1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements sparse conditional constant propagation and merging: 10 // 11 // Specifically, this: 12 // * Assumes values are constant unless proven otherwise 13 // * Assumes BasicBlocks are dead unless proven otherwise 14 // * Proves values to be constant, and replaces them with constants 15 // * Proves conditional branches to be unconditional 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/Transforms/Scalar/SCCP.h" 20 #include "llvm/ADT/ArrayRef.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/DenseSet.h" 23 #include "llvm/ADT/MapVector.h" 24 #include "llvm/ADT/PointerIntPair.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SmallPtrSet.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/Analysis/ConstantFolding.h" 30 #include "llvm/Analysis/DomTreeUpdater.h" 31 #include "llvm/Analysis/GlobalsModRef.h" 32 #include "llvm/Analysis/InstructionSimplify.h" 33 #include "llvm/Analysis/TargetLibraryInfo.h" 34 #include "llvm/Analysis/ValueLattice.h" 35 #include "llvm/Analysis/ValueLatticeUtils.h" 36 #include "llvm/Analysis/ValueTracking.h" 37 #include "llvm/IR/BasicBlock.h" 38 #include "llvm/IR/Constant.h" 39 #include "llvm/IR/Constants.h" 40 #include "llvm/IR/DataLayout.h" 41 #include "llvm/IR/DerivedTypes.h" 42 #include "llvm/IR/Function.h" 43 #include "llvm/IR/GlobalVariable.h" 44 #include "llvm/IR/InstVisitor.h" 45 #include "llvm/IR/InstrTypes.h" 46 #include "llvm/IR/Instruction.h" 47 #include "llvm/IR/Instructions.h" 48 #include "llvm/IR/Module.h" 49 #include "llvm/IR/PassManager.h" 50 #include "llvm/IR/Type.h" 51 #include "llvm/IR/User.h" 52 #include "llvm/IR/Value.h" 53 #include "llvm/InitializePasses.h" 54 #include "llvm/Pass.h" 55 #include "llvm/Support/Casting.h" 56 #include "llvm/Support/Debug.h" 57 #include "llvm/Support/ErrorHandling.h" 58 #include "llvm/Support/raw_ostream.h" 59 #include "llvm/Transforms/Scalar.h" 60 #include "llvm/Transforms/Utils/Local.h" 61 #include "llvm/Transforms/Utils/PredicateInfo.h" 62 #include <cassert> 63 #include <utility> 64 #include <vector> 65 66 using namespace llvm; 67 68 #define DEBUG_TYPE "sccp" 69 70 STATISTIC(NumInstRemoved, "Number of instructions removed"); 71 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 72 STATISTIC(NumInstReplaced, 73 "Number of instructions replaced with (simpler) instruction"); 74 75 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 76 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 77 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 78 STATISTIC( 79 IPNumInstReplaced, 80 "Number of instructions replaced with (simpler) instruction by IPSCCP"); 81 82 // The maximum number of range extensions allowed for operations requiring 83 // widening. 84 static const unsigned MaxNumRangeExtensions = 10; 85 86 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 87 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 88 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 89 MaxNumRangeExtensions); 90 } 91 namespace { 92 93 // Helper to check if \p LV is either a constant or a constant 94 // range with a single element. This should cover exactly the same cases as the 95 // old ValueLatticeElement::isConstant() and is intended to be used in the 96 // transition to ValueLatticeElement. 97 bool isConstant(const ValueLatticeElement &LV) { 98 return LV.isConstant() || 99 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 100 } 101 102 // Helper to check if \p LV is either overdefined or a constant range with more 103 // than a single element. This should cover exactly the same cases as the old 104 // ValueLatticeElement::isOverdefined() and is intended to be used in the 105 // transition to ValueLatticeElement. 106 bool isOverdefined(const ValueLatticeElement &LV) { 107 return !LV.isUnknownOrUndef() && !isConstant(LV); 108 } 109 110 //===----------------------------------------------------------------------===// 111 // 112 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 113 /// Constant Propagation. 114 /// 115 class SCCPSolver : public InstVisitor<SCCPSolver> { 116 const DataLayout &DL; 117 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 118 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 119 DenseMap<Value *, ValueLatticeElement> 120 ValueState; // The state each value is in. 121 122 /// StructValueState - This maintains ValueState for values that have 123 /// StructType, for example for formal arguments, calls, insertelement, etc. 124 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 125 126 /// GlobalValue - If we are tracking any values for the contents of a global 127 /// variable, we keep a mapping from the constant accessor to the element of 128 /// the global, to the currently known value. If the value becomes 129 /// overdefined, it's entry is simply removed from this map. 130 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 131 132 /// TrackedRetVals - If we are tracking arguments into and the return 133 /// value out of a function, it will have an entry in this map, indicating 134 /// what the known return value for the function is. 135 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 136 137 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 138 /// that return multiple values. 139 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 140 TrackedMultipleRetVals; 141 142 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 143 /// represented here for efficient lookup. 144 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 145 146 /// MustTailFunctions - Each function here is a callee of non-removable 147 /// musttail call site. 148 SmallPtrSet<Function *, 16> MustTailCallees; 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 /// AddMustTailCallee - If the SCCP solver finds that this function is called 241 /// from non-removable musttail call site. 242 void AddMustTailCallee(Function *F) { 243 MustTailCallees.insert(F); 244 } 245 246 /// Returns true if the given function is called from non-removable musttail 247 /// call site. 248 bool isMustTailCallee(Function *F) { 249 return MustTailCallees.count(F); 250 } 251 252 void AddArgumentTrackedFunction(Function *F) { 253 TrackingIncomingArguments.insert(F); 254 } 255 256 /// Returns true if the given function is in the solver's set of 257 /// argument-tracked functions. 258 bool isArgumentTrackedFunction(Function *F) { 259 return TrackingIncomingArguments.count(F); 260 } 261 262 /// Solve - Solve for constants and executable blocks. 263 void Solve(); 264 265 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 266 /// that branches on undef values cannot reach any of their successors. 267 /// However, this is not a safe assumption. After we solve dataflow, this 268 /// method should be use to handle this. If this returns true, the solver 269 /// should be rerun. 270 bool ResolvedUndefsIn(Function &F); 271 272 bool isBlockExecutable(BasicBlock *BB) const { 273 return BBExecutable.count(BB); 274 } 275 276 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 277 // block to the 'To' basic block is currently feasible. 278 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 279 280 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 281 std::vector<ValueLatticeElement> StructValues; 282 auto *STy = dyn_cast<StructType>(V->getType()); 283 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 284 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 285 auto I = StructValueState.find(std::make_pair(V, i)); 286 assert(I != StructValueState.end() && "Value not in valuemap!"); 287 StructValues.push_back(I->second); 288 } 289 return StructValues; 290 } 291 292 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 293 294 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 295 assert(!V->getType()->isStructTy() && 296 "Should use getStructLatticeValueFor"); 297 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 298 ValueState.find(V); 299 assert(I != ValueState.end() && 300 "V not found in ValueState nor Paramstate map!"); 301 return I->second; 302 } 303 304 /// getTrackedRetVals - Get the inferred return value map. 305 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 306 return TrackedRetVals; 307 } 308 309 /// getTrackedGlobals - Get and return the set of inferred initializers for 310 /// global variables. 311 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 312 return TrackedGlobals; 313 } 314 315 /// getMRVFunctionsTracked - Get the set of functions which return multiple 316 /// values tracked by the pass. 317 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 318 return MRVFunctionsTracked; 319 } 320 321 /// getMustTailCallees - Get the set of functions which are called 322 /// from non-removable musttail call sites. 323 const SmallPtrSet<Function *, 16> getMustTailCallees() { 324 return MustTailCallees; 325 } 326 327 /// markOverdefined - Mark the specified value overdefined. This 328 /// works with both scalars and structs. 329 void markOverdefined(Value *V) { 330 if (auto *STy = dyn_cast<StructType>(V->getType())) 331 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 332 markOverdefined(getStructValueState(V, i), V); 333 else 334 markOverdefined(ValueState[V], V); 335 } 336 337 // isStructLatticeConstant - Return true if all the lattice values 338 // corresponding to elements of the structure are constants, 339 // false otherwise. 