 PN->eraseFromParent(); 859 DEBUG(verifyRemoved(PN)); 860 861 Phis[BB] = v; 862 return v; 863} 864 865/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 866/// we're analyzing is fully available in the specified block. As we go, keep 867/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 868/// map is actually a tri-state map with the following values: 869/// 0) we know the block *is not* fully available. 870/// 1) we know the block *is* fully available. 871/// 2) we do not know whether the block is fully available or not, but we are 872/// currently speculating that it will be. 873/// 3) we are speculating for this block and have used that to speculate for 874/// other blocks. 875static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 876 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 877 // Optimistically assume that the block is fully available and check to see 878 // if we already know about this block in one lookup. 879 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 880 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 881 882 // If the entry already existed for this block, return the precomputed value. 883 if (!IV.second) { 884 // If this is a speculative "available" value, mark it as being used for 885 // speculation of other blocks. 886 if (IV.first->second == 2) 887 IV.first->second = 3; 888 return IV.first->second != 0; 889 } 890 891 // Otherwise, see if it is fully available in all predecessors. 892 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 893 894 // If this block has no predecessors, it isn't live-in here. 895 if (PI == PE) 896 goto SpeculationFailure; 897 898 for (; PI != PE; ++PI) 899 // If the value isn't fully available in one of our predecessors, then it 900 // isn't fully available in this block either. Undo our previous 901 // optimistic assumption and bail out. 902 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 903 goto SpeculationFailure; 904 905 return true; 906 907// SpeculationFailure - If we get here, we found out that this is not, after 908// all, a fully-available block. We have a problem if we speculated on this and 909// used the speculation to mark other blocks as available. 910SpeculationFailure: 911 char &BBVal = FullyAvailableBlocks[BB]; 912 913 // If we didn't speculate on this, just return with it set to false. 914 if (BBVal == 2) { 915 BBVal = 0; 916 return false; 917 } 918 919 // If we did speculate on this value, we could have blocks set to 1 that are 920 // incorrect. Walk the (transitive) successors of this block and mark them as 921 // 0 if set to one. 922 SmallVector<BasicBlock*, 32> BBWorklist; 923 BBWorklist.push_back(BB); 924 925 while (!BBWorklist.empty()) { 926 BasicBlock *Entry = BBWorklist.pop_back_val(); 927 // Note that this sets blocks to 0 (unavailable) if they happen to not 928 // already be in FullyAvailableBlocks. This is safe. 929 char &EntryVal = FullyAvailableBlocks[Entry]; 930 if (EntryVal == 0) continue; // Already unavailable. 931 932 // Mark as unavailable. 933 EntryVal = 0; 934 935 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 936 BBWorklist.push_back(*I); 937 } 938 939 return false; 940} 941 942 943/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 944/// then a load from a must-aliased pointer of a different type, try to coerce 945/// the stored value. LoadedTy is the type of the load we want to replace and 946/// InsertPt is the place to insert new instructions. 947/// 948/// If we can't do it, return null. 949static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 950 const Type *LoadedTy, 951 Instruction *InsertPt, 952 const TargetData &TD) { 953 const Type *StoredValTy = StoredVal->getType(); 954 955 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy); 956 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy); 957 958 // If the store and reload are the same size, we can always reuse it. 959 if (StoreSize == LoadSize) { 960 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) { 961 // Pointer to Pointer -> use bitcast. 962 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 963 } 964 965 // Convert source pointers to integers, which can be bitcast. 966 if (isa<PointerType>(StoredValTy)) { 967 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 968 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 969 } 970 971 const Type *TypeToCastTo = LoadedTy; 972 if (isa<PointerType>(TypeToCastTo)) 973 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 974 975 if (StoredValTy != TypeToCastTo) 976 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 977 978 // Cast to pointer if the load needs a pointer type. 979 if (isa<PointerType>(LoadedTy)) 980 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 981 982 return StoredVal; 983 } 984 985 // If the loaded value is smaller than the available value, then we can 986 // extract out a piece from it. If the available value is too small, then we 987 // can't do anything. 988 if (StoreSize < LoadSize) 989 return 0; 990 991 // Convert source pointers to integers, which can be manipulated. 992 if (isa<PointerType>(StoredValTy)) { 993 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 994 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 995 } 996 997 // Convert vectors and fp to integer, which can be manipulated. 998 if (!isa<IntegerType>(StoredValTy)) { 999 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 1000 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 1001 } 1002 1003 // If this is a big-endian system, we need to shift the value down to the low 1004 // bits so that a truncate will work. 1005 if (TD.isBigEndian()) { 1006 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 1007 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 1008 } 1009 1010 // Truncate the integer to the right size now. 1011 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 1012 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 1013 1014 if (LoadedTy == NewIntTy) 1015 return StoredVal; 1016 1017 // If the result is a pointer, inttoptr. 