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 // The maximum number of range extensions allowed for operations requiring 76 // widening. 77 static const unsigned MaxNumRangeExtensions = 10; 78 79 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 80 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 81 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 82 MaxNumRangeExtensions); 83 } 84 namespace { 85 86 // Helper to check if \p LV is either a constant or a constant 87 // range with a single element. This should cover exactly the same cases as the 88 // old ValueLatticeElement::isConstant() and is intended to be used in the 89 // transition to ValueLatticeElement. 90 bool isConstant(const ValueLatticeElement &LV) { 91 return LV.isConstant() || 92 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 93 } 94 95 // Helper to check if \p LV is either overdefined or a constant range with more 96 // than a single element. This should cover exactly the same cases as the old 97 // ValueLatticeElement::isOverdefined() and is intended to be used in the 98 // transition to ValueLatticeElement. 99 bool isOverdefined(const ValueLatticeElement &LV) { 100 return LV.isOverdefined() || 101 (LV.isConstantRange() && !LV.getConstantRange().isSingleElement()); 102 } 103 104 //===----------------------------------------------------------------------===// 105 // 106 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 107 /// Constant Propagation. 108 /// 109 class SCCPSolver : public InstVisitor<SCCPSolver> { 110 const DataLayout &DL; 111 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 112 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 113 DenseMap<Value *, ValueLatticeElement> 114 ValueState; // The state each value is in. 115 116 /// StructValueState - This maintains ValueState for values that have 117 /// StructType, for example for formal arguments, calls, insertelement, etc. 118 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 119 120 /// GlobalValue - If we are tracking any values for the contents of a global 121 /// variable, we keep a mapping from the constant accessor to the element of 122 /// the global, to the currently known value. If the value becomes 123 /// overdefined, it's entry is simply removed from this map. 124 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 125 126 /// TrackedRetVals - If we are tracking arguments into and the return 127 /// value out of a function, it will have an entry in this map, indicating 128 /// what the known return value for the function is. 129 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 130 131 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 132 /// that return multiple values. 133 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 134 TrackedMultipleRetVals; 135 136 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 137 /// represented here for efficient lookup. 138 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 139 140 /// MustTailFunctions - Each function here is a callee of non-removable 141 /// musttail call site. 142 SmallPtrSet<Function *, 16> MustTailCallees; 143 144 /// TrackingIncomingArguments - This is the set of functions for whose 145 /// arguments we make optimistic assumptions about and try to prove as 146 /// constants. 147 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 148 149 /// The reason for two worklists is that overdefined is the lowest state 150 /// on the lattice, and moving things to overdefined as fast as possible 151 /// makes SCCP converge much faster. 152 /// 153 /// By having a separate worklist, we accomplish this because everything 154 /// possibly overdefined will become overdefined at the soonest possible 155 /// point. 156 SmallVector<Value *, 64> OverdefinedInstWorkList; 157 SmallVector<Value *, 64> InstWorkList; 158 159 // The BasicBlock work list 160 SmallVector<BasicBlock *, 64> BBWorkList; 161 162 /// KnownFeasibleEdges - Entries in this set are edges which have already had 163 /// PHI nodes retriggered. 164 using Edge = std::pair<BasicBlock *, BasicBlock *>; 165 DenseSet<Edge> KnownFeasibleEdges; 166 167 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 168 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 169 170 LLVMContext &Ctx; 171 172 public: 173 void addAnalysis(Function &F, AnalysisResultsForFn A) { 174 AnalysisResults.insert({&F, std::move(A)}); 175 } 176 177 const PredicateBase *getPredicateInfoFor(Instruction *I) { 178 auto A = AnalysisResults.find(I->getParent()->getParent()); 179 if (A == AnalysisResults.end()) 180 return nullptr; 181 return A->second.PredInfo->getPredicateInfoFor(I); 182 } 183 184 DomTreeUpdater getDTU(Function &F) { 185 auto A = AnalysisResults.find(&F); 186 assert(A != AnalysisResults.end() && "Need analysis results for function."); 187 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 188 } 189 190 SCCPSolver(const DataLayout &DL, 191 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 192 LLVMContext &Ctx) 193 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 194 195 /// MarkBlockExecutable - This method can be used by clients to mark all of 196 /// the blocks that are known to be intrinsically live in the processed unit. 197 /// 198 /// This returns true if the block was not considered live before. 199 bool MarkBlockExecutable(BasicBlock *BB) { 200 if (!BBExecutable.insert(BB).second) 201 return false; 202 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 203 BBWorkList.push_back(BB); // Add the block to the work list! 204 return true; 205 } 206 207 /// TrackValueOfGlobalVariable - Clients can use this method to 208 /// inform the SCCPSolver that it should track loads and stores to the 209 /// specified global variable if it can. This is only legal to call if 210 /// performing Interprocedural SCCP. 211 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 212 // We only track the contents of scalar globals. 213 if (GV->getValueType()->isSingleValueType()) { 214 ValueLatticeElement &IV = TrackedGlobals[GV]; 215 if (!isa<UndefValue>(GV->getInitializer())) 216 IV.markConstant(GV->getInitializer()); 217 } 218 } 219 220 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 221 /// and out of the specified function (which cannot have its address taken), 222 /// this method must be called. 223 void AddTrackedFunction(Function *F) { 224 // Add an entry, F -> undef. 225 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 226 MRVFunctionsTracked.insert(F); 227 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 228 TrackedMultipleRetVals.insert( 229 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 230 } else 231 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 232 } 233 234 /// AddMustTailCallee - If the SCCP solver finds that this function is called 235 /// from non-removable musttail call site. 236 void AddMustTailCallee(Function *F) { 237 MustTailCallees.insert(F); 238 } 239 240 /// Returns true if the given function is called from non-removable musttail 241 /// call site. 242 bool isMustTailCallee(Function *F) { 243 return MustTailCallees.count(F); 244 } 245 246 void AddArgumentTrackedFunction(Function *F) { 247 TrackingIncomingArguments.insert(F); 248 } 249 250 /// Returns true if the given function is in the solver's set of 251 /// argument-tracked functions. 252 bool isArgumentTrackedFunction(Function *F) { 253 return TrackingIncomingArguments.count(F); 254 } 255 256 /// Solve - Solve for constants and executable blocks. 257 void Solve(); 258 259 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 260 /// that branches on undef values cannot reach any of their successors. 261 /// However, this is not a safe assumption. After we solve dataflow, this 262 /// method should be use to handle this. If this returns true, the solver 263 /// should be rerun. 264 bool ResolvedUndefsIn(Function &F); 265 266 bool isBlockExecutable(BasicBlock *BB) const { 267 return BBExecutable.count(BB); 268 } 269 270 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 271 // block to the 'To' basic block is currently feasible. 272 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 273 274 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 275 std::vector<ValueLatticeElement> StructValues; 276 auto *STy = dyn_cast<StructType>(V->getType()); 277 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 278 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 279 auto I = StructValueState.find(std::make_pair(V, i)); 280 assert(I != StructValueState.end() && "Value not in valuemap!"); 281 StructValues.push_back(I->second); 282 } 283 return StructValues; 284 } 285 286 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 287 assert(!V->getType()->isStructTy() && 288 "Should use getStructLatticeValueFor"); 289 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 290 ValueState.find(V); 291 assert(I != ValueState.end() && 292 "V not found in ValueState nor Paramstate map!"); 293 return I->second; 294 } 295 296 /// getTrackedRetVals - Get the inferred return value map. 297 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 298 return TrackedRetVals; 299 } 300 301 /// getTrackedGlobals - Get and return the set of inferred initializers for 302 /// global variables. 