340 bool isStructLatticeConstant(Function *F, StructType *STy) { 341 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 342 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 343 assert(It != TrackedMultipleRetVals.end()); 344 ValueLatticeElement LV = It->second; 345 if (!isConstant(LV)) 346 return false; 347 } 348 return true; 349 } 350 351 /// Helper to return a Constant if \p LV is either a constant or a constant 352 /// range with a single element. 353 Constant *getConstant(const ValueLatticeElement &LV) const { 354 if (LV.isConstant()) 355 return LV.getConstant(); 356 357 if (LV.isConstantRange()) { 358 auto &CR = LV.getConstantRange(); 359 if (CR.getSingleElement()) 360 return ConstantInt::get(Ctx, *CR.getSingleElement()); 361 } 362 return nullptr; 363 } 364 365 private: 366 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 367 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 368 } 369 370 // pushToWorkList - Helper for markConstant/markOverdefined 371 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 372 if (IV.isOverdefined()) 373 return OverdefinedInstWorkList.push_back(V); 374 InstWorkList.push_back(V); 375 } 376 377 // Helper to push \p V to the worklist, after updating it to \p IV. Also 378 // prints a debug message with the updated value. 379 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 380 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 381 pushToWorkList(IV, V); 382 } 383 384 // markConstant - Make a value be marked as "constant". If the value 385 // is not already a constant, add it to the instruction work list so that 386 // the users of the instruction are updated later. 387 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 388 bool MayIncludeUndef = false) { 389 if (!IV.markConstant(C, MayIncludeUndef)) 390 return false; 391 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 392 pushToWorkList(IV, V); 393 return true; 394 } 395 396 bool markConstant(Value *V, Constant *C) { 397 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 398 return markConstant(ValueState[V], V, C); 399 } 400 401 // markOverdefined - Make a value be marked as "overdefined". If the 402 // value is not already overdefined, add it to the overdefined instruction 403 // work list so that the users of the instruction are updated later. 404 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 405 if (!IV.markOverdefined()) return false; 406 407 LLVM_DEBUG(dbgs() << "markOverdefined: "; 408 if (auto *F = dyn_cast<Function>(V)) dbgs() 409 << "Function '" << F->getName() << "'\n"; 410 else dbgs() << *V << '\n'); 411 // Only instructions go on the work list 412 pushToWorkList(IV, V); 413 return true; 414 } 415 416 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 417 /// changes. 418 bool mergeInValue(ValueLatticeElement &IV, Value *V, 419 ValueLatticeElement MergeWithV, 420 ValueLatticeElement::MergeOptions Opts = { 421 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 422 if (IV.mergeIn(MergeWithV, Opts)) { 423 pushToWorkList(IV, V); 424 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 425 << IV << "\n"); 426 return true; 427 } 428 return false; 429 } 430 431 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 432 ValueLatticeElement::MergeOptions Opts = { 433 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 434 assert(!V->getType()->isStructTy() && 435 "non-structs should use markConstant"); 436 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 437 } 438 439 /// getValueState - Return the ValueLatticeElement object that corresponds to 440 /// the value. This function handles the case when the value hasn't been seen 441 /// yet by properly seeding constants etc. 442 ValueLatticeElement &getValueState(Value *V) { 443 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 444 445 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 446 ValueLatticeElement &LV = I.first->second; 447 448 if (!I.second) 449 return LV; // Common case, already in the map. 450 451 if (auto *C = dyn_cast<Constant>(V)) 452 LV.markConstant(C); // Constants are constant 453 454 // All others are unknown by default. 455 return LV; 456 } 457 458 /// getStructValueState - Return the ValueLatticeElement object that 459 /// corresponds to the value/field pair. This function handles the case when 460 /// the value hasn't been seen yet by properly seeding constants etc. 461 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 462 assert(V->getType()->isStructTy() && "Should use getValueState"); 463 assert(i < cast<StructType>(V->getType())->getNumElements() && 464 "Invalid element #"); 465 466 auto I = StructValueState.insert( 467 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 468 ValueLatticeElement &LV = I.first->second; 469 470 if (!I.second) 471 return LV; // Common case, already in the map. 472 473 if (auto *C = dyn_cast<Constant>(V)) { 474 Constant *Elt = C->getAggregateElement(i); 475 476 if (!Elt) 477 LV.markOverdefined(); // Unknown sort of constant. 478 else if (isa<UndefValue>(Elt)) 479 ; // Undef values remain unknown. 480 else 481 LV.markConstant(Elt); // Constants are constant. 482 } 483 484 // All others are underdefined by default. 485 return LV; 486 } 487 488 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 489 /// work list if it is not already executable. 490 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 491 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 492 return false; // This edge is already known to be executable! 493 494 if (!MarkBlockExecutable(Dest)) { 495 // If the destination is already executable, we just made an *edge* 496 // feasible that wasn't before. Revisit the PHI nodes in the block 497 // because they have potentially new operands. 498 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 499 << " -> " << Dest->getName() << '\n'); 500 501 for (PHINode &PN : Dest->phis()) 502 visitPHINode(PN); 503 } 504 return true; 505 } 506 507 // getFeasibleSuccessors - Return a vector of booleans to indicate which 508 // successors are reachable from a given terminator instruction. 509 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 510 511 // OperandChangedState - This method is invoked on all of the users of an 512 // instruction that was just changed state somehow. Based on this 513 // information, we need to update the specified user of this instruction. 514 void OperandChangedState(Instruction *I) { 515 if (BBExecutable.count(I->getParent())) // Inst is executable? 516 visit(*I); 517 } 518 519 // Add U as additional user of V. 520 void addAdditionalUser(Value *V, User *U) { 521 auto Iter = AdditionalUsers.insert({V, {}}); 522 Iter.first->second.insert(U); 523 } 524 525 // Mark I's users as changed, including AdditionalUsers. 526 void markUsersAsChanged(Value *I) { 527 // Functions include their arguments in the use-list. Changed function 528 // values mean that the result of the function changed. We only need to 529 // update the call sites with the new function result and do not have to 530 // propagate the call arguments. 531 if (isa<Function>(I)) { 532 for (User *U : I->users()) { 533 if (auto *CB = dyn_cast<CallBase>(U)) 534 handleCallResult(*CB); 535 } 536 } else { 537 for (User *U : I->users()) 538 if (auto *UI = dyn_cast<Instruction>(U)) 539 OperandChangedState(UI); 540 } 541 542 auto Iter = AdditionalUsers.find(I); 543 if (Iter != AdditionalUsers.end()) { 544 for (User *U : Iter->second) 545 if (auto *UI = dyn_cast<Instruction>(U)) 546 OperandChangedState(UI); 547 } 548 } 549 void handleCallOverdefined(CallBase &CB); 550 void handleCallResult(CallBase &CB); 551 void handleCallArguments(CallBase &CB); 552 553 private: 554 friend class InstVisitor<SCCPSolver>; 555 556 // visit implementations - Something changed in this instruction. Either an 557 // operand made a transition, or the instruction is newly executable. Change 558 // the value type of I to reflect these changes if appropriate. 559 void visitPHINode(PHINode &I); 560 561 // Terminators 562 563 void visitReturnInst(ReturnInst &I); 564 void visitTerminator(Instruction &TI); 565 566 void visitCastInst(CastInst &I); 567 void visitSelectInst(SelectInst &I); 568 void visitUnaryOperator(Instruction &I); 569 void visitBinaryOperator(Instruction &I); 570 void visitCmpInst(CmpInst &I); 571 void visitExtractValueInst(ExtractValueInst &EVI); 572 void visitInsertValueInst(InsertValueInst &IVI); 573 574 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 575 markOverdefined(&CPI); 576 visitTerminator(CPI); 577 } 578 579 // Instructions that cannot be folded away. 580 581 void visitStoreInst (StoreInst &I); 582 void visitLoadInst (LoadInst &I); 583 void visitGetElementPtrInst(GetElementPtrInst &I); 584 585 void visitCallInst (CallInst &I) { 586 visitCallBase(I); 587 } 588 589 void visitInvokeInst (InvokeInst &II) { 590 visitCallBase(II); 591 visitTerminator(II); 592 } 593 594 void visitCallBrInst (CallBrInst &CBI) { 595 visitCallBase(CBI); 596 visitTerminator(CBI); 597 } 598 599 void visitCallBase (CallBase &CB); 600 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 601 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 602 void visitFenceInst (FenceInst &I) { /*returns void*/ } 603 604 void visitInstruction(Instruction &I) { 605 // All the instructions we don't do any special handling for just 606 // go to overdefined. 