1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 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 file implements sparse conditional constant propagation and merging: 11 // 12 // Specifically, this: 13 // * Assumes values are constant unless proven otherwise 14 // * Assumes BasicBlocks are dead unless proven otherwise 15 // * Proves values to be constant, and replaces them with constants 16 // * Proves conditional branches to be unconditional 17 // 18 //===----------------------------------------------------------------------===// 19 20 #define DEBUG_TYPE "sccp" 21 #include "llvm/Transforms/Scalar.h" 22 #include "llvm/Transforms/IPO.h" 23 #include "llvm/Constants.h" 24 #include "llvm/DerivedTypes.h" 25 #include "llvm/Instructions.h" 26 #include "llvm/Pass.h" 27 #include "llvm/Analysis/ConstantFolding.h" 28 #include "llvm/Analysis/ValueTracking.h" 29 #include "llvm/Transforms/Utils/Local.h" 30 #include "llvm/Target/TargetData.h" 31 #include "llvm/Support/CallSite.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/InstVisitor.h" 35 #include "llvm/Support/raw_ostream.h" 36 #include "llvm/ADT/DenseMap.h" 37 #include "llvm/ADT/DenseSet.h" 38 #include "llvm/ADT/PointerIntPair.h" 39 #include "llvm/ADT/SmallPtrSet.h" 40 #include "llvm/ADT/SmallVector.h" 41 #include "llvm/ADT/Statistic.h" 42 #include "llvm/ADT/STLExtras.h" 43 #include <algorithm> 44 #include <map> 45 using namespace llvm; 46 47 STATISTIC(NumInstRemoved, "Number of instructions removed"); 48 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 49 50 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 51 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 52 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 53 54 namespace { 55 /// LatticeVal class - This class represents the different lattice values that 56 /// an LLVM value may occupy. It is a simple class with value semantics. 57 /// 58 class LatticeVal { 59 enum LatticeValueTy { 60 /// undefined - This LLVM Value has no known value yet. 61 undefined, 62 63 /// constant - This LLVM Value has a specific constant value. 64 constant, 65 66 /// forcedconstant - This LLVM Value was thought to be undef until 67 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged 68 /// with another (different) constant, it goes to overdefined, instead of 69 /// asserting. 70 forcedconstant, 71 72 /// overdefined - This instruction is not known to be constant, and we know 73 /// it has a value. 74 overdefined 75 }; 76 77 /// Val: This stores the current lattice value along with the Constant* for 78 /// the constant if this is a 'constant' or 'forcedconstant' value. 79 PointerIntPair<Constant *, 2, LatticeValueTy> Val; 80 81 LatticeValueTy getLatticeValue() const { 82 return Val.getInt(); 83 } 84 85 public: 86 LatticeVal() : Val(0, undefined) {} 87 88 bool isUndefined() const { return getLatticeValue() == undefined; } 89 bool isConstant() const { 90 return getLatticeValue() == constant || getLatticeValue() == forcedconstant; 91 } 92 bool isOverdefined() const { return getLatticeValue() == overdefined; } 93 94 Constant *getConstant() const { 95 assert(isConstant() && "Cannot get the constant of a non-constant!"); 96 return Val.getPointer(); 97 } 98 99 /// markOverdefined - Return true if this is a change in status. 100 bool markOverdefined() { 101 if (isOverdefined()) 102 return false; 103 104 Val.setInt(overdefined); 105 return true; 106 } 107 108 /// markConstant - Return true if this is a change in status. 109 bool markConstant(Constant *V) { 110 if (getLatticeValue() == constant) { // Constant but not forcedconstant. 111 assert(getConstant() == V && "Marking constant with different value"); 112 return false; 113 } 114 115 if (isUndefined()) { 116 Val.setInt(constant); 117 assert(V && "Marking constant with NULL"); 118 Val.setPointer(V); 119 } else { 120 assert(getLatticeValue() == forcedconstant && 121 "Cannot move from overdefined to constant!"); 122 // Stay at forcedconstant if the constant is the same. 123 if (V == getConstant()) return false; 124 125 // Otherwise, we go to overdefined. Assumptions made based on the 126 // forced value are possibly wrong. Assuming this is another constant 127 // could expose a contradiction. 128 Val.setInt(overdefined); 129 } 130 return true; 131 } 132 133 /// getConstantInt - If this is a constant with a ConstantInt value, return it 134 /// otherwise return null. 135 ConstantInt *getConstantInt() const { 136 if (isConstant()) 137 return dyn_cast<ConstantInt>(getConstant()); 138 return 0; 139 } 140 141 void markForcedConstant(Constant *V) { 142 assert(isUndefined() && "Can't force a defined value!"); 143 Val.setInt(forcedconstant); 144 Val.setPointer(V); 145 } 146 }; 147 } // end anonymous namespace. 148 149 150 namespace { 151 152 //===----------------------------------------------------------------------===// 153 // 154 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 155 /// Constant Propagation. 156 /// 157 class SCCPSolver : public InstVisitor<SCCPSolver> { 158 const TargetData *TD; 159 SmallPtrSet<BasicBlock*, 8> BBExecutable;// The BBs that are executable. 160 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in. 161 162 /// StructValueState - This maintains ValueState for values that have 163 /// StructType, for example for formal arguments, calls, insertelement, etc. 164 /// 165 DenseMap<std::pair<Value*, unsigned>, LatticeVal> StructValueState; 166 167 /// GlobalValue - If we are tracking any values for the contents of a global 168 /// variable, we keep a mapping from the constant accessor to the element of 169 /// the global, to the currently known value. If the value becomes 170 /// overdefined, it's entry is simply removed from this map. 171 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals; 172 173 /// TrackedRetVals - If we are tracking arguments into and the return 174 /// value out of a function, it will have an entry in this map, indicating 175 /// what the known return value for the function is. 176 DenseMap<Function*, LatticeVal> TrackedRetVals; 177 178 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 179 /// that return multiple values. 180 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals; 181 182 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 183 /// represented here for efficient lookup. 184 SmallPtrSet<Function*, 16> MRVFunctionsTracked; 185 186 /// TrackingIncomingArguments - This is the set of functions for whose 187 /// arguments we make optimistic assumptions about and try to prove as 188 /// constants. 189 SmallPtrSet<Function*, 16> TrackingIncomingArguments; 190 191 /// The reason for two worklists is that overdefined is the lowest state 192 /// on the lattice, and moving things to overdefined as fast as possible 193 /// makes SCCP converge much faster. 194 /// 195 /// By having a separate worklist, we accomplish this because everything 196 /// possibly overdefined will become overdefined at the soonest possible 197 /// point. 198 SmallVector<Value*, 64> OverdefinedInstWorkList; 199 SmallVector<Value*, 64> InstWorkList; 200 201 202 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list 203 204 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not 205 /// overdefined, despite the fact that the PHI node is overdefined. 206 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs; 207 208 /// KnownFeasibleEdges - Entries in this set are edges which have already had 209 /// PHI nodes retriggered. 210 typedef std::pair<BasicBlock*, BasicBlock*> Edge; 211 DenseSet<Edge> KnownFeasibleEdges; 212 public: 213 SCCPSolver(const TargetData *td) : TD(td) {} 214 215 /// MarkBlockExecutable - This method can be used by clients to mark all of 216 /// the blocks that are known to be intrinsically live in the processed unit. 217 /// 218 /// This returns true if the block was not considered live before. 219 bool MarkBlockExecutable(BasicBlock *BB) { 220 if (!BBExecutable.insert(BB)) return false; 221 DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << "\n"); 222 BBWorkList.push_back(BB); // Add the block to the work list! 223 return true; 224 } 225 226 /// TrackValueOfGlobalVariable - Clients can use this method to 227 /// inform the SCCPSolver that it should track loads and stores to the 228 /// specified global variable if it can. This is only legal to call if 229 /// performing Interprocedural SCCP. 230 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 231 // We only track the contents of scalar globals. 232 if (GV->getType()->getElementType()->isSingleValueType()) { 233 LatticeVal &IV = TrackedGlobals[GV]; 234 if (!isa<UndefValue>(GV->getInitializer())) 235 IV.markConstant(GV->getInitializer()); 236 } 237 } 238 239 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 240 /// and out of the specified function (which cannot have its address taken), 241 /// this method must be called. 242 void AddTrackedFunction(Function *F) { 243 // Add an entry, F -> undef. 244 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) { 245 MRVFunctionsTracked.insert(F); 246 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 247 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i), 248 LatticeVal())); 249 } else 250 TrackedRetVals.insert(std::make_pair(F, LatticeVal())); 251 } 252 253 void AddArgumentTrackedFunction(Function *F) { 254 TrackingIncomingArguments.insert(F); 255 } 256 257 /// Solve - Solve for constants and executable blocks. 