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/SetVector.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/Analysis/ConstantFolding.h" 31 #include "llvm/Analysis/DomTreeUpdater.h" 32 #include "llvm/Analysis/GlobalsModRef.h" 33 #include "llvm/Analysis/InstructionSimplify.h" 34 #include "llvm/Analysis/TargetLibraryInfo.h" 35 #include "llvm/Analysis/ValueLattice.h" 36 #include "llvm/Analysis/ValueLatticeUtils.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/IR/BasicBlock.h" 39 #include "llvm/IR/Constant.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/DerivedTypes.h" 43 #include "llvm/IR/Function.h" 44 #include "llvm/IR/GlobalVariable.h" 45 #include "llvm/IR/InstVisitor.h" 46 #include "llvm/IR/InstrTypes.h" 47 #include "llvm/IR/Instruction.h" 48 #include "llvm/IR/Instructions.h" 49 #include "llvm/IR/Module.h" 50 #include "llvm/IR/PassManager.h" 51 #include "llvm/IR/Type.h" 52 #include "llvm/IR/User.h" 53 #include "llvm/IR/Value.h" 54 #include "llvm/InitializePasses.h" 55 #include "llvm/Pass.h" 56 #include "llvm/Support/Casting.h" 57 #include "llvm/Support/Debug.h" 58 #include "llvm/Support/ErrorHandling.h" 59 #include "llvm/Support/raw_ostream.h" 60 #include "llvm/Transforms/Scalar.h" 61 #include "llvm/Transforms/Utils/Local.h" 62 #include "llvm/Transforms/Utils/PredicateInfo.h" 63 #include <cassert> 64 #include <utility> 65 #include <vector> 66 67 using namespace llvm; 68 69 #define DEBUG_TYPE "sccp" 70 71 STATISTIC(NumInstRemoved, "Number of instructions removed"); 72 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 73 STATISTIC(NumInstReplaced, 74 "Number of instructions replaced with (simpler) instruction"); 75 76 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 77 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 78 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 79 STATISTIC( 80 IPNumInstReplaced, 81 "Number of instructions replaced with (simpler) instruction by IPSCCP"); 82 83 // The maximum number of range extensions allowed for operations requiring 84 // widening. 85 static const unsigned MaxNumRangeExtensions = 10; 86 87 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 88 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 89 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 90 MaxNumRangeExtensions); 91 } 92 namespace { 93 94 // Helper to check if \p LV is either a constant or a constant 95 // range with a single element. This should cover exactly the same cases as the 96 // old ValueLatticeElement::isConstant() and is intended to be used in the 97 // transition to ValueLatticeElement. 98 bool isConstant(const ValueLatticeElement &LV) { 99 return LV.isConstant() || 100 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 101 } 102 103 // Helper to check if \p LV is either overdefined or a constant range with more 104 // than a single element. This should cover exactly the same cases as the old 105 // ValueLatticeElement::isOverdefined() and is intended to be used in the 106 // transition to ValueLatticeElement. 107 bool isOverdefined(const ValueLatticeElement &LV) { 108 return !LV.isUnknownOrUndef() && !isConstant(LV); 109 } 110 111 //===----------------------------------------------------------------------===// 112 // 113 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 114 /// Constant Propagation. 115 /// 116 class SCCPSolver : public InstVisitor<SCCPSolver> { 117 const DataLayout &DL; 118 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 119 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 120 DenseMap<Value *, ValueLatticeElement> 121 ValueState; // The state each value is in. 122 123 /// StructValueState - This maintains ValueState for values that have 124 /// StructType, for example for formal arguments, calls, insertelement, etc. 125 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 126 127 /// GlobalValue - If we are tracking any values for the contents of a global 128 /// variable, we keep a mapping from the constant accessor to the element of 129 /// the global, to the currently known value. If the value becomes 130 /// overdefined, it's entry is simply removed from this map. 131 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 132 133 /// TrackedRetVals - If we are tracking arguments into and the return 134 /// value out of a function, it will have an entry in this map, indicating 135 /// what the known return value for the function is. 136 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 137 138 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 139 /// that return multiple values. 140 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 141 TrackedMultipleRetVals; 142 143 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 144 /// represented here for efficient lookup. 145 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 146 147 /// A list of functions whose return cannot be modified. 148 SmallPtrSet<Function *, 16> MustPreserveReturnsInFunctions; 149 150 /// TrackingIncomingArguments - This is the set of functions for whose 151 /// arguments we make optimistic assumptions about and try to prove as 152 /// constants. 153 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 154 155 /// The reason for two worklists is that overdefined is the lowest state 156 /// on the lattice, and moving things to overdefined as fast as possible 157 /// makes SCCP converge much faster. 158 /// 159 /// By having a separate worklist, we accomplish this because everything 160 /// possibly overdefined will become overdefined at the soonest possible 161 /// point. 162 SmallVector<Value *, 64> OverdefinedInstWorkList; 163 SmallVector<Value *, 64> InstWorkList; 164 165 // The BasicBlock work list 166 SmallVector<BasicBlock *, 64> BBWorkList; 167 168 /// KnownFeasibleEdges - Entries in this set are edges which have already had 169 /// PHI nodes retriggered. 170 using Edge = std::pair<BasicBlock *, BasicBlock *>; 171 DenseSet<Edge> KnownFeasibleEdges; 172 173 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 174 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 175 176 LLVMContext &Ctx; 177 178 public: 179 void addAnalysis(Function &F, AnalysisResultsForFn A) { 180 AnalysisResults.insert({&F, std::move(A)}); 181 } 182 183 const PredicateBase *getPredicateInfoFor(Instruction *I) { 184 auto A = AnalysisResults.find(I->getParent()->getParent()); 185 if (A == AnalysisResults.end()) 186 return nullptr; 187 return A->second.PredInfo->getPredicateInfoFor(I); 188 } 189 190 DomTreeUpdater getDTU(Function &F) { 191 auto A = AnalysisResults.find(&F); 192 assert(A != AnalysisResults.end() && "Need analysis results for function."); 193 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 194 } 195 196 SCCPSolver(const DataLayout &DL, 197 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 198 LLVMContext &Ctx) 199 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 200 201 /// MarkBlockExecutable - This method can be used by clients to mark all of 202 /// the blocks that are known to be intrinsically live in the processed unit. 203 /// 204 /// This returns true if the block was not considered live before. 205 bool MarkBlockExecutable(BasicBlock *BB) { 206 if (!BBExecutable.insert(BB).second) 207 return false; 208 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 209 BBWorkList.push_back(BB); // Add the block to the work list! 210 return true; 211 } 212 213 /// TrackValueOfGlobalVariable - Clients can use this method to 214 /// inform the SCCPSolver that it should track loads and stores to the 215 /// specified global variable if it can. This is only legal to call if 216 /// performing Interprocedural SCCP. 217 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 218 // We only track the contents of scalar globals. 219 if (GV->getValueType()->isSingleValueType()) { 220 ValueLatticeElement &IV = TrackedGlobals[GV]; 221 if (!isa<UndefValue>(GV->getInitializer())) 222 IV.markConstant(GV->getInitializer()); 223 } 224 } 225 226 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 227 /// and out of the specified function (which cannot have its address taken), 228 /// this method must be called. 229 void AddTrackedFunction(Function *F) { 230 // Add an entry, F -> undef. 231 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 232 MRVFunctionsTracked.insert(F); 233 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 234 TrackedMultipleRetVals.insert( 235 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 236 } else if (!F->getReturnType()->isVoidTy()) 237 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 238 } 239 240 /// Add function to the list of functions whose return cannot be modified. 