1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements sparse conditional constant propagation and merging: 10 // 11 // Specifically, this: 12 // * Assumes values are constant unless proven otherwise 13 // * Assumes BasicBlocks are dead unless proven otherwise 14 // * Proves values to be constant, and replaces them with constants 15 // * Proves conditional branches to be unconditional 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/Transforms/Scalar/SCCP.h" 20 #include "llvm/ADT/ArrayRef.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/DenseSet.h" 23 #include "llvm/ADT/PointerIntPair.h" 24 #include "llvm/ADT/STLExtras.h" 25 #include "llvm/ADT/SmallPtrSet.h" 26 #include "llvm/ADT/SmallVector.h" 27 #include "llvm/ADT/Statistic.h" 28 #include "llvm/Analysis/ConstantFolding.h" 29 #include "llvm/Analysis/GlobalsModRef.h" 30 #include "llvm/Analysis/TargetLibraryInfo.h" 31 #include "llvm/Transforms/Utils/Local.h" 32 #include "llvm/Analysis/ValueLattice.h" 33 #include "llvm/Analysis/ValueLatticeUtils.h" 34 #include "llvm/IR/BasicBlock.h" 35 #include "llvm/IR/CallSite.h" 36 #include "llvm/IR/Constant.h" 37 #include "llvm/IR/Constants.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DerivedTypes.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/IR/GlobalVariable.h" 42 #include "llvm/IR/InstVisitor.h" 43 #include "llvm/IR/InstrTypes.h" 44 #include "llvm/IR/Instruction.h" 45 #include "llvm/IR/Instructions.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/PassManager.h" 48 #include "llvm/IR/Type.h" 49 #include "llvm/IR/User.h" 50 #include "llvm/IR/Value.h" 51 #include "llvm/Pass.h" 52 #include "llvm/Support/Casting.h" 53 #include "llvm/Support/Debug.h" 54 #include "llvm/Support/ErrorHandling.h" 55 #include "llvm/Support/raw_ostream.h" 56 #include "llvm/Transforms/Scalar.h" 57 #include "llvm/Transforms/Utils/PredicateInfo.h" 58 #include <cassert> 59 #include <utility> 60 #include <vector> 61 62 using namespace llvm; 63 64 #define DEBUG_TYPE "sccp" 65 66 STATISTIC(NumInstRemoved, "Number of instructions removed"); 67 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 68 69 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 70 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 71 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 72 73 namespace { 74 75 /// LatticeVal class - This class represents the different lattice values that 76 /// an LLVM value may occupy. It is a simple class with value semantics. 77 /// 78 class LatticeVal { 79 enum LatticeValueTy { 80 /// unknown - This LLVM Value has no known value yet. 81 unknown, 82 83 /// constant - This LLVM Value has a specific constant value. 84 constant, 85 86 /// forcedconstant - This LLVM Value was thought to be undef until 87 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged 88 /// with another (different) constant, it goes to overdefined, instead of 89 /// asserting. 90 forcedconstant, 91 92 /// overdefined - This instruction is not known to be constant, and we know 93 /// it has a value. 94 overdefined 95 }; 96 97 /// Val: This stores the current lattice value along with the Constant* for 98 /// the constant if this is a 'constant' or 'forcedconstant' value. 99 PointerIntPair<Constant *, 2, LatticeValueTy> Val; 100 101 LatticeValueTy getLatticeValue() const { 102 return Val.getInt(); 103 } 104 105 public: 106 LatticeVal() : Val(nullptr, unknown) {} 107 108 bool isUnknown() const { return getLatticeValue() == unknown; } 109 110 bool isConstant() const { 111 return getLatticeValue() == constant || getLatticeValue() == forcedconstant; 112 } 113 114 bool isOverdefined() const { return getLatticeValue() == overdefined; } 115 116 Constant *getConstant() const { 117 assert(isConstant() && "Cannot get the constant of a non-constant!"); 118 return Val.getPointer(); 119 } 120 121 /// markOverdefined - Return true if this is a change in status. 122 bool markOverdefined() { 123 if (isOverdefined()) 124 return false; 125 126 Val.setInt(overdefined); 127 return true; 128 } 129 130 /// markConstant - Return true if this is a change in status. 131 bool markConstant(Constant *V) { 132 if (getLatticeValue() == constant) { // Constant but not forcedconstant. 133 assert(getConstant() == V && "Marking constant with different value"); 134 return false; 135 } 136 137 if (isUnknown()) { 138 Val.setInt(constant); 139 assert(V && "Marking constant with NULL"); 140 Val.setPointer(V); 141 } else { 142 assert(getLatticeValue() == forcedconstant && 143 "Cannot move from overdefined to constant!"); 144 // Stay at forcedconstant if the constant is the same. 145 if (V == getConstant()) return false; 146 147 // Otherwise, we go to overdefined. Assumptions made based on the 148 // forced value are possibly wrong. Assuming this is another constant 149 // could expose a contradiction. 150 Val.setInt(overdefined); 151 } 152 return true; 153 } 154 155 /// getConstantInt - If this is a constant with a ConstantInt value, return it 156 /// otherwise return null. 157 ConstantInt *getConstantInt() const { 158 if (isConstant()) 159 return dyn_cast<ConstantInt>(getConstant()); 160 return nullptr; 161 } 162 163 /// getBlockAddress - If this is a constant with a BlockAddress value, return 164 /// it, otherwise return null. 165 BlockAddress *getBlockAddress() const { 166 if (isConstant()) 167 return dyn_cast<BlockAddress>(getConstant()); 168 return nullptr; 169 } 170 171 void markForcedConstant(Constant *V) { 172 assert(isUnknown() && "Can't force a defined value!"); 173 Val.setInt(forcedconstant); 174 Val.setPointer(V); 175 } 176 177 ValueLatticeElement toValueLattice() const { 178 if (isOverdefined()) 179 return ValueLatticeElement::getOverdefined(); 180 if (isConstant()) 181 return ValueLatticeElement::get(getConstant()); 182 return ValueLatticeElement(); 183 } 184 }; 185 186 //===----------------------------------------------------------------------===// 187 // 188 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 189 /// Constant Propagation. 190 /// 191 class SCCPSolver : public InstVisitor<SCCPSolver> { 192 const DataLayout &DL; 193 const TargetLibraryInfo *TLI; 194 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 195 DenseMap<Value *, LatticeVal> ValueState; // The state each value is in. 196 // The state each parameter is in. 197 DenseMap<Value *, ValueLatticeElement> ParamState; 198 199 /// StructValueState - This maintains ValueState for values that have 200 /// StructType, for example for formal arguments, calls, insertelement, etc. 201 DenseMap<std::pair<Value *, unsigned>, LatticeVal> StructValueState; 202 203 /// GlobalValue - If we are tracking any values for the contents of a global 204 /// variable, we keep a mapping from the constant accessor to the element of 205 /// the global, to the currently known value. If the value becomes 206 /// overdefined, it's entry is simply removed from this map. 207 DenseMap<GlobalVariable *, LatticeVal> TrackedGlobals; 208 209 /// TrackedRetVals - If we are tracking arguments into and the return 210 /// value out of a function, it will have an entry in this map, indicating 211 /// what the known return value for the function is. 212 DenseMap<Function *, LatticeVal> TrackedRetVals; 213 214 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 215 /// that return multiple values. 216 DenseMap<std::pair<Function *, unsigned>, LatticeVal> TrackedMultipleRetVals; 217 218 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 219 /// represented here for efficient lookup. 220 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 221 222 /// MustTailFunctions - Each function here is a callee of non-removable 223 /// musttail call site. 224 SmallPtrSet<Function *, 16> MustTailCallees; 225 226 /// TrackingIncomingArguments - This is the set of functions for whose 227 /// arguments we make optimistic assumptions about and try to prove as 228 /// constants. 229 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 230 231 /// The reason for two worklists is that overdefined is the lowest state 232 /// on the lattice, and moving things to overdefined as fast as possible 233 /// makes SCCP converge much faster. 234 /// 235 /// By having a separate worklist, we accomplish this because everything 236 /// possibly overdefined will become overdefined at the soonest possible 237 /// point. 238 SmallVector<Value *, 64> OverdefinedInstWorkList; 239 SmallVector<Value *, 64> InstWorkList; 240 241 // The BasicBlock work list 242 SmallVector<BasicBlock *, 64> BBWorkList; 243 244 /// KnownFeasibleEdges - Entries in this set are edges which have already had 245 /// PHI nodes retriggered. 246 using Edge = std::pair<BasicBlock *, BasicBlock *>; 247 DenseSet<Edge> KnownFeasibleEdges; 248 249 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 250 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 251 252 public: 253 void addAnalysis(Function &F, AnalysisResultsForFn A) { 254 AnalysisResults.insert({&F, std::move(A)}); 255 } 256 257 const PredicateBase *getPredicateInfoFor(Instruction *I) { 258 auto A = AnalysisResults.find(I->getParent()->getParent()); 259 if (A == AnalysisResults.end()) 260 return nullptr; 261 return A->second.PredInfo->getPredicateInfoFor(I); 262 } 263 264 DomTreeUpdater getDTU(Function &F) { 265 auto A = AnalysisResults.find(&F); 266 assert(A != AnalysisResults.end() && "Need analysis results for function."); 267 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 268 } 269 270 SCCPSolver(const DataLayout &DL, const TargetLibraryInfo *tli) 271 : DL(DL), TLI(tli) {} 272 273 /// MarkBlockExecutable - This method can be used by clients to mark all of 274 /// the blocks that are known to be intrinsically live in the processed unit. 275 /// 276 /// This returns true if the block was not considered live before. 