258 /// 259 void Solve(); 260 261 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 262 /// that branches on undef values cannot reach any of their successors. 263 /// However, this is not a safe assumption. After we solve dataflow, this 264 /// method should be use to handle this. If this returns true, the solver 265 /// should be rerun. 266 bool ResolvedUndefsIn(Function &F); 267 268 bool isBlockExecutable(BasicBlock *BB) const { 269 return BBExecutable.count(BB); 270 } 271 272 LatticeVal getLatticeValueFor(Value *V) const { 273 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V); 274 assert(I != ValueState.end() && "V is not in valuemap!"); 275 return I->second; 276 } 277 278 /*LatticeVal getStructLatticeValueFor(Value *V, unsigned i) const { 279 DenseMap<std::pair<Value*, unsigned>, LatticeVal>::const_iterator I = 280 StructValueState.find(std::make_pair(V, i)); 281 assert(I != StructValueState.end() && "V is not in valuemap!"); 282 return I->second; 283 }*/ 284 285 /// getTrackedRetVals - Get the inferred return value map. 286 /// 287 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() { 288 return TrackedRetVals; 289 } 290 291 /// getTrackedGlobals - Get and return the set of inferred initializers for 292 /// global variables. 293 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() { 294 return TrackedGlobals; 295 } 296 297 void markOverdefined(Value *V) { 298 assert(!V->getType()->isStructTy() && "Should use other method"); 299 markOverdefined(ValueState[V], V); 300 } 301 302 /// markAnythingOverdefined - Mark the specified value overdefined. This 303 /// works with both scalars and structs. 304 void markAnythingOverdefined(Value *V) { 305 if (StructType *STy = dyn_cast<StructType>(V->getType())) 306 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 307 markOverdefined(getStructValueState(V, i), V); 308 else 309 markOverdefined(V); 310 } 311 312 private: 313 // markConstant - Make a value be marked as "constant". If the value 314 // is not already a constant, add it to the instruction work list so that 315 // the users of the instruction are updated later. 316 // 317 void markConstant(LatticeVal &IV, Value *V, Constant *C) { 318 if (!IV.markConstant(C)) return; 319 DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 320 if (IV.isOverdefined()) 321 OverdefinedInstWorkList.push_back(V); 322 else 323 InstWorkList.push_back(V); 324 } 325 326 void markConstant(Value *V, Constant *C) { 327 assert(!V->getType()->isStructTy() && "Should use other method"); 328 markConstant(ValueState[V], V, C); 329 } 330 331 void markForcedConstant(Value *V, Constant *C) { 332 assert(!V->getType()->isStructTy() && "Should use other method"); 333 LatticeVal &IV = ValueState[V]; 334 IV.markForcedConstant(C); 335 DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n'); 336 if (IV.isOverdefined()) 337 OverdefinedInstWorkList.push_back(V); 338 else 339 InstWorkList.push_back(V); 340 } 341 342 343 // markOverdefined - Make a value be marked as "overdefined". If the 344 // value is not already overdefined, add it to the overdefined instruction 345 // work list so that the users of the instruction are updated later. 346 void markOverdefined(LatticeVal &IV, Value *V) { 347 if (!IV.markOverdefined()) return; 348 349 DEBUG(dbgs() << "markOverdefined: "; 350 if (Function *F = dyn_cast<Function>(V)) 351 dbgs() << "Function '" << F->getName() << "'\n"; 352 else 353 dbgs() << *V << '\n'); 354 // Only instructions go on the work list 355 OverdefinedInstWorkList.push_back(V); 356 } 357 358 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) { 359 if (IV.isOverdefined() || MergeWithV.isUndefined()) 360 return; // Noop. 361 if (MergeWithV.isOverdefined()) 362 markOverdefined(IV, V); 363 else if (IV.isUndefined()) 364 markConstant(IV, V, MergeWithV.getConstant()); 365 else if (IV.getConstant() != MergeWithV.getConstant()) 366 markOverdefined(IV, V); 367 } 368 369 void mergeInValue(Value *V, LatticeVal MergeWithV) { 370 assert(!V->getType()->isStructTy() && "Should use other method"); 371 mergeInValue(ValueState[V], V, MergeWithV); 372 } 373 374 375 /// getValueState - Return the LatticeVal object that corresponds to the 376 /// value. This function handles the case when the value hasn't been seen yet 377 /// by properly seeding constants etc. 378 LatticeVal &getValueState(Value *V) { 379 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 380 381 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I = 382 ValueState.insert(std::make_pair(V, LatticeVal())); 383 LatticeVal &LV = I.first->second; 384 385 if (!I.second) 386 return LV; // Common case, already in the map. 387 388 if (Constant *C = dyn_cast<Constant>(V)) { 389 // Undef values remain undefined. 390 if (!isa<UndefValue>(V)) 391 LV.markConstant(C); // Constants are constant 392 } 393 394 // All others are underdefined by default. 395 return LV; 396 } 397 398 /// getStructValueState - Return the LatticeVal object that corresponds to the 399 /// value/field pair. This function handles the case when the value hasn't 400 /// been seen yet by properly seeding constants etc. 401 LatticeVal &getStructValueState(Value *V, unsigned i) { 402 assert(V->getType()->isStructTy() && "Should use getValueState"); 403 assert(i < cast<StructType>(V->getType())->getNumElements() && 404 "Invalid element #"); 405 406 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator, 407 bool> I = StructValueState.insert( 408 std::make_pair(std::make_pair(V, i), LatticeVal())); 409 LatticeVal &LV = I.first->second; 410 411 if (!I.second) 412 return LV; // Common case, already in the map. 413 414 if (Constant *C = dyn_cast<Constant>(V)) { 415 if (isa<UndefValue>(C)) 416 ; // Undef values remain undefined. 417 else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) 418 LV.markConstant(CS->getOperand(i)); // Constants are constant. 419 else if (isa<ConstantAggregateZero>(C)) { 420 Type *FieldTy = cast<StructType>(V->getType())->getElementType(i); 421 LV.markConstant(Constant::getNullValue(FieldTy)); 422 } else 423 LV.markOverdefined(); // Unknown sort of constant. 424 } 425 426 // All others are underdefined by default. 427 return LV; 428 } 429 430 431 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 432 /// work list if it is not already executable. 433 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 434 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 435 return; // This edge is already known to be executable! 436 437 if (!MarkBlockExecutable(Dest)) { 438 // If the destination is already executable, we just made an *edge* 439 // feasible that wasn't before. Revisit the PHI nodes in the block 440 // because they have potentially new operands. 441 DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 442 << " -> " << Dest->getName() << "\n"); 443 444 PHINode *PN; 445 for (BasicBlock::iterator I = Dest->begin(); 446 (PN = dyn_cast<PHINode>(I)); ++I) 447 visitPHINode(*PN); 448 } 449 } 450 451 // getFeasibleSuccessors - Return a vector of booleans to indicate which 452 // successors are reachable from a given terminator instruction. 453 // 454 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs); 455 456 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 457 // block to the 'To' basic block is currently feasible. 458 // 459 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 460 461 // OperandChangedState - This method is invoked on all of the users of an 462 // instruction that was just changed state somehow. Based on this 463 // information, we need to update the specified user of this instruction. 464 // 465 void OperandChangedState(Instruction *I) { 466 if (BBExecutable.count(I->getParent())) // Inst is executable? 467 visit(*I); 468 } 469 470 /// RemoveFromOverdefinedPHIs - If I has any entries in the 471 /// UsersOfOverdefinedPHIs map for PN, remove them now. 472 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) { 473 if (UsersOfOverdefinedPHIs.empty()) return; 474 std::multimap<PHINode*, Instruction*>::iterator It, E; 475 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN); 476 while (It != E) { 477 if (It->second == I) 478 UsersOfOverdefinedPHIs.erase(It++); 479 else 480 ++It; 481 } 482 } 483 484 /// InsertInOverdefinedPHIs - Insert an entry in the UsersOfOverdefinedPHIS 485 /// map for I and PN, but if one is there already, do not create another. 486 /// (Duplicate entries do not break anything directly, but can lead to 487 /// exponential growth of the table in rare cases.) 488 void InsertInOverdefinedPHIs(Instruction *I, PHINode *PN) { 489 std::multimap<PHINode*, Instruction*>::iterator J, E; 490 tie(J, E) = UsersOfOverdefinedPHIs.equal_range(PN); 491 for (; J != E; ++J) 492 if (J->second == I) 493 return; 494 UsersOfOverdefinedPHIs.insert(std::make_pair(PN, I)); 495 } 496 497 private: 498 friend class InstVisitor<SCCPSolver>; 499 500 // visit implementations - Something changed in this instruction. Either an 501 // operand made a transition, or the instruction is newly executable. Change 502 // the value type of I to reflect these changes if appropriate. 