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 /// MustTailFunctions - Each function here is a callee of non-removable 148 /// musttail call site. 149 SmallPtrSet<Function *, 16> MustTailCallees; 150 151 /// TrackingIncomingArguments - This is the set of functions for whose 152 /// arguments we make optimistic assumptions about and try to prove as 153 /// constants. 154 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 155 156 /// The reason for two worklists is that overdefined is the lowest state 157 /// on the lattice, and moving things to overdefined as fast as possible 158 /// makes SCCP converge much faster. 159 /// 160 /// By having a separate worklist, we accomplish this because everything 161 /// possibly overdefined will become overdefined at the soonest possible 162 /// point. 163 SmallVector<Value *, 64> OverdefinedInstWorkList; 164 SmallVector<Value *, 64> InstWorkList; 165 166 // The BasicBlock work list 167 SmallVector<BasicBlock *, 64> BBWorkList; 168 169 /// KnownFeasibleEdges - Entries in this set are edges which have already had 170 /// PHI nodes retriggered. 171 using Edge = std::pair<BasicBlock *, BasicBlock *>; 172 DenseSet<Edge> KnownFeasibleEdges; 173 174 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 175 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 176 177 LLVMContext &Ctx; 178 179 public: 180 void addAnalysis(Function &F, AnalysisResultsForFn A) { 181 AnalysisResults.insert({&F, std::move(A)}); 182 } 183 184 const PredicateBase *getPredicateInfoFor(Instruction *I) { 185 auto A = AnalysisResults.find(I->getParent()->getParent()); 186 if (A == AnalysisResults.end()) 187 return nullptr; 188 return A->second.PredInfo->getPredicateInfoFor(I); 189 } 190 191 DomTreeUpdater getDTU(Function &F) { 192 auto A = AnalysisResults.find(&F); 193 assert(A != AnalysisResults.end() && "Need analysis results for function."); 194 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 195 } 196 197 SCCPSolver(const DataLayout &DL, 198 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 199 LLVMContext &Ctx) 200 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 201 202 /// MarkBlockExecutable - This method can be used by clients to mark all of 203 /// the blocks that are known to be intrinsically live in the processed unit. 204 /// 205 /// This returns true if the block was not considered live before. 206 bool MarkBlockExecutable(BasicBlock *BB) { 207 if (!BBExecutable.insert(BB).second) 208 return false; 209 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 210 BBWorkList.push_back(BB); // Add the block to the work list! 211 return true; 212 } 213 214 /// TrackValueOfGlobalVariable - Clients can use this method to 215 /// inform the SCCPSolver that it should track loads and stores to the 216 /// specified global variable if it can. This is only legal to call if 217 /// performing Interprocedural SCCP. 218 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 219 // We only track the contents of scalar globals. 220 if (GV->getValueType()->isSingleValueType()) { 221 ValueLatticeElement &IV = TrackedGlobals[GV]; 222 if (!isa<UndefValue>(GV->getInitializer())) 223 IV.markConstant(GV->getInitializer()); 224 } 225 } 226 227 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 228 /// and out of the specified function (which cannot have its address taken), 229 /// this method must be called. 230 void AddTrackedFunction(Function *F) { 231 // Add an entry, F -> undef. 232 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 233 MRVFunctionsTracked.insert(F); 234 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 235 TrackedMultipleRetVals.insert( 236 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 237 } else if (!F->getReturnType()->isVoidTy()) 238 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 239 } 240 241 /// AddMustTailCallee - If the SCCP solver finds that this function is called 242 /// from non-removable musttail call site. 243 void AddMustTailCallee(Function *F) { 244 MustTailCallees.insert(F); 245 } 246 247 /// Returns true if the given function is called from non-removable musttail 248 /// call site. 249 bool isMustTailCallee(Function *F) { 250 return MustTailCallees.count(F); 251 } 252 253 void AddArgumentTrackedFunction(Function *F) { 254 TrackingIncomingArguments.insert(F); 255 } 256 257 /// Returns true if the given function is in the solver's set of 258 /// argument-tracked functions. 259 bool isArgumentTrackedFunction(Function *F) { 260 return TrackingIncomingArguments.count(F); 261 } 262 263 /// Solve - Solve for constants and executable blocks. 264 void Solve(); 265 266 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 267 /// that branches on undef values cannot reach any of their successors. 268 /// However, this is not a safe assumption. After we solve dataflow, this 269 /// method should be use to handle this. If this returns true, the solver 270 /// should be rerun. 271 bool ResolvedUndefsIn(Function &F); 272 273 bool isBlockExecutable(BasicBlock *BB) const { 274 return BBExecutable.count(BB); 275 } 276 277 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 278 // block to the 'To' basic block is currently feasible. 279 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 280 281 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 282 std::vector<ValueLatticeElement> StructValues; 283 auto *STy = dyn_cast<StructType>(V->getType()); 284 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 285 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 286 auto I = StructValueState.find(std::make_pair(V, i)); 287 assert(I != StructValueState.end() && "Value not in valuemap!"); 288 StructValues.push_back(I->second); 289 } 290 return StructValues; 291 } 292 293 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 294 295 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 296 assert(!V->getType()->isStructTy() && 297 "Should use getStructLatticeValueFor"); 298 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 299 ValueState.find(V); 300 assert(I != ValueState.end() && 301 "V not found in ValueState nor Paramstate map!"); 302 return I->second; 303 } 304 305 /// getTrackedRetVals - Get the inferred return value map. 306 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 307 return TrackedRetVals; 308 } 309 310 /// getTrackedGlobals - Get and return the set of inferred initializers for 311 /// global variables. 312 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 313 return TrackedGlobals; 314 } 315 316 /// getMRVFunctionsTracked - Get the set of functions which return multiple 317 /// values tracked by the pass. 318 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 319 return MRVFunctionsTracked; 320 } 321 322 /// getMustTailCallees - Get the set of functions which are called 323 /// from non-removable musttail call sites. 324 const SmallPtrSet<Function *, 16> getMustTailCallees() { 325 return MustTailCallees; 326 } 327 328 /// markOverdefined - Mark the specified value overdefined. This 329 /// works with both scalars and structs. 330 void markOverdefined(Value *V) { 331 if (auto *STy = dyn_cast<StructType>(V->getType())) 332 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 333 markOverdefined(getStructValueState(V, i), V); 334 else 335 markOverdefined(ValueState[V], V); 336 } 337 338 // isStructLatticeConstant - Return true if all the lattice values 339 // corresponding to elements of the structure are constants, 340 // false otherwise. 341 bool isStructLatticeConstant(Function *F, StructType *STy) { 342 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 343 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 344 assert(It != TrackedMultipleRetVals.end()); 345 ValueLatticeElement LV = It->second; 346 if (!isConstant(LV)) 347 return false; 348 } 349 return true; 350 } 351 352 /// Helper to return a Constant if \p LV is either a constant or a constant 353 /// range with a single element. 354 Constant *getConstant(const ValueLatticeElement &LV) const { 355 if (LV.isConstant()) 356 return LV.getConstant(); 357 358 if (LV.isConstantRange()) { 359 auto &CR = LV.getConstantRange(); 360 if (CR.getSingleElement()) 361 return ConstantInt::get(Ctx, *CR.getSingleElement()); 362 } 363 return nullptr; 364 } 365 366 private: 367 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 368 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 369 } 370 371 // pushToWorkList - Helper for markConstant/markOverdefined 372 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 373 if (IV.isOverdefined()) 374 return OverdefinedInstWorkList.push_back(V); 375 InstWorkList.push_back(V); 376 } 377 378 // Helper to push \p V to the worklist, after updating it to \p IV. Also 379 // prints a debug message with the updated value. 380 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 381 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 382 pushToWorkList(IV, V); 383 } 384 385 // markConstant - Make a value be marked as "constant". If the value 386 // is not already a constant, add it to the instruction work list so that 387 // the users of the instruction are updated later. 