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/MapVector.h" 24 #include "llvm/ADT/PointerIntPair.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SmallPtrSet.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/Analysis/ConstantFolding.h" 30 #include "llvm/Analysis/GlobalsModRef.h" 31 #include "llvm/Analysis/InstructionSimplify.h" 32 #include "llvm/Analysis/TargetLibraryInfo.h" 33 #include "llvm/Analysis/ValueLattice.h" 34 #include "llvm/Analysis/ValueLatticeUtils.h" 35 #include "llvm/IR/BasicBlock.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/InitializePasses.h" 52 #include "llvm/Pass.h" 53 #include "llvm/Support/Casting.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/ErrorHandling.h" 56 #include "llvm/Support/raw_ostream.h" 57 #include "llvm/Transforms/Scalar.h" 58 #include "llvm/Transforms/Utils/Local.h" 59 #include "llvm/Transforms/Utils/PredicateInfo.h" 60 #include <cassert> 61 #include <utility> 62 #include <vector> 63 64 using namespace llvm; 65 66 #define DEBUG_TYPE "sccp" 67 68 STATISTIC(NumInstRemoved, "Number of instructions removed"); 69 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 70 71 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 72 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 73 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 74 75 namespace { 76 77 // Helper to check if \p LV is either a constant or a constant 78 // range with a single element. This should cover exactly the same cases as the 79 // old ValueLatticeElement::isConstant() and is intended to be used in the 80 // transition to ValueLatticeElement. 81 bool isConstant(const ValueLatticeElement &LV) { 82 return LV.isConstant() || 83 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 84 } 85 86 // Helper to check if \p LV is either overdefined or a constant range with more 87 // than a single element. This should cover exactly the same cases as the old 88 // ValueLatticeElement::isOverdefined() and is intended to be used in the 89 // transition to ValueLatticeElement. 90 bool isOverdefined(const ValueLatticeElement &LV) { 91 return LV.isOverdefined() || 92 (LV.isConstantRange() && !LV.getConstantRange().isSingleElement()); 93 } 94 95 //===----------------------------------------------------------------------===// 96 // 97 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 98 /// Constant Propagation. 99 /// 100 class SCCPSolver : public InstVisitor<SCCPSolver> { 101 const DataLayout &DL; 102 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 103 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 104 DenseMap<Value *, ValueLatticeElement> 105 ValueState; // The state each value is in. 106 107 /// StructValueState - This maintains ValueState for values that have 108 /// StructType, for example for formal arguments, calls, insertelement, etc. 109 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 110 111 /// GlobalValue - If we are tracking any values for the contents of a global 112 /// variable, we keep a mapping from the constant accessor to the element of 113 /// the global, to the currently known value. If the value becomes 114 /// overdefined, it's entry is simply removed from this map. 115 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 116 117 /// TrackedRetVals - If we are tracking arguments into and the return 118 /// value out of a function, it will have an entry in this map, indicating 119 /// what the known return value for the function is. 120 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 121 122 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 123 /// that return multiple values. 124 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 125 TrackedMultipleRetVals; 126 127 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 128 /// represented here for efficient lookup. 129 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 130 131 /// MustTailFunctions - Each function here is a callee of non-removable 132 /// musttail call site. 133 SmallPtrSet<Function *, 16> MustTailCallees; 134 135 /// TrackingIncomingArguments - This is the set of functions for whose 136 /// arguments we make optimistic assumptions about and try to prove as 137 /// constants. 138 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 139 140 /// The reason for two worklists is that overdefined is the lowest state 141 /// on the lattice, and moving things to overdefined as fast as possible 142 /// makes SCCP converge much faster. 143 /// 144 /// By having a separate worklist, we accomplish this because everything 145 /// possibly overdefined will become overdefined at the soonest possible 146 /// point. 147 SmallVector<Value *, 64> OverdefinedInstWorkList; 148 SmallVector<Value *, 64> InstWorkList; 149 150 // The BasicBlock work list 151 SmallVector<BasicBlock *, 64> BBWorkList; 152 153 /// KnownFeasibleEdges - Entries in this set are edges which have already had 154 /// PHI nodes retriggered. 155 using Edge = std::pair<BasicBlock *, BasicBlock *>; 156 DenseSet<Edge> KnownFeasibleEdges; 157 158 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 159 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 160 161 LLVMContext &Ctx; 162 163 public: 164 void addAnalysis(Function &F, AnalysisResultsForFn A) { 165 AnalysisResults.insert({&F, std::move(A)}); 166 } 167 168 const PredicateBase *getPredicateInfoFor(Instruction *I) { 169 auto A = AnalysisResults.find(I->getParent()->getParent()); 170 if (A == AnalysisResults.end()) 171 return nullptr; 172 return A->second.PredInfo->getPredicateInfoFor(I); 173 } 174 175 DomTreeUpdater getDTU(Function &F) { 176 auto A = AnalysisResults.find(&F); 177 assert(A != AnalysisResults.end() && "Need analysis results for function."); 178 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 179 } 180 181 SCCPSolver(const DataLayout &DL, 182 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 183 LLVMContext &Ctx) 184 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 185 186 /// MarkBlockExecutable - This method can be used by clients to mark all of 187 /// the blocks that are known to be intrinsically live in the processed unit. 188 /// 189 /// This returns true if the block was not considered live before. 190 bool MarkBlockExecutable(BasicBlock *BB) { 191 if (!BBExecutable.insert(BB).second) 192 return false; 193 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 194 BBWorkList.push_back(BB); // Add the block to the work list! 195 return true; 196 } 197 198 /// TrackValueOfGlobalVariable - Clients can use this method to 199 /// inform the SCCPSolver that it should track loads and stores to the 200 /// specified global variable if it can. This is only legal to call if 201 /// performing Interprocedural SCCP. 202 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 203 // We only track the contents of scalar globals. 204 if (GV->getValueType()->isSingleValueType()) { 205 ValueLatticeElement &IV = TrackedGlobals[GV]; 206 if (!isa<UndefValue>(GV->getInitializer())) 207 IV.markConstant(GV->getInitializer()); 208 } 209 } 210 211 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 212 /// and out of the specified function (which cannot have its address taken), 213 /// this method must be called. 214 void AddTrackedFunction(Function *F) { 215 // Add an entry, F -> undef. 216 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 217 MRVFunctionsTracked.insert(F); 218 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 219 TrackedMultipleRetVals.insert( 220 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 221 } else 222 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 223 } 224 225 /// AddMustTailCallee - If the SCCP solver finds that this function is called 226 /// from non-removable musttail call site. 227 void AddMustTailCallee(Function *F) { 228 MustTailCallees.insert(F); 229 } 230 231 /// Returns true if the given function is called from non-removable musttail 232 /// call site. 233 bool isMustTailCallee(Function *F) { 234 return MustTailCallees.count(F); 235 } 236 237 void AddArgumentTrackedFunction(Function *F) { 238 TrackingIncomingArguments.insert(F); 239 } 240 241 /// Returns true if the given function is in the solver's set of 242 /// argument-tracked functions. 243 bool isArgumentTrackedFunction(Function *F) { 244 return TrackingIncomingArguments.count(F); 245 } 246 247 /// Solve - Solve for constants and executable blocks. 248 void Solve(); 249 250 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 251 /// that branches on undef values cannot reach any of their successors. 252 /// However, this is not a safe assumption. After we solve dataflow, this 253 /// method should be use to handle this. If this returns true, the solver 254 /// should be rerun. 255 bool ResolvedUndefsIn(Function &F); 256 257 bool isBlockExecutable(BasicBlock *BB) const { 258 return BBExecutable.count(BB); 259 } 260 261 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 262 // block to the 'To' basic block is currently feasible. 