1018 if (isa<PointerType>(LoadedTy)) 1019 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 1020 1021 // Otherwise, bitcast. 1022 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 1023} 1024 1025static void 1026GetAvailableBlockValues(DenseMap<BasicBlock*, Value*> &BlockReplValues, 1027 SmallVector<std::pair<BasicBlock*, 1028 Value*>, 16> &ValuesPerBlock, 1029 const Type *LoadTy, 1030 const TargetData *TD) { 1031 1032 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1033 BasicBlock *BB = ValuesPerBlock[i].first; 1034 Value *AvailableVal = ValuesPerBlock[i].second; 1035 1036 Value *&BlockEntry = BlockReplValues[BB]; 1037 if (BlockEntry) continue; 1038 1039 if (AvailableVal->getType() != LoadTy) { 1040 assert(TD && "Need target data to handle type mismatch case"); 1041 AvailableVal = CoerceAvailableValueToLoadType(AvailableVal, LoadTy, 1042 BB->getTerminator(), *TD); 1043 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n" 1044 << *ValuesPerBlock[i].second << '\n' 1045 << *AvailableVal << '\n' << "\n\n\n"); 1046 } 1047 BlockEntry = AvailableVal; 1048 } 1049} 1050 1051/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1052/// non-local by performing PHI construction. 1053bool GVN::processNonLocalLoad(LoadInst *LI, 1054 SmallVectorImpl<Instruction*> &toErase) { 1055 // Find the non-local dependencies of the load. 1056 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps; 1057 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(), 1058 Deps); 1059 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: " 1060 // << Deps.size() << *LI << '\n'); 1061 1062 // If we had to process more than one hundred blocks to find the 1063 // dependencies, this load isn't worth worrying about. Optimizing 1064 // it will be too expensive. 1065 if (Deps.size() > 100) 1066 return false; 1067 1068 // If we had a phi translation failure, we'll have a single entry which is a 1069 // clobber in the current block. Reject this early. 1070 if (Deps.size() == 1 && Deps[0].second.isClobber()) { 1071 DEBUG( 1072 errs() << "GVN: non-local load "; 1073 WriteAsOperand(errs(), LI); 1074 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n'; 1075 ); 1076 return false; 1077 } 1078 1079 // Filter out useless results (non-locals, etc). Keep track of the blocks 1080 // where we have a value available in repl, also keep track of whether we see 1081 // dependencies that produce an unknown value for the load (such as a call 1082 // that could potentially clobber the load). 1083 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock; 1084 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1085 1086 const TargetData *TD = 0; 1087 1088 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1089 BasicBlock *DepBB = Deps[i].first; 1090 MemDepResult DepInfo = Deps[i].second; 1091 1092 if (DepInfo.isClobber()) { 1093 UnavailableBlocks.push_back(DepBB); 1094 continue; 1095 } 1096 1097 Instruction *DepInst = DepInfo.getInst(); 1098 1099 // Loading the allocation -> undef. 1100 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1101 ValuesPerBlock.push_back(std::make_pair(DepBB, 1102 UndefValue::get(LI->getType()))); 1103 continue; 1104 } 1105 1106 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) { 1107 // Reject loads and stores that are to the same address but are of 1108 // different types if we have to. 1109 if (S->getOperand(0)->getType() != LI->getType()) { 1110 if (TD == 0) 1111 TD = getAnalysisIfAvailable<TargetData>(); 1112 1113 // If the stored value is larger or equal to the loaded value, we can 1114 // reuse it. 1115 if (TD == 0 || 1116 TD->getTypeSizeInBits(S->getOperand(0)->getType()) < 1117 TD->getTypeSizeInBits(LI->getType())) { 1118 UnavailableBlocks.push_back(DepBB); 1119 continue; 1120 } 1121 } 1122 1123 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0))); 1124 1125 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) { 1126 // If the types mismatch and we can't handle it, reject reuse of the load. 1127 if (LD->getType() != LI->getType()) { 1128 if (TD == 0) 1129 TD = getAnalysisIfAvailable<TargetData>(); 1130 1131 // If the stored value is larger or equal to the loaded value, we can 1132 // reuse it. 1133 if (TD == 0 || 1134 TD->getTypeSizeInBits(LD->getType()) < 1135 TD->getTypeSizeInBits(LI->getType())) { 1136 UnavailableBlocks.push_back(DepBB); 1137 continue; 1138 } 1139 } 1140 ValuesPerBlock.push_back(std::make_pair(DepBB, LD)); 1141 } else { 1142 // FIXME: Handle memset/memcpy. 1143 UnavailableBlocks.push_back(DepBB); 1144 continue; 1145 } 1146 } 1147 1148 // If we have no predecessors that produce a known value for this load, exit 1149 // early. 1150 if (ValuesPerBlock.empty()) return false; 1151 1152 // If all of the instructions we depend on produce a known value for this 1153 // load, then it is fully redundant and we can use PHI insertion to compute 1154 // its value. Insert PHIs and remove the fully redundant value now. 1155 if (UnavailableBlocks.empty()) { 1156 // Use cached PHI construction information from previous runs 1157 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1158 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated? 1159 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1160 I != E; ++I) { 1161 if ((*I)->getParent() == LI->getParent()) { 1162 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD #1: " << *LI << '\n'); 1163 LI->replaceAllUsesWith(*I); 1164 if (isa<PointerType>((*I)->getType())) 1165 MD->invalidateCachedPointerInfo(*I); 1166 toErase.push_back(LI); 1167 NumGVNLoad++; 1168 return true; 1169 } 1170 1171 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1172 } 1173 1174 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1175 1176 // Convert the block information to a map, and insert coersions as needed. 