607 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 608 markOverdefined(&I); 609 } 610 }; 611 612 } // end anonymous namespace 613 614 // getFeasibleSuccessors - Return a vector of booleans to indicate which 615 // successors are reachable from a given terminator instruction. 616 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 617 SmallVectorImpl<bool> &Succs) { 618 Succs.resize(TI.getNumSuccessors()); 619 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 620 if (BI->isUnconditional()) { 621 Succs[0] = true; 622 return; 623 } 624 625 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 626 ConstantInt *CI = getConstantInt(BCValue); 627 if (!CI) { 628 // Overdefined condition variables, and branches on unfoldable constant 629 // conditions, mean the branch could go either way. 630 if (!BCValue.isUnknownOrUndef()) 631 Succs[0] = Succs[1] = true; 632 return; 633 } 634 635 // Constant condition variables mean the branch can only go a single way. 636 Succs[CI->isZero()] = true; 637 return; 638 } 639 640 // Unwinding instructions successors are always executable. 641 if (TI.isExceptionalTerminator()) { 642 Succs.assign(TI.getNumSuccessors(), true); 643 return; 644 } 645 646 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 647 if (!SI->getNumCases()) { 648 Succs[0] = true; 649 return; 650 } 651 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 652 if (ConstantInt *CI = getConstantInt(SCValue)) { 653 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 654 return; 655 } 656 657 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 658 // is ready. 659 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 660 const ConstantRange &Range = SCValue.getConstantRange(); 661 for (const auto &Case : SI->cases()) { 662 const APInt &CaseValue = Case.getCaseValue()->getValue(); 663 if (Range.contains(CaseValue)) 664 Succs[Case.getSuccessorIndex()] = true; 665 } 666 667 // TODO: Determine whether default case is reachable. 668 Succs[SI->case_default()->getSuccessorIndex()] = true; 669 return; 670 } 671 672 // Overdefined or unknown condition? All destinations are executable! 673 if (!SCValue.isUnknownOrUndef()) 674 Succs.assign(TI.getNumSuccessors(), true); 675 return; 676 } 677 678 // In case of indirect branch and its address is a blockaddress, we mark 679 // the target as executable. 680 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 681 // Casts are folded by visitCastInst. 682 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 683 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 684 if (!Addr) { // Overdefined or unknown condition? 685 // All destinations are executable! 686 if (!IBRValue.isUnknownOrUndef()) 687 Succs.assign(TI.getNumSuccessors(), true); 688 return; 689 } 690 691 BasicBlock* T = Addr->getBasicBlock(); 692 assert(Addr->getFunction() == T->getParent() && 693 "Block address of a different function ?"); 694 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 695 // This is the target. 696 if (IBR->getDestination(i) == T) { 697 Succs[i] = true; 698 return; 699 } 700 } 701 702 // If we didn't find our destination in the IBR successor list, then we 703 // have undefined behavior. Its ok to assume no successor is executable. 704 return; 705 } 706 707 // In case of callbr, we pessimistically assume that all successors are 708 // feasible. 709 if (isa<CallBrInst>(&TI)) { 710 Succs.assign(TI.getNumSuccessors(), true); 711 return; 712 } 713 714 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 715 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 716 } 717 718 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 719 // block to the 'To' basic block is currently feasible. 720 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 721 // Check if we've called markEdgeExecutable on the edge yet. (We could 722 // be more aggressive and try to consider edges which haven't been marked 723 // yet, but there isn't any need.) 724 return KnownFeasibleEdges.count(Edge(From, To)); 725 } 726 727 // visit Implementations - Something changed in this instruction, either an 728 // operand made a transition, or the instruction is newly executable. Change 729 // the value type of I to reflect these changes if appropriate. This method 730 // makes sure to do the following actions: 731 // 732 // 1. If a phi node merges two constants in, and has conflicting value coming 733 // from different branches, or if the PHI node merges in an overdefined 734 // value, then the PHI node becomes overdefined. 735 // 2. If a phi node merges only constants in, and they all agree on value, the 736 // PHI node becomes a constant value equal to that. 737 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 738 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 739 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 740 // 6. If a conditional branch has a value that is constant, make the selected 741 // destination executable 742 // 7. If a conditional branch has a value that is overdefined, make all 743 // successors executable. 744 void SCCPSolver::visitPHINode(PHINode &PN) { 745 // If this PN returns a struct, just mark the result overdefined. 746 // TODO: We could do a lot better than this if code actually uses this. 747 if (PN.getType()->isStructTy()) 748 return (void)markOverdefined(&PN); 749 750 if (getValueState(&PN).isOverdefined()) 751 return; // Quick exit 752 753 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 754 // and slow us down a lot. Just mark them overdefined. 755 if (PN.getNumIncomingValues() > 64) 756 return (void)markOverdefined(&PN); 757 758 unsigned NumActiveIncoming = 0; 759 760 // Look at all of the executable operands of the PHI node. If any of them 761 // are overdefined, the PHI becomes overdefined as well. If they are all 762 // constant, and they agree with each other, the PHI becomes the identical 763 // constant. If they are constant and don't agree, the PHI is a constant 764 // range. If there are no executable operands, the PHI remains unknown. 765 ValueLatticeElement PhiState = getValueState(&PN); 766 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 767 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 768 continue; 769 770 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 771 PhiState.mergeIn(IV); 772 NumActiveIncoming++; 773 if (PhiState.isOverdefined()) 774 break; 775 } 776 777 // We allow up to 1 range extension per active incoming value and one 778 // additional extension. Note that we manually adjust the number of range 779 // extensions to match the number of active incoming values. This helps to 780 // limit multiple extensions caused by the same incoming value, if other 781 // incoming values are equal. 782 mergeInValue(&PN, PhiState, 783 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 784 NumActiveIncoming + 1)); 785 ValueLatticeElement &PhiStateRef = getValueState(&PN); 786 PhiStateRef.setNumRangeExtensions( 787 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 788 } 789 790 void SCCPSolver::visitReturnInst(ReturnInst &I) { 791 if (I.getNumOperands() == 0) return; // ret void 792 793 Function *F = I.getParent()->getParent(); 794 Value *ResultOp = I.getOperand(0); 795 796 // If we are tracking the return value of this function, merge it in. 797 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 798 auto TFRVI = TrackedRetVals.find(F); 799 if (TFRVI != TrackedRetVals.end()) { 800 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 801 return; 802 } 803 } 804 805 // Handle functions that return multiple values. 806 if (!TrackedMultipleRetVals.empty()) { 807 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 808 if (MRVFunctionsTracked.count(F)) 809 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 810 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 811 getStructValueState(ResultOp, i)); 812 } 813 } 814 815 void SCCPSolver::visitTerminator(Instruction &TI) { 816 SmallVector<bool, 16> SuccFeasible; 817 getFeasibleSuccessors(TI, SuccFeasible); 818 819 BasicBlock *BB = TI.getParent(); 820 821 // Mark all feasible successors executable. 822 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 823 if (SuccFeasible[i]) 824 markEdgeExecutable(BB, TI.getSuccessor(i)); 825 } 826 827 void SCCPSolver::visitCastInst(CastInst &I) { 828 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 829 // discover a concrete value later. 830 if (ValueState[&I].isOverdefined()) 831 return; 832 833 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 834 if (Constant *OpC = getConstant(OpSt)) { 835 // Fold the constant as we build. 836 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 837 if (isa<UndefValue>(C)) 838 return; 839 // Propagate constant value 840 markConstant(&I, C); 841 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 842 auto &LV = getValueState(&I); 843 ConstantRange OpRange = OpSt.getConstantRange(); 844 Type *DestTy = I.getDestTy(); 845 ConstantRange Res = 846 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 847 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 848 } else if (!OpSt.isUnknownOrUndef()) 849 markOverdefined(&I); 850 } 851 852 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 853 // If this returns a struct, mark all elements over defined, we don't track 854 // structs in structs. 