241 void addToMustPreserveReturnsInFunctions(Function *F) { 242 MustPreserveReturnsInFunctions.insert(F); 243 } 244 245 /// Returns true if the return of the given function cannot be modified. 246 bool mustPreserveReturn(Function *F) { 247 return MustPreserveReturnsInFunctions.count(F); 248 } 249 250 void AddArgumentTrackedFunction(Function *F) { 251 TrackingIncomingArguments.insert(F); 252 } 253 254 /// Returns true if the given function is in the solver's set of 255 /// argument-tracked functions. 256 bool isArgumentTrackedFunction(Function *F) { 257 return TrackingIncomingArguments.count(F); 258 } 259 260 /// Solve - Solve for constants and executable blocks. 261 void Solve(); 262 263 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 264 /// that branches on undef values cannot reach any of their successors. 265 /// However, this is not a safe assumption. After we solve dataflow, this 266 /// method should be use to handle this. If this returns true, the solver 267 /// should be rerun. 268 bool ResolvedUndefsIn(Function &F); 269 270 bool isBlockExecutable(BasicBlock *BB) const { 271 return BBExecutable.count(BB); 272 } 273 274 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 275 // block to the 'To' basic block is currently feasible. 276 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 277 278 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 279 std::vector<ValueLatticeElement> StructValues; 280 auto *STy = dyn_cast<StructType>(V->getType()); 281 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 282 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 283 auto I = StructValueState.find(std::make_pair(V, i)); 284 assert(I != StructValueState.end() && "Value not in valuemap!"); 285 StructValues.push_back(I->second); 286 } 287 return StructValues; 288 } 289 290 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 291 292 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 293 assert(!V->getType()->isStructTy() && 294 "Should use getStructLatticeValueFor"); 295 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 296 ValueState.find(V); 297 assert(I != ValueState.end() && 298 "V not found in ValueState nor Paramstate map!"); 299 return I->second; 300 } 301 302 /// getTrackedRetVals - Get the inferred return value map. 303 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 304 return TrackedRetVals; 305 } 306 307 /// getTrackedGlobals - Get and return the set of inferred initializers for 308 /// global variables. 309 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 310 return TrackedGlobals; 311 } 312 313 /// getMRVFunctionsTracked - Get the set of functions which return multiple 314 /// values tracked by the pass. 315 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 316 return MRVFunctionsTracked; 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 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 644 if (ConstantInt *CI = getConstantInt(SCValue)) { 645 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 646 return; 647 } 648 649 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 650 // is ready. 651 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 652 const ConstantRange &Range = SCValue.getConstantRange(); 653 for (const auto &Case : SI->cases()) { 654 const APInt &CaseValue = Case.getCaseValue()->getValue(); 655 if (Range.contains(CaseValue)) 656 Succs[Case.getSuccessorIndex()] = true; 657 } 658 659 // TODO: Determine whether default case is reachable. 660 Succs[SI->case_default()->getSuccessorIndex()] = true; 661 return; 662 } 663 664 // Overdefined or unknown condition? All destinations are executable! 665 if (!SCValue.isUnknownOrUndef()) 666 Succs.assign(TI.getNumSuccessors(), true); 667 return; 668 } 669 670 // In case of indirect branch and its address is a blockaddress, we mark 671 // the target as executable. 672 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 673 // Casts are folded by visitCastInst. 674 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 675 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 676 if (!Addr) { // Overdefined or unknown condition? 677 // All destinations are executable! 678 if (!IBRValue.isUnknownOrUndef()) 679 Succs.assign(TI.getNumSuccessors(), true); 680 return; 681 } 682 683 BasicBlock* T = Addr->getBasicBlock(); 684 assert(Addr->getFunction() == T->getParent() && 685 "Block address of a different function ?"); 686 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 687 // This is the target. 688 if (IBR->getDestination(i) == T) { 689 Succs[i] = true; 690 return; 691 } 692 } 693 694 // If we didn't find our destination in the IBR successor list, then we 695 // have undefined behavior. Its ok to assume no successor is executable. 696 return; 697 } 698 699 // In case of callbr, we pessimistically assume that all successors are 700 // feasible. 701 if (isa<CallBrInst>(&TI)) { 702 Succs.assign(TI.getNumSuccessors(), true); 703 return; 704 } 705 706 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 707 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 708 } 709 710 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 711 // block to the 'To' basic block is currently feasible. 712 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 713 // Check if we've called markEdgeExecutable on the edge yet. (We could 714 // be more aggressive and try to consider edges which haven't been marked 715 // yet, but there isn't any need.) 716 return KnownFeasibleEdges.count(Edge(From, To)); 717 } 718 719 // visit Implementations - Something changed in this instruction, either an 720 // operand made a transition, or the instruction is newly executable. Change 721 // the value type of I to reflect these changes if appropriate. This method 722 // makes sure to do the following actions: 723 // 724 // 1. If a phi node merges two constants in, and has conflicting value coming 725 // from different branches, or if the PHI node merges in an overdefined 726 // value, then the PHI node becomes overdefined. 727 // 2. If a phi node merges only constants in, and they all agree on value, the 728 // PHI node becomes a constant value equal to that. 729 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 730 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 731 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 732 // 6. If a conditional branch has a value that is constant, make the selected 733 // destination executable 734 // 7. If a conditional branch has a value that is overdefined, make all 735 // successors executable. 736 void SCCPSolver::visitPHINode(PHINode &PN) { 737 // If this PN returns a struct, just mark the result overdefined. 738 // TODO: We could do a lot better than this if code actually uses this. 739 if (PN.getType()->isStructTy()) 740 return (void)markOverdefined(&PN); 741 742 if (getValueState(&PN).isOverdefined()) 743 return; // Quick exit 744 745 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 746 // and slow us down a lot. Just mark them overdefined. 747 if (PN.getNumIncomingValues() > 64) 748 return (void)markOverdefined(&PN); 749 750 unsigned NumActiveIncoming = 0; 751 752 // Look at all of the executable operands of the PHI node. If any of them 753 // are overdefined, the PHI becomes overdefined as well. If they are all 754 // constant, and they agree with each other, the PHI becomes the identical 755 // constant. If they are constant and don't agree, the PHI is a constant 756 // range. If there are no executable operands, the PHI remains unknown. 757 ValueLatticeElement PhiState = getValueState(&PN); 758 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 759 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 760 continue; 761 762 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 763 PhiState.mergeIn(IV); 764 NumActiveIncoming++; 765 if (PhiState.isOverdefined()) 766 break; 767 } 768 769 // We allow up to 1 range extension per active incoming value and one 770 // additional extension. Note that we manually adjust the number of range 771 // extensions to match the number of active incoming values. This helps to 772 // limit multiple extensions caused by the same incoming value, if other 773 // incoming values are equal. 