1177 DenseMap<BasicBlock*, Value*> BlockReplValues; 1178 GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD); 1179 1180 // Perform PHI construction. 1181 Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1182 LI->replaceAllUsesWith(V); 1183 1184 if (isa<PHINode>(V)) 1185 V->takeName(LI); 1186 if (isa<PointerType>(V->getType())) 1187 MD->invalidateCachedPointerInfo(V); 1188 toErase.push_back(LI); 1189 NumGVNLoad++; 1190 return true; 1191 } 1192 1193 if (!EnablePRE || !EnableLoadPRE) 1194 return false; 1195 1196 // Okay, we have *some* definitions of the value. This means that the value 1197 // is available in some of our (transitive) predecessors. Lets think about 1198 // doing PRE of this load. This will involve inserting a new load into the 1199 // predecessor when it's not available. We could do this in general, but 1200 // prefer to not increase code size. As such, we only do this when we know 1201 // that we only have to insert *one* load (which means we're basically moving 1202 // the load, not inserting a new one). 1203 1204 SmallPtrSet<BasicBlock *, 4> Blockers; 1205 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1206 Blockers.insert(UnavailableBlocks[i]); 1207 1208 // Lets find first basic block with more than one predecessor. Walk backwards 1209 // through predecessors if needed. 1210 BasicBlock *LoadBB = LI->getParent(); 1211 BasicBlock *TmpBB = LoadBB; 1212 1213 bool isSinglePred = false; 1214 bool allSingleSucc = true; 1215 while (TmpBB->getSinglePredecessor()) { 1216 isSinglePred = true; 1217 TmpBB = TmpBB->getSinglePredecessor(); 1218 if (!TmpBB) // If haven't found any, bail now. 1219 return false; 1220 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1221 return false; 1222 if (Blockers.count(TmpBB)) 1223 return false; 1224 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1225 allSingleSucc = false; 1226 } 1227 1228 assert(TmpBB); 1229 LoadBB = TmpBB; 1230 1231 // If we have a repl set with LI itself in it, this means we have a loop where 1232 // at least one of the values is LI. Since this means that we won't be able 1233 // to eliminate LI even if we insert uses in the other predecessors, we will 1234 // end up increasing code size. Reject this by scanning for LI. 1235 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1236 if (ValuesPerBlock[i].second == LI) 1237 return false; 1238 1239 if (isSinglePred) { 1240 bool isHot = false; 1241 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1242 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second)) 1243 // "Hot" Instruction is in some loop (because it dominates its dep. 1244 // instruction). 1245 if (DT->dominates(LI, I)) { 1246 isHot = true; 1247 break; 1248 } 1249 1250 // We are interested only in "hot" instructions. We don't want to do any 1251 // mis-optimizations here. 1252 if (!isHot) 1253 return false; 1254 } 1255 1256 // Okay, we have some hope :). Check to see if the loaded value is fully 1257 // available in all but one predecessor. 1258 // FIXME: If we could restructure the CFG, we could make a common pred with 1259 // all the preds that don't have an available LI and insert a new load into 1260 // that one block. 1261 BasicBlock *UnavailablePred = 0; 1262 1263 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1264 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1265 FullyAvailableBlocks[ValuesPerBlock[i].first] = true; 1266 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1267 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1268 1269 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1270 PI != E; ++PI) { 1271 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 1272 continue; 1273 1274 // If this load is not available in multiple predecessors, reject it. 1275 if (UnavailablePred && UnavailablePred != *PI) 1276 return false; 1277 UnavailablePred = *PI; 1278 } 1279 1280 assert(UnavailablePred != 0 && 1281 "Fully available value should be eliminated above!"); 1282 1283 // If the loaded pointer is PHI node defined in this block, do PHI translation 1284 // to get its value in the predecessor. 1285 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred); 1286 1287 // Make sure the value is live in the predecessor. If it was defined by a 1288 // non-PHI instruction in this block, we don't know how to recompute it above. 1289 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr)) 1290 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) { 1291 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: " 1292 << *LPInst << '\n' << *LI << "\n"); 1293 return false; 1294 } 1295 1296 // We don't currently handle critical edges :( 1297 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { 1298 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" 1299 << UnavailablePred->getName() << "': " << *LI << '\n'); 1300 return false; 1301 } 1302 1303 // Make sure it is valid to move this load here. We have to watch out for: 1304 // @1 = getelementptr (i8* p, ... 1305 // test p and branch if == 0 1306 // load @1 1307 // It is valid to have the getelementptr before the test, even if p can be 0, 1308 // as getelementptr only does address arithmetic. 