503 void visitPHINode(PHINode &I); 504 505 // Terminators 506 void visitReturnInst(ReturnInst &I); 507 void visitTerminatorInst(TerminatorInst &TI); 508 509 void visitCastInst(CastInst &I); 510 void visitSelectInst(SelectInst &I); 511 void visitBinaryOperator(Instruction &I); 512 void visitCmpInst(CmpInst &I); 513 void visitExtractElementInst(ExtractElementInst &I); 514 void visitInsertElementInst(InsertElementInst &I); 515 void visitShuffleVectorInst(ShuffleVectorInst &I); 516 void visitExtractValueInst(ExtractValueInst &EVI); 517 void visitInsertValueInst(InsertValueInst &IVI); 518 519 // Instructions that cannot be folded away. 520 void visitStoreInst (StoreInst &I); 521 void visitLoadInst (LoadInst &I); 522 void visitGetElementPtrInst(GetElementPtrInst &I); 523 void visitCallInst (CallInst &I) { 524 visitCallSite(&I); 525 } 526 void visitInvokeInst (InvokeInst &II) { 527 visitCallSite(&II); 528 visitTerminatorInst(II); 529 } 530 void visitCallSite (CallSite CS); 531 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ } 532 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ } 533 void visitAllocaInst (Instruction &I) { markOverdefined(&I); } 534 void visitVAArgInst (Instruction &I) { markAnythingOverdefined(&I); } 535 536 void visitInstruction(Instruction &I) { 537 // If a new instruction is added to LLVM that we don't handle. 538 dbgs() << "SCCP: Don't know how to handle: " << I; 539 markAnythingOverdefined(&I); // Just in case 540 } 541 }; 542 543 } // end anonymous namespace 544 545 546 // getFeasibleSuccessors - Return a vector of booleans to indicate which 547 // successors are reachable from a given terminator instruction. 548 // 549 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI, 550 SmallVector<bool, 16> &Succs) { 551 Succs.resize(TI.getNumSuccessors()); 552 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) { 553 if (BI->isUnconditional()) { 554 Succs[0] = true; 555 return; 556 } 557 558 LatticeVal BCValue = getValueState(BI->getCondition()); 559 ConstantInt *CI = BCValue.getConstantInt(); 560 if (CI == 0) { 561 // Overdefined condition variables, and branches on unfoldable constant 562 // conditions, mean the branch could go either way. 563 if (!BCValue.isUndefined()) 564 Succs[0] = Succs[1] = true; 565 return; 566 } 567 568 // Constant condition variables mean the branch can only go a single way. 569 Succs[CI->isZero()] = true; 570 return; 571 } 572 573 if (isa<InvokeInst>(TI)) { 574 // Invoke instructions successors are always executable. 575 Succs[0] = Succs[1] = true; 576 return; 577 } 578 579 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) { 580 LatticeVal SCValue = getValueState(SI->getCondition()); 581 ConstantInt *CI = SCValue.getConstantInt(); 582 583 if (CI == 0) { // Overdefined or undefined condition? 584 // All destinations are executable! 585 if (!SCValue.isUndefined()) 586 Succs.assign(TI.getNumSuccessors(), true); 587 return; 588 } 589 590 Succs[SI->findCaseValue(CI)] = true; 591 return; 592 } 593 594 // TODO: This could be improved if the operand is a [cast of a] BlockAddress. 595 if (isa<IndirectBrInst>(&TI)) { 596 // Just mark all destinations executable! 597 Succs.assign(TI.getNumSuccessors(), true); 598 return; 599 } 600 601 #ifndef NDEBUG 602 dbgs() << "Unknown terminator instruction: " << TI << '\n'; 603 #endif 604 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 605 } 606 607 608 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 609 // block to the 'To' basic block is currently feasible. 610 // 611 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 612 assert(BBExecutable.count(To) && "Dest should always be alive!"); 613 614 // Make sure the source basic block is executable!! 615 if (!BBExecutable.count(From)) return false; 616 617 // Check to make sure this edge itself is actually feasible now. 618 TerminatorInst *TI = From->getTerminator(); 619 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 620 if (BI->isUnconditional()) 621 return true; 622 623 LatticeVal BCValue = getValueState(BI->getCondition()); 624 625 // Overdefined condition variables mean the branch could go either way, 626 // undef conditions mean that neither edge is feasible yet. 627 ConstantInt *CI = BCValue.getConstantInt(); 628 if (CI == 0) 629 return !BCValue.isUndefined(); 630 631 // Constant condition variables mean the branch can only go a single way. 632 return BI->getSuccessor(CI->isZero()) == To; 633 } 634 635 // Invoke instructions successors are always executable. 636 if (isa<InvokeInst>(TI)) 637 return true; 638 639 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 640 LatticeVal SCValue = getValueState(SI->getCondition()); 641 ConstantInt *CI = SCValue.getConstantInt(); 642 643 if (CI == 0) 644 return !SCValue.isUndefined(); 645 646 // Make sure to skip the "default value" which isn't a value 647 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i) 648 if (SI->getSuccessorValue(i) == CI) // Found the taken branch. 649 return SI->getSuccessor(i) == To; 650 651 // If the constant value is not equal to any of the branches, we must 652 // execute default branch. 653 return SI->getDefaultDest() == To; 654 } 655 656 // Just mark all destinations executable! 657 // TODO: This could be improved if the operand is a [cast of a] BlockAddress. 658 if (isa<IndirectBrInst>(TI)) 659 return true; 660 661 #ifndef NDEBUG 662 dbgs() << "Unknown terminator instruction: " << *TI << '\n'; 663 #endif 664 llvm_unreachable(0); 665 } 666 667 // visit Implementations - Something changed in this instruction, either an 668 // operand made a transition, or the instruction is newly executable. Change 669 // the value type of I to reflect these changes if appropriate. This method 670 // makes sure to do the following actions: 671 // 672 // 1. If a phi node merges two constants in, and has conflicting value coming 673 // from different branches, or if the PHI node merges in an overdefined 674 // value, then the PHI node becomes overdefined. 675 // 2. If a phi node merges only constants in, and they all agree on value, the 676 // PHI node becomes a constant value equal to that. 677 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 678 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 679 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 680 // 6. If a conditional branch has a value that is constant, make the selected 681 // destination executable 682 // 7. If a conditional branch has a value that is overdefined, make all 683 // successors executable. 684 // 685 void SCCPSolver::visitPHINode(PHINode &PN) { 686 // If this PN returns a struct, just mark the result overdefined. 687 // TODO: We could do a lot better than this if code actually uses this. 688 if (PN.getType()->isStructTy()) 689 return markAnythingOverdefined(&PN); 690 691 if (getValueState(&PN).isOverdefined()) { 692 // There may be instructions using this PHI node that are not overdefined 693 // themselves. If so, make sure that they know that the PHI node operand 694 // changed. 695 std::multimap<PHINode*, Instruction*>::iterator I, E; 696 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN); 697 if (I == E) 698 return; 699 700 SmallVector<Instruction*, 16> Users; 701 for (; I != E; ++I) 702 Users.push_back(I->second); 703 while (!Users.empty()) 704 visit(Users.pop_back_val()); 705 return; // Quick exit 706 } 707 708 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 709 // and slow us down a lot. Just mark them overdefined. 710 if (PN.getNumIncomingValues() > 64) 711 return markOverdefined(&PN); 712 713 // Look at all of the executable operands of the PHI node. If any of them 714 // are overdefined, the PHI becomes overdefined as well. If they are all 715 // constant, and they agree with each other, the PHI becomes the identical 716 // constant. If they are constant and don't agree, the PHI is overdefined. 717 // If there are no executable operands, the PHI remains undefined. 718 // 719 Constant *OperandVal = 0; 720 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 721 LatticeVal IV = getValueState(PN.getIncomingValue(i)); 722 if (IV.isUndefined()) continue; // Doesn't influence PHI node. 723 724 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 725 continue; 726 727 if (IV.isOverdefined()) // PHI node becomes overdefined! 728 return markOverdefined(&PN); 729 730 if (OperandVal == 0) { // Grab the first value. 731 OperandVal = IV.getConstant(); 732 continue; 733 } 734 735 // There is already a reachable operand. If we conflict with it, 736 // then the PHI node becomes overdefined. If we agree with it, we 737 // can continue on. 738 739 // Check to see if there are two different constants merging, if so, the PHI 740 // node is overdefined. 741 if (IV.getConstant() != OperandVal) 742 return markOverdefined(&PN); 743 } 744 745 // If we exited the loop, this means that the PHI node only has constant 746 // arguments that agree with each other(and OperandVal is the constant) or 747 // OperandVal is null because there are no defined incoming arguments. If 748 // this is the case, the PHI remains undefined. 