388 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 389 bool MayIncludeUndef = false) { 390 if (!IV.markConstant(C, MayIncludeUndef)) 391 return false; 392 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 393 pushToWorkList(IV, V); 394 return true; 395 } 396 397 bool markConstant(Value *V, Constant *C) { 398 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 399 return markConstant(ValueState[V], V, C); 400 } 401 402 // markOverdefined - Make a value be marked as "overdefined". If the 403 // value is not already overdefined, add it to the overdefined instruction 404 // work list so that the users of the instruction are updated later. 405 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 406 if (!IV.markOverdefined()) return false; 407 408 LLVM_DEBUG(dbgs() << "markOverdefined: "; 409 if (auto *F = dyn_cast<Function>(V)) dbgs() 410 << "Function '" << F->getName() << "'\n"; 411 else dbgs() << *V << '\n'); 412 // Only instructions go on the work list 413 pushToWorkList(IV, V); 414 return true; 415 } 416 417 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 418 /// changes. 419 bool mergeInValue(ValueLatticeElement &IV, Value *V, 420 ValueLatticeElement MergeWithV, 421 ValueLatticeElement::MergeOptions Opts = { 422 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 423 if (IV.mergeIn(MergeWithV, Opts)) { 424 pushToWorkList(IV, V); 425 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 426 << IV << "\n"); 427 return true; 428 } 429 return false; 430 } 431 432 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 433 ValueLatticeElement::MergeOptions Opts = { 434 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 435 assert(!V->getType()->isStructTy() && 436 "non-structs should use markConstant"); 437 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 438 } 439 440 /// getValueState - Return the ValueLatticeElement object that corresponds to 441 /// the value. This function handles the case when the value hasn't been seen 442 /// yet by properly seeding constants etc. 443 ValueLatticeElement &getValueState(Value *V) { 444 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 445 446 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 447 ValueLatticeElement &LV = I.first->second; 448 449 if (!I.second) 450 return LV; // Common case, already in the map. 451 452 if (auto *C = dyn_cast<Constant>(V)) 453 LV.markConstant(C); // Constants are constant 454 455 // All others are unknown by default. 456 return LV; 457 } 458 459 /// getStructValueState - Return the ValueLatticeElement object that 460 /// corresponds to the value/field pair. This function handles the case when 461 /// the value hasn't been seen yet by properly seeding constants etc. 462 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 463 assert(V->getType()->isStructTy() && "Should use getValueState"); 464 assert(i < cast<StructType>(V->getType())->getNumElements() && 465 "Invalid element #"); 466 467 auto I = StructValueState.insert( 468 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 469 ValueLatticeElement &LV = I.first->second; 470 471 if (!I.second) 472 return LV; // Common case, already in the map. 473 474 if (auto *C = dyn_cast<Constant>(V)) { 475 Constant *Elt = C->getAggregateElement(i); 476 477 if (!Elt) 478 LV.markOverdefined(); // Unknown sort of constant. 479 else if (isa<UndefValue>(Elt)) 480 ; // Undef values remain unknown. 481 else 482 LV.markConstant(Elt); // Constants are constant. 483 } 484 485 // All others are underdefined by default. 486 return LV; 487 } 488 489 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 490 /// work list if it is not already executable. 491 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 492 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 493 return false; // This edge is already known to be executable! 494 495 if (!MarkBlockExecutable(Dest)) { 496 // If the destination is already executable, we just made an *edge* 497 // feasible that wasn't before. Revisit the PHI nodes in the block 498 // because they have potentially new operands. 499 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 500 << " -> " << Dest->getName() << '\n'); 501 502 for (PHINode &PN : Dest->phis()) 503 visitPHINode(PN); 504 } 505 return true; 506 } 507 508 // getFeasibleSuccessors - Return a vector of booleans to indicate which 509 // successors are reachable from a given terminator instruction. 510 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 511 512 // OperandChangedState - This method is invoked on all of the users of an 513 // instruction that was just changed state somehow. Based on this 514 // information, we need to update the specified user of this instruction. 515 void OperandChangedState(Instruction *I) { 516 if (BBExecutable.count(I->getParent())) // Inst is executable? 517 visit(*I); 518 } 519 520 // Add U as additional user of V. 521 void addAdditionalUser(Value *V, User *U) { 522 auto Iter = AdditionalUsers.insert({V, {}}); 523 Iter.first->second.insert(U); 524 } 525 526 // Mark I's users as changed, including AdditionalUsers. 527 void markUsersAsChanged(Value *I) { 528 // Functions include their arguments in the use-list. Changed function 529 // values mean that the result of the function changed. We only need to 530 // update the call sites with the new function result and do not have to 531 // propagate the call arguments. 532 if (isa<Function>(I)) { 533 for (User *U : I->users()) { 534 if (auto *CB = dyn_cast<CallBase>(U)) 535 handleCallResult(*CB); 536 } 537 } else { 538 for (User *U : I->users()) 539 if (auto *UI = dyn_cast<Instruction>(U)) 540 OperandChangedState(UI); 541 } 542 543 auto Iter = AdditionalUsers.find(I); 544 if (Iter != AdditionalUsers.end()) { 545 // Copy additional users before notifying them of changes, because new 546 // users may be added, potentially invalidating the iterator. 547 SmallVector<Instruction *, 2> ToNotify; 548 for (User *U : Iter->second) 549 if (auto *UI = dyn_cast<Instruction>(U)) 550 ToNotify.push_back(UI); 551 for (Instruction *UI : ToNotify) 552 OperandChangedState(UI); 553 } 554 } 555 void handleCallOverdefined(CallBase &CB); 556 void handleCallResult(CallBase &CB); 557 void handleCallArguments(CallBase &CB); 558 559 private: 560 friend class InstVisitor<SCCPSolver>; 561 562 // visit implementations - Something changed in this instruction. Either an 563 // operand made a transition, or the instruction is newly executable. Change 564 // the value type of I to reflect these changes if appropriate. 565 void visitPHINode(PHINode &I); 566 567 // Terminators 568 569 void visitReturnInst(ReturnInst &I); 570 void visitTerminator(Instruction &TI); 571 572 void visitCastInst(CastInst &I); 573 void visitSelectInst(SelectInst &I); 574 void visitUnaryOperator(Instruction &I); 575 void visitBinaryOperator(Instruction &I); 576 void visitCmpInst(CmpInst &I); 577 void visitExtractValueInst(ExtractValueInst &EVI); 578 void visitInsertValueInst(InsertValueInst &IVI); 579 580 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 581 markOverdefined(&CPI); 582 visitTerminator(CPI); 583 } 584 585 // Instructions that cannot be folded away. 586 587 void visitStoreInst (StoreInst &I); 588 void visitLoadInst (LoadInst &I); 589 void visitGetElementPtrInst(GetElementPtrInst &I); 590 591 void visitCallInst (CallInst &I) { 592 visitCallBase(I); 593 } 594 595 void visitInvokeInst (InvokeInst &II) { 596 visitCallBase(II); 597 visitTerminator(II); 598 } 599 600 void visitCallBrInst (CallBrInst &CBI) { 601 visitCallBase(CBI); 602 visitTerminator(CBI); 603 } 604 605 void visitCallBase (CallBase &CB); 606 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 607 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 608 void visitFenceInst (FenceInst &I) { /*returns void*/ } 609 610 void visitInstruction(Instruction &I) { 611 // All the instructions we don't do any special handling for just 612 // go to overdefined. 613 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 614 markOverdefined(&I); 615 } 616 }; 617 618 } // end anonymous namespace 619 620 // getFeasibleSuccessors - Return a vector of booleans to indicate which 621 // successors are reachable from a given terminator instruction. 622 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 623 SmallVectorImpl<bool> &Succs) { 624 Succs.resize(TI.getNumSuccessors()); 625 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 626 if (BI->isUnconditional()) { 627 Succs[0] = true; 628 return; 629 } 630 631 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 632 ConstantInt *CI = getConstantInt(BCValue); 633 if (!CI) { 634 // Overdefined condition variables, and branches on unfoldable constant 635 // conditions, mean the branch could go either way. 636 if (!BCValue.isUnknownOrUndef()) 637 Succs[0] = Succs[1] = true; 638 return; 639 } 640 641 // Constant condition variables mean the branch can only go a single way. 642 Succs[CI->isZero()] = true; 643 return; 644 } 645 646 // Unwinding instructions successors are always executable. 647 if (TI.isExceptionalTerminator()) { 648 Succs.assign(TI.getNumSuccessors(), true); 649 return; 650 } 651 652 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 653 if (!SI->getNumCases()) { 654 Succs[0] = true; 655 return; 656 } 657 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 658 if (ConstantInt *CI = getConstantInt(SCValue)) { 659 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 660 return; 661 } 662 663 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 664 // is ready. 