774 mergeInValue(&PN, PhiState, 775 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 776 NumActiveIncoming + 1)); 777 ValueLatticeElement &PhiStateRef = getValueState(&PN); 778 PhiStateRef.setNumRangeExtensions( 779 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 780 } 781 782 void SCCPSolver::visitReturnInst(ReturnInst &I) { 783 if (I.getNumOperands() == 0) return; // ret void 784 785 Function *F = I.getParent()->getParent(); 786 Value *ResultOp = I.getOperand(0); 787 788 // If we are tracking the return value of this function, merge it in. 789 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 790 auto TFRVI = TrackedRetVals.find(F); 791 if (TFRVI != TrackedRetVals.end()) { 792 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 793 return; 794 } 795 } 796 797 // Handle functions that return multiple values. 798 if (!TrackedMultipleRetVals.empty()) { 799 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 800 if (MRVFunctionsTracked.count(F)) 801 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 802 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 803 getStructValueState(ResultOp, i)); 804 } 805 } 806 807 void SCCPSolver::visitTerminator(Instruction &TI) { 808 SmallVector<bool, 16> SuccFeasible; 809 getFeasibleSuccessors(TI, SuccFeasible); 810 811 BasicBlock *BB = TI.getParent(); 812 813 // Mark all feasible successors executable. 814 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 815 if (SuccFeasible[i]) 816 markEdgeExecutable(BB, TI.getSuccessor(i)); 817 } 818 819 void SCCPSolver::visitCastInst(CastInst &I) { 820 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 821 // discover a concrete value later. 822 if (ValueState[&I].isOverdefined()) 823 return; 824 825 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 826 if (Constant *OpC = getConstant(OpSt)) { 827 // Fold the constant as we build. 828 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 829 if (isa<UndefValue>(C)) 830 return; 831 // Propagate constant value 832 markConstant(&I, C); 833 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 834 auto &LV = getValueState(&I); 835 ConstantRange OpRange = OpSt.getConstantRange(); 836 Type *DestTy = I.getDestTy(); 837 // Vectors where all elements have the same known constant range are treated 838 // as a single constant range in the lattice. When bitcasting such vectors, 839 // there is a mis-match between the width of the lattice value (single 840 // constant range) and the original operands (vector). Go to overdefined in 841 // that case. 842 if (I.getOpcode() == Instruction::BitCast && 843 I.getOperand(0)->getType()->isVectorTy() && 844 OpRange.getBitWidth() < DL.getTypeSizeInBits(DestTy)) 845 return (void)markOverdefined(&I); 846 847 ConstantRange Res = 848 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 849 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 850 } else if (!OpSt.isUnknownOrUndef()) 851 markOverdefined(&I); 852 } 853 854 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 855 // If this returns a struct, mark all elements over defined, we don't track 856 // structs in structs. 857 if (EVI.getType()->isStructTy()) 858 return (void)markOverdefined(&EVI); 859 860 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 861 // discover a concrete value later. 862 if (ValueState[&EVI].isOverdefined()) 863 return (void)markOverdefined(&EVI); 864 865 // If this is extracting from more than one level of struct, we don't know. 866 if (EVI.getNumIndices() != 1) 867 return (void)markOverdefined(&EVI); 868 869 Value *AggVal = EVI.getAggregateOperand(); 870 if (AggVal->getType()->isStructTy()) { 871 unsigned i = *EVI.idx_begin(); 872 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 873 mergeInValue(getValueState(&EVI), &EVI, EltVal); 874 } else { 875 // Otherwise, must be extracting from an array. 876 return (void)markOverdefined(&EVI); 877 } 878 } 879 880 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 881 auto *STy = dyn_cast<StructType>(IVI.getType()); 882 if (!STy) 883 return (void)markOverdefined(&IVI); 884 885 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 886 // discover a concrete value later. 887 if (isOverdefined(ValueState[&IVI])) 888 return (void)markOverdefined(&IVI); 889 890 // If this has more than one index, we can't handle it, drive all results to 891 // undef. 892 if (IVI.getNumIndices() != 1) 893 return (void)markOverdefined(&IVI); 894 895 Value *Aggr = IVI.getAggregateOperand(); 896 unsigned Idx = *IVI.idx_begin(); 897 898 // Compute the result based on what we're inserting. 899 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 900 // This passes through all values that aren't the inserted element. 901 if (i != Idx) { 902 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 903 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 904 continue; 905 } 906 907 Value *Val = IVI.getInsertedValueOperand(); 908 if (Val->getType()->isStructTy()) 909 // We don't track structs in structs. 910 markOverdefined(getStructValueState(&IVI, i), &IVI); 911 else { 912 ValueLatticeElement InVal = getValueState(Val); 913 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 914 } 915 } 916 } 917 918 void SCCPSolver::visitSelectInst(SelectInst &I) { 919 // If this select returns a struct, just mark the result overdefined. 920 // TODO: We could do a lot better than this if code actually uses this. 921 if (I.getType()->isStructTy()) 922 return (void)markOverdefined(&I); 923 924 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 925 // discover a concrete value later. 926 if (ValueState[&I].isOverdefined()) 927 return (void)markOverdefined(&I); 928 929 ValueLatticeElement CondValue = getValueState(I.getCondition()); 930 if (CondValue.isUnknownOrUndef()) 931 return; 932 933 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 934 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 935 mergeInValue(&I, getValueState(OpVal)); 936 return; 937 } 938 939 // Otherwise, the condition is overdefined or a constant we can't evaluate. 940 // See if we can produce something better than overdefined based on the T/F 941 // value. 942 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 943 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 944 945 bool Changed = ValueState[&I].mergeIn(TVal); 946 Changed |= ValueState[&I].mergeIn(FVal); 947 if (Changed) 948 pushToWorkListMsg(ValueState[&I], &I); 949 } 950 951 // Handle Unary Operators. 952 void SCCPSolver::visitUnaryOperator(Instruction &I) { 953 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 954 955 ValueLatticeElement &IV = ValueState[&I]; 956 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 957 // discover a concrete value later. 958 if (isOverdefined(IV)) 959 return (void)markOverdefined(&I); 960 961 if (isConstant(V0State)) { 962 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 963 964 // op Y -> undef. 965 if (isa<UndefValue>(C)) 966 return; 967 return (void)markConstant(IV, &I, C); 968 } 969 970 // If something is undef, wait for it to resolve. 971 if (!isOverdefined(V0State)) 972 return; 973 974 markOverdefined(&I); 975 } 976 977 // Handle Binary Operators. 978 void SCCPSolver::visitBinaryOperator(Instruction &I) { 979 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 980 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 981 982 ValueLatticeElement &IV = ValueState[&I]; 983 if (IV.isOverdefined()) 984 return; 985 986 // If something is undef, wait for it to resolve. 987 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 988 return; 989 990 if (V1State.isOverdefined() && V2State.isOverdefined()) 991 return (void)markOverdefined(&I); 992 993 // If either of the operands is a constant, try to fold it to a constant. 994 // TODO: Use information from notconstant better. 995 if ((V1State.isConstant() || V2State.isConstant())) { 996 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 997 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 998 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 999 auto *C = dyn_cast_or_null<Constant>(R); 1000 if (C) { 1001 // X op Y -> undef. 1002 if (isa<UndefValue>(C)) 1003 return; 1004 // Conservatively assume that the result may be based on operands that may 1005 // be undef. Note that we use mergeInValue to combine the constant with 1006 // the existing lattice value for I, as different constants might be found 1007 // after one of the operands go to overdefined, e.g. due to one operand 1008 // being a special floating value. 