277 bool MarkBlockExecutable(BasicBlock *BB) { 278 if (!BBExecutable.insert(BB).second) 279 return false; 280 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 281 BBWorkList.push_back(BB); // Add the block to the work list! 282 return true; 283 } 284 285 /// TrackValueOfGlobalVariable - Clients can use this method to 286 /// inform the SCCPSolver that it should track loads and stores to the 287 /// specified global variable if it can. This is only legal to call if 288 /// performing Interprocedural SCCP. 289 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 290 // We only track the contents of scalar globals. 291 if (GV->getValueType()->isSingleValueType()) { 292 LatticeVal &IV = TrackedGlobals[GV]; 293 if (!isa<UndefValue>(GV->getInitializer())) 294 IV.markConstant(GV->getInitializer()); 295 } 296 } 297 298 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 299 /// and out of the specified function (which cannot have its address taken), 300 /// this method must be called. 301 void AddTrackedFunction(Function *F) { 302 // Add an entry, F -> undef. 303 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 304 MRVFunctionsTracked.insert(F); 305 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 306 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i), 307 LatticeVal())); 308 } else 309 TrackedRetVals.insert(std::make_pair(F, LatticeVal())); 310 } 311 312 /// AddMustTailCallee - If the SCCP solver finds that this function is called 313 /// from non-removable musttail call site. 314 void AddMustTailCallee(Function *F) { 315 MustTailCallees.insert(F); 316 } 317 318 /// Returns true if the given function is called from non-removable musttail 319 /// call site. 320 bool isMustTailCallee(Function *F) { 321 return MustTailCallees.count(F); 322 } 323 324 void AddArgumentTrackedFunction(Function *F) { 325 TrackingIncomingArguments.insert(F); 326 } 327 328 /// Returns true if the given function is in the solver's set of 329 /// argument-tracked functions. 330 bool isArgumentTrackedFunction(Function *F) { 331 return TrackingIncomingArguments.count(F); 332 } 333 334 /// Solve - Solve for constants and executable blocks. 335 void Solve(); 336 337 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 338 /// that branches on undef values cannot reach any of their successors. 339 /// However, this is not a safe assumption. After we solve dataflow, this 340 /// method should be use to handle this. If this returns true, the solver 341 /// should be rerun. 342 bool ResolvedUndefsIn(Function &F); 343 344 bool isBlockExecutable(BasicBlock *BB) const { 345 return BBExecutable.count(BB); 346 } 347 348 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 349 // block to the 'To' basic block is currently feasible. 350 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 351 352 std::vector<LatticeVal> getStructLatticeValueFor(Value *V) const { 353 std::vector<LatticeVal> StructValues; 354 auto *STy = dyn_cast<StructType>(V->getType()); 355 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 356 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 357 auto I = StructValueState.find(std::make_pair(V, i)); 358 assert(I != StructValueState.end() && "Value not in valuemap!"); 359 StructValues.push_back(I->second); 360 } 361 return StructValues; 362 } 363 364 const LatticeVal &getLatticeValueFor(Value *V) const { 365 assert(!V->getType()->isStructTy() && 366 "Should use getStructLatticeValueFor"); 367 DenseMap<Value *, LatticeVal>::const_iterator I = ValueState.find(V); 368 assert(I != ValueState.end() && 369 "V not found in ValueState nor Paramstate map!"); 370 return I->second; 371 } 372 373 /// getTrackedRetVals - Get the inferred return value map. 374 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() { 375 return TrackedRetVals; 376 } 377 378 /// getTrackedGlobals - Get and return the set of inferred initializers for 379 /// global variables. 380 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() { 381 return TrackedGlobals; 382 } 383 384 /// getMRVFunctionsTracked - Get the set of functions which return multiple 385 /// values tracked by the pass. 386 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 387 return MRVFunctionsTracked; 388 } 389 390 /// getMustTailCallees - Get the set of functions which are called 391 /// from non-removable musttail call sites. 392 const SmallPtrSet<Function *, 16> getMustTailCallees() { 393 return MustTailCallees; 394 } 395 396 /// markOverdefined - Mark the specified value overdefined. This 397 /// works with both scalars and structs. 398 void markOverdefined(Value *V) { 399 if (auto *STy = dyn_cast<StructType>(V->getType())) 400 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 401 markOverdefined(getStructValueState(V, i), V); 402 else 403 markOverdefined(ValueState[V], V); 404 } 405 406 // isStructLatticeConstant - Return true if all the lattice values 407 // corresponding to elements of the structure are not overdefined, 408 // false otherwise. 409 bool isStructLatticeConstant(Function *F, StructType *STy) { 410 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 411 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 412 assert(It != TrackedMultipleRetVals.end()); 413 LatticeVal LV = It->second; 414 if (LV.isOverdefined()) 415 return false; 416 } 417 return true; 418 } 419 420 private: 421 // pushToWorkList - Helper for markConstant/markForcedConstant/markOverdefined 422 void pushToWorkList(LatticeVal &IV, Value *V) { 423 if (IV.isOverdefined()) 424 return OverdefinedInstWorkList.push_back(V); 425 InstWorkList.push_back(V); 426 } 427 428 // markConstant - Make a value be marked as "constant". If the value 429 // is not already a constant, add it to the instruction work list so that 430 // the users of the instruction are updated later. 431 bool markConstant(LatticeVal &IV, Value *V, Constant *C) { 432 if (!IV.markConstant(C)) return false; 433 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 434 pushToWorkList(IV, V); 435 return true; 436 } 437 438 bool markConstant(Value *V, Constant *C) { 439 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 440 return markConstant(ValueState[V], V, C); 441 } 442 443 void markForcedConstant(Value *V, Constant *C) { 444 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 445 LatticeVal &IV = ValueState[V]; 446 IV.markForcedConstant(C); 447 LLVM_DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n'); 448 pushToWorkList(IV, V); 449 } 450 451 // markOverdefined - Make a value be marked as "overdefined". If the 452 // value is not already overdefined, add it to the overdefined instruction 453 // work list so that the users of the instruction are updated later. 454 bool markOverdefined(LatticeVal &IV, Value *V) { 455 if (!IV.markOverdefined()) return false; 456 457 LLVM_DEBUG(dbgs() << "markOverdefined: "; 458 if (auto *F = dyn_cast<Function>(V)) dbgs() 459 << "Function '" << F->getName() << "'\n"; 460 else dbgs() << *V << '\n'); 461 // Only instructions go on the work list 462 pushToWorkList(IV, V); 463 return true; 464 } 465 466 bool mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) { 467 if (IV.isOverdefined() || MergeWithV.isUnknown()) 468 return false; // Noop. 469 if (MergeWithV.isOverdefined()) 470 return markOverdefined(IV, V); 471 if (IV.isUnknown()) 472 return markConstant(IV, V, MergeWithV.getConstant()); 473 if (IV.getConstant() != MergeWithV.getConstant()) 474 return markOverdefined(IV, V); 475 return false; 476 } 477 478 bool mergeInValue(Value *V, LatticeVal MergeWithV) { 479 assert(!V->getType()->isStructTy() && 480 "non-structs should use markConstant"); 481 return mergeInValue(ValueState[V], V, MergeWithV); 482 } 483 484 /// getValueState - Return the LatticeVal object that corresponds to the 485 /// value. This function handles the case when the value hasn't been seen yet 486 /// by properly seeding constants etc. 487 LatticeVal &getValueState(Value *V) { 488 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 489 490 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I = 491 ValueState.insert(std::make_pair(V, LatticeVal())); 492 LatticeVal &LV = I.first->second; 493 494 if (!I.second) 495 return LV; // Common case, already in the map. 496 497 if (auto *C = dyn_cast<Constant>(V)) { 498 // Undef values remain unknown. 499 if (!isa<UndefValue>(V)) 500 LV.markConstant(C); // Constants are constant 501 } 502 503 // All others are underdefined by default. 504 return LV; 505 } 506 507 ValueLatticeElement &getParamState(Value *V) { 508 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 509 510 std::pair<DenseMap<Value*, ValueLatticeElement>::iterator, bool> 511 PI = ParamState.insert(std::make_pair(V, ValueLatticeElement())); 512 ValueLatticeElement &LV = PI.first->second; 513 if (PI.second) 514 LV = getValueState(V).toValueLattice(); 515 516 return LV; 517 } 518 519 /// getStructValueState - Return the LatticeVal object that corresponds to the 520 /// value/field pair. This function handles the case when the value hasn't 521 /// been seen yet by properly seeding constants etc. 522 LatticeVal &getStructValueState(Value *V, unsigned i) { 523 assert(V->getType()->isStructTy() && "Should use getValueState"); 524 assert(i < cast<StructType>(V->getType())->getNumElements() && 525 "Invalid element #"); 526 527 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator, 528 bool> I = StructValueState.insert( 529 std::make_pair(std::make_pair(V, i), LatticeVal())); 530 LatticeVal &LV = I.first->second; 531 532 if (!I.second) 533 return LV; // Common case, already in the map. 534 535 if (auto *C = dyn_cast<Constant>(V)) { 536 Constant *Elt = C->getAggregateElement(i); 537 538 if (!Elt) 539 LV.markOverdefined(); // Unknown sort of constant. 540 else if (isa<UndefValue>(Elt)) 541 ; // Undef values remain unknown. 