749 // 750 if (OperandVal) 751 markConstant(&PN, OperandVal); // Acquire operand value 752 } 753 754 755 756 757 void SCCPSolver::visitReturnInst(ReturnInst &I) { 758 if (I.getNumOperands() == 0) return; // ret void 759 760 Function *F = I.getParent()->getParent(); 761 Value *ResultOp = I.getOperand(0); 762 763 // If we are tracking the return value of this function, merge it in. 764 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 765 DenseMap<Function*, LatticeVal>::iterator TFRVI = 766 TrackedRetVals.find(F); 767 if (TFRVI != TrackedRetVals.end()) { 768 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 769 return; 770 } 771 } 772 773 // Handle functions that return multiple values. 774 if (!TrackedMultipleRetVals.empty()) { 775 if (StructType *STy = dyn_cast<StructType>(ResultOp->getType())) 776 if (MRVFunctionsTracked.count(F)) 777 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 778 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 779 getStructValueState(ResultOp, i)); 780 781 } 782 } 783 784 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) { 785 SmallVector<bool, 16> SuccFeasible; 786 getFeasibleSuccessors(TI, SuccFeasible); 787 788 BasicBlock *BB = TI.getParent(); 789 790 // Mark all feasible successors executable. 791 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 792 if (SuccFeasible[i]) 793 markEdgeExecutable(BB, TI.getSuccessor(i)); 794 } 795 796 void SCCPSolver::visitCastInst(CastInst &I) { 797 LatticeVal OpSt = getValueState(I.getOperand(0)); 798 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand 799 markOverdefined(&I); 800 else if (OpSt.isConstant()) // Propagate constant value 801 markConstant(&I, ConstantExpr::getCast(I.getOpcode(), 802 OpSt.getConstant(), I.getType())); 803 } 804 805 806 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 807 // If this returns a struct, mark all elements over defined, we don't track 808 // structs in structs. 809 if (EVI.getType()->isStructTy()) 810 return markAnythingOverdefined(&EVI); 811 812 // If this is extracting from more than one level of struct, we don't know. 813 if (EVI.getNumIndices() != 1) 814 return markOverdefined(&EVI); 815 816 Value *AggVal = EVI.getAggregateOperand(); 817 if (AggVal->getType()->isStructTy()) { 818 unsigned i = *EVI.idx_begin(); 819 LatticeVal EltVal = getStructValueState(AggVal, i); 820 mergeInValue(getValueState(&EVI), &EVI, EltVal); 821 } else { 822 // Otherwise, must be extracting from an array. 823 return markOverdefined(&EVI); 824 } 825 } 826 827 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 828 StructType *STy = dyn_cast<StructType>(IVI.getType()); 829 if (STy == 0) 830 return markOverdefined(&IVI); 831 832 // If this has more than one index, we can't handle it, drive all results to 833 // undef. 834 if (IVI.getNumIndices() != 1) 835 return markAnythingOverdefined(&IVI); 836 837 Value *Aggr = IVI.getAggregateOperand(); 838 unsigned Idx = *IVI.idx_begin(); 839 840 // Compute the result based on what we're inserting. 841 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 842 // This passes through all values that aren't the inserted element. 843 if (i != Idx) { 844 LatticeVal EltVal = getStructValueState(Aggr, i); 845 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 846 continue; 847 } 848 849 Value *Val = IVI.getInsertedValueOperand(); 850 if (Val->getType()->isStructTy()) 851 // We don't track structs in structs. 852 markOverdefined(getStructValueState(&IVI, i), &IVI); 853 else { 854 LatticeVal InVal = getValueState(Val); 855 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 856 } 857 } 858 } 859 860 void SCCPSolver::visitSelectInst(SelectInst &I) { 861 // If this select returns a struct, just mark the result overdefined. 862 // TODO: We could do a lot better than this if code actually uses this. 863 if (I.getType()->isStructTy()) 864 return markAnythingOverdefined(&I); 865 866 LatticeVal CondValue = getValueState(I.getCondition()); 867 if (CondValue.isUndefined()) 868 return; 869 870 if (ConstantInt *CondCB = CondValue.getConstantInt()) { 871 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 872 mergeInValue(&I, getValueState(OpVal)); 873 return; 874 } 875 876 // Otherwise, the condition is overdefined or a constant we can't evaluate. 877 // See if we can produce something better than overdefined based on the T/F 878 // value. 879 LatticeVal TVal = getValueState(I.getTrueValue()); 880 LatticeVal FVal = getValueState(I.getFalseValue()); 881 882 // select ?, C, C -> C. 883 if (TVal.isConstant() && FVal.isConstant() && 884 TVal.getConstant() == FVal.getConstant()) 885 return markConstant(&I, FVal.getConstant()); 886 887 if (TVal.isUndefined()) // select ?, undef, X -> X. 888 return mergeInValue(&I, FVal); 889 if (FVal.isUndefined()) // select ?, X, undef -> X. 890 return mergeInValue(&I, TVal); 891 markOverdefined(&I); 892 } 893 894 // Handle Binary Operators. 895 void SCCPSolver::visitBinaryOperator(Instruction &I) { 896 LatticeVal V1State = getValueState(I.getOperand(0)); 897 LatticeVal V2State = getValueState(I.getOperand(1)); 898 899 LatticeVal &IV = ValueState[&I]; 900 if (IV.isOverdefined()) return; 901 902 if (V1State.isConstant() && V2State.isConstant()) 903 return markConstant(IV, &I, 904 ConstantExpr::get(I.getOpcode(), V1State.getConstant(), 905 V2State.getConstant())); 906 907 // If something is undef, wait for it to resolve. 908 if (!V1State.isOverdefined() && !V2State.isOverdefined()) 909 return; 910 911 // Otherwise, one of our operands is overdefined. Try to produce something 912 // better than overdefined with some tricks. 913 914 // If this is an AND or OR with 0 or -1, it doesn't matter that the other 915 // operand is overdefined. 916 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) { 917 LatticeVal *NonOverdefVal = 0; 918 if (!V1State.isOverdefined()) 919 NonOverdefVal = &V1State; 920 else if (!V2State.isOverdefined()) 921 NonOverdefVal = &V2State; 922 923 if (NonOverdefVal) { 924 if (NonOverdefVal->isUndefined()) { 925 // Could annihilate value. 926 if (I.getOpcode() == Instruction::And) 927 markConstant(IV, &I, Constant::getNullValue(I.getType())); 928 else if (VectorType *PT = dyn_cast<VectorType>(I.getType())) 929 markConstant(IV, &I, Constant::getAllOnesValue(PT)); 930 else 931 markConstant(IV, &I, 932 Constant::getAllOnesValue(I.getType())); 933 return; 934 } 935 936 if (I.getOpcode() == Instruction::And) { 937 // X and 0 = 0 938 if (NonOverdefVal->getConstant()->isNullValue()) 939 return markConstant(IV, &I, NonOverdefVal->getConstant()); 940 } else { 941 if (ConstantInt *CI = NonOverdefVal->getConstantInt()) 942 if (CI->isAllOnesValue()) // X or -1 = -1 943 return markConstant(IV, &I, NonOverdefVal->getConstant()); 944 } 945 } 946 } 947 948 949 // If both operands are PHI nodes, it is possible that this instruction has 950 // a constant value, despite the fact that the PHI node doesn't. Check for 951 // this condition now. 952 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0))) 953 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1))) 954 if (PN1->getParent() == PN2->getParent()) { 955 // Since the two PHI nodes are in the same basic block, they must have 956 // entries for the same predecessors. Walk the predecessor list, and 957 // if all of the incoming values are constants, and the result of 958 // evaluating this expression with all incoming value pairs is the 959 // same, then this expression is a constant even though the PHI node 960 // is not a constant! 961 LatticeVal Result; 962 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) { 963 LatticeVal In1 = getValueState(PN1->getIncomingValue(i)); 964 BasicBlock *InBlock = PN1->getIncomingBlock(i); 965 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock)); 966 967 if (In1.isOverdefined() || In2.isOverdefined()) { 968 Result.markOverdefined(); 969 break; // Cannot fold this operation over the PHI nodes! 970 } 971 972 if (In1.isConstant() && In2.isConstant()) { 973 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(), 974 In2.getConstant()); 975 if (Result.isUndefined()) 976 Result.markConstant(V); 977 else if (Result.isConstant() && Result.getConstant() != V) { 978 Result.markOverdefined(); 979 break; 980 } 981 } 982 } 983 984 // If we found a constant value here, then we know the instruction is 985 // constant despite the fact that the PHI nodes are overdefined. 986 if (Result.isConstant()) { 987 markConstant(IV, &I, Result.getConstant()); 988 // Remember that this instruction is virtually using the PHI node 989 // operands. 990 InsertInOverdefinedPHIs(&I, PN1); 991 InsertInOverdefinedPHIs(&I, PN2); 992 return; 993 } 994 995 if (Result.isUndefined()) 996 return; 997 998 // Okay, this really is overdefined now. Since we might have 999 // speculatively thought that this was not overdefined before, and 1000 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs, 1001 // make sure to clean out any entries that we put there, for 1002 // efficiency. 1003 RemoveFromOverdefinedPHIs(&I, PN1); 1004 RemoveFromOverdefinedPHIs(&I, PN2); 1005 } 1006 1007 markOverdefined(&I); 1008 } 1009 1010 // Handle ICmpInst instruction. 