855 if (EVI.getType()->isStructTy()) 856 return (void)markOverdefined(&EVI); 857 858 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 859 // discover a concrete value later. 860 if (ValueState[&EVI].isOverdefined()) 861 return (void)markOverdefined(&EVI); 862 863 // If this is extracting from more than one level of struct, we don't know. 864 if (EVI.getNumIndices() != 1) 865 return (void)markOverdefined(&EVI); 866 867 Value *AggVal = EVI.getAggregateOperand(); 868 if (AggVal->getType()->isStructTy()) { 869 unsigned i = *EVI.idx_begin(); 870 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 871 mergeInValue(getValueState(&EVI), &EVI, EltVal); 872 } else { 873 // Otherwise, must be extracting from an array. 874 return (void)markOverdefined(&EVI); 875 } 876 } 877 878 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 879 auto *STy = dyn_cast<StructType>(IVI.getType()); 880 if (!STy) 881 return (void)markOverdefined(&IVI); 882 883 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 884 // discover a concrete value later. 885 if (isOverdefined(ValueState[&IVI])) 886 return (void)markOverdefined(&IVI); 887 888 // If this has more than one index, we can't handle it, drive all results to 889 // undef. 890 if (IVI.getNumIndices() != 1) 891 return (void)markOverdefined(&IVI); 892 893 Value *Aggr = IVI.getAggregateOperand(); 894 unsigned Idx = *IVI.idx_begin(); 895 896 // Compute the result based on what we're inserting. 897 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 898 // This passes through all values that aren't the inserted element. 899 if (i != Idx) { 900 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 901 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 902 continue; 903 } 904 905 Value *Val = IVI.getInsertedValueOperand(); 906 if (Val->getType()->isStructTy()) 907 // We don't track structs in structs. 908 markOverdefined(getStructValueState(&IVI, i), &IVI); 909 else { 910 ValueLatticeElement InVal = getValueState(Val); 911 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 912 } 913 } 914 } 915 916 void SCCPSolver::visitSelectInst(SelectInst &I) { 917 // If this select returns a struct, just mark the result overdefined. 918 // TODO: We could do a lot better than this if code actually uses this. 919 if (I.getType()->isStructTy()) 920 return (void)markOverdefined(&I); 921 922 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 923 // discover a concrete value later. 924 if (ValueState[&I].isOverdefined()) 925 return (void)markOverdefined(&I); 926 927 ValueLatticeElement CondValue = getValueState(I.getCondition()); 928 if (CondValue.isUnknownOrUndef()) 929 return; 930 931 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 932 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 933 mergeInValue(&I, getValueState(OpVal)); 934 return; 935 } 936 937 // Otherwise, the condition is overdefined or a constant we can't evaluate. 938 // See if we can produce something better than overdefined based on the T/F 939 // value. 940 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 941 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 942 943 bool Changed = ValueState[&I].mergeIn(TVal); 944 Changed |= ValueState[&I].mergeIn(FVal); 945 if (Changed) 946 pushToWorkListMsg(ValueState[&I], &I); 947 } 948 949 // Handle Unary Operators. 950 void SCCPSolver::visitUnaryOperator(Instruction &I) { 951 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 952 953 ValueLatticeElement &IV = ValueState[&I]; 954 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 955 // discover a concrete value later. 956 if (isOverdefined(IV)) 957 return (void)markOverdefined(&I); 958 959 if (isConstant(V0State)) { 960 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 961 962 // op Y -> undef. 963 if (isa<UndefValue>(C)) 964 return; 965 return (void)markConstant(IV, &I, C); 966 } 967 968 // If something is undef, wait for it to resolve. 969 if (!isOverdefined(V0State)) 970 return; 971 972 markOverdefined(&I); 973 } 974 975 // Handle Binary Operators. 976 void SCCPSolver::visitBinaryOperator(Instruction &I) { 977 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 978 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 979 980 ValueLatticeElement &IV = ValueState[&I]; 981 if (IV.isOverdefined()) 982 return; 983 984 // If something is undef, wait for it to resolve. 985 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 986 return; 987 988 if (V1State.isOverdefined() && V2State.isOverdefined()) 989 return (void)markOverdefined(&I); 990 991 // If either of the operands is a constant, try to fold it to a constant. 992 // TODO: Use information from notconstant better. 993 if ((V1State.isConstant() || V2State.isConstant())) { 994 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 995 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 996 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 997 auto *C = dyn_cast_or_null<Constant>(R); 998 if (C) { 999 // X op Y -> undef. 1000 if (isa<UndefValue>(C)) 1001 return; 1002 // Conservatively assume that the result may be based on operands that may 1003 // be undef. Note that we use mergeInValue to combine the constant with 1004 // the existing lattice value for I, as different constants might be found 1005 // after one of the operands go to overdefined, e.g. due to one operand 1006 // being a special floating value. 1007 ValueLatticeElement NewV; 1008 NewV.markConstant(C, /*MayIncludeUndef=*/true); 1009 return (void)mergeInValue(&I, NewV); 1010 } 1011 } 1012 1013 // Only use ranges for binary operators on integers. 1014 if (!I.getType()->isIntegerTy()) 1015 return markOverdefined(&I); 1016 1017 // Try to simplify to a constant range. 1018 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1019 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1020 if (V1State.isConstantRange()) 1021 A = V1State.getConstantRange(); 1022 if (V2State.isConstantRange()) 1023 B = V2State.getConstantRange(); 1024 1025 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1026 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1027 1028 // TODO: Currently we do not exploit special values that produce something 1029 // better than overdefined with an overdefined operand for vector or floating 1030 // point types, like and <4 x i32> overdefined, zeroinitializer. 1031 } 1032 1033 // Handle ICmpInst instruction. 1034 void SCCPSolver::visitCmpInst(CmpInst &I) { 1035 // Do not cache this lookup, getValueState calls later in the function might 1036 // invalidate the reference. 1037 if (isOverdefined(ValueState[&I])) 1038 return (void)markOverdefined(&I); 1039 1040 Value *Op1 = I.getOperand(0); 1041 Value *Op2 = I.getOperand(1); 1042 1043 // For parameters, use ParamState which includes constant range info if 1044 // available. 1045 auto V1State = getValueState(Op1); 1046 auto V2State = getValueState(Op2); 1047 1048 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1049 if (C) { 1050 if (isa<UndefValue>(C)) 1051 return; 1052 ValueLatticeElement CV; 1053 CV.markConstant(C); 1054 mergeInValue(&I, CV); 1055 return; 1056 } 1057 1058 // If operands are still unknown, wait for it to resolve. 1059 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1060 !isConstant(ValueState[&I])) 1061 return; 1062 1063 markOverdefined(&I); 1064 } 1065 1066 // Handle getelementptr instructions. If all operands are constants then we 1067 // can turn this into a getelementptr ConstantExpr. 1068 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1069 if (isOverdefined(ValueState[&I])) 1070 return (void)markOverdefined(&I); 1071 1072 SmallVector<Constant*, 8> Operands; 1073 Operands.reserve(I.getNumOperands()); 1074 1075 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1076 ValueLatticeElement State = getValueState(I.getOperand(i)); 1077 if (State.isUnknownOrUndef()) 1078 return; // Operands are not resolved yet. 1079 1080 if (isOverdefined(State)) 1081 return (void)markOverdefined(&I); 1082 1083 if (Constant *C = getConstant(State)) { 1084 Operands.push_back(C); 1085 continue; 1086 } 1087 1088 return (void)markOverdefined(&I); 1089 } 1090 1091 Constant *Ptr = Operands[0]; 1092 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1093 Constant *C = 1094 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1095 if (isa<UndefValue>(C)) 1096 return; 1097 markConstant(&I, C); 1098 } 1099 1100 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1101 // If this store is of a struct, ignore it. 1102 if (SI.getOperand(0)->getType()->isStructTy()) 1103 return; 1104 1105 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1106 return; 1107 1108 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1109 auto I = TrackedGlobals.find(GV); 1110 if (I == TrackedGlobals.end()) 1111 return; 1112 1113 // Get the value we are storing into the global, then merge it. 1114 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1115 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1116 if (I->second.isOverdefined()) 1117 TrackedGlobals.erase(I); // No need to keep tracking this! 1118 } 1119 1120 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1121 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1122 if (I->getType()->isIntegerTy()) 1123 return ValueLatticeElement::getRange( 1124 getConstantRangeFromMetadata(*Ranges)); 1125 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1126 return ValueLatticeElement::getNot( 1127 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1128 return ValueLatticeElement::getOverdefined(); 1129 } 1130 1131 // Handle load instructions. If the operand is a constant pointer to a constant 1132 // global, we can replace the load with the loaded constant value! 