542 else 543 LV.markConstant(Elt); // Constants are constant. 544 } 545 546 // All others are underdefined by default. 547 return LV; 548 } 549 550 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 551 /// work list if it is not already executable. 552 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 553 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 554 return false; // This edge is already known to be executable! 555 556 if (!MarkBlockExecutable(Dest)) { 557 // If the destination is already executable, we just made an *edge* 558 // feasible that wasn't before. Revisit the PHI nodes in the block 559 // because they have potentially new operands. 560 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 561 << " -> " << Dest->getName() << '\n'); 562 563 for (PHINode &PN : Dest->phis()) 564 visitPHINode(PN); 565 } 566 return true; 567 } 568 569 // getFeasibleSuccessors - Return a vector of booleans to indicate which 570 // successors are reachable from a given terminator instruction. 571 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 572 573 // OperandChangedState - This method is invoked on all of the users of an 574 // instruction that was just changed state somehow. Based on this 575 // information, we need to update the specified user of this instruction. 576 void OperandChangedState(Instruction *I) { 577 if (BBExecutable.count(I->getParent())) // Inst is executable? 578 visit(*I); 579 } 580 581 // Add U as additional user of V. 582 void addAdditionalUser(Value *V, User *U) { 583 auto Iter = AdditionalUsers.insert({V, {}}); 584 Iter.first->second.insert(U); 585 } 586 587 // Mark I's users as changed, including AdditionalUsers. 588 void markUsersAsChanged(Value *I) { 589 for (User *U : I->users()) 590 if (auto *UI = dyn_cast<Instruction>(U)) 591 OperandChangedState(UI); 592 593 auto Iter = AdditionalUsers.find(I); 594 if (Iter != AdditionalUsers.end()) { 595 for (User *U : Iter->second) 596 if (auto *UI = dyn_cast<Instruction>(U)) 597 OperandChangedState(UI); 598 } 599 } 600 601 private: 602 friend class InstVisitor<SCCPSolver>; 603 604 // visit implementations - Something changed in this instruction. Either an 605 // operand made a transition, or the instruction is newly executable. Change 606 // the value type of I to reflect these changes if appropriate. 607 void visitPHINode(PHINode &I); 608 609 // Terminators 610 611 void visitReturnInst(ReturnInst &I); 612 void visitTerminator(Instruction &TI); 613 614 void visitCastInst(CastInst &I); 615 void visitSelectInst(SelectInst &I); 616 void visitUnaryOperator(Instruction &I); 617 void visitBinaryOperator(Instruction &I); 618 void visitCmpInst(CmpInst &I); 619 void visitExtractValueInst(ExtractValueInst &EVI); 620 void visitInsertValueInst(InsertValueInst &IVI); 621 622 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 623 markOverdefined(&CPI); 624 visitTerminator(CPI); 625 } 626 627 // Instructions that cannot be folded away. 628 629 void visitStoreInst (StoreInst &I); 630 void visitLoadInst (LoadInst &I); 631 void visitGetElementPtrInst(GetElementPtrInst &I); 632 633 void visitCallInst (CallInst &I) { 634 visitCallSite(&I); 635 } 636 637 void visitInvokeInst (InvokeInst &II) { 638 visitCallSite(&II); 639 visitTerminator(II); 640 } 641 642 void visitCallBrInst (CallBrInst &CBI) { 643 visitCallSite(&CBI); 644 visitTerminator(CBI); 645 } 646 647 void visitCallSite (CallSite CS); 648 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 649 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 650 void visitFenceInst (FenceInst &I) { /*returns void*/ } 651 652 void visitInstruction(Instruction &I) { 653 // All the instructions we don't do any special handling for just 654 // go to overdefined. 655 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 656 markOverdefined(&I); 657 } 658 }; 659 660 } // end anonymous namespace 661 662 // getFeasibleSuccessors - Return a vector of booleans to indicate which 663 // successors are reachable from a given terminator instruction. 664 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 665 SmallVectorImpl<bool> &Succs) { 666 Succs.resize(TI.getNumSuccessors()); 667 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 668 if (BI->isUnconditional()) { 669 Succs[0] = true; 670 return; 671 } 672 673 LatticeVal BCValue = getValueState(BI->getCondition()); 674 ConstantInt *CI = BCValue.getConstantInt(); 675 if (!CI) { 676 // Overdefined condition variables, and branches on unfoldable constant 677 // conditions, mean the branch could go either way. 678 if (!BCValue.isUnknown()) 679 Succs[0] = Succs[1] = true; 680 return; 681 } 682 683 // Constant condition variables mean the branch can only go a single way. 684 Succs[CI->isZero()] = true; 685 return; 686 } 687 688 // Unwinding instructions successors are always executable. 689 if (TI.isExceptionalTerminator()) { 690 Succs.assign(TI.getNumSuccessors(), true); 691 return; 692 } 693 694 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 695 if (!SI->getNumCases()) { 696 Succs[0] = true; 697 return; 698 } 699 LatticeVal SCValue = getValueState(SI->getCondition()); 700 ConstantInt *CI = SCValue.getConstantInt(); 701 702 if (!CI) { // Overdefined or unknown condition? 703 // All destinations are executable! 704 if (!SCValue.isUnknown()) 705 Succs.assign(TI.getNumSuccessors(), true); 706 return; 707 } 708 709 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 710 return; 711 } 712 713 // In case of indirect branch and its address is a blockaddress, we mark 714 // the target as executable. 715 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 716 // Casts are folded by visitCastInst. 717 LatticeVal IBRValue = getValueState(IBR->getAddress()); 718 BlockAddress *Addr = IBRValue.getBlockAddress(); 719 if (!Addr) { // Overdefined or unknown condition? 720 // All destinations are executable! 721 if (!IBRValue.isUnknown()) 722 Succs.assign(TI.getNumSuccessors(), true); 723 return; 724 } 725 726 BasicBlock* T = Addr->getBasicBlock(); 727 assert(Addr->getFunction() == T->getParent() && 728 "Block address of a different function ?"); 729 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 730 // This is the target. 731 if (IBR->getDestination(i) == T) { 732 Succs[i] = true; 733 return; 734 } 735 } 736 737 // If we didn't find our destination in the IBR successor list, then we 738 // have undefined behavior. Its ok to assume no successor is executable. 739 return; 740 } 741 742 // In case of callbr, we pessimistically assume that all successors are 743 // feasible. 744 if (isa<CallBrInst>(&TI)) { 745 Succs.assign(TI.getNumSuccessors(), true); 746 return; 747 } 748 749 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 750 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 751 } 752 753 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 754 // block to the 'To' basic block is currently feasible. 755 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 756 // Check if we've called markEdgeExecutable on the edge yet. (We could 757 // be more aggressive and try to consider edges which haven't been marked 758 // yet, but there isn't any need.) 759 return KnownFeasibleEdges.count(Edge(From, To)); 760 } 761 762 // visit Implementations - Something changed in this instruction, either an 763 // operand made a transition, or the instruction is newly executable. Change 764 // the value type of I to reflect these changes if appropriate. This method 765 // makes sure to do the following actions: 766 // 767 // 1. If a phi node merges two constants in, and has conflicting value coming 768 // from different branches, or if the PHI node merges in an overdefined 769 // value, then the PHI node becomes overdefined. 770 // 2. If a phi node merges only constants in, and they all agree on value, the 771 // PHI node becomes a constant value equal to that. 772 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 773 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 774 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 775 // 6. If a conditional branch has a value that is constant, make the selected 776 // destination executable 777 // 7. If a conditional branch has a value that is overdefined, make all 778 // successors executable. 779 void SCCPSolver::visitPHINode(PHINode &PN) { 780 // If this PN returns a struct, just mark the result overdefined. 781 // TODO: We could do a lot better than this if code actually uses this. 782 if (PN.getType()->isStructTy()) 783 return (void)markOverdefined(&PN); 784 785 if (getValueState(&PN).isOverdefined()) 786 return; // Quick exit 787 788 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 789 // and slow us down a lot. Just mark them overdefined. 790 if (PN.getNumIncomingValues() > 64) 791 return (void)markOverdefined(&PN); 792 793 // Look at all of the executable operands of the PHI node. If any of them 794 // are overdefined, the PHI becomes overdefined as well. If they are all 795 // constant, and they agree with each other, the PHI becomes the identical 796 // constant. If they are constant and don't agree, the PHI is overdefined. 797 // If there are no executable operands, the PHI remains unknown. 798 Constant *OperandVal = nullptr; 799 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 800 LatticeVal IV = getValueState(PN.getIncomingValue(i)); 801 if (IV.isUnknown()) continue; // Doesn't influence PHI node. 802 803 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 804 continue; 805 806 if (IV.isOverdefined()) // PHI node becomes overdefined! 807 return (void)markOverdefined(&PN); 808 809 if (!OperandVal) { // Grab the first value. 