SCCP.cpp revision 54a525d7ae4ce36780d840cfec075d4080f0764a
1//===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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// Notice that: 19// * This pass has a habit of making definitions be dead. It is a good idea 20// to to run a DCE pass sometime after running this pass. 21// 22//===----------------------------------------------------------------------===// 23 24#define DEBUG_TYPE "sccp" 25#include "llvm/Transforms/Scalar.h" 26#include "llvm/Transforms/IPO.h" 27#include "llvm/Constants.h" 28#include "llvm/DerivedTypes.h" 29#include "llvm/Instructions.h" 30#include "llvm/Pass.h" 31#include "llvm/Support/InstVisitor.h" 32#include "llvm/Transforms/Utils/Local.h" 33#include "llvm/Support/CallSite.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/ADT/hash_map" 36#include "llvm/ADT/Statistic.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 39#include <set> 40using namespace llvm; 41 42STATISTIC(NumInstRemoved, "Number of instructions removed"); 43STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 44 45STATISTIC(IPNumInstRemoved, "Number ofinstructions removed by IPSCCP"); 46STATISTIC(IPNumDeadBlocks , "Number of basic blocks unreachable by IPSCCP"); 47STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 48STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 49 50namespace { 51/// LatticeVal class - This class represents the different lattice values that 52/// an LLVM value may occupy. It is a simple class with value semantics. 53/// 54class LatticeVal { 55 enum { 56 /// undefined - This LLVM Value has no known value yet. 57 undefined, 58 59 /// constant - This LLVM Value has a specific constant value. 60 constant, 61 62 /// forcedconstant - This LLVM Value was thought to be undef until 63 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged 64 /// with another (different) constant, it goes to overdefined, instead of 65 /// asserting. 66 forcedconstant, 67 68 /// overdefined - This instruction is not known to be constant, and we know 69 /// it has a value. 70 overdefined 71 } LatticeValue; // The current lattice position 72 73 Constant *ConstantVal; // If Constant value, the current value 74public: 75 inline LatticeVal() : LatticeValue(undefined), ConstantVal(0) {} 76 77 // markOverdefined - Return true if this is a new status to be in... 78 inline bool markOverdefined() { 79 if (LatticeValue != overdefined) { 80 LatticeValue = overdefined; 81 return true; 82 } 83 return false; 84 } 85 86 // markConstant - Return true if this is a new status for us. 87 inline bool markConstant(Constant *V) { 88 if (LatticeValue != constant) { 89 if (LatticeValue == undefined) { 90 LatticeValue = constant; 91 assert(V && "Marking constant with NULL"); 92 ConstantVal = V; 93 } else { 94 assert(LatticeValue == forcedconstant && 95 "Cannot move from overdefined to constant!"); 96 // Stay at forcedconstant if the constant is the same. 97 if (V == ConstantVal) return false; 98 99 // Otherwise, we go to overdefined. Assumptions made based on the 100 // forced value are possibly wrong. Assuming this is another constant 101 // could expose a contradiction. 102 LatticeValue = overdefined; 103 } 104 return true; 105 } else { 106 assert(ConstantVal == V && "Marking constant with different value"); 107 } 108 return false; 109 } 110 111 inline void markForcedConstant(Constant *V) { 112 assert(LatticeValue == undefined && "Can't force a defined value!"); 113 LatticeValue = forcedconstant; 114 ConstantVal = V; 115 } 116 117 inline bool isUndefined() const { return LatticeValue == undefined; } 118 inline bool isConstant() const { 119 return LatticeValue == constant || LatticeValue == forcedconstant; 120 } 121 inline bool isOverdefined() const { return LatticeValue == overdefined; } 122 123 inline Constant *getConstant() const { 124 assert(isConstant() && "Cannot get the constant of a non-constant!"); 125 return ConstantVal; 126 } 127}; 128 129} // end anonymous namespace 130 131 132//===----------------------------------------------------------------------===// 133// 134/// SCCPSolver - This class is a general purpose solver for Sparse Conditional 135/// Constant Propagation. 136/// 137class SCCPSolver : public InstVisitor<SCCPSolver> { 138 std::set<BasicBlock*> BBExecutable;// The basic blocks that are executable 139 hash_map<Value*, LatticeVal> ValueState; // The state each value is in... 140 141 /// GlobalValue - If we are tracking any values for the contents of a global 142 /// variable, we keep a mapping from the constant accessor to the element of 143 /// the global, to the currently known value. If the value becomes 144 /// overdefined, it's entry is simply removed from this map. 145 hash_map<GlobalVariable*, LatticeVal> TrackedGlobals; 146 147 /// TrackedFunctionRetVals - If we are tracking arguments into and the return 148 /// value out of a function, it will have an entry in this map, indicating 149 /// what the known return value for the function is. 150 hash_map<Function*, LatticeVal> TrackedFunctionRetVals; 151 152 // The reason for two worklists is that overdefined is the lowest state 153 // on the lattice, and moving things to overdefined as fast as possible 154 // makes SCCP converge much faster. 155 // By having a separate worklist, we accomplish this because everything 156 // possibly overdefined will become overdefined at the soonest possible 157 // point. 158 std::vector<Value*> OverdefinedInstWorkList; 159 std::vector<Value*> InstWorkList; 160 161 162 std::vector<BasicBlock*> BBWorkList; // The BasicBlock work list 163 164 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not 165 /// overdefined, despite the fact that the PHI node is overdefined. 166 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs; 167 168 /// KnownFeasibleEdges - Entries in this set are edges which have already had 169 /// PHI nodes retriggered. 170 typedef std::pair<BasicBlock*,BasicBlock*> Edge; 171 std::set<Edge> KnownFeasibleEdges; 172public: 173 174 /// MarkBlockExecutable - This method can be used by clients to mark all of 175 /// the blocks that are known to be intrinsically live in the processed unit. 176 void MarkBlockExecutable(BasicBlock *BB) { 177 DOUT << "Marking Block Executable: " << BB->getName() << "\n"; 178 BBExecutable.insert(BB); // Basic block is executable! 179 BBWorkList.push_back(BB); // Add the block to the work list! 180 } 181 182 /// TrackValueOfGlobalVariable - Clients can use this method to 183 /// inform the SCCPSolver that it should track loads and stores to the 184 /// specified global variable if it can. This is only legal to call if 185 /// performing Interprocedural SCCP. 186 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 187 const Type *ElTy = GV->getType()->getElementType(); 188 if (ElTy->isFirstClassType()) { 189 LatticeVal &IV = TrackedGlobals[GV]; 190 if (!isa<UndefValue>(GV->getInitializer())) 191 IV.markConstant(GV->getInitializer()); 192 } 193 } 194 195 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 196 /// and out of the specified function (which cannot have its address taken), 197 /// this method must be called. 198 void AddTrackedFunction(Function *F) { 199 assert(F->hasInternalLinkage() && "Can only track internal functions!"); 200 // Add an entry, F -> undef. 201 TrackedFunctionRetVals[F]; 202 } 203 204 /// Solve - Solve for constants and executable blocks. 205 /// 206 void Solve(); 207 208 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 209 /// that branches on undef values cannot reach any of their successors. 210 /// However, this is not a safe assumption. After we solve dataflow, this 211 /// method should be use to handle this. If this returns true, the solver 212 /// should be rerun. 