1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements sparse conditional constant propagation and merging: 10 // 11 // Specifically, this: 12 // * Assumes values are constant unless proven otherwise 13 // * Assumes BasicBlocks are dead unless proven otherwise 14 // * Proves values to be constant, and replaces them with constants 15 // * Proves conditional branches to be unconditional 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/Transforms/Scalar/SCCP.h" 20 #include "llvm/ADT/ArrayRef.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/DenseSet.h" 23 #include "llvm/ADT/MapVector.h" 24 #include "llvm/ADT/PointerIntPair.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SetVector.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/Analysis/ConstantFolding.h" 31 #include "llvm/Analysis/DomTreeUpdater.h" 32 #include "llvm/Analysis/GlobalsModRef.h" 33 #include "llvm/Analysis/InstructionSimplify.h" 34 #include "llvm/Analysis/TargetLibraryInfo.h" 35 #include "llvm/Analysis/ValueLattice.h" 36 #include "llvm/Analysis/ValueLatticeUtils.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/IR/BasicBlock.h" 39 #include "llvm/IR/Constant.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/DerivedTypes.h" 43 #include "llvm/IR/Function.h" 44 #include "llvm/IR/GlobalVariable.h" 45 #include "llvm/IR/InstVisitor.h" 46 #include "llvm/IR/InstrTypes.h" 47 #include "llvm/IR/Instruction.h" 48 #include "llvm/IR/Instructions.h" 49 #include "llvm/IR/Module.h" 50 #include "llvm/IR/PassManager.h" 51 #include "llvm/IR/Type.h" 52 #include "llvm/IR/User.h" 53 #include "llvm/IR/Value.h" 54 #include "llvm/InitializePasses.h" 55 #include "llvm/Pass.h" 56 #include "llvm/Support/Casting.h" 57 #include "llvm/Support/Debug.h" 58 #include "llvm/Support/ErrorHandling.h" 59 #include "llvm/Support/raw_ostream.h" 60 #include "llvm/Transforms/Scalar.h" 61 #include "llvm/Transforms/Utils/Local.h" 62 #include "llvm/Transforms/Utils/PredicateInfo.h" 63 #include <cassert> 64 #include <utility> 65 #include <vector> 66 67 using namespace llvm; 68 69 #define DEBUG_TYPE "sccp" 70 71 STATISTIC(NumInstRemoved, "Number of instructions removed"); 72 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable"); 73 STATISTIC(NumInstReplaced, 74 "Number of instructions replaced with (simpler) instruction"); 75 76 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP"); 77 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP"); 78 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP"); 79 STATISTIC( 80 IPNumInstReplaced, 81 "Number of instructions replaced with (simpler) instruction by IPSCCP"); 82 83 // The maximum number of range extensions allowed for operations requiring 84 // widening. 85 static const unsigned MaxNumRangeExtensions = 10; 86 87 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 88 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 89 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 90 MaxNumRangeExtensions); 91 } 92 namespace { 93 94 // Helper to check if \p LV is either a constant or a constant 95 // range with a single element. This should cover exactly the same cases as the 96 // old ValueLatticeElement::isConstant() and is intended to be used in the 97 // transition to ValueLatticeElement. 98 bool isConstant(const ValueLatticeElement &LV) { 99 return LV.isConstant() || 100 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 101 } 102 103 // Helper to check if \p LV is either overdefined or a constant range with more 104 // than a single element. This should cover exactly the same cases as the old 105 // ValueLatticeElement::isOverdefined() and is intended to be used in the 106 // transition to ValueLatticeElement. 107 bool isOverdefined(const ValueLatticeElement &LV) { 108 return !LV.isUnknownOrUndef() && !isConstant(LV); 109 } 110 111 //===----------------------------------------------------------------------===// 112 // 113 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional 114 /// Constant Propagation. 115 /// 116 class SCCPSolver : public InstVisitor<SCCPSolver> { 117 const DataLayout &DL; 118 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 119 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 120 DenseMap<Value *, ValueLatticeElement> 121 ValueState; // The state each value is in. 122 123 /// StructValueState - This maintains ValueState for values that have 124 /// StructType, for example for formal arguments, calls, insertelement, etc. 125 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 126 127 /// GlobalValue - If we are tracking any values for the contents of a global 128 /// variable, we keep a mapping from the constant accessor to the element of 129 /// the global, to the currently known value. If the value becomes 130 /// overdefined, it's entry is simply removed from this map. 131 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 132 133 /// TrackedRetVals - If we are tracking arguments into and the return 134 /// value out of a function, it will have an entry in this map, indicating 135 /// what the known return value for the function is. 136 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 137 138 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 139 /// that return multiple values. 140 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 141 TrackedMultipleRetVals; 142 143 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 144 /// represented here for efficient lookup. 145 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 146 147 /// MustTailFunctions - Each function here is a callee of non-removable 148 /// musttail call site. 149 SmallPtrSet<Function *, 16> MustTailCallees; 150 151 /// TrackingIncomingArguments - This is the set of functions for whose 152 /// arguments we make optimistic assumptions about and try to prove as 153 /// constants. 154 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 155 156 /// The reason for two worklists is that overdefined is the lowest state 157 /// on the lattice, and moving things to overdefined as fast as possible 158 /// makes SCCP converge much faster. 159 /// 160 /// By having a separate worklist, we accomplish this because everything 161 /// possibly overdefined will become overdefined at the soonest possible 162 /// point. 163 SmallVector<Value *, 64> OverdefinedInstWorkList; 164 SmallVector<Value *, 64> InstWorkList; 165 166 // The BasicBlock work list 167 SmallVector<BasicBlock *, 64> BBWorkList; 168 169 /// KnownFeasibleEdges - Entries in this set are edges which have already had 170 /// PHI nodes retriggered. 171 using Edge = std::pair<BasicBlock *, BasicBlock *>; 172 DenseSet<Edge> KnownFeasibleEdges; 173 174 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 175 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 176 177 LLVMContext &Ctx; 178 179 public: 180 void addAnalysis(Function &F, AnalysisResultsForFn A) { 181 AnalysisResults.insert({&F, std::move(A)}); 182 } 183 184 const PredicateBase *getPredicateInfoFor(Instruction *I) { 185 auto A = AnalysisResults.find(I->getParent()->getParent()); 186 if (A == AnalysisResults.end()) 187 return nullptr; 188 return A->second.PredInfo->getPredicateInfoFor(I); 189 } 190 191 DomTreeUpdater getDTU(Function &F) { 192 auto A = AnalysisResults.find(&F); 193 assert(A != AnalysisResults.end() && "Need analysis results for function."); 194 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 195 } 196 197 SCCPSolver(const DataLayout &DL, 198 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 199 LLVMContext &Ctx) 200 : DL(DL), GetTLI(std::move(GetTLI)), Ctx(Ctx) {} 201 202 /// MarkBlockExecutable - This method can be used by clients to mark all of 203 /// the blocks that are known to be intrinsically live in the processed unit. 204 /// 205 /// This returns true if the block was not considered live before. 206 bool MarkBlockExecutable(BasicBlock *BB) { 207 if (!BBExecutable.insert(BB).second) 208 return false; 209 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 210 BBWorkList.push_back(BB); // Add the block to the work list! 211 return true; 212 } 213 214 /// TrackValueOfGlobalVariable - Clients can use this method to 215 /// inform the SCCPSolver that it should track loads and stores to the 216 /// specified global variable if it can. This is only legal to call if 217 /// performing Interprocedural SCCP. 218 void TrackValueOfGlobalVariable(GlobalVariable *GV) { 219 // We only track the contents of scalar globals. 220 if (GV->getValueType()->isSingleValueType()) { 221 ValueLatticeElement &IV = TrackedGlobals[GV]; 222 if (!isa<UndefValue>(GV->getInitializer())) 223 IV.markConstant(GV->getInitializer()); 224 } 225 } 226 227 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into 228 /// and out of the specified function (which cannot have its address taken), 229 /// this method must be called. 230 void AddTrackedFunction(Function *F) { 231 // Add an entry, F -> undef. 232 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 233 MRVFunctionsTracked.insert(F); 234 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 235 TrackedMultipleRetVals.insert( 236 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 237 } else if (!F->getReturnType()->isVoidTy()) 238 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 239 } 240 241 /// AddMustTailCallee - If the SCCP solver finds that this function is called 242 /// from non-removable musttail call site. 243 void AddMustTailCallee(Function *F) { 244 MustTailCallees.insert(F); 245 } 246 247 /// Returns true if the given function is called from non-removable musttail 248 /// call site. 249 bool isMustTailCallee(Function *F) { 250 return MustTailCallees.count(F); 251 } 252 253 void AddArgumentTrackedFunction(Function *F) { 254 TrackingIncomingArguments.insert(F); 255 } 256 257 /// Returns true if the given function is in the solver's set of 258 /// argument-tracked functions. 