1011 void SCCPSolver::visitCmpInst(CmpInst &I) { 1012 LatticeVal V1State = getValueState(I.getOperand(0)); 1013 LatticeVal V2State = getValueState(I.getOperand(1)); 1014 1015 LatticeVal &IV = ValueState[&I]; 1016 if (IV.isOverdefined()) return; 1017 1018 if (V1State.isConstant() && V2State.isConstant()) 1019 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(), 1020 V1State.getConstant(), 1021 V2State.getConstant())); 1022 1023 // If operands are still undefined, wait for it to resolve. 1024 if (!V1State.isOverdefined() && !V2State.isOverdefined()) 1025 return; 1026 1027 // If something is overdefined, use some tricks to avoid ending up and over 1028 // defined if we can. 1029 1030 // If both operands are PHI nodes, it is possible that this instruction has 1031 // a constant value, despite the fact that the PHI node doesn't. Check for 1032 // this condition now. 1033 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0))) 1034 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1))) 1035 if (PN1->getParent() == PN2->getParent()) { 1036 // Since the two PHI nodes are in the same basic block, they must have 1037 // entries for the same predecessors. Walk the predecessor list, and 1038 // if all of the incoming values are constants, and the result of 1039 // evaluating this expression with all incoming value pairs is the 1040 // same, then this expression is a constant even though the PHI node 1041 // is not a constant! 1042 LatticeVal Result; 1043 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) { 1044 LatticeVal In1 = getValueState(PN1->getIncomingValue(i)); 1045 BasicBlock *InBlock = PN1->getIncomingBlock(i); 1046 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock)); 1047 1048 if (In1.isOverdefined() || In2.isOverdefined()) { 1049 Result.markOverdefined(); 1050 break; // Cannot fold this operation over the PHI nodes! 1051 } 1052 1053 if (In1.isConstant() && In2.isConstant()) { 1054 Constant *V = ConstantExpr::getCompare(I.getPredicate(), 1055 In1.getConstant(), 1056 In2.getConstant()); 1057 if (Result.isUndefined()) 1058 Result.markConstant(V); 1059 else if (Result.isConstant() && Result.getConstant() != V) { 1060 Result.markOverdefined(); 1061 break; 1062 } 1063 } 1064 } 1065 1066 // If we found a constant value here, then we know the instruction is 1067 // constant despite the fact that the PHI nodes are overdefined. 1068 if (Result.isConstant()) { 1069 markConstant(&I, Result.getConstant()); 1070 // Remember that this instruction is virtually using the PHI node 1071 // operands. 1072 InsertInOverdefinedPHIs(&I, PN1); 1073 InsertInOverdefinedPHIs(&I, PN2); 1074 return; 1075 } 1076 1077 if (Result.isUndefined()) 1078 return; 1079 1080 // Okay, this really is overdefined now. Since we might have 1081 // speculatively thought that this was not overdefined before, and 1082 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs, 1083 // make sure to clean out any entries that we put there, for 1084 // efficiency. 1085 RemoveFromOverdefinedPHIs(&I, PN1); 1086 RemoveFromOverdefinedPHIs(&I, PN2); 1087 } 1088 1089 markOverdefined(&I); 1090 } 1091 1092 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) { 1093 // TODO : SCCP does not handle vectors properly. 1094 return markOverdefined(&I); 1095 1096 #if 0 1097 LatticeVal &ValState = getValueState(I.getOperand(0)); 1098 LatticeVal &IdxState = getValueState(I.getOperand(1)); 1099 1100 if (ValState.isOverdefined() || IdxState.isOverdefined()) 1101 markOverdefined(&I); 1102 else if(ValState.isConstant() && IdxState.isConstant()) 1103 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(), 1104 IdxState.getConstant())); 1105 #endif 1106 } 1107 1108 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) { 1109 // TODO : SCCP does not handle vectors properly. 1110 return markOverdefined(&I); 1111 #if 0 1112 LatticeVal &ValState = getValueState(I.getOperand(0)); 1113 LatticeVal &EltState = getValueState(I.getOperand(1)); 1114 LatticeVal &IdxState = getValueState(I.getOperand(2)); 1115 1116 if (ValState.isOverdefined() || EltState.isOverdefined() || 1117 IdxState.isOverdefined()) 1118 markOverdefined(&I); 1119 else if(ValState.isConstant() && EltState.isConstant() && 1120 IdxState.isConstant()) 1121 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(), 1122 EltState.getConstant(), 1123 IdxState.getConstant())); 1124 else if (ValState.isUndefined() && EltState.isConstant() && 1125 IdxState.isConstant()) 1126 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()), 1127 EltState.getConstant(), 1128 IdxState.getConstant())); 1129 #endif 1130 } 1131 1132 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) { 1133 // TODO : SCCP does not handle vectors properly. 1134 return markOverdefined(&I); 1135 #if 0 1136 LatticeVal &V1State = getValueState(I.getOperand(0)); 1137 LatticeVal &V2State = getValueState(I.getOperand(1)); 1138 LatticeVal &MaskState = getValueState(I.getOperand(2)); 1139 1140 if (MaskState.isUndefined() || 1141 (V1State.isUndefined() && V2State.isUndefined())) 1142 return; // Undefined output if mask or both inputs undefined. 1143 1144 if (V1State.isOverdefined() || V2State.isOverdefined() || 1145 MaskState.isOverdefined()) { 1146 markOverdefined(&I); 1147 } else { 1148 // A mix of constant/undef inputs. 1149 Constant *V1 = V1State.isConstant() ? 1150 V1State.getConstant() : UndefValue::get(I.getType()); 1151 Constant *V2 = V2State.isConstant() ? 1152 V2State.getConstant() : UndefValue::get(I.getType()); 1153 Constant *Mask = MaskState.isConstant() ? 1154 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType()); 1155 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask)); 1156 } 1157 #endif 1158 } 1159 1160 // Handle getelementptr instructions. If all operands are constants then we 1161 // can turn this into a getelementptr ConstantExpr. 1162 // 1163 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1164 if (ValueState[&I].isOverdefined()) return; 1165 1166 SmallVector<Constant*, 8> Operands; 1167 Operands.reserve(I.getNumOperands()); 1168 1169 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1170 LatticeVal State = getValueState(I.getOperand(i)); 1171 if (State.isUndefined()) 1172 return; // Operands are not resolved yet. 1173 1174 if (State.isOverdefined()) 1175 return markOverdefined(&I); 1176 1177 assert(State.isConstant() && "Unknown state!"); 1178 Operands.push_back(State.getConstant()); 1179 } 1180 1181 Constant *Ptr = Operands[0]; 1182 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0]+1, 1183 Operands.size()-1)); 1184 } 1185 1186 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1187 // If this store is of a struct, ignore it. 1188 if (SI.getOperand(0)->getType()->isStructTy()) 1189 return; 1190 1191 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1192 return; 1193 1194 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1195 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV); 1196 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return; 1197 1198 // Get the value we are storing into the global, then merge it. 1199 mergeInValue(I->second, GV, getValueState(SI.getOperand(0))); 1200 if (I->second.isOverdefined()) 1201 TrackedGlobals.erase(I); // No need to keep tracking this! 1202 } 1203 1204 1205 // Handle load instructions. If the operand is a constant pointer to a constant 1206 // global, we can replace the load with the loaded constant value! 1207 void SCCPSolver::visitLoadInst(LoadInst &I) { 1208 // If this load is of a struct, just mark the result overdefined. 1209 if (I.getType()->isStructTy()) 1210 return markAnythingOverdefined(&I); 1211 1212 LatticeVal PtrVal = getValueState(I.getOperand(0)); 1213 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet! 1214 1215 LatticeVal &IV = ValueState[&I]; 1216 if (IV.isOverdefined()) return; 1217 1218 if (!PtrVal.isConstant() || I.isVolatile()) 1219 return markOverdefined(IV, &I); 1220 1221 Constant *Ptr = PtrVal.getConstant(); 1222 1223 // load null -> null 1224 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0) 1225 return markConstant(IV, &I, Constant::getNullValue(I.getType())); 1226 1227 // Transform load (constant global) into the value loaded. 1228 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) { 1229 if (!TrackedGlobals.empty()) { 1230 // If we are tracking this global, merge in the known value for it. 1231 DenseMap<GlobalVariable*, LatticeVal>::iterator It = 1232 TrackedGlobals.find(GV); 1233 if (It != TrackedGlobals.end()) { 1234 mergeInValue(IV, &I, It->second); 1235 return; 1236 } 1237 } 1238 } 1239 1240 // Transform load from a constant into a constant if possible. 1241 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD)) 1242 return markConstant(IV, &I, C); 1243 1244 // Otherwise we cannot say for certain what value this load will produce. 1245 // Bail out. 1246 markOverdefined(IV, &I); 1247 } 1248 1249 void SCCPSolver::visitCallSite(CallSite CS) { 1250 Function *F = CS.getCalledFunction(); 1251 Instruction *I = CS.getInstruction(); 1252 1253 // The common case is that we aren't tracking the callee, either because we 1254 // are not doing interprocedural analysis or the callee is indirect, or is 1255 // external. Handle these cases first. 