213 bool ResolvedUndefsIn(Function &F); 214 215 /// getExecutableBlocks - Once we have solved for constants, return the set of 216 /// blocks that is known to be executable. 217 std::set<BasicBlock*> &getExecutableBlocks() { 218 return BBExecutable; 219 } 220 221 /// getValueMapping - Once we have solved for constants, return the mapping of 222 /// LLVM values to LatticeVals. 223 hash_map<Value*, LatticeVal> &getValueMapping() { 224 return ValueState; 225 } 226 227 /// getTrackedFunctionRetVals - Get the inferred return value map. 228 /// 229 const hash_map<Function*, LatticeVal> &getTrackedFunctionRetVals() { 230 return TrackedFunctionRetVals; 231 } 232 233 /// getTrackedGlobals - Get and return the set of inferred initializers for 234 /// global variables. 235 const hash_map<GlobalVariable*, LatticeVal> &getTrackedGlobals() { 236 return TrackedGlobals; 237 } 238 239 240private: 241 // markConstant - Make a value be marked as "constant". If the value 242 // is not already a constant, add it to the instruction work list so that 243 // the users of the instruction are updated later. 244 // 245 inline void markConstant(LatticeVal &IV, Value *V, Constant *C) { 246 if (IV.markConstant(C)) { 247 DOUT << "markConstant: " << *C << ": " << *V; 248 InstWorkList.push_back(V); 249 } 250 } 251 252 inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) { 253 IV.markForcedConstant(C); 254 DOUT << "markForcedConstant: " << *C << ": " << *V; 255 InstWorkList.push_back(V); 256 } 257 258 inline void markConstant(Value *V, Constant *C) { 259 markConstant(ValueState[V], V, C); 260 } 261 262 // markOverdefined - Make a value be marked as "overdefined". If the 263 // value is not already overdefined, add it to the overdefined instruction 264 // work list so that the users of the instruction are updated later. 265 266 inline void markOverdefined(LatticeVal &IV, Value *V) { 267 if (IV.markOverdefined()) { 268 DEBUG(DOUT << "markOverdefined: "; 269 if (Function *F = dyn_cast<Function>(V)) 270 DOUT << "Function '" << F->getName() << "'\n"; 271 else 272 DOUT << *V); 273 // Only instructions go on the work list 274 OverdefinedInstWorkList.push_back(V); 275 } 276 } 277 inline void markOverdefined(Value *V) { 278 markOverdefined(ValueState[V], V); 279 } 280 281 inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) { 282 if (IV.isOverdefined() || MergeWithV.isUndefined()) 283 return; // Noop. 284 if (MergeWithV.isOverdefined()) 285 markOverdefined(IV, V); 286 else if (IV.isUndefined()) 287 markConstant(IV, V, MergeWithV.getConstant()); 288 else if (IV.getConstant() != MergeWithV.getConstant()) 289 markOverdefined(IV, V); 290 } 291 292 inline void mergeInValue(Value *V, LatticeVal &MergeWithV) { 293 return mergeInValue(ValueState[V], V, MergeWithV); 294 } 295 296 297 // getValueState - Return the LatticeVal object that corresponds to the value. 298 // This function is necessary because not all values should start out in the 299 // underdefined state... Argument's should be overdefined, and 300 // constants should be marked as constants. If a value is not known to be an 301 // Instruction object, then use this accessor to get its value from the map. 302 // 303 inline LatticeVal &getValueState(Value *V) { 304 hash_map<Value*, LatticeVal>::iterator I = ValueState.find(V); 305 if (I != ValueState.end()) return I->second; // Common case, in the map 306 307 if (Constant *C = dyn_cast<Constant>(V)) { 308 if (isa<UndefValue>(V)) { 309 // Nothing to do, remain undefined. 310 } else { 311 ValueState[C].markConstant(C); // Constants are constant 312 } 313 } 314 // All others are underdefined by default... 315 return ValueState[V]; 316 } 317 318 // markEdgeExecutable - Mark a basic block as executable, adding it to the BB 319 // work list if it is not already executable... 320 // 321 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 322 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 323 return; // This edge is already known to be executable! 324 325 if (BBExecutable.count(Dest)) { 326 DOUT << "Marking Edge Executable: " << Source->getName() 327 << " -> " << Dest->getName() << "\n"; 328 329 // The destination is already executable, but we just made an edge 330 // feasible that wasn't before. Revisit the PHI nodes in the block 331 // because they have potentially new operands. 332 for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I) 333 visitPHINode(*cast<PHINode>(I)); 334 335 } else { 336 MarkBlockExecutable(Dest); 337 } 338 } 339 340 // getFeasibleSuccessors - Return a vector of booleans to indicate which 341 // successors are reachable from a given terminator instruction. 342 // 343 void getFeasibleSuccessors(TerminatorInst &TI, std::vector<bool> &Succs); 344 345 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 346 // block to the 'To' basic block is currently feasible... 347 // 348 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 349 350 // OperandChangedState - This method is invoked on all of the users of an 351 // instruction that was just changed state somehow.... Based on this 352 // information, we need to update the specified user of this instruction. 353 // 354 void OperandChangedState(User *U) { 355 // Only instructions use other variable values! 356 Instruction &I = cast<Instruction>(*U); 357 if (BBExecutable.count(I.getParent())) // Inst is executable? 358 visit(I); 359 } 360 361private: 362 friend class InstVisitor<SCCPSolver>; 363 364 // visit implementations - Something changed in this instruction... Either an 365 // operand made a transition, or the instruction is newly executable. Change 366 // the value type of I to reflect these changes if appropriate. 367 // 368 void visitPHINode(PHINode &I); 369 370 // Terminators 371 void visitReturnInst(ReturnInst &I); 372 void visitTerminatorInst(TerminatorInst &TI); 373 374 void visitCastInst(CastInst &I); 375 void visitSelectInst(SelectInst &I); 376 void visitBinaryOperator(Instruction &I); 377 void visitCmpInst(CmpInst &I); 378 void visitShiftInst(ShiftInst &I) { visitBinaryOperator(I); } 379 void visitExtractElementInst(ExtractElementInst &I); 380 void visitInsertElementInst(InsertElementInst &I); 381 void visitShuffleVectorInst(ShuffleVectorInst &I); 382 383 // Instructions that cannot be folded away... 384 void visitStoreInst (Instruction &I); 385 void visitLoadInst (LoadInst &I); 386 void visitGetElementPtrInst(GetElementPtrInst &I); 387 void visitCallInst (CallInst &I) { visitCallSite(CallSite::get(&I)); } 388 void visitInvokeInst (InvokeInst &II) { 389 visitCallSite(CallSite::get(&II)); 390 visitTerminatorInst(II); 391 } 392 void visitCallSite (CallSite CS); 393 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ } 394 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ } 395 void visitAllocationInst(Instruction &I) { markOverdefined(&I); } 396 void visitVANextInst (Instruction &I) { markOverdefined(&I); } 397 void visitVAArgInst (Instruction &I) { markOverdefined(&I); } 398 void visitFreeInst (Instruction &I) { /*returns void*/ } 399 400 void visitInstruction(Instruction &I) { 401 // If a new instruction is added to LLVM that we don't handle... 402 cerr << "SCCP: Don't know how to handle: " << I; 403 markOverdefined(&I); // Just in case 404 } 405}; 406 407// getFeasibleSuccessors - Return a vector of booleans to indicate which 408// successors are reachable from a given terminator instruction. 