810 OperandVal = IV.getConstant(); 811 continue; 812 } 813 814 // There is already a reachable operand. If we conflict with it, 815 // then the PHI node becomes overdefined. If we agree with it, we 816 // can continue on. 817 818 // Check to see if there are two different constants merging, if so, the PHI 819 // node is overdefined. 820 if (IV.getConstant() != OperandVal) 821 return (void)markOverdefined(&PN); 822 } 823 824 // If we exited the loop, this means that the PHI node only has constant 825 // arguments that agree with each other(and OperandVal is the constant) or 826 // OperandVal is null because there are no defined incoming arguments. If 827 // this is the case, the PHI remains unknown. 828 if (OperandVal) 829 markConstant(&PN, OperandVal); // Acquire operand value 830 } 831 832 void SCCPSolver::visitReturnInst(ReturnInst &I) { 833 if (I.getNumOperands() == 0) return; // ret void 834 835 Function *F = I.getParent()->getParent(); 836 Value *ResultOp = I.getOperand(0); 837 838 // If we are tracking the return value of this function, merge it in. 839 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 840 DenseMap<Function*, LatticeVal>::iterator TFRVI = 841 TrackedRetVals.find(F); 842 if (TFRVI != TrackedRetVals.end()) { 843 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 844 return; 845 } 846 } 847 848 // Handle functions that return multiple values. 849 if (!TrackedMultipleRetVals.empty()) { 850 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 851 if (MRVFunctionsTracked.count(F)) 852 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 853 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 854 getStructValueState(ResultOp, i)); 855 } 856 } 857 858 void SCCPSolver::visitTerminator(Instruction &TI) { 859 SmallVector<bool, 16> SuccFeasible; 860 getFeasibleSuccessors(TI, SuccFeasible); 861 862 BasicBlock *BB = TI.getParent(); 863 864 // Mark all feasible successors executable. 865 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 866 if (SuccFeasible[i]) 867 markEdgeExecutable(BB, TI.getSuccessor(i)); 868 } 869 870 void SCCPSolver::visitCastInst(CastInst &I) { 871 LatticeVal OpSt = getValueState(I.getOperand(0)); 872 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand 873 markOverdefined(&I); 874 else if (OpSt.isConstant()) { 875 // Fold the constant as we build. 876 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpSt.getConstant(), 877 I.getType(), DL); 878 if (isa<UndefValue>(C)) 879 return; 880 // Propagate constant value 881 markConstant(&I, C); 882 } 883 } 884 885 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 886 // If this returns a struct, mark all elements over defined, we don't track 887 // structs in structs. 888 if (EVI.getType()->isStructTy()) 889 return (void)markOverdefined(&EVI); 890 891 // If this is extracting from more than one level of struct, we don't know. 892 if (EVI.getNumIndices() != 1) 893 return (void)markOverdefined(&EVI); 894 895 Value *AggVal = EVI.getAggregateOperand(); 896 if (AggVal->getType()->isStructTy()) { 897 unsigned i = *EVI.idx_begin(); 898 LatticeVal EltVal = getStructValueState(AggVal, i); 899 mergeInValue(getValueState(&EVI), &EVI, EltVal); 900 } else { 901 // Otherwise, must be extracting from an array. 902 return (void)markOverdefined(&EVI); 903 } 904 } 905 906 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 907 auto *STy = dyn_cast<StructType>(IVI.getType()); 908 if (!STy) 909 return (void)markOverdefined(&IVI); 910 911 // If this has more than one index, we can't handle it, drive all results to 912 // undef. 913 if (IVI.getNumIndices() != 1) 914 return (void)markOverdefined(&IVI); 915 916 Value *Aggr = IVI.getAggregateOperand(); 917 unsigned Idx = *IVI.idx_begin(); 918 919 // Compute the result based on what we're inserting. 920 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 921 // This passes through all values that aren't the inserted element. 922 if (i != Idx) { 923 LatticeVal EltVal = getStructValueState(Aggr, i); 924 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 925 continue; 926 } 927 928 Value *Val = IVI.getInsertedValueOperand(); 929 if (Val->getType()->isStructTy()) 930 // We don't track structs in structs. 931 markOverdefined(getStructValueState(&IVI, i), &IVI); 932 else { 933 LatticeVal InVal = getValueState(Val); 934 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 935 } 936 } 937 } 938 939 void SCCPSolver::visitSelectInst(SelectInst &I) { 940 // If this select returns a struct, just mark the result overdefined. 941 // TODO: We could do a lot better than this if code actually uses this. 942 if (I.getType()->isStructTy()) 943 return (void)markOverdefined(&I); 944 945 LatticeVal CondValue = getValueState(I.getCondition()); 946 if (CondValue.isUnknown()) 947 return; 948 949 if (ConstantInt *CondCB = CondValue.getConstantInt()) { 950 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 951 mergeInValue(&I, getValueState(OpVal)); 952 return; 953 } 954 955 // Otherwise, the condition is overdefined or a constant we can't evaluate. 956 // See if we can produce something better than overdefined based on the T/F 957 // value. 958 LatticeVal TVal = getValueState(I.getTrueValue()); 959 LatticeVal FVal = getValueState(I.getFalseValue()); 960 961 // select ?, C, C -> C. 962 if (TVal.isConstant() && FVal.isConstant() && 963 TVal.getConstant() == FVal.getConstant()) 964 return (void)markConstant(&I, FVal.getConstant()); 965 966 if (TVal.isUnknown()) // select ?, undef, X -> X. 967 return (void)mergeInValue(&I, FVal); 968 if (FVal.isUnknown()) // select ?, X, undef -> X. 969 return (void)mergeInValue(&I, TVal); 970 markOverdefined(&I); 971 } 972 973 // Handle Unary Operators. 974 void SCCPSolver::visitUnaryOperator(Instruction &I) { 975 LatticeVal V0State = getValueState(I.getOperand(0)); 976 977 LatticeVal &IV = ValueState[&I]; 978 if (IV.isOverdefined()) return; 979 980 if (V0State.isConstant()) { 981 Constant *C = ConstantExpr::get(I.getOpcode(), V0State.getConstant()); 982 983 // op Y -> undef. 984 if (isa<UndefValue>(C)) 985 return; 986 return (void)markConstant(IV, &I, C); 987 } 988 989 // If something is undef, wait for it to resolve. 990 if (!V0State.isOverdefined()) 991 return; 992 993 markOverdefined(&I); 994 } 995 996 // Handle Binary Operators. 997 void SCCPSolver::visitBinaryOperator(Instruction &I) { 998 LatticeVal V1State = getValueState(I.getOperand(0)); 999 LatticeVal V2State = getValueState(I.getOperand(1)); 1000 1001 LatticeVal &IV = ValueState[&I]; 1002 if (IV.isOverdefined()) return; 1003 1004 if (V1State.isConstant() && V2State.isConstant()) { 1005 Constant *C = ConstantExpr::get(I.getOpcode(), V1State.getConstant(), 1006 V2State.getConstant()); 1007 // X op Y -> undef. 1008 if (isa<UndefValue>(C)) 1009 return; 1010 return (void)markConstant(IV, &I, C); 1011 } 1012 1013 // If something is undef, wait for it to resolve. 1014 if (!V1State.isOverdefined() && !V2State.isOverdefined()) 1015 return; 1016 1017 // Otherwise, one of our operands is overdefined. Try to produce something 1018 // better than overdefined with some tricks. 1019 // If this is 0 / Y, it doesn't matter that the second operand is 1020 // overdefined, and we can replace it with zero. 1021 if (I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv) 1022 if (V1State.isConstant() && V1State.getConstant()->isNullValue()) 1023 return (void)markConstant(IV, &I, V1State.getConstant()); 1024 1025 // If this is: 1026 // -> AND/MUL with 0 1027 // -> OR with -1 1028 // it doesn't matter that the other operand is overdefined. 1029 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Mul || 1030 I.getOpcode() == Instruction::Or) { 1031 LatticeVal *NonOverdefVal = nullptr; 1032 if (!V1State.isOverdefined()) 1033 NonOverdefVal = &V1State; 1034 else if (!V2State.isOverdefined()) 1035 NonOverdefVal = &V2State; 1036 1037 if (NonOverdefVal) { 1038 if (NonOverdefVal->isUnknown()) 1039 return; 1040 1041 if (I.getOpcode() == Instruction::And || 1042 I.getOpcode() == Instruction::Mul) { 1043 // X and 0 = 0 1044 // X * 0 = 0 1045 if (NonOverdefVal->getConstant()->isNullValue()) 1046 return (void)markConstant(IV, &I, NonOverdefVal->getConstant()); 1047 } else { 1048 // X or -1 = -1 1049 if (ConstantInt *CI = NonOverdefVal->getConstantInt()) 1050 if (CI->isMinusOne()) 1051 return (void)markConstant(IV, &I, NonOverdefVal->getConstant()); 1052 } 1053 } 1054 } 1055 1056 markOverdefined(&I); 1057 } 1058 1059 // Handle ICmpInst instruction. 1060 void SCCPSolver::visitCmpInst(CmpInst &I) { 1061 // Do not cache this lookup, getValueState calls later in the function might 1062 // invalidate the reference. 1063 if (ValueState[&I].isOverdefined()) return; 1064 1065 Value *Op1 = I.getOperand(0); 1066 Value *Op2 = I.getOperand(1); 1067 1068 // For parameters, use ParamState which includes constant range info if 1069 // available. 1070 auto V1Param = ParamState.find(Op1); 1071 ValueLatticeElement V1State = (V1Param != ParamState.end()) 1072 ? V1Param->second 1073 : getValueState(Op1).toValueLattice(); 1074 1075 auto V2Param = ParamState.find(Op2); 1076 ValueLatticeElement V2State = V2Param != ParamState.end() 1077 ? V2Param->second 1078 : getValueState(Op2).toValueLattice(); 1079 1080 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1081 if (C) { 1082 if (isa<UndefValue>(C)) 1083 return; 1084 LatticeVal CV; 1085 CV.markConstant(C); 1086 mergeInValue(&I, CV); 1087 return; 1088 } 1089 1090 // If operands are still unknown, wait for it to resolve. 1091 if (!V1State.isOverdefined() && !V2State.isOverdefined() && 1092 !ValueState[&I].isConstant()) 1093 return; 1094 1095 markOverdefined(&I); 1096 } 1097 1098 // Handle getelementptr instructions. If all operands are constants then we 1099 // can turn this into a getelementptr ConstantExpr. 