1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements sparse conditional constant propagation and merging: 10 // 11 // Specifically, this: 12 // * Assumes values are constant unless proven otherwise 13 // * Assumes BasicBlocks are dead unless proven otherwise 14 // * Proves values to be constant, and replaces them with constants 15 // * Proves conditional branches to be unconditional 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/Transforms/Scalar/SCCP.h" 20 #include "llvm/ADT/ArrayRef.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/DenseSet.h" 23 #include "llvm/ADT/MapVector.h" 24 #include "llvm/ADT/PointerIntPair.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SmallPtrSet.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/Analysis/ConstantFolding.h" 30 #include "llvm/Analysis/GlobalsModRef.h" 31 #include "llvm/Analysis/InstructionSimplify.h" 32 #include "llvm/Analysis/TargetLibraryInfo.h" 33 #include "llvm/Analysis/ValueLattice.h" 34 #include "llvm/Analysis/ValueLatticeUtils.h" 35 #include "llvm/IR/BasicBlock.h" 36 #include "llvm/IR/Constant.h" 37 #include "llvm/IR/Constants.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DerivedTypes.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/IR/GlobalVariable.h" 42 #include "llvm/IR/InstVisitor.h" 43 #include "llvm/IR/InstrTypes.h" 44 #include "llvm/IR/Instruction.h" 45 #include "llvm/IR/Instructions.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/PassManager.h" 48 #include "llvm/IR/Type.h" 49 #include "llvm/IR/User.h" 50 #include "llvm/IR/Value.h" 51 #include "llvm/InitializePasses.h" 52 #include "llvm/Pass.h" 53 #include "llvm/Support/Casting.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/ErrorHandling.h" 56 #include "llvm/Support/raw_ostream.h" 57 #include "llvm/Transforms/Scalar.h" 58 #include "llvm/Transforms/Utils/Local.h" 59 #include "llvm/Transforms/Utils/PredicateInfo.h" 60 #include <cassert> 61 #include <utility> 62 #include <vector> 63 64 using namespace llvm; 65 66 #define DEBUG_TYPE "sccp" 67 68 STATISTIC(NumInstRemoved, "Number of instructions removed"); 69 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 70 STATISTIC(NumInstReplaced, 71 "Number of instructions replaced with (simpler) instruction"); 72 73 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 74 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 75 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 76 STATISTIC( 77 IPNumInstReplaced, 78 "Number of instructions replaced with (simpler) instruction by IPSCCP"); 79 80 // The maximum number of range extensions allowed for operations requiring 81 // widening. 82 static const unsigned MaxNumRangeExtensions = 10; 83 84 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 85 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 86 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 87 MaxNumRangeExtensions); 88 } 89 namespace { 90 91 // Helper to check if \p LV is either a constant or a constant 92 // range with a single element. This should cover exactly the same cases as the 93 // old ValueLatticeElement::isConstant() and is intended to be used in the 94 // transition to ValueLatticeElement. 95 bool isConstant(const ValueLatticeElement &LV) { 96 return LV.isConstant() || 97 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 98 } 99 100 // Helper to check if \p LV is either overdefined or a constant range with more 101 // than a single element. This should cover exactly the same cases as the old 102 // ValueLatticeElement::isOverdefined() and is intended to be used in the 103 // transition to ValueLatticeElement. 104 bool isOverdefined(const ValueLatticeElement &LV) { 105 return LV.isOverdefined() || 106 (LV.isConstantRange() && !LV.getConstantRange().isSingleElement()); 107 } 108 109 //===----------------------------------------------------------------------===// 110 // 111 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 112 /// Constant Propagation. 113 /// 114 class SCCPSolver : public InstVisitor<SCCPSolver> { 115 const DataLayout &DL; 116 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 117 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 118 DenseMap<Value *, ValueLatticeElement> 119 ValueState; // The state each value is in. 120 121 /// StructValueState - This maintains ValueState for values that have 122 /// StructType, for example for formal arguments, calls, insertelement, etc. 123 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 124 125 /// GlobalValue - If we are tracking any values for the contents of a global 126 /// variable, we keep a mapping from the constant accessor to the element of 127 /// the global, to the currently known value. If the value becomes 128 /// overdefined, it's entry is simply removed from this map. 129 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 130 131 /// TrackedRetVals - If we are tracking arguments into and the return 132 /// value out of a function, it will have an entry in this map, indicating 133 /// what the known return value for the function is. 134 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 135 136 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 137 /// that return multiple values. 138 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 139 TrackedMultipleRetVals; 140 141 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 142 /// represented here for efficient lookup. 143 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 144 145 /// MustTailFunctions - Each function here is a callee of non-removable 146 /// musttail call site. 147 SmallPtrSet<Function *, 16> MustTailCallees; 148 149 /// TrackingIncomingArguments - This is the set of functions for whose 150 /// arguments we make optimistic assumptions about and try to prove as 151 /// constants. 152 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 153 154 /// The reason for two worklists is that overdefined is the lowest state 155 /// on the lattice, and moving things to overdefined as fast as possible 156 /// makes SCCP converge much faster. 157 /// 158 /// By having a separate worklist, we accomplish this because everything 159 /// possibly overdefined will become overdefined at the soonest possible 160 /// point. 161 SmallVector<Value *, 64> OverdefinedInstWorkList; 162 SmallVector<Value *, 64> InstWorkList; 163 164 // The BasicBlock work list 165 SmallVector<BasicBlock *, 64> BBWorkList; 166 167 /// KnownFeasibleEdges - Entries in this set are edges which have already had 168 /// PHI nodes retriggered. 169 using Edge = std::pair<BasicBlock *, BasicBlock *>; 170 DenseSet<Edge> KnownFeasibleEdges; 171 172 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 173 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 174 175 LLVMContext &Ctx; 176 177 public: 178 void addAnalysis(Function &F, AnalysisResultsForFn A) { 179 AnalysisResults.insert({&F, std::move(A)}); 180 } 181 182 const PredicateBase *getPredicateInfoFor(Instruction *I) { 183 auto A = AnalysisResults.find(I->getParent()->getParent()); 184 if (A == AnalysisResults.end()) 185 return nullptr; 186 return A->second.PredInfo->getPredicateInfoFor(I); 187 } 188 189 DomTreeUpdater getDTU(Function &F) { 190 auto A = AnalysisResults.find(&F); 191 assert(A != AnalysisResults.end() && "Need analysis results for function."); 192 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 193 } 194 195 SCCPSolver(const DataLayout &DL, 196 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 197 LLVMContext &Ctx) 198 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 199 200 /// MarkBlockExecutable - This method can be used by clients to mark all of 201 /// the blocks that are known to be intrinsically live in the processed unit. 202 /// 203 /// This returns true if the block was not considered live before. 204 bool MarkBlockExecutable(BasicBlock *BB) { 205 if (!BBExecutable.insert(BB).second) 206 return false; 207 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 208 BBWorkList.push_back(BB); // Add the block to the work list! 209 return true; 210 } 211 212 /// TrackValueOfGlobalVariable - Clients can use this method to 213 /// inform the SCCPSolver that it should track loads and stores to the 214 /// specified global variable if it can. This is only legal to call if 215 /// performing Interprocedural SCCP. 216 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 217 // We only track the contents of scalar globals. 218 if (GV->getValueType()->isSingleValueType()) { 219 ValueLatticeElement &IV = TrackedGlobals[GV]; 220 if (!isa<UndefValue>(GV->getInitializer())) 221 IV.markConstant(GV->getInitializer()); 222 } 223 } 224 225 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 226 /// and out of the specified function (which cannot have its address taken), 227 /// this method must be called. 228 void AddTrackedFunction(Function *F) { 229 // Add an entry, F -> undef. 