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.isOverdefined() || 108 (LV.isConstantRange() && !LV.getConstantRange().isSingleElement()); 109 } 110 111 //===----------------------------------------------------------------------===// 112 // 113 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 114 /// Constant Propagation. 115 /// 116 class SCCPSolver : public InstVisitor<SCCPSolver> { 117 const DataLayout &DL; 118 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 119 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 120 DenseMap<Value *, ValueLatticeElement> 121 ValueState; // The state each value is in. 122 123 /// StructValueState - This maintains ValueState for values that have 124 /// StructType, for example for formal arguments, calls, insertelement, etc. 125 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 126 127 /// GlobalValue - If we are tracking any values for the contents of a global 128 /// variable, we keep a mapping from the constant accessor to the element of 129 /// the global, to the currently known value. If the value becomes 130 /// overdefined, it's entry is simply removed from this map. 131 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 132 133 /// TrackedRetVals - If we are tracking arguments into and the return 134 /// value out of a function, it will have an entry in this map, indicating 135 /// what the known return value for the function is. 136 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 137 138 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 139 /// that return multiple values. 140 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 141 TrackedMultipleRetVals; 142 143 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 144 /// represented here for efficient lookup. 145 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 146 147 /// MustTailFunctions - Each function here is a callee of non-removable 148 /// musttail call site. 149 SmallPtrSet<Function *, 16> MustTailCallees; 150 151 /// TrackingIncomingArguments - This is the set of functions for whose 152 /// arguments we make optimistic assumptions about and try to prove as 153 /// constants. 154 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 155 156 /// The reason for two worklists is that overdefined is the lowest state 157 /// on the lattice, and moving things to overdefined as fast as possible 158 /// makes SCCP converge much faster. 159 /// 160 /// By having a separate worklist, we accomplish this because everything 161 /// possibly overdefined will become overdefined at the soonest possible 162 /// point. 163 SmallVector<Value *, 64> OverdefinedInstWorkList; 164 SmallVector<Value *, 64> InstWorkList; 165 166 // The BasicBlock work list 167 SmallVector<BasicBlock *, 64> BBWorkList; 168 169 /// KnownFeasibleEdges - Entries in this set are edges which have already had 170 /// PHI nodes retriggered. 171 using Edge = std::pair<BasicBlock *, BasicBlock *>; 172 DenseSet<Edge> KnownFeasibleEdges; 173 174 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 175 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 176 177 LLVMContext &Ctx; 178 179 public: 180 void addAnalysis(Function &F, AnalysisResultsForFn A) { 181 AnalysisResults.insert({&F, std::move(A)}); 182 } 183 184 const PredicateBase *getPredicateInfoFor(Instruction *I) { 185 auto A = AnalysisResults.find(I->getParent()->getParent()); 186 if (A == AnalysisResults.end()) 187 return nullptr; 188 return A->second.PredInfo->getPredicateInfoFor(I); 189 } 190 191 DomTreeUpdater getDTU(Function &F) { 192 auto A = AnalysisResults.find(&F); 193 assert(A != AnalysisResults.end() && "Need analysis results for function."); 194 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 195 } 196 197 SCCPSolver(const DataLayout &DL, 198 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 199 LLVMContext &Ctx) 200 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 201 202 /// MarkBlockExecutable - This method can be used by clients to mark all of 203 /// the blocks that are known to be intrinsically live in the processed unit. 204 /// 205 /// This returns true if the block was not considered live before. 206 bool MarkBlockExecutable(BasicBlock *BB) { 207 if (!BBExecutable.insert(BB).second) 208 return false; 209 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 210 BBWorkList.push_back(BB); // Add the block to the work list! 211 return true; 212 } 213 214 /// TrackValueOfGlobalVariable - Clients can use this method to 215 /// inform the SCCPSolver that it should track loads and stores to the 216 /// specified global variable if it can. This is only legal to call if 217 /// performing Interprocedural SCCP. 218 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 219 // We only track the contents of scalar globals. 220 if (GV->getValueType()->isSingleValueType()) { 221 ValueLatticeElement &IV = TrackedGlobals[GV]; 222 if (!isa<UndefValue>(GV->getInitializer())) 223 IV.markConstant(GV->getInitializer()); 224 } 225 } 226 227 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 228 /// and out of the specified function (which cannot have its address taken), 229 /// this method must be called. 230 void AddTrackedFunction(Function *F) { 231 // Add an entry, F -> undef. 232 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 233 MRVFunctionsTracked.insert(F); 234 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 235 TrackedMultipleRetVals.insert( 236 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 237 } else if (!F->getReturnType()->isVoidTy()) 238 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 239 } 240 241 /// AddMustTailCallee - If the SCCP solver finds that this function is called 242 /// from non-removable musttail call site. 243 void AddMustTailCallee(Function *F) { 244 MustTailCallees.insert(F); 245 } 246 247 /// Returns true if the given function is called from non-removable musttail 248 /// call site. 249 bool isMustTailCallee(Function *F) { 250 return MustTailCallees.count(F); 251 } 252 253 void AddArgumentTrackedFunction(Function *F) { 254 TrackingIncomingArguments.insert(F); 255 } 256 257 /// Returns true if the given function is in the solver's set of 258 /// argument-tracked functions. 259 bool isArgumentTrackedFunction(Function *F) { 260 return TrackingIncomingArguments.count(F); 261 } 262 263 /// Solve - Solve for constants and executable blocks. 264 void Solve(); 265 266 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 267 /// that branches on undef values cannot reach any of their successors. 268 /// However, this is not a safe assumption. After we solve dataflow, this 269 /// method should be use to handle this. If this returns true, the solver 270 /// should be rerun. 