1256 if (F == 0 || F->isDeclaration()) { 1257 CallOverdefined: 1258 // Void return and not tracking callee, just bail. 1259 if (I->getType()->isVoidTy()) return; 1260 1261 // Otherwise, if we have a single return value case, and if the function is 1262 // a declaration, maybe we can constant fold it. 1263 if (F && F->isDeclaration() && !I->getType()->isStructTy() && 1264 canConstantFoldCallTo(F)) { 1265 1266 SmallVector<Constant*, 8> Operands; 1267 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end(); 1268 AI != E; ++AI) { 1269 LatticeVal State = getValueState(*AI); 1270 1271 if (State.isUndefined()) 1272 return; // Operands are not resolved yet. 1273 if (State.isOverdefined()) 1274 return markOverdefined(I); 1275 assert(State.isConstant() && "Unknown state!"); 1276 Operands.push_back(State.getConstant()); 1277 } 1278 1279 // If we can constant fold this, mark the result of the call as a 1280 // constant. 1281 if (Constant *C = ConstantFoldCall(F, Operands)) 1282 return markConstant(I, C); 1283 } 1284 1285 // Otherwise, we don't know anything about this call, mark it overdefined. 1286 return markAnythingOverdefined(I); 1287 } 1288 1289 // If this is a local function that doesn't have its address taken, mark its 1290 // entry block executable and merge in the actual arguments to the call into 1291 // the formal arguments of the function. 1292 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){ 1293 MarkBlockExecutable(F->begin()); 1294 1295 // Propagate information from this call site into the callee. 1296 CallSite::arg_iterator CAI = CS.arg_begin(); 1297 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1298 AI != E; ++AI, ++CAI) { 1299 // If this argument is byval, and if the function is not readonly, there 1300 // will be an implicit copy formed of the input aggregate. 1301 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1302 markOverdefined(AI); 1303 continue; 1304 } 1305 1306 if (StructType *STy = dyn_cast<StructType>(AI->getType())) { 1307 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1308 LatticeVal CallArg = getStructValueState(*CAI, i); 1309 mergeInValue(getStructValueState(AI, i), AI, CallArg); 1310 } 1311 } else { 1312 mergeInValue(AI, getValueState(*CAI)); 1313 } 1314 } 1315 } 1316 1317 // If this is a single/zero retval case, see if we're tracking the function. 1318 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) { 1319 if (!MRVFunctionsTracked.count(F)) 1320 goto CallOverdefined; // Not tracking this callee. 1321 1322 // If we are tracking this callee, propagate the result of the function 1323 // into this call site. 1324 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1325 mergeInValue(getStructValueState(I, i), I, 1326 TrackedMultipleRetVals[std::make_pair(F, i)]); 1327 } else { 1328 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F); 1329 if (TFRVI == TrackedRetVals.end()) 1330 goto CallOverdefined; // Not tracking this callee. 1331 1332 // If so, propagate the return value of the callee into this call result. 1333 mergeInValue(I, TFRVI->second); 1334 } 1335 } 1336 1337 void SCCPSolver::Solve() { 1338 // Process the work lists until they are empty! 1339 while (!BBWorkList.empty() || !InstWorkList.empty() || 1340 !OverdefinedInstWorkList.empty()) { 1341 // Process the overdefined instruction's work list first, which drives other 1342 // things to overdefined more quickly. 1343 while (!OverdefinedInstWorkList.empty()) { 1344 Value *I = OverdefinedInstWorkList.pop_back_val(); 1345 1346 DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1347 1348 // "I" got into the work list because it either made the transition from 1349 // bottom to constant 1350 // 1351 // Anything on this worklist that is overdefined need not be visited 1352 // since all of its users will have already been marked as overdefined 1353 // Update all of the users of this instruction's value. 1354 // 1355 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1356 UI != E; ++UI) 1357 if (Instruction *I = dyn_cast<Instruction>(*UI)) 1358 OperandChangedState(I); 1359 } 1360 1361 // Process the instruction work list. 1362 while (!InstWorkList.empty()) { 1363 Value *I = InstWorkList.pop_back_val(); 1364 1365 DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1366 1367 // "I" got into the work list because it made the transition from undef to 1368 // constant. 1369 // 1370 // Anything on this worklist that is overdefined need not be visited 1371 // since all of its users will have already been marked as overdefined. 1372 // Update all of the users of this instruction's value. 1373 // 1374 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1375 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1376 UI != E; ++UI) 1377 if (Instruction *I = dyn_cast<Instruction>(*UI)) 1378 OperandChangedState(I); 1379 } 1380 1381 // Process the basic block work list. 1382 while (!BBWorkList.empty()) { 1383 BasicBlock *BB = BBWorkList.back(); 1384 BBWorkList.pop_back(); 1385 1386 DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1387 1388 // Notify all instructions in this basic block that they are newly 1389 // executable. 1390 visit(BB); 1391 } 1392 } 1393 } 1394 1395 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1396 /// that branches on undef values cannot reach any of their successors. 1397 /// However, this is not a safe assumption. After we solve dataflow, this 1398 /// method should be use to handle this. If this returns true, the solver 1399 /// should be rerun. 1400 /// 1401 /// This method handles this by finding an unresolved branch and marking it one 1402 /// of the edges from the block as being feasible, even though the condition 1403 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1404 /// CFG and only slightly pessimizes the analysis results (by marking one, 1405 /// potentially infeasible, edge feasible). This cannot usefully modify the 1406 /// constraints on the condition of the branch, as that would impact other users 1407 /// of the value. 1408 /// 1409 /// This scan also checks for values that use undefs, whose results are actually 1410 /// defined. For example, 'zext i8 undef to i32' should produce all zeros 1411 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero, 1412 /// even if X isn't defined. 1413 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1414 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 1415 if (!BBExecutable.count(BB)) 1416 continue; 1417 1418 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 1419 // Look for instructions which produce undef values. 1420 if (I->getType()->isVoidTy()) continue; 1421 1422 if (StructType *STy = dyn_cast<StructType>(I->getType())) { 1423 // Only a few things that can be structs matter for undef. Just send 1424 // all their results to overdefined. We could be more precise than this 1425 // but it isn't worth bothering. 1426 if (isa<CallInst>(I) || isa<SelectInst>(I)) { 1427 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1428 LatticeVal &LV = getStructValueState(I, i); 1429 if (LV.isUndefined()) 1430 markOverdefined(LV, I); 1431 } 1432 } 1433 continue; 1434 } 1435 1436 LatticeVal &LV = getValueState(I); 1437 if (!LV.isUndefined()) continue; 1438 1439 // No instructions using structs need disambiguation. 1440 if (I->getOperand(0)->getType()->isStructTy()) 1441 continue; 1442 1443 // Get the lattice values of the first two operands for use below. 1444 LatticeVal Op0LV = getValueState(I->getOperand(0)); 1445 LatticeVal Op1LV; 1446 if (I->getNumOperands() == 2) { 1447 // No instructions using structs need disambiguation. 1448 if (I->getOperand(1)->getType()->isStructTy()) 1449 continue; 1450 1451 // If this is a two-operand instruction, and if both operands are 1452 // undefs, the result stays undef. 1453 Op1LV = getValueState(I->getOperand(1)); 1454 if (Op0LV.isUndefined() && Op1LV.isUndefined()) 1455 continue; 1456 } 1457 1458 // If this is an instructions whose result is defined even if the input is 1459 // not fully defined, propagate the information. 1460 Type *ITy = I->getType(); 1461 switch (I->getOpcode()) { 1462 default: break; // Leave the instruction as an undef. 1463 case Instruction::ZExt: 1464 // After a zero extend, we know the top part is zero. SExt doesn't have 1465 // to be handled here, because we don't know whether the top part is 1's 1466 // or 0's. 1467 case Instruction::SIToFP: // some FP values are not possible, just use 0. 1468 case Instruction::UIToFP: // some FP values are not possible, just use 0. 1469 markForcedConstant(I, Constant::getNullValue(ITy)); 1470 return true; 1471 case Instruction::Mul: 1472 case Instruction::And: 1473 // undef * X -> 0. X could be zero. 1474 // undef & X -> 0. X could be zero. 1475 markForcedConstant(I, Constant::getNullValue(ITy)); 1476 return true; 1477 1478 case Instruction::Or: 1479 // undef | X -> -1. X could be -1. 1480 markForcedConstant(I, Constant::getAllOnesValue(ITy)); 1481 return true; 1482 1483 case Instruction::SDiv: 1484 case Instruction::UDiv: 1485 case Instruction::SRem: 1486 case Instruction::URem: 1487 // X / undef -> undef. No change. 1488 // X % undef -> undef. No change. 1489 if (Op1LV.isUndefined()) break; 1490 1491 // undef / X -> 0. X could be maxint. 1492 // undef % X -> 0. X could be 1. 1493 markForcedConstant(I, Constant::getNullValue(ITy)); 1494 return true; 1495 1496 case Instruction::AShr: 1497 // undef >>s X -> undef. No change. 