259 bool isArgumentTrackedFunction(Function *F) { 260 return TrackingIncomingArguments.count(F); 261 } 262 263 /// Solve - Solve for constants and executable blocks. 264 void Solve(); 265 266 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 267 /// that branches on undef values cannot reach any of their successors. 268 /// However, this is not a safe assumption. After we solve dataflow, this 269 /// method should be use to handle this. If this returns true, the solver 270 /// should be rerun. 271 bool ResolvedUndefsIn(Function &F); 272 273 bool isBlockExecutable(BasicBlock *BB) const { 274 return BBExecutable.count(BB); 275 } 276 277 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 278 // block to the 'To' basic block is currently feasible. 279 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 280 281 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 282 std::vector<ValueLatticeElement> StructValues; 283 auto *STy = dyn_cast<StructType>(V->getType()); 284 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 285 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 286 auto I = StructValueState.find(std::make_pair(V, i)); 287 assert(I != StructValueState.end() && "Value not in valuemap!"); 288 StructValues.push_back(I->second); 289 } 290 return StructValues; 291 } 292 293 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 294 295 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 296 assert(!V->getType()->isStructTy() && 297 "Should use getStructLatticeValueFor"); 298 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 299 ValueState.find(V); 300 assert(I != ValueState.end() && 301 "V not found in ValueState nor Paramstate map!"); 302 return I->second; 303 } 304 305 /// getTrackedRetVals - Get the inferred return value map. 306 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 307 return TrackedRetVals; 308 } 309 310 /// getTrackedGlobals - Get and return the set of inferred initializers for 311 /// global variables. 312 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 313 return TrackedGlobals; 314 } 315 316 /// getMRVFunctionsTracked - Get the set of functions which return multiple 317 /// values tracked by the pass. 318 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 319 return MRVFunctionsTracked; 320 } 321 322 /// getMustTailCallees - Get the set of functions which are called 323 /// from non-removable musttail call sites. 324 const SmallPtrSet<Function *, 16> getMustTailCallees() { 325 return MustTailCallees; 326 } 327 328 /// markOverdefined - Mark the specified value overdefined. This 329 /// works with both scalars and structs. 330 void markOverdefined(Value *V) { 331 if (auto *STy = dyn_cast<StructType>(V->getType())) 332 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 333 markOverdefined(getStructValueState(V, i), V); 334 else 335 markOverdefined(ValueState[V], V); 336 } 337 338 // isStructLatticeConstant - Return true if all the lattice values 339 // corresponding to elements of the structure are constants, 340 // false otherwise. 341 bool isStructLatticeConstant(Function *F, StructType *STy) { 342 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 343 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 344 assert(It != TrackedMultipleRetVals.end()); 345 ValueLatticeElement LV = It->second; 346 if (!isConstant(LV)) 347 return false; 348 } 349 return true; 350 } 351 352 /// Helper to return a Constant if \p LV is either a constant or a constant 353 /// range with a single element. 354 Constant *getConstant(const ValueLatticeElement &LV) const { 355 if (LV.isConstant()) 356 return LV.getConstant(); 357 358 if (LV.isConstantRange()) { 359 auto &CR = LV.getConstantRange(); 360 if (CR.getSingleElement()) 361 return ConstantInt::get(Ctx, *CR.getSingleElement()); 362 } 363 return nullptr; 364 } 365 366 private: 367 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 368 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 369 } 370 371 // pushToWorkList - Helper for markConstant/markOverdefined 372 void pushToWorkList(ValueLatticeElement &IV, Value *V) { 373 if (IV.isOverdefined()) 374 return OverdefinedInstWorkList.push_back(V); 375 InstWorkList.push_back(V); 376 } 377 378 // Helper to push \p V to the worklist, after updating it to \p IV. Also 379 // prints a debug message with the updated value. 380 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 381 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 382 pushToWorkList(IV, V); 383 } 384 385 // markConstant - Make a value be marked as "constant". If the value 386 // is not already a constant, add it to the instruction work list so that 387 // the users of the instruction are updated later. 388 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 389 bool MayIncludeUndef = false) { 390 if (!IV.markConstant(C, MayIncludeUndef)) 391 return false; 392 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 393 pushToWorkList(IV, V); 394 return true; 395 } 396 397 bool markConstant(Value *V, Constant *C) { 398 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 399 return markConstant(ValueState[V], V, C); 400 } 401 402 // markOverdefined - Make a value be marked as "overdefined". If the 403 // value is not already overdefined, add it to the overdefined instruction 404 // work list so that the users of the instruction are updated later. 405 bool markOverdefined(ValueLatticeElement &IV, Value *V) { 406 if (!IV.markOverdefined()) return false; 407 408 LLVM_DEBUG(dbgs() << "markOverdefined: "; 409 if (auto *F = dyn_cast<Function>(V)) dbgs() 410 << "Function '" << F->getName() << "'\n"; 411 else dbgs() << *V << '\n'); 412 // Only instructions go on the work list 413 pushToWorkList(IV, V); 414 return true; 415 } 416 417 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 418 /// changes. 419 bool mergeInValue(ValueLatticeElement &IV, Value *V, 420 ValueLatticeElement MergeWithV, 421 ValueLatticeElement::MergeOptions Opts = { 422 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 423 if (IV.mergeIn(MergeWithV, Opts)) { 424 pushToWorkList(IV, V); 425 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 426 << IV << "\n"); 427 return true; 428 } 429 return false; 430 } 431 432 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 433 ValueLatticeElement::MergeOptions Opts = { 434 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 435 assert(!V->getType()->isStructTy() && 436 "non-structs should use markConstant"); 437 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 438 } 439 440 /// getValueState - Return the ValueLatticeElement object that corresponds to 441 /// the value. This function handles the case when the value hasn't been seen 442 /// yet by properly seeding constants etc. 443 ValueLatticeElement &getValueState(Value *V) { 444 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 445 446 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 447 ValueLatticeElement &LV = I.first->second; 448 449 if (!I.second) 450 return LV; // Common case, already in the map. 451 452 if (auto *C = dyn_cast<Constant>(V)) 453 LV.markConstant(C); // Constants are constant 454 455 // All others are unknown by default. 456 return LV; 457 } 458 459 /// getStructValueState - Return the ValueLatticeElement object that 460 /// corresponds to the value/field pair. This function handles the case when 461 /// the value hasn't been seen yet by properly seeding constants etc. 462 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 463 assert(V->getType()->isStructTy() && "Should use getValueState"); 464 assert(i < cast<StructType>(V->getType())->getNumElements() && 465 "Invalid element #"); 466 467 auto I = StructValueState.insert( 468 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 469 ValueLatticeElement &LV = I.first->second; 470 471 if (!I.second) 472 return LV; // Common case, already in the map. 473 474 if (auto *C = dyn_cast<Constant>(V)) { 475 Constant *Elt = C->getAggregateElement(i); 476 477 if (!Elt) 478 LV.markOverdefined(); // Unknown sort of constant. 479 else if (isa<UndefValue>(Elt)) 480 ; // Undef values remain unknown. 481 else 482 LV.markConstant(Elt); // Constants are constant. 483 } 484 485 // All others are underdefined by default. 486 return LV; 487 } 488 489 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 490 /// work list if it is not already executable. 491 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 492 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 493 return false; // This edge is already known to be executable! 494 495 if (!MarkBlockExecutable(Dest)) { 496 // If the destination is already executable, we just made an *edge* 497 // feasible that wasn't before. Revisit the PHI nodes in the block 498 // because they have potentially new operands. 499 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 500 << " -> " << Dest->getName() << '\n'); 501 502 for (PHINode &PN : Dest->phis()) 503 visitPHINode(PN); 504 } 505 return true; 506 } 507 508 // getFeasibleSuccessors - Return a vector of booleans to indicate which 509 // successors are reachable from a given terminator instruction. 