1133 void SCCPSolver::visitLoadInst(LoadInst &I) { 1134 // If this load is of a struct or the load is volatile, just mark the result 1135 // as overdefined. 1136 if (I.getType()->isStructTy() || I.isVolatile()) 1137 return (void)markOverdefined(&I); 1138 1139 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1140 // discover a concrete value later. 1141 if (ValueState[&I].isOverdefined()) 1142 return (void)markOverdefined(&I); 1143 1144 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1145 if (PtrVal.isUnknownOrUndef()) 1146 return; // The pointer is not resolved yet! 1147 1148 ValueLatticeElement &IV = ValueState[&I]; 1149 1150 if (isConstant(PtrVal)) { 1151 Constant *Ptr = getConstant(PtrVal); 1152 1153 // load null is undefined. 1154 if (isa<ConstantPointerNull>(Ptr)) { 1155 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1156 return (void)markOverdefined(IV, &I); 1157 else 1158 return; 1159 } 1160 1161 // Transform load (constant global) into the value loaded. 1162 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1163 if (!TrackedGlobals.empty()) { 1164 // If we are tracking this global, merge in the known value for it. 1165 auto It = TrackedGlobals.find(GV); 1166 if (It != TrackedGlobals.end()) { 1167 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1168 return; 1169 } 1170 } 1171 } 1172 1173 // Transform load from a constant into a constant if possible. 1174 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1175 if (isa<UndefValue>(C)) 1176 return; 1177 return (void)markConstant(IV, &I, C); 1178 } 1179 } 1180 1181 // Fall back to metadata. 1182 mergeInValue(&I, getValueFromMetadata(&I)); 1183 } 1184 1185 void SCCPSolver::visitCallBase(CallBase &CB) { 1186 handleCallResult(CB); 1187 handleCallArguments(CB); 1188 } 1189 1190 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1191 Function *F = CB.getCalledFunction(); 1192 1193 // Void return and not tracking callee, just bail. 1194 if (CB.getType()->isVoidTy()) 1195 return; 1196 1197 // Always mark struct return as overdefined. 1198 if (CB.getType()->isStructTy()) 1199 return (void)markOverdefined(&CB); 1200 1201 // Otherwise, if we have a single return value case, and if the function is 1202 // a declaration, maybe we can constant fold it. 1203 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1204 SmallVector<Constant *, 8> Operands; 1205 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1206 if (AI->get()->getType()->isStructTy()) 1207 return markOverdefined(&CB); // Can't handle struct args. 1208 ValueLatticeElement State = getValueState(*AI); 1209 1210 if (State.isUnknownOrUndef()) 1211 return; // Operands are not resolved yet. 1212 if (isOverdefined(State)) 1213 return (void)markOverdefined(&CB); 1214 assert(isConstant(State) && "Unknown state!"); 1215 Operands.push_back(getConstant(State)); 1216 } 1217 1218 if (isOverdefined(getValueState(&CB))) 1219 return (void)markOverdefined(&CB); 1220 1221 // If we can constant fold this, mark the result of the call as a 1222 // constant. 1223 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1224 // call -> undef. 1225 if (isa<UndefValue>(C)) 1226 return; 1227 return (void)markConstant(&CB, C); 1228 } 1229 } 1230 1231 // Fall back to metadata. 1232 mergeInValue(&CB, getValueFromMetadata(&CB)); 1233 } 1234 1235 void SCCPSolver::handleCallArguments(CallBase &CB) { 1236 Function *F = CB.getCalledFunction(); 1237 // If this is a local function that doesn't have its address taken, mark its 1238 // entry block executable and merge in the actual arguments to the call into 1239 // the formal arguments of the function. 1240 if (!TrackingIncomingArguments.empty() && 1241 TrackingIncomingArguments.count(F)) { 1242 MarkBlockExecutable(&F->front()); 1243 1244 // Propagate information from this call site into the callee. 1245 auto CAI = CB.arg_begin(); 1246 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1247 ++AI, ++CAI) { 1248 // If this argument is byval, and if the function is not readonly, there 1249 // will be an implicit copy formed of the input aggregate. 1250 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1251 markOverdefined(&*AI); 1252 continue; 1253 } 1254 1255 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1256 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1257 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1258 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1259 getMaxWidenStepsOpts()); 1260 } 1261 } else 1262 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1263 } 1264 } 1265 } 1266 1267 void SCCPSolver::handleCallResult(CallBase &CB) { 1268 Function *F = CB.getCalledFunction(); 1269 1270 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1271 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1272 if (ValueState[&CB].isOverdefined()) 1273 return; 1274 1275 Value *CopyOf = CB.getOperand(0); 1276 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1277 auto *PI = getPredicateInfoFor(&CB); 1278 assert(PI && "Missing predicate info for ssa.copy"); 1279 1280 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1281 if (!Constraint) { 1282 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1283 return; 1284 } 1285 1286 CmpInst::Predicate Pred = Constraint->Predicate; 1287 Value *OtherOp = Constraint->OtherOp; 1288 1289 // Wait until OtherOp is resolved. 1290 if (getValueState(OtherOp).isUnknown()) { 1291 addAdditionalUser(OtherOp, &CB); 1292 return; 1293 } 1294 1295 // TODO: Actually filp MayIncludeUndef for the created range to false, 1296 // once most places in the optimizer respect the branches on 1297 // undef/poison are UB rule. The reason why the new range cannot be 1298 // undef is as follows below: 1299 // The new range is based on a branch condition. That guarantees that 1300 // neither of the compare operands can be undef in the branch targets, 1301 // unless we have conditions that are always true/false (e.g. icmp ule 1302 // i32, %a, i32_max). For the latter overdefined/empty range will be 1303 // inferred, but the branch will get folded accordingly anyways. 1304 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1305 1306 ValueLatticeElement CondVal = getValueState(OtherOp); 1307 ValueLatticeElement &IV = ValueState[&CB]; 1308 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1309 auto ImposedCR = 1310 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1311 1312 // Get the range imposed by the condition. 1313 if (CondVal.isConstantRange()) 1314 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1315 Pred, CondVal.getConstantRange()); 1316 1317 // Combine range info for the original value with the new range from the 1318 // condition. 1319 auto CopyOfCR = CopyOfVal.isConstantRange() 1320 ? CopyOfVal.getConstantRange() 1321 : ConstantRange::getFull( 1322 DL.getTypeSizeInBits(CopyOf->getType())); 1323 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1324 // If the existing information is != x, do not use the information from 1325 // a chained predicate, as the != x information is more likely to be 1326 // helpful in practice. 1327 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1328 NewCR = CopyOfCR; 1329 1330 addAdditionalUser(OtherOp, &CB); 1331 mergeInValue( 1332 IV, &CB, 1333 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1334 return; 1335 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1336 // For non-integer values or integer constant expressions, only 1337 // propagate equal constants. 1338 addAdditionalUser(OtherOp, &CB); 1339 mergeInValue(IV, &CB, CondVal); 1340 return; 1341 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1342 !MayIncludeUndef) { 1343 // Propagate inequalities. 1344 addAdditionalUser(OtherOp, &CB); 1345 mergeInValue(IV, &CB, 1346 ValueLatticeElement::getNot(CondVal.getConstant())); 1347 return; 1348 } 1349 1350 return (void)mergeInValue(IV, &CB, CopyOfVal); 1351 } 1352 } 1353 1354 // The common case is that we aren't tracking the callee, either because we 1355 // are not doing interprocedural analysis or the callee is indirect, or is 1356 // external. Handle these cases first. 1357 if (!F || F->isDeclaration()) 1358 return handleCallOverdefined(CB); 1359 1360 // If this is a single/zero retval case, see if we're tracking the function. 1361 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1362 if (!MRVFunctionsTracked.count(F)) 1363 return handleCallOverdefined(CB); // Not tracking this callee. 