263 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 264 265 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 266 std::vector<ValueLatticeElement> StructValues; 267 auto *STy = dyn_cast<StructType>(V->getType()); 268 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 269 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 270 auto I = StructValueState.find(std::make_pair(V, i)); 271 assert(I != StructValueState.end() && "Value not in valuemap!"); 272 StructValues.push_back(I->second); 273 } 274 return StructValues; 275 } 276 277 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 278 assert(!V->getType()->isStructTy() && 279 "Should use getStructLatticeValueFor"); 280 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 281 ValueState.find(V); 282 assert(I != ValueState.end() && 283 "V not found in ValueState nor Paramstate map!"); 284 return I->second; 285 } 286 287 /// getTrackedRetVals - Get the inferred return value map. 288 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 289 return TrackedRetVals; 290 } 291 292 /// getTrackedGlobals - Get and return the set of inferred initializers for 293 /// global variables. 294 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 295 return TrackedGlobals; 296 } 297 298 /// getMRVFunctionsTracked - Get the set of functions which return multiple 299 /// values tracked by the pass. 300 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 301 return MRVFunctionsTracked; 302 } 303 304 /// getMustTailCallees - Get the set of functions which are called 305 /// from non-removable musttail call sites. 306 const SmallPtrSet<Function *, 16> getMustTailCallees() { 307 return MustTailCallees; 308 } 309 310 /// markOverdefined - Mark the specified value overdefined. This 311 /// works with both scalars and structs. 312 void markOverdefined(Value *V) { 313 if (auto *STy = dyn_cast<StructType>(V->getType())) 314 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 315 markOverdefined(getStructValueState(V, i), V); 316 else 317 markOverdefined(ValueState[V], V); 318 } 319 320 // isStructLatticeConstant - Return true if all the lattice values 321 // corresponding to elements of the structure are constants, 322 // false otherwise. 323 bool isStructLatticeConstant(Function *F, StructType *STy) { 324 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 325 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 326 assert(It != TrackedMultipleRetVals.end()); 327 ValueLatticeElement LV = It->second; 328 if (!isConstant(LV)) 329 return false; 330 } 331 return true; 332 } 333 334 /// Helper to return a Constant if \p LV is either a constant or a constant 335 /// range with a single element. 336 Constant *getConstant(const ValueLatticeElement &LV) const { 337 if (LV.isConstant()) 338 return LV.getConstant(); 339 340 if (LV.isConstantRange()) { 341 auto &CR = LV.getConstantRange(); 342 if (CR.getSingleElement()) 343 return ConstantInt::get(Ctx, *CR.getSingleElement()); 344 } 345 return nullptr; 346 } 347 348 private: 349 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 350 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 351 } 352 353 // pushToWorkList - Helper for markConstant/markOverdefined 354 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 355 if (IV.isOverdefined()) 356 return OverdefinedInstWorkList.push_back(V); 357 InstWorkList.push_back(V); 358 } 359 360 // Helper to push \p V to the worklist, after updating it to \p IV. Also 361 // prints a debug message with the updated value. 362 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 363 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 364 if (IV.isOverdefined()) 365 return OverdefinedInstWorkList.push_back(V); 366 InstWorkList.push_back(V); 367 } 368 369 // markConstant - Make a value be marked as "constant". If the value 370 // is not already a constant, add it to the instruction work list so that 371 // the users of the instruction are updated later. 372 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 373 bool MayIncludeUndef = false) { 374 if (!IV.markConstant(C, MayIncludeUndef)) 375 return false; 376 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 377 pushToWorkList(IV, V); 378 return true; 379 } 380 381 bool markConstant(Value *V, Constant *C) { 382 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 383 return markConstant(ValueState[V], V, C); 384 } 385 386 // markOverdefined - Make a value be marked as "overdefined". If the 387 // value is not already overdefined, add it to the overdefined instruction 388 // work list so that the users of the instruction are updated later. 389 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 390 if (!IV.markOverdefined()) return false; 391 392 LLVM_DEBUG(dbgs() << "markOverdefined: "; 393 if (auto *F = dyn_cast<Function>(V)) dbgs() 394 << "Function '" << F->getName() << "'\n"; 395 else dbgs() << *V << '\n'); 396 // Only instructions go on the work list 397 pushToWorkList(IV, V); 398 return true; 399 } 400 401 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 402 /// changes. 403 bool mergeInValue(ValueLatticeElement &IV, Value *V, 404 ValueLatticeElement MergeWithV, 405 ValueLatticeElement::MergeOptions Opts = { 406 /*MayIncludeUndef=*/false, /*CheckWiden=*/true}) { 407 if (IV.mergeIn(MergeWithV, Opts)) { 408 pushToWorkList(IV, V); 409 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 410 << IV << "\n"); 411 return true; 412 } 413 return false; 414 } 415 416 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 417 ValueLatticeElement::MergeOptions Opts = { 418 /*MayIncludeUndef=*/false, /*CheckWiden=*/true}) { 419 assert(!V->getType()->isStructTy() && 420 "non-structs should use markConstant"); 421 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 422 } 423 424 /// getValueState - Return the ValueLatticeElement object that corresponds to 425 /// the value. This function handles the case when the value hasn't been seen 426 /// yet by properly seeding constants etc. 427 ValueLatticeElement &getValueState(Value *V) { 428 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 429 430 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 431 ValueLatticeElement &LV = I.first->second; 432 433 if (!I.second) 434 return LV; // Common case, already in the map. 435 436 if (auto *C = dyn_cast<Constant>(V)) 437 LV.markConstant(C); // Constants are constant 438 439 // All others are unknown by default. 440 return LV; 441 } 442 443 /// getStructValueState - Return the ValueLatticeElement object that 444 /// corresponds to the value/field pair. This function handles the case when 445 /// the value hasn't been seen yet by properly seeding constants etc. 446 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 447 assert(V->getType()->isStructTy() && "Should use getValueState"); 448 assert(i < cast<StructType>(V->getType())->getNumElements() && 449 "Invalid element #"); 450 451 auto I = StructValueState.insert( 452 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 453 ValueLatticeElement &LV = I.first->second; 454 455 if (!I.second) 456 return LV; // Common case, already in the map. 457 458 if (auto *C = dyn_cast<Constant>(V)) { 459 Constant *Elt = C->getAggregateElement(i); 460 461 if (!Elt) 462 LV.markOverdefined(); // Unknown sort of constant. 463 else if (isa<UndefValue>(Elt)) 464 ; // Undef values remain unknown. 465 else 466 LV.markConstant(Elt); // Constants are constant. 467 } 468 469 // All others are underdefined by default. 470 return LV; 471 } 472 473 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 474 /// work list if it is not already executable. 475 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 476 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 477 return false; // This edge is already known to be executable! 478 479 if (!MarkBlockExecutable(Dest)) { 480 // If the destination is already executable, we just made an *edge* 481 // feasible that wasn't before. Revisit the PHI nodes in the block 482 // because they have potentially new operands. 483 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 484 << " -> " << Dest->getName() << '\n'); 485 486 for (PHINode &PN : Dest->phis()) 487 visitPHINode(PN); 488 } 489 return true; 490 } 491 492 // getFeasibleSuccessors - Return a vector of booleans to indicate which 493 // successors are reachable from a given terminator instruction. 494 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 495 496 // OperandChangedState - This method is invoked on all of the users of an 497 // instruction that was just changed state somehow. Based on this 498 // information, we need to update the specified user of this instruction. 499 void OperandChangedState(Instruction *I) { 500 if (BBExecutable.count(I->getParent())) // Inst is executable? 501 visit(*I); 502 } 503 504 // Add U as additional user of V. 505 void addAdditionalUser(Value *V, User *U) { 506 auto Iter = AdditionalUsers.insert({V, {}}); 507 Iter.first->second.insert(U); 508 } 509 510 // Mark I's users as changed, including AdditionalUsers. 