510 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 511 512 // OperandChangedState - This method is invoked on all of the users of an 513 // instruction that was just changed state somehow. Based on this 514 // information, we need to update the specified user of this instruction. 515 void OperandChangedState(Instruction *I) { 516 if (BBExecutable.count(I->getParent())) // Inst is executable? 517 visit(*I); 518 } 519 520 // Add U as additional user of V. 521 void addAdditionalUser(Value *V, User *U) { 522 auto Iter = AdditionalUsers.insert({V, {}}); 523 Iter.first->second.insert(U); 524 } 525 526 // Mark I's users as changed, including AdditionalUsers. 527 void markUsersAsChanged(Value *I) { 528 // Functions include their arguments in the use-list. Changed function 529 // values mean that the result of the function changed. We only need to 530 // update the call sites with the new function result and do not have to 531 // propagate the call arguments. 532 if (isa<Function>(I)) { 533 for (User *U : I->users()) { 534 if (auto *CB = dyn_cast<CallBase>(U)) 535 handleCallResult(*CB); 536 } 537 } else { 538 for (User *U : I->users()) 539 if (auto *UI = dyn_cast<Instruction>(U)) 540 OperandChangedState(UI); 541 } 542 543 auto Iter = AdditionalUsers.find(I); 544 if (Iter != AdditionalUsers.end()) { 545 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 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 653 if (ConstantInt *CI = getConstantInt(SCValue)) { 654 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 655 return; 656 } 657 658 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 659 // is ready. 660 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 661 const ConstantRange &Range = SCValue.getConstantRange(); 662 for (const auto &Case : SI->cases()) { 663 const APInt &CaseValue = Case.getCaseValue()->getValue(); 664 if (Range.contains(CaseValue)) 665 Succs[Case.getSuccessorIndex()] = true; 666 } 667 668 // TODO: Determine whether default case is reachable. 669 Succs[SI->case_default()->getSuccessorIndex()] = true; 670 return; 671 } 672 673 // Overdefined or unknown condition? All destinations are executable! 674 if (!SCValue.isUnknownOrUndef()) 675 Succs.assign(TI.getNumSuccessors(), true); 676 return; 677 } 678 679 // In case of indirect branch and its address is a blockaddress, we mark 680 // the target as executable. 681 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 682 // Casts are folded by visitCastInst. 683 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 684 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 685 if (!Addr) { // Overdefined or unknown condition? 686 // All destinations are executable! 687 if (!IBRValue.isUnknownOrUndef()) 688 Succs.assign(TI.getNumSuccessors(), true); 689 return; 690 } 691 692 BasicBlock* T = Addr->getBasicBlock(); 693 assert(Addr->getFunction() == T->getParent() && 694 "Block address of a different function ?"); 695 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 696 // This is the target. 697 if (IBR->getDestination(i) == T) { 698 Succs[i] = true; 699 return; 700 } 701 } 702 703 // If we didn't find our destination in the IBR successor list, then we 704 // have undefined behavior. Its ok to assume no successor is executable. 705 return; 706 } 707 708 // In case of callbr, we pessimistically assume that all successors are 709 // feasible. 710 if (isa<CallBrInst>(&TI)) { 711 Succs.assign(TI.getNumSuccessors(), true); 712 return; 713 } 714 715 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 716 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 717 } 718 719 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 720 // block to the 'To' basic block is currently feasible. 721 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 722 // Check if we've called markEdgeExecutable on the edge yet. (We could 723 // be more aggressive and try to consider edges which haven't been marked 724 // yet, but there isn't any need.) 725 return KnownFeasibleEdges.count(Edge(From, To)); 726 } 727 728 // visit Implementations - Something changed in this instruction, either an 729 // operand made a transition, or the instruction is newly executable. Change 730 // the value type of I to reflect these changes if appropriate. This method 731 // makes sure to do the following actions: 732 // 733 // 1. If a phi node merges two constants in, and has conflicting value coming 734 // from different branches, or if the PHI node merges in an overdefined 735 // value, then the PHI node becomes overdefined. 736 // 2. If a phi node merges only constants in, and they all agree on value, the 737 // PHI node becomes a constant value equal to that. 738 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 739 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 740 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 741 // 6. If a conditional branch has a value that is constant, make the selected 742 // destination executable 743 // 7. If a conditional branch has a value that is overdefined, make all 744 // successors executable. 745 void SCCPSolver::visitPHINode(PHINode &PN) { 746 // If this PN returns a struct, just mark the result overdefined. 747 // TODO: We could do a lot better than this if code actually uses this. 748 if (PN.getType()->isStructTy()) 749 return (void)markOverdefined(&PN); 750 751 if (getValueState(&PN).isOverdefined()) 752 return; // Quick exit 753 754 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 755 // and slow us down a lot. Just mark them overdefined. 756 if (PN.getNumIncomingValues() > 64) 757 return (void)markOverdefined(&PN); 758 759 unsigned NumActiveIncoming = 0; 760 761 // Look at all of the executable operands of the PHI node. If any of them 762 // are overdefined, the PHI becomes overdefined as well. If they are all 763 // constant, and they agree with each other, the PHI becomes the identical 764 // constant. If they are constant and don't agree, the PHI is a constant 765 // range. If there are no executable operands, the PHI remains unknown. 766 ValueLatticeElement PhiState = getValueState(&PN); 767 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 768 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 769 continue; 770 771 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 772 PhiState.mergeIn(IV); 773 NumActiveIncoming++; 774 if (PhiState.isOverdefined()) 775 break; 776 } 777 778 // We allow up to 1 range extension per active incoming value and one 779 // additional extension. Note that we manually adjust the number of range 780 // extensions to match the number of active incoming values. This helps to 781 // limit multiple extensions caused by the same incoming value, if other 782 // incoming values are equal. 783 mergeInValue(&PN, PhiState, 784 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 785 NumActiveIncoming + 1)); 786 ValueLatticeElement &PhiStateRef = getValueState(&PN); 787 PhiStateRef.setNumRangeExtensions( 788 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 789 } 790 791 void SCCPSolver::visitReturnInst(ReturnInst &I) { 792 if (I.getNumOperands() == 0) return; // ret void 793 794 Function *F = I.getParent()->getParent(); 795 Value *ResultOp = I.getOperand(0); 796 797 // If we are tracking the return value of this function, merge it in. 798 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 799 auto TFRVI = TrackedRetVals.find(F); 800 if (TFRVI != TrackedRetVals.end()) { 801 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 802 return; 803 } 804 } 805 806 // Handle functions that return multiple values. 807 if (!TrackedMultipleRetVals.empty()) { 808 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 809 if (MRVFunctionsTracked.count(F)) 810 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 811 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 812 getStructValueState(ResultOp, i)); 813 } 814 } 815 816 void SCCPSolver::visitTerminator(Instruction &TI) { 817 SmallVector<bool, 16> SuccFeasible; 818 getFeasibleSuccessors(TI, SuccFeasible); 819 820 BasicBlock *BB = TI.getParent(); 821 822 // Mark all feasible successors executable. 823 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 824 if (SuccFeasible[i]) 825 markEdgeExecutable(BB, TI.getSuccessor(i)); 826 } 827 828 void SCCPSolver::visitCastInst(CastInst &I) { 829 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 830 // discover a concrete value later. 831 if (ValueState[&I].isOverdefined()) 832 return; 833 834 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 835 if (Constant *OpC = getConstant(OpSt)) { 836 // Fold the constant as we build. 837 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 838 if (isa<UndefValue>(C)) 839 return; 840 // Propagate constant value 841 markConstant(&I, C); 842 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 843 auto &LV = getValueState(&I); 844 ConstantRange OpRange = OpSt.getConstantRange(); 845 Type *DestTy = I.getDestTy(); 846 ConstantRange Res = 847 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 848 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 849 } else if (!OpSt.isUnknownOrUndef()) 850 markOverdefined(&I); 851 } 852 853 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) { 854 // If this returns a struct, mark all elements over defined, we don't track 855 // structs in structs. 856 if (EVI.getType()->isStructTy()) 857 return (void)markOverdefined(&EVI); 858 859 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 860 // discover a concrete value later. 