1 //===- SCCPSolver.cpp - SCCP Utility --------------------------- *- C++ -*-===// 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 // \file 10 // This file implements the Sparse Conditional Constant Propagation (SCCP) 11 // utility. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Utils/SCCPSolver.h" 16 #include "llvm/Analysis/ConstantFolding.h" 17 #include "llvm/Analysis/InstructionSimplify.h" 18 #include "llvm/Analysis/ValueTracking.h" 19 #include "llvm/InitializePasses.h" 20 #include "llvm/Pass.h" 21 #include "llvm/Support/Casting.h" 22 #include "llvm/Support/Debug.h" 23 #include "llvm/Support/ErrorHandling.h" 24 #include "llvm/Support/raw_ostream.h" 25 #include "llvm/Transforms/Utils/Local.h" 26 #include <cassert> 27 #include <utility> 28 #include <vector> 29 30 using namespace llvm; 31 32 #define DEBUG_TYPE "sccp" 33 34 // The maximum number of range extensions allowed for operations requiring 35 // widening. 36 static const unsigned MaxNumRangeExtensions = 10; 37 38 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 39 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 40 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 41 MaxNumRangeExtensions); 42 } 43 44 namespace { 45 46 // Helper to check if \p LV is either a constant or a constant 47 // range with a single element. This should cover exactly the same cases as the 48 // old ValueLatticeElement::isConstant() and is intended to be used in the 49 // transition to ValueLatticeElement. 50 bool isConstant(const ValueLatticeElement &LV) { 51 return LV.isConstant() || 52 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 53 } 54 55 // Helper to check if \p LV is either overdefined or a constant range with more 56 // than a single element. This should cover exactly the same cases as the old 57 // ValueLatticeElement::isOverdefined() and is intended to be used in the 58 // transition to ValueLatticeElement. 59 bool isOverdefined(const ValueLatticeElement &LV) { 60 return !LV.isUnknownOrUndef() && !isConstant(LV); 61 } 62 63 } // namespace 64 65 namespace llvm { 66 67 /// Helper class for SCCPSolver. This implements the instruction visitor and 68 /// holds all the state. 69 class SCCPInstVisitor : public InstVisitor<SCCPInstVisitor> { 70 const DataLayout &DL; 71 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 72 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 73 DenseMap<Value *, ValueLatticeElement> 74 ValueState; // The state each value is in. 75 76 /// StructValueState - This maintains ValueState for values that have 77 /// StructType, for example for formal arguments, calls, insertelement, etc. 78 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 79 80 /// GlobalValue - If we are tracking any values for the contents of a global 81 /// variable, we keep a mapping from the constant accessor to the element of 82 /// the global, to the currently known value. If the value becomes 83 /// overdefined, it's entry is simply removed from this map. 84 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 85 86 /// TrackedRetVals - If we are tracking arguments into and the return 87 /// value out of a function, it will have an entry in this map, indicating 88 /// what the known return value for the function is. 89 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 90 91 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 92 /// that return multiple values. 93 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 94 TrackedMultipleRetVals; 95 96 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 97 /// represented here for efficient lookup. 98 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 99 100 /// A list of functions whose return cannot be modified. 101 SmallPtrSet<Function *, 16> MustPreserveReturnsInFunctions; 102 103 /// TrackingIncomingArguments - This is the set of functions for whose 104 /// arguments we make optimistic assumptions about and try to prove as 105 /// constants. 106 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 107 108 /// The reason for two worklists is that overdefined is the lowest state 109 /// on the lattice, and moving things to overdefined as fast as possible 110 /// makes SCCP converge much faster. 111 /// 112 /// By having a separate worklist, we accomplish this because everything 113 /// possibly overdefined will become overdefined at the soonest possible 114 /// point. 115 SmallVector<Value *, 64> OverdefinedInstWorkList; 116 SmallVector<Value *, 64> InstWorkList; 117 118 // The BasicBlock work list 119 SmallVector<BasicBlock *, 64> BBWorkList; 120 121 /// KnownFeasibleEdges - Entries in this set are edges which have already had 122 /// PHI nodes retriggered. 123 using Edge = std::pair<BasicBlock *, BasicBlock *>; 124 DenseSet<Edge> KnownFeasibleEdges; 125 126 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 127 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 128 129 LLVMContext &Ctx; 130 131 private: 132 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 133 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 134 } 135 136 // pushToWorkList - Helper for markConstant/markOverdefined 137 void pushToWorkList(ValueLatticeElement &IV, Value *V); 138 139 // Helper to push \p V to the worklist, after updating it to \p IV. Also 140 // prints a debug message with the updated value. 141 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V); 142 143 // markConstant - Make a value be marked as "constant". If the value 144 // is not already a constant, add it to the instruction work list so that 145 // the users of the instruction are updated later. 146 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 147 bool MayIncludeUndef = false); 148 149 bool markConstant(Value *V, Constant *C) { 150 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 151 return markConstant(ValueState[V], V, C); 152 } 153 154 // markOverdefined - Make a value be marked as "overdefined". If the 155 // value is not already overdefined, add it to the overdefined instruction 156 // work list so that the users of the instruction are updated later. 157 bool markOverdefined(ValueLatticeElement &IV, Value *V); 158 159 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 160 /// changes. 161 bool mergeInValue(ValueLatticeElement &IV, Value *V, 162 ValueLatticeElement MergeWithV, 163 ValueLatticeElement::MergeOptions Opts = { 164 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}); 165 166 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 167 ValueLatticeElement::MergeOptions Opts = { 168 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 169 assert(!V->getType()->isStructTy() && 170 "non-structs should use markConstant"); 171 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 172 } 173 174 /// getValueState - Return the ValueLatticeElement object that corresponds to 175 /// the value. This function handles the case when the value hasn't been seen 176 /// yet by properly seeding constants etc. 177 ValueLatticeElement &getValueState(Value *V) { 178 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 179 180 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 181 ValueLatticeElement &LV = I.first->second; 182 183 if (!I.second) 184 return LV; // Common case, already in the map. 185 186 if (auto *C = dyn_cast<Constant>(V)) 187 LV.markConstant(C); // Constants are constant 188 189 // All others are unknown by default. 