665 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 666 const ConstantRange &Range = SCValue.getConstantRange(); 667 for (const auto &Case : SI->cases()) { 668 const APInt &CaseValue = Case.getCaseValue()->getValue(); 669 if (Range.contains(CaseValue)) 670 Succs[Case.getSuccessorIndex()] = true; 671 } 672 673 // TODO: Determine whether default case is reachable. 674 Succs[SI->case_default()->getSuccessorIndex()] = true; 675 return; 676 } 677 678 // Overdefined or unknown condition? All destinations are executable! 679 if (!SCValue.isUnknownOrUndef()) 680 Succs.assign(TI.getNumSuccessors(), true); 681 return; 682 } 683 684 // In case of indirect branch and its address is a blockaddress, we mark 685 // the target as executable. 686 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 687 // Casts are folded by visitCastInst. 688 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 689 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 690 if (!Addr) { // Overdefined or unknown condition? 691 // All destinations are executable! 692 if (!IBRValue.isUnknownOrUndef()) 693 Succs.assign(TI.getNumSuccessors(), true); 694 return; 695 } 696 697 BasicBlock* T = Addr->getBasicBlock(); 698 assert(Addr->getFunction() == T->getParent() && 699 "Block address of a different function ?"); 700 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 701 // This is the target. 702 if (IBR->getDestination(i) == T) { 703 Succs[i] = true; 704 return; 705 } 706 } 707 708 // If we didn't find our destination in the IBR successor list, then we 709 // have undefined behavior. Its ok to assume no successor is executable. 710 return; 711 } 712 713 // In case of callbr, we pessimistically assume that all successors are 714 // feasible. 715 if (isa<CallBrInst>(&TI)) { 716 Succs.assign(TI.getNumSuccessors(), true); 717 return; 718 } 719 720 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 721 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 722 } 723 724 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 725 // block to the 'To' basic block is currently feasible. 726 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 727 // Check if we've called markEdgeExecutable on the edge yet. (We could 728 // be more aggressive and try to consider edges which haven't been marked 729 // yet, but there isn't any need.) 730 return KnownFeasibleEdges.count(Edge(From, To)); 731 } 732 733 // visit Implementations - Something changed in this instruction, either an 734 // operand made a transition, or the instruction is newly executable. Change 735 // the value type of I to reflect these changes if appropriate. This method 736 // makes sure to do the following actions: 737 // 738 // 1. If a phi node merges two constants in, and has conflicting value coming 739 // from different branches, or if the PHI node merges in an overdefined 740 // value, then the PHI node becomes overdefined. 741 // 2. If a phi node merges only constants in, and they all agree on value, the 742 // PHI node becomes a constant value equal to that. 743 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 744 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 745 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 746 // 6. If a conditional branch has a value that is constant, make the selected 747 // destination executable 748 // 7. If a conditional branch has a value that is overdefined, make all 749 // successors executable. 750 void SCCPSolver::visitPHINode(PHINode &PN) { 751 // If this PN returns a struct, just mark the result overdefined. 752 // TODO: We could do a lot better than this if code actually uses this. 753 if (PN.getType()->isStructTy()) 754 return (void)markOverdefined(&PN); 755 756 if (getValueState(&PN).isOverdefined()) 757 return; // Quick exit 758 759 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 760 // and slow us down a lot. Just mark them overdefined. 761 if (PN.getNumIncomingValues() > 64) 762 return (void)markOverdefined(&PN); 763 764 unsigned NumActiveIncoming = 0; 765 766 // Look at all of the executable operands of the PHI node. If any of them 767 // are overdefined, the PHI becomes overdefined as well. If they are all 768 // constant, and they agree with each other, the PHI becomes the identical 769 // constant. If they are constant and don't agree, the PHI is a constant 770 // range. If there are no executable operands, the PHI remains unknown. 771 ValueLatticeElement PhiState = getValueState(&PN); 772 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 773 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 774 continue; 775 776 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 777 PhiState.mergeIn(IV); 778 NumActiveIncoming++; 779 if (PhiState.isOverdefined()) 780 break; 781 } 782 783 // We allow up to 1 range extension per active incoming value and one 784 // additional extension. Note that we manually adjust the number of range 785 // extensions to match the number of active incoming values. This helps to 786 // limit multiple extensions caused by the same incoming value, if other 787 // incoming values are equal. 788 mergeInValue(&PN, PhiState, 789 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 790 NumActiveIncoming + 1)); 791 ValueLatticeElement &PhiStateRef = getValueState(&PN); 792 PhiStateRef.setNumRangeExtensions( 793 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 794 } 795 796 void SCCPSolver::visitReturnInst(ReturnInst &I) { 797 if (I.getNumOperands() == 0) return; // ret void 798 799 Function *F = I.getParent()->getParent(); 800 Value *ResultOp = I.getOperand(0); 801 802 // If we are tracking the return value of this function, merge it in. 803 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 804 auto TFRVI = TrackedRetVals.find(F); 805 if (TFRVI != TrackedRetVals.end()) { 806 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 807 return; 808 } 809 } 810 811 // Handle functions that return multiple values. 812 if (!TrackedMultipleRetVals.empty()) { 813 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 814 if (MRVFunctionsTracked.count(F)) 815 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 816 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 817 getStructValueState(ResultOp, i)); 818 } 819 } 820 821 void SCCPSolver::visitTerminator(Instruction &TI) { 822 SmallVector<bool, 16> SuccFeasible; 823 getFeasibleSuccessors(TI, SuccFeasible); 824 825 BasicBlock *BB = TI.getParent(); 826 827 // Mark all feasible successors executable. 828 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 829 if (SuccFeasible[i]) 830 markEdgeExecutable(BB, TI.getSuccessor(i)); 831 } 832 833 void SCCPSolver::visitCastInst(CastInst &I) { 834 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 835 // discover a concrete value later. 836 if (ValueState[&I].isOverdefined()) 837 return; 838 839 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 840 if (Constant *OpC = getConstant(OpSt)) { 841 // Fold the constant as we build. 842 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 843 if (isa<UndefValue>(C)) 844 return; 845 // Propagate constant value 846 markConstant(&I, C); 847 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 848 auto &LV = getValueState(&I); 849 ConstantRange OpRange = OpSt.getConstantRange(); 850 Type *DestTy = I.getDestTy(); 851 // Vectors where all elements have the same known constant range are treated 852 // as a single constant range in the lattice. When bitcasting such vectors, 853 // there is a mis-match between the width of the lattice value (single 854 // constant range) and the original operands (vector). Go to overdefined in 855 // that case. 856 if (I.getOpcode() == Instruction::BitCast && 857 I.getOperand(0)->getType()->isVectorTy() && 858 OpRange.getBitWidth() < DL.getTypeSizeInBits(DestTy)) 859 return (void)markOverdefined(&I); 860 861 ConstantRange Res = 862 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 863 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 864 } else if (!OpSt.isUnknownOrUndef()) 865 markOverdefined(&I); 866 } 867 868 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 869 // If this returns a struct, mark all elements over defined, we don't track 870 // structs in structs. 871 if (EVI.getType()->isStructTy()) 872 return (void)markOverdefined(&EVI); 873 874 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 875 // discover a concrete value later. 876 if (ValueState[&EVI].isOverdefined()) 877 return (void)markOverdefined(&EVI); 878 879 // If this is extracting from more than one level of struct, we don't know. 880 if (EVI.getNumIndices() != 1) 881 return (void)markOverdefined(&EVI); 882 883 Value *AggVal = EVI.getAggregateOperand(); 884 if (AggVal->getType()->isStructTy()) { 885 unsigned i = *EVI.idx_begin(); 886 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 887 mergeInValue(getValueState(&EVI), &EVI, EltVal); 888 } else { 889 // Otherwise, must be extracting from an array. 890 return (void)markOverdefined(&EVI); 891 } 892 } 893 894 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 895 auto *STy = dyn_cast<StructType>(IVI.getType()); 896 if (!STy) 897 return (void)markOverdefined(&IVI); 898 899 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 900 // discover a concrete value later. 