861 if (ValueState[&EVI].isOverdefined()) 862 return (void)markOverdefined(&EVI); 863 864 // If this is extracting from more than one level of struct, we don't know. 865 if (EVI.getNumIndices() != 1) 866 return (void)markOverdefined(&EVI); 867 868 Value *AggVal = EVI.getAggregateOperand(); 869 if (AggVal->getType()->isStructTy()) { 870 unsigned i = *EVI.idx_begin(); 871 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 872 mergeInValue(getValueState(&EVI), &EVI, EltVal); 873 } else { 874 // Otherwise, must be extracting from an array. 875 return (void)markOverdefined(&EVI); 876 } 877 } 878 879 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) { 880 auto *STy = dyn_cast<StructType>(IVI.getType()); 881 if (!STy) 882 return (void)markOverdefined(&IVI); 883 884 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 885 // discover a concrete value later. 886 if (isOverdefined(ValueState[&IVI])) 887 return (void)markOverdefined(&IVI); 888 889 // If this has more than one index, we can't handle it, drive all results to 890 // undef. 891 if (IVI.getNumIndices() != 1) 892 return (void)markOverdefined(&IVI); 893 894 Value *Aggr = IVI.getAggregateOperand(); 895 unsigned Idx = *IVI.idx_begin(); 896 897 // Compute the result based on what we're inserting. 898 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 899 // This passes through all values that aren't the inserted element. 900 if (i != Idx) { 901 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 902 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 903 continue; 904 } 905 906 Value *Val = IVI.getInsertedValueOperand(); 907 if (Val->getType()->isStructTy()) 908 // We don't track structs in structs. 909 markOverdefined(getStructValueState(&IVI, i), &IVI); 910 else { 911 ValueLatticeElement InVal = getValueState(Val); 912 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 913 } 914 } 915 } 916 917 void SCCPSolver::visitSelectInst(SelectInst &I) { 918 // If this select returns a struct, just mark the result overdefined. 919 // TODO: We could do a lot better than this if code actually uses this. 920 if (I.getType()->isStructTy()) 921 return (void)markOverdefined(&I); 922 923 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 924 // discover a concrete value later. 925 if (ValueState[&I].isOverdefined()) 926 return (void)markOverdefined(&I); 927 928 ValueLatticeElement CondValue = getValueState(I.getCondition()); 929 if (CondValue.isUnknownOrUndef()) 930 return; 931 932 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 933 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 934 mergeInValue(&I, getValueState(OpVal)); 935 return; 936 } 937 938 // Otherwise, the condition is overdefined or a constant we can't evaluate. 939 // See if we can produce something better than overdefined based on the T/F 940 // value. 941 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 942 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 943 944 bool Changed = ValueState[&I].mergeIn(TVal); 945 Changed |= ValueState[&I].mergeIn(FVal); 946 if (Changed) 947 pushToWorkListMsg(ValueState[&I], &I); 948 } 949 950 // Handle Unary Operators. 951 void SCCPSolver::visitUnaryOperator(Instruction &I) { 952 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 953 954 ValueLatticeElement &IV = ValueState[&I]; 955 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 956 // discover a concrete value later. 957 if (isOverdefined(IV)) 958 return (void)markOverdefined(&I); 959 960 if (isConstant(V0State)) { 961 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 962 963 // op Y -> undef. 964 if (isa<UndefValue>(C)) 965 return; 966 return (void)markConstant(IV, &I, C); 967 } 968 969 // If something is undef, wait for it to resolve. 970 if (!isOverdefined(V0State)) 971 return; 972 973 markOverdefined(&I); 974 } 975 976 // Handle Binary Operators. 977 void SCCPSolver::visitBinaryOperator(Instruction &I) { 978 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 979 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 980 981 ValueLatticeElement &IV = ValueState[&I]; 982 if (IV.isOverdefined()) 983 return; 984 985 // If something is undef, wait for it to resolve. 986 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 987 return; 988 989 if (V1State.isOverdefined() && V2State.isOverdefined()) 990 return (void)markOverdefined(&I); 991 992 // If either of the operands is a constant, try to fold it to a constant. 993 // TODO: Use information from notconstant better. 994 if ((V1State.isConstant() || V2State.isConstant())) { 995 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 996 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 997 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 998 auto *C = dyn_cast_or_null<Constant>(R); 999 if (C) { 1000 // X op Y -> undef. 1001 if (isa<UndefValue>(C)) 1002 return; 1003 // Conservatively assume that the result may be based on operands that may 1004 // be undef. Note that we use mergeInValue to combine the constant with 1005 // the existing lattice value for I, as different constants might be found 1006 // after one of the operands go to overdefined, e.g. due to one operand 1007 // being a special floating value. 1008 ValueLatticeElement NewV; 1009 NewV.markConstant(C, /*MayIncludeUndef=*/true); 1010 return (void)mergeInValue(&I, NewV); 1011 } 1012 } 1013 1014 // Only use ranges for binary operators on integers. 1015 if (!I.getType()->isIntegerTy()) 1016 return markOverdefined(&I); 1017 1018 // Try to simplify to a constant range. 1019 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1020 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1021 if (V1State.isConstantRange()) 1022 A = V1State.getConstantRange(); 1023 if (V2State.isConstantRange()) 1024 B = V2State.getConstantRange(); 1025 1026 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1027 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1028 1029 // TODO: Currently we do not exploit special values that produce something 1030 // better than overdefined with an overdefined operand for vector or floating 1031 // point types, like and <4 x i32> overdefined, zeroinitializer. 1032 } 1033 1034 // Handle ICmpInst instruction. 1035 void SCCPSolver::visitCmpInst(CmpInst &I) { 1036 // Do not cache this lookup, getValueState calls later in the function might 1037 // invalidate the reference. 1038 if (isOverdefined(ValueState[&I])) 1039 return (void)markOverdefined(&I); 1040 1041 Value *Op1 = I.getOperand(0); 1042 Value *Op2 = I.getOperand(1); 1043 1044 // For parameters, use ParamState which includes constant range info if 1045 // available. 1046 auto V1State = getValueState(Op1); 1047 auto V2State = getValueState(Op2); 1048 1049 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1050 if (C) { 1051 if (isa<UndefValue>(C)) 1052 return; 1053 ValueLatticeElement CV; 1054 CV.markConstant(C); 1055 mergeInValue(&I, CV); 1056 return; 1057 } 1058 1059 // If operands are still unknown, wait for it to resolve. 1060 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1061 !isConstant(ValueState[&I])) 1062 return; 1063 1064 markOverdefined(&I); 1065 } 1066 1067 // Handle getelementptr instructions. If all operands are constants then we 1068 // can turn this into a getelementptr ConstantExpr. 1069 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) { 1070 if (isOverdefined(ValueState[&I])) 1071 return (void)markOverdefined(&I); 1072 1073 SmallVector<Constant*, 8> Operands; 1074 Operands.reserve(I.getNumOperands()); 1075 1076 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1077 ValueLatticeElement State = getValueState(I.getOperand(i)); 1078 if (State.isUnknownOrUndef()) 1079 return; // Operands are not resolved yet. 1080 1081 if (isOverdefined(State)) 1082 return (void)markOverdefined(&I); 1083 1084 if (Constant *C = getConstant(State)) { 1085 Operands.push_back(C); 1086 continue; 1087 } 1088 1089 return (void)markOverdefined(&I); 1090 } 1091 1092 Constant *Ptr = Operands[0]; 1093 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1094 Constant *C = 1095 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1096 if (isa<UndefValue>(C)) 1097 return; 1098 markConstant(&I, C); 1099 } 1100 1101 void SCCPSolver::visitStoreInst(StoreInst &SI) { 1102 // If this store is of a struct, ignore it. 1103 if (SI.getOperand(0)->getType()->isStructTy()) 1104 return; 1105 1106 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1107 return; 1108 1109 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1110 auto I = TrackedGlobals.find(GV); 1111 if (I == TrackedGlobals.end()) 1112 return; 1113 1114 // Get the value we are storing into the global, then merge it. 1115 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1116 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1117 if (I->second.isOverdefined()) 1118 TrackedGlobals.erase(I); // No need to keep tracking this! 1119 } 1120 1121 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1122 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1123 if (I->getType()->isIntegerTy()) 1124 return ValueLatticeElement::getRange( 1125 getConstantRangeFromMetadata(*Ranges)); 1126 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1127 return ValueLatticeElement::getNot( 1128 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1129 return ValueLatticeElement::getOverdefined(); 1130 } 1131 1132 // Handle load instructions. If the operand is a constant pointer to a constant 1133 // global, we can replace the load with the loaded constant value! 