511 void markUsersAsChanged(Value *I) { 512 // Functions include their arguments in the use-list. Changed function 513 // values mean that the result of the function changed. We only need to 514 // update the call sites with the new function result and do not have to 515 // propagate the call arguments. 516 if (isa<Function>(I)) { 517 for (User *U : I->users()) { 518 if (auto *CB = dyn_cast<CallBase>(U)) 519 handleCallResult(*CB); 520 } 521 } else { 522 for (User *U : I->users()) 523 if (auto *UI = dyn_cast<Instruction>(U)) 524 OperandChangedState(UI); 525 } 526 527 auto Iter = AdditionalUsers.find(I); 528 if (Iter != AdditionalUsers.end()) { 529 for (User *U : Iter->second) 530 if (auto *UI = dyn_cast<Instruction>(U)) 531 OperandChangedState(UI); 532 } 533 } 534 void handleCallOverdefined(CallBase &CB); 535 void handleCallResult(CallBase &CB); 536 void handleCallArguments(CallBase &CB); 537 538 private: 539 friend class InstVisitor<SCCPSolver>; 540 541 // visit implementations - Something changed in this instruction. Either an 542 // operand made a transition, or the instruction is newly executable. Change 543 // the value type of I to reflect these changes if appropriate. 544 void visitPHINode(PHINode &I); 545 546 // Terminators 547 548 void visitReturnInst(ReturnInst &I); 549 void visitTerminator(Instruction &TI); 550 551 void visitCastInst(CastInst &I); 552 void visitSelectInst(SelectInst &I); 553 void visitUnaryOperator(Instruction &I); 554 void visitBinaryOperator(Instruction &I); 555 void visitCmpInst(CmpInst &I); 556 void visitExtractValueInst(ExtractValueInst &EVI); 557 void visitInsertValueInst(InsertValueInst &IVI); 558 559 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 560 markOverdefined(&CPI); 561 visitTerminator(CPI); 562 } 563 564 // Instructions that cannot be folded away. 565 566 void visitStoreInst (StoreInst &I); 567 void visitLoadInst (LoadInst &I); 568 void visitGetElementPtrInst(GetElementPtrInst &I); 569 570 void visitCallInst (CallInst &I) { 571 visitCallBase(I); 572 } 573 574 void visitInvokeInst (InvokeInst &II) { 575 visitCallBase(II); 576 visitTerminator(II); 577 } 578 579 void visitCallBrInst (CallBrInst &CBI) { 580 visitCallBase(CBI); 581 visitTerminator(CBI); 582 } 583 584 void visitCallBase (CallBase &CB); 585 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 586 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 587 void visitFenceInst (FenceInst &I) { /*returns void*/ } 588 589 void visitInstruction(Instruction &I) { 590 // All the instructions we don't do any special handling for just 591 // go to overdefined. 592 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 593 markOverdefined(&I); 594 } 595 }; 596 597 } // end anonymous namespace 598 599 // getFeasibleSuccessors - Return a vector of booleans to indicate which 600 // successors are reachable from a given terminator instruction. 601 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 602 SmallVectorImpl<bool> &Succs) { 603 Succs.resize(TI.getNumSuccessors()); 604 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 605 if (BI->isUnconditional()) { 606 Succs[0] = true; 607 return; 608 } 609 610 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 611 ConstantInt *CI = getConstantInt(BCValue); 612 if (!CI) { 613 // Overdefined condition variables, and branches on unfoldable constant 614 // conditions, mean the branch could go either way. 615 if (!BCValue.isUnknownOrUndef()) 616 Succs[0] = Succs[1] = true; 617 return; 618 } 619 620 // Constant condition variables mean the branch can only go a single way. 621 Succs[CI->isZero()] = true; 622 return; 623 } 624 625 // Unwinding instructions successors are always executable. 626 if (TI.isExceptionalTerminator()) { 627 Succs.assign(TI.getNumSuccessors(), true); 628 return; 629 } 630 631 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 632 if (!SI->getNumCases()) { 633 Succs[0] = true; 634 return; 635 } 636 ValueLatticeElement SCValue = getValueState(SI->getCondition()); 637 ConstantInt *CI = getConstantInt(SCValue); 638 639 if (!CI) { // Overdefined or unknown condition? 640 // All destinations are executable! 641 if (!SCValue.isUnknownOrUndef()) 642 Succs.assign(TI.getNumSuccessors(), true); 643 return; 644 } 645 646 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 647 return; 648 } 649 650 // In case of indirect branch and its address is a blockaddress, we mark 651 // the target as executable. 652 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 653 // Casts are folded by visitCastInst. 654 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 655 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 656 if (!Addr) { // Overdefined or unknown condition? 657 // All destinations are executable! 658 if (!IBRValue.isUnknownOrUndef()) 659 Succs.assign(TI.getNumSuccessors(), true); 660 return; 661 } 662 663 BasicBlock* T = Addr->getBasicBlock(); 664 assert(Addr->getFunction() == T->getParent() && 665 "Block address of a different function ?"); 666 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 667 // This is the target. 668 if (IBR->getDestination(i) == T) { 669 Succs[i] = true; 670 return; 671 } 672 } 673 674 // If we didn't find our destination in the IBR successor list, then we 675 // have undefined behavior. Its ok to assume no successor is executable. 676 return; 677 } 678 679 // In case of callbr, we pessimistically assume that all successors are 680 // feasible. 681 if (isa<CallBrInst>(&TI)) { 682 Succs.assign(TI.getNumSuccessors(), true); 683 return; 684 } 685 686 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 687 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 688 } 689 690 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 691 // block to the 'To' basic block is currently feasible. 692 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 693 // Check if we've called markEdgeExecutable on the edge yet. (We could 694 // be more aggressive and try to consider edges which haven't been marked 695 // yet, but there isn't any need.) 696 return KnownFeasibleEdges.count(Edge(From, To)); 697 } 698 699 // visit Implementations - Something changed in this instruction, either an 700 // operand made a transition, or the instruction is newly executable. Change 701 // the value type of I to reflect these changes if appropriate. This method 702 // makes sure to do the following actions: 703 // 704 // 1. If a phi node merges two constants in, and has conflicting value coming 705 // from different branches, or if the PHI node merges in an overdefined 706 // value, then the PHI node becomes overdefined. 707 // 2. If a phi node merges only constants in, and they all agree on value, the 708 // PHI node becomes a constant value equal to that. 709 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 710 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 711 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 712 // 6. If a conditional branch has a value that is constant, make the selected 713 // destination executable 714 // 7. If a conditional branch has a value that is overdefined, make all 715 // successors executable. 716 void SCCPSolver::visitPHINode(PHINode &PN) { 717 // If this PN returns a struct, just mark the result overdefined. 718 // TODO: We could do a lot better than this if code actually uses this. 719 if (PN.getType()->isStructTy()) 720 return (void)markOverdefined(&PN); 721 722 if (getValueState(&PN).isOverdefined()) 723 return; // Quick exit 724 725 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 726 // and slow us down a lot. Just mark them overdefined. 727 if (PN.getNumIncomingValues() > 64) 728 return (void)markOverdefined(&PN); 729 730 // Look at all of the executable operands of the PHI node. If any of them 731 // are overdefined, the PHI becomes overdefined as well. If they are all 732 // constant, and they agree with each other, the PHI becomes the identical 733 // constant. If they are constant and don't agree, the PHI is overdefined. 734 // If there are no executable operands, the PHI remains unknown. 735 bool Changed = false; 736 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 737 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 738 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 739 continue; 740 741 ValueLatticeElement &Res = getValueState(&PN); 742 Changed |= Res.mergeIn(IV); 743 if (Res.isOverdefined()) 744 break; 745 } 746 if (Changed) 747 pushToWorkListMsg(ValueState[&PN], &PN); 748 } 749 750 void SCCPSolver::visitReturnInst(ReturnInst &I) { 751 if (I.getNumOperands() == 0) return; // ret void 752 753 Function *F = I.getParent()->getParent(); 754 Value *ResultOp = I.getOperand(0); 755 756 // If we are tracking the return value of this function, merge it in. 757 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 758 auto TFRVI = TrackedRetVals.find(F); 759 if (TFRVI != TrackedRetVals.end()) { 760 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 761 return; 762 } 763 } 764 765 // Handle functions that return multiple values. 