901 if (isOverdefined(ValueState[&IVI])) 902 return (void)markOverdefined(&IVI); 903 904 // If this has more than one index, we can't handle it, drive all results to 905 // undef. 906 if (IVI.getNumIndices() != 1) 907 return (void)markOverdefined(&IVI); 908 909 Value *Aggr = IVI.getAggregateOperand(); 910 unsigned Idx = *IVI.idx_begin(); 911 912 // Compute the result based on what we're inserting. 913 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 914 // This passes through all values that aren't the inserted element. 915 if (i != Idx) { 916 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 917 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 918 continue; 919 } 920 921 Value *Val = IVI.getInsertedValueOperand(); 922 if (Val->getType()->isStructTy()) 923 // We don't track structs in structs. 924 markOverdefined(getStructValueState(&IVI, i), &IVI); 925 else { 926 ValueLatticeElement InVal = getValueState(Val); 927 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 928 } 929 } 930 } 931 932 void SCCPSolver::visitSelectInst(SelectInst &I) { 933 // If this select returns a struct, just mark the result overdefined. 934 // TODO: We could do a lot better than this if code actually uses this. 935 if (I.getType()->isStructTy()) 936 return (void)markOverdefined(&I); 937 938 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 939 // discover a concrete value later. 940 if (ValueState[&I].isOverdefined()) 941 return (void)markOverdefined(&I); 942 943 ValueLatticeElement CondValue = getValueState(I.getCondition()); 944 if (CondValue.isUnknownOrUndef()) 945 return; 946 947 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 948 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 949 mergeInValue(&I, getValueState(OpVal)); 950 return; 951 } 952 953 // Otherwise, the condition is overdefined or a constant we can't evaluate. 954 // See if we can produce something better than overdefined based on the T/F 955 // value. 956 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 957 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 958 959 bool Changed = ValueState[&I].mergeIn(TVal); 960 Changed |= ValueState[&I].mergeIn(FVal); 961 if (Changed) 962 pushToWorkListMsg(ValueState[&I], &I); 963 } 964 965 // Handle Unary Operators. 966 void SCCPSolver::visitUnaryOperator(Instruction &I) { 967 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 968 969 ValueLatticeElement &IV = ValueState[&I]; 970 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 971 // discover a concrete value later. 972 if (isOverdefined(IV)) 973 return (void)markOverdefined(&I); 974 975 if (isConstant(V0State)) { 976 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 977 978 // op Y -> undef. 979 if (isa<UndefValue>(C)) 980 return; 981 return (void)markConstant(IV, &I, C); 982 } 983 984 // If something is undef, wait for it to resolve. 985 if (!isOverdefined(V0State)) 986 return; 987 988 markOverdefined(&I); 989 } 990 991 // Handle Binary Operators. 992 void SCCPSolver::visitBinaryOperator(Instruction &I) { 993 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 994 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 995 996 ValueLatticeElement &IV = ValueState[&I]; 997 if (IV.isOverdefined()) 998 return; 999 1000 // If something is undef, wait for it to resolve. 1001 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 1002 return; 1003 1004 if (V1State.isOverdefined() && V2State.isOverdefined()) 1005 return (void)markOverdefined(&I); 1006 1007 // If either of the operands is a constant, try to fold it to a constant. 1008 // TODO: Use information from notconstant better. 1009 if ((V1State.isConstant() || V2State.isConstant())) { 1010 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 1011 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 1012 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 1013 auto *C = dyn_cast_or_null<Constant>(R); 1014 if (C) { 1015 // X op Y -> undef. 1016 if (isa<UndefValue>(C)) 1017 return; 1018 // Conservatively assume that the result may be based on operands that may 1019 // be undef. Note that we use mergeInValue to combine the constant with 1020 // the existing lattice value for I, as different constants might be found 1021 // after one of the operands go to overdefined, e.g. due to one operand 1022 // being a special floating value. 1023 ValueLatticeElement NewV; 1024 NewV.markConstant(C, /*MayIncludeUndef=*/true); 1025 return (void)mergeInValue(&I, NewV); 1026 } 1027 } 1028 1029 // Only use ranges for binary operators on integers. 1030 if (!I.getType()->isIntegerTy()) 1031 return markOverdefined(&I); 1032 1033 // Try to simplify to a constant range. 1034 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1035 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1036 if (V1State.isConstantRange()) 1037 A = V1State.getConstantRange(); 1038 if (V2State.isConstantRange()) 1039 B = V2State.getConstantRange(); 1040 1041 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1042 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1043 1044 // TODO: Currently we do not exploit special values that produce something 1045 // better than overdefined with an overdefined operand for vector or floating 1046 // point types, like and <4 x i32> overdefined, zeroinitializer. 1047 } 1048 1049 // Handle ICmpInst instruction. 1050 void SCCPSolver::visitCmpInst(CmpInst &I) { 1051 // Do not cache this lookup, getValueState calls later in the function might 1052 // invalidate the reference. 1053 if (isOverdefined(ValueState[&I])) 1054 return (void)markOverdefined(&I); 1055 1056 Value *Op1 = I.getOperand(0); 1057 Value *Op2 = I.getOperand(1); 1058 1059 // For parameters, use ParamState which includes constant range info if 1060 // available. 1061 auto V1State = getValueState(Op1); 1062 auto V2State = getValueState(Op2); 1063 1064 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1065 if (C) { 1066 if (isa<UndefValue>(C)) 1067 return; 1068 ValueLatticeElement CV; 1069 CV.markConstant(C); 1070 mergeInValue(&I, CV); 1071 return; 1072 } 1073 1074 // If operands are still unknown, wait for it to resolve. 1075 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1076 !isConstant(ValueState[&I])) 1077 return; 1078 1079 markOverdefined(&I); 1080 } 1081 1082 // Handle getelementptr instructions. If all operands are constants then we 1083 // can turn this into a getelementptr ConstantExpr. 1084 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1085 if (isOverdefined(ValueState[&I])) 1086 return (void)markOverdefined(&I); 1087 1088 SmallVector<Constant*, 8> Operands; 1089 Operands.reserve(I.getNumOperands()); 1090 1091 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1092 ValueLatticeElement State = getValueState(I.getOperand(i)); 1093 if (State.isUnknownOrUndef()) 1094 return; // Operands are not resolved yet. 1095 1096 if (isOverdefined(State)) 1097 return (void)markOverdefined(&I); 1098 1099 if (Constant *C = getConstant(State)) { 1100 Operands.push_back(C); 1101 continue; 1102 } 1103 1104 return (void)markOverdefined(&I); 1105 } 1106 1107 Constant *Ptr = Operands[0]; 1108 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1109 Constant *C = 1110 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1111 if (isa<UndefValue>(C)) 1112 return; 1113 markConstant(&I, C); 1114 } 1115 1116 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1117 // If this store is of a struct, ignore it. 1118 if (SI.getOperand(0)->getType()->isStructTy()) 1119 return; 1120 1121 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1122 return; 1123 1124 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1125 auto I = TrackedGlobals.find(GV); 1126 if (I == TrackedGlobals.end()) 1127 return; 1128 1129 // Get the value we are storing into the global, then merge it. 1130 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1131 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1132 if (I->second.isOverdefined()) 1133 TrackedGlobals.erase(I); // No need to keep tracking this! 1134 } 1135 1136 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1137 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1138 if (I->getType()->isIntegerTy()) 1139 return ValueLatticeElement::getRange( 1140 getConstantRangeFromMetadata(*Ranges)); 1141 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1142 return ValueLatticeElement::getNot( 1143 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1144 return ValueLatticeElement::getOverdefined(); 1145 } 1146 1147 // Handle load instructions. If the operand is a constant pointer to a constant 1148 // global, we can replace the load with the loaded constant value! 1149 void SCCPSolver::visitLoadInst(LoadInst &I) { 1150 // If this load is of a struct or the load is volatile, just mark the result 1151 // as overdefined. 1152 if (I.getType()->isStructTy() || I.isVolatile()) 1153 return (void)markOverdefined(&I); 1154 1155 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1156 // discover a concrete value later. 1157 if (ValueState[&I].isOverdefined()) 1158 return (void)markOverdefined(&I); 1159 1160 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1161 if (PtrVal.isUnknownOrUndef()) 1162 return; // The pointer is not resolved yet! 1163 1164 ValueLatticeElement &IV = ValueState[&I]; 1165 1166 if (isConstant(PtrVal)) { 1167 Constant *Ptr = getConstant(PtrVal); 1168 1169 // load null is undefined. 