1134 void SCCPSolver::visitLoadInst(LoadInst &I) { 1135 // If this load is of a struct or the load is volatile, just mark the result 1136 // as overdefined. 1137 if (I.getType()->isStructTy() || I.isVolatile()) 1138 return (void)markOverdefined(&I); 1139 1140 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1141 // discover a concrete value later. 1142 if (ValueState[&I].isOverdefined()) 1143 return (void)markOverdefined(&I); 1144 1145 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1146 if (PtrVal.isUnknownOrUndef()) 1147 return; // The pointer is not resolved yet! 1148 1149 ValueLatticeElement &IV = ValueState[&I]; 1150 1151 if (isConstant(PtrVal)) { 1152 Constant *Ptr = getConstant(PtrVal); 1153 1154 // load null is undefined. 1155 if (isa<ConstantPointerNull>(Ptr)) { 1156 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1157 return (void)markOverdefined(IV, &I); 1158 else 1159 return; 1160 } 1161 1162 // Transform load (constant global) into the value loaded. 1163 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1164 if (!TrackedGlobals.empty()) { 1165 // If we are tracking this global, merge in the known value for it. 1166 auto It = TrackedGlobals.find(GV); 1167 if (It != TrackedGlobals.end()) { 1168 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1169 return; 1170 } 1171 } 1172 } 1173 1174 // Transform load from a constant into a constant if possible. 1175 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1176 if (isa<UndefValue>(C)) 1177 return; 1178 return (void)markConstant(IV, &I, C); 1179 } 1180 } 1181 1182 // Fall back to metadata. 1183 mergeInValue(&I, getValueFromMetadata(&I)); 1184 } 1185 1186 void SCCPSolver::visitCallBase(CallBase &CB) { 1187 handleCallResult(CB); 1188 handleCallArguments(CB); 1189 } 1190 1191 void SCCPSolver::handleCallOverdefined(CallBase &CB) { 1192 Function *F = CB.getCalledFunction(); 1193 1194 // Void return and not tracking callee, just bail. 1195 if (CB.getType()->isVoidTy()) 1196 return; 1197 1198 // Always mark struct return as overdefined. 1199 if (CB.getType()->isStructTy()) 1200 return (void)markOverdefined(&CB); 1201 1202 // Otherwise, if we have a single return value case, and if the function is 1203 // a declaration, maybe we can constant fold it. 1204 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1205 SmallVector<Constant *, 8> Operands; 1206 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1207 if (AI->get()->getType()->isStructTy()) 1208 return markOverdefined(&CB); // Can't handle struct args. 1209 ValueLatticeElement State = getValueState(*AI); 1210 1211 if (State.isUnknownOrUndef()) 1212 return; // Operands are not resolved yet. 1213 if (isOverdefined(State)) 1214 return (void)markOverdefined(&CB); 1215 assert(isConstant(State) && "Unknown state!"); 1216 Operands.push_back(getConstant(State)); 1217 } 1218 1219 if (isOverdefined(getValueState(&CB))) 1220 return (void)markOverdefined(&CB); 1221 1222 // If we can constant fold this, mark the result of the call as a 1223 // constant. 1224 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1225 // call -> undef. 1226 if (isa<UndefValue>(C)) 1227 return; 1228 return (void)markConstant(&CB, C); 1229 } 1230 } 1231 1232 // Fall back to metadata. 1233 mergeInValue(&CB, getValueFromMetadata(&CB)); 1234 } 1235 1236 void SCCPSolver::handleCallArguments(CallBase &CB) { 1237 Function *F = CB.getCalledFunction(); 1238 // If this is a local function that doesn't have its address taken, mark its 1239 // entry block executable and merge in the actual arguments to the call into 1240 // the formal arguments of the function. 1241 if (!TrackingIncomingArguments.empty() && 1242 TrackingIncomingArguments.count(F)) { 1243 MarkBlockExecutable(&F->front()); 1244 1245 // Propagate information from this call site into the callee. 1246 auto CAI = CB.arg_begin(); 1247 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1248 ++AI, ++CAI) { 1249 // If this argument is byval, and if the function is not readonly, there 1250 // will be an implicit copy formed of the input aggregate. 1251 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1252 markOverdefined(&*AI); 1253 continue; 1254 } 1255 1256 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1257 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1258 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1259 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1260 getMaxWidenStepsOpts()); 1261 } 1262 } else 1263 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1264 } 1265 } 1266 } 1267 1268 void SCCPSolver::handleCallResult(CallBase &CB) { 1269 Function *F = CB.getCalledFunction(); 1270 1271 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1272 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1273 if (ValueState[&CB].isOverdefined()) 1274 return; 1275 1276 Value *CopyOf = CB.getOperand(0); 1277 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1278 auto *PI = getPredicateInfoFor(&CB); 1279 assert(PI && "Missing predicate info for ssa.copy"); 1280 1281 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1282 if (!Constraint) { 1283 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1284 return; 1285 } 1286 1287 CmpInst::Predicate Pred = Constraint->Predicate; 1288 Value *OtherOp = Constraint->OtherOp; 1289 1290 // Wait until OtherOp is resolved. 1291 if (getValueState(OtherOp).isUnknown()) { 1292 addAdditionalUser(OtherOp, &CB); 1293 return; 1294 } 1295 1296 // TODO: Actually filp MayIncludeUndef for the created range to false, 1297 // once most places in the optimizer respect the branches on 1298 // undef/poison are UB rule. The reason why the new range cannot be 1299 // undef is as follows below: 1300 // The new range is based on a branch condition. That guarantees that 1301 // neither of the compare operands can be undef in the branch targets, 1302 // unless we have conditions that are always true/false (e.g. icmp ule 1303 // i32, %a, i32_max). For the latter overdefined/empty range will be 1304 // inferred, but the branch will get folded accordingly anyways. 1305 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1306 1307 ValueLatticeElement CondVal = getValueState(OtherOp); 1308 ValueLatticeElement &IV = ValueState[&CB]; 1309 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1310 auto ImposedCR = 1311 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1312 1313 // Get the range imposed by the condition. 1314 if (CondVal.isConstantRange()) 1315 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1316 Pred, CondVal.getConstantRange()); 1317 1318 // Combine range info for the original value with the new range from the 1319 // condition. 1320 auto CopyOfCR = CopyOfVal.isConstantRange() 1321 ? CopyOfVal.getConstantRange() 1322 : ConstantRange::getFull( 1323 DL.getTypeSizeInBits(CopyOf->getType())); 1324 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1325 // If the existing information is != x, do not use the information from 1326 // a chained predicate, as the != x information is more likely to be 1327 // helpful in practice. 1328 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1329 NewCR = CopyOfCR; 1330 1331 addAdditionalUser(OtherOp, &CB); 1332 mergeInValue( 1333 IV, &CB, 1334 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1335 return; 1336 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1337 // For non-integer values or integer constant expressions, only 1338 // propagate equal constants. 1339 addAdditionalUser(OtherOp, &CB); 1340 mergeInValue(IV, &CB, CondVal); 1341 return; 1342 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1343 !MayIncludeUndef) { 1344 // Propagate inequalities. 1345 addAdditionalUser(OtherOp, &CB); 1346 mergeInValue(IV, &CB, 1347 ValueLatticeElement::getNot(CondVal.getConstant())); 1348 return; 1349 } 1350 1351 return (void)mergeInValue(IV, &CB, CopyOfVal); 1352 } 1353 1354 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { 1355 // Compute result range for intrinsics supported by ConstantRange. 1356 // Do this even if we don't know a range for all operands, as we may 1357 // still know something about the result range, e.g. of abs(x). 1358 SmallVector<ConstantRange, 2> OpRanges; 1359 for (Value *Op : II->args()) { 1360 const ValueLatticeElement &State = getValueState(Op); 1361 if (State.isConstantRange()) 1362 OpRanges.push_back(State.getConstantRange()); 1363 else 1364 OpRanges.push_back( 1365 ConstantRange::getFull(Op->getType()->getScalarSizeInBits())); 1366 } 1367 1368 ConstantRange Result = 1369 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges); 1370 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result)); 1371 } 1372 } 1373 1374 // The common case is that we aren't tracking the callee, either because we 1375 // are not doing interprocedural analysis or the callee is indirect, or is 1376 // external. Handle these cases first. 1377 if (!F || F->isDeclaration()) 1378 return handleCallOverdefined(CB); 1379 1380 // If this is a single/zero retval case, see if we're tracking the function. 1381 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1382 if (!MRVFunctionsTracked.count(F)) 1383 return handleCallOverdefined(CB); // Not tracking this callee. 1384 1385 // If we are tracking this callee, propagate the result of the function 1386 // into this call site. 1387 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1388 mergeInValue(getStructValueState(&CB, i), &CB, 1389 TrackedMultipleRetVals[std::make_pair(F, i)], 1390 getMaxWidenStepsOpts()); 1391 } else { 1392 auto TFRVI = TrackedRetVals.find(F); 1393 if (TFRVI == TrackedRetVals.end()) 1394 return handleCallOverdefined(CB); // Not tracking this callee. 