409// 410void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI, 411 std::vector<bool> &Succs) { 412 Succs.resize(TI.getNumSuccessors()); 413 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) { 414 if (BI->isUnconditional()) { 415 Succs[0] = true; 416 } else { 417 LatticeVal &BCValue = getValueState(BI->getCondition()); 418 if (BCValue.isOverdefined() || 419 (BCValue.isConstant() && !isa<ConstantInt>(BCValue.getConstant()))) { 420 // Overdefined condition variables, and branches on unfoldable constant 421 // conditions, mean the branch could go either way. 422 Succs[0] = Succs[1] = true; 423 } else if (BCValue.isConstant()) { 424 // Constant condition variables mean the branch can only go a single way 425 Succs[BCValue.getConstant() == ConstantInt::getFalse()] = true; 426 } 427 } 428 } else if (isa<InvokeInst>(&TI)) { 429 // Invoke instructions successors are always executable. 430 Succs[0] = Succs[1] = true; 431 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) { 432 LatticeVal &SCValue = getValueState(SI->getCondition()); 433 if (SCValue.isOverdefined() || // Overdefined condition? 434 (SCValue.isConstant() && !isa<ConstantInt>(SCValue.getConstant()))) { 435 // All destinations are executable! 436 Succs.assign(TI.getNumSuccessors(), true); 437 } else if (SCValue.isConstant()) { 438 Constant *CPV = SCValue.getConstant(); 439 // Make sure to skip the "default value" which isn't a value 440 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i) { 441 if (SI->getSuccessorValue(i) == CPV) {// Found the right branch... 442 Succs[i] = true; 443 return; 444 } 445 } 446 447 // Constant value not equal to any of the branches... must execute 448 // default branch then... 449 Succs[0] = true; 450 } 451 } else { 452 cerr << "SCCP: Don't know how to handle: " << TI; 453 Succs.assign(TI.getNumSuccessors(), true); 454 } 455} 456 457 458// isEdgeFeasible - Return true if the control flow edge from the 'From' basic 459// block to the 'To' basic block is currently feasible... 460// 461bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 462 assert(BBExecutable.count(To) && "Dest should always be alive!"); 463 464 // Make sure the source basic block is executable!! 465 if (!BBExecutable.count(From)) return false; 466 467 // Check to make sure this edge itself is actually feasible now... 468 TerminatorInst *TI = From->getTerminator(); 469 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 470 if (BI->isUnconditional()) 471 return true; 472 else { 473 LatticeVal &BCValue = getValueState(BI->getCondition()); 474 if (BCValue.isOverdefined()) { 475 // Overdefined condition variables mean the branch could go either way. 476 return true; 477 } else if (BCValue.isConstant()) { 478 // Not branching on an evaluatable constant? 479 if (!isa<ConstantInt>(BCValue.getConstant())) return true; 480 481 // Constant condition variables mean the branch can only go a single way 482 return BI->getSuccessor(BCValue.getConstant() == 483 ConstantInt::getFalse()) == To; 484 } 485 return false; 486 } 487 } else if (isa<InvokeInst>(TI)) { 488 // Invoke instructions successors are always executable. 489 return true; 490 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 491 LatticeVal &SCValue = getValueState(SI->getCondition()); 492 if (SCValue.isOverdefined()) { // Overdefined condition? 493 // All destinations are executable! 494 return true; 495 } else if (SCValue.isConstant()) { 496 Constant *CPV = SCValue.getConstant(); 497 if (!isa<ConstantInt>(CPV)) 498 return true; // not a foldable constant? 499 500 // Make sure to skip the "default value" which isn't a value 501 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i) 502 if (SI->getSuccessorValue(i) == CPV) // Found the taken branch... 503 return SI->getSuccessor(i) == To; 504 505 // Constant value not equal to any of the branches... must execute 506 // default branch then... 507 return SI->getDefaultDest() == To; 508 } 509 return false; 510 } else { 511 cerr << "Unknown terminator instruction: " << *TI; 512 abort(); 513 } 514} 515 516// visit Implementations - Something changed in this instruction... Either an 517// operand made a transition, or the instruction is newly executable. Change 518// the value type of I to reflect these changes if appropriate. This method 519// makes sure to do the following actions: 520// 521// 1. If a phi node merges two constants in, and has conflicting value coming 522// from different branches, or if the PHI node merges in an overdefined 523// value, then the PHI node becomes overdefined. 524// 2. If a phi node merges only constants in, and they all agree on value, the 525// PHI node becomes a constant value equal to that. 526// 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 527// 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 528// 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 529// 6. If a conditional branch has a value that is constant, make the selected 530// destination executable 531// 7. If a conditional branch has a value that is overdefined, make all 532// successors executable. 533// 534void SCCPSolver::visitPHINode(PHINode &PN) { 535 LatticeVal &PNIV = getValueState(&PN); 536 if (PNIV.isOverdefined()) { 537 // There may be instructions using this PHI node that are not overdefined 538 // themselves. If so, make sure that they know that the PHI node operand 539 // changed. 540 std::multimap<PHINode*, Instruction*>::iterator I, E; 541 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN); 542 if (I != E) { 543 std::vector<Instruction*> Users; 544 Users.reserve(std::distance(I, E)); 545 for (; I != E; ++I) Users.push_back(I->second); 546 while (!Users.empty()) { 547 visit(Users.back()); 548 Users.pop_back(); 549 } 550 } 551 return; // Quick exit 552 } 553 554 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 555 // and slow us down a lot. Just mark them overdefined. 556 if (PN.getNumIncomingValues() > 64) { 557 markOverdefined(PNIV, &PN); 558 return; 559 } 560 561 // Look at all of the executable operands of the PHI node. If any of them 562 // are overdefined, the PHI becomes overdefined as well. If they are all 563 // constant, and they agree with each other, the PHI becomes the identical 564 // constant. If they are constant and don't agree, the PHI is overdefined. 565 // If there are no executable operands, the PHI remains undefined. 566 // 567 Constant *OperandVal = 0; 568 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 569 LatticeVal &IV = getValueState(PN.getIncomingValue(i)); 570 if (IV.isUndefined()) continue; // Doesn't influence PHI node. 571 572 if (isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) { 573 if (IV.isOverdefined()) { // PHI node becomes overdefined! 574 markOverdefined(PNIV, &PN); 575 return; 576 } 577 578 if (OperandVal == 0) { // Grab the first value... 579 OperandVal = IV.getConstant(); 580 } else { // Another value is being merged in! 581 // There is already a reachable operand. If we conflict with it, 582 // then the PHI node becomes overdefined. If we agree with it, we 583 // can continue on. 584 585 // Check to see if there are two different constants merging... 586 if (IV.getConstant() != OperandVal) { 587 // Yes there is. This means the PHI node is not constant. 588 // You must be overdefined poor PHI. 589 // 590 markOverdefined(PNIV, &PN); // The PHI node now becomes overdefined 591 return; // I'm done analyzing you 592 } 593 } 594 } 595 } 596 597 // If we exited the loop, this means that the PHI node only has constant 598 // arguments that agree with each other(and OperandVal is the constant) or 599 // OperandVal is null because there are no defined incoming arguments. If 600 // this is the case, the PHI remains undefined. 