230 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 231 MRVFunctionsTracked.insert(F); 232 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 233 TrackedMultipleRetVals.insert( 234 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 235 } else 236 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 237 } 238 239 /// AddMustTailCallee - If the SCCP solver finds that this function is called 240 /// from non-removable musttail call site. 241 void AddMustTailCallee(Function *F) { 242 MustTailCallees.insert(F); 243 } 244 245 /// Returns true if the given function is called from non-removable musttail 246 /// call site. 247 bool isMustTailCallee(Function *F) { 248 return MustTailCallees.count(F); 249 } 250 251 void AddArgumentTrackedFunction(Function *F) { 252 TrackingIncomingArguments.insert(F); 253 } 254 255 /// Returns true if the given function is in the solver's set of 256 /// argument-tracked functions. 257 bool isArgumentTrackedFunction(Function *F) { 258 return TrackingIncomingArguments.count(F); 259 } 260 261 /// Solve - Solve for constants and executable blocks. 262 void Solve(); 263 264 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 265 /// that branches on undef values cannot reach any of their successors. 266 /// However, this is not a safe assumption. After we solve dataflow, this 267 /// method should be use to handle this. If this returns true, the solver 268 /// should be rerun. 269 bool ResolvedUndefsIn(Function &F); 270 271 bool isBlockExecutable(BasicBlock *BB) const { 272 return BBExecutable.count(BB); 273 } 274 275 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 276 // block to the 'To' basic block is currently feasible. 277 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To); 278 279 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 280 std::vector<ValueLatticeElement> StructValues; 281 auto *STy = dyn_cast<StructType>(V->getType()); 282 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 283 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 284 auto I = StructValueState.find(std::make_pair(V, i)); 285 assert(I != StructValueState.end() && "Value not in valuemap!"); 286 StructValues.push_back(I->second); 287 } 288 return StructValues; 289 } 290 291 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 292 293 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 294 assert(!V->getType()->isStructTy() && 295 "Should use getStructLatticeValueFor"); 296 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 297 ValueState.find(V); 298 assert(I != ValueState.end() && 299 "V not found in ValueState nor Paramstate map!"); 300 return I->second; 301 } 302 303 /// getTrackedRetVals - Get the inferred return value map. 304 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 305 return TrackedRetVals; 306 } 307 308 /// getTrackedGlobals - Get and return the set of inferred initializers for 309 /// global variables. 310 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 311 return TrackedGlobals; 312 } 313 314 /// getMRVFunctionsTracked - Get the set of functions which return multiple 315 /// values tracked by the pass. 316 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 317 return MRVFunctionsTracked; 318 } 319 320 /// getMustTailCallees - Get the set of functions which are called 321 /// from non-removable musttail call sites. 322 const SmallPtrSet<Function *, 16> getMustTailCallees() { 323 return MustTailCallees; 324 } 325 326 /// markOverdefined - Mark the specified value overdefined. This 327 /// works with both scalars and structs. 328 void markOverdefined(Value *V) { 329 if (auto *STy = dyn_cast<StructType>(V->getType())) 330 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 331 markOverdefined(getStructValueState(V, i), V); 332 else 333 markOverdefined(ValueState[V], V); 334 } 335 336 // isStructLatticeConstant - Return true if all the lattice values 337 // corresponding to elements of the structure are constants, 338 // false otherwise. 339 bool isStructLatticeConstant(Function *F, StructType *STy) { 340 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 341 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 342 assert(It != TrackedMultipleRetVals.end()); 343 ValueLatticeElement LV = It->second; 344 if (!isConstant(LV)) 345 return false; 346 } 347 return true; 348 } 349 350 /// Helper to return a Constant if \p LV is either a constant or a constant 351 /// range with a single element. 352 Constant *getConstant(const ValueLatticeElement &LV) const { 353 if (LV.isConstant()) 354 return LV.getConstant(); 355 356 if (LV.isConstantRange()) { 357 auto &CR = LV.getConstantRange(); 358 if (CR.getSingleElement()) 359 return ConstantInt::get(Ctx, *CR.getSingleElement()); 360 } 361 return nullptr; 362 } 363 364 private: 365 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 366 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 367 } 368 369 // pushToWorkList - Helper for markConstant/markOverdefined 370 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 371 if (IV.isOverdefined()) 372 return OverdefinedInstWorkList.push_back(V); 373 InstWorkList.push_back(V); 374 } 375 376 // Helper to push \p V to the worklist, after updating it to \p IV. Also 377 // prints a debug message with the updated value. 378 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 379 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 380 pushToWorkList(IV, V); 381 } 382 383 // markConstant - Make a value be marked as "constant". If the value 384 // is not already a constant, add it to the instruction work list so that 385 // the users of the instruction are updated later. 386 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 387 bool MayIncludeUndef = false) { 388 if (!IV.markConstant(C, MayIncludeUndef)) 389 return false; 390 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 391 pushToWorkList(IV, V); 392 return true; 393 } 394 395 bool markConstant(Value *V, Constant *C) { 396 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 397 return markConstant(ValueState[V], V, C); 398 } 399 400 // markOverdefined - Make a value be marked as "overdefined". If the 401 // value is not already overdefined, add it to the overdefined instruction 402 // work list so that the users of the instruction are updated later. 403 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 404 if (!IV.markOverdefined()) return false; 405 406 LLVM_DEBUG(dbgs() << "markOverdefined: "; 407 if (auto *F = dyn_cast<Function>(V)) dbgs() 408 << "Function '" << F->getName() << "'\n"; 409 else dbgs() << *V << '\n'); 410 // Only instructions go on the work list 411 pushToWorkList(IV, V); 412 return true; 413 } 414 415 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 416 /// changes. 417 bool mergeInValue(ValueLatticeElement &IV, Value *V, 418 ValueLatticeElement MergeWithV, 419 ValueLatticeElement::MergeOptions Opts = { 420 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 421 if (IV.mergeIn(MergeWithV, Opts)) { 422 pushToWorkList(IV, V); 423 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 424 << IV << "\n"); 425 return true; 426 } 427 return false; 428 } 429 430 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 431 ValueLatticeElement::MergeOptions Opts = { 432 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 433 assert(!V->getType()->isStructTy() && 434 "non-structs should use markConstant"); 435 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 436 } 437 438 /// getValueState - Return the ValueLatticeElement object that corresponds to 439 /// the value. This function handles the case when the value hasn't been seen 440 /// yet by properly seeding constants etc. 441 ValueLatticeElement &getValueState(Value *V) { 442 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 443 444 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 445 ValueLatticeElement &LV = I.first->second; 446 447 if (!I.second) 448 return LV; // Common case, already in the map. 449 450 if (auto *C = dyn_cast<Constant>(V)) 451 LV.markConstant(C); // Constants are constant 452 453 // All others are unknown by default. 454 return LV; 455 } 456 457 /// getStructValueState - Return the ValueLatticeElement object that 458 /// corresponds to the value/field pair. This function handles the case when 459 /// the value hasn't been seen yet by properly seeding constants etc. 460 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 461 assert(V->getType()->isStructTy() && "Should use getValueState"); 462 assert(i < cast<StructType>(V->getType())->getNumElements() && 463 "Invalid element #"); 464 465 auto I = StructValueState.insert( 466 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 467 ValueLatticeElement &LV = I.first->second; 468 469 if (!I.second) 470 return LV; // Common case, already in the map. 471 472 if (auto *C = dyn_cast<Constant>(V)) { 473 Constant *Elt = C->getAggregateElement(i); 474 475 if (!Elt) 476 LV.markOverdefined(); // Unknown sort of constant. 477 else if (isa<UndefValue>(Elt)) 478 ; // Undef values remain unknown. 479 else 480 LV.markConstant(Elt); // Constants are constant. 481 } 482 483 // All others are underdefined by default. 484 return LV; 485 } 486 487 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 488 /// work list if it is not already executable. 489 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 490 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 491 return false; // This edge is already known to be executable! 492 493 if (!MarkBlockExecutable(Dest)) { 494 // If the destination is already executable, we just made an *edge* 495 // feasible that wasn't before. Revisit the PHI nodes in the block 496 // because they have potentially new operands. 