1395 1396 // If so, propagate the return value of the callee into this call result. 1397 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1398 } 1399 } 1400 1401 void SCCPSolver::Solve() { 1402 // Process the work lists until they are empty! 1403 while (!BBWorkList.empty() || !InstWorkList.empty() || 1404 !OverdefinedInstWorkList.empty()) { 1405 // Process the overdefined instruction's work list first, which drives other 1406 // things to overdefined more quickly. 1407 while (!OverdefinedInstWorkList.empty()) { 1408 Value *I = OverdefinedInstWorkList.pop_back_val(); 1409 1410 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1411 1412 // "I" got into the work list because it either made the transition from 1413 // bottom to constant, or to overdefined. 1414 // 1415 // Anything on this worklist that is overdefined need not be visited 1416 // since all of its users will have already been marked as overdefined 1417 // Update all of the users of this instruction's value. 1418 // 1419 markUsersAsChanged(I); 1420 } 1421 1422 // Process the instruction work list. 1423 while (!InstWorkList.empty()) { 1424 Value *I = InstWorkList.pop_back_val(); 1425 1426 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1427 1428 // "I" got into the work list because it made the transition from undef to 1429 // constant. 1430 // 1431 // Anything on this worklist that is overdefined need not be visited 1432 // since all of its users will have already been marked as overdefined. 1433 // Update all of the users of this instruction's value. 1434 // 1435 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1436 markUsersAsChanged(I); 1437 } 1438 1439 // Process the basic block work list. 1440 while (!BBWorkList.empty()) { 1441 BasicBlock *BB = BBWorkList.back(); 1442 BBWorkList.pop_back(); 1443 1444 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1445 1446 // Notify all instructions in this basic block that they are newly 1447 // executable. 1448 visit(BB); 1449 } 1450 } 1451 } 1452 1453 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume 1454 /// that branches on undef values cannot reach any of their successors. 1455 /// However, this is not a safe assumption. After we solve dataflow, this 1456 /// method should be use to handle this. If this returns true, the solver 1457 /// should be rerun. 1458 /// 1459 /// This method handles this by finding an unresolved branch and marking it one 1460 /// of the edges from the block as being feasible, even though the condition 1461 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1462 /// CFG and only slightly pessimizes the analysis results (by marking one, 1463 /// potentially infeasible, edge feasible). This cannot usefully modify the 1464 /// constraints on the condition of the branch, as that would impact other users 1465 /// of the value. 1466 /// 1467 /// This scan also checks for values that use undefs. It conservatively marks 1468 /// them as overdefined. 1469 bool SCCPSolver::ResolvedUndefsIn(Function &F) { 1470 bool MadeChange = false; 1471 for (BasicBlock &BB : F) { 1472 if (!BBExecutable.count(&BB)) 1473 continue; 1474 1475 for (Instruction &I : BB) { 1476 // Look for instructions which produce undef values. 1477 if (I.getType()->isVoidTy()) continue; 1478 1479 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1480 // Only a few things that can be structs matter for undef. 1481 1482 // Tracked calls must never be marked overdefined in ResolvedUndefsIn. 1483 if (auto *CB = dyn_cast<CallBase>(&I)) 1484 if (Function *F = CB->getCalledFunction()) 1485 if (MRVFunctionsTracked.count(F)) 1486 continue; 1487 1488 // extractvalue and insertvalue don't need to be marked; they are 1489 // tracked as precisely as their operands. 1490 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1491 continue; 1492 // Send the results of everything else to overdefined. We could be 1493 // more precise than this but it isn't worth bothering. 1494 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1495 ValueLatticeElement &LV = getStructValueState(&I, i); 1496 if (LV.isUnknownOrUndef()) { 1497 markOverdefined(LV, &I); 1498 MadeChange = true; 1499 } 1500 } 1501 continue; 1502 } 1503 1504 ValueLatticeElement &LV = getValueState(&I); 1505 if (!LV.isUnknownOrUndef()) 1506 continue; 1507 1508 // There are two reasons a call can have an undef result 1509 // 1. It could be tracked. 1510 // 2. It could be constant-foldable. 1511 // Because of the way we solve return values, tracked calls must 1512 // never be marked overdefined in ResolvedUndefsIn. 1513 if (auto *CB = dyn_cast<CallBase>(&I)) 1514 if (Function *F = CB->getCalledFunction()) 1515 if (TrackedRetVals.count(F)) 1516 continue; 1517 1518 if (isa<LoadInst>(I)) { 1519 // A load here means one of two things: a load of undef from a global, 1520 // a load from an unknown pointer. Either way, having it return undef 1521 // is okay. 1522 continue; 1523 } 1524 1525 markOverdefined(&I); 1526 MadeChange = true; 1527 } 1528 1529 // Check to see if we have a branch or switch on an undefined value. If so 1530 // we force the branch to go one way or the other to make the successor 1531 // values live. It doesn't really matter which way we force it. 1532 Instruction *TI = BB.getTerminator(); 1533 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1534 if (!BI->isConditional()) continue; 1535 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1536 continue; 1537 1538 // If the input to SCCP is actually branch on undef, fix the undef to 1539 // false. 1540 if (isa<UndefValue>(BI->getCondition())) { 1541 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1542 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1543 MadeChange = true; 1544 continue; 1545 } 1546 1547 // Otherwise, it is a branch on a symbolic value which is currently 1548 // considered to be undef. Make sure some edge is executable, so a 1549 // branch on "undef" always flows somewhere. 1550 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1551 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1552 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1553 MadeChange = true; 1554 1555 continue; 1556 } 1557 1558 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1559 // Indirect branch with no successor ?. Its ok to assume it branches 1560 // to no target. 1561 if (IBR->getNumSuccessors() < 1) 1562 continue; 1563 1564 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1565 continue; 1566 1567 // If the input to SCCP is actually branch on undef, fix the undef to 1568 // the first successor of the indirect branch. 1569 if (isa<UndefValue>(IBR->getAddress())) { 1570 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1571 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1572 MadeChange = true; 1573 continue; 1574 } 1575 1576 // Otherwise, it is a branch on a symbolic value which is currently 1577 // considered to be undef. Make sure some edge is executable, so a 1578 // branch on "undef" always flows somewhere. 1579 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1580 // we can assume the branch has undefined behavior instead. 1581 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1582 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1583 MadeChange = true; 1584 1585 continue; 1586 } 1587 1588 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1589 if (!SI->getNumCases() || 1590 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1591 continue; 1592 1593 // If the input to SCCP is actually switch on undef, fix the undef to 1594 // the first constant. 1595 if (isa<UndefValue>(SI->getCondition())) { 1596 SI->setCondition(SI->case_begin()->getCaseValue()); 1597 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1598 MadeChange = true; 1599 continue; 1600 } 1601 1602 // Otherwise, it is a branch on a symbolic value which is currently 1603 // considered to be undef. Make sure some edge is executable, so a 1604 // branch on "undef" always flows somewhere. 1605 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1606 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1607 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1608 MadeChange = true; 1609 1610 continue; 1611 } 1612 } 1613 1614 return MadeChange; 1615 } 1616 1617 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) { 1618 Constant *Const = nullptr; 1619 if (V->getType()->isStructTy()) { 1620 std::vector<ValueLatticeElement> IVs = Solver.getStructLatticeValueFor(V); 1621 if (any_of(IVs, 1622 [](const ValueLatticeElement &LV) { return isOverdefined(LV); })) 1623 return false; 1624 std::vector<Constant *> ConstVals; 1625 auto *ST = cast<StructType>(V->getType()); 1626 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1627 ValueLatticeElement V = IVs[i]; 1628 ConstVals.push_back(isConstant(V) 1629 ? Solver.getConstant(V) 1630 : UndefValue::get(ST->getElementType(i))); 1631 } 1632 Const = ConstantStruct::get(ST, ConstVals); 1633 } else { 1634 const ValueLatticeElement &IV = Solver.getLatticeValueFor(V); 1635 if (isOverdefined(IV)) 1636 return false; 1637 1638 Const = 1639 isConstant(IV) ? Solver.getConstant(IV) : UndefValue::get(V->getType()); 1640 } 1641 assert(Const && "Constant is nullptr here!"); 1642 1643 // Replacing `musttail` instructions with constant breaks `musttail` invariant 1644 // unless the call itself can be removed 1645 CallInst *CI = dyn_cast<CallInst>(V); 1646 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) { 1647 Function *F = CI->getCalledFunction(); 1648 1649 // Don't zap returns of the callee 1650 if (F) 1651 Solver.AddMustTailCallee(F); 1652 1653 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI 1654 << " as a constant\n"); 1655 return false; 1656 } 1657 1658 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n'); 1659 1660 // Replaces all of the uses of a variable with uses of the constant. 