601 // 602 if (OperandVal) 603 markConstant(PNIV, &PN, OperandVal); // Acquire operand value 604} 605 606void SCCPSolver::visitReturnInst(ReturnInst &I) { 607 if (I.getNumOperands() == 0) return; // Ret void 608 609 // If we are tracking the return value of this function, merge it in. 610 Function *F = I.getParent()->getParent(); 611 if (F->hasInternalLinkage() && !TrackedFunctionRetVals.empty()) { 612 hash_map<Function*, LatticeVal>::iterator TFRVI = 613 TrackedFunctionRetVals.find(F); 614 if (TFRVI != TrackedFunctionRetVals.end() && 615 !TFRVI->second.isOverdefined()) { 616 LatticeVal &IV = getValueState(I.getOperand(0)); 617 mergeInValue(TFRVI->second, F, IV); 618 } 619 } 620} 621 622 623void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) { 624 std::vector<bool> SuccFeasible; 625 getFeasibleSuccessors(TI, SuccFeasible); 626 627 BasicBlock *BB = TI.getParent(); 628 629 // Mark all feasible successors executable... 630 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 631 if (SuccFeasible[i]) 632 markEdgeExecutable(BB, TI.getSuccessor(i)); 633} 634 635void SCCPSolver::visitCastInst(CastInst &I) { 636 Value *V = I.getOperand(0); 637 LatticeVal &VState = getValueState(V); 638 if (VState.isOverdefined()) // Inherit overdefinedness of operand 639 markOverdefined(&I); 640 else if (VState.isConstant()) // Propagate constant value 641 markConstant(&I, ConstantExpr::getCast(I.getOpcode(), 642 VState.getConstant(), I.getType())); 643} 644 645void SCCPSolver::visitSelectInst(SelectInst &I) { 646 LatticeVal &CondValue = getValueState(I.getCondition()); 647 if (CondValue.isUndefined()) 648 return; 649 if (CondValue.isConstant()) { 650 if (ConstantInt *CondCB = dyn_cast<ConstantInt>(CondValue.getConstant())){ 651 mergeInValue(&I, getValueState(CondCB->getZExtValue() ? I.getTrueValue() 652 : I.getFalseValue())); 653 return; 654 } 655 } 656 657 // Otherwise, the condition is overdefined or a constant we can't evaluate. 658 // See if we can produce something better than overdefined based on the T/F 659 // value. 660 LatticeVal &TVal = getValueState(I.getTrueValue()); 661 LatticeVal &FVal = getValueState(I.getFalseValue()); 662 663 // select ?, C, C -> C. 664 if (TVal.isConstant() && FVal.isConstant() && 665 TVal.getConstant() == FVal.getConstant()) { 666 markConstant(&I, FVal.getConstant()); 667 return; 668 } 669 670 if (TVal.isUndefined()) { // select ?, undef, X -> X. 671 mergeInValue(&I, FVal); 672 } else if (FVal.isUndefined()) { // select ?, X, undef -> X. 673 mergeInValue(&I, TVal); 674 } else { 675 markOverdefined(&I); 676 } 677} 678 679// Handle BinaryOperators and Shift Instructions... 680void SCCPSolver::visitBinaryOperator(Instruction &I) { 681 LatticeVal &IV = ValueState[&I]; 682 if (IV.isOverdefined()) return; 683 684 LatticeVal &V1State = getValueState(I.getOperand(0)); 685 LatticeVal &V2State = getValueState(I.getOperand(1)); 686 687 if (V1State.isOverdefined() || V2State.isOverdefined()) { 688 // If this is an AND or OR with 0 or -1, it doesn't matter that the other 689 // operand is overdefined. 690 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) { 691 LatticeVal *NonOverdefVal = 0; 692 if (!V1State.isOverdefined()) { 693 NonOverdefVal = &V1State; 694 } else if (!V2State.isOverdefined()) { 695 NonOverdefVal = &V2State; 696 } 697 698 if (NonOverdefVal) { 699 if (NonOverdefVal->isUndefined()) { 700 // Could annihilate value. 701 if (I.getOpcode() == Instruction::And) 702 markConstant(IV, &I, Constant::getNullValue(I.getType())); 703 else if (const PackedType *PT = dyn_cast<PackedType>(I.getType())) 704 markConstant(IV, &I, ConstantPacked::getAllOnesValue(PT)); 705 else 706 markConstant(IV, &I, ConstantInt::getAllOnesValue(I.getType())); 707 return; 708 } else { 709 if (I.getOpcode() == Instruction::And) { 710 if (NonOverdefVal->getConstant()->isNullValue()) { 711 markConstant(IV, &I, NonOverdefVal->getConstant()); 712 return; // X and 0 = 0 713 } 714 } else { 715 if (ConstantInt *CI = 716 dyn_cast<ConstantInt>(NonOverdefVal->getConstant())) 717 if (CI->isAllOnesValue()) { 718 markConstant(IV, &I, NonOverdefVal->getConstant()); 719 return; // X or -1 = -1 720 } 721 } 722 } 723 } 724 } 725 726 727 // If both operands are PHI nodes, it is possible that this instruction has 728 // a constant value, despite the fact that the PHI node doesn't. Check for 729 // this condition now. 730 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0))) 731 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1))) 732 if (PN1->getParent() == PN2->getParent()) { 733 // Since the two PHI nodes are in the same basic block, they must have 734 // entries for the same predecessors. Walk the predecessor list, and 735 // if all of the incoming values are constants, and the result of 736 // evaluating this expression with all incoming value pairs is the 737 // same, then this expression is a constant even though the PHI node 738 // is not a constant! 739 LatticeVal Result; 740 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) { 741 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i)); 742 BasicBlock *InBlock = PN1->getIncomingBlock(i); 743 LatticeVal &In2 = 744 getValueState(PN2->getIncomingValueForBlock(InBlock)); 745 746 if (In1.isOverdefined() || In2.isOverdefined()) { 747 Result.markOverdefined(); 748 break; // Cannot fold this operation over the PHI nodes! 749 } else if (In1.isConstant() && In2.isConstant()) { 750 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(), 751 In2.getConstant()); 752 if (Result.isUndefined()) 753 Result.markConstant(V); 754 else if (Result.isConstant() && Result.getConstant() != V) { 755 Result.markOverdefined(); 756 break; 757 } 758 } 759 } 760 761 // If we found a constant value here, then we know the instruction is 762 // constant despite the fact that the PHI nodes are overdefined. 763 if (Result.isConstant()) { 764 markConstant(IV, &I, Result.getConstant()); 765 // Remember that this instruction is virtually using the PHI node 766 // operands. 767 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I)); 768 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I)); 769 return; 770 } else if (Result.isUndefined()) { 771 return; 772 } 773 774 // Okay, this really is overdefined now. Since we might have 775 // speculatively thought that this was not overdefined before, and 776 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs, 777 // make sure to clean out any entries that we put there, for 778 // efficiency. 779 std::multimap<PHINode*, Instruction*>::iterator It, E; 780 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1); 781 while (It != E) { 782 if (It->second == &I) { 783 UsersOfOverdefinedPHIs.erase(It++); 784 } else 785 ++It; 786 } 787 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2); 788 while (It != E) { 789 if (It->second == &I) { 790 UsersOfOverdefinedPHIs.erase(It++); 791 } else 792 ++It; 793 } 794 } 795 796 markOverdefined(IV, &I); 797 } else if (V1State.isConstant() && V2State.isConstant()) { 798 markConstant(IV, &I, ConstantExpr::get(I.getOpcode(), V1State.getConstant(), 799 V2State.getConstant())); 800 } 801} 802 803// Handle ICmpInst instruction... 804void SCCPSolver::visitCmpInst(CmpInst &I) { 805 LatticeVal &IV = ValueState[&I]; 806 if (IV.isOverdefined()) return; 807 808 LatticeVal &V1State = getValueState(I.getOperand(0)); 809 LatticeVal &V2State = getValueState(I.getOperand(1)); 810 811 if (V1State.isOverdefined() || V2State.isOverdefined()) { 812 // If both operands are PHI nodes, it is possible that this instruction has 813 // a constant value, despite the fact that the PHI node doesn't. Check for 814 // this condition now. 815 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0))) 816 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1))) 817 if (PN1->getParent() == PN2->getParent()) { 818 // Since the two PHI nodes are in the same basic block, they must have 819 // entries for the same predecessors. Walk the predecessor list, and 820 // if all of the incoming values are constants, and the result of 821 // evaluating this expression with all incoming value pairs is the 822 // same, then this expression is a constant even though the PHI node 823 // is not a constant! 