303 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 304 return TrackedGlobals; 305 } 306 307 /// getMRVFunctionsTracked - Get the set of functions which return multiple 308 /// values tracked by the pass. 309 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 310 return MRVFunctionsTracked; 311 } 312 313 /// getMustTailCallees - Get the set of functions which are called 314 /// from non-removable musttail call sites. 315 const SmallPtrSet<Function *, 16> getMustTailCallees() { 316 return MustTailCallees; 317 } 318 319 /// markOverdefined - Mark the specified value overdefined. This 320 /// works with both scalars and structs. 321 void markOverdefined(Value *V) { 322 if (auto *STy = dyn_cast<StructType>(V->getType())) 323 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 324 markOverdefined(getStructValueState(V, i), V); 325 else 326 markOverdefined(ValueState[V], V); 327 } 328 329 // isStructLatticeConstant - Return true if all the lattice values 330 // corresponding to elements of the structure are constants, 331 // false otherwise. 332 bool isStructLatticeConstant(Function *F, StructType *STy) { 333 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 334 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 335 assert(It != TrackedMultipleRetVals.end()); 336 ValueLatticeElement LV = It->second; 337 if (!isConstant(LV)) 338 return false; 339 } 340 return true; 341 } 342 343 /// Helper to return a Constant if \p LV is either a constant or a constant 344 /// range with a single element. 345 Constant *getConstant(const ValueLatticeElement &LV) const { 346 if (LV.isConstant()) 347 return LV.getConstant(); 348 349 if (LV.isConstantRange()) { 350 auto &CR = LV.getConstantRange(); 351 if (CR.getSingleElement()) 352 return ConstantInt::get(Ctx, *CR.getSingleElement()); 353 } 354 return nullptr; 355 } 356 357 private: 358 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 359 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 360 } 361 362 // pushToWorkList - Helper for markConstant/markOverdefined 363 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 364 if (IV.isOverdefined()) 365 return OverdefinedInstWorkList.push_back(V); 366 InstWorkList.push_back(V); 367 } 368 369 // Helper to push \p V to the worklist, after updating it to \p IV. Also 370 // prints a debug message with the updated value. 371 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 372 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 373 pushToWorkList(IV, V); 374 } 375 376 // markConstant - Make a value be marked as "constant". If the value 377 // is not already a constant, add it to the instruction work list so that 378 // the users of the instruction are updated later. 379 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 380 bool MayIncludeUndef = false) { 381 if (!IV.markConstant(C, MayIncludeUndef)) 382 return false; 383 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 384 pushToWorkList(IV, V); 385 return true; 386 } 387 388 bool markConstant(Value *V, Constant *C) { 389 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 390 return markConstant(ValueState[V], V, C); 391 } 392 393 // markOverdefined - Make a value be marked as "overdefined". If the 394 // value is not already overdefined, add it to the overdefined instruction 395 // work list so that the users of the instruction are updated later. 396 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 397 if (!IV.markOverdefined()) return false; 398 399 LLVM_DEBUG(dbgs() << "markOverdefined: "; 400 if (auto *F = dyn_cast<Function>(V)) dbgs() 401 << "Function '" << F->getName() << "'\n"; 402 else dbgs() << *V << '\n'); 403 // Only instructions go on the work list 404 pushToWorkList(IV, V); 405 return true; 406 } 407 408 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 409 /// changes. 410 bool mergeInValue(ValueLatticeElement &IV, Value *V, 411 ValueLatticeElement MergeWithV, 412 ValueLatticeElement::MergeOptions Opts = { 413 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 414 if (IV.mergeIn(MergeWithV, Opts)) { 415 pushToWorkList(IV, V); 416 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 417 << IV << "\n"); 418 return true; 419 } 420 return false; 421 } 422 423 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 424 ValueLatticeElement::MergeOptions Opts = { 425 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 426 assert(!V->getType()->isStructTy() && 427 "non-structs should use markConstant"); 428 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 429 } 430 431 /// getValueState - Return the ValueLatticeElement object that corresponds to 432 /// the value. This function handles the case when the value hasn't been seen 433 /// yet by properly seeding constants etc. 434 ValueLatticeElement &getValueState(Value *V) { 435 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 436 437 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 438 ValueLatticeElement &LV = I.first->second; 439 440 if (!I.second) 441 return LV; // Common case, already in the map. 442 443 if (auto *C = dyn_cast<Constant>(V)) 444 LV.markConstant(C); // Constants are constant 445 446 // All others are unknown by default. 447 return LV; 448 } 449 450 /// getStructValueState - Return the ValueLatticeElement object that 451 /// corresponds to the value/field pair. This function handles the case when 452 /// the value hasn't been seen yet by properly seeding constants etc. 453 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 454 assert(V->getType()->isStructTy() && "Should use getValueState"); 455 assert(i < cast<StructType>(V->getType())->getNumElements() && 456 "Invalid element #"); 457 458 auto I = StructValueState.insert( 459 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 460 ValueLatticeElement &LV = I.first->second; 461 462 if (!I.second) 463 return LV; // Common case, already in the map. 464 465 if (auto *C = dyn_cast<Constant>(V)) { 466 Constant *Elt = C->getAggregateElement(i); 467 468 if (!Elt) 469 LV.markOverdefined(); // Unknown sort of constant. 470 else if (isa<UndefValue>(Elt)) 471 ; // Undef values remain unknown. 472 else 473 LV.markConstant(Elt); // Constants are constant. 474 } 475 476 // All others are underdefined by default. 477 return LV; 478 } 479 480 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 481 /// work list if it is not already executable. 482 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 483 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 484 return false; // This edge is already known to be executable! 485 486 if (!MarkBlockExecutable(Dest)) { 487 // If the destination is already executable, we just made an *edge* 488 // feasible that wasn't before. Revisit the PHI nodes in the block 489 // because they have potentially new operands. 490 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 491 << " -> " << Dest->getName() << '\n'); 492 493 for (PHINode &PN : Dest->phis()) 494 visitPHINode(PN); 495 } 496 return true; 497 } 498 499 // getFeasibleSuccessors - Return a vector of booleans to indicate which 500 // successors are reachable from a given terminator instruction. 501 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 502 503 // OperandChangedState - This method is invoked on all of the users of an 504 // instruction that was just changed state somehow. Based on this 505 // information, we need to update the specified user of this instruction. 506 void OperandChangedState(Instruction *I) { 507 if (BBExecutable.count(I->getParent())) // Inst is executable? 508 visit(*I); 509 } 510 511 // Add U as additional user of V. 512 void addAdditionalUser(Value *V, User *U) { 513 auto Iter = AdditionalUsers.insert({V, {}}); 514 Iter.first->second.insert(U); 515 } 516 517 // Mark I's users as changed, including AdditionalUsers. 518 void markUsersAsChanged(Value *I) { 519 // Functions include their arguments in the use-list. Changed function 520 // values mean that the result of the function changed. We only need to 521 // update the call sites with the new function result and do not have to 522 // propagate the call arguments. 523 if (isa<Function>(I)) { 524 for (User *U : I->users()) { 525 if (auto *CB = dyn_cast<CallBase>(U)) 526 handleCallResult(*CB); 527 } 528 } else { 529 for (User *U : I->users()) 530 if (auto *UI = dyn_cast<Instruction>(U)) 531 OperandChangedState(UI); 532 } 533 534 auto Iter = AdditionalUsers.find(I); 535 if (Iter != AdditionalUsers.end()) { 536 for (User *U : Iter->second) 537 if (auto *UI = dyn_cast<Instruction>(U)) 538 OperandChangedState(UI); 539 } 540 } 541 void handleCallOverdefined(CallBase &CB); 542 void handleCallResult(CallBase &CB); 543 void handleCallArguments(CallBase &CB); 544 545 private: 546 friend class InstVisitor<SCCPSolver>; 547 548 // visit implementations - Something changed in this instruction. Either an 549 // operand made a transition, or the instruction is newly executable. Change 550 // the value type of I to reflect these changes if appropriate. 