1309 // If we are not pushing the value through any multiple-successor blocks 1310 // we do not have this case. Otherwise, check that the load is safe to 1311 // put anywhere; this can be improved, but should be conservatively safe. 1312 if (!allSingleSucc && 1313 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) 1314 return false; 1315 1316 // Okay, we can eliminate this load by inserting a reload in the predecessor 1317 // and using PHI construction to get the value in the other predecessors, do 1318 // it. 1319 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1320 1321 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1322 LI->getAlignment(), 1323 UnavailablePred->getTerminator()); 1324 1325 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()]; 1326 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end(); 1327 I != E; ++I) 1328 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I)); 1329 1330 DenseMap<BasicBlock*, Value*> BlockReplValues; 1331 GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD); 1332 BlockReplValues[UnavailablePred] = NewLoad; 1333 1334 // Perform PHI construction. 1335 Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true); 1336 LI->replaceAllUsesWith(V); 1337 if (isa<PHINode>(V)) 1338 V->takeName(LI); 1339 if (isa<PointerType>(V->getType())) 1340 MD->invalidateCachedPointerInfo(V); 1341 toErase.push_back(LI); 1342 NumPRELoad++; 1343 return true; 1344} 1345 1346/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can 1347/// be expressed as a base pointer plus a constant offset. Return the base and 1348/// offset to the caller. 1349static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 1350 const TargetData *TD) { 1351 Operator *PtrOp = dyn_cast<Operator>(Ptr); 1352 if (PtrOp == 0) return Ptr; 1353 1354 // Just look through bitcasts. 1355 if (PtrOp->getOpcode() == Instruction::BitCast) 1356 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD); 1357 1358 // If this is a GEP with constant indices, we can look through it. 1359 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp); 1360 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr; 1361 1362 gep_type_iterator GTI = gep_type_begin(GEP); 1363 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E; 1364 ++I, ++GTI) { 1365 ConstantInt *OpC = cast<ConstantInt>(*I); 1366 if (OpC->isZero()) continue; 1367 1368 // Handle a struct and array indices which add their offset to the pointer. 1369 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 1370 Offset += TD->getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 1371 } else { 1372 uint64_t Size = TD->getTypeAllocSize(GTI.getIndexedType()); 1373 Offset += OpC->getSExtValue()*Size; 1374 } 1375 } 1376 1377 // Re-sign extend from the pointer size if needed to get overflow edge cases 1378 // right. 1379 unsigned PtrSize = TD->getPointerSizeInBits(); 1380 if (PtrSize < 64) 1381 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize); 1382 1383 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD); 1384} 1385 1386 1387/// AnalyzeLoadFromClobberingStore - This function is called when we have a 1388/// memdep query of a load that ends up being a clobbering store. This means 1389/// that the store *may* provide bits used by the load but we can't be sure 1390/// because the pointers don't mustalias. Check this case to see if there is 1391/// anything more we can do before we give up. This returns -1 if we have to 1392/// give up, or a byte number in the stored value of the piece that feeds the 1393/// load. 1394static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI, 1395 const TargetData *TD) { 1396 int64_t StoreOffset = 0, LoadOffset = 0; 1397 Value *StoreBase = 1398 GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD); 1399 Value *LoadBase = 1400 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD); 1401 if (StoreBase != LoadBase) 1402 return -1; 1403 1404 // If the load and store are to the exact same address, they should have been 1405 // a must alias. AA must have gotten confused. 1406 // FIXME: Study to see if/when this happens. 1407 if (LoadOffset == StoreOffset) { 1408#if 0 1409 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 1410 << "Base = " << *StoreBase << "\n" 1411 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 1412 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 1413 << "Load Ptr = " << *L->getPointerOperand() << "\n" 1414 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 1415 errs() << "'" << L->getParent()->getParent()->getName() << "'" 1416 << *L->getParent(); 1417#endif 1418 return -1; 1419 } 1420 1421 // If the load and store don't overlap at all, the store doesn't provide 1422 // anything to the load. In this case, they really don't alias at all, AA 1423 // must have gotten confused. 1424 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then 1425 // remove this check, as it is duplicated with what we have below. 1426 uint64_t StoreSize = TD->getTypeSizeInBits(DepSI->getOperand(0)->getType()); 1427 uint64_t LoadSize = TD->getTypeSizeInBits(L->getType()); 1428 1429 if ((StoreSize & 7) | (LoadSize & 7)) 1430 return -1; 1431 StoreSize >>= 3; // Convert to bytes. 1432 LoadSize >>= 3; 1433 1434 1435 bool isAAFailure = false; 1436 if (StoreOffset < LoadOffset) { 1437 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 1438 } else { 1439 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 1440 } 1441 if (isAAFailure) { 1442#if 0 1443 errs() << "STORE LOAD DEP WITH COMMON BASE:\n" 1444 << "Base = " << *StoreBase << "\n" 1445 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n" 1446 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n" 1447 << "Load Ptr = " << *L->getPointerOperand() << "\n" 1448 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n"; 1449 errs() << "'" << L->getParent()->getParent()->getName() << "'" 1450 << *L->getParent(); 1451#endif 1452 return -1; 1453 } 1454 1455 // If the Load isn't completely contained within the stored bits, we don't 1456 // have all the bits to feed it. We could do something crazy in the future 1457 // (issue a smaller load then merge the bits in) but this seems unlikely to be 1458 // valuable. 