190 return LV; 191 } 192 193 /// getStructValueState - Return the ValueLatticeElement object that 194 /// corresponds to the value/field pair. This function handles the case when 195 /// the value hasn't been seen yet by properly seeding constants etc. 196 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 197 assert(V->getType()->isStructTy() && "Should use getValueState"); 198 assert(i < cast<StructType>(V->getType())->getNumElements() && 199 "Invalid element #"); 200 201 auto I = StructValueState.insert( 202 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 203 ValueLatticeElement &LV = I.first->second; 204 205 if (!I.second) 206 return LV; // Common case, already in the map. 207 208 if (auto *C = dyn_cast<Constant>(V)) { 209 Constant *Elt = C->getAggregateElement(i); 210 211 if (!Elt) 212 LV.markOverdefined(); // Unknown sort of constant. 213 else if (isa<UndefValue>(Elt)) 214 ; // Undef values remain unknown. 215 else 216 LV.markConstant(Elt); // Constants are constant. 217 } 218 219 // All others are underdefined by default. 220 return LV; 221 } 222 223 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 224 /// work list if it is not already executable. 225 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest); 226 227 // getFeasibleSuccessors - Return a vector of booleans to indicate which 228 // successors are reachable from a given terminator instruction. 229 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 230 231 // OperandChangedState - This method is invoked on all of the users of an 232 // instruction that was just changed state somehow. Based on this 233 // information, we need to update the specified user of this instruction. 234 void operandChangedState(Instruction *I) { 235 if (BBExecutable.count(I->getParent())) // Inst is executable? 236 visit(*I); 237 } 238 239 // Add U as additional user of V. 240 void addAdditionalUser(Value *V, User *U) { 241 auto Iter = AdditionalUsers.insert({V, {}}); 242 Iter.first->second.insert(U); 243 } 244 245 // Mark I's users as changed, including AdditionalUsers. 246 void markUsersAsChanged(Value *I) { 247 // Functions include their arguments in the use-list. Changed function 248 // values mean that the result of the function changed. We only need to 249 // update the call sites with the new function result and do not have to 250 // propagate the call arguments. 251 if (isa<Function>(I)) { 252 for (User *U : I->users()) { 253 if (auto *CB = dyn_cast<CallBase>(U)) 254 handleCallResult(*CB); 255 } 256 } else { 257 for (User *U : I->users()) 258 if (auto *UI = dyn_cast<Instruction>(U)) 259 operandChangedState(UI); 260 } 261 262 auto Iter = AdditionalUsers.find(I); 263 if (Iter != AdditionalUsers.end()) { 264 // Copy additional users before notifying them of changes, because new 265 // users may be added, potentially invalidating the iterator. 266 SmallVector<Instruction *, 2> ToNotify; 267 for (User *U : Iter->second) 268 if (auto *UI = dyn_cast<Instruction>(U)) 269 ToNotify.push_back(UI); 270 for (Instruction *UI : ToNotify) 271 operandChangedState(UI); 272 } 273 } 274 void handleCallOverdefined(CallBase &CB); 275 void handleCallResult(CallBase &CB); 276 void handleCallArguments(CallBase &CB); 277 278 private: 279 friend class InstVisitor<SCCPInstVisitor>; 280 281 // visit implementations - Something changed in this instruction. Either an 282 // operand made a transition, or the instruction is newly executable. Change 283 // the value type of I to reflect these changes if appropriate. 284 void visitPHINode(PHINode &I); 285 286 // Terminators 287 288 void visitReturnInst(ReturnInst &I); 289 void visitTerminator(Instruction &TI); 290 291 void visitCastInst(CastInst &I); 292 void visitSelectInst(SelectInst &I); 293 void visitUnaryOperator(Instruction &I); 294 void visitBinaryOperator(Instruction &I); 295 void visitCmpInst(CmpInst &I); 296 void visitExtractValueInst(ExtractValueInst &EVI); 297 void visitInsertValueInst(InsertValueInst &IVI); 298 299 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 300 markOverdefined(&CPI); 301 visitTerminator(CPI); 302 } 303 304 // Instructions that cannot be folded away. 305 306 void visitStoreInst(StoreInst &I); 307 void visitLoadInst(LoadInst &I); 308 void visitGetElementPtrInst(GetElementPtrInst &I); 309 310 void visitInvokeInst(InvokeInst &II) { 311 visitCallBase(II); 312 visitTerminator(II); 313 } 314 315 void visitCallBrInst(CallBrInst &CBI) { 316 visitCallBase(CBI); 317 visitTerminator(CBI); 318 } 319 320 void visitCallBase(CallBase &CB); 321 void visitResumeInst(ResumeInst &I) { /*returns void*/ 322 } 323 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ 324 } 325 void visitFenceInst(FenceInst &I) { /*returns void*/ 326 } 327 328 void visitInstruction(Instruction &I); 329 330 public: 331 void addAnalysis(Function &F, AnalysisResultsForFn A) { 332 AnalysisResults.insert({&F, std::move(A)}); 333 } 334 335 void visitCallInst(CallInst &I) { visitCallBase(I); } 336 337 bool markBlockExecutable(BasicBlock *BB); 338 339 const PredicateBase *getPredicateInfoFor(Instruction *I) { 340 auto A = AnalysisResults.find(I->getParent()->getParent()); 341 if (A == AnalysisResults.end()) 342 return nullptr; 343 return A->second.PredInfo->getPredicateInfoFor(I); 344 } 345 346 DomTreeUpdater getDTU(Function &F) { 347 auto A = AnalysisResults.find(&F); 348 assert(A != AnalysisResults.end() && "Need analysis results for function."); 349 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 350 } 351 352 SCCPInstVisitor(const DataLayout &DL, 353 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 354 LLVMContext &Ctx) 355 : DL(DL), GetTLI(GetTLI), Ctx(Ctx) {} 356 357 void trackValueOfGlobalVariable(GlobalVariable *GV) { 358 // We only track the contents of scalar globals. 359 if (GV->getValueType()->isSingleValueType()) { 360 ValueLatticeElement &IV = TrackedGlobals[GV]; 361 if (!isa<UndefValue>(GV->getInitializer())) 362 IV.markConstant(GV->getInitializer()); 363 } 364 } 365 366 void addTrackedFunction(Function *F) { 367 // Add an entry, F -> undef. 368 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 369 MRVFunctionsTracked.insert(F); 370 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 371 TrackedMultipleRetVals.insert( 372 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 373 } else if (!F->getReturnType()->isVoidTy()) 374 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 375 } 376 377 void addToMustPreserveReturnsInFunctions(Function *F) { 378 MustPreserveReturnsInFunctions.insert(F); 379 } 380 381 bool mustPreserveReturn(Function *F) { 382 return MustPreserveReturnsInFunctions.count(F); 383 } 384 385 void addArgumentTrackedFunction(Function *F) { 386 TrackingIncomingArguments.insert(F); 387 } 388 389 bool isArgumentTrackedFunction(Function *F) { 390 return TrackingIncomingArguments.count(F); 391 } 392 393 void solve(); 394 395 bool resolvedUndefsIn(Function &F); 396 397 bool isBlockExecutable(BasicBlock *BB) const { 398 return BBExecutable.count(BB); 399 } 400 401 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 402 403 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 404 std::vector<ValueLatticeElement> StructValues; 405 auto *STy = dyn_cast<StructType>(V->getType()); 406 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 407 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 408 auto I = StructValueState.find(std::make_pair(V, i)); 409 assert(I != StructValueState.end() && "Value not in valuemap!"); 410 StructValues.push_back(I->second); 411 } 412 return StructValues; 413 } 414 415 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 416 417 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 418 assert(!V->getType()->isStructTy() && 419 "Should use getStructLatticeValueFor"); 420 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 421 ValueState.find(V); 422 assert(I != ValueState.end() && 423 "V not found in ValueState nor Paramstate map!"); 424 return I->second; 425 } 426 427 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 428 return TrackedRetVals; 429 } 430 431 