551 void visitPHINode(PHINode &I); 552 553 // Terminators 554 555 void visitReturnInst(ReturnInst &I); 556 void visitTerminator(Instruction &TI); 557 558 void visitCastInst(CastInst &I); 559 void visitSelectInst(SelectInst &I); 560 void visitUnaryOperator(Instruction &I); 561 void visitBinaryOperator(Instruction &I); 562 void visitCmpInst(CmpInst &I); 563 void visitExtractValueInst(ExtractValueInst &EVI); 564 void visitInsertValueInst(InsertValueInst &IVI); 565 566 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 567 markOverdefined(&CPI); 568 visitTerminator(CPI); 569 } 570 571 // Instructions that cannot be folded away. 572 573 void visitStoreInst (StoreInst &I); 574 void visitLoadInst (LoadInst &I); 575 void visitGetElementPtrInst(GetElementPtrInst &I); 576 577 void visitCallInst (CallInst &I) { 578 visitCallBase(I); 579 } 580 581 void visitInvokeInst (InvokeInst &II) { 582 visitCallBase(II); 583 visitTerminator(II); 584 } 585 586 void visitCallBrInst (CallBrInst &CBI) { 587 visitCallBase(CBI); 588 visitTerminator(CBI); 589 } 590 591 void visitCallBase (CallBase &CB); 592 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 593 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 594 void visitFenceInst (FenceInst &I) { /*returns void*/ } 595 596 void visitInstruction(Instruction &I) { 597 // All the instructions we don't do any special handling for just 598 // go to overdefined. 599 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 600 markOverdefined(&I); 601 } 602 }; 603 604 } // end anonymous namespace 605 606 // getFeasibleSuccessors - Return a vector of booleans to indicate which 607 // successors are reachable from a given terminator instruction. 608 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 609 SmallVectorImpl<bool> &Succs) { 610 Succs.resize(TI.getNumSuccessors()); 611 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 612 if (BI->isUnconditional()) { 613 Succs[0] = true; 614 return; 615 } 616 617 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 618 ConstantInt *CI = getConstantInt(BCValue); 619 if (!CI) { 620 // Overdefined condition variables, and branches on unfoldable constant 621 // conditions, mean the branch could go either way. 622 if (!BCValue.isUnknownOrUndef()) 623 Succs[0] = Succs[1] = true; 624 return; 625 } 626 627 // Constant condition variables mean the branch can only go a single way. 628 Succs[CI->isZero()] = true; 629 return; 630 } 631 632 // Unwinding instructions successors are always executable. 633 if (TI.isExceptionalTerminator()) { 634 Succs.assign(TI.getNumSuccessors(), true); 635 return; 636 } 637 638 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 639 if (!SI->getNumCases()) { 640 Succs[0] = true; 641 return; 642 } 643 ValueLatticeElement SCValue = getValueState(SI->getCondition()); 644 ConstantInt *CI = getConstantInt(SCValue); 645 646 if (!CI) { // Overdefined or unknown condition? 647 // All destinations are executable! 648 if (!SCValue.isUnknownOrUndef()) 649 Succs.assign(TI.getNumSuccessors(), true); 650 return; 651 } 652 653 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 654 return; 655 } 656 657 // In case of indirect branch and its address is a blockaddress, we mark 658 // the target as executable. 659 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 660 // Casts are folded by visitCastInst. 661 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 662 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 663 if (!Addr) { // Overdefined or unknown condition? 664 // All destinations are executable! 665 if (!IBRValue.isUnknownOrUndef()) 666 Succs.assign(TI.getNumSuccessors(), true); 667 return; 668 } 669 670 BasicBlock* T = Addr->getBasicBlock(); 671 assert(Addr->getFunction() == T->getParent() && 672 "Block address of a different function ?"); 673 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 674 // This is the target. 675 if (IBR->getDestination(i) == T) { 676 Succs[i] = true; 677 return; 678 } 679 } 680 681 // If we didn't find our destination in the IBR successor list, then we 682 // have undefined behavior. Its ok to assume no successor is executable. 683 return; 684 } 685 686 // In case of callbr, we pessimistically assume that all successors are 687 // feasible. 688 if (isa<CallBrInst>(&TI)) { 689 Succs.assign(TI.getNumSuccessors(), true); 690 return; 691 } 692 693 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 694 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 695 } 696 697 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 698 // block to the 'To' basic block is currently feasible. 699 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 700 // Check if we've called markEdgeExecutable on the edge yet. (We could 701 // be more aggressive and try to consider edges which haven't been marked 702 // yet, but there isn't any need.) 703 return KnownFeasibleEdges.count(Edge(From, To)); 704 } 705 706 // visit Implementations - Something changed in this instruction, either an 707 // operand made a transition, or the instruction is newly executable. Change 708 // the value type of I to reflect these changes if appropriate. This method 709 // makes sure to do the following actions: 710 // 711 // 1. If a phi node merges two constants in, and has conflicting value coming 712 // from different branches, or if the PHI node merges in an overdefined 713 // value, then the PHI node becomes overdefined. 714 // 2. If a phi node merges only constants in, and they all agree on value, the 715 // PHI node becomes a constant value equal to that. 716 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 717 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 718 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 719 // 6. If a conditional branch has a value that is constant, make the selected 720 // destination executable 721 // 7. If a conditional branch has a value that is overdefined, make all 722 // successors executable. 723 void SCCPSolver::visitPHINode(PHINode &PN) { 724 // If this PN returns a struct, just mark the result overdefined. 725 // TODO: We could do a lot better than this if code actually uses this. 726 if (PN.getType()->isStructTy()) 727 return (void)markOverdefined(&PN); 728 729 if (getValueState(&PN).isOverdefined()) 730 return; // Quick exit 731 732 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 733 // and slow us down a lot. Just mark them overdefined. 734 if (PN.getNumIncomingValues() > 64) 735 return (void)markOverdefined(&PN); 736 737 unsigned NumActiveIncoming = 0; 738 739 // Look at all of the executable operands of the PHI node. If any of them 740 // are overdefined, the PHI becomes overdefined as well. If they are all 741 // constant, and they agree with each other, the PHI becomes the identical 742 // constant. If they are constant and don't agree, the PHI is a constant 743 // range. If there are no executable operands, the PHI remains unknown. 744 ValueLatticeElement PhiState = getValueState(&PN); 745 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 746 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 747 continue; 748 749 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 750 PhiState.mergeIn(IV); 751 NumActiveIncoming++; 752 if (PhiState.isOverdefined()) 753 break; 754 } 755 756 // We allow up to 1 range extension per active incoming value and one 757 // additional extension. Note that we manually adjust the number of range 758 // extensions to match the number of active incoming values. This helps to 759 // limit multiple extensions caused by the same incoming value, if other 760 // incoming values are equal. 761 mergeInValue(&PN, PhiState, 762 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 763 NumActiveIncoming + 1)); 764 ValueLatticeElement &PhiStateRef = getValueState(&PN); 765 PhiStateRef.setNumRangeExtensions( 766 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 767 } 768 769 void SCCPSolver::visitReturnInst(ReturnInst &I) { 770 if (I.getNumOperands() == 0) return; // ret void 771 772 Function *F = I.getParent()->getParent(); 773 Value *ResultOp = I.getOperand(0); 774 775 // If we are tracking the return value of this function, merge it in. 776 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 777 auto TFRVI = TrackedRetVals.find(F); 778 if (TFRVI != TrackedRetVals.end()) { 779 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 780 return; 781 } 782 } 783 784 // Handle functions that return multiple values. 