1100 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1101 if (ValueState[&I].isOverdefined()) return; 1102 1103 SmallVector<Constant*, 8> Operands; 1104 Operands.reserve(I.getNumOperands()); 1105 1106 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1107 LatticeVal State = getValueState(I.getOperand(i)); 1108 if (State.isUnknown()) 1109 return; // Operands are not resolved yet. 1110 1111 if (State.isOverdefined()) 1112 return (void)markOverdefined(&I); 1113 1114 assert(State.isConstant() && "Unknown state!"); 1115 Operands.push_back(State.getConstant()); 1116 } 1117 1118 Constant *Ptr = Operands[0]; 1119 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1120 Constant *C = 1121 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1122 if (isa<UndefValue>(C)) 1123 return; 1124 markConstant(&I, C); 1125 } 1126 1127 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1128 // If this store is of a struct, ignore it. 1129 if (SI.getOperand(0)->getType()->isStructTy()) 1130 return; 1131 1132 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1133 return; 1134 1135 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1136 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV); 1137 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return; 1138 1139 // Get the value we are storing into the global, then merge it. 1140 mergeInValue(I->second, GV, getValueState(SI.getOperand(0))); 1141 if (I->second.isOverdefined()) 1142 TrackedGlobals.erase(I); // No need to keep tracking this! 1143 } 1144 1145 // Handle load instructions. If the operand is a constant pointer to a constant 1146 // global, we can replace the load with the loaded constant value! 1147 void SCCPSolver::visitLoadInst(LoadInst &I) { 1148 // If this load is of a struct, just mark the result overdefined. 1149 if (I.getType()->isStructTy()) 1150 return (void)markOverdefined(&I); 1151 1152 LatticeVal PtrVal = getValueState(I.getOperand(0)); 1153 if (PtrVal.isUnknown()) return; // The pointer is not resolved yet! 1154 1155 LatticeVal &IV = ValueState[&I]; 1156 if (IV.isOverdefined()) return; 1157 1158 if (!PtrVal.isConstant() || I.isVolatile()) 1159 return (void)markOverdefined(IV, &I); 1160 1161 Constant *Ptr = PtrVal.getConstant(); 1162 1163 // load null is undefined. 1164 if (isa<ConstantPointerNull>(Ptr)) { 1165 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1166 return (void)markOverdefined(IV, &I); 1167 else 1168 return; 1169 } 1170 1171 // Transform load (constant global) into the value loaded. 1172 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1173 if (!TrackedGlobals.empty()) { 1174 // If we are tracking this global, merge in the known value for it. 1175 DenseMap<GlobalVariable*, LatticeVal>::iterator It = 1176 TrackedGlobals.find(GV); 1177 if (It != TrackedGlobals.end()) { 1178 mergeInValue(IV, &I, It->second); 1179 return; 1180 } 1181 } 1182 } 1183 1184 // Transform load from a constant into a constant if possible. 1185 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1186 if (isa<UndefValue>(C)) 1187 return; 1188 return (void)markConstant(IV, &I, C); 1189 } 1190 1191 // Otherwise we cannot say for certain what value this load will produce. 1192 // Bail out. 1193 markOverdefined(IV, &I); 1194 } 1195 1196 void SCCPSolver::visitCallSite(CallSite CS) { 1197 Function *F = CS.getCalledFunction(); 1198 Instruction *I = CS.getInstruction(); 1199 1200 if (auto *II = dyn_cast<IntrinsicInst>(I)) { 1201 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1202 if (ValueState[I].isOverdefined()) 1203 return; 1204 1205 auto *PI = getPredicateInfoFor(I); 1206 if (!PI) 1207 return; 1208 1209 Value *CopyOf = I->getOperand(0); 1210 auto *PBranch = dyn_cast<PredicateBranch>(PI); 1211 if (!PBranch) { 1212 mergeInValue(ValueState[I], I, getValueState(CopyOf)); 1213 return; 1214 } 1215 1216 Value *Cond = PBranch->Condition; 1217 1218 // Everything below relies on the condition being a comparison. 1219 auto *Cmp = dyn_cast<CmpInst>(Cond); 1220 if (!Cmp) { 1221 mergeInValue(ValueState[I], I, getValueState(CopyOf)); 1222 return; 1223 } 1224 1225 Value *CmpOp0 = Cmp->getOperand(0); 1226 Value *CmpOp1 = Cmp->getOperand(1); 1227 if (CopyOf != CmpOp0 && CopyOf != CmpOp1) { 1228 mergeInValue(ValueState[I], I, getValueState(CopyOf)); 1229 return; 1230 } 1231 1232 if (CmpOp0 != CopyOf) 1233 std::swap(CmpOp0, CmpOp1); 1234 1235 LatticeVal OriginalVal = getValueState(CopyOf); 1236 LatticeVal EqVal = getValueState(CmpOp1); 1237 LatticeVal &IV = ValueState[I]; 1238 if (PBranch->TrueEdge && Cmp->getPredicate() == CmpInst::ICMP_EQ) { 1239 addAdditionalUser(CmpOp1, I); 1240 if (OriginalVal.isConstant()) 1241 mergeInValue(IV, I, OriginalVal); 1242 else 1243 mergeInValue(IV, I, EqVal); 1244 return; 1245 } 1246 if (!PBranch->TrueEdge && Cmp->getPredicate() == CmpInst::ICMP_NE) { 1247 addAdditionalUser(CmpOp1, I); 1248 if (OriginalVal.isConstant()) 1249 mergeInValue(IV, I, OriginalVal); 1250 else 1251 mergeInValue(IV, I, EqVal); 1252 return; 1253 } 1254 1255 return (void)mergeInValue(IV, I, getValueState(CopyOf)); 1256 } 1257 } 1258 1259 // The common case is that we aren't tracking the callee, either because we 1260 // are not doing interprocedural analysis or the callee is indirect, or is 1261 // external. Handle these cases first. 1262 if (!F || F->isDeclaration()) { 1263 CallOverdefined: 1264 // Void return and not tracking callee, just bail. 1265 if (I->getType()->isVoidTy()) return; 1266 1267 // Otherwise, if we have a single return value case, and if the function is 1268 // a declaration, maybe we can constant fold it. 1269 if (F && F->isDeclaration() && !I->getType()->isStructTy() && 1270 canConstantFoldCallTo(cast<CallBase>(CS.getInstruction()), F)) { 1271 SmallVector<Constant*, 8> Operands; 1272 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end(); 1273 AI != E; ++AI) { 1274 if (AI->get()->getType()->isStructTy()) 1275 return markOverdefined(I); // Can't handle struct args. 1276 LatticeVal State = getValueState(*AI); 1277 1278 if (State.isUnknown()) 1279 return; // Operands are not resolved yet. 1280 if (State.isOverdefined()) 1281 return (void)markOverdefined(I); 1282 assert(State.isConstant() && "Unknown state!"); 1283 Operands.push_back(State.getConstant()); 1284 } 1285 1286 if (getValueState(I).isOverdefined()) 1287 return; 1288 1289 // If we can constant fold this, mark the result of the call as a 1290 // constant. 1291 if (Constant *C = ConstantFoldCall(cast<CallBase>(CS.getInstruction()), F, 1292 Operands, TLI)) { 1293 // call -> undef. 1294 if (isa<UndefValue>(C)) 1295 return; 1296 return (void)markConstant(I, C); 1297 } 1298 } 1299 1300 // Otherwise, we don't know anything about this call, mark it overdefined. 1301 return (void)markOverdefined(I); 1302 } 1303 1304 // If this is a local function that doesn't have its address taken, mark its 1305 // entry block executable and merge in the actual arguments to the call into 1306 // the formal arguments of the function. 1307 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){ 1308 MarkBlockExecutable(&F->front()); 1309 1310 // Propagate information from this call site into the callee. 1311 CallSite::arg_iterator CAI = CS.arg_begin(); 1312 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); 1313 AI != E; ++AI, ++CAI) { 1314 // If this argument is byval, and if the function is not readonly, there 1315 // will be an implicit copy formed of the input aggregate. 1316 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1317 markOverdefined(&*AI); 1318 continue; 1319 } 1320 1321 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1322 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1323 LatticeVal CallArg = getStructValueState(*CAI, i); 1324 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg); 1325 } 1326 } else { 1327 // Most other parts of the Solver still only use the simpler value 1328 // lattice, so we propagate changes for parameters to both lattices. 1329 LatticeVal ConcreteArgument = getValueState(*CAI); 1330 bool ParamChanged = 1331 getParamState(&*AI).mergeIn(ConcreteArgument.toValueLattice(), DL); 1332 bool ValueChanged = mergeInValue(&*AI, ConcreteArgument); 1333 // Add argument to work list, if the state of a parameter changes but 1334 // ValueState does not change (because it is already overdefined there), 1335 // We have to take changes in ParamState into account, as it is used 1336 // when evaluating Cmp instructions. 1337 if (!ValueChanged && ParamChanged) 1338 pushToWorkList(ValueState[&*AI], &*AI); 1339 } 1340 } 1341 } 1342 1343 // If this is a single/zero retval case, see if we're tracking the function. 1344 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1345 if (!MRVFunctionsTracked.count(F)) 1346 goto CallOverdefined; // Not tracking this callee. 1347 1348 // If we are tracking this callee, propagate the result of the function 1349 // into this call site. 1350 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1351 mergeInValue(getStructValueState(I, i), I, 1352 TrackedMultipleRetVals[std::make_pair(F, i)]); 1353 } else { 1354 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F); 1355 if (TFRVI == TrackedRetVals.end()) 1356 goto CallOverdefined; // Not tracking this callee. 1357 1358 // If so, propagate the return value of the callee into this call result. 1359 mergeInValue(I, TFRVI->second); 1360 } 1361 } 1362 1363 void SCCPSolver::Solve() { 1364 // Process the work lists until they are empty! 1365 while (!BBWorkList.empty() || !InstWorkList.empty() || 1366 !OverdefinedInstWorkList.empty()) { 1367 // Process the overdefined instruction's work list first, which drives other 1368 // things to overdefined more quickly. 1369 while (!OverdefinedInstWorkList.empty()) { 1370 Value *I = OverdefinedInstWorkList.pop_back_val(); 1371 1372 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1373 1374 // "I" got into the work list because it either made the transition from 1375 // bottom to constant, or to overdefined. 