271 bool ResolvedUndefsIn(Function &F); 272 273 bool isBlockExecutable(BasicBlock *BB) const { 274 return BBExecutable.count(BB); 275 } 276 277 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 278 // block to the 'To' basic block is currently feasible. 279 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 280 281 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 282 std::vector<ValueLatticeElement> StructValues; 283 auto *STy = dyn_cast<StructType>(V->getType()); 284 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 285 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 286 auto I = StructValueState.find(std::make_pair(V, i)); 287 assert(I != StructValueState.end() && "Value not in valuemap!"); 288 StructValues.push_back(I->second); 289 } 290 return StructValues; 291 } 292 293 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 294 295 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 296 assert(!V->getType()->isStructTy() && 297 "Should use getStructLatticeValueFor"); 298 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 299 ValueState.find(V); 300 assert(I != ValueState.end() && 301 "V not found in ValueState nor Paramstate map!"); 302 return I->second; 303 } 304 305 /// getTrackedRetVals - Get the inferred return value map. 306 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 307 return TrackedRetVals; 308 } 309 310 /// getTrackedGlobals - Get and return the set of inferred initializers for 311 /// global variables. 312 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 313 return TrackedGlobals; 314 } 315 316 /// getMRVFunctionsTracked - Get the set of functions which return multiple 317 /// values tracked by the pass. 318 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 319 return MRVFunctionsTracked; 320 } 321 322 /// getMustTailCallees - Get the set of functions which are called 323 /// from non-removable musttail call sites. 324 const SmallPtrSet<Function *, 16> getMustTailCallees() { 325 return MustTailCallees; 326 } 327 328 /// markOverdefined - Mark the specified value overdefined. This 329 /// works with both scalars and structs. 330 void markOverdefined(Value *V) { 331 if (auto *STy = dyn_cast<StructType>(V->getType())) 332 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 333 markOverdefined(getStructValueState(V, i), V); 334 else 335 markOverdefined(ValueState[V], V); 336 } 337 338 // isStructLatticeConstant - Return true if all the lattice values 339 // corresponding to elements of the structure are constants, 340 // false otherwise. 341 bool isStructLatticeConstant(Function *F, StructType *STy) { 342 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 343 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 344 assert(It != TrackedMultipleRetVals.end()); 345 ValueLatticeElement LV = It->second; 346 if (!isConstant(LV)) 347 return false; 348 } 349 return true; 350 } 351 352 /// Helper to return a Constant if \p LV is either a constant or a constant 353 /// range with a single element. 354 Constant *getConstant(const ValueLatticeElement &LV) const { 355 if (LV.isConstant()) 356 return LV.getConstant(); 357 358 if (LV.isConstantRange()) { 359 auto &CR = LV.getConstantRange(); 360 if (CR.getSingleElement()) 361 return ConstantInt::get(Ctx, *CR.getSingleElement()); 362 } 363 return nullptr; 364 } 365 366 private: 367 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 368 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 369 } 370 371 // pushToWorkList - Helper for markConstant/markOverdefined 372 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 373 if (IV.isOverdefined()) 374 return OverdefinedInstWorkList.push_back(V); 375 InstWorkList.push_back(V); 376 } 377 378 // Helper to push \p V to the worklist, after updating it to \p IV. Also 379 // prints a debug message with the updated value. 380 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 381 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 382 pushToWorkList(IV, V); 383 } 384 385 // markConstant - Make a value be marked as "constant". If the value 386 // is not already a constant, add it to the instruction work list so that 387 // the users of the instruction are updated later. 388 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 389 bool MayIncludeUndef = false) { 390 if (!IV.markConstant(C, MayIncludeUndef)) 391 return false; 392 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 393 pushToWorkList(IV, V); 394 return true; 395 } 396 397 bool markConstant(Value *V, Constant *C) { 398 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 399 return markConstant(ValueState[V], V, C); 400 } 401 402 // markOverdefined - Make a value be marked as "overdefined". If the 403 // value is not already overdefined, add it to the overdefined instruction 404 // work list so that the users of the instruction are updated later. 405 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 406 if (!IV.markOverdefined()) return false; 407 408 LLVM_DEBUG(dbgs() << "markOverdefined: "; 409 if (auto *F = dyn_cast<Function>(V)) dbgs() 410 << "Function '" << F->getName() << "'\n"; 411 else dbgs() << *V << '\n'); 412 // Only instructions go on the work list 413 pushToWorkList(IV, V); 414 return true; 415 } 416 417 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 418 /// changes. 419 bool mergeInValue(ValueLatticeElement &IV, Value *V, 420 ValueLatticeElement MergeWithV, 421 ValueLatticeElement::MergeOptions Opts = { 422 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 423 if (IV.mergeIn(MergeWithV, Opts)) { 424 pushToWorkList(IV, V); 425 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 426 << IV << "\n"); 427 return true; 428 } 429 return false; 430 } 431 432 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 433 ValueLatticeElement::MergeOptions Opts = { 434 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 435 assert(!V->getType()->isStructTy() && 436 "non-structs should use markConstant"); 437 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 438 } 439 440 /// getValueState - Return the ValueLatticeElement object that corresponds to 441 /// the value. This function handles the case when the value hasn't been seen 442 /// yet by properly seeding constants etc. 443 ValueLatticeElement &getValueState(Value *V) { 444 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 445 446 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 447 ValueLatticeElement &LV = I.first->second; 448 449 if (!I.second) 450 return LV; // Common case, already in the map. 451 452 if (auto *C = dyn_cast<Constant>(V)) 453 LV.markConstant(C); // Constants are constant 454 455 // All others are unknown by default. 