1459 if (StoreOffset > LoadOffset || 1460 StoreOffset+StoreSize < LoadOffset+LoadSize) 1461 return -1; 1462 1463 // Okay, we can do this transformation. Return the number of bytes into the 1464 // store that the load is. 1465 return LoadOffset-StoreOffset; 1466} 1467 1468 1469/// GetStoreValueForLoad - This function is called when we have a 1470/// memdep query of a load that ends up being a clobbering store. This means 1471/// that the store *may* provide bits used by the load but we can't be sure 1472/// because the pointers don't mustalias. Check this case to see if there is 1473/// anything more we can do before we give up. 1474static Value *GetStoreValueForLoad(Value *SrcVal, int Offset,const Type *LoadTy, 1475 Instruction *InsertPt, const TargetData *TD){ 1476 LLVMContext &Ctx = SrcVal->getType()->getContext(); 1477 1478 uint64_t StoreSize = TD->getTypeSizeInBits(SrcVal->getType())/8; 1479 uint64_t LoadSize = TD->getTypeSizeInBits(LoadTy)/8; 1480 1481 1482 // Compute which bits of the stored value are being used by the load. Convert 1483 // to an integer type to start with. 1484 if (isa<PointerType>(SrcVal->getType())) 1485 SrcVal = new PtrToIntInst(SrcVal, TD->getIntPtrType(Ctx), "tmp", InsertPt); 1486 if (!isa<IntegerType>(SrcVal->getType())) 1487 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8), 1488 "tmp", InsertPt); 1489 1490 // Shift the bits to the least significant depending on endianness. 1491 unsigned ShiftAmt; 1492 if (TD->isLittleEndian()) { 1493 ShiftAmt = Offset*8; 1494 } else { 1495 ShiftAmt = StoreSize-LoadSize-Offset; 1496 } 1497 1498 SrcVal = BinaryOperator::CreateLShr(SrcVal, 1499 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt); 1500 1501 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8), 1502 "tmp", InsertPt); 1503 1504 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, *TD); 1505} 1506 1507 1508 1509/// processLoad - Attempt to eliminate a load, first by eliminating it 1510/// locally, and then attempting non-local elimination if that fails. 1511bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) { 1512 if (L->isVolatile()) 1513 return false; 1514 1515 // ... to a pointer that has been loaded from before... 1516 MemDepResult Dep = MD->getDependency(L); 1517 1518 // If the value isn't available, don't do anything! 1519 if (Dep.isClobber()) { 1520 // FIXME: We should handle memset/memcpy/memmove as dependent instructions 1521 // to forward the value if available. 1522 //if (isa<MemIntrinsic>(Dep.getInst())) 1523 //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n"; 1524 1525 // Check to see if we have something like this: 1526 // store i32 123, i32* %P 1527 // %A = bitcast i32* %P to i8* 1528 // %B = gep i8* %A, i32 1 1529 // %C = load i8* %B 1530 // 1531 // We could do that by recognizing if the clobber instructions are obviously 1532 // a common base + constant offset, and if the previous store (or memset) 1533 // completely covers this load. This sort of thing can happen in bitfield 1534 // access code. 1535 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) 1536 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) { 1537 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, TD); 1538 if (Offset != -1) { 1539 Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset, 1540 L->getType(), L, TD); 1541 DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n' 1542 << *AvailVal << '\n' << *L << "\n\n\n"); 1543 1544 // Replace the load! 1545 L->replaceAllUsesWith(AvailVal); 1546 if (isa<PointerType>(AvailVal->getType())) 1547 MD->invalidateCachedPointerInfo(AvailVal); 1548 toErase.push_back(L); 1549 NumGVNLoad++; 1550 return true; 1551 } 1552 } 1553 1554 DEBUG( 1555 // fast print dep, using operator<< on instruction would be too slow 1556 errs() << "GVN: load "; 1557 WriteAsOperand(errs(), L); 1558 Instruction *I = Dep.getInst(); 1559 errs() << " is clobbered by " << *I << '\n'; 1560 ); 1561 return false; 1562 } 1563 1564 // If it is defined in another block, try harder. 1565 if (Dep.isNonLocal()) 1566 return processNonLocalLoad(L, toErase); 1567 1568 Instruction *DepInst = Dep.getInst(); 1569 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1570 Value *StoredVal = DepSI->getOperand(0); 1571 1572 // The store and load are to a must-aliased pointer, but they may not 1573 // actually have the same type. See if we know how to reuse the stored 1574 // value (depending on its type). 1575 const TargetData *TD = 0; 1576 if (StoredVal->getType() != L->getType() && 1577 (TD = getAnalysisIfAvailable<TargetData>())) { 1578 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), L, *TD); 1579 if (StoredVal == 0) 1580 return false; 1581 1582 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1583 << '\n' << *L << "\n\n\n"); 1584 } 1585 1586 // Remove it! 1587 L->replaceAllUsesWith(StoredVal); 1588 if (isa<PointerType>(StoredVal->getType())) 1589 MD->invalidateCachedPointerInfo(StoredVal); 1590 toErase.push_back(L); 1591 NumGVNLoad++; 1592 return true; 1593 } 1594 1595 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1596 Value *AvailableVal = DepLI; 1597 1598 // The loads are of a must-aliased pointer, but they may not actually have 1599 // the same type. See if we know how to reuse the previously loaded value 1600 // (depending on its type). 