1376 // 1377 // Anything on this worklist that is overdefined need not be visited 1378 // since all of its users will have already been marked as overdefined 1379 // Update all of the users of this instruction's value. 1380 // 1381 markUsersAsChanged(I); 1382 } 1383 1384 // Process the instruction work list. 1385 while (!InstWorkList.empty()) { 1386 Value *I = InstWorkList.pop_back_val(); 1387 1388 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1389 1390 // "I" got into the work list because it made the transition from undef to 1391 // constant. 1392 // 1393 // Anything on this worklist that is overdefined need not be visited 1394 // since all of its users will have already been marked as overdefined. 1395 // Update all of the users of this instruction's value. 1396 // 1397 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1398 markUsersAsChanged(I); 1399 } 1400 1401 // Process the basic block work list. 1402 while (!BBWorkList.empty()) { 1403 BasicBlock *BB = BBWorkList.back(); 1404 BBWorkList.pop_back(); 1405 1406 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1407 1408 // Notify all instructions in this basic block that they are newly 1409 // executable. 1410 visit(BB); 1411 } 1412 } 1413 } 1414 1415 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1416 /// that branches on undef values cannot reach any of their successors. 1417 /// However, this is not a safe assumption. After we solve dataflow, this 1418 /// method should be use to handle this. If this returns true, the solver 1419 /// should be rerun. 1420 /// 1421 /// This method handles this by finding an unresolved branch and marking it one 1422 /// of the edges from the block as being feasible, even though the condition 1423 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1424 /// CFG and only slightly pessimizes the analysis results (by marking one, 1425 /// potentially infeasible, edge feasible). This cannot usefully modify the 1426 /// constraints on the condition of the branch, as that would impact other users 1427 /// of the value. 1428 /// 1429 /// This scan also checks for values that use undefs, whose results are actually 1430 /// defined. For example, 'zext i8 undef to i32' should produce all zeros 1431 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero, 1432 /// even if X isn't defined. 1433 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1434 for (BasicBlock &BB : F) { 1435 if (!BBExecutable.count(&BB)) 1436 continue; 1437 1438 for (Instruction &I : BB) { 1439 // Look for instructions which produce undef values. 1440 if (I.getType()->isVoidTy()) continue; 1441 1442 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1443 // Only a few things that can be structs matter for undef. 1444 1445 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1446 if (CallSite CS = CallSite(&I)) 1447 if (Function *F = CS.getCalledFunction()) 1448 if (MRVFunctionsTracked.count(F)) 1449 continue; 1450 1451 // extractvalue and insertvalue don't need to be marked; they are 1452 // tracked as precisely as their operands. 1453 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1454 continue; 1455 1456 // Send the results of everything else to overdefined. We could be 1457 // more precise than this but it isn't worth bothering. 1458 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1459 LatticeVal &LV = getStructValueState(&I, i); 1460 if (LV.isUnknown()) 1461 markOverdefined(LV, &I); 1462 } 1463 continue; 1464 } 1465 1466 LatticeVal &LV = getValueState(&I); 1467 if (!LV.isUnknown()) continue; 1468 1469 // extractvalue is safe; check here because the argument is a struct. 1470 if (isa<ExtractValueInst>(I)) 1471 continue; 1472 1473 // Compute the operand LatticeVals, for convenience below. 1474 // Anything taking a struct is conservatively assumed to require 1475 // overdefined markings. 1476 if (I.getOperand(0)->getType()->isStructTy()) { 1477 markOverdefined(&I); 1478 return true; 1479 } 1480 LatticeVal Op0LV = getValueState(I.getOperand(0)); 1481 LatticeVal Op1LV; 1482 if (I.getNumOperands() == 2) { 1483 if (I.getOperand(1)->getType()->isStructTy()) { 1484 markOverdefined(&I); 1485 return true; 1486 } 1487 1488 Op1LV = getValueState(I.getOperand(1)); 1489 } 1490 // If this is an instructions whose result is defined even if the input is 1491 // not fully defined, propagate the information. 1492 Type *ITy = I.getType(); 1493 switch (I.getOpcode()) { 1494 case Instruction::Add: 1495 case Instruction::Sub: 1496 case Instruction::Trunc: 1497 case Instruction::FPTrunc: 1498 case Instruction::BitCast: 1499 break; // Any undef -> undef 1500 case Instruction::FSub: 1501 case Instruction::FAdd: 1502 case Instruction::FMul: 1503 case Instruction::FDiv: 1504 case Instruction::FRem: 1505 // Floating-point binary operation: be conservative. 1506 if (Op0LV.isUnknown() && Op1LV.isUnknown()) 1507 markForcedConstant(&I, Constant::getNullValue(ITy)); 1508 else 1509 markOverdefined(&I); 1510 return true; 1511 case Instruction::FNeg: 1512 break; // fneg undef -> undef 1513 case Instruction::ZExt: 1514 case Instruction::SExt: 1515 case Instruction::FPToUI: 1516 case Instruction::FPToSI: 1517 case Instruction::FPExt: 1518 case Instruction::PtrToInt: 1519 case Instruction::IntToPtr: 1520 case Instruction::SIToFP: 1521 case Instruction::UIToFP: 1522 // undef -> 0; some outputs are impossible 1523 markForcedConstant(&I, Constant::getNullValue(ITy)); 1524 return true; 1525 case Instruction::Mul: 1526 case Instruction::And: 1527 // Both operands undef -> undef 1528 if (Op0LV.isUnknown() && Op1LV.isUnknown()) 1529 break; 1530 // undef * X -> 0. X could be zero. 1531 // undef & X -> 0. X could be zero. 1532 markForcedConstant(&I, Constant::getNullValue(ITy)); 1533 return true; 1534 case Instruction::Or: 1535 // Both operands undef -> undef 1536 if (Op0LV.isUnknown() && Op1LV.isUnknown()) 1537 break; 1538 // undef | X -> -1. X could be -1. 1539 markForcedConstant(&I, Constant::getAllOnesValue(ITy)); 1540 return true; 1541 case Instruction::Xor: 1542 // undef ^ undef -> 0; strictly speaking, this is not strictly 1543 // necessary, but we try to be nice to people who expect this 1544 // behavior in simple cases 1545 if (Op0LV.isUnknown() && Op1LV.isUnknown()) { 1546 markForcedConstant(&I, Constant::getNullValue(ITy)); 1547 return true; 1548 } 1549 // undef ^ X -> undef 1550 break; 1551 case Instruction::SDiv: 1552 case Instruction::UDiv: 1553 case Instruction::SRem: 1554 case Instruction::URem: 1555 // X / undef -> undef. No change. 1556 // X % undef -> undef. No change. 1557 if (Op1LV.isUnknown()) break; 1558 1559 // X / 0 -> undef. No change. 1560 // X % 0 -> undef. No change. 1561 if (Op1LV.isConstant() && Op1LV.getConstant()->isZeroValue()) 1562 break; 1563 1564 // undef / X -> 0. X could be maxint. 1565 // undef % X -> 0. X could be 1. 1566 markForcedConstant(&I, Constant::getNullValue(ITy)); 1567 return true; 1568 case Instruction::AShr: 1569 // X >>a undef -> undef. 1570 if (Op1LV.isUnknown()) break; 1571 1572 // Shifting by the bitwidth or more is undefined. 1573 if (Op1LV.isConstant()) { 1574 if (auto *ShiftAmt = Op1LV.getConstantInt()) 1575 if (ShiftAmt->getLimitedValue() >= 1576 ShiftAmt->getType()->getScalarSizeInBits()) 1577 break; 1578 } 1579 1580 // undef >>a X -> 0 1581 markForcedConstant(&I, Constant::getNullValue(ITy)); 1582 return true; 1583 case Instruction::LShr: 1584 case Instruction::Shl: 1585 // X << undef -> undef. 1586 // X >> undef -> undef. 1587 if (Op1LV.isUnknown()) break; 1588 1589 // Shifting by the bitwidth or more is undefined. 1590 if (Op1LV.isConstant()) { 1591 if (auto *ShiftAmt = Op1LV.getConstantInt()) 1592 if (ShiftAmt->getLimitedValue() >= 1593 ShiftAmt->getType()->getScalarSizeInBits()) 1594 break; 1595 } 1596 1597 // undef << X -> 0 1598 // undef >> X -> 0 1599 markForcedConstant(&I, Constant::getNullValue(ITy)); 1600 return true; 1601 case Instruction::Select: 1602 Op1LV = getValueState(I.getOperand(1)); 1603 // undef ? X : Y -> X or Y. There could be commonality between X/Y. 1604 if (Op0LV.isUnknown()) { 1605 if (!Op1LV.isConstant()) // Pick the constant one if there is any. 1606 Op1LV = getValueState(I.getOperand(2)); 1607 } else if (Op1LV.isUnknown()) { 1608 // c ? undef : undef -> undef. No change. 1609 Op1LV = getValueState(I.getOperand(2)); 1610 if (Op1LV.isUnknown()) 1611 break; 1612 // Otherwise, c ? undef : x -> x. 1613 } else { 1614 // Leave Op1LV as Operand(1)'s LatticeValue. 1615 } 1616 1617 if (Op1LV.isConstant()) 1618 markForcedConstant(&I, Op1LV.getConstant()); 1619 else 1620 markOverdefined(&I); 1621 return true; 1622 case Instruction::Load: 1623 // A load here means one of two things: a load of undef from a global, 1624 // a load from an unknown pointer. Either way, having it return undef 1625 // is okay. 1626 break; 1627 case Instruction::ICmp: 1628 // X == undef -> undef. Other comparisons get more complicated. 1629 Op0LV = getValueState(I.getOperand(0)); 1630 Op1LV = getValueState(I.getOperand(1)); 1631 1632 if ((Op0LV.isUnknown() || Op1LV.isUnknown()) && 1633 cast<ICmpInst>(&I)->isEquality()) 1634 break; 1635 markOverdefined(&I); 1636 return true; 1637 case Instruction::Call: 1638 case Instruction::Invoke: 1639 case Instruction::CallBr: 1640 // There are two reasons a call can have an undef result 1641 // 1. It could be tracked. 1642 // 2. It could be constant-foldable. 1643 // Because of the way we solve return values, tracked calls must 1644 // never be marked overdefined in ResolvedUndefsIn. 1645 if (Function *F = CallSite(&I).getCalledFunction()) 1646 if (TrackedRetVals.count(F)) 1647 break; 1648 1649 // If the call is constant-foldable, we mark it overdefined because 1650 // we do not know what return values are valid. 