456 return LV; 457 } 458 459 /// getStructValueState - Return the ValueLatticeElement object that 460 /// corresponds to the value/field pair. This function handles the case when 461 /// the value hasn't been seen yet by properly seeding constants etc. 462 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 463 assert(V->getType()->isStructTy() && "Should use getValueState"); 464 assert(i < cast<StructType>(V->getType())->getNumElements() && 465 "Invalid element #"); 466 467 auto I = StructValueState.insert( 468 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 469 ValueLatticeElement &LV = I.first->second; 470 471 if (!I.second) 472 return LV; // Common case, already in the map. 473 474 if (auto *C = dyn_cast<Constant>(V)) { 475 Constant *Elt = C->getAggregateElement(i); 476 477 if (!Elt) 478 LV.markOverdefined(); // Unknown sort of constant. 479 else if (isa<UndefValue>(Elt)) 480 ; // Undef values remain unknown. 481 else 482 LV.markConstant(Elt); // Constants are constant. 483 } 484 485 // All others are underdefined by default. 486 return LV; 487 } 488 489 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 490 /// work list if it is not already executable. 491 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 492 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 493 return false; // This edge is already known to be executable! 494 495 if (!MarkBlockExecutable(Dest)) { 496 // If the destination is already executable, we just made an *edge* 497 // feasible that wasn't before. Revisit the PHI nodes in the block 498 // because they have potentially new operands. 499 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 500 << " -> " << Dest->getName() << '\n'); 501 502 for (PHINode &PN : Dest->phis()) 503 visitPHINode(PN); 504 } 505 return true; 506 } 507 508 // getFeasibleSuccessors - Return a vector of booleans to indicate which 509 // successors are reachable from a given terminator instruction. 510 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 511 512 // OperandChangedState - This method is invoked on all of the users of an 513 // instruction that was just changed state somehow. Based on this 514 // information, we need to update the specified user of this instruction. 515 void OperandChangedState(Instruction *I) { 516 if (BBExecutable.count(I->getParent())) // Inst is executable? 517 visit(*I); 518 } 519 520 // Add U as additional user of V. 521 void addAdditionalUser(Value *V, User *U) { 522 auto Iter = AdditionalUsers.insert({V, {}}); 523 Iter.first->second.insert(U); 524 } 525 526 // Mark I's users as changed, including AdditionalUsers. 527 void markUsersAsChanged(Value *I) { 528 // Functions include their arguments in the use-list. Changed function 529 // values mean that the result of the function changed. We only need to 530 // update the call sites with the new function result and do not have to 531 // propagate the call arguments. 532 if (isa<Function>(I)) { 533 for (User *U : I->users()) { 534 if (auto *CB = dyn_cast<CallBase>(U)) 535 handleCallResult(*CB); 536 } 537 } else { 538 for (User *U : I->users()) 539 if (auto *UI = dyn_cast<Instruction>(U)) 540 OperandChangedState(UI); 541 } 542 543 auto Iter = AdditionalUsers.find(I); 544 if (Iter != AdditionalUsers.end()) { 545 for (User *U : Iter->second) 546 if (auto *UI = dyn_cast<Instruction>(U)) 547 OperandChangedState(UI); 548 } 549 } 550 void handleCallOverdefined(CallBase &CB); 551 void handleCallResult(CallBase &CB); 552 void handleCallArguments(CallBase &CB); 553 554 private: 555 friend class InstVisitor<SCCPSolver>; 556 557 // visit implementations - Something changed in this instruction. Either an 558 // operand made a transition, or the instruction is newly executable. Change 559 // the value type of I to reflect these changes if appropriate. 560 void visitPHINode(PHINode &I); 561 562 // Terminators 563 564 void visitReturnInst(ReturnInst &I); 565 void visitTerminator(Instruction &TI); 566 567 void visitCastInst(CastInst &I); 568 void visitSelectInst(SelectInst &I); 569 void visitUnaryOperator(Instruction &I); 570 void visitBinaryOperator(Instruction &I); 571 void visitCmpInst(CmpInst &I); 572 void visitExtractValueInst(ExtractValueInst &EVI); 573 void visitInsertValueInst(InsertValueInst &IVI); 574 575 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 576 markOverdefined(&CPI); 577 visitTerminator(CPI); 578 } 579 580 // Instructions that cannot be folded away. 581 582 void visitStoreInst (StoreInst &I); 583 void visitLoadInst (LoadInst &I); 584 void visitGetElementPtrInst(GetElementPtrInst &I); 585 586 void visitCallInst (CallInst &I) { 587 visitCallBase(I); 588 } 589 590 void visitInvokeInst (InvokeInst &II) { 591 visitCallBase(II); 592 visitTerminator(II); 593 } 594 595 void visitCallBrInst (CallBrInst &CBI) { 596 visitCallBase(CBI); 597 visitTerminator(CBI); 598 } 599 600 void visitCallBase (CallBase &CB); 601 void visitResumeInst (ResumeInst &I) { /*returns void*/ } 602 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ } 603 void visitFenceInst (FenceInst &I) { /*returns void*/ } 604 605 void visitInstruction(Instruction &I) { 606 // All the instructions we don't do any special handling for just 607 // go to overdefined. 608 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 609 markOverdefined(&I); 610 } 611 }; 612 613 } // end anonymous namespace 614 615 // getFeasibleSuccessors - Return a vector of booleans to indicate which 616 // successors are reachable from a given terminator instruction. 617 void SCCPSolver::getFeasibleSuccessors(Instruction &TI, 618 SmallVectorImpl<bool> &Succs) { 619 Succs.resize(TI.getNumSuccessors()); 620 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 621 if (BI->isUnconditional()) { 622 Succs[0] = true; 623 return; 624 } 625 626 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 627 ConstantInt *CI = getConstantInt(BCValue); 628 if (!CI) { 629 // Overdefined condition variables, and branches on unfoldable constant 630 // conditions, mean the branch could go either way. 631 if (!BCValue.isUnknownOrUndef()) 632 Succs[0] = Succs[1] = true; 633 return; 634 } 635 636 // Constant condition variables mean the branch can only go a single way. 637 Succs[CI->isZero()] = true; 638 return; 639 } 640 641 // Unwinding instructions successors are always executable. 642 if (TI.isExceptionalTerminator()) { 643 Succs.assign(TI.getNumSuccessors(), true); 644 return; 645 } 646 647 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 648 if (!SI->getNumCases()) { 649 Succs[0] = true; 650 return; 651 } 652 ValueLatticeElement SCValue = getValueState(SI->getCondition()); 653 ConstantInt *CI = getConstantInt(SCValue); 654 655 if (!CI) { // Overdefined or unknown condition? 656 // All destinations are executable! 657 if (!SCValue.isUnknownOrUndef()) 658 Succs.assign(TI.getNumSuccessors(), true); 659 return; 660 } 661 662 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 663 return; 664 } 665 666 // In case of indirect branch and its address is a blockaddress, we mark 667 // the target as executable. 