766 if (!TrackedMultipleRetVals.empty()) { 767 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 768 if (MRVFunctionsTracked.count(F)) 769 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 770 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 771 getStructValueState(ResultOp, i)); 772 } 773 } 774 775 void SCCPSolver::visitTerminator(Instruction &TI) { 776 SmallVector<bool, 16> SuccFeasible; 777 getFeasibleSuccessors(TI, SuccFeasible); 778 779 BasicBlock *BB = TI.getParent(); 780 781 // Mark all feasible successors executable. 782 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 783 if (SuccFeasible[i]) 784 markEdgeExecutable(BB, TI.getSuccessor(i)); 785 } 786 787 void SCCPSolver::visitCastInst(CastInst &I) { 788 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 789 // discover a concrete value later. 790 if (ValueState[&I].isOverdefined()) 791 return; 792 793 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 794 if (Constant *OpC = getConstant(OpSt)) { 795 // Fold the constant as we build. 796 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 797 if (isa<UndefValue>(C)) 798 return; 799 // Propagate constant value 800 markConstant(&I, C); 801 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 802 auto &LV = getValueState(&I); 803 ConstantRange OpRange = OpSt.getConstantRange(); 804 Type *DestTy = I.getDestTy(); 805 ConstantRange Res = 806 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 807 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 808 } else if (!OpSt.isUnknownOrUndef()) 809 markOverdefined(&I); 810 } 811 812 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 813 // If this returns a struct, mark all elements over defined, we don't track 814 // structs in structs. 815 if (EVI.getType()->isStructTy()) 816 return (void)markOverdefined(&EVI); 817 818 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 819 // discover a concrete value later. 820 if (ValueState[&EVI].isOverdefined()) 821 return (void)markOverdefined(&EVI); 822 823 // If this is extracting from more than one level of struct, we don't know. 824 if (EVI.getNumIndices() != 1) 825 return (void)markOverdefined(&EVI); 826 827 Value *AggVal = EVI.getAggregateOperand(); 828 if (AggVal->getType()->isStructTy()) { 829 unsigned i = *EVI.idx_begin(); 830 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 831 mergeInValue(getValueState(&EVI), &EVI, EltVal); 832 } else { 833 // Otherwise, must be extracting from an array. 834 return (void)markOverdefined(&EVI); 835 } 836 } 837 838 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 839 auto *STy = dyn_cast<StructType>(IVI.getType()); 840 if (!STy) 841 return (void)markOverdefined(&IVI); 842 843 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 844 // discover a concrete value later. 845 if (isOverdefined(ValueState[&IVI])) 846 return (void)markOverdefined(&IVI); 847 848 // If this has more than one index, we can't handle it, drive all results to 849 // undef. 850 if (IVI.getNumIndices() != 1) 851 return (void)markOverdefined(&IVI); 852 853 Value *Aggr = IVI.getAggregateOperand(); 854 unsigned Idx = *IVI.idx_begin(); 855 856 // Compute the result based on what we're inserting. 857 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 858 // This passes through all values that aren't the inserted element. 859 if (i != Idx) { 860 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 861 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 862 continue; 863 } 864 865 Value *Val = IVI.getInsertedValueOperand(); 866 if (Val->getType()->isStructTy()) 867 // We don't track structs in structs. 868 markOverdefined(getStructValueState(&IVI, i), &IVI); 869 else { 870 ValueLatticeElement InVal = getValueState(Val); 871 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 872 } 873 } 874 } 875 876 void SCCPSolver::visitSelectInst(SelectInst &I) { 877 // If this select returns a struct, just mark the result overdefined. 878 // TODO: We could do a lot better than this if code actually uses this. 879 if (I.getType()->isStructTy()) 880 return (void)markOverdefined(&I); 881 882 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 883 // discover a concrete value later. 884 if (ValueState[&I].isOverdefined()) 885 return (void)markOverdefined(&I); 886 887 ValueLatticeElement CondValue = getValueState(I.getCondition()); 888 if (CondValue.isUnknownOrUndef()) 889 return; 890 891 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 892 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 893 mergeInValue(&I, getValueState(OpVal)); 894 return; 895 } 896 897 // Otherwise, the condition is overdefined or a constant we can't evaluate. 898 // See if we can produce something better than overdefined based on the T/F 899 // value. 900 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 901 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 902 903 bool Changed = ValueState[&I].mergeIn(TVal); 904 Changed |= ValueState[&I].mergeIn(FVal); 905 if (Changed) 906 pushToWorkListMsg(ValueState[&I], &I); 907 } 908 909 // Handle Unary Operators. 910 void SCCPSolver::visitUnaryOperator(Instruction &I) { 911 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 912 913 ValueLatticeElement &IV = ValueState[&I]; 914 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 915 // discover a concrete value later. 916 if (isOverdefined(IV)) 917 return (void)markOverdefined(&I); 918 919 if (isConstant(V0State)) { 920 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 921 922 // op Y -> undef. 923 if (isa<UndefValue>(C)) 924 return; 925 return (void)markConstant(IV, &I, C); 926 } 927 928 // If something is undef, wait for it to resolve. 929 if (!isOverdefined(V0State)) 930 return; 931 932 markOverdefined(&I); 933 } 934 935 // Handle Binary Operators. 936 void SCCPSolver::visitBinaryOperator(Instruction &I) { 937 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 938 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 939 940 ValueLatticeElement &IV = ValueState[&I]; 941 if (IV.isOverdefined()) 942 return; 943 944 // If something is undef, wait for it to resolve. 945 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 946 return; 947 948 if (V1State.isOverdefined() && V2State.isOverdefined()) 949 return (void)markOverdefined(&I); 950 951 // If either of the operands is a constant, try to fold it to a constant. 952 // TODO: Use information from notconstant better. 953 if ((V1State.isConstant() || V2State.isConstant())) { 954 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 955 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 956 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 957 auto *C = dyn_cast_or_null<Constant>(R); 958 if (C) { 959 // X op Y -> undef. 960 if (isa<UndefValue>(C)) 961 return; 962 // Conservatively assume that the result may be based on operands that may 963 // be undef. Note that we use mergeInValue to combine the constant with 964 // the existing lattice value for I, as different constants might be found 965 // after one of the operands go to overdefined, e.g. due to one operand 966 // being a special floating value. 967 ValueLatticeElement NewV; 968 NewV.markConstant(C, /*MayIncludeUndef=*/true); 969 return (void)mergeInValue(&I, NewV); 970 } 971 } 972 973 // Only use ranges for binary operators on integers. 974 if (!I.getType()->isIntegerTy()) 975 return markOverdefined(&I); 976 977 // Try to simplify to a constant range. 978 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 979 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 980 if (V1State.isConstantRange()) 981 A = V1State.getConstantRange(); 982 if (V2State.isConstantRange()) 983 B = V2State.getConstantRange(); 984 985 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 986 mergeInValue(&I, ValueLatticeElement::getRange(R)); 987 988 // TODO: Currently we do not exploit special values that produce something 989 // better than overdefined with an overdefined operand for vector or floating 990 // point types, like and <4 x i32> overdefined, zeroinitializer. 991 } 992 993 // Handle ICmpInst instruction. 994 void SCCPSolver::visitCmpInst(CmpInst &I) { 995 // Do not cache this lookup, getValueState calls later in the function might 996 // invalidate the reference. 997 if (isOverdefined(ValueState[&I])) 998 return (void)markOverdefined(&I); 999 1000 Value *Op1 = I.getOperand(0); 1001 Value *Op2 = I.getOperand(1); 1002 1003 // For parameters, use ParamState which includes constant range info if 1004 // available. 1005 auto V1State = getValueState(Op1); 1006 auto V2State = getValueState(Op2); 1007 1008 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1009 if (C) { 1010 if (isa<UndefValue>(C)) 1011 return; 1012 ValueLatticeElement CV; 1013 CV.markConstant(C); 1014 mergeInValue(&I, CV); 1015 return; 1016 } 1017 1018 // If operands are still unknown, wait for it to resolve. 