1498 if (Op0LV.isUndefined()) break; 1499 1500 // X >>s undef -> X. X could be 0, X could have the high-bit known set. 1501 if (Op0LV.isConstant()) 1502 markForcedConstant(I, Op0LV.getConstant()); 1503 else 1504 markOverdefined(I); 1505 return true; 1506 case Instruction::LShr: 1507 case Instruction::Shl: 1508 // undef >> X -> undef. No change. 1509 // undef << X -> undef. No change. 1510 if (Op0LV.isUndefined()) break; 1511 1512 // X >> undef -> 0. X could be 0. 1513 // X << undef -> 0. X could be 0. 1514 markForcedConstant(I, Constant::getNullValue(ITy)); 1515 return true; 1516 case Instruction::Select: 1517 // undef ? X : Y -> X or Y. There could be commonality between X/Y. 1518 if (Op0LV.isUndefined()) { 1519 if (!Op1LV.isConstant()) // Pick the constant one if there is any. 1520 Op1LV = getValueState(I->getOperand(2)); 1521 } else if (Op1LV.isUndefined()) { 1522 // c ? undef : undef -> undef. No change. 1523 Op1LV = getValueState(I->getOperand(2)); 1524 if (Op1LV.isUndefined()) 1525 break; 1526 // Otherwise, c ? undef : x -> x. 1527 } else { 1528 // Leave Op1LV as Operand(1)'s LatticeValue. 1529 } 1530 1531 if (Op1LV.isConstant()) 1532 markForcedConstant(I, Op1LV.getConstant()); 1533 else 1534 markOverdefined(I); 1535 return true; 1536 case Instruction::Call: 1537 // If a call has an undef result, it is because it is constant foldable 1538 // but one of the inputs was undef. Just force the result to 1539 // overdefined. 1540 markOverdefined(I); 1541 return true; 1542 } 1543 } 1544 1545 // Check to see if we have a branch or switch on an undefined value. If so 1546 // we force the branch to go one way or the other to make the successor 1547 // values live. It doesn't really matter which way we force it. 1548 TerminatorInst *TI = BB->getTerminator(); 1549 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1550 if (!BI->isConditional()) continue; 1551 if (!getValueState(BI->getCondition()).isUndefined()) 1552 continue; 1553 1554 // If the input to SCCP is actually branch on undef, fix the undef to 1555 // false. 1556 if (isa<UndefValue>(BI->getCondition())) { 1557 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1558 markEdgeExecutable(BB, TI->getSuccessor(1)); 1559 return true; 1560 } 1561 1562 // Otherwise, it is a branch on a symbolic value which is currently 1563 // considered to be undef. Handle this by forcing the input value to the 1564 // branch to false. 1565 markForcedConstant(BI->getCondition(), 1566 ConstantInt::getFalse(TI->getContext())); 1567 return true; 1568 } 1569 1570 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1571 if (SI->getNumSuccessors() < 2) // no cases 1572 continue; 1573 if (!getValueState(SI->getCondition()).isUndefined()) 1574 continue; 1575 1576 // If the input to SCCP is actually switch on undef, fix the undef to 1577 // the first constant. 1578 if (isa<UndefValue>(SI->getCondition())) { 1579 SI->setCondition(SI->getCaseValue(1)); 1580 markEdgeExecutable(BB, TI->getSuccessor(1)); 1581 return true; 1582 } 1583 1584 markForcedConstant(SI->getCondition(), SI->getCaseValue(1)); 1585 return true; 1586 } 1587 } 1588 1589 return false; 1590 } 1591 1592 1593 namespace { 1594 //===--------------------------------------------------------------------===// 1595 // 1596 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1597 /// Sparse Conditional Constant Propagator. 1598 /// 1599 struct SCCP : public FunctionPass { 1600 static char ID; // Pass identification, replacement for typeid 1601 SCCP() : FunctionPass(ID) { 1602 initializeSCCPPass(*PassRegistry::getPassRegistry()); 1603 } 1604 1605 // runOnFunction - Run the Sparse Conditional Constant Propagation 1606 // algorithm, and return true if the function was modified. 1607 // 1608 bool runOnFunction(Function &F); 1609 }; 1610 } // end anonymous namespace 1611 1612 char SCCP::ID = 0; 1613 INITIALIZE_PASS(SCCP, "sccp", 1614 "Sparse Conditional Constant Propagation", false, false) 1615 1616 // createSCCPPass - This is the public interface to this file. 1617 FunctionPass *llvm::createSCCPPass() { 1618 return new SCCP(); 1619 } 1620 1621 static void DeleteInstructionInBlock(BasicBlock *BB) { 1622 DEBUG(dbgs() << " BasicBlock Dead:" << *BB); 1623 ++NumDeadBlocks; 1624 1625 // Delete the instructions backwards, as it has a reduced likelihood of 1626 // having to update as many def-use and use-def chains. 1627 while (!isa<TerminatorInst>(BB->begin())) { 1628 Instruction *I = --BasicBlock::iterator(BB->getTerminator()); 1629 1630 if (!I->use_empty()) 1631 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1632 BB->getInstList().erase(I); 1633 ++NumInstRemoved; 1634 } 1635 } 1636 1637 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm, 1638 // and return true if the function was modified. 1639 // 1640 bool SCCP::runOnFunction(Function &F) { 1641 DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1642 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>()); 1643 1644 // Mark the first block of the function as being executable. 1645 Solver.MarkBlockExecutable(F.begin()); 1646 1647 // Mark all arguments to the function as being overdefined. 1648 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI) 1649 Solver.markAnythingOverdefined(AI); 1650 1651 // Solve for constants. 1652 bool ResolvedUndefs = true; 1653 while (ResolvedUndefs) { 1654 Solver.Solve(); 1655 DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1656 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1657 } 1658 1659 bool MadeChanges = false; 1660 1661 // If we decided that there are basic blocks that are dead in this function, 1662 // delete their contents now. Note that we cannot actually delete the blocks, 1663 // as we cannot modify the CFG of the function. 1664 1665 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 1666 if (!Solver.isBlockExecutable(BB)) { 1667 DeleteInstructionInBlock(BB); 1668 MadeChanges = true; 1669 continue; 1670 } 1671 1672 // Iterate over all of the instructions in a function, replacing them with 1673 // constants if we have found them to be of constant values. 1674 // 1675 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1676 Instruction *Inst = BI++; 1677 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst)) 1678 continue; 1679 1680 // TODO: Reconstruct structs from their elements. 1681 if (Inst->getType()->isStructTy()) 1682 continue; 1683 1684 LatticeVal IV = Solver.getLatticeValueFor(Inst); 1685 if (IV.isOverdefined()) 1686 continue; 1687 1688 Constant *Const = IV.isConstant() 1689 ? IV.getConstant() : UndefValue::get(Inst->getType()); 1690 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst); 1691 1692 // Replaces all of the uses of a variable with uses of the constant. 1693 Inst->replaceAllUsesWith(Const); 1694 1695 // Delete the instruction. 1696 Inst->eraseFromParent(); 1697 1698 // Hey, we just changed something! 1699 MadeChanges = true; 1700 ++NumInstRemoved; 1701 } 1702 } 1703 1704 return MadeChanges; 1705 } 1706 1707 namespace { 1708 //===--------------------------------------------------------------------===// 1709 // 1710 /// IPSCCP Class - This class implements interprocedural Sparse Conditional 1711 /// Constant Propagation. 1712 /// 1713 struct IPSCCP : public ModulePass { 1714 static char ID; 1715 IPSCCP() : ModulePass(ID) { 1716 initializeIPSCCPPass(*PassRegistry::getPassRegistry()); 1717 } 1718 bool runOnModule(Module &M); 1719 }; 1720 } // end anonymous namespace 1721 1722 char IPSCCP::ID = 0; 1723 INITIALIZE_PASS(IPSCCP, "ipsccp", 1724 "Interprocedural Sparse Conditional Constant Propagation", 1725 false, false) 1726 1727 // createIPSCCPPass - This is the public interface to this file. 1728 ModulePass *llvm::createIPSCCPPass() { 1729 return new IPSCCP(); 1730 } 1731 1732 1733 static bool AddressIsTaken(const GlobalValue *GV) { 1734 // Delete any dead constantexpr klingons. 1735 GV->removeDeadConstantUsers(); 1736 1737 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end(); 1738 UI != E; ++UI) { 1739 const User *U = *UI; 1740 if (const StoreInst *SI = dyn_cast<StoreInst>(U)) { 1741 if (SI->getOperand(0) == GV || SI->isVolatile()) 1742 return true; // Storing addr of GV. 1743 } else if (isa<InvokeInst>(U) || isa<CallInst>(U)) { 1744 // Make sure we are calling the function, not passing the address. 1745 ImmutableCallSite CS(cast<Instruction>(U)); 1746 if (!CS.isCallee(UI)) 1747 return true; 1748 } else if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 1749 if (LI->isVolatile()) 1750 return true; 1751 } else if (isa<BlockAddress>(U)) { 1752 // blockaddress doesn't take the address of the function, it takes addr 1753 // of label. 1754 } else { 1755 return true; 1756 } 1757 } 1758 return false; 1759 } 1760 1761 bool IPSCCP::runOnModule(Module &M) { 1762 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>()); 1763 1764 // AddressTakenFunctions - This set keeps track of the address-taken functions 1765 // that are in the input. As IPSCCP runs through and simplifies code, 1766 // functions that were address taken can end up losing their 1767 // address-taken-ness. Because of this, we keep track of their addresses from 1768 // the first pass so we can use them for the later simplification pass. 1769 SmallPtrSet<Function*, 32> AddressTakenFunctions; 1770 1771 // Loop over all functions, marking arguments to those with their addresses 1772 // taken or that are external as overdefined. 