1170 if (isa<ConstantPointerNull>(Ptr)) { 1171 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1172 return (void)markOverdefined(IV, &I); 1173 else 1174 return; 1175 } 1176 1177 // Transform load (constant global) into the value loaded. 1178 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1179 if (!TrackedGlobals.empty()) { 1180 // If we are tracking this global, merge in the known value for it. 1181 auto It = TrackedGlobals.find(GV); 1182 if (It != TrackedGlobals.end()) { 1183 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1184 return; 1185 } 1186 } 1187 } 1188 1189 // Transform load from a constant into a constant if possible. 1190 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1191 if (isa<UndefValue>(C)) 1192 return; 1193 return (void)markConstant(IV, &I, C); 1194 } 1195 } 1196 1197 // Fall back to metadata. 1198 mergeInValue(&I, getValueFromMetadata(&I)); 1199 } 1200 1201 void SCCPSolver::visitCallBase(CallBase &CB) { 1202 handleCallResult(CB); 1203 handleCallArguments(CB); 1204 } 1205 1206 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1207 Function *F = CB.getCalledFunction(); 1208 1209 // Void return and not tracking callee, just bail. 1210 if (CB.getType()->isVoidTy()) 1211 return; 1212 1213 // Always mark struct return as overdefined. 1214 if (CB.getType()->isStructTy()) 1215 return (void)markOverdefined(&CB); 1216 1217 // Otherwise, if we have a single return value case, and if the function is 1218 // a declaration, maybe we can constant fold it. 1219 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1220 SmallVector<Constant *, 8> Operands; 1221 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1222 if (AI->get()->getType()->isStructTy()) 1223 return markOverdefined(&CB); // Can't handle struct args. 1224 ValueLatticeElement State = getValueState(*AI); 1225 1226 if (State.isUnknownOrUndef()) 1227 return; // Operands are not resolved yet. 1228 if (isOverdefined(State)) 1229 return (void)markOverdefined(&CB); 1230 assert(isConstant(State) && "Unknown state!"); 1231 Operands.push_back(getConstant(State)); 1232 } 1233 1234 if (isOverdefined(getValueState(&CB))) 1235 return (void)markOverdefined(&CB); 1236 1237 // If we can constant fold this, mark the result of the call as a 1238 // constant. 1239 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1240 // call -> undef. 1241 if (isa<UndefValue>(C)) 1242 return; 1243 return (void)markConstant(&CB, C); 1244 } 1245 } 1246 1247 // Fall back to metadata. 1248 mergeInValue(&CB, getValueFromMetadata(&CB)); 1249 } 1250 1251 void SCCPSolver::handleCallArguments(CallBase &CB) { 1252 Function *F = CB.getCalledFunction(); 1253 // If this is a local function that doesn't have its address taken, mark its 1254 // entry block executable and merge in the actual arguments to the call into 1255 // the formal arguments of the function. 1256 if (!TrackingIncomingArguments.empty() && 1257 TrackingIncomingArguments.count(F)) { 1258 MarkBlockExecutable(&F->front()); 1259 1260 // Propagate information from this call site into the callee. 1261 auto CAI = CB.arg_begin(); 1262 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1263 ++AI, ++CAI) { 1264 // If this argument is byval, and if the function is not readonly, there 1265 // will be an implicit copy formed of the input aggregate. 1266 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1267 markOverdefined(&*AI); 1268 continue; 1269 } 1270 1271 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1272 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1273 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1274 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1275 getMaxWidenStepsOpts()); 1276 } 1277 } else 1278 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1279 } 1280 } 1281 } 1282 1283 void SCCPSolver::handleCallResult(CallBase &CB) { 1284 Function *F = CB.getCalledFunction(); 1285 1286 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1287 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1288 if (ValueState[&CB].isOverdefined()) 1289 return; 1290 1291 Value *CopyOf = CB.getOperand(0); 1292 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1293 auto *PI = getPredicateInfoFor(&CB); 1294 assert(PI && "Missing predicate info for ssa.copy"); 1295 1296 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1297 if (!Constraint) { 1298 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1299 return; 1300 } 1301 1302 CmpInst::Predicate Pred = Constraint->Predicate; 1303 Value *OtherOp = Constraint->OtherOp; 1304 1305 // Wait until OtherOp is resolved. 1306 if (getValueState(OtherOp).isUnknown()) { 1307 addAdditionalUser(OtherOp, &CB); 1308 return; 1309 } 1310 1311 // TODO: Actually filp MayIncludeUndef for the created range to false, 1312 // once most places in the optimizer respect the branches on 1313 // undef/poison are UB rule. The reason why the new range cannot be 1314 // undef is as follows below: 1315 // The new range is based on a branch condition. That guarantees that 1316 // neither of the compare operands can be undef in the branch targets, 1317 // unless we have conditions that are always true/false (e.g. icmp ule 1318 // i32, %a, i32_max). For the latter overdefined/empty range will be 1319 // inferred, but the branch will get folded accordingly anyways. 1320 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1321 1322 ValueLatticeElement CondVal = getValueState(OtherOp); 1323 ValueLatticeElement &IV = ValueState[&CB]; 1324 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1325 auto ImposedCR = 1326 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1327 1328 // Get the range imposed by the condition. 1329 if (CondVal.isConstantRange()) 1330 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1331 Pred, CondVal.getConstantRange()); 1332 1333 // Combine range info for the original value with the new range from the 1334 // condition. 1335 auto CopyOfCR = CopyOfVal.isConstantRange() 1336 ? CopyOfVal.getConstantRange() 1337 : ConstantRange::getFull( 1338 DL.getTypeSizeInBits(CopyOf->getType())); 1339 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1340 // If the existing information is != x, do not use the information from 1341 // a chained predicate, as the != x information is more likely to be 1342 // helpful in practice. 1343 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1344 NewCR = CopyOfCR; 1345 1346 addAdditionalUser(OtherOp, &CB); 1347 mergeInValue( 1348 IV, &CB, 1349 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1350 return; 1351 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1352 // For non-integer values or integer constant expressions, only 1353 // propagate equal constants. 1354 addAdditionalUser(OtherOp, &CB); 1355 mergeInValue(IV, &CB, CondVal); 1356 return; 1357 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1358 !MayIncludeUndef) { 1359 // Propagate inequalities. 1360 addAdditionalUser(OtherOp, &CB); 1361 mergeInValue(IV, &CB, 1362 ValueLatticeElement::getNot(CondVal.getConstant())); 1363 return; 1364 } 1365 1366 return (void)mergeInValue(IV, &CB, CopyOfVal); 1367 } 1368 1369 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { 1370 // Compute result range for intrinsics supported by ConstantRange. 1371 // Do this even if we don't know a range for all operands, as we may 1372 // still know something about the result range, e.g. of abs(x). 1373 SmallVector<ConstantRange, 2> OpRanges; 1374 for (Value *Op : II->args()) { 1375 const ValueLatticeElement &State = getValueState(Op); 1376 if (State.isConstantRange()) 1377 OpRanges.push_back(State.getConstantRange()); 1378 else 1379 OpRanges.push_back( 1380 ConstantRange::getFull(Op->getType()->getScalarSizeInBits())); 1381 } 1382 1383 ConstantRange Result = 1384 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges); 1385 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result)); 1386 } 1387 } 1388 1389 // The common case is that we aren't tracking the callee, either because we 1390 // are not doing interprocedural analysis or the callee is indirect, or is 1391 // external. Handle these cases first. 1392 if (!F || F->isDeclaration()) 1393 return handleCallOverdefined(CB); 1394 1395 // If this is a single/zero retval case, see if we're tracking the function. 1396 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1397 if (!MRVFunctionsTracked.count(F)) 1398 return handleCallOverdefined(CB); // Not tracking this callee. 1399 1400 // If we are tracking this callee, propagate the result of the function 1401 // into this call site. 1402 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1403 mergeInValue(getStructValueState(&CB, i), &CB, 1404 TrackedMultipleRetVals[std::make_pair(F, i)], 1405 getMaxWidenStepsOpts()); 1406 } else { 1407 auto TFRVI = TrackedRetVals.find(F); 1408 if (TFRVI == TrackedRetVals.end()) 1409 return handleCallOverdefined(CB); // Not tracking this callee. 1410 1411 // If so, propagate the return value of the callee into this call result. 1412 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1413 } 1414 } 1415 1416 void SCCPSolver::Solve() { 1417 // Process the work lists until they are empty! 