1364 1365 // If we are tracking this callee, propagate the result of the function 1366 // into this call site. 1367 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1368 mergeInValue(getStructValueState(&CB, i), &CB, 1369 TrackedMultipleRetVals[std::make_pair(F, i)], 1370 getMaxWidenStepsOpts()); 1371 } else { 1372 auto TFRVI = TrackedRetVals.find(F); 1373 if (TFRVI == TrackedRetVals.end()) 1374 return handleCallOverdefined(CB); // Not tracking this callee. 1375 1376 // If so, propagate the return value of the callee into this call result. 1377 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1378 } 1379 } 1380 1381 void SCCPSolver::Solve() { 1382 // Process the work lists until they are empty! 1383 while (!BBWorkList.empty() || !InstWorkList.empty() || 1384 !OverdefinedInstWorkList.empty()) { 1385 // Process the overdefined instruction's work list first, which drives other 1386 // things to overdefined more quickly. 1387 while (!OverdefinedInstWorkList.empty()) { 1388 Value *I = OverdefinedInstWorkList.pop_back_val(); 1389 1390 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1391 1392 // "I" got into the work list because it either made the transition from 1393 // bottom to constant, or to overdefined. 1394 // 1395 // Anything on this worklist that is overdefined need not be visited 1396 // since all of its users will have already been marked as overdefined 1397 // Update all of the users of this instruction's value. 1398 // 1399 markUsersAsChanged(I); 1400 } 1401 1402 // Process the instruction work list. 1403 while (!InstWorkList.empty()) { 1404 Value *I = InstWorkList.pop_back_val(); 1405 1406 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1407 1408 // "I" got into the work list because it made the transition from undef to 1409 // constant. 1410 // 1411 // Anything on this worklist that is overdefined need not be visited 1412 // since all of its users will have already been marked as overdefined. 1413 // Update all of the users of this instruction's value. 1414 // 1415 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1416 markUsersAsChanged(I); 1417 } 1418 1419 // Process the basic block work list. 1420 while (!BBWorkList.empty()) { 1421 BasicBlock *BB = BBWorkList.back(); 1422 BBWorkList.pop_back(); 1423 1424 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1425 1426 // Notify all instructions in this basic block that they are newly 1427 // executable. 1428 visit(BB); 1429 } 1430 } 1431 } 1432 1433 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1434 /// that branches on undef values cannot reach any of their successors. 1435 /// However, this is not a safe assumption. After we solve dataflow, this 1436 /// method should be use to handle this. If this returns true, the solver 1437 /// should be rerun. 1438 /// 1439 /// This method handles this by finding an unresolved branch and marking it one 1440 /// of the edges from the block as being feasible, even though the condition 1441 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1442 /// CFG and only slightly pessimizes the analysis results (by marking one, 1443 /// potentially infeasible, edge feasible). This cannot usefully modify the 1444 /// constraints on the condition of the branch, as that would impact other users 1445 /// of the value. 1446 /// 1447 /// This scan also checks for values that use undefs. It conservatively marks 1448 /// them as overdefined. 1449 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1450 for (BasicBlock &BB : F) { 1451 if (!BBExecutable.count(&BB)) 1452 continue; 1453 1454 for (Instruction &I : BB) { 1455 // Look for instructions which produce undef values. 1456 if (I.getType()->isVoidTy()) continue; 1457 1458 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1459 // Only a few things that can be structs matter for undef. 1460 1461 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1462 if (auto *CB = dyn_cast<CallBase>(&I)) 1463 if (Function *F = CB->getCalledFunction()) 1464 if (MRVFunctionsTracked.count(F)) 1465 continue; 1466 1467 // extractvalue and insertvalue don't need to be marked; they are 1468 // tracked as precisely as their operands. 1469 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1470 continue; 1471 // Send the results of everything else to overdefined. We could be 1472 // more precise than this but it isn't worth bothering. 1473 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1474 ValueLatticeElement &LV = getStructValueState(&I, i); 1475 if (LV.isUnknownOrUndef()) 1476 markOverdefined(LV, &I); 1477 } 1478 continue; 1479 } 1480 1481 ValueLatticeElement &LV = getValueState(&I); 1482 if (!LV.isUnknownOrUndef()) 1483 continue; 1484 1485 // There are two reasons a call can have an undef result 1486 // 1. It could be tracked. 1487 // 2. It could be constant-foldable. 1488 // Because of the way we solve return values, tracked calls must 1489 // never be marked overdefined in ResolvedUndefsIn. 1490 if (auto *CB = dyn_cast<CallBase>(&I)) 1491 if (Function *F = CB->getCalledFunction()) 1492 if (TrackedRetVals.count(F)) 1493 continue; 1494 1495 if (isa<LoadInst>(I)) { 1496 // A load here means one of two things: a load of undef from a global, 1497 // a load from an unknown pointer. Either way, having it return undef 1498 // is okay. 1499 continue; 1500 } 1501 1502 markOverdefined(&I); 1503 return true; 1504 } 1505 1506 // Check to see if we have a branch or switch on an undefined value. If so 1507 // we force the branch to go one way or the other to make the successor 1508 // values live. It doesn't really matter which way we force it. 1509 Instruction *TI = BB.getTerminator(); 1510 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1511 if (!BI->isConditional()) continue; 1512 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1513 continue; 1514 1515 // If the input to SCCP is actually branch on undef, fix the undef to 1516 // false. 1517 if (isa<UndefValue>(BI->getCondition())) { 1518 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1519 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1520 return true; 1521 } 1522 1523 // Otherwise, it is a branch on a symbolic value which is currently 1524 // considered to be undef. Make sure some edge is executable, so a 1525 // branch on "undef" always flows somewhere. 1526 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1527 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1528 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1529 return true; 1530 1531 continue; 1532 } 1533 1534 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1535 // Indirect branch with no successor ?. Its ok to assume it branches 1536 // to no target. 1537 if (IBR->getNumSuccessors() < 1) 1538 continue; 1539 1540 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1541 continue; 1542 1543 // If the input to SCCP is actually branch on undef, fix the undef to 1544 // the first successor of the indirect branch. 1545 if (isa<UndefValue>(IBR->getAddress())) { 1546 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1547 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1548 return true; 1549 } 1550 1551 // Otherwise, it is a branch on a symbolic value which is currently 1552 // considered to be undef. Make sure some edge is executable, so a 1553 // branch on "undef" always flows somewhere. 1554 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1555 // we can assume the branch has undefined behavior instead. 1556 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1557 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1558 return true; 1559 1560 continue; 1561 } 1562 1563 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1564 if (!SI->getNumCases() || 1565 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1566 continue; 1567 1568 // If the input to SCCP is actually switch on undef, fix the undef to 1569 // the first constant. 1570 if (isa<UndefValue>(SI->getCondition())) { 1571 SI->setCondition(SI->case_begin()->getCaseValue()); 1572 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1573 return true; 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: Distinguish between dead code and an LLVM "undef" value. 1580 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1581 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1582 return true; 1583 1584 continue; 1585 } 1586 } 1587 1588 return false; 1589 } 1590 1591 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1592 Constant *Const = nullptr; 1593 if (V->getType()->isStructTy()) { 1594 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1595 if (any_of(IVs, 1596 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1597 return false; 1598 std::vector<Constant *> ConstVals; 1599 auto *ST = cast<StructType>(V->getType()); 1600 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1601 ValueLatticeElement V = IVs[i]; 1602 ConstVals.push_back(isConstant(V) 1603 ? Solver.getConstant(V) 1604 : UndefValue::get(ST->getElementType(i))); 1605 } 1606 Const = ConstantStruct::get(ST, ConstVals); 1607 } else { 1608 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1609 if (isOverdefined(IV)) 1610 return false; 1611 1612 Const = 1613 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1614 } 1615 assert(Const && "Constant is nullptr here!"); 1616 1617 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1618 // unless the call itself can be removed 1619 CallInst *CI = dyn_cast<CallInst>(V); 1620 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1621 Function *F = CI->getCalledFunction(); 1622 1623 // Don't zap returns of the callee 1624 if (F) 1625 Solver.AddMustTailCallee(F); 1626 1627 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1628 << " as a constant\n"); 1629 return false; 1630 } 1631 1632 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1633 1634 // Replaces all of the uses of a variable with uses of the constant. 