824 LatticeVal Result; 825 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) { 826 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i)); 827 BasicBlock *InBlock = PN1->getIncomingBlock(i); 828 LatticeVal &In2 = 829 getValueState(PN2->getIncomingValueForBlock(InBlock)); 830 831 if (In1.isOverdefined() || In2.isOverdefined()) { 832 Result.markOverdefined(); 833 break; // Cannot fold this operation over the PHI nodes! 834 } else if (In1.isConstant() && In2.isConstant()) { 835 Constant *V = ConstantExpr::getCompare(I.getPredicate(), 836 In1.getConstant(), 837 In2.getConstant()); 838 if (Result.isUndefined()) 839 Result.markConstant(V); 840 else if (Result.isConstant() && Result.getConstant() != V) { 841 Result.markOverdefined(); 842 break; 843 } 844 } 845 } 846 847 // If we found a constant value here, then we know the instruction is 848 // constant despite the fact that the PHI nodes are overdefined. 849 if (Result.isConstant()) { 850 markConstant(IV, &I, Result.getConstant()); 851 // Remember that this instruction is virtually using the PHI node 852 // operands. 853 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I)); 854 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I)); 855 return; 856 } else if (Result.isUndefined()) { 857 return; 858 } 859 860 // Okay, this really is overdefined now. Since we might have 861 // speculatively thought that this was not overdefined before, and 862 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs, 863 // make sure to clean out any entries that we put there, for 864 // efficiency. 865 std::multimap<PHINode*, Instruction*>::iterator It, E; 866 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1); 867 while (It != E) { 868 if (It->second == &I) { 869 UsersOfOverdefinedPHIs.erase(It++); 870 } else 871 ++It; 872 } 873 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2); 874 while (It != E) { 875 if (It->second == &I) { 876 UsersOfOverdefinedPHIs.erase(It++); 877 } else 878 ++It; 879 } 880 } 881 882 markOverdefined(IV, &I); 883 } else if (V1State.isConstant() && V2State.isConstant()) { 884 markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(), 885 V1State.getConstant(), 886 V2State.getConstant())); 887 } 888} 889 890void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) { 891 // FIXME : SCCP does not handle vectors properly. 892 markOverdefined(&I); 893 return; 894 895#if 0 896 LatticeVal &ValState = getValueState(I.getOperand(0)); 897 LatticeVal &IdxState = getValueState(I.getOperand(1)); 898 899 if (ValState.isOverdefined() || IdxState.isOverdefined()) 900 markOverdefined(&I); 901 else if(ValState.isConstant() && IdxState.isConstant()) 902 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(), 903 IdxState.getConstant())); 904#endif 905} 906 907void SCCPSolver::visitInsertElementInst(InsertElementInst &I) { 908 // FIXME : SCCP does not handle vectors properly. 909 markOverdefined(&I); 910 return; 911#if 0 912 LatticeVal &ValState = getValueState(I.getOperand(0)); 913 LatticeVal &EltState = getValueState(I.getOperand(1)); 914 LatticeVal &IdxState = getValueState(I.getOperand(2)); 915 916 if (ValState.isOverdefined() || EltState.isOverdefined() || 917 IdxState.isOverdefined()) 918 markOverdefined(&I); 919 else if(ValState.isConstant() && EltState.isConstant() && 920 IdxState.isConstant()) 921 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(), 922 EltState.getConstant(), 923 IdxState.getConstant())); 924 else if (ValState.isUndefined() && EltState.isConstant() && 925 IdxState.isConstant()) 926 markConstant(&I, ConstantExpr::getInsertElement(UndefValue::get(I.getType()), 927 EltState.getConstant(), 928 IdxState.getConstant())); 929#endif 930} 931 932void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) { 933 // FIXME : SCCP does not handle vectors properly. 934 markOverdefined(&I); 935 return; 936#if 0 937 LatticeVal &V1State = getValueState(I.getOperand(0)); 938 LatticeVal &V2State = getValueState(I.getOperand(1)); 939 LatticeVal &MaskState = getValueState(I.getOperand(2)); 940 941 if (MaskState.isUndefined() || 942 (V1State.isUndefined() && V2State.isUndefined())) 943 return; // Undefined output if mask or both inputs undefined. 944 945 if (V1State.isOverdefined() || V2State.isOverdefined() || 946 MaskState.isOverdefined()) { 947 markOverdefined(&I); 948 } else { 949 // A mix of constant/undef inputs. 950 Constant *V1 = V1State.isConstant() ? 951 V1State.getConstant() : UndefValue::get(I.getType()); 952 Constant *V2 = V2State.isConstant() ? 953 V2State.getConstant() : UndefValue::get(I.getType()); 954 Constant *Mask = MaskState.isConstant() ? 955 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType()); 956 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask)); 957 } 958#endif 959} 960 961// Handle getelementptr instructions... if all operands are constants then we 962// can turn this into a getelementptr ConstantExpr. 963// 964void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 965 LatticeVal &IV = ValueState[&I]; 966 if (IV.isOverdefined()) return; 967 968 std::vector<Constant*> Operands; 969 Operands.reserve(I.getNumOperands()); 970 971 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 972 LatticeVal &State = getValueState(I.getOperand(i)); 973 if (State.isUndefined()) 974 return; // Operands are not resolved yet... 975 else if (State.isOverdefined()) { 976 markOverdefined(IV, &I); 977 return; 978 } 979 assert(State.isConstant() && "Unknown state!"); 980 Operands.push_back(State.getConstant()); 981 } 982 983 Constant *Ptr = Operands[0]; 984 Operands.erase(Operands.begin()); // Erase the pointer from idx list... 985 986 markConstant(IV, &I, ConstantExpr::getGetElementPtr(Ptr, Operands)); 987} 988 989void SCCPSolver::visitStoreInst(Instruction &SI) { 990 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 991 return; 992 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 993 hash_map<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV); 994 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return; 995 996 // Get the value we are storing into the global. 997 LatticeVal &PtrVal = getValueState(SI.getOperand(0)); 998 999 mergeInValue(I->second, GV, PtrVal); 1000 if (I->second.isOverdefined()) 1001 TrackedGlobals.erase(I); // No need to keep tracking this! 1002} 1003 1004 1005// Handle load instructions. If the operand is a constant pointer to a constant 1006// global, we can replace the load with the loaded constant value! 1007void SCCPSolver::visitLoadInst(LoadInst &I) { 1008 LatticeVal &IV = ValueState[&I]; 1009 if (IV.isOverdefined()) return; 1010 1011 LatticeVal &PtrVal = getValueState(I.getOperand(0)); 1012 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet! 1013 if (PtrVal.isConstant() && !I.isVolatile()) { 1014 Value *Ptr = PtrVal.getConstant(); 1015 if (isa<ConstantPointerNull>(Ptr)) { 1016 // load null -> null 1017 markConstant(IV, &I, Constant::getNullValue(I.getType())); 1018 return; 1019 } 1020 1021 // Transform load (constant global) into the value loaded. 1022 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) { 1023 if (GV->isConstant()) { 1024 if (!GV->isExternal()) { 1025 markConstant(IV, &I, GV->getInitializer()); 1026 return; 1027 } 1028 } else if (!TrackedGlobals.empty()) { 1029 // If we are tracking this global, merge in the known value for it. 1030 hash_map<GlobalVariable*, LatticeVal>::iterator It = 1031 TrackedGlobals.find(GV); 1032 if (It != TrackedGlobals.end()) { 1033 mergeInValue(IV, &I, It->second); 1034 return; 1035 } 1036 } 1037 } 1038 1039 // Transform load (constantexpr_GEP global, 0, ...) into the value loaded. 