1019 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1020 !isConstant(ValueState[&I])) 1021 return; 1022 1023 markOverdefined(&I); 1024 } 1025 1026 // Handle getelementptr instructions. If all operands are constants then we 1027 // can turn this into a getelementptr ConstantExpr. 1028 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1029 if (isOverdefined(ValueState[&I])) 1030 return (void)markOverdefined(&I); 1031 1032 SmallVector<Constant*, 8> Operands; 1033 Operands.reserve(I.getNumOperands()); 1034 1035 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1036 ValueLatticeElement State = getValueState(I.getOperand(i)); 1037 if (State.isUnknownOrUndef()) 1038 return; // Operands are not resolved yet. 1039 1040 if (isOverdefined(State)) 1041 return (void)markOverdefined(&I); 1042 1043 if (Constant *C = getConstant(State)) { 1044 Operands.push_back(C); 1045 continue; 1046 } 1047 1048 return (void)markOverdefined(&I); 1049 } 1050 1051 Constant *Ptr = Operands[0]; 1052 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1053 Constant *C = 1054 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1055 if (isa<UndefValue>(C)) 1056 return; 1057 markConstant(&I, C); 1058 } 1059 1060 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1061 // If this store is of a struct, ignore it. 1062 if (SI.getOperand(0)->getType()->isStructTy()) 1063 return; 1064 1065 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1066 return; 1067 1068 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1069 auto I = TrackedGlobals.find(GV); 1070 if (I == TrackedGlobals.end()) 1071 return; 1072 1073 // Get the value we are storing into the global, then merge it. 1074 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1075 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1076 if (I->second.isOverdefined()) 1077 TrackedGlobals.erase(I); // No need to keep tracking this! 1078 } 1079 1080 // Handle load instructions. If the operand is a constant pointer to a constant 1081 // global, we can replace the load with the loaded constant value! 1082 void SCCPSolver::visitLoadInst(LoadInst &I) { 1083 // If this load is of a struct, just mark the result overdefined. 1084 if (I.getType()->isStructTy()) 1085 return (void)markOverdefined(&I); 1086 1087 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1088 // discover a concrete value later. 1089 if (ValueState[&I].isOverdefined()) 1090 return (void)markOverdefined(&I); 1091 1092 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1093 if (PtrVal.isUnknownOrUndef()) 1094 return; // The pointer is not resolved yet! 1095 1096 ValueLatticeElement &IV = ValueState[&I]; 1097 1098 if (!isConstant(PtrVal) || I.isVolatile()) 1099 return (void)markOverdefined(IV, &I); 1100 1101 Constant *Ptr = getConstant(PtrVal); 1102 1103 // load null is undefined. 1104 if (isa<ConstantPointerNull>(Ptr)) { 1105 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1106 return (void)markOverdefined(IV, &I); 1107 else 1108 return; 1109 } 1110 1111 // Transform load (constant global) into the value loaded. 1112 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1113 if (!TrackedGlobals.empty()) { 1114 // If we are tracking this global, merge in the known value for it. 1115 auto It = TrackedGlobals.find(GV); 1116 if (It != TrackedGlobals.end()) { 1117 mergeInValue(IV, &I, It->second, 1118 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1119 return; 1120 } 1121 } 1122 } 1123 1124 // Transform load from a constant into a constant if possible. 1125 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1126 if (isa<UndefValue>(C)) 1127 return; 1128 return (void)markConstant(IV, &I, C); 1129 } 1130 1131 // Otherwise we cannot say for certain what value this load will produce. 1132 // Bail out. 1133 markOverdefined(IV, &I); 1134 } 1135 1136 void SCCPSolver::visitCallBase(CallBase &CB) { 1137 handleCallResult(CB); 1138 handleCallArguments(CB); 1139 } 1140 1141 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1142 Function *F = CB.getCalledFunction(); 1143 1144 // Void return and not tracking callee, just bail. 1145 if (CB.getType()->isVoidTy()) 1146 return; 1147 1148 // Otherwise, if we have a single return value case, and if the function is 1149 // a declaration, maybe we can constant fold it. 1150 if (F && F->isDeclaration() && !CB.getType()->isStructTy() && 1151 canConstantFoldCallTo(&CB, F)) { 1152 SmallVector<Constant *, 8> Operands; 1153 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1154 if (AI->get()->getType()->isStructTy()) 1155 return markOverdefined(&CB); // Can't handle struct args. 1156 ValueLatticeElement State = getValueState(*AI); 1157 1158 if (State.isUnknownOrUndef()) 1159 return; // Operands are not resolved yet. 1160 if (isOverdefined(State)) 1161 return (void)markOverdefined(&CB); 1162 assert(isConstant(State) && "Unknown state!"); 1163 Operands.push_back(getConstant(State)); 1164 } 1165 1166 if (isOverdefined(getValueState(&CB))) 1167 return (void)markOverdefined(&CB); 1168 1169 // If we can constant fold this, mark the result of the call as a 1170 // constant. 1171 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1172 // call -> undef. 1173 if (isa<UndefValue>(C)) 1174 return; 1175 return (void)markConstant(&CB, C); 1176 } 1177 } 1178 1179 // Otherwise, we don't know anything about this call, mark it overdefined. 1180 return (void)markOverdefined(&CB); 1181 } 1182 1183 void SCCPSolver::handleCallArguments(CallBase &CB) { 1184 Function *F = CB.getCalledFunction(); 1185 // If this is a local function that doesn't have its address taken, mark its 1186 // entry block executable and merge in the actual arguments to the call into 1187 // the formal arguments of the function. 1188 if (!TrackingIncomingArguments.empty() && 1189 TrackingIncomingArguments.count(F)) { 1190 MarkBlockExecutable(&F->front()); 1191 1192 // Propagate information from this call site into the callee. 1193 auto CAI = CB.arg_begin(); 1194 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1195 ++AI, ++CAI) { 1196 // If this argument is byval, and if the function is not readonly, there 1197 // will be an implicit copy formed of the input aggregate. 1198 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1199 markOverdefined(&*AI); 1200 continue; 1201 } 1202 1203 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1204 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1205 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1206 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg); 1207 } 1208 } else 1209 mergeInValue(&*AI, getValueState(*CAI), 1210 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1211 } 1212 } 1213 } 1214 1215 void SCCPSolver::handleCallResult(CallBase &CB) { 1216 Function *F = CB.getCalledFunction(); 1217 1218 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1219 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1220 if (ValueState[&CB].isOverdefined()) 1221 return; 1222 1223 Value *CopyOf = CB.getOperand(0); 1224 auto *PI = getPredicateInfoFor(&CB); 1225 auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI); 1226 if (!PI || !PBranch) { 1227 mergeInValue(ValueState[&CB], &CB, getValueState(CopyOf)); 1228 return; 1229 } 1230 1231 // Everything below relies on the condition being a comparison. 1232 auto *Cmp = dyn_cast<CmpInst>(PBranch->Condition); 1233 if (!Cmp) { 1234 mergeInValue(ValueState[&CB], &CB, getValueState(CopyOf)); 1235 return; 1236 } 1237 1238 Value *CmpOp0 = Cmp->getOperand(0); 1239 Value *CmpOp1 = Cmp->getOperand(1); 1240 if (CopyOf != CmpOp0 && CopyOf != CmpOp1) { 1241 mergeInValue(ValueState[&CB], &CB, getValueState(CopyOf)); 1242 return; 1243 } 1244 1245 auto Pred = Cmp->getPredicate(); 1246 if (CmpOp0 != CopyOf) { 1247 std::swap(CmpOp0, CmpOp1); 1248 Pred = Cmp->getSwappedPredicate(); 1249 } 1250 1251 // Wait until CmpOp1 is resolved. 1252 if (getValueState(CmpOp1).isUnknown()) { 1253 addAdditionalUser(CmpOp1, &CB); 1254 return; 1255 } 1256 1257 if (!PBranch->TrueEdge) 1258 Pred = CmpInst::getInversePredicate(Pred); 1259 1260 ValueLatticeElement CondVal = getValueState(CmpOp1); 1261 ValueLatticeElement &IV = ValueState[&CB]; 1262 ValueLatticeElement OriginalVal = getValueState(CopyOf); 1263 if (CondVal.isConstantRange() || OriginalVal.isConstantRange()) { 1264 auto NewCR = 1265 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1266 1267 // Get the range imposed by the condition. 1268 if (CondVal.isConstantRange()) 1269 NewCR = ConstantRange::makeAllowedICmpRegion( 1270 Pred, CondVal.getConstantRange()); 1271 1272 // Combine range info for the original value with the new range from the 1273 // condition. 