1635 V->replaceAllUsesWith(Const); 1636 return true; 1637 } 1638 1639 static bool simplifyInstsInBlock(SCCPSolver &Solver, BasicBlock &BB, 1640 SmallPtrSetImpl<Value *> &InsertedValues, 1641 Statistic &InstRemovedStat, 1642 Statistic &InstReplacedStat) { 1643 bool MadeChanges = false; 1644 for (Instruction &Inst : make_early_inc_range(BB)) { 1645 if (Inst.getType()->isVoidTy()) 1646 continue; 1647 if (tryToReplaceWithConstant(Solver, &Inst)) { 1648 if (Inst.isSafeToRemove()) 1649 Inst.eraseFromParent(); 1650 // Hey, we just changed something! 1651 MadeChanges = true; 1652 ++InstRemovedStat; 1653 } else if (isa<SExtInst>(&Inst)) { 1654 Value *ExtOp = Inst.getOperand(0); 1655 if (isa<Constant>(ExtOp) || InsertedValues.count(ExtOp)) 1656 continue; 1657 const ValueLatticeElement &IV = Solver.getLatticeValueFor(ExtOp); 1658 if (!IV.isConstantRange(/*UndefAllowed=*/false)) 1659 continue; 1660 if (IV.getConstantRange().isAllNonNegative()) { 1661 auto *ZExt = new ZExtInst(ExtOp, Inst.getType(), "", &Inst); 1662 InsertedValues.insert(ZExt); 1663 Inst.replaceAllUsesWith(ZExt); 1664 Solver.removeLatticeValueFor(&Inst); 1665 Inst.eraseFromParent(); 1666 InstReplacedStat++; 1667 MadeChanges = true; 1668 } 1669 } 1670 } 1671 return MadeChanges; 1672 } 1673 1674 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1675 // and return true if the function was modified. 1676 static bool runSCCP(Function &F, const DataLayout &DL, 1677 const TargetLibraryInfo *TLI) { 1678 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1679 SCCPSolver Solver( 1680 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1681 F.getContext()); 1682 1683 // Mark the first block of the function as being executable. 1684 Solver.MarkBlockExecutable(&F.front()); 1685 1686 // Mark all arguments to the function as being overdefined. 1687 for (Argument &AI : F.args()) 1688 Solver.markOverdefined(&AI); 1689 1690 // Solve for constants. 1691 bool ResolvedUndefs = true; 1692 while (ResolvedUndefs) { 1693 Solver.Solve(); 1694 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1695 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1696 } 1697 1698 bool MadeChanges = false; 1699 1700 // If we decided that there are basic blocks that are dead in this function, 1701 // delete their contents now. Note that we cannot actually delete the blocks, 1702 // as we cannot modify the CFG of the function. 1703 1704 SmallPtrSet<Value *, 32> InsertedValues; 1705 for (BasicBlock &BB : F) { 1706 if (!Solver.isBlockExecutable(&BB)) { 1707 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1708 1709 ++NumDeadBlocks; 1710 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB); 1711 1712 MadeChanges = true; 1713 continue; 1714 } 1715 1716 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1717 NumInstRemoved, NumInstReplaced); 1718 } 1719 1720 return MadeChanges; 1721 } 1722 1723 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1724 const DataLayout &DL = F.getParent()->getDataLayout(); 1725 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1726 if (!runSCCP(F, DL, &TLI)) 1727 return PreservedAnalyses::all(); 1728 1729 auto PA = PreservedAnalyses(); 1730 PA.preserve<GlobalsAA>(); 1731 PA.preserveSet<CFGAnalyses>(); 1732 return PA; 1733 } 1734 1735 namespace { 1736 1737 //===--------------------------------------------------------------------===// 1738 // 1739 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1740 /// Sparse Conditional Constant Propagator. 1741 /// 1742 class SCCPLegacyPass : public FunctionPass { 1743 public: 1744 // Pass identification, replacement for typeid 1745 static char ID; 1746 1747 SCCPLegacyPass() : FunctionPass(ID) { 1748 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1749 } 1750 1751 void getAnalysisUsage(AnalysisUsage &AU) const override { 1752 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1753 AU.addPreserved<GlobalsAAWrapperPass>(); 1754 AU.setPreservesCFG(); 1755 } 1756 1757 // runOnFunction - Run the Sparse Conditional Constant Propagation 1758 // algorithm, and return true if the function was modified. 1759 bool runOnFunction(Function &F) override { 1760 if (skipFunction(F)) 1761 return false; 1762 const DataLayout &DL = F.getParent()->getDataLayout(); 1763 const TargetLibraryInfo *TLI = 1764 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1765 return runSCCP(F, DL, TLI); 1766 } 1767 }; 1768 1769 } // end anonymous namespace 1770 1771 char SCCPLegacyPass::ID = 0; 1772 1773 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1774 "Sparse Conditional Constant Propagation", false, false) 1775 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1776 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1777 "Sparse Conditional Constant Propagation", false, false) 1778 1779 // createSCCPPass - This is the public interface to this file. 1780 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1781 1782 static void findReturnsToZap(Function &F, 1783 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1784 SCCPSolver &Solver) { 1785 // We can only do this if we know that nothing else can call the function. 1786 if (!Solver.isArgumentTrackedFunction(&F)) 1787 return; 1788 1789 // There is a non-removable musttail call site of this function. Zapping 1790 // returns is not allowed. 1791 if (Solver.isMustTailCallee(&F)) { 1792 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1793 << " due to present musttail call of it\n"); 1794 return; 1795 } 1796 1797 assert( 1798 all_of(F.users(), 1799 [&Solver](User *U) { 1800 if (isa<Instruction>(U) && 1801 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1802 return true; 1803 // Non-callsite uses are not impacted by zapping. Also, constant 1804 // uses (like blockaddresses) could stuck around, without being 1805 // used in the underlying IR, meaning we do not have lattice 1806 // values for them. 1807 if (!isa<CallBase>(U)) 1808 return true; 1809 if (U->getType()->isStructTy()) { 1810 return all_of(Solver.getStructLatticeValueFor(U), 1811 [](const ValueLatticeElement &LV) { 1812 return !isOverdefined(LV); 1813 }); 1814 } 1815 return !isOverdefined(Solver.getLatticeValueFor(U)); 1816 }) && 1817 "We can only zap functions where all live users have a concrete value"); 1818 1819 for (BasicBlock &BB : F) { 1820 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1821 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1822 << "musttail call : " << *CI << "\n"); 1823 (void)CI; 1824 return; 1825 } 1826 1827 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1828 if (!isa<UndefValue>(RI->getOperand(0))) 1829 ReturnsToZap.push_back(RI); 1830 } 1831 } 1832 1833 static bool removeNonFeasibleEdges(const SCCPSolver &Solver, BasicBlock *BB, 1834 DomTreeUpdater &DTU) { 1835 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors; 1836 bool HasNonFeasibleEdges = false; 1837 for (BasicBlock *Succ : successors(BB)) { 1838 if (Solver.isEdgeFeasible(BB, Succ)) 1839 FeasibleSuccessors.insert(Succ); 1840 else 1841 HasNonFeasibleEdges = true; 1842 } 1843 1844 // All edges feasible, nothing to do. 1845 if (!HasNonFeasibleEdges) 1846 return false; 1847 1848 // SCCP can only determine non-feasible edges for br, switch and indirectbr. 1849 Instruction *TI = BB->getTerminator(); 1850 assert((isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1851 isa<IndirectBrInst>(TI)) && 1852 "Terminator must be a br, switch or indirectbr"); 1853 1854 if (FeasibleSuccessors.size() == 1) { 1855 // Replace with an unconditional branch to the only feasible successor. 1856 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin(); 1857 SmallVector<DominatorTree::UpdateType, 8> Updates; 1858 bool HaveSeenOnlyFeasibleSuccessor = false; 1859 for (BasicBlock *Succ : successors(BB)) { 1860 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) { 1861 // Don't remove the edge to the only feasible successor the first time 1862 // we see it. We still do need to remove any multi-edges to it though. 