668 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 669 // Casts are folded by visitCastInst. 670 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 671 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 672 if (!Addr) { // Overdefined or unknown condition? 673 // All destinations are executable! 674 if (!IBRValue.isUnknownOrUndef()) 675 Succs.assign(TI.getNumSuccessors(), true); 676 return; 677 } 678 679 BasicBlock* T = Addr->getBasicBlock(); 680 assert(Addr->getFunction() == T->getParent() && 681 "Block address of a different function ?"); 682 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 683 // This is the target. 684 if (IBR->getDestination(i) == T) { 685 Succs[i] = true; 686 return; 687 } 688 } 689 690 // If we didn't find our destination in the IBR successor list, then we 691 // have undefined behavior. Its ok to assume no successor is executable. 692 return; 693 } 694 695 // In case of callbr, we pessimistically assume that all successors are 696 // feasible. 697 if (isa<CallBrInst>(&TI)) { 698 Succs.assign(TI.getNumSuccessors(), true); 699 return; 700 } 701 702 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 703 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 704 } 705 706 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 707 // block to the 'To' basic block is currently feasible. 708 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 709 // Check if we've called markEdgeExecutable on the edge yet. (We could 710 // be more aggressive and try to consider edges which haven't been marked 711 // yet, but there isn't any need.) 712 return KnownFeasibleEdges.count(Edge(From, To)); 713 } 714 715 // visit Implementations - Something changed in this instruction, either an 716 // operand made a transition, or the instruction is newly executable. Change 717 // the value type of I to reflect these changes if appropriate. This method 718 // makes sure to do the following actions: 719 // 720 // 1. If a phi node merges two constants in, and has conflicting value coming 721 // from different branches, or if the PHI node merges in an overdefined 722 // value, then the PHI node becomes overdefined. 723 // 2. If a phi node merges only constants in, and they all agree on value, the 724 // PHI node becomes a constant value equal to that. 725 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 726 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 727 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 728 // 6. If a conditional branch has a value that is constant, make the selected 729 // destination executable 730 // 7. If a conditional branch has a value that is overdefined, make all 731 // successors executable. 732 void SCCPSolver::visitPHINode(PHINode &PN) { 733 // If this PN returns a struct, just mark the result overdefined. 734 // TODO: We could do a lot better than this if code actually uses this. 735 if (PN.getType()->isStructTy()) 736 return (void)markOverdefined(&PN); 737 738 if (getValueState(&PN).isOverdefined()) 739 return; // Quick exit 740 741 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 742 // and slow us down a lot. Just mark them overdefined. 743 if (PN.getNumIncomingValues() > 64) 744 return (void)markOverdefined(&PN); 745 746 unsigned NumActiveIncoming = 0; 747 748 // Look at all of the executable operands of the PHI node. If any of them 749 // are overdefined, the PHI becomes overdefined as well. If they are all 750 // constant, and they agree with each other, the PHI becomes the identical 751 // constant. If they are constant and don't agree, the PHI is a constant 752 // range. If there are no executable operands, the PHI remains unknown. 753 ValueLatticeElement PhiState = getValueState(&PN); 754 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 755 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 756 continue; 757 758 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 759 PhiState.mergeIn(IV); 760 NumActiveIncoming++; 761 if (PhiState.isOverdefined()) 762 break; 763 } 764 765 // We allow up to 1 range extension per active incoming value and one 766 // additional extension. Note that we manually adjust the number of range 767 // extensions to match the number of active incoming values. This helps to 768 // limit multiple extensions caused by the same incoming value, if other 769 // incoming values are equal. 770 mergeInValue(&PN, PhiState, 771 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 772 NumActiveIncoming + 1)); 773 ValueLatticeElement &PhiStateRef = getValueState(&PN); 774 PhiStateRef.setNumRangeExtensions( 775 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 776 } 777 778 void SCCPSolver::visitReturnInst(ReturnInst &I) { 779 if (I.getNumOperands() == 0) return; // ret void 780 781 Function *F = I.getParent()->getParent(); 782 Value *ResultOp = I.getOperand(0); 783 784 // If we are tracking the return value of this function, merge it in. 785 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 786 auto TFRVI = TrackedRetVals.find(F); 787 if (TFRVI != TrackedRetVals.end()) { 788 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 789 return; 790 } 791 } 792 793 // Handle functions that return multiple values. 794 if (!TrackedMultipleRetVals.empty()) { 795 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 796 if (MRVFunctionsTracked.count(F)) 797 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 798 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 799 getStructValueState(ResultOp, i)); 800 } 801 } 802 803 void SCCPSolver::visitTerminator(Instruction &TI) { 804 SmallVector<bool, 16> SuccFeasible; 805 getFeasibleSuccessors(TI, SuccFeasible); 806 807 BasicBlock *BB = TI.getParent(); 808 809 // Mark all feasible successors executable. 810 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 811 if (SuccFeasible[i]) 812 markEdgeExecutable(BB, TI.getSuccessor(i)); 813 } 814 815 void SCCPSolver::visitCastInst(CastInst &I) { 816 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 817 // discover a concrete value later. 818 if (ValueState[&I].isOverdefined()) 819 return; 820 821 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 822 if (Constant *OpC = getConstant(OpSt)) { 823 // Fold the constant as we build. 824 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 825 if (isa<UndefValue>(C)) 826 return; 827 // Propagate constant value 828 markConstant(&I, C); 829 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 830 auto &LV = getValueState(&I); 831 ConstantRange OpRange = OpSt.getConstantRange(); 832 Type *DestTy = I.getDestTy(); 833 ConstantRange Res = 834 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 835 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 836 } else if (!OpSt.isUnknownOrUndef()) 837 markOverdefined(&I); 838 } 839 840 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 841 // If this returns a struct, mark all elements over defined, we don't track 842 // structs in structs. 