1418 while (!BBWorkList.empty() || !InstWorkList.empty() || 1419 !OverdefinedInstWorkList.empty()) { 1420 // Process the overdefined instruction's work list first, which drives other 1421 // things to overdefined more quickly. 1422 while (!OverdefinedInstWorkList.empty()) { 1423 Value *I = OverdefinedInstWorkList.pop_back_val(); 1424 1425 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1426 1427 // "I" got into the work list because it either made the transition from 1428 // bottom to constant, or to overdefined. 1429 // 1430 // Anything on this worklist that is overdefined need not be visited 1431 // since all of its users will have already been marked as overdefined 1432 // Update all of the users of this instruction's value. 1433 // 1434 markUsersAsChanged(I); 1435 } 1436 1437 // Process the instruction work list. 1438 while (!InstWorkList.empty()) { 1439 Value *I = InstWorkList.pop_back_val(); 1440 1441 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1442 1443 // "I" got into the work list because it made the transition from undef to 1444 // constant. 1445 // 1446 // Anything on this worklist that is overdefined need not be visited 1447 // since all of its users will have already been marked as overdefined. 1448 // Update all of the users of this instruction's value. 1449 // 1450 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1451 markUsersAsChanged(I); 1452 } 1453 1454 // Process the basic block work list. 1455 while (!BBWorkList.empty()) { 1456 BasicBlock *BB = BBWorkList.pop_back_val(); 1457 1458 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1459 1460 // Notify all instructions in this basic block that they are newly 1461 // executable. 1462 visit(BB); 1463 } 1464 } 1465 } 1466 1467 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1468 /// that branches on undef values cannot reach any of their successors. 1469 /// However, this is not a safe assumption. After we solve dataflow, this 1470 /// method should be use to handle this. If this returns true, the solver 1471 /// should be rerun. 1472 /// 1473 /// This method handles this by finding an unresolved branch and marking it one 1474 /// of the edges from the block as being feasible, even though the condition 1475 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1476 /// CFG and only slightly pessimizes the analysis results (by marking one, 1477 /// potentially infeasible, edge feasible). This cannot usefully modify the 1478 /// constraints on the condition of the branch, as that would impact other users 1479 /// of the value. 1480 /// 1481 /// This scan also checks for values that use undefs. It conservatively marks 1482 /// them as overdefined. 1483 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1484 bool MadeChange = false; 1485 for (BasicBlock &BB : F) { 1486 if (!BBExecutable.count(&BB)) 1487 continue; 1488 1489 for (Instruction &I : BB) { 1490 // Look for instructions which produce undef values. 1491 if (I.getType()->isVoidTy()) continue; 1492 1493 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1494 // Only a few things that can be structs matter for undef. 1495 1496 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1497 if (auto *CB = dyn_cast<CallBase>(&I)) 1498 if (Function *F = CB->getCalledFunction()) 1499 if (MRVFunctionsTracked.count(F)) 1500 continue; 1501 1502 // extractvalue and insertvalue don't need to be marked; they are 1503 // tracked as precisely as their operands. 1504 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1505 continue; 1506 // Send the results of everything else to overdefined. We could be 1507 // more precise than this but it isn't worth bothering. 1508 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1509 ValueLatticeElement &LV = getStructValueState(&I, i); 1510 if (LV.isUnknownOrUndef()) { 1511 markOverdefined(LV, &I); 1512 MadeChange = true; 1513 } 1514 } 1515 continue; 1516 } 1517 1518 ValueLatticeElement &LV = getValueState(&I); 1519 if (!LV.isUnknownOrUndef()) 1520 continue; 1521 1522 // There are two reasons a call can have an undef result 1523 // 1. It could be tracked. 1524 // 2. It could be constant-foldable. 1525 // Because of the way we solve return values, tracked calls must 1526 // never be marked overdefined in ResolvedUndefsIn. 1527 if (auto *CB = dyn_cast<CallBase>(&I)) 1528 if (Function *F = CB->getCalledFunction()) 1529 if (TrackedRetVals.count(F)) 1530 continue; 1531 1532 if (isa<LoadInst>(I)) { 1533 // A load here means one of two things: a load of undef from a global, 1534 // a load from an unknown pointer. Either way, having it return undef 1535 // is okay. 1536 continue; 1537 } 1538 1539 markOverdefined(&I); 1540 MadeChange = true; 1541 } 1542 1543 // Check to see if we have a branch or switch on an undefined value. If so 1544 // we force the branch to go one way or the other to make the successor 1545 // values live. It doesn't really matter which way we force it. 1546 Instruction *TI = BB.getTerminator(); 1547 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1548 if (!BI->isConditional()) continue; 1549 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1550 continue; 1551 1552 // If the input to SCCP is actually branch on undef, fix the undef to 1553 // false. 1554 if (isa<UndefValue>(BI->getCondition())) { 1555 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1556 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1557 MadeChange = true; 1558 continue; 1559 } 1560 1561 // Otherwise, it is a branch on a symbolic value which is currently 1562 // considered to be undef. Make sure some edge is executable, so a 1563 // branch on "undef" always flows somewhere. 1564 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1565 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1566 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1567 MadeChange = true; 1568 1569 continue; 1570 } 1571 1572 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1573 // Indirect branch with no successor ?. Its ok to assume it branches 1574 // to no target. 1575 if (IBR->getNumSuccessors() < 1) 1576 continue; 1577 1578 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1579 continue; 1580 1581 // If the input to SCCP is actually branch on undef, fix the undef to 1582 // the first successor of the indirect branch. 1583 if (isa<UndefValue>(IBR->getAddress())) { 1584 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1585 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1586 MadeChange = true; 1587 continue; 1588 } 1589 1590 // Otherwise, it is a branch on a symbolic value which is currently 1591 // considered to be undef. Make sure some edge is executable, so a 1592 // branch on "undef" always flows somewhere. 1593 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1594 // we can assume the branch has undefined behavior instead. 1595 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1596 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1597 MadeChange = true; 1598 1599 continue; 1600 } 1601 1602 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1603 if (!SI->getNumCases() || 1604 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1605 continue; 1606 1607 // If the input to SCCP is actually switch on undef, fix the undef to 1608 // the first constant. 1609 if (isa<UndefValue>(SI->getCondition())) { 1610 SI->setCondition(SI->case_begin()->getCaseValue()); 1611 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1612 MadeChange = true; 1613 continue; 1614 } 1615 1616 // Otherwise, it is a branch on a symbolic value which is currently 1617 // considered to be undef. Make sure some edge is executable, so a 1618 // branch on "undef" always flows somewhere. 1619 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1620 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1621 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1622 MadeChange = true; 1623 1624 continue; 1625 } 1626 } 1627 1628 return MadeChange; 1629 } 1630 1631 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1632 Constant *Const = nullptr; 1633 if (V->getType()->isStructTy()) { 1634 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1635 if (any_of(IVs, 1636 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1637 return false; 1638 std::vector<Constant *> ConstVals; 1639 auto *ST = cast<StructType>(V->getType()); 1640 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1641 ValueLatticeElement V = IVs[i]; 1642 ConstVals.push_back(isConstant(V) 1643 ? Solver.getConstant(V) 1644 : UndefValue::get(ST->getElementType(i))); 1645 } 1646 Const = ConstantStruct::get(ST, ConstVals); 1647 } else { 1648 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1649 if (isOverdefined(IV)) 1650 return false; 1651 1652 Const = 1653 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1654 } 1655 assert(Const && "Constant is nullptr here!"); 1656 1657 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1658 // unless the call itself can be removed 1659 CallInst *CI = dyn_cast<CallInst>(V); 1660 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1661 Function *F = CI->getCalledFunction(); 1662 1663 // Don't zap returns of the callee 1664 if (F) 1665 Solver.AddMustTailCallee(F); 1666 1667 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1668 << " as a constant\n"); 1669 return false; 1670 } 1671 1672 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1673 1674 // Replaces all of the uses of a variable with uses of the constant. 