1009 ValueLatticeElement NewV; 1010 NewV.markConstant(C, /*MayIncludeUndef=*/true); 1011 return (void)mergeInValue(&I, NewV); 1012 } 1013 } 1014 1015 // Only use ranges for binary operators on integers. 1016 if (!I.getType()->isIntegerTy()) 1017 return markOverdefined(&I); 1018 1019 // Try to simplify to a constant range. 1020 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1021 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1022 if (V1State.isConstantRange()) 1023 A = V1State.getConstantRange(); 1024 if (V2State.isConstantRange()) 1025 B = V2State.getConstantRange(); 1026 1027 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1028 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1029 1030 // TODO: Currently we do not exploit special values that produce something 1031 // better than overdefined with an overdefined operand for vector or floating 1032 // point types, like and <4 x i32> overdefined, zeroinitializer. 1033 } 1034 1035 // Handle ICmpInst instruction. 1036 void SCCPSolver::visitCmpInst(CmpInst &I) { 1037 // Do not cache this lookup, getValueState calls later in the function might 1038 // invalidate the reference. 1039 if (isOverdefined(ValueState[&I])) 1040 return (void)markOverdefined(&I); 1041 1042 Value *Op1 = I.getOperand(0); 1043 Value *Op2 = I.getOperand(1); 1044 1045 // For parameters, use ParamState which includes constant range info if 1046 // available. 1047 auto V1State = getValueState(Op1); 1048 auto V2State = getValueState(Op2); 1049 1050 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1051 if (C) { 1052 if (isa<UndefValue>(C)) 1053 return; 1054 ValueLatticeElement CV; 1055 CV.markConstant(C); 1056 mergeInValue(&I, CV); 1057 return; 1058 } 1059 1060 // If operands are still unknown, wait for it to resolve. 1061 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1062 !isConstant(ValueState[&I])) 1063 return; 1064 1065 markOverdefined(&I); 1066 } 1067 1068 // Handle getelementptr instructions. If all operands are constants then we 1069 // can turn this into a getelementptr ConstantExpr. 1070 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1071 if (isOverdefined(ValueState[&I])) 1072 return (void)markOverdefined(&I); 1073 1074 SmallVector<Constant*, 8> Operands; 1075 Operands.reserve(I.getNumOperands()); 1076 1077 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1078 ValueLatticeElement State = getValueState(I.getOperand(i)); 1079 if (State.isUnknownOrUndef()) 1080 return; // Operands are not resolved yet. 1081 1082 if (isOverdefined(State)) 1083 return (void)markOverdefined(&I); 1084 1085 if (Constant *C = getConstant(State)) { 1086 Operands.push_back(C); 1087 continue; 1088 } 1089 1090 return (void)markOverdefined(&I); 1091 } 1092 1093 Constant *Ptr = Operands[0]; 1094 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1095 Constant *C = 1096 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1097 if (isa<UndefValue>(C)) 1098 return; 1099 markConstant(&I, C); 1100 } 1101 1102 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1103 // If this store is of a struct, ignore it. 1104 if (SI.getOperand(0)->getType()->isStructTy()) 1105 return; 1106 1107 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1108 return; 1109 1110 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1111 auto I = TrackedGlobals.find(GV); 1112 if (I == TrackedGlobals.end()) 1113 return; 1114 1115 // Get the value we are storing into the global, then merge it. 1116 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1117 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1118 if (I->second.isOverdefined()) 1119 TrackedGlobals.erase(I); // No need to keep tracking this! 1120 } 1121 1122 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1123 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1124 if (I->getType()->isIntegerTy()) 1125 return ValueLatticeElement::getRange( 1126 getConstantRangeFromMetadata(*Ranges)); 1127 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1128 return ValueLatticeElement::getNot( 1129 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1130 return ValueLatticeElement::getOverdefined(); 1131 } 1132 1133 // Handle load instructions. If the operand is a constant pointer to a constant 1134 // global, we can replace the load with the loaded constant value! 1135 void SCCPSolver::visitLoadInst(LoadInst &I) { 1136 // If this load is of a struct or the load is volatile, just mark the result 1137 // as overdefined. 1138 if (I.getType()->isStructTy() || I.isVolatile()) 1139 return (void)markOverdefined(&I); 1140 1141 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1142 // discover a concrete value later. 1143 if (ValueState[&I].isOverdefined()) 1144 return (void)markOverdefined(&I); 1145 1146 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1147 if (PtrVal.isUnknownOrUndef()) 1148 return; // The pointer is not resolved yet! 1149 1150 ValueLatticeElement &IV = ValueState[&I]; 1151 1152 if (isConstant(PtrVal)) { 1153 Constant *Ptr = getConstant(PtrVal); 1154 1155 // load null is undefined. 1156 if (isa<ConstantPointerNull>(Ptr)) { 1157 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1158 return (void)markOverdefined(IV, &I); 1159 else 1160 return; 1161 } 1162 1163 // Transform load (constant global) into the value loaded. 1164 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1165 if (!TrackedGlobals.empty()) { 1166 // If we are tracking this global, merge in the known value for it. 1167 auto It = TrackedGlobals.find(GV); 1168 if (It != TrackedGlobals.end()) { 1169 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1170 return; 1171 } 1172 } 1173 } 1174 1175 // Transform load from a constant into a constant if possible. 1176 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1177 if (isa<UndefValue>(C)) 1178 return; 1179 return (void)markConstant(IV, &I, C); 1180 } 1181 } 1182 1183 // Fall back to metadata. 1184 mergeInValue(&I, getValueFromMetadata(&I)); 1185 } 1186 1187 void SCCPSolver::visitCallBase(CallBase &CB) { 1188 handleCallResult(CB); 1189 handleCallArguments(CB); 1190 } 1191 1192 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1193 Function *F = CB.getCalledFunction(); 1194 1195 // Void return and not tracking callee, just bail. 1196 if (CB.getType()->isVoidTy()) 1197 return; 1198 1199 // Always mark struct return as overdefined. 1200 if (CB.getType()->isStructTy()) 1201 return (void)markOverdefined(&CB); 1202 1203 // Otherwise, if we have a single return value case, and if the function is 1204 // a declaration, maybe we can constant fold it. 1205 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1206 SmallVector<Constant *, 8> Operands; 1207 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1208 if (AI->get()->getType()->isStructTy()) 1209 return markOverdefined(&CB); // Can't handle struct args. 1210 ValueLatticeElement State = getValueState(*AI); 1211 1212 if (State.isUnknownOrUndef()) 1213 return; // Operands are not resolved yet. 1214 if (isOverdefined(State)) 1215 return (void)markOverdefined(&CB); 1216 assert(isConstant(State) && "Unknown state!"); 1217 Operands.push_back(getConstant(State)); 1218 } 1219 1220 if (isOverdefined(getValueState(&CB))) 1221 return (void)markOverdefined(&CB); 1222 1223 // If we can constant fold this, mark the result of the call as a 1224 // constant. 1225 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1226 // call -> undef. 1227 if (isa<UndefValue>(C)) 1228 return; 1229 return (void)markConstant(&CB, C); 1230 } 1231 } 1232 1233 // Fall back to metadata. 1234 mergeInValue(&CB, getValueFromMetadata(&CB)); 1235 } 1236 1237 void SCCPSolver::handleCallArguments(CallBase &CB) { 1238 Function *F = CB.getCalledFunction(); 1239 // If this is a local function that doesn't have its address taken, mark its 1240 // entry block executable and merge in the actual arguments to the call into 1241 // the formal arguments of the function. 1242 if (!TrackingIncomingArguments.empty() && 1243 TrackingIncomingArguments.count(F)) { 1244 MarkBlockExecutable(&F->front()); 1245 1246 // Propagate information from this call site into the callee. 1247 auto CAI = CB.arg_begin(); 1248 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1249 ++AI, ++CAI) { 1250 // If this argument is byval, and if the function is not readonly, there 1251 // will be an implicit copy formed of the input aggregate. 