497 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 498 << " -> " << Dest->getName() << '\n'); 499 500 for (PHINode &PN : Dest->phis()) 501 visitPHINode(PN); 502 } 503 return true; 504 } 505 506 // getFeasibleSuccessors - Return a vector of booleans to indicate which 507 // successors are reachable from a given terminator instruction. 508 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 509 510 // OperandChangedState - This method is invoked on all of the users of an 511 // instruction that was just changed state somehow. Based on this 512 // information, we need to update the specified user of this instruction. 513 void OperandChangedState(Instruction *I) { 514 if (BBExecutable.count(I->getParent())) // Inst is executable? 515 visit(*I); 516 } 517 518 // Add U as additional user of V. 519 void addAdditionalUser(Value *V, User *U) { 520 auto Iter = AdditionalUsers.insert({V, {}}); 521 Iter.first->second.insert(U); 522 } 523 524 // Mark I's users as changed, including AdditionalUsers. 525 void markUsersAsChanged(Value *I) { 526 // Functions include their arguments in the use-list. Changed function 527 // values mean that the result of the function changed. We only need to 528 // update the call sites with the new function result and do not have to 529 // propagate the call arguments. 530 if (isa<Function>(I)) { 531 for (User *U : I->users()) { 532 if (auto *CB = dyn_cast<CallBase>(U)) 533 handleCallResult(*CB); 534 } 535 } else { 536 for (User *U : I->users()) 537 if (auto *UI = dyn_cast<Instruction>(U)) 538 OperandChangedState(UI); 539 } 540 541 auto Iter = AdditionalUsers.find(I); 542 if (Iter != AdditionalUsers.end()) { 543 for (User *U : Iter->second) 544 if (auto *UI = dyn_cast<Instruction>(U)) 545 OperandChangedState(UI); 546 } 547 } 548 void handleCallOverdefined(CallBase &CB); 549 void handleCallResult(CallBase &CB); 550 void handleCallArguments(CallBase &CB); 551 552 private: 553 friend class InstVisitor<SCCPSolver>; 554 555 // visit implementations - Something changed in this instruction. Either an 556 // operand made a transition, or the instruction is newly executable. Change 557 // the value type of I to reflect these changes if appropriate. 558 void visitPHINode(PHINode &I); 559 560 // Terminators 561 562 void visitReturnInst(ReturnInst &I); 563 void visitTerminator(Instruction &TI); 564 565 void visitCastInst(CastInst &I); 566 void visitSelectInst(SelectInst &I); 567 void visitUnaryOperator(Instruction &I); 568 void visitBinaryOperator(Instruction &I); 569 void visitCmpInst(CmpInst &I); 570 void visitExtractValueInst(ExtractValueInst &EVI); 571 void visitInsertValueInst(InsertValueInst &IVI); 572 573 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 574 markOverdefined(&CPI); 575 visitTerminator(CPI); 576 } 577 578 // Instructions that cannot be folded away. 579 580 void visitStoreInst (StoreInst &I); 581 void visitLoadInst (LoadInst &I); 582 void visitGetElementPtrInst(GetElementPtrInst &I); 583 584 void visitCallInst (CallInst &I) { 585 visitCallBase(I); 586 } 587 588 void visitInvokeInst (InvokeInst &II) { 589 visitCallBase(II); 590 visitTerminator(II); 591 } 592 593 void visitCallBrInst (CallBrInst &CBI) { 594 visitCallBase(CBI); 595 visitTerminator(CBI); 596 } 597 598 void visitCallBase (CallBase &CB); 599 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 600 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 601 void visitFenceInst (FenceInst &I) { /*returns void*/ } 602 603 void visitInstruction(Instruction &I) { 604 // All the instructions we don't do any special handling for just 605 // go to overdefined. 606 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 607 markOverdefined(&I); 608 } 609 }; 610 611 } // end anonymous namespace 612 613 // getFeasibleSuccessors - Return a vector of booleans to indicate which 614 // successors are reachable from a given terminator instruction. 615 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 616 SmallVectorImpl<bool> &Succs) { 617 Succs.resize(TI.getNumSuccessors()); 618 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 619 if (BI->isUnconditional()) { 620 Succs[0] = true; 621 return; 622 } 623 624 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 625 ConstantInt *CI = getConstantInt(BCValue); 626 if (!CI) { 627 // Overdefined condition variables, and branches on unfoldable constant 628 // conditions, mean the branch could go either way. 629 if (!BCValue.isUnknownOrUndef()) 630 Succs[0] = Succs[1] = true; 631 return; 632 } 633 634 // Constant condition variables mean the branch can only go a single way. 635 Succs[CI->isZero()] = true; 636 return; 637 } 638 639 // Unwinding instructions successors are always executable. 640 if (TI.isExceptionalTerminator()) { 641 Succs.assign(TI.getNumSuccessors(), true); 642 return; 643 } 644 645 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 646 if (!SI->getNumCases()) { 647 Succs[0] = true; 648 return; 649 } 650 ValueLatticeElement SCValue = getValueState(SI->getCondition()); 651 ConstantInt *CI = getConstantInt(SCValue); 652 653 if (!CI) { // Overdefined or unknown condition? 654 // All destinations are executable! 655 if (!SCValue.isUnknownOrUndef()) 656 Succs.assign(TI.getNumSuccessors(), true); 657 return; 658 } 659 660 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 661 return; 662 } 663 664 // In case of indirect branch and its address is a blockaddress, we mark 665 // the target as executable. 666 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 667 // Casts are folded by visitCastInst. 668 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 669 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 670 if (!Addr) { // Overdefined or unknown condition? 671 // All destinations are executable! 672 if (!IBRValue.isUnknownOrUndef()) 673 Succs.assign(TI.getNumSuccessors(), true); 674 return; 675 } 676 677 BasicBlock* T = Addr->getBasicBlock(); 678 assert(Addr->getFunction() == T->getParent() && 679 "Block address of a different function ?"); 680 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 681 // This is the target. 682 if (IBR->getDestination(i) == T) { 683 Succs[i] = true; 684 return; 685 } 686 } 687 688 // If we didn't find our destination in the IBR successor list, then we 689 // have undefined behavior. Its ok to assume no successor is executable. 690 return; 691 } 692 693 // In case of callbr, we pessimistically assume that all successors are 694 // feasible. 695 if (isa<CallBrInst>(&TI)) { 696 Succs.assign(TI.getNumSuccessors(), true); 697 return; 698 } 699 700 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 701 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 702 } 703 704 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 705 // block to the 'To' basic block is currently feasible. 706 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) { 707 // Check if we've called markEdgeExecutable on the edge yet. (We could 708 // be more aggressive and try to consider edges which haven't been marked 709 // yet, but there isn't any need.) 710 return KnownFeasibleEdges.count(Edge(From, To)); 711 } 712 713 // visit Implementations - Something changed in this instruction, either an 714 // operand made a transition, or the instruction is newly executable. Change 715 // the value type of I to reflect these changes if appropriate. This method 716 // makes sure to do the following actions: 717 // 718 // 1. If a phi node merges two constants in, and has conflicting value coming 719 // from different branches, or if the PHI node merges in an overdefined 720 // value, then the PHI node becomes overdefined. 721 // 2. If a phi node merges only constants in, and they all agree on value, the 722 // PHI node becomes a constant value equal to that. 723 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 724 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 725 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 726 // 6. If a conditional branch has a value that is constant, make the selected 727 // destination executable 728 // 7. If a conditional branch has a value that is overdefined, make all 729 // successors executable. 730 void SCCPSolver::visitPHINode(PHINode &PN) { 731 // If this PN returns a struct, just mark the result overdefined. 732 // TODO: We could do a lot better than this if code actually uses this. 733 if (PN.getType()->isStructTy()) 734 return (void)markOverdefined(&PN); 735 736 if (getValueState(&PN).isOverdefined()) 737 return; // Quick exit 738 739 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 740 // and slow us down a lot. Just mark them overdefined. 741 if (PN.getNumIncomingValues() > 64) 742 return (void)markOverdefined(&PN); 743 744 unsigned NumActiveIncoming = 0; 745 746 // Look at all of the executable operands of the PHI node. If any of them 747 // are overdefined, the PHI becomes overdefined as well. If they are all 748 // constant, and they agree with each other, the PHI becomes the identical 749 // constant. If they are constant and don't agree, the PHI is a constant 750 // range. If there are no executable operands, the PHI remains unknown. 