843 if (EVI.getType()->isStructTy()) 844 return (void)markOverdefined(&EVI); 845 846 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 847 // discover a concrete value later. 848 if (ValueState[&EVI].isOverdefined()) 849 return (void)markOverdefined(&EVI); 850 851 // If this is extracting from more than one level of struct, we don't know. 852 if (EVI.getNumIndices() != 1) 853 return (void)markOverdefined(&EVI); 854 855 Value *AggVal = EVI.getAggregateOperand(); 856 if (AggVal->getType()->isStructTy()) { 857 unsigned i = *EVI.idx_begin(); 858 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 859 mergeInValue(getValueState(&EVI), &EVI, EltVal); 860 } else { 861 // Otherwise, must be extracting from an array. 862 return (void)markOverdefined(&EVI); 863 } 864 } 865 866 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 867 auto *STy = dyn_cast<StructType>(IVI.getType()); 868 if (!STy) 869 return (void)markOverdefined(&IVI); 870 871 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 872 // discover a concrete value later. 873 if (isOverdefined(ValueState[&IVI])) 874 return (void)markOverdefined(&IVI); 875 876 // If this has more than one index, we can't handle it, drive all results to 877 // undef. 878 if (IVI.getNumIndices() != 1) 879 return (void)markOverdefined(&IVI); 880 881 Value *Aggr = IVI.getAggregateOperand(); 882 unsigned Idx = *IVI.idx_begin(); 883 884 // Compute the result based on what we're inserting. 885 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 886 // This passes through all values that aren't the inserted element. 887 if (i != Idx) { 888 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 889 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 890 continue; 891 } 892 893 Value *Val = IVI.getInsertedValueOperand(); 894 if (Val->getType()->isStructTy()) 895 // We don't track structs in structs. 896 markOverdefined(getStructValueState(&IVI, i), &IVI); 897 else { 898 ValueLatticeElement InVal = getValueState(Val); 899 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 900 } 901 } 902 } 903 904 void SCCPSolver::visitSelectInst(SelectInst &I) { 905 // If this select returns a struct, just mark the result overdefined. 906 // TODO: We could do a lot better than this if code actually uses this. 907 if (I.getType()->isStructTy()) 908 return (void)markOverdefined(&I); 909 910 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 911 // discover a concrete value later. 912 if (ValueState[&I].isOverdefined()) 913 return (void)markOverdefined(&I); 914 915 ValueLatticeElement CondValue = getValueState(I.getCondition()); 916 if (CondValue.isUnknownOrUndef()) 917 return; 918 919 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 920 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 921 mergeInValue(&I, getValueState(OpVal)); 922 return; 923 } 924 925 // Otherwise, the condition is overdefined or a constant we can't evaluate. 926 // See if we can produce something better than overdefined based on the T/F 927 // value. 928 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 929 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 930 931 bool Changed = ValueState[&I].mergeIn(TVal); 932 Changed |= ValueState[&I].mergeIn(FVal); 933 if (Changed) 934 pushToWorkListMsg(ValueState[&I], &I); 935 } 936 937 // Handle Unary Operators. 938 void SCCPSolver::visitUnaryOperator(Instruction &I) { 939 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 940 941 ValueLatticeElement &IV = ValueState[&I]; 942 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 943 // discover a concrete value later. 944 if (isOverdefined(IV)) 945 return (void)markOverdefined(&I); 946 947 if (isConstant(V0State)) { 948 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 949 950 // op Y -> undef. 951 if (isa<UndefValue>(C)) 952 return; 953 return (void)markConstant(IV, &I, C); 954 } 955 956 // If something is undef, wait for it to resolve. 957 if (!isOverdefined(V0State)) 958 return; 959 960 markOverdefined(&I); 961 } 962 963 // Handle Binary Operators. 964 void SCCPSolver::visitBinaryOperator(Instruction &I) { 965 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 966 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 967 968 ValueLatticeElement &IV = ValueState[&I]; 969 if (IV.isOverdefined()) 970 return; 971 972 // If something is undef, wait for it to resolve. 973 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 974 return; 975 976 if (V1State.isOverdefined() && V2State.isOverdefined()) 977 return (void)markOverdefined(&I); 978 979 // If either of the operands is a constant, try to fold it to a constant. 980 // TODO: Use information from notconstant better. 981 if ((V1State.isConstant() || V2State.isConstant())) { 982 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 983 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 984 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 985 auto *C = dyn_cast_or_null<Constant>(R); 986 if (C) { 987 // X op Y -> undef. 988 if (isa<UndefValue>(C)) 989 return; 990 // Conservatively assume that the result may be based on operands that may 991 // be undef. Note that we use mergeInValue to combine the constant with 992 // the existing lattice value for I, as different constants might be found 993 // after one of the operands go to overdefined, e.g. due to one operand 994 // being a special floating value. 995 ValueLatticeElement NewV; 996 NewV.markConstant(C, /*MayIncludeUndef=*/true); 997 return (void)mergeInValue(&I, NewV); 998 } 999 } 1000 1001 // Only use ranges for binary operators on integers. 1002 if (!I.getType()->isIntegerTy()) 1003 return markOverdefined(&I); 1004 1005 // Try to simplify to a constant range. 1006 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1007 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1008 if (V1State.isConstantRange()) 1009 A = V1State.getConstantRange(); 1010 if (V2State.isConstantRange()) 1011 B = V2State.getConstantRange(); 1012 1013 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1014 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1015 1016 // TODO: Currently we do not exploit special values that produce something 1017 // better than overdefined with an overdefined operand for vector or floating 1018 // point types, like and <4 x i32> overdefined, zeroinitializer. 1019 } 1020 1021 // Handle ICmpInst instruction. 1022 void SCCPSolver::visitCmpInst(CmpInst &I) { 1023 // Do not cache this lookup, getValueState calls later in the function might 1024 // invalidate the reference. 