1252 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1253 markOverdefined(&*AI); 1254 continue; 1255 } 1256 1257 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1258 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1259 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1260 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1261 getMaxWidenStepsOpts()); 1262 } 1263 } else 1264 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1265 } 1266 } 1267 } 1268 1269 void SCCPSolver::handleCallResult(CallBase &CB) { 1270 Function *F = CB.getCalledFunction(); 1271 1272 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1273 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1274 if (ValueState[&CB].isOverdefined()) 1275 return; 1276 1277 Value *CopyOf = CB.getOperand(0); 1278 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1279 auto *PI = getPredicateInfoFor(&CB); 1280 assert(PI && "Missing predicate info for ssa.copy"); 1281 1282 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1283 if (!Constraint) { 1284 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1285 return; 1286 } 1287 1288 CmpInst::Predicate Pred = Constraint->Predicate; 1289 Value *OtherOp = Constraint->OtherOp; 1290 1291 // Wait until OtherOp is resolved. 1292 if (getValueState(OtherOp).isUnknown()) { 1293 addAdditionalUser(OtherOp, &CB); 1294 return; 1295 } 1296 1297 // TODO: Actually filp MayIncludeUndef for the created range to false, 1298 // once most places in the optimizer respect the branches on 1299 // undef/poison are UB rule. The reason why the new range cannot be 1300 // undef is as follows below: 1301 // The new range is based on a branch condition. That guarantees that 1302 // neither of the compare operands can be undef in the branch targets, 1303 // unless we have conditions that are always true/false (e.g. icmp ule 1304 // i32, %a, i32_max). For the latter overdefined/empty range will be 1305 // inferred, but the branch will get folded accordingly anyways. 1306 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1307 1308 ValueLatticeElement CondVal = getValueState(OtherOp); 1309 ValueLatticeElement &IV = ValueState[&CB]; 1310 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1311 auto ImposedCR = 1312 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1313 1314 // Get the range imposed by the condition. 1315 if (CondVal.isConstantRange()) 1316 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1317 Pred, CondVal.getConstantRange()); 1318 1319 // Combine range info for the original value with the new range from the 1320 // condition. 1321 auto CopyOfCR = CopyOfVal.isConstantRange() 1322 ? CopyOfVal.getConstantRange() 1323 : ConstantRange::getFull( 1324 DL.getTypeSizeInBits(CopyOf->getType())); 1325 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1326 // If the existing information is != x, do not use the information from 1327 // a chained predicate, as the != x information is more likely to be 1328 // helpful in practice. 1329 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1330 NewCR = CopyOfCR; 1331 1332 addAdditionalUser(OtherOp, &CB); 1333 mergeInValue( 1334 IV, &CB, 1335 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1336 return; 1337 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1338 // For non-integer values or integer constant expressions, only 1339 // propagate equal constants. 1340 addAdditionalUser(OtherOp, &CB); 1341 mergeInValue(IV, &CB, CondVal); 1342 return; 1343 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1344 !MayIncludeUndef) { 1345 // Propagate inequalities. 1346 addAdditionalUser(OtherOp, &CB); 1347 mergeInValue(IV, &CB, 1348 ValueLatticeElement::getNot(CondVal.getConstant())); 1349 return; 1350 } 1351 1352 return (void)mergeInValue(IV, &CB, CopyOfVal); 1353 } 1354 1355 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { 1356 // Compute result range for intrinsics supported by ConstantRange. 1357 // Do this even if we don't know a range for all operands, as we may 1358 // still know something about the result range, e.g. of abs(x). 1359 SmallVector<ConstantRange, 2> OpRanges; 1360 for (Value *Op : II->args()) { 1361 const ValueLatticeElement &State = getValueState(Op); 1362 if (State.isConstantRange()) 1363 OpRanges.push_back(State.getConstantRange()); 1364 else 1365 OpRanges.push_back( 1366 ConstantRange::getFull(Op->getType()->getScalarSizeInBits())); 1367 } 1368 1369 ConstantRange Result = 1370 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges); 1371 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result)); 1372 } 1373 } 1374 1375 // The common case is that we aren't tracking the callee, either because we 1376 // are not doing interprocedural analysis or the callee is indirect, or is 1377 // external. Handle these cases first. 1378 if (!F || F->isDeclaration()) 1379 return handleCallOverdefined(CB); 1380 1381 // If this is a single/zero retval case, see if we're tracking the function. 1382 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1383 if (!MRVFunctionsTracked.count(F)) 1384 return handleCallOverdefined(CB); // Not tracking this callee. 1385 1386 // If we are tracking this callee, propagate the result of the function 1387 // into this call site. 1388 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1389 mergeInValue(getStructValueState(&CB, i), &CB, 1390 TrackedMultipleRetVals[std::make_pair(F, i)], 1391 getMaxWidenStepsOpts()); 1392 } else { 1393 auto TFRVI = TrackedRetVals.find(F); 1394 if (TFRVI == TrackedRetVals.end()) 1395 return handleCallOverdefined(CB); // Not tracking this callee. 1396 1397 // If so, propagate the return value of the callee into this call result. 1398 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1399 } 1400 } 1401 1402 void SCCPSolver::Solve() { 1403 // Process the work lists until they are empty! 1404 while (!BBWorkList.empty() || !InstWorkList.empty() || 1405 !OverdefinedInstWorkList.empty()) { 1406 // Process the overdefined instruction's work list first, which drives other 1407 // things to overdefined more quickly. 1408 while (!OverdefinedInstWorkList.empty()) { 1409 Value *I = OverdefinedInstWorkList.pop_back_val(); 1410 1411 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1412 1413 // "I" got into the work list because it either made the transition from 1414 // bottom to constant, or to overdefined. 1415 // 1416 // Anything on this worklist that is overdefined need not be visited 1417 // since all of its users will have already been marked as overdefined 1418 // Update all of the users of this instruction's value. 1419 // 1420 markUsersAsChanged(I); 1421 } 1422 1423 // Process the instruction work list. 1424 while (!InstWorkList.empty()) { 1425 Value *I = InstWorkList.pop_back_val(); 1426 1427 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1428 1429 // "I" got into the work list because it made the transition from undef to 1430 // constant. 1431 // 1432 // Anything on this worklist that is overdefined need not be visited 1433 // since all of its users will have already been marked as overdefined. 1434 // Update all of the users of this instruction's value. 1435 // 1436 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1437 markUsersAsChanged(I); 1438 } 1439 1440 // Process the basic block work list. 1441 while (!BBWorkList.empty()) { 1442 BasicBlock *BB = BBWorkList.pop_back_val(); 1443 1444 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1445 1446 // Notify all instructions in this basic block that they are newly 1447 // executable. 1448 visit(BB); 1449 } 1450 } 1451 } 1452 1453 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1454 /// that branches on undef values cannot reach any of their successors. 1455 /// However, this is not a safe assumption. After we solve dataflow, this 1456 /// method should be use to handle this. If this returns true, the solver 1457 /// should be rerun. 1458 /// 1459 /// This method handles this by finding an unresolved branch and marking it one 1460 /// of the edges from the block as being feasible, even though the condition 1461 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1462 /// CFG and only slightly pessimizes the analysis results (by marking one, 1463 /// potentially infeasible, edge feasible). This cannot usefully modify the 1464 /// constraints on the condition of the branch, as that would impact other users 1465 /// of the value. 1466 /// 1467 /// This scan also checks for values that use undefs. It conservatively marks 1468 /// them as overdefined. 