1040 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 1041 if (CE->getOpcode() == Instruction::GetElementPtr) 1042 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 1043 if (GV->isConstant() && !GV->isExternal()) 1044 if (Constant *V = 1045 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) { 1046 markConstant(IV, &I, V); 1047 return; 1048 } 1049 } 1050 1051 // Otherwise we cannot say for certain what value this load will produce. 1052 // Bail out. 1053 markOverdefined(IV, &I); 1054} 1055 1056void SCCPSolver::visitCallSite(CallSite CS) { 1057 Function *F = CS.getCalledFunction(); 1058 1059 // If we are tracking this function, we must make sure to bind arguments as 1060 // appropriate. 1061 hash_map<Function*, LatticeVal>::iterator TFRVI =TrackedFunctionRetVals.end(); 1062 if (F && F->hasInternalLinkage()) 1063 TFRVI = TrackedFunctionRetVals.find(F); 1064 1065 if (TFRVI != TrackedFunctionRetVals.end()) { 1066 // If this is the first call to the function hit, mark its entry block 1067 // executable. 1068 if (!BBExecutable.count(F->begin())) 1069 MarkBlockExecutable(F->begin()); 1070 1071 CallSite::arg_iterator CAI = CS.arg_begin(); 1072 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1073 AI != E; ++AI, ++CAI) { 1074 LatticeVal &IV = ValueState[AI]; 1075 if (!IV.isOverdefined()) 1076 mergeInValue(IV, AI, getValueState(*CAI)); 1077 } 1078 } 1079 Instruction *I = CS.getInstruction(); 1080 if (I->getType() == Type::VoidTy) return; 1081 1082 LatticeVal &IV = ValueState[I]; 1083 if (IV.isOverdefined()) return; 1084 1085 // Propagate the return value of the function to the value of the instruction. 1086 if (TFRVI != TrackedFunctionRetVals.end()) { 1087 mergeInValue(IV, I, TFRVI->second); 1088 return; 1089 } 1090 1091 if (F == 0 || !F->isExternal() || !canConstantFoldCallTo(F)) { 1092 markOverdefined(IV, I); 1093 return; 1094 } 1095 1096 std::vector<Constant*> Operands; 1097 Operands.reserve(I->getNumOperands()-1); 1098 1099 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end(); 1100 AI != E; ++AI) { 1101 LatticeVal &State = getValueState(*AI); 1102 if (State.isUndefined()) 1103 return; // Operands are not resolved yet... 1104 else if (State.isOverdefined()) { 1105 markOverdefined(IV, I); 1106 return; 1107 } 1108 assert(State.isConstant() && "Unknown state!"); 1109 Operands.push_back(State.getConstant()); 1110 } 1111 1112 if (Constant *C = ConstantFoldCall(F, Operands)) 1113 markConstant(IV, I, C); 1114 else 1115 markOverdefined(IV, I); 1116} 1117 1118 1119void SCCPSolver::Solve() { 1120 // Process the work lists until they are empty! 1121 while (!BBWorkList.empty() || !InstWorkList.empty() || 1122 !OverdefinedInstWorkList.empty()) { 1123 // Process the instruction work list... 1124 while (!OverdefinedInstWorkList.empty()) { 1125 Value *I = OverdefinedInstWorkList.back(); 1126 OverdefinedInstWorkList.pop_back(); 1127 1128 DOUT << "\nPopped off OI-WL: " << *I; 1129 1130 // "I" got into the work list because it either made the transition from 1131 // bottom to constant 1132 // 1133 // Anything on this worklist that is overdefined need not be visited 1134 // since all of its users will have already been marked as overdefined 1135 // Update all of the users of this instruction's value... 1136 // 1137 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1138 UI != E; ++UI) 1139 OperandChangedState(*UI); 1140 } 1141 // Process the instruction work list... 1142 while (!InstWorkList.empty()) { 1143 Value *I = InstWorkList.back(); 1144 InstWorkList.pop_back(); 1145 1146 DOUT << "\nPopped off I-WL: " << *I; 1147 1148 // "I" got into the work list because it either made the transition from 1149 // bottom to constant 1150 // 1151 // Anything on this worklist that is overdefined need not be visited 1152 // since all of its users will have already been marked as overdefined. 1153 // Update all of the users of this instruction's value... 1154 // 1155 if (!getValueState(I).isOverdefined()) 1156 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1157 UI != E; ++UI) 1158 OperandChangedState(*UI); 1159 } 1160 1161 // Process the basic block work list... 1162 while (!BBWorkList.empty()) { 1163 BasicBlock *BB = BBWorkList.back(); 1164 BBWorkList.pop_back(); 1165 1166 DOUT << "\nPopped off BBWL: " << *BB; 1167 1168 // Notify all instructions in this basic block that they are newly 1169 // executable. 1170 visit(BB); 1171 } 1172 } 1173} 1174 1175/// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1176/// that branches on undef values cannot reach any of their successors. 1177/// However, this is not a safe assumption. After we solve dataflow, this 1178/// method should be use to handle this. If this returns true, the solver 1179/// should be rerun. 1180/// 1181/// This method handles this by finding an unresolved branch and marking it one 1182/// of the edges from the block as being feasible, even though the condition 1183/// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1184/// CFG and only slightly pessimizes the analysis results (by marking one, 1185/// potentially infeasible, edge feasible). This cannot usefully modify the 1186/// constraints on the condition of the branch, as that would impact other users 1187/// of the value. 1188/// 1189/// This scan also checks for values that use undefs, whose results are actually 1190/// defined. For example, 'zext i8 undef to i32' should produce all zeros 1191/// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero, 1192/// even if X isn't defined. 1193bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1194 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 1195 if (!BBExecutable.count(BB)) 1196 continue; 1197 1198 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 1199 // Look for instructions which produce undef values. 1200 if (I->getType() == Type::VoidTy) continue; 1201 1202 LatticeVal &LV = getValueState(I); 1203 if (!LV.isUndefined()) continue; 1204 1205 // Get the lattice values of the first two operands for use below. 1206 LatticeVal &Op0LV = getValueState(I->getOperand(0)); 1207 LatticeVal Op1LV; 1208 if (I->getNumOperands() == 2) { 1209 // If this is a two-operand instruction, and if both operands are 1210 // undefs, the result stays undef. 1211 Op1LV = getValueState(I->getOperand(1)); 1212 if (Op0LV.isUndefined() && Op1LV.isUndefined()) 1213 continue; 1214 } 1215 1216 // If this is an instructions whose result is defined even if the input is 1217 // not fully defined, propagate the information. 1218 const Type *ITy = I->getType(); 1219 switch (I->getOpcode()) { 1220 default: break; // Leave the instruction as an undef. 1221 case Instruction::ZExt: 1222 // After a zero extend, we know the top part is zero. SExt doesn't have 1223 // to be handled here, because we don't know whether the top part is 1's 1224 // or 0's. 1225 assert(Op0LV.isUndefined()); 1226 markForcedConstant(LV, I, Constant::getNullValue(ITy)); 1227 return true; 1228 case Instruction::Mul: 1229 case Instruction::And: 1230 // undef * X -> 0. X could be zero. 1231 // undef & X -> 0. X could be zero. 1232 markForcedConstant(LV, I, Constant::getNullValue(ITy)); 1233 return true; 1234 1235 case Instruction::Or: 1236 // undef | X -> -1. X could be -1. 1237 if (const PackedType *PTy = dyn_cast<PackedType>(ITy)) 1238 markForcedConstant(LV, I, ConstantPacked::getAllOnesValue(PTy)); 1239 else 1240 markForcedConstant(LV, I, ConstantInt::getAllOnesValue(ITy)); 1241 return true; 1242 1243 case Instruction::SDiv: 1244 case Instruction::UDiv: 1245 case Instruction::SRem: 1246 case Instruction::URem: 1247 // X / undef -> undef. No change. 