1025 if (isOverdefined(ValueState[&I])) 1026 return (void)markOverdefined(&I); 1027 1028 Value *Op1 = I.getOperand(0); 1029 Value *Op2 = I.getOperand(1); 1030 1031 // For parameters, use ParamState which includes constant range info if 1032 // available. 1033 auto V1State = getValueState(Op1); 1034 auto V2State = getValueState(Op2); 1035 1036 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1037 if (C) { 1038 if (isa<UndefValue>(C)) 1039 return; 1040 ValueLatticeElement CV; 1041 CV.markConstant(C); 1042 mergeInValue(&I, CV); 1043 return; 1044 } 1045 1046 // If operands are still unknown, wait for it to resolve. 1047 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1048 !isConstant(ValueState[&I])) 1049 return; 1050 1051 markOverdefined(&I); 1052 } 1053 1054 // Handle getelementptr instructions. If all operands are constants then we 1055 // can turn this into a getelementptr ConstantExpr. 1056 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1057 if (isOverdefined(ValueState[&I])) 1058 return (void)markOverdefined(&I); 1059 1060 SmallVector<Constant*, 8> Operands; 1061 Operands.reserve(I.getNumOperands()); 1062 1063 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1064 ValueLatticeElement State = getValueState(I.getOperand(i)); 1065 if (State.isUnknownOrUndef()) 1066 return; // Operands are not resolved yet. 1067 1068 if (isOverdefined(State)) 1069 return (void)markOverdefined(&I); 1070 1071 if (Constant *C = getConstant(State)) { 1072 Operands.push_back(C); 1073 continue; 1074 } 1075 1076 return (void)markOverdefined(&I); 1077 } 1078 1079 Constant *Ptr = Operands[0]; 1080 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1081 Constant *C = 1082 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1083 if (isa<UndefValue>(C)) 1084 return; 1085 markConstant(&I, C); 1086 } 1087 1088 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1089 // If this store is of a struct, ignore it. 1090 if (SI.getOperand(0)->getType()->isStructTy()) 1091 return; 1092 1093 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1094 return; 1095 1096 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1097 auto I = TrackedGlobals.find(GV); 1098 if (I == TrackedGlobals.end()) 1099 return; 1100 1101 // Get the value we are storing into the global, then merge it. 1102 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1103 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1104 if (I->second.isOverdefined()) 1105 TrackedGlobals.erase(I); // No need to keep tracking this! 1106 } 1107 1108 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1109 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1110 if (I->getType()->isIntegerTy()) 1111 return ValueLatticeElement::getRange( 1112 getConstantRangeFromMetadata(*Ranges)); 1113 // TODO: Also handle MD_nonnull. 1114 return ValueLatticeElement::getOverdefined(); 1115 } 1116 1117 // Handle load instructions. If the operand is a constant pointer to a constant 1118 // global, we can replace the load with the loaded constant value! 1119 void SCCPSolver::visitLoadInst(LoadInst &I) { 1120 // If this load is of a struct or the load is volatile, just mark the result 1121 // as overdefined. 1122 if (I.getType()->isStructTy() || I.isVolatile()) 1123 return (void)markOverdefined(&I); 1124 1125 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1126 // discover a concrete value later. 1127 if (ValueState[&I].isOverdefined()) 1128 return (void)markOverdefined(&I); 1129 1130 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1131 if (PtrVal.isUnknownOrUndef()) 1132 return; // The pointer is not resolved yet! 1133 1134 ValueLatticeElement &IV = ValueState[&I]; 1135 1136 if (isConstant(PtrVal)) { 1137 Constant *Ptr = getConstant(PtrVal); 1138 1139 // load null is undefined. 1140 if (isa<ConstantPointerNull>(Ptr)) { 1141 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1142 return (void)markOverdefined(IV, &I); 1143 else 1144 return; 1145 } 1146 1147 // Transform load (constant global) into the value loaded. 1148 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1149 if (!TrackedGlobals.empty()) { 1150 // If we are tracking this global, merge in the known value for it. 1151 auto It = TrackedGlobals.find(GV); 1152 if (It != TrackedGlobals.end()) { 1153 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1154 return; 1155 } 1156 } 1157 } 1158 1159 // Transform load from a constant into a constant if possible. 1160 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1161 if (isa<UndefValue>(C)) 1162 return; 1163 return (void)markConstant(IV, &I, C); 1164 } 1165 } 1166 1167 // Fall back to metadata. 1168 mergeInValue(&I, getValueFromMetadata(&I)); 1169 } 1170 1171 void SCCPSolver::visitCallBase(CallBase &CB) { 1172 handleCallResult(CB); 1173 handleCallArguments(CB); 1174 } 1175 1176 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1177 Function *F = CB.getCalledFunction(); 1178 1179 // Void return and not tracking callee, just bail. 1180 if (CB.getType()->isVoidTy()) 1181 return; 1182 1183 // Always mark struct return as overdefined. 1184 if (CB.getType()->isStructTy()) 1185 return (void)markOverdefined(&CB); 1186 1187 // Otherwise, if we have a single return value case, and if the function is 1188 // a declaration, maybe we can constant fold it. 1189 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1190 SmallVector<Constant *, 8> Operands; 1191 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1192 if (AI->get()->getType()->isStructTy()) 1193 return markOverdefined(&CB); // Can't handle struct args. 1194 ValueLatticeElement State = getValueState(*AI); 1195 1196 if (State.isUnknownOrUndef()) 1197 return; // Operands are not resolved yet. 1198 if (isOverdefined(State)) 1199 return (void)markOverdefined(&CB); 1200 assert(isConstant(State) && "Unknown state!"); 1201 Operands.push_back(getConstant(State)); 1202 } 1203 1204 if (isOverdefined(getValueState(&CB))) 1205 return (void)markOverdefined(&CB); 1206 1207 // If we can constant fold this, mark the result of the call as a 1208 // constant. 1209 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1210 // call -> undef. 1211 if (isa<UndefValue>(C)) 1212 return; 1213 return (void)markConstant(&CB, C); 1214 } 1215 } 1216 1217 // Fall back to metadata. 1218 mergeInValue(&CB, getValueFromMetadata(&CB)); 1219 } 1220 1221 void SCCPSolver::handleCallArguments(CallBase &CB) { 1222 Function *F = CB.getCalledFunction(); 1223 // If this is a local function that doesn't have its address taken, mark its 1224 // entry block executable and merge in the actual arguments to the call into 1225 // the formal arguments of the function. 1226 if (!TrackingIncomingArguments.empty() && 1227 TrackingIncomingArguments.count(F)) { 1228 MarkBlockExecutable(&F->front()); 1229 1230 // Propagate information from this call site into the callee. 