751 ValueLatticeElement PhiState = getValueState(&PN); 752 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 753 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 754 continue; 755 756 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 757 PhiState.mergeIn(IV); 758 NumActiveIncoming++; 759 if (PhiState.isOverdefined()) 760 break; 761 } 762 763 // We allow up to 1 range extension per active incoming value and one 764 // additional extension. Note that we manually adjust the number of range 765 // extensions to match the number of active incoming values. This helps to 766 // limit multiple extensions caused by the same incoming value, if other 767 // incoming values are equal. 768 mergeInValue(&PN, PhiState, 769 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 770 NumActiveIncoming + 1)); 771 ValueLatticeElement &PhiStateRef = getValueState(&PN); 772 PhiStateRef.setNumRangeExtensions( 773 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 774 } 775 776 void SCCPSolver::visitReturnInst(ReturnInst &I) { 777 if (I.getNumOperands() == 0) return; // ret void 778 779 Function *F = I.getParent()->getParent(); 780 Value *ResultOp = I.getOperand(0); 781 782 // If we are tracking the return value of this function, merge it in. 783 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 784 auto TFRVI = TrackedRetVals.find(F); 785 if (TFRVI != TrackedRetVals.end()) { 786 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 787 return; 788 } 789 } 790 791 // Handle functions that return multiple values. 792 if (!TrackedMultipleRetVals.empty()) { 793 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 794 if (MRVFunctionsTracked.count(F)) 795 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 796 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 797 getStructValueState(ResultOp, i)); 798 } 799 } 800 801 void SCCPSolver::visitTerminator(Instruction &TI) { 802 SmallVector<bool, 16> SuccFeasible; 803 getFeasibleSuccessors(TI, SuccFeasible); 804 805 BasicBlock *BB = TI.getParent(); 806 807 // Mark all feasible successors executable. 808 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 809 if (SuccFeasible[i]) 810 markEdgeExecutable(BB, TI.getSuccessor(i)); 811 } 812 813 void SCCPSolver::visitCastInst(CastInst &I) { 814 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 815 // discover a concrete value later. 816 if (ValueState[&I].isOverdefined()) 817 return; 818 819 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 820 if (Constant *OpC = getConstant(OpSt)) { 821 // Fold the constant as we build. 822 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 823 if (isa<UndefValue>(C)) 824 return; 825 // Propagate constant value 826 markConstant(&I, C); 827 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 828 auto &LV = getValueState(&I); 829 ConstantRange OpRange = OpSt.getConstantRange(); 830 Type *DestTy = I.getDestTy(); 831 ConstantRange Res = 832 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 833 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 834 } else if (!OpSt.isUnknownOrUndef()) 835 markOverdefined(&I); 836 } 837 838 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 839 // If this returns a struct, mark all elements over defined, we don't track 840 // structs in structs. 841 if (EVI.getType()->isStructTy()) 842 return (void)markOverdefined(&EVI); 843 844 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 845 // discover a concrete value later. 846 if (ValueState[&EVI].isOverdefined()) 847 return (void)markOverdefined(&EVI); 848 849 // If this is extracting from more than one level of struct, we don't know. 850 if (EVI.getNumIndices() != 1) 851 return (void)markOverdefined(&EVI); 852 853 Value *AggVal = EVI.getAggregateOperand(); 854 if (AggVal->getType()->isStructTy()) { 855 unsigned i = *EVI.idx_begin(); 856 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 857 mergeInValue(getValueState(&EVI), &EVI, EltVal); 858 } else { 859 // Otherwise, must be extracting from an array. 860 return (void)markOverdefined(&EVI); 861 } 862 } 863 864 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 865 auto *STy = dyn_cast<StructType>(IVI.getType()); 866 if (!STy) 867 return (void)markOverdefined(&IVI); 868 869 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 870 // discover a concrete value later. 871 if (isOverdefined(ValueState[&IVI])) 872 return (void)markOverdefined(&IVI); 873 874 // If this has more than one index, we can't handle it, drive all results to 875 // undef. 876 if (IVI.getNumIndices() != 1) 877 return (void)markOverdefined(&IVI); 878 879 Value *Aggr = IVI.getAggregateOperand(); 880 unsigned Idx = *IVI.idx_begin(); 881 882 // Compute the result based on what we're inserting. 883 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 884 // This passes through all values that aren't the inserted element. 885 if (i != Idx) { 886 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 887 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 888 continue; 889 } 890 891 Value *Val = IVI.getInsertedValueOperand(); 892 if (Val->getType()->isStructTy()) 893 // We don't track structs in structs. 894 markOverdefined(getStructValueState(&IVI, i), &IVI); 895 else { 896 ValueLatticeElement InVal = getValueState(Val); 897 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 898 } 899 } 900 } 901 902 void SCCPSolver::visitSelectInst(SelectInst &I) { 903 // If this select returns a struct, just mark the result overdefined. 904 // TODO: We could do a lot better than this if code actually uses this. 905 if (I.getType()->isStructTy()) 906 return (void)markOverdefined(&I); 907 908 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 909 // discover a concrete value later. 910 if (ValueState[&I].isOverdefined()) 911 return (void)markOverdefined(&I); 912 913 ValueLatticeElement CondValue = getValueState(I.getCondition()); 914 if (CondValue.isUnknownOrUndef()) 915 return; 916 917 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 918 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 919 mergeInValue(&I, getValueState(OpVal)); 920 return; 921 } 922 923 // Otherwise, the condition is overdefined or a constant we can't evaluate. 924 // See if we can produce something better than overdefined based on the T/F 925 // value. 926 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 927 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 928 929 bool Changed = ValueState[&I].mergeIn(TVal); 930 Changed |= ValueState[&I].mergeIn(FVal); 931 if (Changed) 932 pushToWorkListMsg(ValueState[&I], &I); 933 } 934 935 // Handle Unary Operators. 936 void SCCPSolver::visitUnaryOperator(Instruction &I) { 937 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 938 939 ValueLatticeElement &IV = ValueState[&I]; 940 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 941 // discover a concrete value later. 942 if (isOverdefined(IV)) 943 return (void)markOverdefined(&I); 944 945 if (isConstant(V0State)) { 946 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 947 948 // op Y -> undef. 949 if (isa<UndefValue>(C)) 950 return; 951 return (void)markConstant(IV, &I, C); 952 } 953 954 // If something is undef, wait for it to resolve. 955 if (!isOverdefined(V0State)) 956 return; 957 958 markOverdefined(&I); 959 } 960 961 // Handle Binary Operators. 962 void SCCPSolver::visitBinaryOperator(Instruction &I) { 963 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 964 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 965 966 ValueLatticeElement &IV = ValueState[&I]; 967 if (IV.isOverdefined()) 968 return; 969 970 // If something is undef, wait for it to resolve. 971 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 972 return; 973 974 if (V1State.isOverdefined() && V2State.isOverdefined()) 975 return (void)markOverdefined(&I); 976 977 // If either of the operands is a constant, try to fold it to a constant. 978 // TODO: Use information from notconstant better. 979 if ((V1State.isConstant() || V2State.isConstant())) { 980 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 981 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 982 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 983 auto *C = dyn_cast_or_null<Constant>(R); 984 if (C) { 985 // X op Y -> undef. 986 if (isa<UndefValue>(C)) 987 return; 988 // Conservatively assume that the result may be based on operands that may 989 // be undef. Note that we use mergeInValue to combine the constant with 990 // the existing lattice value for I, as different constants might be found 991 // after one of the operands go to overdefined, e.g. due to one operand 992 // being a special floating value. 993 ValueLatticeElement NewV; 994 NewV.markConstant(C, /*MayIncludeUndef=*/true); 995 return (void)mergeInValue(&I, NewV); 996 } 997 } 998 999 // Only use ranges for binary operators on integers. 1000 if (!I.getType()->isIntegerTy()) 1001 return markOverdefined(&I); 1002 1003 // Try to simplify to a constant range. 1004 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1005 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1006 if (V1State.isConstantRange()) 1007 A = V1State.getConstantRange(); 1008 if (V2State.isConstantRange()) 1009 B = V2State.getConstantRange(); 1010 1011 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1012 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1013 1014 // TODO: Currently we do not exploit special values that produce something 1015 // better than overdefined with an overdefined operand for vector or floating 1016 // point types, like and <4 x i32> overdefined, zeroinitializer. 