1274 auto OriginalCR = OriginalVal.isConstantRange() 1275 ? OriginalVal.getConstantRange() 1276 : ConstantRange::getFull( 1277 DL.getTypeSizeInBits(CopyOf->getType())); 1278 NewCR = NewCR.intersectWith(OriginalCR); 1279 1280 addAdditionalUser(CmpOp1, &CB); 1281 // TODO: Actually filp MayIncludeUndef for the created range to false, 1282 // once most places in the optimizer respect the branches on 1283 // undef/poison are UB rule. The reason why the new range cannot be 1284 // undef is as follows below: 1285 // The new range is based on a branch condition. That guarantees that 1286 // neither of the compare operands can be undef in the branch targets, 1287 // unless we have conditions that are always true/false (e.g. icmp ule 1288 // i32, %a, i32_max). For the latter overdefined/empty range will be 1289 // inferred, but the branch will get folded accordingly anyways. 1290 mergeInValue( 1291 IV, &CB, 1292 ValueLatticeElement::getRange(NewCR, /*MayIncludeUndef=*/true)); 1293 return; 1294 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1295 // For non-integer values or integer constant expressions, only 1296 // propagate equal constants. 1297 addAdditionalUser(CmpOp1, &CB); 1298 mergeInValue(IV, &CB, CondVal); 1299 return; 1300 } 1301 1302 return (void)mergeInValue(IV, &CB, getValueState(CopyOf)); 1303 } 1304 } 1305 1306 // The common case is that we aren't tracking the callee, either because we 1307 // are not doing interprocedural analysis or the callee is indirect, or is 1308 // external. Handle these cases first. 1309 if (!F || F->isDeclaration()) 1310 return handleCallOverdefined(CB); 1311 1312 // If this is a single/zero retval case, see if we're tracking the function. 1313 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1314 if (!MRVFunctionsTracked.count(F)) 1315 return handleCallOverdefined(CB); // Not tracking this callee. 1316 1317 // If we are tracking this callee, propagate the result of the function 1318 // into this call site. 1319 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1320 mergeInValue(getStructValueState(&CB, i), &CB, 1321 TrackedMultipleRetVals[std::make_pair(F, i)]); 1322 } else { 1323 auto TFRVI = TrackedRetVals.find(F); 1324 if (TFRVI == TrackedRetVals.end()) 1325 return handleCallOverdefined(CB); // Not tracking this callee. 1326 1327 // If so, propagate the return value of the callee into this call result. 1328 mergeInValue(&CB, TFRVI->second); 1329 } 1330 } 1331 1332 void SCCPSolver::Solve() { 1333 // Process the work lists until they are empty! 1334 while (!BBWorkList.empty() || !InstWorkList.empty() || 1335 !OverdefinedInstWorkList.empty()) { 1336 // Process the overdefined instruction's work list first, which drives other 1337 // things to overdefined more quickly. 1338 while (!OverdefinedInstWorkList.empty()) { 1339 Value *I = OverdefinedInstWorkList.pop_back_val(); 1340 1341 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1342 1343 // "I" got into the work list because it either made the transition from 1344 // bottom to constant, or to overdefined. 1345 // 1346 // Anything on this worklist that is overdefined need not be visited 1347 // since all of its users will have already been marked as overdefined 1348 // Update all of the users of this instruction's value. 1349 // 1350 markUsersAsChanged(I); 1351 } 1352 1353 // Process the instruction work list. 1354 while (!InstWorkList.empty()) { 1355 Value *I = InstWorkList.pop_back_val(); 1356 1357 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1358 1359 // "I" got into the work list because it made the transition from undef to 1360 // constant. 1361 // 1362 // Anything on this worklist that is overdefined need not be visited 1363 // since all of its users will have already been marked as overdefined. 1364 // Update all of the users of this instruction's value. 1365 // 1366 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1367 markUsersAsChanged(I); 1368 } 1369 1370 // Process the basic block work list. 1371 while (!BBWorkList.empty()) { 1372 BasicBlock *BB = BBWorkList.back(); 1373 BBWorkList.pop_back(); 1374 1375 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1376 1377 // Notify all instructions in this basic block that they are newly 1378 // executable. 1379 visit(BB); 1380 } 1381 } 1382 } 1383 1384 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1385 /// that branches on undef values cannot reach any of their successors. 1386 /// However, this is not a safe assumption. After we solve dataflow, this 1387 /// method should be use to handle this. If this returns true, the solver 1388 /// should be rerun. 1389 /// 1390 /// This method handles this by finding an unresolved branch and marking it one 1391 /// of the edges from the block as being feasible, even though the condition 1392 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1393 /// CFG and only slightly pessimizes the analysis results (by marking one, 1394 /// potentially infeasible, edge feasible). This cannot usefully modify the 1395 /// constraints on the condition of the branch, as that would impact other users 1396 /// of the value. 1397 /// 1398 /// This scan also checks for values that use undefs. It conservatively marks 1399 /// them as overdefined. 1400 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1401 for (BasicBlock &BB : F) { 1402 if (!BBExecutable.count(&BB)) 1403 continue; 1404 1405 for (Instruction &I : BB) { 1406 // Look for instructions which produce undef values. 1407 if (I.getType()->isVoidTy()) continue; 1408 1409 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1410 // Only a few things that can be structs matter for undef. 1411 1412 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1413 if (auto *CB = dyn_cast<CallBase>(&I)) 1414 if (Function *F = CB->getCalledFunction()) 1415 if (MRVFunctionsTracked.count(F)) 1416 continue; 1417 1418 // extractvalue and insertvalue don't need to be marked; they are 1419 // tracked as precisely as their operands. 1420 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1421 continue; 1422 // Send the results of everything else to overdefined. We could be 1423 // more precise than this but it isn't worth bothering. 1424 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1425 ValueLatticeElement &LV = getStructValueState(&I, i); 1426 if (LV.isUnknownOrUndef()) 1427 markOverdefined(LV, &I); 1428 } 1429 continue; 1430 } 1431 1432 ValueLatticeElement &LV = getValueState(&I); 1433 if (!LV.isUnknownOrUndef()) 1434 continue; 1435 1436 // There are two reasons a call can have an undef result 1437 // 1. It could be tracked. 1438 // 2. It could be constant-foldable. 1439 // Because of the way we solve return values, tracked calls must 1440 // never be marked overdefined in ResolvedUndefsIn. 1441 if (auto *CB = dyn_cast<CallBase>(&I)) 1442 if (Function *F = CB->getCalledFunction()) 1443 if (TrackedRetVals.count(F)) 1444 continue; 1445 1446 if (isa<LoadInst>(I)) { 1447 // A load here means one of two things: a load of undef from a global, 1448 // a load from an unknown pointer. Either way, having it return undef 1449 // is okay. 1450 continue; 1451 } 1452 1453 markOverdefined(&I); 1454 return true; 1455 } 1456 1457 // Check to see if we have a branch or switch on an undefined value. If so 1458 // we force the branch to go one way or the other to make the successor 1459 // values live. It doesn't really matter which way we force it. 1460 Instruction *TI = BB.getTerminator(); 1461 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1462 if (!BI->isConditional()) continue; 1463 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1464 continue; 1465 1466 // If the input to SCCP is actually branch on undef, fix the undef to 1467 // false. 1468 if (isa<UndefValue>(BI->getCondition())) { 1469 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1470 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1471 return true; 1472 } 1473 1474 // Otherwise, it is a branch on a symbolic value which is currently 1475 // considered to be undef. Make sure some edge is executable, so a 1476 // branch on "undef" always flows somewhere. 1477 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1478 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1479 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1480 return true; 1481 1482 continue; 1483 } 1484 1485 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1486 // Indirect branch with no successor ?. Its ok to assume it branches 1487 // to no target. 1488 if (IBR->getNumSuccessors() < 1) 1489 continue; 1490 1491 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1492 continue; 1493 1494 // If the input to SCCP is actually branch on undef, fix the undef to 1495 // the first successor of the indirect branch. 1496 if (isa<UndefValue>(IBR->getAddress())) { 1497 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1498 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1499 return true; 1500 } 1501 1502 // Otherwise, it is a branch on a symbolic value which is currently 1503 // considered to be undef. Make sure some edge is executable, so a 1504 // branch on "undef" always flows somewhere. 1505 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1506 // we can assume the branch has undefined behavior instead. 