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 432 return TrackedGlobals; 433 } 434 435 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 436 return MRVFunctionsTracked; 437 } 438 439 void markOverdefined(Value *V) { 440 if (auto *STy = dyn_cast<StructType>(V->getType())) 441 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 442 markOverdefined(getStructValueState(V, i), V); 443 else 444 markOverdefined(ValueState[V], V); 445 } 446 447 bool isStructLatticeConstant(Function *F, StructType *STy); 448 449 Constant *getConstant(const ValueLatticeElement &LV) const; 450 451 SmallPtrSetImpl<Function *> &getArgumentTrackedFunctions() { 452 return TrackingIncomingArguments; 453 } 454 455 void markArgInFuncSpecialization(Function *F, Argument *A, Constant *C); 456 457 void markFunctionUnreachable(Function *F) { 458 for (auto &BB : *F) 459 BBExecutable.erase(&BB); 460 } 461 }; 462 463 } // namespace llvm 464 465 bool SCCPInstVisitor::markBlockExecutable(BasicBlock *BB) { 466 if (!BBExecutable.insert(BB).second) 467 return false; 468 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 469 BBWorkList.push_back(BB); // Add the block to the work list! 470 return true; 471 } 472 473 void SCCPInstVisitor::pushToWorkList(ValueLatticeElement &IV, Value *V) { 474 if (IV.isOverdefined()) 475 return OverdefinedInstWorkList.push_back(V); 476 InstWorkList.push_back(V); 477 } 478 479 void SCCPInstVisitor::pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 480 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 481 pushToWorkList(IV, V); 482 } 483 484 bool SCCPInstVisitor::markConstant(ValueLatticeElement &IV, Value *V, 485 Constant *C, bool MayIncludeUndef) { 486 if (!IV.markConstant(C, MayIncludeUndef)) 487 return false; 488 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 489 pushToWorkList(IV, V); 490 return true; 491 } 492 493 bool SCCPInstVisitor::markOverdefined(ValueLatticeElement &IV, Value *V) { 494 if (!IV.markOverdefined()) 495 return false; 496 497 LLVM_DEBUG(dbgs() << "markOverdefined: "; 498 if (auto *F = dyn_cast<Function>(V)) dbgs() 499 << "Function '" << F->getName() << "'\n"; 500 else dbgs() << *V << '\n'); 501 // Only instructions go on the work list 502 pushToWorkList(IV, V); 503 return true; 504 } 505 506 bool SCCPInstVisitor::isStructLatticeConstant(Function *F, StructType *STy) { 507 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 508 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 509 assert(It != TrackedMultipleRetVals.end()); 510 ValueLatticeElement LV = It->second; 511 if (!isConstant(LV)) 512 return false; 513 } 514 return true; 515 } 516 517 Constant *SCCPInstVisitor::getConstant(const ValueLatticeElement &LV) const { 518 if (LV.isConstant()) 519 return LV.getConstant(); 520 521 if (LV.isConstantRange()) { 522 const auto &CR = LV.getConstantRange(); 523 if (CR.getSingleElement()) 524 return ConstantInt::get(Ctx, *CR.getSingleElement()); 525 } 526 return nullptr; 527 } 528 529 void SCCPInstVisitor::markArgInFuncSpecialization(Function *F, Argument *A, 530 Constant *C) { 531 assert(F->arg_size() == A->getParent()->arg_size() && 532 "Functions should have the same number of arguments"); 533 534 // Mark the argument constant in the new function. 535 markConstant(A, C); 536 537 // For the remaining arguments in the new function, copy the lattice state 538 // over from the old function. 539 for (auto I = F->arg_begin(), J = A->getParent()->arg_begin(), 540 E = F->arg_end(); 541 I != E; ++I, ++J) 542 if (J != A && ValueState.count(I)) { 543 // Note: This previously looked like this: 544 // ValueState[J] = ValueState[I]; 545 // This is incorrect because the DenseMap class may resize the underlying 546 // memory when inserting `J`, which will invalidate the reference to `I`. 547 // Instead, we make sure `J` exists, then set it to `I` afterwards. 548 auto &NewValue = ValueState[J]; 549 NewValue = ValueState[I]; 550 pushToWorkList(NewValue, J); 551 } 552 } 553 554 void SCCPInstVisitor::visitInstruction(Instruction &I) { 555 // All the instructions we don't do any special handling for just 556 // go to overdefined. 557 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 558 markOverdefined(&I); 559 } 560 561 bool SCCPInstVisitor::mergeInValue(ValueLatticeElement &IV, Value *V, 562 ValueLatticeElement MergeWithV, 563 ValueLatticeElement::MergeOptions Opts) { 564 if (IV.mergeIn(MergeWithV, Opts)) { 565 pushToWorkList(IV, V); 566 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 567 << IV << "\n"); 568 return true; 569 } 570 return false; 571 } 572 573 bool SCCPInstVisitor::markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 574 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 575 return false; // This edge is already known to be executable! 576 577 if (!markBlockExecutable(Dest)) { 578 // If the destination is already executable, we just made an *edge* 579 // feasible that wasn't before. Revisit the PHI nodes in the block 580 // because they have potentially new operands. 581 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 582 << " -> " << Dest->getName() << '\n'); 583 584 for (PHINode &PN : Dest->phis()) 585 visitPHINode(PN); 586 } 587 return true; 588 } 589 590 // getFeasibleSuccessors - Return a vector of booleans to indicate which 591 // successors are reachable from a given terminator instruction. 592 void SCCPInstVisitor::getFeasibleSuccessors(Instruction &TI, 593 SmallVectorImpl<bool> &Succs) { 594 Succs.resize(TI.getNumSuccessors()); 595 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 596 if (BI->isUnconditional()) { 597 Succs[0] = true; 598 return; 599 } 600 601 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 602 ConstantInt *CI = getConstantInt(BCValue); 603 if (!CI) { 604 // Overdefined condition variables, and branches on unfoldable constant 605 // conditions, mean the branch could go either way. 606 if (!BCValue.isUnknownOrUndef()) 607 Succs[0] = Succs[1] = true; 608 return; 609 } 610 611 // Constant condition variables mean the branch can only go a single way. 612 Succs[CI->isZero()] = true; 613 return; 614 } 615 616 // Unwinding instructions successors are always executable. 617 if (TI.isExceptionalTerminator()) { 618 Succs.assign(TI.getNumSuccessors(), true); 619 return; 620 } 621 622 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 623 if (!SI->getNumCases()) { 624 Succs[0] = true; 625 return; 626 } 627 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 628 if (ConstantInt *CI = getConstantInt(SCValue)) { 629 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 630 return; 631 } 632 633 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 634 // is ready. 635 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 636 const ConstantRange &Range = SCValue.getConstantRange(); 637 for (const auto &Case : SI->cases()) { 638 const APInt &CaseValue = Case.getCaseValue()->getValue(); 639 if (Range.contains(CaseValue)) 640 Succs[Case.getSuccessorIndex()] = true; 641 } 642 643 // TODO: Determine whether default case is reachable. 644 Succs[SI->case_default()->getSuccessorIndex()] = true; 645 return; 646 } 647 648 // Overdefined or unknown condition? All destinations are executable! 649 if (!SCValue.isUnknownOrUndef()) 650 Succs.assign(TI.getNumSuccessors(), true); 651 return; 652 } 653 654 // In case of indirect branch and its address is a blockaddress, we mark 655 // the target as executable. 656 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 657 // Casts are folded by visitCastInst. 658 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 659 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 660 if (!Addr) { // Overdefined or unknown condition? 661 // All destinations are executable! 