1248 // X % undef -> undef. No change. 1249 if (Op1LV.isUndefined()) break; 1250 1251 // undef / X -> 0. X could be maxint. 1252 // undef % X -> 0. X could be 1. 1253 markForcedConstant(LV, I, Constant::getNullValue(ITy)); 1254 return true; 1255 1256 case Instruction::AShr: 1257 // undef >>s X -> undef. No change. 1258 if (Op0LV.isUndefined()) break; 1259 1260 // X >>s undef -> X. X could be 0, X could have the high-bit known set. 1261 if (Op0LV.isConstant()) 1262 markForcedConstant(LV, I, Op0LV.getConstant()); 1263 else 1264 markOverdefined(LV, I); 1265 return true; 1266 case Instruction::LShr: 1267 case Instruction::Shl: 1268 // undef >> X -> undef. No change. 1269 // undef << X -> undef. No change. 1270 if (Op0LV.isUndefined()) break; 1271 1272 // X >> undef -> 0. X could be 0. 1273 // X << undef -> 0. X could be 0. 1274 markForcedConstant(LV, I, Constant::getNullValue(ITy)); 1275 return true; 1276 case Instruction::Select: 1277 // undef ? X : Y -> X or Y. There could be commonality between X/Y. 1278 if (Op0LV.isUndefined()) { 1279 if (!Op1LV.isConstant()) // Pick the constant one if there is any. 1280 Op1LV = getValueState(I->getOperand(2)); 1281 } else if (Op1LV.isUndefined()) { 1282 // c ? undef : undef -> undef. No change. 1283 Op1LV = getValueState(I->getOperand(2)); 1284 if (Op1LV.isUndefined()) 1285 break; 1286 // Otherwise, c ? undef : x -> x. 1287 } else { 1288 // Leave Op1LV as Operand(1)'s LatticeValue. 1289 } 1290 1291 if (Op1LV.isConstant()) 1292 markForcedConstant(LV, I, Op1LV.getConstant()); 1293 else 1294 markOverdefined(LV, I); 1295 return true; 1296 } 1297 } 1298 1299 TerminatorInst *TI = BB->getTerminator(); 1300 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1301 if (!BI->isConditional()) continue; 1302 if (!getValueState(BI->getCondition()).isUndefined()) 1303 continue; 1304 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1305 if (!getValueState(SI->getCondition()).isUndefined()) 1306 continue; 1307 } else { 1308 continue; 1309 } 1310 1311 // If the edge to the first successor isn't thought to be feasible yet, mark 1312 // it so now. 1313 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(0)))) 1314 continue; 1315 1316 // Otherwise, it isn't already thought to be feasible. Mark it as such now 1317 // and return. This will make other blocks reachable, which will allow new 1318 // values to be discovered and existing ones to be moved in the lattice. 1319 markEdgeExecutable(BB, TI->getSuccessor(0)); 1320 return true; 1321 } 1322 1323 return false; 1324} 1325 1326 1327namespace { 1328 //===--------------------------------------------------------------------===// 1329 // 1330 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1331 /// Sparse Conditional Constant Propagator. 1332 /// 1333 struct SCCP : public FunctionPass { 1334 // runOnFunction - Run the Sparse Conditional Constant Propagation 1335 // algorithm, and return true if the function was modified. 1336 // 1337 bool runOnFunction(Function &F); 1338 1339 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 1340 AU.setPreservesCFG(); 1341 } 1342 }; 1343 1344 RegisterPass<SCCP> X("sccp", "Sparse Conditional Constant Propagation"); 1345} // end anonymous namespace 1346 1347 1348// createSCCPPass - This is the public interface to this file... 1349FunctionPass *llvm::createSCCPPass() { 1350 return new SCCP(); 1351} 1352 1353 1354// runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm, 1355// and return true if the function was modified. 1356// 1357bool SCCP::runOnFunction(Function &F) { 1358 DOUT << "SCCP on function '" << F.getName() << "'\n"; 1359 SCCPSolver Solver; 1360 1361 // Mark the first block of the function as being executable. 1362 Solver.MarkBlockExecutable(F.begin()); 1363 1364 // Mark all arguments to the function as being overdefined. 1365 hash_map<Value*, LatticeVal> &Values = Solver.getValueMapping(); 1366 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E; ++AI) 1367 Values[AI].markOverdefined(); 1368 1369 // Solve for constants. 1370 bool ResolvedUndefs = true; 1371 while (ResolvedUndefs) { 1372 Solver.Solve(); 1373 DOUT << "RESOLVING UNDEFs\n"; 1374 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1375 } 1376 1377 bool MadeChanges = false; 1378 1379 // If we decided that there are basic blocks that are dead in this function, 1380 // delete their contents now. Note that we cannot actually delete the blocks, 1381 // as we cannot modify the CFG of the function. 1382 // 1383 std::set<BasicBlock*> &ExecutableBBs = Solver.getExecutableBlocks(); 1384 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 1385 if (!ExecutableBBs.count(BB)) { 1386 DOUT << " BasicBlock Dead:" << *BB; 1387 ++NumDeadBlocks; 1388 1389 // Delete the instructions backwards, as it has a reduced likelihood of 1390 // having to update as many def-use and use-def chains. 1391 std::vector<Instruction*> Insts; 1392 for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator(); 1393 I != E; ++I) 1394 Insts.push_back(I); 1395 while (!Insts.empty()) { 1396 Instruction *I = Insts.back(); 1397 Insts.pop_back(); 1398 if (!I->use_empty()) 1399 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1400 BB->getInstList().erase(I); 1401 MadeChanges = true; 1402 ++NumInstRemoved; 1403 } 1404 } else { 1405 // Iterate over all of the instructions in a function, replacing them with 1406 // constants if we have found them to be of constant values. 1407 // 1408 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1409 Instruction *Inst = BI++; 1410 if (Inst->getType() != Type::VoidTy) { 1411 LatticeVal &IV = Values[Inst]; 1412 if (IV.isConstant() || IV.isUndefined() && 1413 !isa<TerminatorInst>(Inst)) { 1414 Constant *Const = IV.isConstant() 1415 ? IV.getConstant() : UndefValue::get(Inst->getType()); 1416 DOUT << " Constant: " << *Const << " = " << *Inst; 1417 1418 // Replaces all of the uses of a variable with uses of the constant. 1419 Inst->replaceAllUsesWith(Const); 1420 1421 // Delete the instruction. 1422 BB->getInstList().erase(Inst); 1423 1424 // Hey, we just changed something! 1425 MadeChanges = true; 1426 ++NumInstRemoved; 1427 } 1428 } 1429 } 1430 } 1431 1432 return MadeChanges; 1433} 1434 1435namespace { 1436 //===--------------------------------------------------------------------===// 1437 // 1438 /// IPSCCP Class - This class implements interprocedural Sparse Conditional 1439 /// Constant Propagation. 1440 /// 1441 struct IPSCCP : public ModulePass { 1442 bool runOnModule(Module &M); 1443 }; 1444 1445 RegisterPass<IPSCCP> 1446 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation"); 1447} // end anonymous namespace 1448 1449// createIPSCCPPass - This is the public interface to this file... 1450ModulePass *llvm::createIPSCCPPass() { 1451 return new IPSCCP(); 1452} 1453 1454 1455static bool AddressIsTaken(GlobalValue *GV) { 1456 // Delete any dead constantexpr klingons. 1457 GV->removeDeadConstantUsers(); 1458 1459 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); 1460 UI != E; ++UI) 1461 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 1462 if (SI->getOperand(0) == GV || SI->isVolatile()) 1463 return true; // Storing addr of GV. 1464 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) { 1465 // Make sure we are calling the function, not passing the address. 