1601 const TargetData *TD = 0; 1602 if (DepLI->getType() != L->getType() && 1603 (TD = getAnalysisIfAvailable<TargetData>())) { 1604 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L, *TD); 1605 if (AvailableVal == 0) 1606 return false; 1607 1608 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1609 << "\n" << *L << "\n\n\n"); 1610 } 1611 1612 // Remove it! 1613 L->replaceAllUsesWith(AvailableVal); 1614 if (isa<PointerType>(DepLI->getType())) 1615 MD->invalidateCachedPointerInfo(DepLI); 1616 toErase.push_back(L); 1617 NumGVNLoad++; 1618 return true; 1619 } 1620 1621 // If this load really doesn't depend on anything, then we must be loading an 1622 // undef value. This can happen when loading for a fresh allocation with no 1623 // intervening stores, for example. 1624 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) { 1625 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1626 toErase.push_back(L); 1627 NumGVNLoad++; 1628 return true; 1629 } 1630 1631 return false; 1632} 1633 1634Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { 1635 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB); 1636 if (I == localAvail.end()) 1637 return 0; 1638 1639 ValueNumberScope *Locals = I->second; 1640 while (Locals) { 1641 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num); 1642 if (I != Locals->table.end()) 1643 return I->second; 1644 Locals = Locals->parent; 1645 } 1646 1647 return 0; 1648} 1649 1650/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination 1651/// by inheritance from the dominator fails, see if we can perform phi 1652/// construction to eliminate the redundancy. 1653Value *GVN::AttemptRedundancyElimination(Instruction *orig, unsigned valno) { 1654 BasicBlock *BaseBlock = orig->getParent(); 1655 1656 SmallPtrSet<BasicBlock*, 4> Visited; 1657 SmallVector<BasicBlock*, 8> Stack; 1658 Stack.push_back(BaseBlock); 1659 1660 DenseMap<BasicBlock*, Value*> Results; 1661 1662 // Walk backwards through our predecessors, looking for instances of the 1663 // value number we're looking for. Instances are recorded in the Results 1664 // map, which is then used to perform phi construction. 1665 while (!Stack.empty()) { 1666 BasicBlock *Current = Stack.back(); 1667 Stack.pop_back(); 1668 1669 // If we've walked all the way to a proper dominator, then give up. Cases 1670 // where the instance is in the dominator will have been caught by the fast 1671 // path, and any cases that require phi construction further than this are 1672 // probably not worth it anyways. Note that this is a SIGNIFICANT compile 1673 // time improvement. 1674 if (DT->properlyDominates(Current, orig->getParent())) return 0; 1675 1676 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA = 1677 localAvail.find(Current); 1678 if (LA == localAvail.end()) return 0; 1679 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno); 1680 1681 if (V != LA->second->table.end()) { 1682 // Found an instance, record it. 1683 Results.insert(std::make_pair(Current, V->second)); 1684 continue; 1685 } 1686 1687 // If we reach the beginning of the function, then give up. 1688 if (pred_begin(Current) == pred_end(Current)) 1689 return 0; 1690 1691 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current); 1692 PI != PE; ++PI) 1693 if (Visited.insert(*PI)) 1694 Stack.push_back(*PI); 1695 } 1696 1697 // If we didn't find instances, give up. Otherwise, perform phi construction. 1698 if (Results.size() == 0) 1699 return 0; 1700 else 1701 return GetValueForBlock(BaseBlock, orig, Results, true); 1702} 1703 1704/// processInstruction - When calculating availability, handle an instruction 1705/// by inserting it into the appropriate sets 1706bool GVN::processInstruction(Instruction *I, 1707 SmallVectorImpl<Instruction*> &toErase) { 1708 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1709 bool Changed = processLoad(LI, toErase); 1710 1711 if (!Changed) { 1712 unsigned Num = VN.lookup_or_add(LI); 1713 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI)); 1714 } 1715 1716 return Changed; 1717 } 1718 1719 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1720 unsigned Num = VN.lookup_or_add(I); 1721 1722 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1723 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1724 1725 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1726 return false; 1727 1728 Value *BranchCond = BI->getCondition(); 1729 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1730 1731 BasicBlock *TrueSucc = BI->getSuccessor(0); 1732 BasicBlock *FalseSucc = BI->getSuccessor(1); 1733 1734 if (TrueSucc->getSinglePredecessor()) 1735 localAvail[TrueSucc]->table[CondVN] = 1736 ConstantInt::getTrue(TrueSucc->getContext()); 1737 if (FalseSucc->getSinglePredecessor()) 1738 localAvail[FalseSucc]->table[CondVN] = 1739 ConstantInt::getFalse(TrueSucc->getContext()); 1740 1741 return false; 1742 1743 // Allocations are always uniquely numbered, so we can save time and memory 1744 // by fast failing them. 1745 } else if (isa<AllocationInst>(I) || isMalloc(I) || isa<TerminatorInst>(I)) { 1746 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1747 return false; 1748 } 1749 1750 // Collapse PHI nodes 1751 if (PHINode* p = dyn_cast<PHINode>(I)) { 1752 Value *constVal = CollapsePhi(p); 1753 1754 if (constVal) { 1755 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end(); 1756 PI != PE; ++PI) 1757 PI->second.erase(p); 1758 1759 p->replaceAllUsesWith(constVal); 1760 if (isa<PointerType>(constVal->getType())) 1761 MD->invalidateCachedPointerInfo(constVal); 1762 VN.erase(p); 1763 1764 toErase.push_back(p); 1765 } else { 1766 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1767 } 1768 1769 // If the number we were assigned was a brand new VN, then we don't 1770 // need to do a lookup to see if the number already exists 1771 // somewhere in the domtree: it can't! 1772 } else if (Num == NextNum) { 1773 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1774 1775 // Perform fast-path value-number based elimination of values inherited from 1776 // dominators. 