662 if (!IBRValue.isUnknownOrUndef()) 663 Succs.assign(TI.getNumSuccessors(), true); 664 return; 665 } 666 667 BasicBlock *T = Addr->getBasicBlock(); 668 assert(Addr->getFunction() == T->getParent() && 669 "Block address of a different function ?"); 670 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 671 // This is the target. 672 if (IBR->getDestination(i) == T) { 673 Succs[i] = true; 674 return; 675 } 676 } 677 678 // If we didn't find our destination in the IBR successor list, then we 679 // have undefined behavior. Its ok to assume no successor is executable. 680 return; 681 } 682 683 // In case of callbr, we pessimistically assume that all successors are 684 // feasible. 685 if (isa<CallBrInst>(&TI)) { 686 Succs.assign(TI.getNumSuccessors(), true); 687 return; 688 } 689 690 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 691 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 692 } 693 694 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 695 // block to the 'To' basic block is currently feasible. 696 bool SCCPInstVisitor::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 697 // Check if we've called markEdgeExecutable on the edge yet. (We could 698 // be more aggressive and try to consider edges which haven't been marked 699 // yet, but there isn't any need.) 700 return KnownFeasibleEdges.count(Edge(From, To)); 701 } 702 703 // visit Implementations - Something changed in this instruction, either an 704 // operand made a transition, or the instruction is newly executable. Change 705 // the value type of I to reflect these changes if appropriate. This method 706 // makes sure to do the following actions: 707 // 708 // 1. If a phi node merges two constants in, and has conflicting value coming 709 // from different branches, or if the PHI node merges in an overdefined 710 // value, then the PHI node becomes overdefined. 711 // 2. If a phi node merges only constants in, and they all agree on value, the 712 // PHI node becomes a constant value equal to that. 713 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 714 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 715 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 716 // 6. If a conditional branch has a value that is constant, make the selected 717 // destination executable 718 // 7. If a conditional branch has a value that is overdefined, make all 719 // successors executable. 720 void SCCPInstVisitor::visitPHINode(PHINode &PN) { 721 // If this PN returns a struct, just mark the result overdefined. 722 // TODO: We could do a lot better than this if code actually uses this. 723 if (PN.getType()->isStructTy()) 724 return (void)markOverdefined(&PN); 725 726 if (getValueState(&PN).isOverdefined()) 727 return; // Quick exit 728 729 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 730 // and slow us down a lot. Just mark them overdefined. 731 if (PN.getNumIncomingValues() > 64) 732 return (void)markOverdefined(&PN); 733 734 unsigned NumActiveIncoming = 0; 735 736 // Look at all of the executable operands of the PHI node. If any of them 737 // are overdefined, the PHI becomes overdefined as well. If they are all 738 // constant, and they agree with each other, the PHI becomes the identical 739 // constant. If they are constant and don't agree, the PHI is a constant 740 // range. If there are no executable operands, the PHI remains unknown. 741 ValueLatticeElement PhiState = getValueState(&PN); 742 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 743 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 744 continue; 745 746 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 747 PhiState.mergeIn(IV); 748 NumActiveIncoming++; 749 if (PhiState.isOverdefined()) 750 break; 751 } 752 753 // We allow up to 1 range extension per active incoming value and one 754 // additional extension. Note that we manually adjust the number of range 755 // extensions to match the number of active incoming values. This helps to 756 // limit multiple extensions caused by the same incoming value, if other 757 // incoming values are equal. 758 mergeInValue(&PN, PhiState, 759 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 760 NumActiveIncoming + 1)); 761 ValueLatticeElement &PhiStateRef = getValueState(&PN); 762 PhiStateRef.setNumRangeExtensions( 763 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 764 } 765 766 void SCCPInstVisitor::visitReturnInst(ReturnInst &I) { 767 if (I.getNumOperands() == 0) 768 return; // ret void 769 770 Function *F = I.getParent()->getParent(); 771 Value *ResultOp = I.getOperand(0); 772 773 // If we are tracking the return value of this function, merge it in. 774 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 775 auto TFRVI = TrackedRetVals.find(F); 776 if (TFRVI != TrackedRetVals.end()) { 777 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 778 return; 779 } 780 } 781 782 // Handle functions that return multiple values. 783 if (!TrackedMultipleRetVals.empty()) { 784 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 785 if (MRVFunctionsTracked.count(F)) 786 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 787 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 788 getStructValueState(ResultOp, i)); 789 } 790 } 791 792 void SCCPInstVisitor::visitTerminator(Instruction &TI) { 793 SmallVector<bool, 16> SuccFeasible; 794 getFeasibleSuccessors(TI, SuccFeasible); 795 796 BasicBlock *BB = TI.getParent(); 797 798 // Mark all feasible successors executable. 799 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 800 if (SuccFeasible[i]) 801 markEdgeExecutable(BB, TI.getSuccessor(i)); 802 } 803 804 void SCCPInstVisitor::visitCastInst(CastInst &I) { 805 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 806 // discover a concrete value later. 807 if (ValueState[&I].isOverdefined()) 808 return; 809 810 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 811 if (Constant *OpC = getConstant(OpSt)) { 812 // Fold the constant as we build. 813 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 814 if (isa<UndefValue>(C)) 815 return; 816 // Propagate constant value 817 markConstant(&I, C); 818 } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) { 819 auto &LV = getValueState(&I); 820 ConstantRange OpRange = OpSt.getConstantRange(); 821 Type *DestTy = I.getDestTy(); 822 // Vectors where all elements have the same known constant range are treated 823 // as a single constant range in the lattice. When bitcasting such vectors, 824 // there is a mis-match between the width of the lattice value (single 825 // constant range) and the original operands (vector). Go to overdefined in 826 // that case. 827 if (I.getOpcode() == Instruction::BitCast && 828 I.getOperand(0)->getType()->isVectorTy() && 829 OpRange.getBitWidth() < DL.getTypeSizeInBits(DestTy)) 830 return (void)markOverdefined(&I); 831 832 ConstantRange Res = 833 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 834 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 835 } else if (!OpSt.isUnknownOrUndef()) 836 markOverdefined(&I); 837 } 838 839 void SCCPInstVisitor::visitExtractValueInst(ExtractValueInst &EVI) { 840 // If this returns a struct, mark all elements over defined, we don't track 841 // structs in structs. 842 if (EVI.getType()->isStructTy()) 843 return (void)markOverdefined(&EVI); 844 845 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 846 // discover a concrete value later. 847 if (ValueState[&EVI].isOverdefined()) 848 return (void)markOverdefined(&EVI); 849 850 // If this is extracting from more than one level of struct, we don't know. 851 if (EVI.getNumIndices() != 1) 852 return (void)markOverdefined(&EVI); 853 854 Value *AggVal = EVI.getAggregateOperand(); 855 if (AggVal->getType()->isStructTy()) { 856 unsigned i = *EVI.idx_begin(); 857 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 858 mergeInValue(getValueState(&EVI), &EVI, EltVal); 859 } else { 860 // Otherwise, must be extracting from an array. 