1469 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1470 bool MadeChange = false; 1471 for (BasicBlock &BB : F) { 1472 if (!BBExecutable.count(&BB)) 1473 continue; 1474 1475 for (Instruction &I : BB) { 1476 // Look for instructions which produce undef values. 1477 if (I.getType()->isVoidTy()) continue; 1478 1479 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1480 // Only a few things that can be structs matter for undef. 1481 1482 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1483 if (auto *CB = dyn_cast<CallBase>(&I)) 1484 if (Function *F = CB->getCalledFunction()) 1485 if (MRVFunctionsTracked.count(F)) 1486 continue; 1487 1488 // extractvalue and insertvalue don't need to be marked; they are 1489 // tracked as precisely as their operands. 1490 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1491 continue; 1492 // Send the results of everything else to overdefined. We could be 1493 // more precise than this but it isn't worth bothering. 1494 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1495 ValueLatticeElement &LV = getStructValueState(&I, i); 1496 if (LV.isUnknownOrUndef()) { 1497 markOverdefined(LV, &I); 1498 MadeChange = true; 1499 } 1500 } 1501 continue; 1502 } 1503 1504 ValueLatticeElement &LV = getValueState(&I); 1505 if (!LV.isUnknownOrUndef()) 1506 continue; 1507 1508 // There are two reasons a call can have an undef result 1509 // 1. It could be tracked. 1510 // 2. It could be constant-foldable. 1511 // Because of the way we solve return values, tracked calls must 1512 // never be marked overdefined in ResolvedUndefsIn. 1513 if (auto *CB = dyn_cast<CallBase>(&I)) 1514 if (Function *F = CB->getCalledFunction()) 1515 if (TrackedRetVals.count(F)) 1516 continue; 1517 1518 if (isa<LoadInst>(I)) { 1519 // A load here means one of two things: a load of undef from a global, 1520 // a load from an unknown pointer. Either way, having it return undef 1521 // is okay. 1522 continue; 1523 } 1524 1525 markOverdefined(&I); 1526 MadeChange = true; 1527 } 1528 1529 // Check to see if we have a branch or switch on an undefined value. If so 1530 // we force the branch to go one way or the other to make the successor 1531 // values live. It doesn't really matter which way we force it. 1532 Instruction *TI = BB.getTerminator(); 1533 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1534 if (!BI->isConditional()) continue; 1535 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1536 continue; 1537 1538 // If the input to SCCP is actually branch on undef, fix the undef to 1539 // false. 1540 if (isa<UndefValue>(BI->getCondition())) { 1541 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1542 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1543 MadeChange = true; 1544 continue; 1545 } 1546 1547 // Otherwise, it is a branch on a symbolic value which is currently 1548 // considered to be undef. Make sure some edge is executable, so a 1549 // branch on "undef" always flows somewhere. 1550 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1551 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1552 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1553 MadeChange = true; 1554 1555 continue; 1556 } 1557 1558 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1559 // Indirect branch with no successor ?. Its ok to assume it branches 1560 // to no target. 1561 if (IBR->getNumSuccessors() < 1) 1562 continue; 1563 1564 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1565 continue; 1566 1567 // If the input to SCCP is actually branch on undef, fix the undef to 1568 // the first successor of the indirect branch. 1569 if (isa<UndefValue>(IBR->getAddress())) { 1570 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1571 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1572 MadeChange = true; 1573 continue; 1574 } 1575 1576 // Otherwise, it is a branch on a symbolic value which is currently 1577 // considered to be undef. Make sure some edge is executable, so a 1578 // branch on "undef" always flows somewhere. 1579 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1580 // we can assume the branch has undefined behavior instead. 1581 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1582 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1583 MadeChange = true; 1584 1585 continue; 1586 } 1587 1588 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1589 if (!SI->getNumCases() || 1590 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1591 continue; 1592 1593 // If the input to SCCP is actually switch on undef, fix the undef to 1594 // the first constant. 1595 if (isa<UndefValue>(SI->getCondition())) { 1596 SI->setCondition(SI->case_begin()->getCaseValue()); 1597 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1598 MadeChange = true; 1599 continue; 1600 } 1601 1602 // Otherwise, it is a branch on a symbolic value which is currently 1603 // considered to be undef. Make sure some edge is executable, so a 1604 // branch on "undef" always flows somewhere. 1605 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1606 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1607 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1608 MadeChange = true; 1609 1610 continue; 1611 } 1612 } 1613 1614 return MadeChange; 1615 } 1616 1617 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1618 Constant *Const = nullptr; 1619 if (V->getType()->isStructTy()) { 1620 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1621 if (any_of(IVs, 1622 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1623 return false; 1624 std::vector<Constant *> ConstVals; 1625 auto *ST = cast<StructType>(V->getType()); 1626 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1627 ValueLatticeElement V = IVs[i]; 1628 ConstVals.push_back(isConstant(V) 1629 ? Solver.getConstant(V) 1630 : UndefValue::get(ST->getElementType(i))); 1631 } 1632 Const = ConstantStruct::get(ST, ConstVals); 1633 } else { 1634 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1635 if (isOverdefined(IV)) 1636 return false; 1637 1638 Const = 1639 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1640 } 1641 assert(Const && "Constant is nullptr here!"); 1642 1643 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1644 // unless the call itself can be removed. 1645 // Calls with "clang.arc.attachedcall" implicitly use the return value and 1646 // those uses cannot be updated with a constant. 1647 CallBase *CB = dyn_cast<CallBase>(V); 1648 if (CB && ((CB->isMustTailCall() && !CB->isSafeToRemove()) || 1649 CB->getOperandBundle(LLVMContext::OB_clang_arc_attachedcall))) { 1650 Function *F = CB->getCalledFunction(); 1651 1652 // Don't zap returns of the callee 1653 if (F) 1654 Solver.addToMustPreserveReturnsInFunctions(F); 1655 1656 LLVM_DEBUG(dbgs() << " Can\'t treat the result of call " << *CB 1657 << " as a constant\n"); 1658 return false; 1659 } 1660 1661 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1662 1663 // Replaces all of the uses of a variable with uses of the constant. 1664 V->replaceAllUsesWith(Const); 1665 return true; 1666 } 1667 1668 static bool simplifyInstsInBlock(SCCPSolver &Solver, BasicBlock &BB, 1669 SmallPtrSetImpl<Value *> &InsertedValues, 1670 Statistic &InstRemovedStat, 1671 Statistic &InstReplacedStat) { 1672 bool MadeChanges = false; 1673 for (Instruction &Inst : make_early_inc_range(BB)) { 1674 if (Inst.getType()->isVoidTy()) 1675 continue; 1676 if (tryToReplaceWithConstant(Solver, &Inst)) { 1677 if (Inst.isSafeToRemove()) 1678 Inst.eraseFromParent(); 1679 // Hey, we just changed something! 1680 MadeChanges = true; 1681 ++InstRemovedStat; 1682 } else if (isa<SExtInst>(&Inst)) { 1683 Value *ExtOp = Inst.getOperand(0); 1684 if (isa<Constant>(ExtOp) || InsertedValues.count(ExtOp)) 1685 continue; 1686 const ValueLatticeElement &IV = Solver.getLatticeValueFor(ExtOp); 1687 if (!IV.isConstantRange(/*UndefAllowed=*/false)) 1688 continue; 1689 if (IV.getConstantRange().isAllNonNegative()) { 1690 auto *ZExt = new ZExtInst(ExtOp, Inst.getType(), "", &Inst); 1691 InsertedValues.insert(ZExt); 1692 Inst.replaceAllUsesWith(ZExt); 1693 Solver.removeLatticeValueFor(&Inst); 1694 Inst.eraseFromParent(); 1695 InstReplacedStat++; 1696 MadeChanges = true; 1697 } 1698 } 1699 } 1700 return MadeChanges; 1701 } 1702 1703 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1704 // and return true if the function was modified. 