1777 } else if (Value *repl = lookupNumber(I->getParent(), Num)) { 1778 // Remove it! 1779 VN.erase(I); 1780 I->replaceAllUsesWith(repl); 1781 if (isa<PointerType>(repl->getType())) 1782 MD->invalidateCachedPointerInfo(repl); 1783 toErase.push_back(I); 1784 return true; 1785 1786#if 0 1787 // Perform slow-pathvalue-number based elimination with phi construction. 1788 } else if (Value *repl = AttemptRedundancyElimination(I, Num)) { 1789 // Remove it! 1790 VN.erase(I); 1791 I->replaceAllUsesWith(repl); 1792 if (isa<PointerType>(repl->getType())) 1793 MD->invalidateCachedPointerInfo(repl); 1794 toErase.push_back(I); 1795 return true; 1796#endif 1797 } else { 1798 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I)); 1799 } 1800 1801 return false; 1802} 1803 1804/// runOnFunction - This is the main transformation entry point for a function. 1805bool GVN::runOnFunction(Function& F) { 1806 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1807 DT = &getAnalysis<DominatorTree>(); 1808 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1809 VN.setMemDep(MD); 1810 VN.setDomTree(DT); 1811 1812 bool Changed = false; 1813 bool ShouldContinue = true; 1814 1815 // Merge unconditional branches, allowing PRE to catch more 1816 // optimization opportunities. 1817 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1818 BasicBlock *BB = FI; 1819 ++FI; 1820 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1821 if (removedBlock) NumGVNBlocks++; 1822 1823 Changed |= removedBlock; 1824 } 1825 1826 unsigned Iteration = 0; 1827 1828 while (ShouldContinue) { 1829 DEBUG(errs() << "GVN iteration: " << Iteration << "\n"); 1830 ShouldContinue = iterateOnFunction(F); 1831 Changed |= ShouldContinue; 1832 ++Iteration; 1833 } 1834 1835 if (EnablePRE) { 1836 bool PREChanged = true; 1837 while (PREChanged) { 1838 PREChanged = performPRE(F); 1839 Changed |= PREChanged; 1840 } 1841 } 1842 // FIXME: Should perform GVN again after PRE does something. PRE can move 1843 // computations into blocks where they become fully redundant. Note that 1844 // we can't do this until PRE's critical edge splitting updates memdep. 1845 // Actually, when this happens, we should just fully integrate PRE into GVN. 1846 1847 cleanupGlobalSets(); 1848 1849 return Changed; 1850} 1851 1852 1853bool GVN::processBlock(BasicBlock *BB) { 1854 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and 1855 // incrementing BI before processing an instruction). 1856 SmallVector<Instruction*, 8> toErase; 1857 bool ChangedFunction = false; 1858 1859 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1860 BI != BE;) { 1861 ChangedFunction |= processInstruction(BI, toErase); 1862 if (toErase.empty()) { 1863 ++BI; 1864 continue; 1865 } 1866 1867 // If we need some instructions deleted, do it now. 1868 NumGVNInstr += toErase.size(); 1869 1870 // Avoid iterator invalidation. 1871 bool AtStart = BI == BB->begin(); 1872 if (!AtStart) 1873 --BI; 1874 1875 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(), 1876 E = toErase.end(); I != E; ++I) { 1877 DEBUG(errs() << "GVN removed: " << **I << '\n'); 1878 MD->removeInstruction(*I); 1879 (*I)->eraseFromParent(); 1880 DEBUG(verifyRemoved(*I)); 1881 } 1882 toErase.clear(); 1883 1884 if (AtStart) 1885 BI = BB->begin(); 1886 else 1887 ++BI; 1888 } 1889 1890 return ChangedFunction; 1891} 1892 1893/// performPRE - Perform a purely local form of PRE that looks for diamond 1894/// control flow patterns and attempts to perform simple PRE at the join point. 1895bool GVN::performPRE(Function& F) { 1896 bool Changed = false; 1897 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 1898 DenseMap<BasicBlock*, Value*> predMap; 1899 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1900 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1901 BasicBlock *CurrentBlock = *DI; 1902 1903 // Nothing to PRE in the entry block. 1904 if (CurrentBlock == &F.getEntryBlock()) continue; 1905 1906 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1907 BE = CurrentBlock->end(); BI != BE; ) { 1908 Instruction *CurInst = BI++; 1909 1910 if (isa<AllocationInst>(CurInst) || isMalloc(CurInst) || 1911 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1912 (CurInst->getType() == Type::getVoidTy(F.getContext())) || 1913 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1914 isa<DbgInfoIntrinsic>(CurInst)) 1915 continue; 1916 1917 uint32_t ValNo = VN.lookup(CurInst); 1918 1919 // Look for the predecessors for PRE opportunities. We're 1920 // only trying to solve the basic diamond case, where 1921 // a value is computed in the successor and one predecessor, 1922 // but not the other. We also explicitly disallow cases 1923 // where the successor is its own predecessor, because they're 1924 // more complicated to get right. 1925 unsigned NumWith = 0; 1926 unsigned NumWithout = 0; 1927 BasicBlock *PREPred = 0; 1928 predMap.clear(); 1929 1930 for (pred_iterator PI = pred_begin(CurrentBlock), 1931 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 1932 // We're not interested in PRE where the block is its 1933 // own predecessor, on in blocks with predecessors 1934 // that are not reachable. 