861 return (void)markOverdefined(&EVI); 862 } 863 } 864 865 void SCCPInstVisitor::visitInsertValueInst(InsertValueInst &IVI) { 866 auto *STy = dyn_cast<StructType>(IVI.getType()); 867 if (!STy) 868 return (void)markOverdefined(&IVI); 869 870 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 871 // discover a concrete value later. 872 if (isOverdefined(ValueState[&IVI])) 873 return (void)markOverdefined(&IVI); 874 875 // If this has more than one index, we can't handle it, drive all results to 876 // undef. 877 if (IVI.getNumIndices() != 1) 878 return (void)markOverdefined(&IVI); 879 880 Value *Aggr = IVI.getAggregateOperand(); 881 unsigned Idx = *IVI.idx_begin(); 882 883 // Compute the result based on what we're inserting. 884 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 885 // This passes through all values that aren't the inserted element. 886 if (i != Idx) { 887 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 888 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 889 continue; 890 } 891 892 Value *Val = IVI.getInsertedValueOperand(); 893 if (Val->getType()->isStructTy()) 894 // We don't track structs in structs. 895 markOverdefined(getStructValueState(&IVI, i), &IVI); 896 else { 897 ValueLatticeElement InVal = getValueState(Val); 898 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 899 } 900 } 901 } 902 903 void SCCPInstVisitor::visitSelectInst(SelectInst &I) { 904 // If this select returns a struct, just mark the result overdefined. 905 // TODO: We could do a lot better than this if code actually uses this. 906 if (I.getType()->isStructTy()) 907 return (void)markOverdefined(&I); 908 909 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 910 // discover a concrete value later. 911 if (ValueState[&I].isOverdefined()) 912 return (void)markOverdefined(&I); 913 914 ValueLatticeElement CondValue = getValueState(I.getCondition()); 915 if (CondValue.isUnknownOrUndef()) 916 return; 917 918 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 919 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 920 mergeInValue(&I, getValueState(OpVal)); 921 return; 922 } 923 924 // Otherwise, the condition is overdefined or a constant we can't evaluate. 925 // See if we can produce something better than overdefined based on the T/F 926 // value. 927 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 928 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 929 930 bool Changed = ValueState[&I].mergeIn(TVal); 931 Changed |= ValueState[&I].mergeIn(FVal); 932 if (Changed) 933 pushToWorkListMsg(ValueState[&I], &I); 934 } 935 936 // Handle Unary Operators. 937 void SCCPInstVisitor::visitUnaryOperator(Instruction &I) { 938 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 939 940 ValueLatticeElement &IV = ValueState[&I]; 941 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 942 // discover a concrete value later. 943 if (isOverdefined(IV)) 944 return (void)markOverdefined(&I); 945 946 if (isConstant(V0State)) { 947 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 948 949 // op Y -> undef. 950 if (isa<UndefValue>(C)) 951 return; 952 return (void)markConstant(IV, &I, C); 953 } 954 955 // If something is undef, wait for it to resolve. 956 if (!isOverdefined(V0State)) 957 return; 958 959 markOverdefined(&I); 960 } 961 962 // Handle Binary Operators. 963 void SCCPInstVisitor::visitBinaryOperator(Instruction &I) { 964 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 965 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 966 967 ValueLatticeElement &IV = ValueState[&I]; 968 if (IV.isOverdefined()) 969 return; 970 971 // If something is undef, wait for it to resolve. 972 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 973 return; 974 975 if (V1State.isOverdefined() && V2State.isOverdefined()) 976 return (void)markOverdefined(&I); 977 978 // If either of the operands is a constant, try to fold it to a constant. 979 // TODO: Use information from notconstant better. 980 if ((V1State.isConstant() || V2State.isConstant())) { 981 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 982 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 983 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 984 auto *C = dyn_cast_or_null<Constant>(R); 985 if (C) { 986 // X op Y -> undef. 987 if (isa<UndefValue>(C)) 988 return; 989 // Conservatively assume that the result may be based on operands that may 990 // be undef. Note that we use mergeInValue to combine the constant with 991 // the existing lattice value for I, as different constants might be found 992 // after one of the operands go to overdefined, e.g. due to one operand 993 // being a special floating value. 994 ValueLatticeElement NewV; 995 NewV.markConstant(C, /*MayIncludeUndef=*/true); 996 return (void)mergeInValue(&I, NewV); 997 } 998 } 999 1000 // Only use ranges for binary operators on integers. 1001 if (!I.getType()->isIntegerTy()) 1002 return markOverdefined(&I); 1003 1004 // Try to simplify to a constant range. 1005 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1006 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1007 if (V1State.isConstantRange()) 1008 A = V1State.getConstantRange(); 1009 if (V2State.isConstantRange()) 1010 B = V2State.getConstantRange(); 1011 1012 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1013 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1014 1015 // TODO: Currently we do not exploit special values that produce something 1016 // better than overdefined with an overdefined operand for vector or floating 1017 // point types, like and <4 x i32> overdefined, zeroinitializer. 1018 } 1019 1020 // Handle ICmpInst instruction. 1021 void SCCPInstVisitor::visitCmpInst(CmpInst &I) { 1022 // Do not cache this lookup, getValueState calls later in the function might 1023 // invalidate the reference. 1024 if (isOverdefined(ValueState[&I])) 1025 return (void)markOverdefined(&I); 1026 1027 Value *Op1 = I.getOperand(0); 1028 Value *Op2 = I.getOperand(1); 1029 1030 // For parameters, use ParamState which includes constant range info if 1031 // available. 1032 auto V1State = getValueState(Op1); 1033 auto V2State = getValueState(Op2); 1034 1035 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1036 if (C) { 1037 if (isa<UndefValue>(C)) 1038 return; 1039 ValueLatticeElement CV; 1040 CV.markConstant(C); 1041 mergeInValue(&I, CV); 1042 return; 1043 } 1044 1045 // If operands are still unknown, wait for it to resolve. 1046 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1047 !isConstant(ValueState[&I])) 1048 return; 1049 1050 markOverdefined(&I); 1051 } 1052 1053 // Handle getelementptr instructions. If all operands are constants then we 1054 // can turn this into a getelementptr ConstantExpr. 