1231 auto CAI = CB.arg_begin(); 1232 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1233 ++AI, ++CAI) { 1234 // If this argument is byval, and if the function is not readonly, there 1235 // will be an implicit copy formed of the input aggregate. 1236 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1237 markOverdefined(&*AI); 1238 continue; 1239 } 1240 1241 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1242 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1243 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1244 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1245 getMaxWidenStepsOpts()); 1246 } 1247 } else 1248 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1249 } 1250 } 1251 } 1252 1253 void SCCPSolver::handleCallResult(CallBase &CB) { 1254 Function *F = CB.getCalledFunction(); 1255 1256 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1257 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1258 if (ValueState[&CB].isOverdefined()) 1259 return; 1260 1261 Value *CopyOf = CB.getOperand(0); 1262 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1263 auto *PI = getPredicateInfoFor(&CB); 1264 assert(PI && "Missing predicate info for ssa.copy"); 1265 1266 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1267 if (!Constraint) { 1268 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1269 return; 1270 } 1271 1272 CmpInst::Predicate Pred = Constraint->Predicate; 1273 Value *OtherOp = Constraint->OtherOp; 1274 1275 // Wait until OtherOp is resolved. 1276 if (getValueState(OtherOp).isUnknown()) { 1277 addAdditionalUser(OtherOp, &CB); 1278 return; 1279 } 1280 1281 ValueLatticeElement CondVal = getValueState(OtherOp); 1282 ValueLatticeElement &IV = ValueState[&CB]; 1283 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1284 auto ImposedCR = 1285 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1286 1287 // Get the range imposed by the condition. 1288 if (CondVal.isConstantRange()) 1289 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1290 Pred, CondVal.getConstantRange()); 1291 1292 // Combine range info for the original value with the new range from the 1293 // condition. 1294 auto CopyOfCR = CopyOfVal.isConstantRange() 1295 ? CopyOfVal.getConstantRange() 1296 : ConstantRange::getFull( 1297 DL.getTypeSizeInBits(CopyOf->getType())); 1298 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1299 // If the existing information is != x, do not use the information from 1300 // a chained predicate, as the != x information is more likely to be 1301 // helpful in practice. 1302 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1303 NewCR = CopyOfCR; 1304 1305 addAdditionalUser(OtherOp, &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(OtherOp, &CB); 1323 mergeInValue(IV, &CB, CondVal); 1324 return; 1325 } 1326 1327 return (void)mergeInValue(IV, &CB, CopyOfVal); 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 static bool removeNonFeasibleEdges(const SCCPSolver &Solver, BasicBlock *BB, 1811 DomTreeUpdater &DTU) { 1812 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors; 1813 bool HasNonFeasibleEdges = false; 1814 for (BasicBlock *Succ : successors(BB)) { 1815 if (Solver.isEdgeFeasible(BB, Succ)) 1816 FeasibleSuccessors.insert(Succ); 1817 else 1818 HasNonFeasibleEdges = true; 1819 } 1820 1821 // All edges feasible, nothing to do. 1822 if (!HasNonFeasibleEdges) 1823 return false; 1824 1825 // SCCP can only determine non-feasible edges for br, switch and indirectbr. 1826 Instruction *TI = BB->getTerminator(); 1827 assert((isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1828 isa<IndirectBrInst>(TI)) && 1829 "Terminator must be a br, switch or indirectbr"); 1830 1831 if (FeasibleSuccessors.size() == 1) { 1832 // Replace with an unconditional branch to the only feasible successor. 1833 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin(); 1834 SmallVector<DominatorTree::UpdateType, 8> Updates; 1835 bool HaveSeenOnlyFeasibleSuccessor = false; 1836 for (BasicBlock *Succ : successors(BB)) { 1837 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) { 1838 // Don't remove the edge to the only feasible successor the first time 1839 // we see it. We still do need to remove any multi-edges to it though. 1840 HaveSeenOnlyFeasibleSuccessor = true; 1841 continue; 1842 } 1843 1844 Succ->removePredecessor(BB); 1845 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1846 } 1847 1848 BranchInst::Create(OnlyFeasibleSuccessor, BB); 1849 TI->eraseFromParent(); 1850 DTU.applyUpdatesPermissive(Updates); 1851 } else { 1852 llvm_unreachable("Either all successors are feasible, or exactly one is"); 1853 } 1854 return true; 1855 } 1856 1857 bool llvm::runIPSCCP( 1858 Module &M, const DataLayout &DL, 1859 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1860 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1861 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1862 1863 // Loop over all functions, marking arguments to those with their addresses 1864 // taken or that are external as overdefined. 1865 for (Function &F : M) { 1866 if (F.isDeclaration()) 1867 continue; 1868 1869 Solver.addAnalysis(F, getAnalysis(F)); 1870 1871 // Determine if we can track the function's return values. If so, add the 1872 // function to the solver's set of return-tracked functions. 1873 if (canTrackReturnsInterprocedurally(&F)) 1874 Solver.AddTrackedFunction(&F); 1875 1876 // Determine if we can track the function's arguments. If so, add the 1877 // function to the solver's set of argument-tracked functions. 1878 if (canTrackArgumentsInterprocedurally(&F)) { 1879 Solver.AddArgumentTrackedFunction(&F); 1880 continue; 1881 } 1882 1883 // Assume the function is called. 1884 Solver.MarkBlockExecutable(&F.front()); 1885 1886 // Assume nothing about the incoming arguments. 1887 for (Argument &AI : F.args()) 1888 Solver.markOverdefined(&AI); 1889 } 1890 1891 // Determine if we can track any of the module's global variables. If so, add 1892 // the global variables we can track to the solver's set of tracked global 1893 // variables. 1894 for (GlobalVariable &G : M.globals()) { 1895 G.removeDeadConstantUsers(); 1896 if (canTrackGlobalVariableInterprocedurally(&G)) 1897 Solver.TrackValueOfGlobalVariable(&G); 1898 } 1899 1900 // Solve for constants. 1901 bool ResolvedUndefs = true; 1902 Solver.Solve(); 1903 while (ResolvedUndefs) { 1904 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1905 ResolvedUndefs = false; 1906 for (Function &F : M) 1907 if (Solver.ResolvedUndefsIn(F)) { 1908 // We run Solve() after we resolved an undef in a function, because 1909 // we might deduce a fact that eliminates an undef in another function. 