1017 } 1018 1019 // Handle ICmpInst instruction. 1020 void SCCPSolver::visitCmpInst(CmpInst &I) { 1021 // Do not cache this lookup, getValueState calls later in the function might 1022 // invalidate the reference. 1023 if (isOverdefined(ValueState[&I])) 1024 return (void)markOverdefined(&I); 1025 1026 Value *Op1 = I.getOperand(0); 1027 Value *Op2 = I.getOperand(1); 1028 1029 // For parameters, use ParamState which includes constant range info if 1030 // available. 1031 auto V1State = getValueState(Op1); 1032 auto V2State = getValueState(Op2); 1033 1034 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1035 if (C) { 1036 if (isa<UndefValue>(C)) 1037 return; 1038 ValueLatticeElement CV; 1039 CV.markConstant(C); 1040 mergeInValue(&I, CV); 1041 return; 1042 } 1043 1044 // If operands are still unknown, wait for it to resolve. 1045 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1046 !isConstant(ValueState[&I])) 1047 return; 1048 1049 markOverdefined(&I); 1050 } 1051 1052 // Handle getelementptr instructions. If all operands are constants then we 1053 // can turn this into a getelementptr ConstantExpr. 1054 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1055 if (isOverdefined(ValueState[&I])) 1056 return (void)markOverdefined(&I); 1057 1058 SmallVector<Constant*, 8> Operands; 1059 Operands.reserve(I.getNumOperands()); 1060 1061 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1062 ValueLatticeElement State = getValueState(I.getOperand(i)); 1063 if (State.isUnknownOrUndef()) 1064 return; // Operands are not resolved yet. 1065 1066 if (isOverdefined(State)) 1067 return (void)markOverdefined(&I); 1068 1069 if (Constant *C = getConstant(State)) { 1070 Operands.push_back(C); 1071 continue; 1072 } 1073 1074 return (void)markOverdefined(&I); 1075 } 1076 1077 Constant *Ptr = Operands[0]; 1078 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1079 Constant *C = 1080 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1081 if (isa<UndefValue>(C)) 1082 return; 1083 markConstant(&I, C); 1084 } 1085 1086 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1087 // If this store is of a struct, ignore it. 1088 if (SI.getOperand(0)->getType()->isStructTy()) 1089 return; 1090 1091 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1092 return; 1093 1094 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1095 auto I = TrackedGlobals.find(GV); 1096 if (I == TrackedGlobals.end()) 1097 return; 1098 1099 // Get the value we are storing into the global, then merge it. 1100 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1101 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1102 if (I->second.isOverdefined()) 1103 TrackedGlobals.erase(I); // No need to keep tracking this! 1104 } 1105 1106 // Handle load instructions. If the operand is a constant pointer to a constant 1107 // global, we can replace the load with the loaded constant value! 1108 void SCCPSolver::visitLoadInst(LoadInst &I) { 1109 // If this load is of a struct, just mark the result overdefined. 1110 if (I.getType()->isStructTy()) 1111 return (void)markOverdefined(&I); 1112 1113 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1114 // discover a concrete value later. 1115 if (ValueState[&I].isOverdefined()) 1116 return (void)markOverdefined(&I); 1117 1118 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1119 if (PtrVal.isUnknownOrUndef()) 1120 return; // The pointer is not resolved yet! 1121 1122 ValueLatticeElement &IV = ValueState[&I]; 1123 1124 if (!isConstant(PtrVal) || I.isVolatile()) 1125 return (void)markOverdefined(IV, &I); 1126 1127 Constant *Ptr = getConstant(PtrVal); 1128 1129 // load null is undefined. 1130 if (isa<ConstantPointerNull>(Ptr)) { 1131 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1132 return (void)markOverdefined(IV, &I); 1133 else 1134 return; 1135 } 1136 1137 // Transform load (constant global) into the value loaded. 1138 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1139 if (!TrackedGlobals.empty()) { 1140 // If we are tracking this global, merge in the known value for it. 1141 auto It = TrackedGlobals.find(GV); 1142 if (It != TrackedGlobals.end()) { 1143 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1144 return; 1145 } 1146 } 1147 } 1148 1149 // Transform load from a constant into a constant if possible. 1150 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1151 if (isa<UndefValue>(C)) 1152 return; 1153 return (void)markConstant(IV, &I, C); 1154 } 1155 1156 // Otherwise we cannot say for certain what value this load will produce. 1157 // Bail out. 1158 markOverdefined(IV, &I); 1159 } 1160 1161 void SCCPSolver::visitCallBase(CallBase &CB) { 1162 handleCallResult(CB); 1163 handleCallArguments(CB); 1164 } 1165 1166 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1167 Function *F = CB.getCalledFunction(); 1168 1169 // Void return and not tracking callee, just bail. 1170 if (CB.getType()->isVoidTy()) 1171 return; 1172 1173 // Otherwise, if we have a single return value case, and if the function is 1174 // a declaration, maybe we can constant fold it. 1175 if (F && F->isDeclaration() && !CB.getType()->isStructTy() && 1176 canConstantFoldCallTo(&CB, F)) { 1177 SmallVector<Constant *, 8> Operands; 1178 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1179 if (AI->get()->getType()->isStructTy()) 1180 return markOverdefined(&CB); // Can't handle struct args. 1181 ValueLatticeElement State = getValueState(*AI); 1182 1183 if (State.isUnknownOrUndef()) 1184 return; // Operands are not resolved yet. 1185 if (isOverdefined(State)) 1186 return (void)markOverdefined(&CB); 1187 assert(isConstant(State) && "Unknown state!"); 1188 Operands.push_back(getConstant(State)); 1189 } 1190 1191 if (isOverdefined(getValueState(&CB))) 1192 return (void)markOverdefined(&CB); 1193 1194 // If we can constant fold this, mark the result of the call as a 1195 // constant. 1196 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1197 // call -> undef. 1198 if (isa<UndefValue>(C)) 1199 return; 1200 return (void)markConstant(&CB, C); 1201 } 1202 } 1203 1204 // Otherwise, we don't know anything about this call, mark it overdefined. 1205 return (void)markOverdefined(&CB); 1206 } 1207 1208 void SCCPSolver::handleCallArguments(CallBase &CB) { 1209 Function *F = CB.getCalledFunction(); 1210 // If this is a local function that doesn't have its address taken, mark its 1211 // entry block executable and merge in the actual arguments to the call into 1212 // the formal arguments of the function. 1213 if (!TrackingIncomingArguments.empty() && 1214 TrackingIncomingArguments.count(F)) { 1215 MarkBlockExecutable(&F->front()); 1216 1217 // Propagate information from this call site into the callee. 1218 auto CAI = CB.arg_begin(); 1219 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1220 ++AI, ++CAI) { 1221 // If this argument is byval, and if the function is not readonly, there 1222 // will be an implicit copy formed of the input aggregate. 1223 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1224 markOverdefined(&*AI); 1225 continue; 1226 } 1227 1228 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1229 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1230 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1231 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1232 getMaxWidenStepsOpts()); 1233 } 1234 } else 1235 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1236 } 1237 } 1238 } 1239 1240 void SCCPSolver::handleCallResult(CallBase &CB) { 1241 Function *F = CB.getCalledFunction(); 1242 1243 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1244 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1245 if (ValueState[&CB].isOverdefined()) 1246 return; 1247 1248 Value *CopyOf = CB.getOperand(0); 1249 auto *PI = getPredicateInfoFor(&CB); 1250 auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI); 1251 ValueLatticeElement OriginalVal = getValueState(CopyOf); 1252 if (!PI || !PBranch) { 1253 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1254 return; 1255 } 1256 1257 // Everything below relies on the condition being a comparison. 1258 auto *Cmp = dyn_cast<CmpInst>(PBranch->Condition); 1259 if (!Cmp) { 1260 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1261 return; 1262 } 1263 1264 Value *CmpOp0 = Cmp->getOperand(0); 1265 Value *CmpOp1 = Cmp->getOperand(1); 1266 if (CopyOf != CmpOp0 && CopyOf != CmpOp1) { 1267 mergeInValue(ValueState[&CB], &CB, OriginalVal); 1268 return; 1269 } 1270 1271 auto Pred = Cmp->getPredicate(); 1272 if (CmpOp0 != CopyOf) { 1273 std::swap(CmpOp0, CmpOp1); 1274 Pred = Cmp->getSwappedPredicate(); 1275 } 1276 1277 // Wait until CmpOp1 is resolved. 1278 if (getValueState(CmpOp1).isUnknown()) { 1279 addAdditionalUser(CmpOp1, &CB); 1280 return; 1281 } 1282 1283 if (!PBranch->TrueEdge) 1284 Pred = CmpInst::getInversePredicate(Pred); 1285 1286 ValueLatticeElement CondVal = getValueState(CmpOp1); 1287 ValueLatticeElement &IV = ValueState[&CB]; 1288 if (CondVal.isConstantRange() || OriginalVal.isConstantRange()) { 1289 auto NewCR = 1290 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1291 1292 // Get the range imposed by the condition. 1293 if (CondVal.isConstantRange()) 1294 NewCR = ConstantRange::makeAllowedICmpRegion( 1295 Pred, CondVal.getConstantRange()); 1296 1297 // Combine range info for the original value with the new range from the 1298 // condition. 