1055 void SCCPInstVisitor::visitGetElementPtrInst(GetElementPtrInst &I) { 1056 if (isOverdefined(ValueState[&I])) 1057 return (void)markOverdefined(&I); 1058 1059 SmallVector<Constant *, 8> Operands; 1060 Operands.reserve(I.getNumOperands()); 1061 1062 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1063 ValueLatticeElement State = getValueState(I.getOperand(i)); 1064 if (State.isUnknownOrUndef()) 1065 return; // Operands are not resolved yet. 1066 1067 if (isOverdefined(State)) 1068 return (void)markOverdefined(&I); 1069 1070 if (Constant *C = getConstant(State)) { 1071 Operands.push_back(C); 1072 continue; 1073 } 1074 1075 return (void)markOverdefined(&I); 1076 } 1077 1078 Constant *Ptr = Operands[0]; 1079 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1080 Constant *C = 1081 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1082 if (isa<UndefValue>(C)) 1083 return; 1084 markConstant(&I, C); 1085 } 1086 1087 void SCCPInstVisitor::visitStoreInst(StoreInst &SI) { 1088 // If this store is of a struct, ignore it. 1089 if (SI.getOperand(0)->getType()->isStructTy()) 1090 return; 1091 1092 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1093 return; 1094 1095 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1096 auto I = TrackedGlobals.find(GV); 1097 if (I == TrackedGlobals.end()) 1098 return; 1099 1100 // Get the value we are storing into the global, then merge it. 1101 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1102 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1103 if (I->second.isOverdefined()) 1104 TrackedGlobals.erase(I); // No need to keep tracking this! 1105 } 1106 1107 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1108 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1109 if (I->getType()->isIntegerTy()) 1110 return ValueLatticeElement::getRange( 1111 getConstantRangeFromMetadata(*Ranges)); 1112 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1113 return ValueLatticeElement::getNot( 1114 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1115 return ValueLatticeElement::getOverdefined(); 1116 } 1117 1118 // Handle load instructions. If the operand is a constant pointer to a constant 1119 // global, we can replace the load with the loaded constant value! 1120 void SCCPInstVisitor::visitLoadInst(LoadInst &I) { 1121 // If this load is of a struct or the load is volatile, just mark the result 1122 // as overdefined. 1123 if (I.getType()->isStructTy() || I.isVolatile()) 1124 return (void)markOverdefined(&I); 1125 1126 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1127 // discover a concrete value later. 1128 if (ValueState[&I].isOverdefined()) 1129 return (void)markOverdefined(&I); 1130 1131 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1132 if (PtrVal.isUnknownOrUndef()) 1133 return; // The pointer is not resolved yet! 1134 1135 ValueLatticeElement &IV = ValueState[&I]; 1136 1137 if (isConstant(PtrVal)) { 1138 Constant *Ptr = getConstant(PtrVal); 1139 1140 // load null is undefined. 1141 if (isa<ConstantPointerNull>(Ptr)) { 1142 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1143 return (void)markOverdefined(IV, &I); 1144 else 1145 return; 1146 } 1147 1148 // Transform load (constant global) into the value loaded. 1149 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1150 if (!TrackedGlobals.empty()) { 1151 // If we are tracking this global, merge in the known value for it. 1152 auto It = TrackedGlobals.find(GV); 1153 if (It != TrackedGlobals.end()) { 1154 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1155 return; 1156 } 1157 } 1158 } 1159 1160 // Transform load from a constant into a constant if possible. 1161 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1162 if (isa<UndefValue>(C)) 1163 return; 1164 return (void)markConstant(IV, &I, C); 1165 } 1166 } 1167 1168 // Fall back to metadata. 1169 mergeInValue(&I, getValueFromMetadata(&I)); 1170 } 1171 1172 void SCCPInstVisitor::visitCallBase(CallBase &CB) { 1173 handleCallResult(CB); 1174 handleCallArguments(CB); 1175 } 1176 1177 void SCCPInstVisitor::handleCallOverdefined(CallBase &CB) { 1178 Function *F = CB.getCalledFunction(); 1179 1180 // Void return and not tracking callee, just bail. 1181 if (CB.getType()->isVoidTy()) 1182 return; 1183 1184 // Always mark struct return as overdefined. 1185 if (CB.getType()->isStructTy()) 1186 return (void)markOverdefined(&CB); 1187 1188 // Otherwise, if we have a single return value case, and if the function is 1189 // a declaration, maybe we can constant fold it. 1190 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1191 SmallVector<Constant *, 8> Operands; 1192 for (auto AI = CB.arg_begin(), E = CB.arg_end(); AI != E; ++AI) { 1193 if (AI->get()->getType()->isStructTy()) 1194 return markOverdefined(&CB); // Can't handle struct args. 1195 ValueLatticeElement State = getValueState(*AI); 1196 1197 if (State.isUnknownOrUndef()) 1198 return; // Operands are not resolved yet. 1199 if (isOverdefined(State)) 1200 return (void)markOverdefined(&CB); 1201 assert(isConstant(State) && "Unknown state!"); 1202 Operands.push_back(getConstant(State)); 1203 } 1204 1205 if (isOverdefined(getValueState(&CB))) 1206 return (void)markOverdefined(&CB); 1207 1208 // If we can constant fold this, mark the result of the call as a 1209 // constant. 1210 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1211 // call -> undef. 1212 if (isa<UndefValue>(C)) 1213 return; 1214 return (void)markConstant(&CB, C); 1215 } 1216 } 1217 1218 // Fall back to metadata. 1219 mergeInValue(&CB, getValueFromMetadata(&CB)); 1220 } 1221 1222 void SCCPInstVisitor::handleCallArguments(CallBase &CB) { 1223 Function *F = CB.getCalledFunction(); 1224 // If this is a local function that doesn't have its address taken, mark its 1225 // entry block executable and merge in the actual arguments to the call into 1226 // the formal arguments of the function. 1227 if (!TrackingIncomingArguments.empty() && 1228 TrackingIncomingArguments.count(F)) { 1229 markBlockExecutable(&F->front()); 1230 1231 // Propagate information from this call site into the callee. 1232 auto CAI = CB.arg_begin(); 1233 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1234 ++AI, ++CAI) { 1235 // If this argument is byval, and if the function is not readonly, there 1236 // will be an implicit copy formed of the input aggregate. 1237 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1238 markOverdefined(&*AI); 1239 continue; 1240 } 1241 1242 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1243 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1244 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1245 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1246 getMaxWidenStepsOpts()); 1247 } 1248 } else 1249 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1250 } 1251 } 1252 } 1253 1254 void SCCPInstVisitor::handleCallResult(CallBase &CB) { 1255 Function *F = CB.getCalledFunction(); 1256 1257 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1258 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1259 if (ValueState[&CB].isOverdefined()) 1260 return; 1261 1262 Value *CopyOf = CB.getOperand(0); 1263 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1264 const auto *PI = getPredicateInfoFor(&CB); 1265 assert(PI && "Missing predicate info for ssa.copy"); 1266 1267 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1268 if (!Constraint) { 1269 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1270 return; 1271 } 1272 1273 CmpInst::Predicate Pred = Constraint->Predicate; 1274 Value *OtherOp = Constraint->OtherOp; 1275 1276 // Wait until OtherOp is resolved. 1277 if (getValueState(OtherOp).isUnknown()) { 1278 addAdditionalUser(OtherOp, &CB); 1279 return; 1280 } 1281 1282 // TODO: Actually filp MayIncludeUndef for the created range to false, 1283 // once most places in the optimizer respect the branches on 1284 // undef/poison are UB rule. The reason why the new range cannot be 1285 // undef is as follows below: 1286 // The new range is based on a branch condition. That guarantees that 1287 // neither of the compare operands can be undef in the branch targets, 1288 // unless we have conditions that are always true/false (e.g. icmp ule 1289 // i32, %a, i32_max). For the latter overdefined/empty range will be 1290 // inferred, but the branch will get folded accordingly anyways. 