1651 markOverdefined(&I); 1652 return true; 1653 default: 1654 // If we don't know what should happen here, conservatively mark it 1655 // overdefined. 1656 markOverdefined(&I); 1657 return true; 1658 } 1659 } 1660 1661 // Check to see if we have a branch or switch on an undefined value. If so 1662 // we force the branch to go one way or the other to make the successor 1663 // values live. It doesn't really matter which way we force it. 1664 Instruction *TI = BB.getTerminator(); 1665 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1666 if (!BI->isConditional()) continue; 1667 if (!getValueState(BI->getCondition()).isUnknown()) 1668 continue; 1669 1670 // If the input to SCCP is actually branch on undef, fix the undef to 1671 // false. 1672 if (isa<UndefValue>(BI->getCondition())) { 1673 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1674 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1675 return true; 1676 } 1677 1678 // Otherwise, it is a branch on a symbolic value which is currently 1679 // considered to be undef. Make sure some edge is executable, so a 1680 // branch on "undef" always flows somewhere. 1681 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1682 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1683 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1684 return true; 1685 1686 continue; 1687 } 1688 1689 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1690 // Indirect branch with no successor ?. Its ok to assume it branches 1691 // to no target. 1692 if (IBR->getNumSuccessors() < 1) 1693 continue; 1694 1695 if (!getValueState(IBR->getAddress()).isUnknown()) 1696 continue; 1697 1698 // If the input to SCCP is actually branch on undef, fix the undef to 1699 // the first successor of the indirect branch. 1700 if (isa<UndefValue>(IBR->getAddress())) { 1701 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1702 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1703 return true; 1704 } 1705 1706 // Otherwise, it is a branch on a symbolic value which is currently 1707 // considered to be undef. Make sure some edge is executable, so a 1708 // branch on "undef" always flows somewhere. 1709 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1710 // we can assume the branch has undefined behavior instead. 1711 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1712 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1713 return true; 1714 1715 continue; 1716 } 1717 1718 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1719 if (!SI->getNumCases() || !getValueState(SI->getCondition()).isUnknown()) 1720 continue; 1721 1722 // If the input to SCCP is actually switch on undef, fix the undef to 1723 // the first constant. 1724 if (isa<UndefValue>(SI->getCondition())) { 1725 SI->setCondition(SI->case_begin()->getCaseValue()); 1726 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1727 return true; 1728 } 1729 1730 // Otherwise, it is a branch on a symbolic value which is currently 1731 // considered to be undef. Make sure some edge is executable, so a 1732 // branch on "undef" always flows somewhere. 1733 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1734 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1735 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1736 return true; 1737 1738 continue; 1739 } 1740 } 1741 1742 return false; 1743 } 1744 1745 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1746 Constant *Const = nullptr; 1747 if (V->getType()->isStructTy()) { 1748 std::vector<LatticeVal> IVs = Solver.getStructLatticeValueFor(V); 1749 if (llvm::any_of(IVs, 1750 [](const LatticeVal &LV) { return LV.isOverdefined(); })) 1751 return false; 1752 std::vector<Constant *> ConstVals; 1753 auto *ST = dyn_cast<StructType>(V->getType()); 1754 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1755 LatticeVal V = IVs[i]; 1756 ConstVals.push_back(V.isConstant() 1757 ? V.getConstant() 1758 : UndefValue::get(ST->getElementType(i))); 1759 } 1760 Const = ConstantStruct::get(ST, ConstVals); 1761 } else { 1762 const LatticeVal &IV = Solver.getLatticeValueFor(V); 1763 if (IV.isOverdefined()) 1764 return false; 1765 1766 Const = IV.isConstant() ? IV.getConstant() : UndefValue::get(V->getType()); 1767 } 1768 assert(Const && "Constant is nullptr here!"); 1769 1770 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1771 // unless the call itself can be removed 1772 CallInst *CI = dyn_cast<CallInst>(V); 1773 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1774 CallSite CS(CI); 1775 Function *F = CS.getCalledFunction(); 1776 1777 // Don't zap returns of the callee 1778 if (F) 1779 Solver.AddMustTailCallee(F); 1780 1781 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1782 << " as a constant\n"); 1783 return false; 1784 } 1785 1786 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1787 1788 // Replaces all of the uses of a variable with uses of the constant. 1789 V->replaceAllUsesWith(Const); 1790 return true; 1791 } 1792 1793 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1794 // and return true if the function was modified. 1795 static bool runSCCP(Function &F, const DataLayout &DL, 1796 const TargetLibraryInfo *TLI) { 1797 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1798 SCCPSolver Solver(DL, TLI); 1799 1800 // Mark the first block of the function as being executable. 1801 Solver.MarkBlockExecutable(&F.front()); 1802 1803 // Mark all arguments to the function as being overdefined. 1804 for (Argument &AI : F.args()) 1805 Solver.markOverdefined(&AI); 1806 1807 // Solve for constants. 1808 bool ResolvedUndefs = true; 1809 while (ResolvedUndefs) { 1810 Solver.Solve(); 1811 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1812 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1813 } 1814 1815 bool MadeChanges = false; 1816 1817 // If we decided that there are basic blocks that are dead in this function, 1818 // delete their contents now. Note that we cannot actually delete the blocks, 1819 // as we cannot modify the CFG of the function. 1820 1821 for (BasicBlock &BB : F) { 1822 if (!Solver.isBlockExecutable(&BB)) { 1823 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1824 1825 ++NumDeadBlocks; 1826 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB); 1827 1828 MadeChanges = true; 1829 continue; 1830 } 1831 1832 // Iterate over all of the instructions in a function, replacing them with 1833 // constants if we have found them to be of constant values. 1834 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1835 Instruction *Inst = &*BI++; 1836 if (Inst->getType()->isVoidTy() || Inst->isTerminator()) 1837 continue; 1838 1839 if (tryToReplaceWithConstant(Solver, Inst)) { 1840 if (isInstructionTriviallyDead(Inst)) 1841 Inst->eraseFromParent(); 1842 // Hey, we just changed something! 1843 MadeChanges = true; 1844 ++NumInstRemoved; 1845 } 1846 } 1847 } 1848 1849 return MadeChanges; 1850 } 1851 1852 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1853 const DataLayout &DL = F.getParent()->getDataLayout(); 1854 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1855 if (!runSCCP(F, DL, &TLI)) 1856 return PreservedAnalyses::all(); 1857 1858 auto PA = PreservedAnalyses(); 1859 PA.preserve<GlobalsAA>(); 1860 PA.preserveSet<CFGAnalyses>(); 1861 return PA; 1862 } 1863 1864 namespace { 1865 1866 //===--------------------------------------------------------------------===// 1867 // 1868 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1869 /// Sparse Conditional Constant Propagator. 1870 /// 1871 class SCCPLegacyPass : public FunctionPass { 1872 public: 1873 // Pass identification, replacement for typeid 1874 static char ID; 1875 1876 SCCPLegacyPass() : FunctionPass(ID) { 1877 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1878 } 1879 1880 void getAnalysisUsage(AnalysisUsage &AU) const override { 1881 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1882 AU.addPreserved<GlobalsAAWrapperPass>(); 1883 AU.setPreservesCFG(); 1884 } 1885 1886 // runOnFunction - Run the Sparse Conditional Constant Propagation 1887 // algorithm, and return true if the function was modified. 1888 bool runOnFunction(Function &F) override { 1889 if (skipFunction(F)) 1890 return false; 1891 const DataLayout &DL = F.getParent()->getDataLayout(); 1892 const TargetLibraryInfo *TLI = 1893 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 1894 return runSCCP(F, DL, TLI); 1895 } 1896 }; 1897 1898 } // end anonymous namespace 1899 1900 char SCCPLegacyPass::ID = 0; 1901 1902 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1903 "Sparse Conditional Constant Propagation", false, false) 1904 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1905 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1906 "Sparse Conditional Constant Propagation", false, false) 1907 1908 // createSCCPPass - This is the public interface to this file. 1909 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1910 1911 static void findReturnsToZap(Function &F, 1912 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1913 SCCPSolver &Solver) { 1914 // We can only do this if we know that nothing else can call the function. 1915 if (!Solver.isArgumentTrackedFunction(&F)) 1916 return; 1917 1918 // There is a non-removable musttail call site of this function. Zapping 1919 // returns is not allowed. 1920 if (Solver.isMustTailCallee(&F)) { 1921 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1922 << " due to present musttail call of it\n"); 1923 return; 1924 } 1925 1926 for (BasicBlock &BB : F) { 1927 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1928 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1929 << "musttail call : " << *CI << "\n"); 1930 (void)CI; 1931 return; 1932 } 1933 1934 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1935 if (!isa<UndefValue>(RI->getOperand(0))) 1936 ReturnsToZap.push_back(RI); 1937 } 1938 } 1939 1940 // Update the condition for terminators that are branching on indeterminate 1941 // values, forcing them to use a specific edge. 1942 static void forceIndeterminateEdge(Instruction* I, SCCPSolver &Solver) { 1943 BasicBlock *Dest = nullptr; 1944 Constant *C = nullptr; 1945 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1946 if (!isa<ConstantInt>(SI->getCondition())) { 1947 // Indeterminate switch; use first case value. 