1299 auto OriginalCR = OriginalVal.isConstantRange() 1300 ? OriginalVal.getConstantRange() 1301 : ConstantRange::getFull( 1302 DL.getTypeSizeInBits(CopyOf->getType())); 1303 NewCR = NewCR.intersectWith(OriginalCR); 1304 1305 addAdditionalUser(CmpOp1, &CB); 1306 // TODO: Actually filp MayIncludeUndef for the created range to false, 1307 // once most places in the optimizer respect the branches on 1308 // undef/poison are UB rule. The reason why the new range cannot be 1309 // undef is as follows below: 1310 // The new range is based on a branch condition. That guarantees that 1311 // neither of the compare operands can be undef in the branch targets, 1312 // unless we have conditions that are always true/false (e.g. icmp ule 1313 // i32, %a, i32_max). For the latter overdefined/empty range will be 1314 // inferred, but the branch will get folded accordingly anyways. 1315 mergeInValue( 1316 IV, &CB, 1317 ValueLatticeElement::getRange(NewCR, /*MayIncludeUndef=*/true)); 1318 return; 1319 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1320 // For non-integer values or integer constant expressions, only 1321 // propagate equal constants. 1322 addAdditionalUser(CmpOp1, &CB); 1323 mergeInValue(IV, &CB, CondVal); 1324 return; 1325 } 1326 1327 return (void)mergeInValue(IV, &CB, OriginalVal); 1328 } 1329 } 1330 1331 // The common case is that we aren't tracking the callee, either because we 1332 // are not doing interprocedural analysis or the callee is indirect, or is 1333 // external. Handle these cases first. 1334 if (!F || F->isDeclaration()) 1335 return handleCallOverdefined(CB); 1336 1337 // If this is a single/zero retval case, see if we're tracking the function. 1338 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1339 if (!MRVFunctionsTracked.count(F)) 1340 return handleCallOverdefined(CB); // Not tracking this callee. 1341 1342 // If we are tracking this callee, propagate the result of the function 1343 // into this call site. 1344 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1345 mergeInValue(getStructValueState(&CB, i), &CB, 1346 TrackedMultipleRetVals[std::make_pair(F, i)], 1347 getMaxWidenStepsOpts()); 1348 } else { 1349 auto TFRVI = TrackedRetVals.find(F); 1350 if (TFRVI == TrackedRetVals.end()) 1351 return handleCallOverdefined(CB); // Not tracking this callee. 1352 1353 // If so, propagate the return value of the callee into this call result. 1354 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1355 } 1356 } 1357 1358 void SCCPSolver::Solve() { 1359 // Process the work lists until they are empty! 1360 while (!BBWorkList.empty() || !InstWorkList.empty() || 1361 !OverdefinedInstWorkList.empty()) { 1362 // Process the overdefined instruction's work list first, which drives other 1363 // things to overdefined more quickly. 1364 while (!OverdefinedInstWorkList.empty()) { 1365 Value *I = OverdefinedInstWorkList.pop_back_val(); 1366 1367 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1368 1369 // "I" got into the work list because it either made the transition from 1370 // bottom to constant, or to overdefined. 1371 // 1372 // Anything on this worklist that is overdefined need not be visited 1373 // since all of its users will have already been marked as overdefined 1374 // Update all of the users of this instruction's value. 1375 // 1376 markUsersAsChanged(I); 1377 } 1378 1379 // Process the instruction work list. 1380 while (!InstWorkList.empty()) { 1381 Value *I = InstWorkList.pop_back_val(); 1382 1383 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1384 1385 // "I" got into the work list because it made the transition from undef to 1386 // constant. 1387 // 1388 // Anything on this worklist that is overdefined need not be visited 1389 // since all of its users will have already been marked as overdefined. 1390 // Update all of the users of this instruction's value. 1391 // 1392 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1393 markUsersAsChanged(I); 1394 } 1395 1396 // Process the basic block work list. 1397 while (!BBWorkList.empty()) { 1398 BasicBlock *BB = BBWorkList.back(); 1399 BBWorkList.pop_back(); 1400 1401 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1402 1403 // Notify all instructions in this basic block that they are newly 1404 // executable. 1405 visit(BB); 1406 } 1407 } 1408 } 1409 1410 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1411 /// that branches on undef values cannot reach any of their successors. 1412 /// However, this is not a safe assumption. After we solve dataflow, this 1413 /// method should be use to handle this. If this returns true, the solver 1414 /// should be rerun. 1415 /// 1416 /// This method handles this by finding an unresolved branch and marking it one 1417 /// of the edges from the block as being feasible, even though the condition 1418 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1419 /// CFG and only slightly pessimizes the analysis results (by marking one, 1420 /// potentially infeasible, edge feasible). This cannot usefully modify the 1421 /// constraints on the condition of the branch, as that would impact other users 1422 /// of the value. 1423 /// 1424 /// This scan also checks for values that use undefs. It conservatively marks 1425 /// them as overdefined. 1426 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1427 for (BasicBlock &BB : F) { 1428 if (!BBExecutable.count(&BB)) 1429 continue; 1430 1431 for (Instruction &I : BB) { 1432 // Look for instructions which produce undef values. 1433 if (I.getType()->isVoidTy()) continue; 1434 1435 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1436 // Only a few things that can be structs matter for undef. 1437 1438 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1439 if (auto *CB = dyn_cast<CallBase>(&I)) 1440 if (Function *F = CB->getCalledFunction()) 1441 if (MRVFunctionsTracked.count(F)) 1442 continue; 1443 1444 // extractvalue and insertvalue don't need to be marked; they are 1445 // tracked as precisely as their operands. 1446 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1447 continue; 1448 // Send the results of everything else to overdefined. We could be 1449 // more precise than this but it isn't worth bothering. 1450 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1451 ValueLatticeElement &LV = getStructValueState(&I, i); 1452 if (LV.isUnknownOrUndef()) 1453 markOverdefined(LV, &I); 1454 } 1455 continue; 1456 } 1457 1458 ValueLatticeElement &LV = getValueState(&I); 1459 if (!LV.isUnknownOrUndef()) 1460 continue; 1461 1462 // There are two reasons a call can have an undef result 1463 // 1. It could be tracked. 1464 // 2. It could be constant-foldable. 1465 // Because of the way we solve return values, tracked calls must 1466 // never be marked overdefined in ResolvedUndefsIn. 1467 if (auto *CB = dyn_cast<CallBase>(&I)) 1468 if (Function *F = CB->getCalledFunction()) 1469 if (TrackedRetVals.count(F)) 1470 continue; 1471 1472 if (isa<LoadInst>(I)) { 1473 // A load here means one of two things: a load of undef from a global, 1474 // a load from an unknown pointer. Either way, having it return undef 1475 // is okay. 1476 continue; 1477 } 1478 1479 markOverdefined(&I); 1480 return true; 1481 } 1482 1483 // Check to see if we have a branch or switch on an undefined value. If so 1484 // we force the branch to go one way or the other to make the successor 1485 // values live. It doesn't really matter which way we force it. 1486 Instruction *TI = BB.getTerminator(); 1487 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1488 if (!BI->isConditional()) continue; 1489 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1490 continue; 1491 1492 // If the input to SCCP is actually branch on undef, fix the undef to 1493 // false. 1494 if (isa<UndefValue>(BI->getCondition())) { 1495 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1496 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1497 return true; 1498 } 1499 1500 // Otherwise, it is a branch on a symbolic value which is currently 1501 // considered to be undef. Make sure some edge is executable, so a 1502 // branch on "undef" always flows somewhere. 1503 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1504 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1505 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1506 return true; 1507 1508 continue; 1509 } 1510 1511 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1512 // Indirect branch with no successor ?. Its ok to assume it branches 1513 // to no target. 1514 if (IBR->getNumSuccessors() < 1) 1515 continue; 1516 1517 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1518 continue; 1519 1520 // If the input to SCCP is actually branch on undef, fix the undef to 1521 // the first successor of the indirect branch. 1522 if (isa<UndefValue>(IBR->getAddress())) { 1523 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1524 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1525 return true; 1526 } 1527 1528 // Otherwise, it is a branch on a symbolic value which is currently 1529 // considered to be undef. Make sure some edge is executable, so a 1530 // branch on "undef" always flows somewhere. 1531 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1532 // we can assume the branch has undefined behavior instead. 