1705 static bool runSCCP(Function &F, const DataLayout &DL, 1706 const TargetLibraryInfo *TLI) { 1707 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1708 SCCPSolver Solver( 1709 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1710 F.getContext()); 1711 1712 // Mark the first block of the function as being executable. 1713 Solver.MarkBlockExecutable(&F.front()); 1714 1715 // Mark all arguments to the function as being overdefined. 1716 for (Argument &AI : F.args()) 1717 Solver.markOverdefined(&AI); 1718 1719 // Solve for constants. 1720 bool ResolvedUndefs = true; 1721 while (ResolvedUndefs) { 1722 Solver.Solve(); 1723 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1724 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1725 } 1726 1727 bool MadeChanges = false; 1728 1729 // If we decided that there are basic blocks that are dead in this function, 1730 // delete their contents now. Note that we cannot actually delete the blocks, 1731 // as we cannot modify the CFG of the function. 1732 1733 SmallPtrSet<Value *, 32> InsertedValues; 1734 for (BasicBlock &BB : F) { 1735 if (!Solver.isBlockExecutable(&BB)) { 1736 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1737 1738 ++NumDeadBlocks; 1739 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB).first; 1740 1741 MadeChanges = true; 1742 continue; 1743 } 1744 1745 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1746 NumInstRemoved, NumInstReplaced); 1747 } 1748 1749 return MadeChanges; 1750 } 1751 1752 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1753 const DataLayout &DL = F.getParent()->getDataLayout(); 1754 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1755 if (!runSCCP(F, DL, &TLI)) 1756 return PreservedAnalyses::all(); 1757 1758 auto PA = PreservedAnalyses(); 1759 PA.preserve<GlobalsAA>(); 1760 PA.preserveSet<CFGAnalyses>(); 1761 return PA; 1762 } 1763 1764 namespace { 1765 1766 //===--------------------------------------------------------------------===// 1767 // 1768 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1769 /// Sparse Conditional Constant Propagator. 1770 /// 1771 class SCCPLegacyPass : public FunctionPass { 1772 public: 1773 // Pass identification, replacement for typeid 1774 static char ID; 1775 1776 SCCPLegacyPass() : FunctionPass(ID) { 1777 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1778 } 1779 1780 void getAnalysisUsage(AnalysisUsage &AU) const override { 1781 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1782 AU.addPreserved<GlobalsAAWrapperPass>(); 1783 AU.setPreservesCFG(); 1784 } 1785 1786 // runOnFunction - Run the Sparse Conditional Constant Propagation 1787 // algorithm, and return true if the function was modified. 1788 bool runOnFunction(Function &F) override { 1789 if (skipFunction(F)) 1790 return false; 1791 const DataLayout &DL = F.getParent()->getDataLayout(); 1792 const TargetLibraryInfo *TLI = 1793 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1794 return runSCCP(F, DL, TLI); 1795 } 1796 }; 1797 1798 } // end anonymous namespace 1799 1800 char SCCPLegacyPass::ID = 0; 1801 1802 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1803 "Sparse Conditional Constant Propagation", false, false) 1804 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1805 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1806 "Sparse Conditional Constant Propagation", false, false) 1807 1808 // createSCCPPass - This is the public interface to this file. 1809 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1810 1811 static void findReturnsToZap(Function &F, 1812 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1813 SCCPSolver &Solver) { 1814 // We can only do this if we know that nothing else can call the function. 1815 if (!Solver.isArgumentTrackedFunction(&F)) 1816 return; 1817 1818 if (Solver.mustPreserveReturn(&F)) { 1819 LLVM_DEBUG( 1820 dbgs() 1821 << "Can't zap returns of the function : " << F.getName() 1822 << " due to present musttail or \"clang.arc.attachedcall\" call of " 1823 "it\n"); 1824 return; 1825 } 1826 1827 assert( 1828 all_of(F.users(), 1829 [&Solver](User *U) { 1830 if (isa<Instruction>(U) && 1831 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1832 return true; 1833 // Non-callsite uses are not impacted by zapping. Also, constant 1834 // uses (like blockaddresses) could stuck around, without being 1835 // used in the underlying IR, meaning we do not have lattice 1836 // values for them. 1837 if (!isa<CallBase>(U)) 1838 return true; 1839 if (U->getType()->isStructTy()) { 1840 return all_of(Solver.getStructLatticeValueFor(U), 1841 [](const ValueLatticeElement &LV) { 1842 return !isOverdefined(LV); 1843 }); 1844 } 1845 return !isOverdefined(Solver.getLatticeValueFor(U)); 1846 }) && 1847 "We can only zap functions where all live users have a concrete value"); 1848 1849 for (BasicBlock &BB : F) { 1850 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1851 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1852 << "musttail call : " << *CI << "\n"); 1853 (void)CI; 1854 return; 1855 } 1856 1857 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1858 if (!isa<UndefValue>(RI->getOperand(0))) 1859 ReturnsToZap.push_back(RI); 1860 } 1861 } 1862 1863 static bool removeNonFeasibleEdges(const SCCPSolver &Solver, BasicBlock *BB, 1864 DomTreeUpdater &DTU) { 1865 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors; 1866 bool HasNonFeasibleEdges = false; 1867 for (BasicBlock *Succ : successors(BB)) { 1868 if (Solver.isEdgeFeasible(BB, Succ)) 1869 FeasibleSuccessors.insert(Succ); 1870 else 1871 HasNonFeasibleEdges = true; 1872 } 1873 1874 // All edges feasible, nothing to do. 1875 if (!HasNonFeasibleEdges) 1876 return false; 1877 1878 // SCCP can only determine non-feasible edges for br, switch and indirectbr. 1879 Instruction *TI = BB->getTerminator(); 1880 assert((isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1881 isa<IndirectBrInst>(TI)) && 1882 "Terminator must be a br, switch or indirectbr"); 1883 1884 if (FeasibleSuccessors.size() == 1) { 1885 // Replace with an unconditional branch to the only feasible successor. 1886 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin(); 1887 SmallVector<DominatorTree::UpdateType, 8> Updates; 1888 bool HaveSeenOnlyFeasibleSuccessor = false; 1889 for (BasicBlock *Succ : successors(BB)) { 1890 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) { 1891 // Don't remove the edge to the only feasible successor the first time 1892 // we see it. We still do need to remove any multi-edges to it though. 1893 HaveSeenOnlyFeasibleSuccessor = true; 1894 continue; 1895 } 1896 1897 Succ->removePredecessor(BB); 1898 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1899 } 1900 1901 BranchInst::Create(OnlyFeasibleSuccessor, BB); 1902 TI->eraseFromParent(); 1903 DTU.applyUpdatesPermissive(Updates); 1904 } else if (FeasibleSuccessors.size() > 1) { 1905 SwitchInstProfUpdateWrapper SI(*cast<SwitchInst>(TI)); 1906 SmallVector<DominatorTree::UpdateType, 8> Updates; 1907 for (auto CI = SI->case_begin(); CI != SI->case_end();) { 1908 if (FeasibleSuccessors.contains(CI->getCaseSuccessor())) { 1909 ++CI; 1910 continue; 1911 } 1912 1913 BasicBlock *Succ = CI->getCaseSuccessor(); 1914 Succ->removePredecessor(BB); 1915 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1916 SI.removeCase(CI); 1917 // Don't increment CI, as we removed a case. 1918 } 1919 1920 DTU.applyUpdatesPermissive(Updates); 1921 } else { 1922 llvm_unreachable("Must have at least one feasible successor"); 1923 } 1924 return true; 1925 } 1926 1927 bool llvm::runIPSCCP( 1928 Module &M, const DataLayout &DL, 1929 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1930 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1931 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1932 1933 // Loop over all functions, marking arguments to those with their addresses 1934 // taken or that are external as overdefined. 1935 for (Function &F : M) { 1936 if (F.isDeclaration()) 1937 continue; 1938 1939 Solver.addAnalysis(F, getAnalysis(F)); 1940 1941 // Determine if we can track the function's return values. If so, add the 1942 // function to the solver's set of return-tracked functions. 1943 if (canTrackReturnsInterprocedurally(&F)) 1944 Solver.AddTrackedFunction(&F); 1945 1946 // Determine if we can track the function's arguments. If so, add the 1947 // function to the solver's set of argument-tracked functions. 