785 if (!TrackedMultipleRetVals.empty()) { 786 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 787 if (MRVFunctionsTracked.count(F)) 788 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 789 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 790 getStructValueState(ResultOp, i)); 791 } 792 } 793 794 void SCCPSolver::visitTerminator(Instruction &TI) { 795 SmallVector<bool, 16> SuccFeasible; 796 getFeasibleSuccessors(TI, SuccFeasible); 797 798 BasicBlock *BB = TI.getParent(); 799 800 // Mark all feasible successors executable. 801 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 802 if (SuccFeasible[i]) 803 markEdgeExecutable(BB, TI.getSuccessor(i)); 804 } 805 806 void SCCPSolver::visitCastInst(CastInst &I) { 807 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 808 // discover a concrete value later. 809 if (ValueState[&I].isOverdefined()) 810 return; 811 812 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 813 if (Constant *OpC = getConstant(OpSt)) { 814 // Fold the constant as we build. 815 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 816 if (isa<UndefValue>(C)) 817 return; 818 // Propagate constant value 819 markConstant(&I, C); 820 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 821 auto &LV = getValueState(&I); 822 ConstantRange OpRange = OpSt.getConstantRange(); 823 Type *DestTy = I.getDestTy(); 824 ConstantRange Res = 825 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 826 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 827 } else if (!OpSt.isUnknownOrUndef()) 828 markOverdefined(&I); 829 } 830 831 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 832 // If this returns a struct, mark all elements over defined, we don't track 833 // structs in structs. 834 if (EVI.getType()->isStructTy()) 835 return (void)markOverdefined(&EVI); 836 837 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 838 // discover a concrete value later. 839 if (ValueState[&EVI].isOverdefined()) 840 return (void)markOverdefined(&EVI); 841 842 // If this is extracting from more than one level of struct, we don't know. 843 if (EVI.getNumIndices() != 1) 844 return (void)markOverdefined(&EVI); 845 846 Value *AggVal = EVI.getAggregateOperand(); 847 if (AggVal->getType()->isStructTy()) { 848 unsigned i = *EVI.idx_begin(); 849 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 850 mergeInValue(getValueState(&EVI), &EVI, EltVal); 851 } else { 852 // Otherwise, must be extracting from an array. 853 return (void)markOverdefined(&EVI); 854 } 855 } 856 857 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 858 auto *STy = dyn_cast<StructType>(IVI.getType()); 859 if (!STy) 860 return (void)markOverdefined(&IVI); 861 862 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 863 // discover a concrete value later. 864 if (isOverdefined(ValueState[&IVI])) 865 return (void)markOverdefined(&IVI); 866 867 // If this has more than one index, we can't handle it, drive all results to 868 // undef. 869 if (IVI.getNumIndices() != 1) 870 return (void)markOverdefined(&IVI); 871 872 Value *Aggr = IVI.getAggregateOperand(); 873 unsigned Idx = *IVI.idx_begin(); 874 875 // Compute the result based on what we're inserting. 876 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 877 // This passes through all values that aren't the inserted element. 878 if (i != Idx) { 879 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 880 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 881 continue; 882 } 883 884 Value *Val = IVI.getInsertedValueOperand(); 885 if (Val->getType()->isStructTy()) 886 // We don't track structs in structs. 887 markOverdefined(getStructValueState(&IVI, i), &IVI); 888 else { 889 ValueLatticeElement InVal = getValueState(Val); 890 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 891 } 892 } 893 } 894 895 void SCCPSolver::visitSelectInst(SelectInst &I) { 896 // If this select returns a struct, just mark the result overdefined. 897 // TODO: We could do a lot better than this if code actually uses this. 898 if (I.getType()->isStructTy()) 899 return (void)markOverdefined(&I); 900 901 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 902 // discover a concrete value later. 903 if (ValueState[&I].isOverdefined()) 904 return (void)markOverdefined(&I); 905 906 ValueLatticeElement CondValue = getValueState(I.getCondition()); 907 if (CondValue.isUnknownOrUndef()) 908 return; 909 910 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 911 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 912 mergeInValue(&I, getValueState(OpVal)); 913 return; 914 } 915 916 // Otherwise, the condition is overdefined or a constant we can't evaluate. 917 // See if we can produce something better than overdefined based on the T/F 918 // value. 919 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 920 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 921 922 bool Changed = ValueState[&I].mergeIn(TVal); 923 Changed |= ValueState[&I].mergeIn(FVal); 924 if (Changed) 925 pushToWorkListMsg(ValueState[&I], &I); 926 } 927 928 // Handle Unary Operators. 929 void SCCPSolver::visitUnaryOperator(Instruction &I) { 930 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 931 932 ValueLatticeElement &IV = ValueState[&I]; 933 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 934 // discover a concrete value later. 935 if (isOverdefined(IV)) 936 return (void)markOverdefined(&I); 937 938 if (isConstant(V0State)) { 939 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 940 941 // op Y -> undef. 942 if (isa<UndefValue>(C)) 943 return; 944 return (void)markConstant(IV, &I, C); 945 } 946 947 // If something is undef, wait for it to resolve. 948 if (!isOverdefined(V0State)) 949 return; 950 951 markOverdefined(&I); 952 } 953 954 // Handle Binary Operators. 955 void SCCPSolver::visitBinaryOperator(Instruction &I) { 956 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 957 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 958 959 ValueLatticeElement &IV = ValueState[&I]; 960 if (IV.isOverdefined()) 961 return; 962 963 // If something is undef, wait for it to resolve. 964 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 965 return; 966 967 if (V1State.isOverdefined() && V2State.isOverdefined()) 968 return (void)markOverdefined(&I); 969 970 // If either of the operands is a constant, try to fold it to a constant. 971 // TODO: Use information from notconstant better. 972 if ((V1State.isConstant() || V2State.isConstant())) { 973 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 974 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 975 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 976 auto *C = dyn_cast_or_null<Constant>(R); 977 if (C) { 978 // X op Y -> undef. 979 if (isa<UndefValue>(C)) 980 return; 981 // Conservatively assume that the result may be based on operands that may 982 // be undef. Note that we use mergeInValue to combine the constant with 983 // the existing lattice value for I, as different constants might be found 984 // after one of the operands go to overdefined, e.g. due to one operand 985 // being a special floating value. 986 ValueLatticeElement NewV; 987 NewV.markConstant(C, /*MayIncludeUndef=*/true); 988 return (void)mergeInValue(&I, NewV); 989 } 990 } 991 992 // Only use ranges for binary operators on integers. 993 if (!I.getType()->isIntegerTy()) 994 return markOverdefined(&I); 995 996 // Try to simplify to a constant range. 997 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 998 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 999 if (V1State.isConstantRange()) 1000 A = V1State.getConstantRange(); 1001 if (V2State.isConstantRange()) 1002 B = V2State.getConstantRange(); 1003 1004 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1005 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1006 1007 // TODO: Currently we do not exploit special values that produce something 1008 // better than overdefined with an overdefined operand for vector or floating 1009 // point types, like and <4 x i32> overdefined, zeroinitializer. 1010 } 1011 1012 // Handle ICmpInst instruction. 1013 void SCCPSolver::visitCmpInst(CmpInst &I) { 1014 // Do not cache this lookup, getValueState calls later in the function might 1015 // invalidate the reference. 1016 if (isOverdefined(ValueState[&I])) 1017 return (void)markOverdefined(&I); 1018 1019 Value *Op1 = I.getOperand(0); 1020 Value *Op2 = I.getOperand(1); 1021 1022 // For parameters, use ParamState which includes constant range info if 1023 // available. 1024 auto V1State = getValueState(Op1); 1025 auto V2State = getValueState(Op2); 1026 1027 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1028 if (C) { 1029 if (isa<UndefValue>(C)) 1030 return; 1031 ValueLatticeElement CV; 1032 CV.markConstant(C); 1033 mergeInValue(&I, CV); 1034 return; 1035 } 1036 1037 // If operands are still unknown, wait for it to resolve. 1038 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1039 !isConstant(ValueState[&I])) 1040 return; 1041 1042 markOverdefined(&I); 1043 } 1044 1045 // Handle getelementptr instructions. If all operands are constants then we 1046 // can turn this into a getelementptr ConstantExpr. 