1948 Dest = SI->case_begin()->getCaseSuccessor(); 1949 C = SI->case_begin()->getCaseValue(); 1950 } 1951 } else if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1952 if (!isa<ConstantInt>(BI->getCondition())) { 1953 // Indeterminate branch; use false. 1954 Dest = BI->getSuccessor(1); 1955 C = ConstantInt::getFalse(BI->getContext()); 1956 } 1957 } else if (IndirectBrInst *IBR = dyn_cast<IndirectBrInst>(I)) { 1958 if (!isa<BlockAddress>(IBR->getAddress()->stripPointerCasts())) { 1959 // Indeterminate indirectbr; use successor 0. 1960 Dest = IBR->getSuccessor(0); 1961 C = BlockAddress::get(IBR->getSuccessor(0)); 1962 } 1963 } else { 1964 llvm_unreachable("Unexpected terminator instruction"); 1965 } 1966 if (C) { 1967 assert(Solver.isEdgeFeasible(I->getParent(), Dest) && 1968 "Didn't find feasible edge?"); 1969 (void)Dest; 1970 1971 I->setOperand(0, C); 1972 } 1973 } 1974 1975 bool llvm::runIPSCCP( 1976 Module &M, const DataLayout &DL, const TargetLibraryInfo *TLI, 1977 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1978 SCCPSolver Solver(DL, TLI); 1979 1980 // Loop over all functions, marking arguments to those with their addresses 1981 // taken or that are external as overdefined. 1982 for (Function &F : M) { 1983 if (F.isDeclaration()) 1984 continue; 1985 1986 Solver.addAnalysis(F, getAnalysis(F)); 1987 1988 // Determine if we can track the function's return values. If so, add the 1989 // function to the solver's set of return-tracked functions. 1990 if (canTrackReturnsInterprocedurally(&F)) 1991 Solver.AddTrackedFunction(&F); 1992 1993 // Determine if we can track the function's arguments. If so, add the 1994 // function to the solver's set of argument-tracked functions. 1995 if (canTrackArgumentsInterprocedurally(&F)) { 1996 Solver.AddArgumentTrackedFunction(&F); 1997 continue; 1998 } 1999 2000 // Assume the function is called. 2001 Solver.MarkBlockExecutable(&F.front()); 2002 2003 // Assume nothing about the incoming arguments. 2004 for (Argument &AI : F.args()) 2005 Solver.markOverdefined(&AI); 2006 } 2007 2008 // Determine if we can track any of the module's global variables. If so, add 2009 // the global variables we can track to the solver's set of tracked global 2010 // variables. 2011 for (GlobalVariable &G : M.globals()) { 2012 G.removeDeadConstantUsers(); 2013 if (canTrackGlobalVariableInterprocedurally(&G)) 2014 Solver.TrackValueOfGlobalVariable(&G); 2015 } 2016 2017 // Solve for constants. 2018 bool ResolvedUndefs = true; 2019 Solver.Solve(); 2020 while (ResolvedUndefs) { 2021 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 2022 ResolvedUndefs = false; 2023 for (Function &F : M) 2024 if (Solver.ResolvedUndefsIn(F)) { 2025 // We run Solve() after we resolved an undef in a function, because 2026 // we might deduce a fact that eliminates an undef in another function. 2027 Solver.Solve(); 2028 ResolvedUndefs = true; 2029 } 2030 } 2031 2032 bool MadeChanges = false; 2033 2034 // Iterate over all of the instructions in the module, replacing them with 2035 // constants if we have found them to be of constant values. 2036 2037 for (Function &F : M) { 2038 if (F.isDeclaration()) 2039 continue; 2040 2041 SmallVector<BasicBlock *, 512> BlocksToErase; 2042 2043 if (Solver.isBlockExecutable(&F.front())) 2044 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E; 2045 ++AI) { 2046 if (!AI->use_empty() && tryToReplaceWithConstant(Solver, &*AI)) { 2047 ++IPNumArgsElimed; 2048 continue; 2049 } 2050 } 2051 2052 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 2053 if (!Solver.isBlockExecutable(&*BB)) { 2054 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB); 2055 ++NumDeadBlocks; 2056 2057 MadeChanges = true; 2058 2059 if (&*BB != &F.front()) 2060 BlocksToErase.push_back(&*BB); 2061 continue; 2062 } 2063 2064 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 2065 Instruction *Inst = &*BI++; 2066 if (Inst->getType()->isVoidTy()) 2067 continue; 2068 if (tryToReplaceWithConstant(Solver, Inst)) { 2069 if (Inst->isSafeToRemove()) 2070 Inst->eraseFromParent(); 2071 // Hey, we just changed something! 2072 MadeChanges = true; 2073 ++IPNumInstRemoved; 2074 } 2075 } 2076 } 2077 2078 DomTreeUpdater DTU = Solver.getDTU(F); 2079 // Change dead blocks to unreachable. We do it after replacing constants 2080 // in all executable blocks, because changeToUnreachable may remove PHI 2081 // nodes in executable blocks we found values for. The function's entry 2082 // block is not part of BlocksToErase, so we have to handle it separately. 2083 for (BasicBlock *BB : BlocksToErase) { 2084 NumInstRemoved += 2085 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 2086 /*PreserveLCSSA=*/false, &DTU); 2087 } 2088 if (!Solver.isBlockExecutable(&F.front())) 2089 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 2090 /*UseLLVMTrap=*/false, 2091 /*PreserveLCSSA=*/false, &DTU); 2092 2093 // Now that all instructions in the function are constant folded, 2094 // use ConstantFoldTerminator to get rid of in-edges, record DT updates and 2095 // delete dead BBs. 2096 for (BasicBlock *DeadBB : BlocksToErase) { 2097 // If there are any PHI nodes in this successor, drop entries for BB now. 2098 for (Value::user_iterator UI = DeadBB->user_begin(), 2099 UE = DeadBB->user_end(); 2100 UI != UE;) { 2101 // Grab the user and then increment the iterator early, as the user 2102 // will be deleted. Step past all adjacent uses from the same user. 2103 auto *I = dyn_cast<Instruction>(*UI); 2104 do { ++UI; } while (UI != UE && *UI == I); 2105 2106 // Ignore blockaddress users; BasicBlock's dtor will handle them. 2107 if (!I) continue; 2108 2109 // If we have forced an edge for an indeterminate value, then force the 2110 // terminator to fold to that edge. 2111 forceIndeterminateEdge(I, Solver); 2112 BasicBlock *InstBB = I->getParent(); 2113 bool Folded = ConstantFoldTerminator(InstBB, 2114 /*DeleteDeadConditions=*/false, 2115 /*TLI=*/nullptr, &DTU); 2116 assert(Folded && 2117 "Expect TermInst on constantint or blockaddress to be folded"); 2118 (void) Folded; 2119 // If we folded the terminator to an unconditional branch to another 2120 // dead block, replace it with Unreachable, to avoid trying to fold that 2121 // branch again. 2122 BranchInst *BI = cast<BranchInst>(InstBB->getTerminator()); 2123 if (BI && BI->isUnconditional() && 2124 !Solver.isBlockExecutable(BI->getSuccessor(0))) { 2125 InstBB->getTerminator()->eraseFromParent(); 2126 new UnreachableInst(InstBB->getContext(), InstBB); 2127 } 2128 } 2129 // Mark dead BB for deletion. 2130 DTU.deleteBB(DeadBB); 2131 } 2132 2133 for (BasicBlock &BB : F) { 2134 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 2135 Instruction *Inst = &*BI++; 2136 if (Solver.getPredicateInfoFor(Inst)) { 2137 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2138 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2139 Value *Op = II->getOperand(0); 2140 Inst->replaceAllUsesWith(Op); 2141 Inst->eraseFromParent(); 2142 } 2143 } 2144 } 2145 } 2146 } 2147 } 2148 2149 // If we inferred constant or undef return values for a function, we replaced 2150 // all call uses with the inferred value. This means we don't need to bother 2151 // actually returning anything from the function. Replace all return 2152 // instructions with return undef. 2153 // 2154 // Do this in two stages: first identify the functions we should process, then 2155 // actually zap their returns. This is important because we can only do this 2156 // if the address of the function isn't taken. In cases where a return is the 2157 // last use of a function, the order of processing functions would affect 2158 // whether other functions are optimizable. 2159 SmallVector<ReturnInst*, 8> ReturnsToZap; 2160 2161 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals(); 2162 for (const auto &I : RV) { 2163 Function *F = I.first; 2164 if (I.second.isOverdefined() || F->getReturnType()->isVoidTy()) 2165 continue; 2166 findReturnsToZap(*F, ReturnsToZap, Solver); 2167 } 2168 2169 for (const auto &F : Solver.getMRVFunctionsTracked()) { 2170 assert(F->getReturnType()->isStructTy() && 2171 "The return type should be a struct"); 2172 StructType *STy = cast<StructType>(F->getReturnType()); 2173 if (Solver.isStructLatticeConstant(F, STy)) 2174 findReturnsToZap(*F, ReturnsToZap, Solver); 2175 } 2176 2177 // Zap all returns which we've identified as zap to change. 2178 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2179 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2180 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2181 } 2182 2183 // If we inferred constant or undef values for globals variables, we can 2184 // delete the global and any stores that remain to it. 2185 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals(); 2186 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(), 2187 E = TG.end(); I != E; ++I) { 2188 GlobalVariable *GV = I->first; 2189 assert(!I->second.isOverdefined() && 2190 "Overdefined values should have been taken out of the map!"); 2191 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2192 << "' is constant!\n"); 2193 while (!GV->use_empty()) { 2194 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2195 SI->eraseFromParent(); 2196 } 2197 M.getGlobalList().erase(GV); 2198 ++IPNumGlobalConst; 2199 } 2200 2201 return MadeChanges; 2202 } 2203