1507 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1508 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1509 return true; 1510 1511 continue; 1512 } 1513 1514 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1515 if (!SI->getNumCases() || 1516 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1517 continue; 1518 1519 // If the input to SCCP is actually switch on undef, fix the undef to 1520 // the first constant. 1521 if (isa<UndefValue>(SI->getCondition())) { 1522 SI->setCondition(SI->case_begin()->getCaseValue()); 1523 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1524 return true; 1525 } 1526 1527 // Otherwise, it is a branch on a symbolic value which is currently 1528 // considered to be undef. Make sure some edge is executable, so a 1529 // branch on "undef" always flows somewhere. 1530 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1531 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1532 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1533 return true; 1534 1535 continue; 1536 } 1537 } 1538 1539 return false; 1540 } 1541 1542 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1543 Constant *Const = nullptr; 1544 if (V->getType()->isStructTy()) { 1545 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1546 if (any_of(IVs, 1547 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1548 return false; 1549 std::vector<Constant *> ConstVals; 1550 auto *ST = cast<StructType>(V->getType()); 1551 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1552 ValueLatticeElement V = IVs[i]; 1553 ConstVals.push_back(isConstant(V) 1554 ? Solver.getConstant(V) 1555 : UndefValue::get(ST->getElementType(i))); 1556 } 1557 Const = ConstantStruct::get(ST, ConstVals); 1558 } else { 1559 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1560 if (isOverdefined(IV)) 1561 return false; 1562 1563 Const = 1564 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1565 } 1566 assert(Const && "Constant is nullptr here!"); 1567 1568 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1569 // unless the call itself can be removed 1570 CallInst *CI = dyn_cast<CallInst>(V); 1571 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1572 Function *F = CI->getCalledFunction(); 1573 1574 // Don't zap returns of the callee 1575 if (F) 1576 Solver.AddMustTailCallee(F); 1577 1578 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1579 << " as a constant\n"); 1580 return false; 1581 } 1582 1583 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1584 1585 // Replaces all of the uses of a variable with uses of the constant. 1586 V->replaceAllUsesWith(Const); 1587 return true; 1588 } 1589 1590 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1591 // and return true if the function was modified. 1592 static bool runSCCP(Function &F, const DataLayout &DL, 1593 const TargetLibraryInfo *TLI) { 1594 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1595 SCCPSolver Solver( 1596 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1597 F.getContext()); 1598 1599 // Mark the first block of the function as being executable. 1600 Solver.MarkBlockExecutable(&F.front()); 1601 1602 // Mark all arguments to the function as being overdefined. 1603 for (Argument &AI : F.args()) 1604 Solver.markOverdefined(&AI); 1605 1606 // Solve for constants. 1607 bool ResolvedUndefs = true; 1608 while (ResolvedUndefs) { 1609 Solver.Solve(); 1610 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1611 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1612 } 1613 1614 bool MadeChanges = false; 1615 1616 // If we decided that there are basic blocks that are dead in this function, 1617 // delete their contents now. Note that we cannot actually delete the blocks, 1618 // as we cannot modify the CFG of the function. 1619 1620 for (BasicBlock &BB : F) { 1621 if (!Solver.isBlockExecutable(&BB)) { 1622 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1623 1624 ++NumDeadBlocks; 1625 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB); 1626 1627 MadeChanges = true; 1628 continue; 1629 } 1630 1631 // Iterate over all of the instructions in a function, replacing them with 1632 // constants if we have found them to be of constant values. 1633 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1634 Instruction *Inst = &*BI++; 1635 if (Inst->getType()->isVoidTy() || Inst->isTerminator()) 1636 continue; 1637 1638 if (tryToReplaceWithConstant(Solver, Inst)) { 1639 if (isInstructionTriviallyDead(Inst)) 1640 Inst->eraseFromParent(); 1641 // Hey, we just changed something! 1642 MadeChanges = true; 1643 ++NumInstRemoved; 1644 } 1645 } 1646 } 1647 1648 return MadeChanges; 1649 } 1650 1651 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1652 const DataLayout &DL = F.getParent()->getDataLayout(); 1653 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1654 if (!runSCCP(F, DL, &TLI)) 1655 return PreservedAnalyses::all(); 1656 1657 auto PA = PreservedAnalyses(); 1658 PA.preserve<GlobalsAA>(); 1659 PA.preserveSet<CFGAnalyses>(); 1660 return PA; 1661 } 1662 1663 namespace { 1664 1665 //===--------------------------------------------------------------------===// 1666 // 1667 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1668 /// Sparse Conditional Constant Propagator. 1669 /// 1670 class SCCPLegacyPass : public FunctionPass { 1671 public: 1672 // Pass identification, replacement for typeid 1673 static char ID; 1674 1675 SCCPLegacyPass() : FunctionPass(ID) { 1676 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1677 } 1678 1679 void getAnalysisUsage(AnalysisUsage &AU) const override { 1680 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1681 AU.addPreserved<GlobalsAAWrapperPass>(); 1682 AU.setPreservesCFG(); 1683 } 1684 1685 // runOnFunction - Run the Sparse Conditional Constant Propagation 1686 // algorithm, and return true if the function was modified. 1687 bool runOnFunction(Function &F) override { 1688 if (skipFunction(F)) 1689 return false; 1690 const DataLayout &DL = F.getParent()->getDataLayout(); 1691 const TargetLibraryInfo *TLI = 1692 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1693 return runSCCP(F, DL, TLI); 1694 } 1695 }; 1696 1697 } // end anonymous namespace 1698 1699 char SCCPLegacyPass::ID = 0; 1700 1701 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1702 "Sparse Conditional Constant Propagation", false, false) 1703 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1704 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1705 "Sparse Conditional Constant Propagation", false, false) 1706 1707 // createSCCPPass - This is the public interface to this file. 1708 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1709 1710 static void findReturnsToZap(Function &F, 1711 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1712 SCCPSolver &Solver) { 1713 // We can only do this if we know that nothing else can call the function. 1714 if (!Solver.isArgumentTrackedFunction(&F)) 1715 return; 1716 1717 // There is a non-removable musttail call site of this function. Zapping 1718 // returns is not allowed. 1719 if (Solver.isMustTailCallee(&F)) { 1720 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1721 << " due to present musttail call of it\n"); 1722 return; 1723 } 1724 1725 assert( 1726 all_of(F.users(), 1727 [&Solver](User *U) { 1728 if (isa<Instruction>(U) && 1729 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1730 return true; 1731 // Non-callsite uses are not impacted by zapping. Also, constant 1732 // uses (like blockaddresses) could stuck around, without being 1733 // used in the underlying IR, meaning we do not have lattice 1734 // values for them. 1735 if (!isa<CallBase>(U)) 1736 return true; 1737 if (U->getType()->isStructTy()) { 1738 return all_of(Solver.getStructLatticeValueFor(U), 1739 [](const ValueLatticeElement &LV) { 1740 return !isOverdefined(LV); 1741 }); 1742 } 1743 return !isOverdefined(Solver.getLatticeValueFor(U)); 1744 }) && 1745 "We can only zap functions where all live users have a concrete value"); 1746 1747 for (BasicBlock &BB : F) { 1748 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1749 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1750 << "musttail call : " << *CI << "\n"); 1751 (void)CI; 1752 return; 1753 } 1754 1755 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1756 if (!isa<UndefValue>(RI->getOperand(0))) 1757 ReturnsToZap.push_back(RI); 1758 } 1759 } 1760 1761 // Update the condition for terminators that are branching on indeterminate 1762 // values, forcing them to use a specific edge. 1763 static void forceIndeterminateEdge(Instruction* I, SCCPSolver &Solver) { 1764 BasicBlock *Dest = nullptr; 1765 Constant *C = nullptr; 1766 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1767 if (!isa<ConstantInt>(SI->getCondition())) { 1768 // Indeterminate switch; use first case value. 1769 Dest = SI->case_begin()->getCaseSuccessor(); 1770 C = SI->case_begin()->getCaseValue(); 1771 } 1772 } else if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1773 if (!isa<ConstantInt>(BI->getCondition())) { 1774 // Indeterminate branch; use false. 