1910 Solver.Solve(); 1911 ResolvedUndefs = true; 1912 } 1913 } 1914 1915 bool MadeChanges = false; 1916 1917 // Iterate over all of the instructions in the module, replacing them with 1918 // constants if we have found them to be of constant values. 1919 1920 for (Function &F : M) { 1921 if (F.isDeclaration()) 1922 continue; 1923 1924 SmallVector<BasicBlock *, 512> BlocksToErase; 1925 1926 if (Solver.isBlockExecutable(&F.front())) { 1927 bool ReplacedPointerArg = false; 1928 for (Argument &Arg : F.args()) { 1929 if (!Arg.use_empty() && tryToReplaceWithConstant(Solver, &Arg)) { 1930 ReplacedPointerArg |= Arg.getType()->isPointerTy(); 1931 ++IPNumArgsElimed; 1932 } 1933 } 1934 1935 // If we replaced an argument, the argmemonly and 1936 // inaccessiblemem_or_argmemonly attributes do not hold any longer. Remove 1937 // them from both the function and callsites. 1938 if (ReplacedPointerArg) { 1939 SmallVector<Attribute::AttrKind, 2> AttributesToRemove = { 1940 Attribute::ArgMemOnly, Attribute::InaccessibleMemOrArgMemOnly}; 1941 for (auto Attr : AttributesToRemove) 1942 F.removeFnAttr(Attr); 1943 1944 for (User *U : F.users()) { 1945 auto *CB = dyn_cast<CallBase>(U); 1946 if (!CB || CB->getCalledFunction() != &F) 1947 continue; 1948 1949 for (auto Attr : AttributesToRemove) 1950 CB->removeAttribute(AttributeList::FunctionIndex, Attr); 1951 } 1952 } 1953 } 1954 1955 SmallPtrSet<Value *, 32> InsertedValues; 1956 for (BasicBlock &BB : F) { 1957 if (!Solver.isBlockExecutable(&BB)) { 1958 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1959 ++NumDeadBlocks; 1960 1961 MadeChanges = true; 1962 1963 if (&BB != &F.front()) 1964 BlocksToErase.push_back(&BB); 1965 continue; 1966 } 1967 1968 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1969 IPNumInstRemoved, IPNumInstReplaced); 1970 } 1971 1972 DomTreeUpdater DTU = Solver.getDTU(F); 1973 // Change dead blocks to unreachable. We do it after replacing constants 1974 // in all executable blocks, because changeToUnreachable may remove PHI 1975 // nodes in executable blocks we found values for. The function's entry 1976 // block is not part of BlocksToErase, so we have to handle it separately. 1977 for (BasicBlock *BB : BlocksToErase) { 1978 NumInstRemoved += 1979 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 1980 /*PreserveLCSSA=*/false, &DTU); 1981 } 1982 if (!Solver.isBlockExecutable(&F.front())) 1983 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 1984 /*UseLLVMTrap=*/false, 1985 /*PreserveLCSSA=*/false, &DTU); 1986 1987 for (BasicBlock &BB : F) 1988 MadeChanges |= removeNonFeasibleEdges(Solver, &BB, DTU); 1989 1990 for (BasicBlock *DeadBB : BlocksToErase) 1991 DTU.deleteBB(DeadBB); 1992 1993 for (BasicBlock &BB : F) { 1994 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 1995 Instruction *Inst = &*BI++; 1996 if (Solver.getPredicateInfoFor(Inst)) { 1997 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 1998 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1999 Value *Op = II->getOperand(0); 2000 Inst->replaceAllUsesWith(Op); 2001 Inst->eraseFromParent(); 2002 } 2003 } 2004 } 2005 } 2006 } 2007 } 2008 2009 // If we inferred constant or undef return values for a function, we replaced 2010 // all call uses with the inferred value. This means we don't need to bother 2011 // actually returning anything from the function. Replace all return 2012 // instructions with return undef. 2013 // 2014 // Do this in two stages: first identify the functions we should process, then 2015 // actually zap their returns. This is important because we can only do this 2016 // if the address of the function isn't taken. In cases where a return is the 2017 // last use of a function, the order of processing functions would affect 2018 // whether other functions are optimizable. 2019 SmallVector<ReturnInst*, 8> ReturnsToZap; 2020 2021 for (const auto &I : Solver.getTrackedRetVals()) { 2022 Function *F = I.first; 2023 const ValueLatticeElement &ReturnValue = I.second; 2024 2025 // If there is a known constant range for the return value, add !range 2026 // metadata to the function's call sites. 2027 if (ReturnValue.isConstantRange() && 2028 !ReturnValue.getConstantRange().isSingleElement()) { 2029 // Do not add range metadata if the return value may include undef. 2030 if (ReturnValue.isConstantRangeIncludingUndef()) 2031 continue; 2032 2033 auto &CR = ReturnValue.getConstantRange(); 2034 for (User *User : F->users()) { 2035 auto *CB = dyn_cast<CallBase>(User); 2036 if (!CB || CB->getCalledFunction() != F) 2037 continue; 2038 2039 // Limit to cases where the return value is guaranteed to be neither 2040 // poison nor undef. Poison will be outside any range and currently 2041 // values outside of the specified range cause immediate undefined 2042 // behavior. 2043 if (!isGuaranteedNotToBeUndefOrPoison(CB, CB)) 2044 continue; 2045 2046 // Do not touch existing metadata for now. 2047 // TODO: We should be able to take the intersection of the existing 2048 // metadata and the inferred range. 2049 if (CB->getMetadata(LLVMContext::MD_range)) 2050 continue; 2051 2052 LLVMContext &Context = CB->getParent()->getContext(); 2053 Metadata *RangeMD[] = { 2054 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getLower())), 2055 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getUpper()))}; 2056 CB->setMetadata(LLVMContext::MD_range, MDNode::get(Context, RangeMD)); 2057 } 2058 continue; 2059 } 2060 if (F->getReturnType()->isVoidTy()) 2061 continue; 2062 if (isConstant(ReturnValue) || ReturnValue.isUnknownOrUndef()) 2063 findReturnsToZap(*F, ReturnsToZap, Solver); 2064 } 2065 2066 for (auto F : Solver.getMRVFunctionsTracked()) { 2067 assert(F->getReturnType()->isStructTy() && 2068 "The return type should be a struct"); 2069 StructType *STy = cast<StructType>(F->getReturnType()); 2070 if (Solver.isStructLatticeConstant(F, STy)) 2071 findReturnsToZap(*F, ReturnsToZap, Solver); 2072 } 2073 2074 // Zap all returns which we've identified as zap to change. 2075 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2076 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2077 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2078 } 2079 2080 // If we inferred constant or undef values for globals variables, we can 2081 // delete the global and any stores that remain to it. 2082 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2083 GlobalVariable *GV = I.first; 2084 if (isOverdefined(I.second)) 2085 continue; 2086 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2087 << "' is constant!\n"); 2088 while (!GV->use_empty()) { 2089 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2090 SI->eraseFromParent(); 2091 MadeChanges = true; 2092 } 2093 M.getGlobalList().erase(GV); 2094 ++IPNumGlobalConst; 2095 } 2096 2097 return MadeChanges; 2098 } 2099