1948 if (canTrackArgumentsInterprocedurally(&F)) { 1949 Solver.AddArgumentTrackedFunction(&F); 1950 continue; 1951 } 1952 1953 // Assume the function is called. 1954 Solver.MarkBlockExecutable(&F.front()); 1955 1956 // Assume nothing about the incoming arguments. 1957 for (Argument &AI : F.args()) 1958 Solver.markOverdefined(&AI); 1959 } 1960 1961 // Determine if we can track any of the module's global variables. If so, add 1962 // the global variables we can track to the solver's set of tracked global 1963 // variables. 1964 for (GlobalVariable &G : M.globals()) { 1965 G.removeDeadConstantUsers(); 1966 if (canTrackGlobalVariableInterprocedurally(&G)) 1967 Solver.TrackValueOfGlobalVariable(&G); 1968 } 1969 1970 // Solve for constants. 1971 bool ResolvedUndefs = true; 1972 Solver.Solve(); 1973 while (ResolvedUndefs) { 1974 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1975 ResolvedUndefs = false; 1976 for (Function &F : M) { 1977 if (Solver.ResolvedUndefsIn(F)) 1978 ResolvedUndefs = true; 1979 } 1980 if (ResolvedUndefs) 1981 Solver.Solve(); 1982 } 1983 1984 bool MadeChanges = false; 1985 1986 // Iterate over all of the instructions in the module, replacing them with 1987 // constants if we have found them to be of constant values. 1988 1989 for (Function &F : M) { 1990 if (F.isDeclaration()) 1991 continue; 1992 1993 SmallVector<BasicBlock *, 512> BlocksToErase; 1994 1995 if (Solver.isBlockExecutable(&F.front())) { 1996 bool ReplacedPointerArg = false; 1997 for (Argument &Arg : F.args()) { 1998 if (!Arg.use_empty() && tryToReplaceWithConstant(Solver, &Arg)) { 1999 ReplacedPointerArg |= Arg.getType()->isPointerTy(); 2000 ++IPNumArgsElimed; 2001 } 2002 } 2003 2004 // If we replaced an argument, the argmemonly and 2005 // inaccessiblemem_or_argmemonly attributes do not hold any longer. Remove 2006 // them from both the function and callsites. 2007 if (ReplacedPointerArg) { 2008 AttrBuilder AttributesToRemove; 2009 AttributesToRemove.addAttribute(Attribute::ArgMemOnly); 2010 AttributesToRemove.addAttribute(Attribute::InaccessibleMemOrArgMemOnly); 2011 F.removeAttributes(AttributeList::FunctionIndex, AttributesToRemove); 2012 2013 for (User *U : F.users()) { 2014 auto *CB = dyn_cast<CallBase>(U); 2015 if (!CB || CB->getCalledFunction() != &F) 2016 continue; 2017 2018 CB->removeAttributes(AttributeList::FunctionIndex, 2019 AttributesToRemove); 2020 } 2021 } 2022 } 2023 2024 SmallPtrSet<Value *, 32> InsertedValues; 2025 for (BasicBlock &BB : F) { 2026 if (!Solver.isBlockExecutable(&BB)) { 2027 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 2028 ++NumDeadBlocks; 2029 2030 MadeChanges = true; 2031 2032 if (&BB != &F.front()) 2033 BlocksToErase.push_back(&BB); 2034 continue; 2035 } 2036 2037 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 2038 IPNumInstRemoved, IPNumInstReplaced); 2039 } 2040 2041 DomTreeUpdater DTU = Solver.getDTU(F); 2042 // Change dead blocks to unreachable. We do it after replacing constants 2043 // in all executable blocks, because changeToUnreachable may remove PHI 2044 // nodes in executable blocks we found values for. The function's entry 2045 // block is not part of BlocksToErase, so we have to handle it separately. 2046 for (BasicBlock *BB : BlocksToErase) { 2047 NumInstRemoved += 2048 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 2049 /*PreserveLCSSA=*/false, &DTU); 2050 } 2051 if (!Solver.isBlockExecutable(&F.front())) 2052 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 2053 /*UseLLVMTrap=*/false, 2054 /*PreserveLCSSA=*/false, &DTU); 2055 2056 for (BasicBlock &BB : F) 2057 MadeChanges |= removeNonFeasibleEdges(Solver, &BB, DTU); 2058 2059 for (BasicBlock *DeadBB : BlocksToErase) 2060 DTU.deleteBB(DeadBB); 2061 2062 for (BasicBlock &BB : F) { 2063 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 2064 Instruction *Inst = &*BI++; 2065 if (Solver.getPredicateInfoFor(Inst)) { 2066 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2067 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2068 Value *Op = II->getOperand(0); 2069 Inst->replaceAllUsesWith(Op); 2070 Inst->eraseFromParent(); 2071 } 2072 } 2073 } 2074 } 2075 } 2076 } 2077 2078 // If we inferred constant or undef return values for a function, we replaced 2079 // all call uses with the inferred value. This means we don't need to bother 2080 // actually returning anything from the function. Replace all return 2081 // instructions with return undef. 2082 // 2083 // Do this in two stages: first identify the functions we should process, then 2084 // actually zap their returns. This is important because we can only do this 2085 // if the address of the function isn't taken. In cases where a return is the 2086 // last use of a function, the order of processing functions would affect 2087 // whether other functions are optimizable. 2088 SmallVector<ReturnInst*, 8> ReturnsToZap; 2089 2090 for (const auto &I : Solver.getTrackedRetVals()) { 2091 Function *F = I.first; 2092 const ValueLatticeElement &ReturnValue = I.second; 2093 2094 // If there is a known constant range for the return value, add !range 2095 // metadata to the function's call sites. 2096 if (ReturnValue.isConstantRange() && 2097 !ReturnValue.getConstantRange().isSingleElement()) { 2098 // Do not add range metadata if the return value may include undef. 2099 if (ReturnValue.isConstantRangeIncludingUndef()) 2100 continue; 2101 2102 auto &CR = ReturnValue.getConstantRange(); 2103 for (User *User : F->users()) { 2104 auto *CB = dyn_cast<CallBase>(User); 2105 if (!CB || CB->getCalledFunction() != F) 2106 continue; 2107 2108 // Limit to cases where the return value is guaranteed to be neither 2109 // poison nor undef. Poison will be outside any range and currently 2110 // values outside of the specified range cause immediate undefined 2111 // behavior. 2112 if (!isGuaranteedNotToBeUndefOrPoison(CB, nullptr, CB)) 2113 continue; 2114 2115 // Do not touch existing metadata for now. 2116 // TODO: We should be able to take the intersection of the existing 2117 // metadata and the inferred range. 2118 if (CB->getMetadata(LLVMContext::MD_range)) 2119 continue; 2120 2121 LLVMContext &Context = CB->getParent()->getContext(); 2122 Metadata *RangeMD[] = { 2123 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getLower())), 2124 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getUpper()))}; 2125 CB->setMetadata(LLVMContext::MD_range, MDNode::get(Context, RangeMD)); 2126 } 2127 continue; 2128 } 2129 if (F->getReturnType()->isVoidTy()) 2130 continue; 2131 if (isConstant(ReturnValue) || ReturnValue.isUnknownOrUndef()) 2132 findReturnsToZap(*F, ReturnsToZap, Solver); 2133 } 2134 2135 for (auto F : Solver.getMRVFunctionsTracked()) { 2136 assert(F->getReturnType()->isStructTy() && 2137 "The return type should be a struct"); 2138 StructType *STy = cast<StructType>(F->getReturnType()); 2139 if (Solver.isStructLatticeConstant(F, STy)) 2140 findReturnsToZap(*F, ReturnsToZap, Solver); 2141 } 2142 2143 // Zap all returns which we've identified as zap to change. 2144 SmallSetVector<Function *, 8> FuncZappedReturn; 2145 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2146 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2147 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2148 // Record all functions that are zapped. 2149 FuncZappedReturn.insert(F); 2150 } 2151 2152 // Remove the returned attribute for zapped functions and the 2153 // corresponding call sites. 2154 for (Function *F : FuncZappedReturn) { 2155 for (Argument &A : F->args()) 2156 F->removeParamAttr(A.getArgNo(), Attribute::Returned); 2157 for (Use &U : F->uses()) { 2158 // Skip over blockaddr users. 2159 if (isa<BlockAddress>(U.getUser())) 2160 continue; 2161 CallBase *CB = cast<CallBase>(U.getUser()); 2162 for (Use &Arg : CB->args()) 2163 CB->removeParamAttr(CB->getArgOperandNo(&Arg), Attribute::Returned); 2164 } 2165 } 2166 2167 // If we inferred constant or undef values for globals variables, we can 2168 // delete the global and any stores that remain to it. 2169 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2170 GlobalVariable *GV = I.first; 2171 if (isOverdefined(I.second)) 2172 continue; 2173 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2174 << "' is constant!\n"); 2175 while (!GV->use_empty()) { 2176 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2177 SI->eraseFromParent(); 2178 MadeChanges = true; 2179 } 2180 M.getGlobalList().erase(GV); 2181 ++IPNumGlobalConst; 2182 } 2183 2184 return MadeChanges; 2185 } 2186