1533 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1534 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1535 return true; 1536 1537 continue; 1538 } 1539 1540 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1541 if (!SI->getNumCases() || 1542 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1543 continue; 1544 1545 // If the input to SCCP is actually switch on undef, fix the undef to 1546 // the first constant. 1547 if (isa<UndefValue>(SI->getCondition())) { 1548 SI->setCondition(SI->case_begin()->getCaseValue()); 1549 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1550 return true; 1551 } 1552 1553 // Otherwise, it is a branch on a symbolic value which is currently 1554 // considered to be undef. Make sure some edge is executable, so a 1555 // branch on "undef" always flows somewhere. 1556 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1557 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1558 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1559 return true; 1560 1561 continue; 1562 } 1563 } 1564 1565 return false; 1566 } 1567 1568 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1569 Constant *Const = nullptr; 1570 if (V->getType()->isStructTy()) { 1571 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1572 if (any_of(IVs, 1573 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1574 return false; 1575 std::vector<Constant *> ConstVals; 1576 auto *ST = cast<StructType>(V->getType()); 1577 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1578 ValueLatticeElement V = IVs[i]; 1579 ConstVals.push_back(isConstant(V) 1580 ? Solver.getConstant(V) 1581 : UndefValue::get(ST->getElementType(i))); 1582 } 1583 Const = ConstantStruct::get(ST, ConstVals); 1584 } else { 1585 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1586 if (isOverdefined(IV)) 1587 return false; 1588 1589 Const = 1590 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1591 } 1592 assert(Const && "Constant is nullptr here!"); 1593 1594 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1595 // unless the call itself can be removed 1596 CallInst *CI = dyn_cast<CallInst>(V); 1597 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1598 Function *F = CI->getCalledFunction(); 1599 1600 // Don't zap returns of the callee 1601 if (F) 1602 Solver.AddMustTailCallee(F); 1603 1604 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1605 << " as a constant\n"); 1606 return false; 1607 } 1608 1609 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1610 1611 // Replaces all of the uses of a variable with uses of the constant. 1612 V->replaceAllUsesWith(Const); 1613 return true; 1614 } 1615 1616 static bool simplifyInstsInBlock(SCCPSolver &Solver, BasicBlock &BB, 1617 SmallPtrSetImpl<Value *> &InsertedValues, 1618 Statistic &InstRemovedStat, 1619 Statistic &InstReplacedStat) { 1620 bool MadeChanges = false; 1621 for (Instruction &Inst : make_early_inc_range(BB)) { 1622 if (Inst.getType()->isVoidTy()) 1623 continue; 1624 if (tryToReplaceWithConstant(Solver, &Inst)) { 1625 if (Inst.isSafeToRemove()) 1626 Inst.eraseFromParent(); 1627 // Hey, we just changed something! 1628 MadeChanges = true; 1629 ++InstRemovedStat; 1630 } else if (isa<SExtInst>(&Inst)) { 1631 Value *ExtOp = Inst.getOperand(0); 1632 if (isa<Constant>(ExtOp) || InsertedValues.count(ExtOp)) 1633 continue; 1634 const ValueLatticeElement &IV = Solver.getLatticeValueFor(ExtOp); 1635 if (!IV.isConstantRange(/*UndefAllowed=*/false)) 1636 continue; 1637 if (IV.getConstantRange().isAllNonNegative()) { 1638 auto *ZExt = new ZExtInst(ExtOp, Inst.getType(), "", &Inst); 1639 InsertedValues.insert(ZExt); 1640 Inst.replaceAllUsesWith(ZExt); 1641 Solver.removeLatticeValueFor(&Inst); 1642 Inst.eraseFromParent(); 1643 InstReplacedStat++; 1644 MadeChanges = true; 1645 } 1646 } 1647 } 1648 return MadeChanges; 1649 } 1650 1651 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1652 // and return true if the function was modified. 1653 static bool runSCCP(Function &F, const DataLayout &DL, 1654 const TargetLibraryInfo *TLI) { 1655 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1656 SCCPSolver Solver( 1657 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1658 F.getContext()); 1659 1660 // Mark the first block of the function as being executable. 1661 Solver.MarkBlockExecutable(&F.front()); 1662 1663 // Mark all arguments to the function as being overdefined. 1664 for (Argument &AI : F.args()) 1665 Solver.markOverdefined(&AI); 1666 1667 // Solve for constants. 1668 bool ResolvedUndefs = true; 1669 while (ResolvedUndefs) { 1670 Solver.Solve(); 1671 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1672 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1673 } 1674 1675 bool MadeChanges = false; 1676 1677 // If we decided that there are basic blocks that are dead in this function, 1678 // delete their contents now. Note that we cannot actually delete the blocks, 1679 // as we cannot modify the CFG of the function. 1680 1681 SmallPtrSet<Value *, 32> InsertedValues; 1682 for (BasicBlock &BB : F) { 1683 if (!Solver.isBlockExecutable(&BB)) { 1684 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1685 1686 ++NumDeadBlocks; 1687 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB); 1688 1689 MadeChanges = true; 1690 continue; 1691 } 1692 1693 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1694 NumInstRemoved, NumInstReplaced); 1695 } 1696 1697 return MadeChanges; 1698 } 1699 1700 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1701 const DataLayout &DL = F.getParent()->getDataLayout(); 1702 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1703 if (!runSCCP(F, DL, &TLI)) 1704 return PreservedAnalyses::all(); 1705 1706 auto PA = PreservedAnalyses(); 1707 PA.preserve<GlobalsAA>(); 1708 PA.preserveSet<CFGAnalyses>(); 1709 return PA; 1710 } 1711 1712 namespace { 1713 1714 //===--------------------------------------------------------------------===// 1715 // 1716 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1717 /// Sparse Conditional Constant Propagator. 1718 /// 1719 class SCCPLegacyPass : public FunctionPass { 1720 public: 1721 // Pass identification, replacement for typeid 1722 static char ID; 1723 1724 SCCPLegacyPass() : FunctionPass(ID) { 1725 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1726 } 1727 1728 void getAnalysisUsage(AnalysisUsage &AU) const override { 1729 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1730 AU.addPreserved<GlobalsAAWrapperPass>(); 1731 AU.setPreservesCFG(); 1732 } 1733 1734 // runOnFunction - Run the Sparse Conditional Constant Propagation 1735 // algorithm, and return true if the function was modified. 1736 bool runOnFunction(Function &F) override { 1737 if (skipFunction(F)) 1738 return false; 1739 const DataLayout &DL = F.getParent()->getDataLayout(); 1740 const TargetLibraryInfo *TLI = 1741 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1742 return runSCCP(F, DL, TLI); 1743 } 1744 }; 1745 1746 } // end anonymous namespace 1747 1748 char SCCPLegacyPass::ID = 0; 1749 1750 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1751 "Sparse Conditional Constant Propagation", false, false) 1752 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1753 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1754 "Sparse Conditional Constant Propagation", false, false) 1755 1756 // createSCCPPass - This is the public interface to this file. 1757 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1758 1759 static void findReturnsToZap(Function &F, 1760 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1761 SCCPSolver &Solver) { 1762 // We can only do this if we know that nothing else can call the function. 1763 if (!Solver.isArgumentTrackedFunction(&F)) 1764 return; 1765 1766 // There is a non-removable musttail call site of this function. Zapping 1767 // returns is not allowed. 1768 if (Solver.isMustTailCallee(&F)) { 1769 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1770 << " due to present musttail call of it\n"); 1771 return; 1772 } 1773 1774 assert( 1775 all_of(F.users(), 1776 [&Solver](User *U) { 1777 if (isa<Instruction>(U) && 1778 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1779 return true; 1780 // Non-callsite uses are not impacted by zapping. Also, constant 1781 // uses (like blockaddresses) could stuck around, without being 1782 // used in the underlying IR, meaning we do not have lattice 1783 // values for them. 1784 if (!isa<CallBase>(U)) 1785 return true; 1786 if (U->getType()->isStructTy()) { 1787 return all_of(Solver.getStructLatticeValueFor(U), 1788 [](const ValueLatticeElement &LV) { 1789 return !isOverdefined(LV); 1790 }); 1791 } 1792 return !isOverdefined(Solver.getLatticeValueFor(U)); 1793 }) && 1794 "We can only zap functions where all live users have a concrete value"); 1795 1796 for (BasicBlock &BB : F) { 1797 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1798 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1799 << "musttail call : " << *CI << "\n"); 1800 (void)CI; 1801 return; 1802 } 1803 1804 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1805 if (!isa<UndefValue>(RI->getOperand(0))) 1806 ReturnsToZap.push_back(RI); 1807 } 1808 } 1809 1810 // Update the condition for terminators that are branching on indeterminate 1811 // values, forcing them to use a specific edge. 1812 static void forceIndeterminateEdge(Instruction* I, SCCPSolver &Solver) { 1813 BasicBlock *Dest = nullptr; 1814 Constant *C = nullptr; 1815 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { 1816 if (!isa<ConstantInt>(SI->getCondition())) { 1817 // Indeterminate switch; use first case value. 