1675 V->replaceAllUsesWith(Const); 1676 return true; 1677 } 1678 1679 static bool simplifyInstsInBlock(SCCPSolver &Solver, BasicBlock &BB, 1680 SmallPtrSetImpl<Value *> &InsertedValues, 1681 Statistic &InstRemovedStat, 1682 Statistic &InstReplacedStat) { 1683 bool MadeChanges = false; 1684 for (Instruction &Inst : make_early_inc_range(BB)) { 1685 if (Inst.getType()->isVoidTy()) 1686 continue; 1687 if (tryToReplaceWithConstant(Solver, &Inst)) { 1688 if (Inst.isSafeToRemove()) 1689 Inst.eraseFromParent(); 1690 // Hey, we just changed something! 1691 MadeChanges = true; 1692 ++InstRemovedStat; 1693 } else if (isa<SExtInst>(&Inst)) { 1694 Value *ExtOp = Inst.getOperand(0); 1695 if (isa<Constant>(ExtOp) || InsertedValues.count(ExtOp)) 1696 continue; 1697 const ValueLatticeElement &IV = Solver.getLatticeValueFor(ExtOp); 1698 if (!IV.isConstantRange(/*UndefAllowed=*/false)) 1699 continue; 1700 if (IV.getConstantRange().isAllNonNegative()) { 1701 auto *ZExt = new ZExtInst(ExtOp, Inst.getType(), "", &Inst); 1702 InsertedValues.insert(ZExt); 1703 Inst.replaceAllUsesWith(ZExt); 1704 Solver.removeLatticeValueFor(&Inst); 1705 Inst.eraseFromParent(); 1706 InstReplacedStat++; 1707 MadeChanges = true; 1708 } 1709 } 1710 } 1711 return MadeChanges; 1712 } 1713 1714 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1715 // and return true if the function was modified. 1716 static bool runSCCP(Function &F, const DataLayout &DL, 1717 const TargetLibraryInfo *TLI) { 1718 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1719 SCCPSolver Solver( 1720 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1721 F.getContext()); 1722 1723 // Mark the first block of the function as being executable. 1724 Solver.MarkBlockExecutable(&F.front()); 1725 1726 // Mark all arguments to the function as being overdefined. 1727 for (Argument &AI : F.args()) 1728 Solver.markOverdefined(&AI); 1729 1730 // Solve for constants. 1731 bool ResolvedUndefs = true; 1732 while (ResolvedUndefs) { 1733 Solver.Solve(); 1734 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1735 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1736 } 1737 1738 bool MadeChanges = false; 1739 1740 // If we decided that there are basic blocks that are dead in this function, 1741 // delete their contents now. Note that we cannot actually delete the blocks, 1742 // as we cannot modify the CFG of the function. 1743 1744 SmallPtrSet<Value *, 32> InsertedValues; 1745 for (BasicBlock &BB : F) { 1746 if (!Solver.isBlockExecutable(&BB)) { 1747 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1748 1749 ++NumDeadBlocks; 1750 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB).first; 1751 1752 MadeChanges = true; 1753 continue; 1754 } 1755 1756 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1757 NumInstRemoved, NumInstReplaced); 1758 } 1759 1760 return MadeChanges; 1761 } 1762 1763 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1764 const DataLayout &DL = F.getParent()->getDataLayout(); 1765 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1766 if (!runSCCP(F, DL, &TLI)) 1767 return PreservedAnalyses::all(); 1768 1769 auto PA = PreservedAnalyses(); 1770 PA.preserve<GlobalsAA>(); 1771 PA.preserveSet<CFGAnalyses>(); 1772 return PA; 1773 } 1774 1775 namespace { 1776 1777 //===--------------------------------------------------------------------===// 1778 // 1779 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1780 /// Sparse Conditional Constant Propagator. 1781 /// 1782 class SCCPLegacyPass : public FunctionPass { 1783 public: 1784 // Pass identification, replacement for typeid 1785 static char ID; 1786 1787 SCCPLegacyPass() : FunctionPass(ID) { 1788 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1789 } 1790 1791 void getAnalysisUsage(AnalysisUsage &AU) const override { 1792 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1793 AU.addPreserved<GlobalsAAWrapperPass>(); 1794 AU.setPreservesCFG(); 1795 } 1796 1797 // runOnFunction - Run the Sparse Conditional Constant Propagation 1798 // algorithm, and return true if the function was modified. 1799 bool runOnFunction(Function &F) override { 1800 if (skipFunction(F)) 1801 return false; 1802 const DataLayout &DL = F.getParent()->getDataLayout(); 1803 const TargetLibraryInfo *TLI = 1804 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1805 return runSCCP(F, DL, TLI); 1806 } 1807 }; 1808 1809 } // end anonymous namespace 1810 1811 char SCCPLegacyPass::ID = 0; 1812 1813 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1814 "Sparse Conditional Constant Propagation", false, false) 1815 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1816 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1817 "Sparse Conditional Constant Propagation", false, false) 1818 1819 // createSCCPPass - This is the public interface to this file. 1820 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1821 1822 static void findReturnsToZap(Function &F, 1823 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1824 SCCPSolver &Solver) { 1825 // We can only do this if we know that nothing else can call the function. 1826 if (!Solver.isArgumentTrackedFunction(&F)) 1827 return; 1828 1829 // There is a non-removable musttail call site of this function. Zapping 1830 // returns is not allowed. 1831 if (Solver.isMustTailCallee(&F)) { 1832 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1833 << " due to present musttail call of it\n"); 1834 return; 1835 } 1836 1837 assert( 1838 all_of(F.users(), 1839 [&Solver](User *U) { 1840 if (isa<Instruction>(U) && 1841 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1842 return true; 1843 // Non-callsite uses are not impacted by zapping. Also, constant 1844 // uses (like blockaddresses) could stuck around, without being 1845 // used in the underlying IR, meaning we do not have lattice 1846 // values for them. 1847 if (!isa<CallBase>(U)) 1848 return true; 1849 if (U->getType()->isStructTy()) { 1850 return all_of(Solver.getStructLatticeValueFor(U), 1851 [](const ValueLatticeElement &LV) { 1852 return !isOverdefined(LV); 1853 }); 1854 } 1855 return !isOverdefined(Solver.getLatticeValueFor(U)); 1856 }) && 1857 "We can only zap functions where all live users have a concrete value"); 1858 1859 for (BasicBlock &BB : F) { 1860 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1861 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1862 << "musttail call : " << *CI << "\n"); 1863 (void)CI; 1864 return; 1865 } 1866 1867 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1868 if (!isa<UndefValue>(RI->getOperand(0))) 1869 ReturnsToZap.push_back(RI); 1870 } 1871 } 1872 1873 static bool removeNonFeasibleEdges(const SCCPSolver &Solver, BasicBlock *BB, 1874 DomTreeUpdater &DTU) { 1875 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors; 1876 bool HasNonFeasibleEdges = false; 1877 for (BasicBlock *Succ : successors(BB)) { 1878 if (Solver.isEdgeFeasible(BB, Succ)) 1879 FeasibleSuccessors.insert(Succ); 1880 else 1881 HasNonFeasibleEdges = true; 1882 } 1883 1884 // All edges feasible, nothing to do. 1885 if (!HasNonFeasibleEdges) 1886 return false; 1887 1888 // SCCP can only determine non-feasible edges for br, switch and indirectbr. 1889 Instruction *TI = BB->getTerminator(); 1890 assert((isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1891 isa<IndirectBrInst>(TI)) && 1892 "Terminator must be a br, switch or indirectbr"); 1893 1894 if (FeasibleSuccessors.size() == 1) { 1895 // Replace with an unconditional branch to the only feasible successor. 1896 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin(); 1897 SmallVector<DominatorTree::UpdateType, 8> Updates; 1898 bool HaveSeenOnlyFeasibleSuccessor = false; 1899 for (BasicBlock *Succ : successors(BB)) { 1900 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) { 1901 // Don't remove the edge to the only feasible successor the first time 1902 // we see it. We still do need to remove any multi-edges to it though. 1903 HaveSeenOnlyFeasibleSuccessor = true; 1904 continue; 1905 } 1906 1907 Succ->removePredecessor(BB); 1908 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1909 } 1910 1911 BranchInst::Create(OnlyFeasibleSuccessor, BB); 1912 TI->eraseFromParent(); 1913 DTU.applyUpdatesPermissive(Updates); 1914 } else if (FeasibleSuccessors.size() > 1) { 1915 SwitchInstProfUpdateWrapper SI(*cast<SwitchInst>(TI)); 1916 SmallVector<DominatorTree::UpdateType, 8> Updates; 1917 for (auto CI = SI->case_begin(); CI != SI->case_end();) { 1918 if (FeasibleSuccessors.contains(CI->getCaseSuccessor())) { 1919 ++CI; 1920 continue; 1921 } 1922 1923 BasicBlock *Succ = CI->getCaseSuccessor(); 1924 Succ->removePredecessor(BB); 1925 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1926 SI.removeCase(CI); 1927 // Don't increment CI, as we removed a case. 1928 } 1929 1930 DTU.applyUpdatesPermissive(Updates); 1931 } else { 1932 llvm_unreachable("Must have at least one feasible successor"); 1933 } 1934 return true; 1935 } 1936 1937 bool llvm::runIPSCCP( 1938 Module &M, const DataLayout &DL, 1939 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1940 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1941 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1942 1943 // Loop over all functions, marking arguments to those with their addresses 1944 // taken or that are external as overdefined. 