1775 Dest = BI->getSuccessor(1); 1776 C = ConstantInt::getFalse(BI->getContext()); 1777 } 1778 } else if (IndirectBrInst *IBR = dyn_cast<IndirectBrInst>(I)) { 1779 if (!isa<BlockAddress>(IBR->getAddress()->stripPointerCasts())) { 1780 // Indeterminate indirectbr; use successor 0. 1781 Dest = IBR->getSuccessor(0); 1782 C = BlockAddress::get(IBR->getSuccessor(0)); 1783 } 1784 } else { 1785 llvm_unreachable("Unexpected terminator instruction"); 1786 } 1787 if (C) { 1788 assert(Solver.isEdgeFeasible(I->getParent(), Dest) && 1789 "Didn't find feasible edge?"); 1790 (void)Dest; 1791 1792 I->setOperand(0, C); 1793 } 1794 } 1795 1796 bool llvm::runIPSCCP( 1797 Module &M, const DataLayout &DL, 1798 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1799 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1800 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1801 1802 // Loop over all functions, marking arguments to those with their addresses 1803 // taken or that are external as overdefined. 1804 for (Function &F : M) { 1805 if (F.isDeclaration()) 1806 continue; 1807 1808 Solver.addAnalysis(F, getAnalysis(F)); 1809 1810 // Determine if we can track the function's return values. If so, add the 1811 // function to the solver's set of return-tracked functions. 1812 if (canTrackReturnsInterprocedurally(&F)) 1813 Solver.AddTrackedFunction(&F); 1814 1815 // Determine if we can track the function's arguments. If so, add the 1816 // function to the solver's set of argument-tracked functions. 1817 if (canTrackArgumentsInterprocedurally(&F)) { 1818 Solver.AddArgumentTrackedFunction(&F); 1819 continue; 1820 } 1821 1822 // Assume the function is called. 1823 Solver.MarkBlockExecutable(&F.front()); 1824 1825 // Assume nothing about the incoming arguments. 1826 for (Argument &AI : F.args()) 1827 Solver.markOverdefined(&AI); 1828 } 1829 1830 // Determine if we can track any of the module's global variables. If so, add 1831 // the global variables we can track to the solver's set of tracked global 1832 // variables. 1833 for (GlobalVariable &G : M.globals()) { 1834 G.removeDeadConstantUsers(); 1835 if (canTrackGlobalVariableInterprocedurally(&G)) 1836 Solver.TrackValueOfGlobalVariable(&G); 1837 } 1838 1839 // Solve for constants. 1840 bool ResolvedUndefs = true; 1841 Solver.Solve(); 1842 while (ResolvedUndefs) { 1843 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1844 ResolvedUndefs = false; 1845 for (Function &F : M) 1846 if (Solver.ResolvedUndefsIn(F)) { 1847 // We run Solve() after we resolved an undef in a function, because 1848 // we might deduce a fact that eliminates an undef in another function. 1849 Solver.Solve(); 1850 ResolvedUndefs = true; 1851 } 1852 } 1853 1854 bool MadeChanges = false; 1855 1856 // Iterate over all of the instructions in the module, replacing them with 1857 // constants if we have found them to be of constant values. 1858 1859 for (Function &F : M) { 1860 if (F.isDeclaration()) 1861 continue; 1862 1863 SmallVector<BasicBlock *, 512> BlocksToErase; 1864 1865 if (Solver.isBlockExecutable(&F.front())) 1866 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E; 1867 ++AI) { 1868 if (!AI->use_empty() && tryToReplaceWithConstant(Solver, &*AI)) { 1869 ++IPNumArgsElimed; 1870 continue; 1871 } 1872 } 1873 1874 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 1875 if (!Solver.isBlockExecutable(&*BB)) { 1876 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB); 1877 ++NumDeadBlocks; 1878 1879 MadeChanges = true; 1880 1881 if (&*BB != &F.front()) 1882 BlocksToErase.push_back(&*BB); 1883 continue; 1884 } 1885 1886 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1887 Instruction *Inst = &*BI++; 1888 if (Inst->getType()->isVoidTy()) 1889 continue; 1890 if (tryToReplaceWithConstant(Solver, Inst)) { 1891 if (Inst->isSafeToRemove()) 1892 Inst->eraseFromParent(); 1893 // Hey, we just changed something! 1894 MadeChanges = true; 1895 ++IPNumInstRemoved; 1896 } 1897 } 1898 } 1899 1900 DomTreeUpdater DTU = Solver.getDTU(F); 1901 // Change dead blocks to unreachable. We do it after replacing constants 1902 // in all executable blocks, because changeToUnreachable may remove PHI 1903 // nodes in executable blocks we found values for. The function's entry 1904 // block is not part of BlocksToErase, so we have to handle it separately. 1905 for (BasicBlock *BB : BlocksToErase) { 1906 NumInstRemoved += 1907 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 1908 /*PreserveLCSSA=*/false, &DTU); 1909 } 1910 if (!Solver.isBlockExecutable(&F.front())) 1911 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 1912 /*UseLLVMTrap=*/false, 1913 /*PreserveLCSSA=*/false, &DTU); 1914 1915 // Now that all instructions in the function are constant folded, 1916 // use ConstantFoldTerminator to get rid of in-edges, record DT updates and 1917 // delete dead BBs. 1918 for (BasicBlock *DeadBB : BlocksToErase) { 1919 // If there are any PHI nodes in this successor, drop entries for BB now. 1920 for (Value::user_iterator UI = DeadBB->user_begin(), 1921 UE = DeadBB->user_end(); 1922 UI != UE;) { 1923 // Grab the user and then increment the iterator early, as the user 1924 // will be deleted. Step past all adjacent uses from the same user. 1925 auto *I = dyn_cast<Instruction>(*UI); 1926 do { ++UI; } while (UI != UE && *UI == I); 1927 1928 // Ignore blockaddress users; BasicBlock's dtor will handle them. 1929 if (!I) continue; 1930 1931 // If we have forced an edge for an indeterminate value, then force the 1932 // terminator to fold to that edge. 1933 forceIndeterminateEdge(I, Solver); 1934 BasicBlock *InstBB = I->getParent(); 1935 bool Folded = ConstantFoldTerminator(InstBB, 1936 /*DeleteDeadConditions=*/false, 1937 /*TLI=*/nullptr, &DTU); 1938 assert(Folded && 1939 "Expect TermInst on constantint or blockaddress to be folded"); 1940 (void) Folded; 1941 // If we folded the terminator to an unconditional branch to another 1942 // dead block, replace it with Unreachable, to avoid trying to fold that 1943 // branch again. 1944 BranchInst *BI = cast<BranchInst>(InstBB->getTerminator()); 1945 if (BI && BI->isUnconditional() && 1946 !Solver.isBlockExecutable(BI->getSuccessor(0))) { 1947 InstBB->getTerminator()->eraseFromParent(); 1948 new UnreachableInst(InstBB->getContext(), InstBB); 1949 } 1950 } 1951 // Mark dead BB for deletion. 1952 DTU.deleteBB(DeadBB); 1953 } 1954 1955 for (BasicBlock &BB : F) { 1956 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1957 Instruction *Inst = &*BI++; 1958 if (Solver.getPredicateInfoFor(Inst)) { 1959 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 1960 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1961 Value *Op = II->getOperand(0); 1962 Inst->replaceAllUsesWith(Op); 1963 Inst->eraseFromParent(); 1964 } 1965 } 1966 } 1967 } 1968 } 1969 } 1970 1971 // If we inferred constant or undef return values for a function, we replaced 1972 // all call uses with the inferred value. This means we don't need to bother 1973 // actually returning anything from the function. Replace all return 1974 // instructions with return undef. 1975 // 1976 // Do this in two stages: first identify the functions we should process, then 1977 // actually zap their returns. This is important because we can only do this 1978 // if the address of the function isn't taken. In cases where a return is the 1979 // last use of a function, the order of processing functions would affect 1980 // whether other functions are optimizable. 1981 SmallVector<ReturnInst*, 8> ReturnsToZap; 1982 1983 for (const auto &I : Solver.getTrackedRetVals()) { 1984 Function *F = I.first; 1985 if (isOverdefined(I.second) || F->getReturnType()->isVoidTy()) 1986 continue; 1987 findReturnsToZap(*F, ReturnsToZap, Solver); 1988 } 1989 1990 for (auto F : Solver.getMRVFunctionsTracked()) { 1991 assert(F->getReturnType()->isStructTy() && 1992 "The return type should be a struct"); 1993 StructType *STy = cast<StructType>(F->getReturnType()); 1994 if (Solver.isStructLatticeConstant(F, STy)) 1995 findReturnsToZap(*F, ReturnsToZap, Solver); 1996 } 1997 1998 // Zap all returns which we've identified as zap to change. 1999 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2000 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2001 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2002 } 2003 2004 // If we inferred constant or undef values for globals variables, we can 2005 // delete the global and any stores that remain to it. 2006 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2007 GlobalVariable *GV = I.first; 2008 if (isOverdefined(I.second)) 2009 continue; 2010 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2011 << "' is constant!\n"); 2012 while (!GV->use_empty()) { 2013 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2014 SI->eraseFromParent(); 2015 } 2016 M.getGlobalList().erase(GV); 2017 ++IPNumGlobalConst; 2018 } 2019 2020 return MadeChanges; 2021 } 2022