1935 if (*PI == CurrentBlock) { 1936 NumWithout = 2; 1937 break; 1938 } else if (!localAvail.count(*PI)) { 1939 NumWithout = 2; 1940 break; 1941 } 1942 1943 DenseMap<uint32_t, Value*>::iterator predV = 1944 localAvail[*PI]->table.find(ValNo); 1945 if (predV == localAvail[*PI]->table.end()) { 1946 PREPred = *PI; 1947 NumWithout++; 1948 } else if (predV->second == CurInst) { 1949 NumWithout = 2; 1950 } else { 1951 predMap[*PI] = predV->second; 1952 NumWith++; 1953 } 1954 } 1955 1956 // Don't do PRE when it might increase code size, i.e. when 1957 // we would need to insert instructions in more than one pred. 1958 if (NumWithout != 1 || NumWith == 0) 1959 continue; 1960 1961 // We can't do PRE safely on a critical edge, so instead we schedule 1962 // the edge to be split and perform the PRE the next time we iterate 1963 // on the function. 1964 unsigned SuccNum = 0; 1965 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors(); 1966 i != e; ++i) 1967 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) { 1968 SuccNum = i; 1969 break; 1970 } 1971 1972 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 1973 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 1974 continue; 1975 } 1976 1977 // Instantiate the expression the in predecessor that lacked it. 1978 // Because we are going top-down through the block, all value numbers 1979 // will be available in the predecessor by the time we need them. Any 1980 // that weren't original present will have been instantiated earlier 1981 // in this loop. 1982 Instruction *PREInstr = CurInst->clone(CurInst->getContext()); 1983 bool success = true; 1984 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 1985 Value *Op = PREInstr->getOperand(i); 1986 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 1987 continue; 1988 1989 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) { 1990 PREInstr->setOperand(i, V); 1991 } else { 1992 success = false; 1993 break; 1994 } 1995 } 1996 1997 // Fail out if we encounter an operand that is not available in 1998 // the PRE predecessor. This is typically because of loads which 1999 // are not value numbered precisely. 2000 if (!success) { 2001 delete PREInstr; 2002 DEBUG(verifyRemoved(PREInstr)); 2003 continue; 2004 } 2005 2006 PREInstr->insertBefore(PREPred->getTerminator()); 2007 PREInstr->setName(CurInst->getName() + ".pre"); 2008 predMap[PREPred] = PREInstr; 2009 VN.add(PREInstr, ValNo); 2010 NumGVNPRE++; 2011 2012 // Update the availability map to include the new instruction. 2013 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr)); 2014 2015 // Create a PHI to make the value available in this block. 2016 PHINode* Phi = PHINode::Create(CurInst->getType(), 2017 CurInst->getName() + ".pre-phi", 2018 CurrentBlock->begin()); 2019 for (pred_iterator PI = pred_begin(CurrentBlock), 2020 PE = pred_end(CurrentBlock); PI != PE; ++PI) 2021 Phi->addIncoming(predMap[*PI], *PI); 2022 2023 VN.add(Phi, ValNo); 2024 localAvail[CurrentBlock]->table[ValNo] = Phi; 2025 2026 CurInst->replaceAllUsesWith(Phi); 2027 if (isa<PointerType>(Phi->getType())) 2028 MD->invalidateCachedPointerInfo(Phi); 2029 VN.erase(CurInst); 2030 2031 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n'); 2032 MD->removeInstruction(CurInst); 2033 CurInst->eraseFromParent(); 2034 DEBUG(verifyRemoved(CurInst)); 2035 Changed = true; 2036 } 2037 } 2038 2039 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator 2040 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) 2041 SplitCriticalEdge(I->first, I->second, this); 2042 2043 return Changed || toSplit.size(); 2044} 2045 2046/// iterateOnFunction - Executes one iteration of GVN 2047bool GVN::iterateOnFunction(Function &F) { 2048 cleanupGlobalSets(); 2049 2050 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2051 DE = df_end(DT->getRootNode()); DI != DE; ++DI) { 2052 if (DI->getIDom()) 2053 localAvail[DI->getBlock()] = 2054 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]); 2055 else 2056 localAvail[DI->getBlock()] = new ValueNumberScope(0); 2057 } 2058 2059 // Top-down walk of the dominator tree 2060 bool Changed = false; 2061#if 0 2062 // Needed for value numbering with phi construction to work. 2063 ReversePostOrderTraversal<Function*> RPOT(&F); 2064 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2065 RE = RPOT.end(); RI != RE; ++RI) 2066 Changed |= processBlock(*RI); 2067#else 2068 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2069 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2070 Changed |= processBlock(DI->getBlock()); 2071#endif 2072 2073 return Changed; 2074} 2075 2076void GVN::cleanupGlobalSets() { 2077 VN.clear(); 2078 phiMap.clear(); 2079 2080 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2081 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) 2082 delete I->second; 2083 localAvail.clear(); 2084} 2085 2086/// verifyRemoved - Verify that the specified instruction does not occur in our 2087/// internal data structures. 2088void GVN::verifyRemoved(const Instruction *Inst) const { 2089 VN.verifyRemoved(Inst); 2090 2091 // Walk through the PHI map to make sure the instruction isn't hiding in there 2092 // somewhere. 2093 for (PhiMapType::iterator 2094 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) { 2095 assert(I->first != Inst && "Inst is still a key in PHI map!"); 2096 2097 for (SmallPtrSet<Instruction*, 4>::iterator 2098 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) { 2099 assert(*II != Inst && "Inst is still a value in PHI map!"); 2100 } 2101 } 2102 2103 // Walk through the value number scope to make sure the instruction isn't 2104 // ferreted away in it. 2105 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator 2106 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) { 2107 const ValueNumberScope *VNS = I->second; 2108 2109 while (VNS) { 2110 for (DenseMap<uint32_t, Value*>::iterator 2111 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) { 2112 assert(II->second != Inst && "Inst still in value numbering scope!"); 2113 } 2114 2115 VNS = VNS->parent; 2116 } 2117 } 2118} 2119