1945 for (Function &F : M) { 1946 if (F.isDeclaration()) 1947 continue; 1948 1949 Solver.addAnalysis(F, getAnalysis(F)); 1950 1951 // Determine if we can track the function's return values. If so, add the 1952 // function to the solver's set of return-tracked functions. 1953 if (canTrackReturnsInterprocedurally(&F)) 1954 Solver.AddTrackedFunction(&F); 1955 1956 // Determine if we can track the function's arguments. If so, add the 1957 // function to the solver's set of argument-tracked functions. 1958 if (canTrackArgumentsInterprocedurally(&F)) { 1959 Solver.AddArgumentTrackedFunction(&F); 1960 continue; 1961 } 1962 1963 // Assume the function is called. 1964 Solver.MarkBlockExecutable(&F.front()); 1965 1966 // Assume nothing about the incoming arguments. 1967 for (Argument &AI : F.args()) 1968 Solver.markOverdefined(&AI); 1969 } 1970 1971 // Determine if we can track any of the module's global variables. If so, add 1972 // the global variables we can track to the solver's set of tracked global 1973 // variables. 1974 for (GlobalVariable &G : M.globals()) { 1975 G.removeDeadConstantUsers(); 1976 if (canTrackGlobalVariableInterprocedurally(&G)) 1977 Solver.TrackValueOfGlobalVariable(&G); 1978 } 1979 1980 // Solve for constants. 1981 bool ResolvedUndefs = true; 1982 Solver.Solve(); 1983 while (ResolvedUndefs) { 1984 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1985 ResolvedUndefs = false; 1986 for (Function &F : M) { 1987 if (Solver.ResolvedUndefsIn(F)) 1988 ResolvedUndefs = true; 1989 } 1990 if (ResolvedUndefs) 1991 Solver.Solve(); 1992 } 1993 1994 bool MadeChanges = false; 1995 1996 // Iterate over all of the instructions in the module, replacing them with 1997 // constants if we have found them to be of constant values. 1998 1999 for (Function &F : M) { 2000 if (F.isDeclaration()) 2001 continue; 2002 2003 SmallVector<BasicBlock *, 512> BlocksToErase; 2004 2005 if (Solver.isBlockExecutable(&F.front())) { 2006 bool ReplacedPointerArg = false; 2007 for (Argument &Arg : F.args()) { 2008 if (!Arg.use_empty() && tryToReplaceWithConstant(Solver, &Arg)) { 2009 ReplacedPointerArg |= Arg.getType()->isPointerTy(); 2010 ++IPNumArgsElimed; 2011 } 2012 } 2013 2014 // If we replaced an argument, the argmemonly and 2015 // inaccessiblemem_or_argmemonly attributes do not hold any longer. Remove 2016 // them from both the function and callsites. 2017 if (ReplacedPointerArg) { 2018 AttrBuilder AttributesToRemove; 2019 AttributesToRemove.addAttribute(Attribute::ArgMemOnly); 2020 AttributesToRemove.addAttribute(Attribute::InaccessibleMemOrArgMemOnly); 2021 F.removeAttributes(AttributeList::FunctionIndex, AttributesToRemove); 2022 2023 for (User *U : F.users()) { 2024 auto *CB = dyn_cast<CallBase>(U); 2025 if (!CB || CB->getCalledFunction() != &F) 2026 continue; 2027 2028 CB->removeAttributes(AttributeList::FunctionIndex, 2029 AttributesToRemove); 2030 } 2031 } 2032 } 2033 2034 SmallPtrSet<Value *, 32> InsertedValues; 2035 for (BasicBlock &BB : F) { 2036 if (!Solver.isBlockExecutable(&BB)) { 2037 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 2038 ++NumDeadBlocks; 2039 2040 MadeChanges = true; 2041 2042 if (&BB != &F.front()) 2043 BlocksToErase.push_back(&BB); 2044 continue; 2045 } 2046 2047 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 2048 IPNumInstRemoved, IPNumInstReplaced); 2049 } 2050 2051 DomTreeUpdater DTU = Solver.getDTU(F); 2052 // Change dead blocks to unreachable. We do it after replacing constants 2053 // in all executable blocks, because changeToUnreachable may remove PHI 2054 // nodes in executable blocks we found values for. The function's entry 2055 // block is not part of BlocksToErase, so we have to handle it separately. 2056 for (BasicBlock *BB : BlocksToErase) { 2057 NumInstRemoved += 2058 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 2059 /*PreserveLCSSA=*/false, &DTU); 2060 } 2061 if (!Solver.isBlockExecutable(&F.front())) 2062 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 2063 /*UseLLVMTrap=*/false, 2064 /*PreserveLCSSA=*/false, &DTU); 2065 2066 for (BasicBlock &BB : F) 2067 MadeChanges |= removeNonFeasibleEdges(Solver, &BB, DTU); 2068 2069 for (BasicBlock *DeadBB : BlocksToErase) 2070 DTU.deleteBB(DeadBB); 2071 2072 for (BasicBlock &BB : F) { 2073 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 2074 Instruction *Inst = &*BI++; 2075 if (Solver.getPredicateInfoFor(Inst)) { 2076 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2077 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2078 Value *Op = II->getOperand(0); 2079 Inst->replaceAllUsesWith(Op); 2080 Inst->eraseFromParent(); 2081 } 2082 } 2083 } 2084 } 2085 } 2086 } 2087 2088 // If we inferred constant or undef return values for a function, we replaced 2089 // all call uses with the inferred value. This means we don't need to bother 2090 // actually returning anything from the function. Replace all return 2091 // instructions with return undef. 2092 // 2093 // Do this in two stages: first identify the functions we should process, then 2094 // actually zap their returns. This is important because we can only do this 2095 // if the address of the function isn't taken. In cases where a return is the 2096 // last use of a function, the order of processing functions would affect 2097 // whether other functions are optimizable. 2098 SmallVector<ReturnInst*, 8> ReturnsToZap; 2099 2100 for (const auto &I : Solver.getTrackedRetVals()) { 2101 Function *F = I.first; 2102 const ValueLatticeElement &ReturnValue = I.second; 2103 2104 // If there is a known constant range for the return value, add !range 2105 // metadata to the function's call sites. 2106 if (ReturnValue.isConstantRange() && 2107 !ReturnValue.getConstantRange().isSingleElement()) { 2108 // Do not add range metadata if the return value may include undef. 2109 if (ReturnValue.isConstantRangeIncludingUndef()) 2110 continue; 2111 2112 auto &CR = ReturnValue.getConstantRange(); 2113 for (User *User : F->users()) { 2114 auto *CB = dyn_cast<CallBase>(User); 2115 if (!CB || CB->getCalledFunction() != F) 2116 continue; 2117 2118 // Limit to cases where the return value is guaranteed to be neither 2119 // poison nor undef. Poison will be outside any range and currently 2120 // values outside of the specified range cause immediate undefined 2121 // behavior. 2122 if (!isGuaranteedNotToBeUndefOrPoison(CB, nullptr, CB)) 2123 continue; 2124 2125 // Do not touch existing metadata for now. 2126 // TODO: We should be able to take the intersection of the existing 2127 // metadata and the inferred range. 2128 if (CB->getMetadata(LLVMContext::MD_range)) 2129 continue; 2130 2131 LLVMContext &Context = CB->getParent()->getContext(); 2132 Metadata *RangeMD[] = { 2133 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getLower())), 2134 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getUpper()))}; 2135 CB->setMetadata(LLVMContext::MD_range, MDNode::get(Context, RangeMD)); 2136 } 2137 continue; 2138 } 2139 if (F->getReturnType()->isVoidTy()) 2140 continue; 2141 if (isConstant(ReturnValue) || ReturnValue.isUnknownOrUndef()) 2142 findReturnsToZap(*F, ReturnsToZap, Solver); 2143 } 2144 2145 for (auto F : Solver.getMRVFunctionsTracked()) { 2146 assert(F->getReturnType()->isStructTy() && 2147 "The return type should be a struct"); 2148 StructType *STy = cast<StructType>(F->getReturnType()); 2149 if (Solver.isStructLatticeConstant(F, STy)) 2150 findReturnsToZap(*F, ReturnsToZap, Solver); 2151 } 2152 2153 // Zap all returns which we've identified as zap to change. 2154 SmallSetVector<Function *, 8> FuncZappedReturn; 2155 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2156 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2157 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2158 // Record all functions that are zapped. 2159 FuncZappedReturn.insert(F); 2160 } 2161 2162 // Remove the returned attribute for zapped functions and the 2163 // corresponding call sites. 2164 for (Function *F : FuncZappedReturn) { 2165 for (Argument &A : F->args()) 2166 F->removeParamAttr(A.getArgNo(), Attribute::Returned); 2167 for (Use &U : F->uses()) { 2168 // Skip over blockaddr users. 2169 if (isa<BlockAddress>(U.getUser())) 2170 continue; 2171 CallBase *CB = cast<CallBase>(U.getUser()); 2172 for (Use &Arg : CB->args()) 2173 CB->removeParamAttr(CB->getArgOperandNo(&Arg), Attribute::Returned); 2174 } 2175 } 2176 2177 // If we inferred constant or undef values for globals variables, we can 2178 // delete the global and any stores that remain to it. 2179 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2180 GlobalVariable *GV = I.first; 2181 if (isOverdefined(I.second)) 2182 continue; 2183 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2184 << "' is constant!\n"); 2185 while (!GV->use_empty()) { 2186 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2187 SI->eraseFromParent(); 2188 MadeChanges = true; 2189 } 2190 M.getGlobalList().erase(GV); 2191 ++IPNumGlobalConst; 2192 } 2193 2194 return MadeChanges; 2195 } 2196