1818 Dest = SI->case_begin()->getCaseSuccessor(); 1819 C = SI->case_begin()->getCaseValue(); 1820 } 1821 } else if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1822 if (!isa<ConstantInt>(BI->getCondition())) { 1823 // Indeterminate branch; use false. 1824 Dest = BI->getSuccessor(1); 1825 C = ConstantInt::getFalse(BI->getContext()); 1826 } 1827 } else if (IndirectBrInst *IBR = dyn_cast<IndirectBrInst>(I)) { 1828 if (!isa<BlockAddress>(IBR->getAddress()->stripPointerCasts())) { 1829 // Indeterminate indirectbr; use successor 0. 1830 Dest = IBR->getSuccessor(0); 1831 C = BlockAddress::get(IBR->getSuccessor(0)); 1832 } 1833 } else { 1834 llvm_unreachable("Unexpected terminator instruction"); 1835 } 1836 if (C) { 1837 assert(Solver.isEdgeFeasible(I->getParent(), Dest) && 1838 "Didn't find feasible edge?"); 1839 (void)Dest; 1840 1841 I->setOperand(0, C); 1842 } 1843 } 1844 1845 bool llvm::runIPSCCP( 1846 Module &M, const DataLayout &DL, 1847 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1848 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1849 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1850 1851 // Loop over all functions, marking arguments to those with their addresses 1852 // taken or that are external as overdefined. 1853 for (Function &F : M) { 1854 if (F.isDeclaration()) 1855 continue; 1856 1857 Solver.addAnalysis(F, getAnalysis(F)); 1858 1859 // Determine if we can track the function's return values. If so, add the 1860 // function to the solver's set of return-tracked functions. 1861 if (canTrackReturnsInterprocedurally(&F)) 1862 Solver.AddTrackedFunction(&F); 1863 1864 // Determine if we can track the function's arguments. If so, add the 1865 // function to the solver's set of argument-tracked functions. 1866 if (canTrackArgumentsInterprocedurally(&F)) { 1867 Solver.AddArgumentTrackedFunction(&F); 1868 continue; 1869 } 1870 1871 // Assume the function is called. 1872 Solver.MarkBlockExecutable(&F.front()); 1873 1874 // Assume nothing about the incoming arguments. 1875 for (Argument &AI : F.args()) 1876 Solver.markOverdefined(&AI); 1877 } 1878 1879 // Determine if we can track any of the module's global variables. If so, add 1880 // the global variables we can track to the solver's set of tracked global 1881 // variables. 1882 for (GlobalVariable &G : M.globals()) { 1883 G.removeDeadConstantUsers(); 1884 if (canTrackGlobalVariableInterprocedurally(&G)) 1885 Solver.TrackValueOfGlobalVariable(&G); 1886 } 1887 1888 // Solve for constants. 1889 bool ResolvedUndefs = true; 1890 Solver.Solve(); 1891 while (ResolvedUndefs) { 1892 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1893 ResolvedUndefs = false; 1894 for (Function &F : M) 1895 if (Solver.ResolvedUndefsIn(F)) { 1896 // We run Solve() after we resolved an undef in a function, because 1897 // we might deduce a fact that eliminates an undef in another function. 1898 Solver.Solve(); 1899 ResolvedUndefs = true; 1900 } 1901 } 1902 1903 bool MadeChanges = false; 1904 1905 // Iterate over all of the instructions in the module, replacing them with 1906 // constants if we have found them to be of constant values. 1907 1908 for (Function &F : M) { 1909 if (F.isDeclaration()) 1910 continue; 1911 1912 SmallVector<BasicBlock *, 512> BlocksToErase; 1913 1914 if (Solver.isBlockExecutable(&F.front())) 1915 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E; 1916 ++AI) { 1917 if (!AI->use_empty() && tryToReplaceWithConstant(Solver, &*AI)) { 1918 ++IPNumArgsElimed; 1919 continue; 1920 } 1921 } 1922 1923 SmallPtrSet<Value *, 32> InsertedValues; 1924 for (BasicBlock &BB : F) { 1925 if (!Solver.isBlockExecutable(&BB)) { 1926 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1927 ++NumDeadBlocks; 1928 1929 MadeChanges = true; 1930 1931 if (&BB != &F.front()) 1932 BlocksToErase.push_back(&BB); 1933 continue; 1934 } 1935 1936 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1937 IPNumInstRemoved, IPNumInstReplaced); 1938 } 1939 1940 DomTreeUpdater DTU = Solver.getDTU(F); 1941 // Change dead blocks to unreachable. We do it after replacing constants 1942 // in all executable blocks, because changeToUnreachable may remove PHI 1943 // nodes in executable blocks we found values for. The function's entry 1944 // block is not part of BlocksToErase, so we have to handle it separately. 1945 for (BasicBlock *BB : BlocksToErase) { 1946 NumInstRemoved += 1947 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 1948 /*PreserveLCSSA=*/false, &DTU); 1949 } 1950 if (!Solver.isBlockExecutable(&F.front())) 1951 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 1952 /*UseLLVMTrap=*/false, 1953 /*PreserveLCSSA=*/false, &DTU); 1954 1955 // Now that all instructions in the function are constant folded, 1956 // use ConstantFoldTerminator to get rid of in-edges, record DT updates and 1957 // delete dead BBs. 1958 for (BasicBlock *DeadBB : BlocksToErase) { 1959 // If there are any PHI nodes in this successor, drop entries for BB now. 1960 for (Value::user_iterator UI = DeadBB->user_begin(), 1961 UE = DeadBB->user_end(); 1962 UI != UE;) { 1963 // Grab the user and then increment the iterator early, as the user 1964 // will be deleted. Step past all adjacent uses from the same user. 1965 auto *I = dyn_cast<Instruction>(*UI); 1966 do { ++UI; } while (UI != UE && *UI == I); 1967 1968 // Ignore blockaddress users; BasicBlock's dtor will handle them. 1969 if (!I) continue; 1970 1971 // If we have forced an edge for an indeterminate value, then force the 1972 // terminator to fold to that edge. 1973 forceIndeterminateEdge(I, Solver); 1974 BasicBlock *InstBB = I->getParent(); 1975 bool Folded = ConstantFoldTerminator(InstBB, 1976 /*DeleteDeadConditions=*/false, 1977 /*TLI=*/nullptr, &DTU); 1978 assert(Folded && 1979 "Expect TermInst on constantint or blockaddress to be folded"); 1980 (void) Folded; 1981 // If we folded the terminator to an unconditional branch to another 1982 // dead block, replace it with Unreachable, to avoid trying to fold that 1983 // branch again. 1984 BranchInst *BI = cast<BranchInst>(InstBB->getTerminator()); 1985 if (BI && BI->isUnconditional() && 1986 !Solver.isBlockExecutable(BI->getSuccessor(0))) { 1987 InstBB->getTerminator()->eraseFromParent(); 1988 new UnreachableInst(InstBB->getContext(), InstBB); 1989 } 1990 } 1991 // Mark dead BB for deletion. 1992 DTU.deleteBB(DeadBB); 1993 } 1994 1995 for (BasicBlock &BB : F) { 1996 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1997 Instruction *Inst = &*BI++; 1998 if (Solver.getPredicateInfoFor(Inst)) { 1999 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2000 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2001 Value *Op = II->getOperand(0); 2002 Inst->replaceAllUsesWith(Op); 2003 Inst->eraseFromParent(); 2004 } 2005 } 2006 } 2007 } 2008 } 2009 } 2010 2011 // If we inferred constant or undef return values for a function, we replaced 2012 // all call uses with the inferred value. This means we don't need to bother 2013 // actually returning anything from the function. Replace all return 2014 // instructions with return undef. 2015 // 2016 // Do this in two stages: first identify the functions we should process, then 2017 // actually zap their returns. This is important because we can only do this 2018 // if the address of the function isn't taken. In cases where a return is the 2019 // last use of a function, the order of processing functions would affect 2020 // whether other functions are optimizable. 2021 SmallVector<ReturnInst*, 8> ReturnsToZap; 2022 2023 for (const auto &I : Solver.getTrackedRetVals()) { 2024 Function *F = I.first; 2025 if (isOverdefined(I.second) || F->getReturnType()->isVoidTy()) 2026 continue; 2027 findReturnsToZap(*F, ReturnsToZap, Solver); 2028 } 2029 2030 for (auto F : Solver.getMRVFunctionsTracked()) { 2031 assert(F->getReturnType()->isStructTy() && 2032 "The return type should be a struct"); 2033 StructType *STy = cast<StructType>(F->getReturnType()); 2034 if (Solver.isStructLatticeConstant(F, STy)) 2035 findReturnsToZap(*F, ReturnsToZap, Solver); 2036 } 2037 2038 // Zap all returns which we've identified as zap to change. 2039 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2040 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2041 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2042 } 2043 2044 // If we inferred constant or undef values for globals variables, we can 2045 // delete the global and any stores that remain to it. 2046 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2047 GlobalVariable *GV = I.first; 2048 if (isOverdefined(I.second)) 2049 continue; 2050 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2051 << "' is constant!\n"); 2052 while (!GV->use_empty()) { 2053 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2054 SI->eraseFromParent(); 2055 MadeChanges = true; 2056 } 2057 M.getGlobalList().erase(GV); 2058 ++IPNumGlobalConst; 2059 } 2060 2061 return MadeChanges; 2062 } 2063