1661 V->replaceAllUsesWith(Const); 1662 return true; 1663 } 1664 1665 static bool simplifyInstsInBlock(SCCPSolver &Solver, BasicBlock &BB, 1666 SmallPtrSetImpl<Value *> &InsertedValues, 1667 Statistic &InstRemovedStat, 1668 Statistic &InstReplacedStat) { 1669 bool MadeChanges = false; 1670 for (Instruction &Inst : make_early_inc_range(BB)) { 1671 if (Inst.getType()->isVoidTy()) 1672 continue; 1673 if (tryToReplaceWithConstant(Solver, &Inst)) { 1674 if (Inst.isSafeToRemove()) 1675 Inst.eraseFromParent(); 1676 // Hey, we just changed something! 1677 MadeChanges = true; 1678 ++InstRemovedStat; 1679 } else if (isa<SExtInst>(&Inst)) { 1680 Value *ExtOp = Inst.getOperand(0); 1681 if (isa<Constant>(ExtOp) || InsertedValues.count(ExtOp)) 1682 continue; 1683 const ValueLatticeElement &IV = Solver.getLatticeValueFor(ExtOp); 1684 if (!IV.isConstantRange(/*UndefAllowed=*/false)) 1685 continue; 1686 if (IV.getConstantRange().isAllNonNegative()) { 1687 auto *ZExt = new ZExtInst(ExtOp, Inst.getType(), "", &Inst); 1688 InsertedValues.insert(ZExt); 1689 Inst.replaceAllUsesWith(ZExt); 1690 Solver.removeLatticeValueFor(&Inst); 1691 Inst.eraseFromParent(); 1692 InstReplacedStat++; 1693 MadeChanges = true; 1694 } 1695 } 1696 } 1697 return MadeChanges; 1698 } 1699 1700 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm, 1701 // and return true if the function was modified. 1702 static bool runSCCP(Function &F, const DataLayout &DL, 1703 const TargetLibraryInfo *TLI) { 1704 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); 1705 SCCPSolver Solver( 1706 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; }, 1707 F.getContext()); 1708 1709 // Mark the first block of the function as being executable. 1710 Solver.MarkBlockExecutable(&F.front()); 1711 1712 // Mark all arguments to the function as being overdefined. 1713 for (Argument &AI : F.args()) 1714 Solver.markOverdefined(&AI); 1715 1716 // Solve for constants. 1717 bool ResolvedUndefs = true; 1718 while (ResolvedUndefs) { 1719 Solver.Solve(); 1720 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n"); 1721 ResolvedUndefs = Solver.ResolvedUndefsIn(F); 1722 } 1723 1724 bool MadeChanges = false; 1725 1726 // If we decided that there are basic blocks that are dead in this function, 1727 // delete their contents now. Note that we cannot actually delete the blocks, 1728 // as we cannot modify the CFG of the function. 1729 1730 SmallPtrSet<Value *, 32> InsertedValues; 1731 for (BasicBlock &BB : F) { 1732 if (!Solver.isBlockExecutable(&BB)) { 1733 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 1734 1735 ++NumDeadBlocks; 1736 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB).first; 1737 1738 MadeChanges = true; 1739 continue; 1740 } 1741 1742 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 1743 NumInstRemoved, NumInstReplaced); 1744 } 1745 1746 return MadeChanges; 1747 } 1748 1749 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) { 1750 const DataLayout &DL = F.getParent()->getDataLayout(); 1751 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1752 if (!runSCCP(F, DL, &TLI)) 1753 return PreservedAnalyses::all(); 1754 1755 auto PA = PreservedAnalyses(); 1756 PA.preserve<GlobalsAA>(); 1757 PA.preserveSet<CFGAnalyses>(); 1758 return PA; 1759 } 1760 1761 namespace { 1762 1763 //===--------------------------------------------------------------------===// 1764 // 1765 /// SCCP Class - This class uses the SCCPSolver to implement a per-function 1766 /// Sparse Conditional Constant Propagator. 1767 /// 1768 class SCCPLegacyPass : public FunctionPass { 1769 public: 1770 // Pass identification, replacement for typeid 1771 static char ID; 1772 1773 SCCPLegacyPass() : FunctionPass(ID) { 1774 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry()); 1775 } 1776 1777 void getAnalysisUsage(AnalysisUsage &AU) const override { 1778 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1779 AU.addPreserved<GlobalsAAWrapperPass>(); 1780 AU.setPreservesCFG(); 1781 } 1782 1783 // runOnFunction - Run the Sparse Conditional Constant Propagation 1784 // algorithm, and return true if the function was modified. 1785 bool runOnFunction(Function &F) override { 1786 if (skipFunction(F)) 1787 return false; 1788 const DataLayout &DL = F.getParent()->getDataLayout(); 1789 const TargetLibraryInfo *TLI = 1790 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1791 return runSCCP(F, DL, TLI); 1792 } 1793 }; 1794 1795 } // end anonymous namespace 1796 1797 char SCCPLegacyPass::ID = 0; 1798 1799 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp", 1800 "Sparse Conditional Constant Propagation", false, false) 1801 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1802 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp", 1803 "Sparse Conditional Constant Propagation", false, false) 1804 1805 // createSCCPPass - This is the public interface to this file. 1806 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); } 1807 1808 static void findReturnsToZap(Function &F, 1809 SmallVector<ReturnInst *, 8> &ReturnsToZap, 1810 SCCPSolver &Solver) { 1811 // We can only do this if we know that nothing else can call the function. 1812 if (!Solver.isArgumentTrackedFunction(&F)) 1813 return; 1814 1815 // There is a non-removable musttail call site of this function. Zapping 1816 // returns is not allowed. 1817 if (Solver.isMustTailCallee(&F)) { 1818 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName() 1819 << " due to present musttail call of it\n"); 1820 return; 1821 } 1822 1823 assert( 1824 all_of(F.users(), 1825 [&Solver](User *U) { 1826 if (isa<Instruction>(U) && 1827 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) 1828 return true; 1829 // Non-callsite uses are not impacted by zapping. Also, constant 1830 // uses (like blockaddresses) could stuck around, without being 1831 // used in the underlying IR, meaning we do not have lattice 1832 // values for them. 1833 if (!isa<CallBase>(U)) 1834 return true; 1835 if (U->getType()->isStructTy()) { 1836 return all_of(Solver.getStructLatticeValueFor(U), 1837 [](const ValueLatticeElement &LV) { 1838 return !isOverdefined(LV); 1839 }); 1840 } 1841 return !isOverdefined(Solver.getLatticeValueFor(U)); 1842 }) && 1843 "We can only zap functions where all live users have a concrete value"); 1844 1845 for (BasicBlock &BB : F) { 1846 if (CallInst *CI = BB.getTerminatingMustTailCall()) { 1847 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " 1848 << "musttail call : " << *CI << "\n"); 1849 (void)CI; 1850 return; 1851 } 1852 1853 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1854 if (!isa<UndefValue>(RI->getOperand(0))) 1855 ReturnsToZap.push_back(RI); 1856 } 1857 } 1858 1859 static bool removeNonFeasibleEdges(const SCCPSolver &Solver, BasicBlock *BB, 1860 DomTreeUpdater &DTU) { 1861 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors; 1862 bool HasNonFeasibleEdges = false; 1863 for (BasicBlock *Succ : successors(BB)) { 1864 if (Solver.isEdgeFeasible(BB, Succ)) 1865 FeasibleSuccessors.insert(Succ); 1866 else 1867 HasNonFeasibleEdges = true; 1868 } 1869 1870 // All edges feasible, nothing to do. 1871 if (!HasNonFeasibleEdges) 1872 return false; 1873 1874 // SCCP can only determine non-feasible edges for br, switch and indirectbr. 1875 Instruction *TI = BB->getTerminator(); 1876 assert((isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1877 isa<IndirectBrInst>(TI)) && 1878 "Terminator must be a br, switch or indirectbr"); 1879 1880 if (FeasibleSuccessors.size() == 1) { 1881 // Replace with an unconditional branch to the only feasible successor. 1882 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin(); 1883 SmallVector<DominatorTree::UpdateType, 8> Updates; 1884 bool HaveSeenOnlyFeasibleSuccessor = false; 1885 for (BasicBlock *Succ : successors(BB)) { 1886 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) { 1887 // Don't remove the edge to the only feasible successor the first time 1888 // we see it. We still do need to remove any multi-edges to it though. 1889 HaveSeenOnlyFeasibleSuccessor = true; 1890 continue; 1891 } 1892 1893 Succ->removePredecessor(BB); 1894 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1895 } 1896 1897 BranchInst::Create(OnlyFeasibleSuccessor, BB); 1898 TI->eraseFromParent(); 1899 DTU.applyUpdatesPermissive(Updates); 1900 } else if (FeasibleSuccessors.size() > 1) { 1901 SwitchInstProfUpdateWrapper SI(*cast<SwitchInst>(TI)); 1902 SmallVector<DominatorTree::UpdateType, 8> Updates; 1903 for (auto CI = SI->case_begin(); CI != SI->case_end();) { 1904 if (FeasibleSuccessors.contains(CI->getCaseSuccessor())) { 1905 ++CI; 1906 continue; 1907 } 1908 1909 BasicBlock *Succ = CI->getCaseSuccessor(); 1910 Succ->removePredecessor(BB); 1911 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1912 SI.removeCase(CI); 1913 // Don't increment CI, as we removed a case. 1914 } 1915 1916 DTU.applyUpdatesPermissive(Updates); 1917 } else { 1918 llvm_unreachable("Must have at least one feasible successor"); 1919 } 1920 return true; 1921 } 1922 1923 bool llvm::runIPSCCP( 1924 Module &M, const DataLayout &DL, 1925 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1926 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) { 1927 SCCPSolver Solver(DL, GetTLI, M.getContext()); 1928 1929 // Loop over all functions, marking arguments to those with their addresses 1930 // taken or that are external as overdefined. 