1047 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1048 if (isOverdefined(ValueState[&I])) 1049 return (void)markOverdefined(&I); 1050 1051 SmallVector<Constant*, 8> Operands; 1052 Operands.reserve(I.getNumOperands()); 1053 1054 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1055 ValueLatticeElement State = getValueState(I.getOperand(i)); 1056 if (State.isUnknownOrUndef()) 1057 return; // Operands are not resolved yet. 1058 1059 if (isOverdefined(State)) 1060 return (void)markOverdefined(&I); 1061 1062 if (Constant *C = getConstant(State)) { 1063 Operands.push_back(C); 1064 continue; 1065 } 1066 1067 return (void)markOverdefined(&I); 1068 } 1069 1070 Constant *Ptr = Operands[0]; 1071 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1072 Constant *C = 1073 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1074 if (isa<UndefValue>(C)) 1075 return; 1076 markConstant(&I, C); 1077 } 1078 1079 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1080 // If this store is of a struct, ignore it. 1081 if (SI.getOperand(0)->getType()->isStructTy()) 1082 return; 1083 1084 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1085 return; 1086 1087 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1088 auto I = TrackedGlobals.find(GV); 1089 if (I == TrackedGlobals.end()) 1090 return; 1091 1092 // Get the value we are storing into the global, then merge it. 1093 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1094 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1095 if (I->second.isOverdefined()) 1096 TrackedGlobals.erase(I); // No need to keep tracking this! 1097 } 1098 1099 // Handle load instructions. If the operand is a constant pointer to a constant 1100 // global, we can replace the load with the loaded constant value! 1101 void SCCPSolver::visitLoadInst(LoadInst &I) { 1102 // If this load is of a struct, just mark the result overdefined. 1103 if (I.getType()->isStructTy()) 1104 return (void)markOverdefined(&I); 1105 1106 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1107 // discover a concrete value later. 1108 if (ValueState[&I].isOverdefined()) 1109 return (void)markOverdefined(&I); 1110 1111 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1112 if (PtrVal.isUnknownOrUndef()) 1113 return; // The pointer is not resolved yet! 1114 1115 ValueLatticeElement &IV = ValueState[&I]; 1116 1117 if (!isConstant(PtrVal) || I.isVolatile()) 1118 return (void)markOverdefined(IV, &I); 1119 1120 Constant *Ptr = getConstant(PtrVal); 1121 1122 // load null is undefined. 1123 if (isa<ConstantPointerNull>(Ptr)) { 1124 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1125 return (void)markOverdefined(IV, &I); 1126 else 1127 return; 1128 } 1129 1130 // Transform load (constant global) into the value loaded. 1131 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1132 if (!TrackedGlobals.empty()) { 1133 // If we are tracking this global, merge in the known value for it. 1134 auto It = TrackedGlobals.find(GV); 1135 if (It != TrackedGlobals.end()) { 1136 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1137 return; 1138 } 1139 } 1140 } 1141 1142 // Transform load from a constant into a constant if possible. 1143 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1144 if (isa<UndefValue>(C)) 1145 return; 1146 return (void)markConstant(IV, &I, C); 1147 } 1148 1149 // Otherwise we cannot say for certain what value this load will produce. 1150 // Bail out. 1151 markOverdefined(IV, &I); 1152 } 1153 1154 void SCCPSolver::visitCallBase(CallBase &CB) { 1155 handleCallResult(CB); 1156 handleCallArguments(CB); 1157 } 1158 1159 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1160 Function *F = CB.getCalledFunction(); 1161 1162 // Void return and not tracking callee, just bail. 1163 if (CB.getType()->isVoidTy()) 1164 return; 1165 1166 // Otherwise, if we have a single return value case, and if the function is 1167 // a declaration, maybe we can constant fold it. 1168 if (F && F->isDeclaration() && !CB.getType()->isStructTy() && 1169 canConstantFoldCallTo(&CB, F)) { 1170 SmallVector<Constant *, 8> Operands; 1171 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1172 if (AI->get()->getType()->isStructTy()) 1173 return markOverdefined(&CB); // Can't handle struct args. 1174 ValueLatticeElement State = getValueState(*AI); 1175 1176 if (State.isUnknownOrUndef()) 1177 return; // Operands are not resolved yet. 1178 if (isOverdefined(State)) 1179 return (void)markOverdefined(&CB); 1180 assert(isConstant(State) && "Unknown state!"); 1181 Operands.push_back(getConstant(State)); 1182 } 1183 1184 if (isOverdefined(getValueState(&CB))) 1185 return (void)markOverdefined(&CB); 1186 1187 // If we can constant fold this, mark the result of the call as a 1188 // constant. 1189 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1190 // call -> undef. 1191 if (isa<UndefValue>(C)) 1192 return; 1193 return (void)markConstant(&CB, C); 1194 } 1195 } 1196 1197 // Otherwise, we don't know anything about this call, mark it overdefined. 1198 return (void)markOverdefined(&CB); 1199 } 1200 1201 void SCCPSolver::handleCallArguments(CallBase &CB) { 1202 Function *F = CB.getCalledFunction(); 1203 // If this is a local function that doesn't have its address taken, mark its 1204 // entry block executable and merge in the actual arguments to the call into 1205 // the formal arguments of the function. 1206 if (!TrackingIncomingArguments.empty() && 1207 TrackingIncomingArguments.count(F)) { 1208 MarkBlockExecutable(&F->front()); 1209 1210 // Propagate information from this call site into the callee. 1211 auto CAI = CB.arg_begin(); 1212 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1213 ++AI, ++CAI) { 1214 // If this argument is byval, and if the function is not readonly, there 1215 // will be an implicit copy formed of the input aggregate. 1216 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1217 markOverdefined(&*AI); 1218 continue; 1219 } 1220 1221 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1222 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1223 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1224 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1225 getMaxWidenStepsOpts()); 1226 } 1227 } else 1228 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1229 } 1230 } 1231 } 1232 1233 void SCCPSolver::handleCallResult(CallBase &CB) { 1234 Function *F = CB.getCalledFunction(); 1235 1236 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1237 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1238 if (ValueState[&CB].isOverdefined()) 1239 return; 1240 1241 Value *CopyOf = CB.getOperand(0); 1242 auto *PI = getPredicateInfoFor(&CB); 1243 auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI); 1244 ValueLatticeElement OriginalVal = getValueState(CopyOf); 1245 if (!PI || !PBranch) { 1246 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1247 return; 1248 } 1249 1250 // Everything below relies on the condition being a comparison. 1251 auto *Cmp = dyn_cast<CmpInst>(PBranch->Condition); 1252 if (!Cmp) { 1253 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1254 return; 1255 } 1256 1257 Value *CmpOp0 = Cmp->getOperand(0); 1258 Value *CmpOp1 = Cmp->getOperand(1); 1259 if (CopyOf != CmpOp0 && CopyOf != CmpOp1) { 1260 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1261 return; 1262 } 1263 1264 auto Pred = Cmp->getPredicate(); 1265 if (CmpOp0 != CopyOf) { 1266 std::swap(CmpOp0, CmpOp1); 1267 Pred = Cmp->getSwappedPredicate(); 1268 } 1269 1270 // Wait until CmpOp1 is resolved. 1271 if (getValueState(CmpOp1).isUnknown()) { 1272 addAdditionalUser(CmpOp1, &CB); 1273 return; 1274 } 1275 1276 if (!PBranch->TrueEdge) 1277 Pred = CmpInst::getInversePredicate(Pred); 1278 1279 ValueLatticeElement CondVal = getValueState(CmpOp1); 1280 ValueLatticeElement &IV = ValueState[&CB]; 1281 if (CondVal.isConstantRange() || OriginalVal.isConstantRange()) { 1282 auto NewCR = 1283 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1284 1285 // Get the range imposed by the condition. 1286 if (CondVal.isConstantRange()) 1287 NewCR = ConstantRange::makeAllowedICmpRegion( 1288 Pred, CondVal.getConstantRange()); 1289 1290 // Combine range info for the original value with the new range from the 1291 // condition. 1292 auto OriginalCR = OriginalVal.isConstantRange() 1293 ? OriginalVal.getConstantRange() 1294 : ConstantRange::getFull( 1295 DL.getTypeSizeInBits(CopyOf->getType())); 1296 NewCR = NewCR.intersectWith(OriginalCR); 1297 1298 addAdditionalUser(CmpOp1, &CB); 1299 // TODO: Actually filp MayIncludeUndef for the created range to false, 1300 // once most places in the optimizer respect the branches on 1301 // undef/poison are UB rule. The reason why the new range cannot be 1302 // undef is as follows below: 1303 // The new range is based on a branch condition. That guarantees that 1304 // neither of the compare operands can be undef in the branch targets, 1305 // unless we have conditions that are always true/false (e.g. icmp ule 1306 // i32, %a, i32_max). For the latter overdefined/empty range will be 1307 // inferred, but the branch will get folded accordingly anyways. 