1291 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1292 1293 ValueLatticeElement CondVal = getValueState(OtherOp); 1294 ValueLatticeElement &IV = ValueState[&CB]; 1295 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1296 auto ImposedCR = 1297 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1298 1299 // Get the range imposed by the condition. 1300 if (CondVal.isConstantRange()) 1301 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1302 Pred, CondVal.getConstantRange()); 1303 1304 // Combine range info for the original value with the new range from the 1305 // condition. 1306 auto CopyOfCR = CopyOfVal.isConstantRange() 1307 ? CopyOfVal.getConstantRange() 1308 : ConstantRange::getFull( 1309 DL.getTypeSizeInBits(CopyOf->getType())); 1310 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1311 // If the existing information is != x, do not use the information from 1312 // a chained predicate, as the != x information is more likely to be 1313 // helpful in practice. 1314 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1315 NewCR = CopyOfCR; 1316 1317 addAdditionalUser(OtherOp, &CB); 1318 mergeInValue(IV, &CB, 1319 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1320 return; 1321 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1322 // For non-integer values or integer constant expressions, only 1323 // propagate equal constants. 1324 addAdditionalUser(OtherOp, &CB); 1325 mergeInValue(IV, &CB, CondVal); 1326 return; 1327 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1328 !MayIncludeUndef) { 1329 // Propagate inequalities. 1330 addAdditionalUser(OtherOp, &CB); 1331 mergeInValue(IV, &CB, 1332 ValueLatticeElement::getNot(CondVal.getConstant())); 1333 return; 1334 } 1335 1336 return (void)mergeInValue(IV, &CB, CopyOfVal); 1337 } 1338 1339 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { 1340 // Compute result range for intrinsics supported by ConstantRange. 1341 // Do this even if we don't know a range for all operands, as we may 1342 // still know something about the result range, e.g. of abs(x). 1343 SmallVector<ConstantRange, 2> OpRanges; 1344 for (Value *Op : II->args()) { 1345 const ValueLatticeElement &State = getValueState(Op); 1346 if (State.isConstantRange()) 1347 OpRanges.push_back(State.getConstantRange()); 1348 else 1349 OpRanges.push_back( 1350 ConstantRange::getFull(Op->getType()->getScalarSizeInBits())); 1351 } 1352 1353 ConstantRange Result = 1354 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges); 1355 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result)); 1356 } 1357 } 1358 1359 // The common case is that we aren't tracking the callee, either because we 1360 // are not doing interprocedural analysis or the callee is indirect, or is 1361 // external. Handle these cases first. 1362 if (!F || F->isDeclaration()) 1363 return handleCallOverdefined(CB); 1364 1365 // If this is a single/zero retval case, see if we're tracking the function. 1366 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1367 if (!MRVFunctionsTracked.count(F)) 1368 return handleCallOverdefined(CB); // Not tracking this callee. 1369 1370 // If we are tracking this callee, propagate the result of the function 1371 // into this call site. 1372 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1373 mergeInValue(getStructValueState(&CB, i), &CB, 1374 TrackedMultipleRetVals[std::make_pair(F, i)], 1375 getMaxWidenStepsOpts()); 1376 } else { 1377 auto TFRVI = TrackedRetVals.find(F); 1378 if (TFRVI == TrackedRetVals.end()) 1379 return handleCallOverdefined(CB); // Not tracking this callee. 1380 1381 // If so, propagate the return value of the callee into this call result. 1382 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1383 } 1384 } 1385 1386 void SCCPInstVisitor::solve() { 1387 // Process the work lists until they are empty! 1388 while (!BBWorkList.empty() || !InstWorkList.empty() || 1389 !OverdefinedInstWorkList.empty()) { 1390 // Process the overdefined instruction's work list first, which drives other 1391 // things to overdefined more quickly. 1392 while (!OverdefinedInstWorkList.empty()) { 1393 Value *I = OverdefinedInstWorkList.pop_back_val(); 1394 1395 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1396 1397 // "I" got into the work list because it either made the transition from 1398 // bottom to constant, or to overdefined. 1399 // 1400 // Anything on this worklist that is overdefined need not be visited 1401 // since all of its users will have already been marked as overdefined 1402 // Update all of the users of this instruction's value. 1403 // 1404 markUsersAsChanged(I); 1405 } 1406 1407 // Process the instruction work list. 1408 while (!InstWorkList.empty()) { 1409 Value *I = InstWorkList.pop_back_val(); 1410 1411 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1412 1413 // "I" got into the work list because it made the transition from undef to 1414 // constant. 1415 // 1416 // Anything on this worklist that is overdefined need not be visited 1417 // since all of its users will have already been marked as overdefined. 1418 // Update all of the users of this instruction's value. 1419 // 1420 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1421 markUsersAsChanged(I); 1422 } 1423 1424 // Process the basic block work list. 1425 while (!BBWorkList.empty()) { 1426 BasicBlock *BB = BBWorkList.pop_back_val(); 1427 1428 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1429 1430 // Notify all instructions in this basic block that they are newly 1431 // executable. 1432 visit(BB); 1433 } 1434 } 1435 } 1436 1437 /// resolvedUndefsIn - While solving the dataflow for a function, we assume 1438 /// that branches on undef values cannot reach any of their successors. 1439 /// However, this is not a safe assumption. After we solve dataflow, this 1440 /// method should be use to handle this. If this returns true, the solver 1441 /// should be rerun. 1442 /// 1443 /// This method handles this by finding an unresolved branch and marking it one 1444 /// of the edges from the block as being feasible, even though the condition 1445 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1446 /// CFG and only slightly pessimizes the analysis results (by marking one, 1447 /// potentially infeasible, edge feasible). This cannot usefully modify the 1448 /// constraints on the condition of the branch, as that would impact other users 1449 /// of the value. 1450 /// 1451 /// This scan also checks for values that use undefs. It conservatively marks 1452 /// them as overdefined. 1453 bool SCCPInstVisitor::resolvedUndefsIn(Function &F) { 1454 bool MadeChange = false; 1455 for (BasicBlock &BB : F) { 1456 if (!BBExecutable.count(&BB)) 1457 continue; 1458 1459 for (Instruction &I : BB) { 1460 // Look for instructions which produce undef values. 1461 if (I.getType()->isVoidTy()) 1462 continue; 1463 1464 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1465 // Only a few things that can be structs matter for undef. 1466 1467 // Tracked calls must never be marked overdefined in resolvedUndefsIn. 1468 if (auto *CB = dyn_cast<CallBase>(&I)) 1469 if (Function *F = CB->getCalledFunction()) 1470 if (MRVFunctionsTracked.count(F)) 1471 continue; 1472 1473 // extractvalue and insertvalue don't need to be marked; they are 1474 // tracked as precisely as their operands. 1475 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1476 continue; 1477 // Send the results of everything else to overdefined. We could be 1478 // more precise than this but it isn't worth bothering. 1479 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1480 ValueLatticeElement &LV = getStructValueState(&I, i); 1481 if (LV.isUnknownOrUndef()) { 1482 markOverdefined(LV, &I); 1483 MadeChange = true; 1484 } 1485 } 1486 continue; 1487 } 1488 1489 ValueLatticeElement &LV = getValueState(&I); 1490 if (!LV.isUnknownOrUndef()) 1491 continue; 1492 1493 // There are two reasons a call can have an undef result 1494 // 1. It could be tracked. 