1931 for (Function &F : M) { 1932 if (F.isDeclaration()) 1933 continue; 1934 1935 Solver.addAnalysis(F, getAnalysis(F)); 1936 1937 // Determine if we can track the function's return values. If so, add the 1938 // function to the solver's set of return-tracked functions. 1939 if (canTrackReturnsInterprocedurally(&F)) 1940 Solver.AddTrackedFunction(&F); 1941 1942 // Determine if we can track the function's arguments. If so, add the 1943 // function to the solver's set of argument-tracked functions. 1944 if (canTrackArgumentsInterprocedurally(&F)) { 1945 Solver.AddArgumentTrackedFunction(&F); 1946 continue; 1947 } 1948 1949 // Assume the function is called. 1950 Solver.MarkBlockExecutable(&F.front()); 1951 1952 // Assume nothing about the incoming arguments. 1953 for (Argument &AI : F.args()) 1954 Solver.markOverdefined(&AI); 1955 } 1956 1957 // Determine if we can track any of the module's global variables. If so, add 1958 // the global variables we can track to the solver's set of tracked global 1959 // variables. 1960 for (GlobalVariable &G : M.globals()) { 1961 G.removeDeadConstantUsers(); 1962 if (canTrackGlobalVariableInterprocedurally(&G)) 1963 Solver.TrackValueOfGlobalVariable(&G); 1964 } 1965 1966 // Solve for constants. 1967 bool ResolvedUndefs = true; 1968 Solver.Solve(); 1969 while (ResolvedUndefs) { 1970 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n"); 1971 ResolvedUndefs = false; 1972 for (Function &F : M) { 1973 if (Solver.ResolvedUndefsIn(F)) 1974 ResolvedUndefs = true; 1975 } 1976 if (ResolvedUndefs) 1977 Solver.Solve(); 1978 } 1979 1980 bool MadeChanges = false; 1981 1982 // Iterate over all of the instructions in the module, replacing them with 1983 // constants if we have found them to be of constant values. 1984 1985 for (Function &F : M) { 1986 if (F.isDeclaration()) 1987 continue; 1988 1989 SmallVector<BasicBlock *, 512> BlocksToErase; 1990 1991 if (Solver.isBlockExecutable(&F.front())) { 1992 bool ReplacedPointerArg = false; 1993 for (Argument &Arg : F.args()) { 1994 if (!Arg.use_empty() && tryToReplaceWithConstant(Solver, &Arg)) { 1995 ReplacedPointerArg |= Arg.getType()->isPointerTy(); 1996 ++IPNumArgsElimed; 1997 } 1998 } 1999 2000 // If we replaced an argument, the argmemonly and 2001 // inaccessiblemem_or_argmemonly attributes do not hold any longer. Remove 2002 // them from both the function and callsites. 2003 if (ReplacedPointerArg) { 2004 AttrBuilder AttributesToRemove; 2005 AttributesToRemove.addAttribute(Attribute::ArgMemOnly); 2006 AttributesToRemove.addAttribute(Attribute::InaccessibleMemOrArgMemOnly); 2007 F.removeAttributes(AttributeList::FunctionIndex, AttributesToRemove); 2008 2009 for (User *U : F.users()) { 2010 auto *CB = dyn_cast<CallBase>(U); 2011 if (!CB || CB->getCalledFunction() != &F) 2012 continue; 2013 2014 CB->removeAttributes(AttributeList::FunctionIndex, 2015 AttributesToRemove); 2016 } 2017 } 2018 } 2019 2020 SmallPtrSet<Value *, 32> InsertedValues; 2021 for (BasicBlock &BB : F) { 2022 if (!Solver.isBlockExecutable(&BB)) { 2023 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); 2024 ++NumDeadBlocks; 2025 2026 MadeChanges = true; 2027 2028 if (&BB != &F.front()) 2029 BlocksToErase.push_back(&BB); 2030 continue; 2031 } 2032 2033 MadeChanges |= simplifyInstsInBlock(Solver, BB, InsertedValues, 2034 IPNumInstRemoved, IPNumInstReplaced); 2035 } 2036 2037 DomTreeUpdater DTU = Solver.getDTU(F); 2038 // Change dead blocks to unreachable. We do it after replacing constants 2039 // in all executable blocks, because changeToUnreachable may remove PHI 2040 // nodes in executable blocks we found values for. The function's entry 2041 // block is not part of BlocksToErase, so we have to handle it separately. 2042 for (BasicBlock *BB : BlocksToErase) { 2043 NumInstRemoved += 2044 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false, 2045 /*PreserveLCSSA=*/false, &DTU); 2046 } 2047 if (!Solver.isBlockExecutable(&F.front())) 2048 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(), 2049 /*UseLLVMTrap=*/false, 2050 /*PreserveLCSSA=*/false, &DTU); 2051 2052 for (BasicBlock &BB : F) 2053 MadeChanges |= removeNonFeasibleEdges(Solver, &BB, DTU); 2054 2055 for (BasicBlock *DeadBB : BlocksToErase) 2056 DTU.deleteBB(DeadBB); 2057 2058 for (BasicBlock &BB : F) { 2059 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) { 2060 Instruction *Inst = &*BI++; 2061 if (Solver.getPredicateInfoFor(Inst)) { 2062 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) { 2063 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 2064 Value *Op = II->getOperand(0); 2065 Inst->replaceAllUsesWith(Op); 2066 Inst->eraseFromParent(); 2067 } 2068 } 2069 } 2070 } 2071 } 2072 } 2073 2074 // If we inferred constant or undef return values for a function, we replaced 2075 // all call uses with the inferred value. This means we don't need to bother 2076 // actually returning anything from the function. Replace all return 2077 // instructions with return undef. 2078 // 2079 // Do this in two stages: first identify the functions we should process, then 2080 // actually zap their returns. This is important because we can only do this 2081 // if the address of the function isn't taken. In cases where a return is the 2082 // last use of a function, the order of processing functions would affect 2083 // whether other functions are optimizable. 2084 SmallVector<ReturnInst*, 8> ReturnsToZap; 2085 2086 for (const auto &I : Solver.getTrackedRetVals()) { 2087 Function *F = I.first; 2088 const ValueLatticeElement &ReturnValue = I.second; 2089 2090 // If there is a known constant range for the return value, add !range 2091 // metadata to the function's call sites. 2092 if (ReturnValue.isConstantRange() && 2093 !ReturnValue.getConstantRange().isSingleElement()) { 2094 // Do not add range metadata if the return value may include undef. 2095 if (ReturnValue.isConstantRangeIncludingUndef()) 2096 continue; 2097 2098 auto &CR = ReturnValue.getConstantRange(); 2099 for (User *User : F->users()) { 2100 auto *CB = dyn_cast<CallBase>(User); 2101 if (!CB || CB->getCalledFunction() != F) 2102 continue; 2103 2104 // Limit to cases where the return value is guaranteed to be neither 2105 // poison nor undef. Poison will be outside any range and currently 2106 // values outside of the specified range cause immediate undefined 2107 // behavior. 2108 if (!isGuaranteedNotToBeUndefOrPoison(CB, nullptr, CB)) 2109 continue; 2110 2111 // Do not touch existing metadata for now. 2112 // TODO: We should be able to take the intersection of the existing 2113 // metadata and the inferred range. 2114 if (CB->getMetadata(LLVMContext::MD_range)) 2115 continue; 2116 2117 LLVMContext &Context = CB->getParent()->getContext(); 2118 Metadata *RangeMD[] = { 2119 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getLower())), 2120 ConstantAsMetadata::get(ConstantInt::get(Context, CR.getUpper()))}; 2121 CB->setMetadata(LLVMContext::MD_range, MDNode::get(Context, RangeMD)); 2122 } 2123 continue; 2124 } 2125 if (F->getReturnType()->isVoidTy()) 2126 continue; 2127 if (isConstant(ReturnValue) || ReturnValue.isUnknownOrUndef()) 2128 findReturnsToZap(*F, ReturnsToZap, Solver); 2129 } 2130 2131 for (auto F : Solver.getMRVFunctionsTracked()) { 2132 assert(F->getReturnType()->isStructTy() && 2133 "The return type should be a struct"); 2134 StructType *STy = cast<StructType>(F->getReturnType()); 2135 if (Solver.isStructLatticeConstant(F, STy)) 2136 findReturnsToZap(*F, ReturnsToZap, Solver); 2137 } 2138 2139 // Zap all returns which we've identified as zap to change. 2140 SmallSetVector<Function *, 8> FuncZappedReturn; 2141 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) { 2142 Function *F = ReturnsToZap[i]->getParent()->getParent(); 2143 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType())); 2144 // Record all functions that are zapped. 2145 FuncZappedReturn.insert(F); 2146 } 2147 2148 // Remove the returned attribute for zapped functions and the 2149 // corresponding call sites. 2150 for (Function *F : FuncZappedReturn) { 2151 for (Argument &A : F->args()) 2152 F->removeParamAttr(A.getArgNo(), Attribute::Returned); 2153 for (Use &U : F->uses()) { 2154 // Skip over blockaddr users. 2155 if (isa<BlockAddress>(U.getUser())) 2156 continue; 2157 CallBase *CB = cast<CallBase>(U.getUser()); 2158 for (Use &Arg : CB->args()) 2159 CB->removeParamAttr(CB->getArgOperandNo(&Arg), Attribute::Returned); 2160 } 2161 } 2162 2163 // If we inferred constant or undef values for globals variables, we can 2164 // delete the global and any stores that remain to it. 2165 for (auto &I : make_early_inc_range(Solver.getTrackedGlobals())) { 2166 GlobalVariable *GV = I.first; 2167 if (isOverdefined(I.second)) 2168 continue; 2169 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() 2170 << "' is constant!\n"); 2171 while (!GV->use_empty()) { 2172 StoreInst *SI = cast<StoreInst>(GV->user_back()); 2173 SI->eraseFromParent(); 2174 MadeChanges = true; 2175 } 2176 M.getGlobalList().erase(GV); 2177 ++IPNumGlobalConst; 2178 } 2179 2180 return MadeChanges; 2181 } 2182