1308 mergeInValue( 1309 IV, &CB, 1310 ValueLatticeElement::getRange(NewCR, /*MayIncludeUndef=*/true)); 1311 return; 1312 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1313 // For non-integer values or integer constant expressions, only 1314 // propagate equal constants. 1315 addAdditionalUser(CmpOp1, &CB); 1316 mergeInValue(IV, &CB, CondVal); 1317 return; 1318 } 1319 1320 return (void)mergeInValue(IV, &CB, OriginalVal); 1321 } 1322 } 1323 1324 // The common case is that we aren't tracking the callee, either because we 1325 // are not doing interprocedural analysis or the callee is indirect, or is 1326 // external. Handle these cases first. 1327 if (!F || F->isDeclaration()) 1328 return handleCallOverdefined(CB); 1329 1330 // If this is a single/zero retval case, see if we're tracking the function. 1331 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1332 if (!MRVFunctionsTracked.count(F)) 1333 return handleCallOverdefined(CB); // Not tracking this callee. 1334 1335 // If we are tracking this callee, propagate the result of the function 1336 // into this call site. 1337 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1338 mergeInValue(getStructValueState(&CB, i), &CB, 1339 TrackedMultipleRetVals[std::make_pair(F, i)], 1340 getMaxWidenStepsOpts()); 1341 } else { 1342 auto TFRVI = TrackedRetVals.find(F); 1343 if (TFRVI == TrackedRetVals.end()) 1344 return handleCallOverdefined(CB); // Not tracking this callee. 1345 1346 // If so, propagate the return value of the callee into this call result. 1347 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1348 } 1349 } 1350 1351 void SCCPSolver::Solve() { 1352 // Process the work lists until they are empty! 1353 while (!BBWorkList.empty() || !InstWorkList.empty() || 1354 !OverdefinedInstWorkList.empty()) { 1355 // Process the overdefined instruction's work list first, which drives other 1356 // things to overdefined more quickly. 1357 while (!OverdefinedInstWorkList.empty()) { 1358 Value *I = OverdefinedInstWorkList.pop_back_val(); 1359 1360 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1361 1362 // "I" got into the work list because it either made the transition from 1363 // bottom to constant, or to overdefined. 1364 // 1365 // Anything on this worklist that is overdefined need not be visited 1366 // since all of its users will have already been marked as overdefined 1367 // Update all of the users of this instruction's value. 1368 // 1369 markUsersAsChanged(I); 1370 } 1371 1372 // Process the instruction work list. 1373 while (!InstWorkList.empty()) { 1374 Value *I = InstWorkList.pop_back_val(); 1375 1376 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1377 1378 // "I" got into the work list because it made the transition from undef to 1379 // constant. 1380 // 1381 // Anything on this worklist that is overdefined need not be visited 1382 // since all of its users will have already been marked as overdefined. 1383 // Update all of the users of this instruction's value. 1384 // 1385 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1386 markUsersAsChanged(I); 1387 } 1388 1389 // Process the basic block work list. 1390 while (!BBWorkList.empty()) { 1391 BasicBlock *BB = BBWorkList.back(); 1392 BBWorkList.pop_back(); 1393 1394 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1395 1396 // Notify all instructions in this basic block that they are newly 1397 // executable. 1398 visit(BB); 1399 } 1400 } 1401 } 1402 1403 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1404 /// that branches on undef values cannot reach any of their successors. 1405 /// However, this is not a safe assumption. After we solve dataflow, this 1406 /// method should be use to handle this. If this returns true, the solver 1407 /// should be rerun. 1408 /// 1409 /// This method handles this by finding an unresolved branch and marking it one 1410 /// of the edges from the block as being feasible, even though the condition 1411 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1412 /// CFG and only slightly pessimizes the analysis results (by marking one, 1413 /// potentially infeasible, edge feasible). This cannot usefully modify the 1414 /// constraints on the condition of the branch, as that would impact other users 1415 /// of the value. 1416 /// 1417 /// This scan also checks for values that use undefs. It conservatively marks 1418 /// them as overdefined. 1419 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1420 for (BasicBlock &BB : F) { 1421 if (!BBExecutable.count(&BB)) 1422 continue; 1423 1424 for (Instruction &I : BB) { 1425 // Look for instructions which produce undef values. 1426 if (I.getType()->isVoidTy()) continue; 1427 1428 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1429 // Only a few things that can be structs matter for undef. 1430 1431 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1432 if (auto *CB = dyn_cast<CallBase>(&I)) 1433 if (Function *F = CB->getCalledFunction()) 1434 if (MRVFunctionsTracked.count(F)) 1435 continue; 1436 1437 // extractvalue and insertvalue don't need to be marked; they are 1438 // tracked as precisely as their operands. 1439 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1440 continue; 1441 // Send the results of everything else to overdefined. We could be 1442 // more precise than this but it isn't worth bothering. 1443 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1444 ValueLatticeElement &LV = getStructValueState(&I, i); 1445 if (LV.isUnknownOrUndef()) 1446 markOverdefined(LV, &I); 1447 } 1448 continue; 1449 } 1450 1451 ValueLatticeElement &LV = getValueState(&I); 1452 if (!LV.isUnknownOrUndef()) 1453 continue; 1454 1455 // There are two reasons a call can have an undef result 1456 // 1. It could be tracked. 1457 // 2. It could be constant-foldable. 1458 // Because of the way we solve return values, tracked calls must 1459 // never be marked overdefined in ResolvedUndefsIn. 1460 if (auto *CB = dyn_cast<CallBase>(&I)) 1461 if (Function *F = CB->getCalledFunction()) 1462 if (TrackedRetVals.count(F)) 1463 continue; 1464 1465 if (isa<LoadInst>(I)) { 1466 // A load here means one of two things: a load of undef from a global, 1467 // a load from an unknown pointer. Either way, having it return undef 1468 // is okay. 1469 continue; 1470 } 1471 1472 markOverdefined(&I); 1473 return true; 1474 } 1475 1476 // Check to see if we have a branch or switch on an undefined value. If so 1477 // we force the branch to go one way or the other to make the successor 1478 // values live. It doesn't really matter which way we force it. 1479 Instruction *TI = BB.getTerminator(); 1480 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1481 if (!BI->isConditional()) continue; 1482 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1483 continue; 1484 1485 // If the input to SCCP is actually branch on undef, fix the undef to 1486 // false. 1487 if (isa<UndefValue>(BI->getCondition())) { 1488 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1489 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1490 return true; 1491 } 1492 1493 // Otherwise, it is a branch on a symbolic value which is currently 1494 // considered to be undef. Make sure some edge is executable, so a 1495 // branch on "undef" always flows somewhere. 1496 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1497 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1498 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1499 return true; 1500 1501 continue; 1502 } 1503 1504 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1505 // Indirect branch with no successor ?. Its ok to assume it branches 1506 // to no target. 1507 if (IBR->getNumSuccessors() < 1) 1508 continue; 1509 1510 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1511 continue; 1512 1513 // If the input to SCCP is actually branch on undef, fix the undef to 1514 // the first successor of the indirect branch. 1515 if (isa<UndefValue>(IBR->getAddress())) { 1516 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1517 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1518 return true; 1519 } 1520 1521 // Otherwise, it is a branch on a symbolic value which is currently 1522 // considered to be undef. Make sure some edge is executable, so a 1523 // branch on "undef" always flows somewhere. 1524 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1525 // we can assume the branch has undefined behavior instead. 