1863 HaveSeenOnlyFeasibleSuccessor = true; 1864 continue; 1865 } 1866 1867 Succ->removePredecessor(BB); 1868 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1869 } 1870 1871 BranchInst::Create(OnlyFeasibleSuccessor, BB); 1872 TI->eraseFromParent(); 1873 DTU.applyUpdatesPermissive(Updates); 1874 } else if (FeasibleSuccessors.size() > 1) { 1875 SwitchInstProfUpdateWrapper SI(*cast<SwitchInst>(TI)); 1876 SmallVector<DominatorTree::UpdateType, 8> Updates; 1877 for (auto CI = SI->case_begin(); CI != SI->case_end();) { 1878 if (FeasibleSuccessors.contains(CI->getCaseSuccessor())) { 1879 ++CI; 1880 continue; 1881 } 1882 1883 BasicBlock *Succ = CI->getCaseSuccessor(); 1884 Succ->removePredecessor(BB); 1885 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1886 SI.removeCase(CI); 1887 // Don't increment CI, as we removed a case. 1888 } 1889 1890 DTU.applyUpdatesPermissive(Updates); 1891 } else { 1892 llvm_unreachable("Must have at least one feasible successor"); 1893 } 1894 return true; 1895 } 1896 1897 bool llvm::runIPSCCP( 1898 Module &M, const DataLayout &DL, 1899 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1900 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1901 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1902 1903 // Loop over all functions, marking arguments to those with their addresses 1904 // taken or that are external as overdefined. 1905 for (Function &F : M) { 1906 if (F.isDeclaration()) 1907 continue; 1908 1909 Solver.addAnalysis(F, getAnalysis(F)); 1910 1911 // Determine if we can track the function's return values. If so, add the 1912 // function to the solver's set of return-tracked functions. 1913 if (canTrackReturnsInterprocedurally(&F)) 1914 Solver.AddTrackedFunction(&F); 1915 1916 // Determine if we can track the function's arguments. If so, add the 1917 // function to the solver's set of argument-tracked functions. 1918 if (canTrackArgumentsInterprocedurally(&F)) { 1919 Solver.AddArgumentTrackedFunction(&F); 1920 continue; 1921 } 1922 1923 // Assume the function is called. 1924 Solver.MarkBlockExecutable(&F.front()); 1925 1926 // Assume nothing about the incoming arguments. 1927 for (Argument &AI : F.args()) 1928 Solver.markOverdefined(&AI); 1929 } 1930 1931 // Determine if we can track any of the module's global variables. If so, add 1932 // the global variables we can track to the solver's set of tracked global 1933 // variables. 1934 for (GlobalVariable &G : M.globals()) { 1935 G.removeDeadConstantUsers(); 1936 if (canTrackGlobalVariableInterprocedurally(&G)) 1937 Solver.TrackValueOfGlobalVariable(&G); 1938 } 1939 1940 // Solve for constants. 1941 bool ResolvedUndefs = true; 1942 Solver.Solve(); 1943 while (ResolvedUndefs) { 1944 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1945 ResolvedUndefs = false; 1946 for (Function &F : M) 1947 if (Solver.ResolvedUndefsIn(F)) { 1948 // We run Solve() after we resolved an undef in a function, because 1949 // we might deduce a fact that eliminates an undef in another function. 1950 Solver.Solve(); 1951 ResolvedUndefs = true; 1952 } 1953 } 1954 1955 bool MadeChanges = false; 1956 1957 // Iterate over all of the instructions in the module, replacing them with 1958 // constants if we have found them to be of constant values. 1959 1960 for (Function &F : M) { 1961 if (F.isDeclaration()) 1962 continue; 1963 1964 SmallVector<BasicBlock *, 512> BlocksToErase; 1965 1966 if (Solver.isBlockExecutable(&F.front())) { 1967 bool ReplacedPointerArg = false; 1968 for (Argument &Arg : F.args()) { 1969 if (!Arg.use_empty() && tryToReplaceWithConstant(Solver, &Arg)) { 1970 ReplacedPointerArg |= Arg.getType()->isPointerTy(); 1971 ++IPNumArgsElimed; 1972 } 1973 } 1974 1975 // If we replaced an argument, the argmemonly and 1976 // inaccessiblemem_or_argmemonly attributes do not hold any longer. Remove 1977 // them from both the function and callsites. 1978 if (ReplacedPointerArg) { 1979 SmallVector<Attribute::AttrKind, 2> AttributesToRemove = { 1980 Attribute::ArgMemOnly, Attribute::InaccessibleMemOrArgMemOnly}; 1981 for (auto Attr : AttributesToRemove) 1982 F.removeFnAttr(Attr); 1983 1984 for (User *U : F.users()) { 1985 auto *CB = dyn_cast<CallBase>(U); 1986 if (!CB || CB->getCalledFunction() != &F) 1987 continue; 1988 1989 for (auto Attr : AttributesToRemove) 1990 CB->removeAttribute(AttributeList::FunctionIndex, Attr); 1991 } 1992 } 1993 } 1994 1995 SmallPtrSet<Value *, 32> InsertedValues; 1996 for (BasicBlock &BB : F) { 1997 if (!Solver.isBlockExecutable(&BB)) { 1998 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1999 ++NumDeadBlocks; 2000 2001 MadeChanges = true; 2002 2003 if (&BB != &F.front()) 2004 BlocksToErase.push_back(&BB); 2005 continue; 2006 } 2007 2008 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 2009 IPNumInstRemoved, IPNumInstReplaced); 2010 } 2011 2012 DomTreeUpdater DTU = Solver.getDTU(F); 2013 // Change dead blocks to unreachable. We do it after replacing constants 2014 // in all executable blocks, because changeToUnreachable may remove PHI 2015 // nodes in executable blocks we found values for. The function's entry 2016 // block is not part of BlocksToErase, so we have to handle it separately. 2017 for (BasicBlock *BB : BlocksToErase) { 2018 NumInstRemoved += 2019 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 2020 /*PreserveLCSSA=*/false, &DTU); 2021 } 2022 if (!Solver.isBlockExecutable(&F.front())) 2023 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 2024 /*UseLLVMTrap=*/false, 2025 /*PreserveLCSSA=*/false, &DTU); 2026 2027 for (BasicBlock &BB : F) 2028 MadeChanges |= removeNonFeasibleEdges(Solver, &BB, DTU); 2029 2030 for (BasicBlock *DeadBB : BlocksToErase) 2031 DTU.deleteBB(DeadBB); 2032 2033 for (BasicBlock &BB : F) { 2034 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 2035 Instruction *Inst = &*BI++; 2036 if (Solver.getPredicateInfoFor(Inst)) { 2037 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2038 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2039 Value *Op = II->getOperand(0); 2040 Inst->replaceAllUsesWith(Op); 2041 Inst->eraseFromParent(); 2042 } 2043 } 2044 } 2045 } 2046 } 2047 } 2048 2049 // If we inferred constant or undef return values for a function, we replaced 2050 // all call uses with the inferred value. This means we don't need to bother 2051 // actually returning anything from the function. Replace all return 2052 // instructions with return undef. 2053 // 2054 // Do this in two stages: first identify the functions we should process, then 2055 // actually zap their returns. This is important because we can only do this 2056 // if the address of the function isn't taken. In cases where a return is the 2057 // last use of a function, the order of processing functions would affect 2058 // whether other functions are optimizable. 2059 SmallVector<ReturnInst*, 8> ReturnsToZap; 2060 2061 for (const auto &I : Solver.getTrackedRetVals()) { 2062 Function *F = I.first; 2063 const ValueLatticeElement &ReturnValue = I.second; 2064 2065 // If there is a known constant range for the return value, add !range 2066 // metadata to the function's call sites. 2067 if (ReturnValue.isConstantRange() && 2068 !ReturnValue.getConstantRange().isSingleElement()) { 2069 // Do not add range metadata if the return value may include undef. 2070 if (ReturnValue.isConstantRangeIncludingUndef()) 2071 continue; 2072 2073 auto &CR = ReturnValue.getConstantRange(); 2074 for (User *User : F->users()) { 2075 auto *CB = dyn_cast<CallBase>(User); 2076 if (!CB || CB->getCalledFunction() != F) 2077 continue; 2078 2079 // Limit to cases where the return value is guaranteed to be neither 2080 // poison nor undef. Poison will be outside any range and currently 2081 // values outside of the specified range cause immediate undefined 2082 // behavior. 2083 if (!isGuaranteedNotToBeUndefOrPoison(CB, CB)) 2084 continue; 2085 2086 // Do not touch existing metadata for now. 2087 // TODO: We should be able to take the intersection of the existing 2088 // metadata and the inferred range. 2089 if (CB->getMetadata(LLVMContext::MD_range)) 2090 continue; 2091 2092 LLVMContext &Context = CB->getParent()->getContext(); 2093 Metadata *RangeMD[] = { 2094 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getLower())), 2095 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getUpper()))}; 2096 CB->setMetadata(LLVMContext::MD_range, MDNode::get(Context, RangeMD)); 2097 } 2098 continue; 2099 } 2100 if (F->getReturnType()->isVoidTy()) 2101 continue; 2102 if (isConstant(ReturnValue) || ReturnValue.isUnknownOrUndef()) 2103 findReturnsToZap(*F, ReturnsToZap, Solver); 2104 } 2105 2106 for (auto F : Solver.getMRVFunctionsTracked()) { 2107 assert(F->getReturnType()->isStructTy() && 2108 "The return type should be a struct"); 2109 StructType *STy = cast<StructType>(F->getReturnType()); 2110 if (Solver.isStructLatticeConstant(F, STy)) 2111 findReturnsToZap(*F, ReturnsToZap, Solver); 2112 } 2113 2114 // Zap all returns which we've identified as zap to change. 2115 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2116 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2117 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2118 } 2119 2120 // If we inferred constant or undef values for globals variables, we can 2121 // delete the global and any stores that remain to it. 2122 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2123 GlobalVariable *GV = I.first; 2124 if (isOverdefined(I.second)) 2125 continue; 2126 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2127 << "' is constant!\n"); 2128 while (!GV->use_empty()) { 2129 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2130 SI->eraseFromParent(); 2131 MadeChanges = true; 2132 } 2133 M.getGlobalList().erase(GV); 2134 ++IPNumGlobalConst; 2135 } 2136 2137 return MadeChanges; 2138 } 2139