1773 // 1774 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { 1775 if (F->isDeclaration()) 1776 continue; 1777 1778 // If this is a strong or ODR definition of this function, then we can 1779 // propagate information about its result into callsites of it. 1780 if (!F->mayBeOverridden()) 1781 Solver.AddTrackedFunction(F); 1782 1783 // If this function only has direct calls that we can see, we can track its 1784 // arguments and return value aggressively, and can assume it is not called 1785 // unless we see evidence to the contrary. 1786 if (F->hasLocalLinkage()) { 1787 if (AddressIsTaken(F)) 1788 AddressTakenFunctions.insert(F); 1789 else { 1790 Solver.AddArgumentTrackedFunction(F); 1791 continue; 1792 } 1793 } 1794 1795 // Assume the function is called. 1796 Solver.MarkBlockExecutable(F->begin()); 1797 1798 // Assume nothing about the incoming arguments. 1799 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1800 AI != E; ++AI) 1801 Solver.markAnythingOverdefined(AI); 1802 } 1803 1804 // Loop over global variables. We inform the solver about any internal global 1805 // variables that do not have their 'addresses taken'. If they don't have 1806 // their addresses taken, we can propagate constants through them. 1807 for (Module::global_iterator G = M.global_begin(), E = M.global_end(); 1808 G != E; ++G) 1809 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G)) 1810 Solver.TrackValueOfGlobalVariable(G); 1811 1812 // Solve for constants. 1813 bool ResolvedUndefs = true; 1814 while (ResolvedUndefs) { 1815 Solver.Solve(); 1816 1817 DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1818 ResolvedUndefs = false; 1819 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) 1820 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F); 1821 } 1822 1823 bool MadeChanges = false; 1824 1825 // Iterate over all of the instructions in the module, replacing them with 1826 // constants if we have found them to be of constant values. 1827 // 1828 SmallVector<BasicBlock*, 512> BlocksToErase; 1829 1830 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { 1831 if (Solver.isBlockExecutable(F->begin())) { 1832 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1833 AI != E; ++AI) { 1834 if (AI->use_empty() || AI->getType()->isStructTy()) continue; 1835 1836 // TODO: Could use getStructLatticeValueFor to find out if the entire 1837 // result is a constant and replace it entirely if so. 1838 1839 LatticeVal IV = Solver.getLatticeValueFor(AI); 1840 if (IV.isOverdefined()) continue; 1841 1842 Constant *CST = IV.isConstant() ? 1843 IV.getConstant() : UndefValue::get(AI->getType()); 1844 DEBUG(dbgs() << "*** Arg " << *AI << " = " << *CST <<"\n"); 1845 1846 // Replaces all of the uses of a variable with uses of the 1847 // constant. 1848 AI->replaceAllUsesWith(CST); 1849 ++IPNumArgsElimed; 1850 } 1851 } 1852 1853 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) { 1854 if (!Solver.isBlockExecutable(BB)) { 1855 DeleteInstructionInBlock(BB); 1856 MadeChanges = true; 1857 1858 TerminatorInst *TI = BB->getTerminator(); 1859 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1860 BasicBlock *Succ = TI->getSuccessor(i); 1861 if (!Succ->empty() && isa<PHINode>(Succ->begin())) 1862 TI->getSuccessor(i)->removePredecessor(BB); 1863 } 1864 if (!TI->use_empty()) 1865 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 1866 TI->eraseFromParent(); 1867 1868 if (&*BB != &F->front()) 1869 BlocksToErase.push_back(BB); 1870 else 1871 new UnreachableInst(M.getContext(), BB); 1872 continue; 1873 } 1874 1875 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1876 Instruction *Inst = BI++; 1877 if (Inst->getType()->isVoidTy() || Inst->getType()->isStructTy()) 1878 continue; 1879 1880 // TODO: Could use getStructLatticeValueFor to find out if the entire 1881 // result is a constant and replace it entirely if so. 1882 1883 LatticeVal IV = Solver.getLatticeValueFor(Inst); 1884 if (IV.isOverdefined()) 1885 continue; 1886 1887 Constant *Const = IV.isConstant() 1888 ? IV.getConstant() : UndefValue::get(Inst->getType()); 1889 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst); 1890 1891 // Replaces all of the uses of a variable with uses of the 1892 // constant. 1893 Inst->replaceAllUsesWith(Const); 1894 1895 // Delete the instruction. 1896 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst)) 1897 Inst->eraseFromParent(); 1898 1899 // Hey, we just changed something! 1900 MadeChanges = true; 1901 ++IPNumInstRemoved; 1902 } 1903 } 1904 1905 // Now that all instructions in the function are constant folded, erase dead 1906 // blocks, because we can now use ConstantFoldTerminator to get rid of 1907 // in-edges. 1908 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) { 1909 // If there are any PHI nodes in this successor, drop entries for BB now. 1910 BasicBlock *DeadBB = BlocksToErase[i]; 1911 for (Value::use_iterator UI = DeadBB->use_begin(), UE = DeadBB->use_end(); 1912 UI != UE; ) { 1913 // Grab the user and then increment the iterator early, as the user 1914 // will be deleted. Step past all adjacent uses from the same user. 1915 Instruction *I = dyn_cast<Instruction>(*UI); 1916 do { ++UI; } while (UI != UE && *UI == I); 1917 1918 // Ignore blockaddress users; BasicBlock's dtor will handle them. 1919 if (!I) continue; 1920 1921 bool Folded = ConstantFoldTerminator(I->getParent()); 1922 if (!Folded) { 1923 // The constant folder may not have been able to fold the terminator 1924 // if this is a branch or switch on undef. Fold it manually as a 1925 // branch to the first successor. 1926 #ifndef NDEBUG 1927 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1928 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) && 1929 "Branch should be foldable!"); 1930 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1931 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold"); 1932 } else { 1933 llvm_unreachable("Didn't fold away reference to block!"); 1934 } 1935 #endif 1936 1937 // Make this an uncond branch to the first successor. 1938 TerminatorInst *TI = I->getParent()->getTerminator(); 1939 BranchInst::Create(TI->getSuccessor(0), TI); 1940 1941 // Remove entries in successor phi nodes to remove edges. 1942 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i) 1943 TI->getSuccessor(i)->removePredecessor(TI->getParent()); 1944 1945 // Remove the old terminator. 1946 TI->eraseFromParent(); 1947 } 1948 } 1949 1950 // Finally, delete the basic block. 1951 F->getBasicBlockList().erase(DeadBB); 1952 } 1953 BlocksToErase.clear(); 1954 } 1955 1956 // If we inferred constant or undef return values for a function, we replaced 1957 // all call uses with the inferred value. This means we don't need to bother 1958 // actually returning anything from the function. Replace all return 1959 // instructions with return undef. 1960 // 1961 // Do this in two stages: first identify the functions we should process, then 1962 // actually zap their returns. This is important because we can only do this 1963 // if the address of the function isn't taken. In cases where a return is the 1964 // last use of a function, the order of processing functions would affect 1965 // whether other functions are optimizable. 1966 SmallVector<ReturnInst*, 8> ReturnsToZap; 1967 1968 // TODO: Process multiple value ret instructions also. 1969 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals(); 1970 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(), 1971 E = RV.end(); I != E; ++I) { 1972 Function *F = I->first; 1973 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy()) 1974 continue; 1975 1976 // We can only do this if we know that nothing else can call the function. 1977 if (!F->hasLocalLinkage() || AddressTakenFunctions.count(F)) 1978 continue; 1979 1980 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 1981 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 1982 if (!isa<UndefValue>(RI->getOperand(0))) 1983 ReturnsToZap.push_back(RI); 1984 } 1985 1986 // Zap all returns which we've identified as zap to change. 1987 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 1988 Function *F = ReturnsToZap[i]->getParent()->getParent(); 1989 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 1990 } 1991 1992 // If we inferred constant or undef values for globals variables, we can delete 1993 // the global and any stores that remain to it. 1994 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals(); 1995 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(), 1996 E = TG.end(); I != E; ++I) { 1997 GlobalVariable *GV = I->first; 1998 assert(!I->second.isOverdefined() && 1999 "Overdefined values should have been taken out of the map!"); 2000 DEBUG(dbgs() << "Found that GV '" << GV->getName() << "' is constant!\n"); 2001 while (!GV->use_empty()) { 2002 StoreInst *SI = cast<StoreInst>(GV->use_back()); 2003 SI->eraseFromParent(); 2004 } 2005 M.getGlobalList().erase(GV); 2006 ++IPNumGlobalConst; 2007 } 2008 2009 return MadeChanges; 2010 } 2011