1526 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1527 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1528 return true; 1529 1530 continue; 1531 } 1532 1533 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1534 if (!SI->getNumCases() || 1535 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1536 continue; 1537 1538 // If the input to SCCP is actually switch on undef, fix the undef to 1539 // the first constant. 1540 if (isa<UndefValue>(SI->getCondition())) { 1541 SI->setCondition(SI->case_begin()->getCaseValue()); 1542 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1543 return true; 1544 } 1545 1546 // Otherwise, it is a branch on a symbolic value which is currently 1547 // considered to be undef. Make sure some edge is executable, so a 1548 // branch on "undef" always flows somewhere. 1549 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1550 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1551 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1552 return true; 1553 1554 continue; 1555 } 1556 } 1557 1558 return false; 1559 } 1560 1561 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1562 Constant *Const = nullptr; 1563 if (V->getType()->isStructTy()) { 1564 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1565 if (any_of(IVs, 1566 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1567 return false; 1568 std::vector<Constant *> ConstVals; 1569 auto *ST = cast<StructType>(V->getType()); 1570 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1571 ValueLatticeElement V = IVs[i]; 1572 ConstVals.push_back(isConstant(V) 1573 ? Solver.getConstant(V) 1574 : UndefValue::get(ST->getElementType(i))); 1575 } 1576 Const = ConstantStruct::get(ST, ConstVals); 1577 } else { 1578 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1579 if (isOverdefined(IV)) 1580 return false; 1581 1582 Const = 1583 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1584 } 1585 assert(Const && "Constant is nullptr here!"); 1586 1587 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1588 // unless the call itself can be removed 1589 CallInst *CI = dyn_cast<CallInst>(V); 1590 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1591 Function *F = CI->getCalledFunction(); 1592 1593 // Don't zap returns of the callee 1594 if (F) 1595 Solver.AddMustTailCallee(F); 1596 1597 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1598 << " as a constant\n"); 1599 return false; 1600 } 1601 1602 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1603 1604 // Replaces all of the uses of a variable with uses of the constant. 1605 V->replaceAllUsesWith(Const); 1606 return true; 1607 } 1608 1609 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1610 // and return true if the function was modified. 1611 static bool runSCCP(Function &F, const DataLayout &DL, 1612 const TargetLibraryInfo *TLI) { 1613 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1614 SCCPSolver Solver( 1615 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1616 F.getContext()); 1617 1618 // Mark the first block of the function as being executable. 1619 Solver.MarkBlockExecutable(&F.front()); 1620 1621 // Mark all arguments to the function as being overdefined. 1622 for (Argument &AI : F.args()) 1623 Solver.markOverdefined(&AI); 1624 1625 // Solve for constants. 1626 bool ResolvedUndefs = true; 1627 while (ResolvedUndefs) { 1628 Solver.Solve(); 1629 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1630 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1631 } 1632 1633 bool MadeChanges = false; 1634 1635 // If we decided that there are basic blocks that are dead in this function, 1636 // delete their contents now. Note that we cannot actually delete the blocks, 1637 // as we cannot modify the CFG of the function. 1638 1639 for (BasicBlock &BB : F) { 1640 if (!Solver.isBlockExecutable(&BB)) { 1641 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1642 1643 ++NumDeadBlocks; 1644 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB); 1645 1646 MadeChanges = true; 1647 continue; 1648 } 1649 1650 // Iterate over all of the instructions in a function, replacing them with 1651 // constants if we have found them to be of constant values. 1652 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1653 Instruction *Inst = &*BI++; 1654 if (Inst->getType()->isVoidTy() || Inst->isTerminator()) 1655 continue; 1656 1657 if (tryToReplaceWithConstant(Solver, Inst)) { 1658 if (isInstructionTriviallyDead(Inst)) 1659 Inst->eraseFromParent(); 1660 // Hey, we just changed something! 1661 MadeChanges = true; 1662 ++NumInstRemoved; 1663 } 1664 } 1665 } 1666 1667 return MadeChanges; 1668 } 1669 1670 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1671 const DataLayout &DL = F.getParent()->getDataLayout(); 1672 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1673 if (!runSCCP(F, DL, &TLI)) 1674 return PreservedAnalyses::all(); 1675 1676 auto PA = PreservedAnalyses(); 1677 PA.preserve<GlobalsAA>(); 1678 PA.preserveSet<CFGAnalyses>(); 1679 return PA; 1680 } 1681 1682 namespace { 1683 1684 //===--------------------------------------------------------------------===// 1685 // 1686 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1687 /// Sparse Conditional Constant Propagator. 1688 /// 1689 class SCCPLegacyPass : public FunctionPass { 1690 public: 1691 // Pass identification, replacement for typeid 1692 static char ID; 1693 1694 SCCPLegacyPass() : FunctionPass(ID) { 1695 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1696 } 1697 1698 void getAnalysisUsage(AnalysisUsage &AU) const override { 1699 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1700 AU.addPreserved<GlobalsAAWrapperPass>(); 1701 AU.setPreservesCFG(); 1702 } 1703 1704 // runOnFunction - Run the Sparse Conditional Constant Propagation 1705 // algorithm, and return true if the function was modified. 1706 bool runOnFunction(Function &F) override { 1707 if (skipFunction(F)) 1708 return false; 1709 const DataLayout &DL = F.getParent()->getDataLayout(); 1710 const TargetLibraryInfo *TLI = 1711 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1712 return runSCCP(F, DL, TLI); 1713 } 1714 }; 1715 1716 } // end anonymous namespace 1717 1718 char SCCPLegacyPass::ID = 0; 1719 1720 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1721 "Sparse Conditional Constant Propagation", false, false) 1722 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1723 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1724 "Sparse Conditional Constant Propagation", false, false) 1725 1726 // createSCCPPass - This is the public interface to this file. 1727 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1728 1729 static void findReturnsToZap(Function &F, 1730 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1731 SCCPSolver &Solver) { 1732 // We can only do this if we know that nothing else can call the function. 1733 if (!Solver.isArgumentTrackedFunction(&F)) 1734 return; 1735 1736 // There is a non-removable musttail call site of this function. Zapping 1737 // returns is not allowed. 1738 if (Solver.isMustTailCallee(&F)) { 1739 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1740 << " due to present musttail call of it\n"); 1741 return; 1742 } 1743 1744 assert( 1745 all_of(F.users(), 1746 [&Solver](User *U) { 1747 if (isa<Instruction>(U) && 1748 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1749 return true; 1750 // Non-callsite uses are not impacted by zapping. Also, constant 1751 // uses (like blockaddresses) could stuck around, without being 1752 // used in the underlying IR, meaning we do not have lattice 1753 // values for them. 1754 if (!isa<CallBase>(U)) 1755 return true; 1756 if (U->getType()->isStructTy()) { 1757 return all_of(Solver.getStructLatticeValueFor(U), 1758 [](const ValueLatticeElement &LV) { 1759 return !isOverdefined(LV); 1760 }); 1761 } 1762 return !isOverdefined(Solver.getLatticeValueFor(U)); 1763 }) && 1764 "We can only zap functions where all live users have a concrete value"); 1765 1766 for (BasicBlock &BB : F) { 1767 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1768 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1769 << "musttail call : " << *CI << "\n"); 1770 (void)CI; 1771 return; 1772 } 1773 1774 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1775 if (!isa<UndefValue>(RI->getOperand(0))) 1776 ReturnsToZap.push_back(RI); 1777 } 1778 } 1779 1780 // Update the condition for terminators that are branching on indeterminate 1781 // values, forcing them to use a specific edge. 1782 static void forceIndeterminateEdge(Instruction* I, SCCPSolver &Solver) { 1783 BasicBlock *Dest = nullptr; 1784 Constant *C = nullptr; 1785 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1786 if (!isa<ConstantInt>(SI->getCondition())) { 1787 // Indeterminate switch; use first case value. 1788 Dest = SI->case_begin()->getCaseSuccessor(); 1789 C = SI->case_begin()->getCaseValue(); 1790 } 1791 } else if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1792 if (!isa<ConstantInt>(BI->getCondition())) { 1793 // Indeterminate branch; use false. 