1495 // 2. It could be constant-foldable. 1496 // Because of the way we solve return values, tracked calls must 1497 // never be marked overdefined in resolvedUndefsIn. 1498 if (auto *CB = dyn_cast<CallBase>(&I)) 1499 if (Function *F = CB->getCalledFunction()) 1500 if (TrackedRetVals.count(F)) 1501 continue; 1502 1503 if (isa<LoadInst>(I)) { 1504 // A load here means one of two things: a load of undef from a global, 1505 // a load from an unknown pointer. Either way, having it return undef 1506 // is okay. 1507 continue; 1508 } 1509 1510 markOverdefined(&I); 1511 MadeChange = true; 1512 } 1513 1514 // Check to see if we have a branch or switch on an undefined value. If so 1515 // we force the branch to go one way or the other to make the successor 1516 // values live. It doesn't really matter which way we force it. 1517 Instruction *TI = BB.getTerminator(); 1518 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1519 if (!BI->isConditional()) 1520 continue; 1521 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1522 continue; 1523 1524 // If the input to SCCP is actually branch on undef, fix the undef to 1525 // false. 1526 if (isa<UndefValue>(BI->getCondition())) { 1527 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1528 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1529 MadeChange = true; 1530 continue; 1531 } 1532 1533 // Otherwise, it is a branch on a symbolic value which is currently 1534 // considered to be undef. Make sure some edge is executable, so a 1535 // branch on "undef" always flows somewhere. 1536 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1537 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1538 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1539 MadeChange = true; 1540 1541 continue; 1542 } 1543 1544 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1545 // Indirect branch with no successor ?. Its ok to assume it branches 1546 // to no target. 1547 if (IBR->getNumSuccessors() < 1) 1548 continue; 1549 1550 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1551 continue; 1552 1553 // If the input to SCCP is actually branch on undef, fix the undef to 1554 // the first successor of the indirect branch. 1555 if (isa<UndefValue>(IBR->getAddress())) { 1556 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1557 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1558 MadeChange = true; 1559 continue; 1560 } 1561 1562 // Otherwise, it is a branch on a symbolic value which is currently 1563 // considered to be undef. Make sure some edge is executable, so a 1564 // branch on "undef" always flows somewhere. 1565 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1566 // we can assume the branch has undefined behavior instead. 1567 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1568 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1569 MadeChange = true; 1570 1571 continue; 1572 } 1573 1574 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1575 if (!SI->getNumCases() || 1576 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1577 continue; 1578 1579 // If the input to SCCP is actually switch on undef, fix the undef to 1580 // the first constant. 1581 if (isa<UndefValue>(SI->getCondition())) { 1582 SI->setCondition(SI->case_begin()->getCaseValue()); 1583 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1584 MadeChange = true; 1585 continue; 1586 } 1587 1588 // Otherwise, it is a branch on a symbolic value which is currently 1589 // considered to be undef. Make sure some edge is executable, so a 1590 // branch on "undef" always flows somewhere. 1591 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1592 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1593 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1594 MadeChange = true; 1595 1596 continue; 1597 } 1598 } 1599 1600 return MadeChange; 1601 } 1602 1603 //===----------------------------------------------------------------------===// 1604 // 1605 // SCCPSolver implementations 1606 // 1607 SCCPSolver::SCCPSolver( 1608 const DataLayout &DL, 1609 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1610 LLVMContext &Ctx) 1611 : Visitor(new SCCPInstVisitor(DL, std::move(GetTLI), Ctx)) {} 1612 1613 SCCPSolver::~SCCPSolver() {} 1614 1615 void SCCPSolver::addAnalysis(Function &F, AnalysisResultsForFn A) { 1616 return Visitor->addAnalysis(F, std::move(A)); 1617 } 1618 1619 bool SCCPSolver::markBlockExecutable(BasicBlock *BB) { 1620 return Visitor->markBlockExecutable(BB); 1621 } 1622 1623 const PredicateBase *SCCPSolver::getPredicateInfoFor(Instruction *I) { 1624 return Visitor->getPredicateInfoFor(I); 1625 } 1626 1627 DomTreeUpdater SCCPSolver::getDTU(Function &F) { return Visitor->getDTU(F); } 1628 1629 void SCCPSolver::trackValueOfGlobalVariable(GlobalVariable *GV) { 1630 Visitor->trackValueOfGlobalVariable(GV); 1631 } 1632 1633 void SCCPSolver::addTrackedFunction(Function *F) { 1634 Visitor->addTrackedFunction(F); 1635 } 1636 1637 void SCCPSolver::addToMustPreserveReturnsInFunctions(Function *F) { 1638 Visitor->addToMustPreserveReturnsInFunctions(F); 1639 } 1640 1641 bool SCCPSolver::mustPreserveReturn(Function *F) { 1642 return Visitor->mustPreserveReturn(F); 1643 } 1644 1645 void SCCPSolver::addArgumentTrackedFunction(Function *F) { 1646 Visitor->addArgumentTrackedFunction(F); 1647 } 1648 1649 bool SCCPSolver::isArgumentTrackedFunction(Function *F) { 1650 return Visitor->isArgumentTrackedFunction(F); 1651 } 1652 1653 void SCCPSolver::solve() { Visitor->solve(); } 1654 1655 bool SCCPSolver::resolvedUndefsIn(Function &F) { 1656 return Visitor->resolvedUndefsIn(F); 1657 } 1658 1659 bool SCCPSolver::isBlockExecutable(BasicBlock *BB) const { 1660 return Visitor->isBlockExecutable(BB); 1661 } 1662 1663 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 1664 return Visitor->isEdgeFeasible(From, To); 1665 } 1666 1667 std::vector<ValueLatticeElement> 1668 SCCPSolver::getStructLatticeValueFor(Value *V) const { 1669 return Visitor->getStructLatticeValueFor(V); 1670 } 1671 1672 void SCCPSolver::removeLatticeValueFor(Value *V) { 1673 return Visitor->removeLatticeValueFor(V); 1674 } 1675 1676 const ValueLatticeElement &SCCPSolver::getLatticeValueFor(Value *V) const { 1677 return Visitor->getLatticeValueFor(V); 1678 } 1679 1680 const MapVector<Function *, ValueLatticeElement> & 1681 SCCPSolver::getTrackedRetVals() { 1682 return Visitor->getTrackedRetVals(); 1683 } 1684 1685 const DenseMap<GlobalVariable *, ValueLatticeElement> & 1686 SCCPSolver::getTrackedGlobals() { 1687 return Visitor->getTrackedGlobals(); 1688 } 1689 1690 const SmallPtrSet<Function *, 16> SCCPSolver::getMRVFunctionsTracked() { 1691 return Visitor->getMRVFunctionsTracked(); 1692 } 1693 1694 void SCCPSolver::markOverdefined(Value *V) { Visitor->markOverdefined(V); } 1695 1696 bool SCCPSolver::isStructLatticeConstant(Function *F, StructType *STy) { 1697 return Visitor->isStructLatticeConstant(F, STy); 1698 } 1699 1700 Constant *SCCPSolver::getConstant(const ValueLatticeElement &LV) const { 1701 return Visitor->getConstant(LV); 1702 } 1703 1704 SmallPtrSetImpl<Function *> &SCCPSolver::getArgumentTrackedFunctions() { 1705 return Visitor->getArgumentTrackedFunctions(); 1706 } 1707 1708 void SCCPSolver::markArgInFuncSpecialization(Function *F, Argument *A, 1709 Constant *C) { 1710 Visitor->markArgInFuncSpecialization(F, A, C); 1711 } 1712 1713 void SCCPSolver::markFunctionUnreachable(Function *F) { 1714 Visitor->markFunctionUnreachable(F); 1715 } 1716 1717 void SCCPSolver::visit(Instruction *I) { Visitor->visit(I); } 1718 1719 void SCCPSolver::visitCall(CallInst &I) { Visitor->visitCall(I); } 1720