1466 CallSite CS = CallSite::get(cast<Instruction>(*UI)); 1467 for (CallSite::arg_iterator AI = CS.arg_begin(), 1468 E = CS.arg_end(); AI != E; ++AI) 1469 if (*AI == GV) 1470 return true; 1471 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 1472 if (LI->isVolatile()) 1473 return true; 1474 } else { 1475 return true; 1476 } 1477 return false; 1478} 1479 1480bool IPSCCP::runOnModule(Module &M) { 1481 SCCPSolver Solver; 1482 1483 // Loop over all functions, marking arguments to those with their addresses 1484 // taken or that are external as overdefined. 1485 // 1486 hash_map<Value*, LatticeVal> &Values = Solver.getValueMapping(); 1487 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) 1488 if (!F->hasInternalLinkage() || AddressIsTaken(F)) { 1489 if (!F->isExternal()) 1490 Solver.MarkBlockExecutable(F->begin()); 1491 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1492 AI != E; ++AI) 1493 Values[AI].markOverdefined(); 1494 } else { 1495 Solver.AddTrackedFunction(F); 1496 } 1497 1498 // Loop over global variables. We inform the solver about any internal global 1499 // variables that do not have their 'addresses taken'. If they don't have 1500 // their addresses taken, we can propagate constants through them. 1501 for (Module::global_iterator G = M.global_begin(), E = M.global_end(); 1502 G != E; ++G) 1503 if (!G->isConstant() && G->hasInternalLinkage() && !AddressIsTaken(G)) 1504 Solver.TrackValueOfGlobalVariable(G); 1505 1506 // Solve for constants. 1507 bool ResolvedUndefs = true; 1508 while (ResolvedUndefs) { 1509 Solver.Solve(); 1510 1511 DOUT << "RESOLVING UNDEFS\n"; 1512 ResolvedUndefs = false; 1513 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) 1514 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F); 1515 } 1516 1517 bool MadeChanges = false; 1518 1519 // Iterate over all of the instructions in the module, replacing them with 1520 // constants if we have found them to be of constant values. 1521 // 1522 std::set<BasicBlock*> &ExecutableBBs = Solver.getExecutableBlocks(); 1523 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { 1524 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1525 AI != E; ++AI) 1526 if (!AI->use_empty()) { 1527 LatticeVal &IV = Values[AI]; 1528 if (IV.isConstant() || IV.isUndefined()) { 1529 Constant *CST = IV.isConstant() ? 1530 IV.getConstant() : UndefValue::get(AI->getType()); 1531 DOUT << "*** Arg " << *AI << " = " << *CST <<"\n"; 1532 1533 // Replaces all of the uses of a variable with uses of the 1534 // constant. 1535 AI->replaceAllUsesWith(CST); 1536 ++IPNumArgsElimed; 1537 } 1538 } 1539 1540 std::vector<BasicBlock*> BlocksToErase; 1541 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 1542 if (!ExecutableBBs.count(BB)) { 1543 DOUT << " BasicBlock Dead:" << *BB; 1544 ++IPNumDeadBlocks; 1545 1546 // Delete the instructions backwards, as it has a reduced likelihood of 1547 // having to update as many def-use and use-def chains. 1548 std::vector<Instruction*> Insts; 1549 TerminatorInst *TI = BB->getTerminator(); 1550 for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I) 1551 Insts.push_back(I); 1552 1553 while (!Insts.empty()) { 1554 Instruction *I = Insts.back(); 1555 Insts.pop_back(); 1556 if (!I->use_empty()) 1557 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1558 BB->getInstList().erase(I); 1559 MadeChanges = true; 1560 ++IPNumInstRemoved; 1561 } 1562 1563 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1564 BasicBlock *Succ = TI->getSuccessor(i); 1565 if (Succ->begin() != Succ->end() && isa<PHINode>(Succ->begin())) 1566 TI->getSuccessor(i)->removePredecessor(BB); 1567 } 1568 if (!TI->use_empty()) 1569 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 1570 BB->getInstList().erase(TI); 1571 1572 if (&*BB != &F->front()) 1573 BlocksToErase.push_back(BB); 1574 else 1575 new UnreachableInst(BB); 1576 1577 } else { 1578 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1579 Instruction *Inst = BI++; 1580 if (Inst->getType() != Type::VoidTy) { 1581 LatticeVal &IV = Values[Inst]; 1582 if (IV.isConstant() || IV.isUndefined() && 1583 !isa<TerminatorInst>(Inst)) { 1584 Constant *Const = IV.isConstant() 1585 ? IV.getConstant() : UndefValue::get(Inst->getType()); 1586 DOUT << " Constant: " << *Const << " = " << *Inst; 1587 1588 // Replaces all of the uses of a variable with uses of the 1589 // constant. 1590 Inst->replaceAllUsesWith(Const); 1591 1592 // Delete the instruction. 1593 if (!isa<TerminatorInst>(Inst) && !isa<CallInst>(Inst)) 1594 BB->getInstList().erase(Inst); 1595 1596 // Hey, we just changed something! 1597 MadeChanges = true; 1598 ++IPNumInstRemoved; 1599 } 1600 } 1601 } 1602 } 1603 1604 // Now that all instructions in the function are constant folded, erase dead 1605 // blocks, because we can now use ConstantFoldTerminator to get rid of 1606 // in-edges. 1607 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) { 1608 // If there are any PHI nodes in this successor, drop entries for BB now. 1609 BasicBlock *DeadBB = BlocksToErase[i]; 1610 while (!DeadBB->use_empty()) { 1611 Instruction *I = cast<Instruction>(DeadBB->use_back()); 1612 bool Folded = ConstantFoldTerminator(I->getParent()); 1613 if (!Folded) { 1614 // The constant folder may not have been able to fold the terminator 1615 // if this is a branch or switch on undef. Fold it manually as a 1616 // branch to the first successor. 1617 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1618 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) && 1619 "Branch should be foldable!"); 1620 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1621 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold"); 1622 } else { 1623 assert(0 && "Didn't fold away reference to block!"); 1624 } 1625 1626 // Make this an uncond branch to the first successor. 1627 TerminatorInst *TI = I->getParent()->getTerminator(); 1628 new BranchInst(TI->getSuccessor(0), TI); 1629 1630 // Remove entries in successor phi nodes to remove edges. 1631 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i) 1632 TI->getSuccessor(i)->removePredecessor(TI->getParent()); 1633 1634 // Remove the old terminator. 1635 TI->eraseFromParent(); 1636 } 1637 } 1638 1639 // Finally, delete the basic block. 1640 F->getBasicBlockList().erase(DeadBB); 1641 } 1642 } 1643 1644 // If we inferred constant or undef return values for a function, we replaced 1645 // all call uses with the inferred value. This means we don't need to bother 1646 // actually returning anything from the function. Replace all return 1647 // instructions with return undef. 1648 const hash_map<Function*, LatticeVal> &RV =Solver.getTrackedFunctionRetVals(); 1649 for (hash_map<Function*, LatticeVal>::const_iterator I = RV.begin(), 1650 E = RV.end(); I != E; ++I) 1651 if (!I->second.isOverdefined() && 1652 I->first->getReturnType() != Type::VoidTy) { 1653 Function *F = I->first; 1654 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 1655 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 1656 if (!isa<UndefValue>(RI->getOperand(0))) 1657 RI->setOperand(0, UndefValue::get(F->getReturnType())); 1658 } 1659 1660 // If we infered constant or undef values for globals variables, we can delete 1661 // the global and any stores that remain to it. 1662 const hash_map<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals(); 1663 for (hash_map<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(), 1664 E = TG.end(); I != E; ++I) { 1665 GlobalVariable *GV = I->first; 1666 assert(!I->second.isOverdefined() && 1667 "Overdefined values should have been taken out of the map!"); 1668 DOUT << "Found that GV '" << GV->getName()<< "' is constant!\n"; 1669 while (!GV->use_empty()) { 1670 StoreInst *SI = cast<StoreInst>(GV->use_back()); 1671 SI->eraseFromParent(); 1672 } 1673 M.getGlobalList().erase(GV); 1674 ++IPNumGlobalConst; 1675 } 1676 1677 return MadeChanges; 1678} 1679