1794 Dest = BI->getSuccessor(1); 1795 C = ConstantInt::getFalse(BI->getContext()); 1796 } 1797 } else if (IndirectBrInst *IBR = dyn_cast<IndirectBrInst>(I)) { 1798 if (!isa<BlockAddress>(IBR->getAddress()->stripPointerCasts())) { 1799 // Indeterminate indirectbr; use successor 0. 1800 Dest = IBR->getSuccessor(0); 1801 C = BlockAddress::get(IBR->getSuccessor(0)); 1802 } 1803 } else { 1804 llvm_unreachable("Unexpected terminator instruction"); 1805 } 1806 if (C) { 1807 assert(Solver.isEdgeFeasible(I->getParent(), Dest) && 1808 "Didn't find feasible edge?"); 1809 (void)Dest; 1810 1811 I->setOperand(0, C); 1812 } 1813 } 1814 1815 bool llvm::runIPSCCP( 1816 Module &M, const DataLayout &DL, 1817 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1818 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1819 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1820 1821 // Loop over all functions, marking arguments to those with their addresses 1822 // taken or that are external as overdefined. 1823 for (Function &F : M) { 1824 if (F.isDeclaration()) 1825 continue; 1826 1827 Solver.addAnalysis(F, getAnalysis(F)); 1828 1829 // Determine if we can track the function's return values. If so, add the 1830 // function to the solver's set of return-tracked functions. 1831 if (canTrackReturnsInterprocedurally(&F)) 1832 Solver.AddTrackedFunction(&F); 1833 1834 // Determine if we can track the function's arguments. If so, add the 1835 // function to the solver's set of argument-tracked functions. 1836 if (canTrackArgumentsInterprocedurally(&F)) { 1837 Solver.AddArgumentTrackedFunction(&F); 1838 continue; 1839 } 1840 1841 // Assume the function is called. 1842 Solver.MarkBlockExecutable(&F.front()); 1843 1844 // Assume nothing about the incoming arguments. 1845 for (Argument &AI : F.args()) 1846 Solver.markOverdefined(&AI); 1847 } 1848 1849 // Determine if we can track any of the module's global variables. If so, add 1850 // the global variables we can track to the solver's set of tracked global 1851 // variables. 1852 for (GlobalVariable &G : M.globals()) { 1853 G.removeDeadConstantUsers(); 1854 if (canTrackGlobalVariableInterprocedurally(&G)) 1855 Solver.TrackValueOfGlobalVariable(&G); 1856 } 1857 1858 // Solve for constants. 1859 bool ResolvedUndefs = true; 1860 Solver.Solve(); 1861 while (ResolvedUndefs) { 1862 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1863 ResolvedUndefs = false; 1864 for (Function &F : M) 1865 if (Solver.ResolvedUndefsIn(F)) { 1866 // We run Solve() after we resolved an undef in a function, because 1867 // we might deduce a fact that eliminates an undef in another function. 1868 Solver.Solve(); 1869 ResolvedUndefs = true; 1870 } 1871 } 1872 1873 bool MadeChanges = false; 1874 1875 // Iterate over all of the instructions in the module, replacing them with 1876 // constants if we have found them to be of constant values. 1877 1878 for (Function &F : M) { 1879 if (F.isDeclaration()) 1880 continue; 1881 1882 SmallVector<BasicBlock *, 512> BlocksToErase; 1883 1884 if (Solver.isBlockExecutable(&F.front())) 1885 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E; 1886 ++AI) { 1887 if (!AI->use_empty() && tryToReplaceWithConstant(Solver, &*AI)) { 1888 ++IPNumArgsElimed; 1889 continue; 1890 } 1891 } 1892 1893 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 1894 if (!Solver.isBlockExecutable(&*BB)) { 1895 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB); 1896 ++NumDeadBlocks; 1897 1898 MadeChanges = true; 1899 1900 if (&*BB != &F.front()) 1901 BlocksToErase.push_back(&*BB); 1902 continue; 1903 } 1904 1905 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 1906 Instruction *Inst = &*BI++; 1907 if (Inst->getType()->isVoidTy()) 1908 continue; 1909 if (tryToReplaceWithConstant(Solver, Inst)) { 1910 if (Inst->isSafeToRemove()) 1911 Inst->eraseFromParent(); 1912 // Hey, we just changed something! 1913 MadeChanges = true; 1914 ++IPNumInstRemoved; 1915 } 1916 } 1917 } 1918 1919 DomTreeUpdater DTU = Solver.getDTU(F); 1920 // Change dead blocks to unreachable. We do it after replacing constants 1921 // in all executable blocks, because changeToUnreachable may remove PHI 1922 // nodes in executable blocks we found values for. The function's entry 1923 // block is not part of BlocksToErase, so we have to handle it separately. 1924 for (BasicBlock *BB : BlocksToErase) { 1925 NumInstRemoved += 1926 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 1927 /*PreserveLCSSA=*/false, &DTU); 1928 } 1929 if (!Solver.isBlockExecutable(&F.front())) 1930 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 1931 /*UseLLVMTrap=*/false, 1932 /*PreserveLCSSA=*/false, &DTU); 1933 1934 // Now that all instructions in the function are constant folded, 1935 // use ConstantFoldTerminator to get rid of in-edges, record DT updates and 1936 // delete dead BBs. 1937 for (BasicBlock *DeadBB : BlocksToErase) { 1938 // If there are any PHI nodes in this successor, drop entries for BB now. 1939 for (Value::user_iterator UI = DeadBB->user_begin(), 1940 UE = DeadBB->user_end(); 1941 UI != UE;) { 1942 // Grab the user and then increment the iterator early, as the user 1943 // will be deleted. Step past all adjacent uses from the same user. 1944 auto *I = dyn_cast<Instruction>(*UI); 1945 do { ++UI; } while (UI != UE && *UI == I); 1946 1947 // Ignore blockaddress users; BasicBlock's dtor will handle them. 1948 if (!I) continue; 1949 1950 // If we have forced an edge for an indeterminate value, then force the 1951 // terminator to fold to that edge. 1952 forceIndeterminateEdge(I, Solver); 1953 BasicBlock *InstBB = I->getParent(); 1954 bool Folded = ConstantFoldTerminator(InstBB, 1955 /*DeleteDeadConditions=*/false, 1956 /*TLI=*/nullptr, &DTU); 1957 assert(Folded && 1958 "Expect TermInst on constantint or blockaddress to be folded"); 1959 (void) Folded; 1960 // If we folded the terminator to an unconditional branch to another 1961 // dead block, replace it with Unreachable, to avoid trying to fold that 1962 // branch again. 1963 BranchInst *BI = cast<BranchInst>(InstBB->getTerminator()); 1964 if (BI && BI->isUnconditional() && 1965 !Solver.isBlockExecutable(BI->getSuccessor(0))) { 1966 InstBB->getTerminator()->eraseFromParent(); 1967 new UnreachableInst(InstBB->getContext(), InstBB); 1968 } 1969 } 1970 // Mark dead BB for deletion. 1971 DTU.deleteBB(DeadBB); 1972 } 1973 1974 for (BasicBlock &BB : F) { 1975 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1976 Instruction *Inst = &*BI++; 1977 if (Solver.getPredicateInfoFor(Inst)) { 1978 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 1979 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1980 Value *Op = II->getOperand(0); 1981 Inst->replaceAllUsesWith(Op); 1982 Inst->eraseFromParent(); 1983 } 1984 } 1985 } 1986 } 1987 } 1988 } 1989 1990 // If we inferred constant or undef return values for a function, we replaced 1991 // all call uses with the inferred value. This means we don't need to bother 1992 // actually returning anything from the function. Replace all return 1993 // instructions with return undef. 1994 // 1995 // Do this in two stages: first identify the functions we should process, then 1996 // actually zap their returns. This is important because we can only do this 1997 // if the address of the function isn't taken. In cases where a return is the 1998 // last use of a function, the order of processing functions would affect 1999 // whether other functions are optimizable. 2000 SmallVector<ReturnInst*, 8> ReturnsToZap; 2001 2002 for (const auto &I : Solver.getTrackedRetVals()) { 2003 Function *F = I.first; 2004 if (isOverdefined(I.second) || F->getReturnType()->isVoidTy()) 2005 continue; 2006 findReturnsToZap(*F, ReturnsToZap, Solver); 2007 } 2008 2009 for (auto F : Solver.getMRVFunctionsTracked()) { 2010 assert(F->getReturnType()->isStructTy() && 2011 "The return type should be a struct"); 2012 StructType *STy = cast<StructType>(F->getReturnType()); 2013 if (Solver.isStructLatticeConstant(F, STy)) 2014 findReturnsToZap(*F, ReturnsToZap, Solver); 2015 } 2016 2017 // Zap all returns which we've identified as zap to change. 2018 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2019 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2020 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2021 } 2022 2023 // If we inferred constant or undef values for globals variables, we can 2024 // delete the global and any stores that remain to it. 2025 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2026 GlobalVariable *GV = I.first; 2027 if (isOverdefined(I.second)) 2028 continue; 2029 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2030 << "' is constant!\n"); 2031 while (!GV->use_empty()) { 2032 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2033 SI->eraseFromParent(); 2034 } 2035 M.getGlobalList().erase(GV); 2036 ++IPNumGlobalConst; 2037 } 2038 2039 return MadeChanges; 2040 } 2041