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/ValueLattice.h" 19 #include "llvm/IR/InstVisitor.h" 20 #include "llvm/Support/Casting.h" 21 #include "llvm/Support/Debug.h" 22 #include "llvm/Support/ErrorHandling.h" 23 #include "llvm/Support/raw_ostream.h" 24 #include <cassert> 25 #include <utility> 26 #include <vector> 27 28 using namespace llvm; 29 30 #define DEBUG_TYPE "sccp" 31 32 // The maximum number of range extensions allowed for operations requiring 33 // widening. 34 static const unsigned MaxNumRangeExtensions = 10; 35 36 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions. 37 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() { 38 return ValueLatticeElement::MergeOptions().setMaxWidenSteps( 39 MaxNumRangeExtensions); 40 } 41 42 namespace { 43 44 // Helper to check if \p LV is either a constant or a constant 45 // range with a single element. This should cover exactly the same cases as the 46 // old ValueLatticeElement::isConstant() and is intended to be used in the 47 // transition to ValueLatticeElement. 48 bool isConstant(const ValueLatticeElement &LV) { 49 return LV.isConstant() || 50 (LV.isConstantRange() && LV.getConstantRange().isSingleElement()); 51 } 52 53 // Helper to check if \p LV is either overdefined or a constant range with more 54 // than a single element. This should cover exactly the same cases as the old 55 // ValueLatticeElement::isOverdefined() and is intended to be used in the 56 // transition to ValueLatticeElement. 57 bool isOverdefined(const ValueLatticeElement &LV) { 58 return !LV.isUnknownOrUndef() && !isConstant(LV); 59 } 60 61 } // namespace 62 63 namespace llvm { 64 65 /// Helper class for SCCPSolver. This implements the instruction visitor and 66 /// holds all the state. 67 class SCCPInstVisitor : public InstVisitor<SCCPInstVisitor> { 68 const DataLayout &DL; 69 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 70 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable. 71 DenseMap<Value *, ValueLatticeElement> 72 ValueState; // The state each value is in. 73 74 /// StructValueState - This maintains ValueState for values that have 75 /// StructType, for example for formal arguments, calls, insertelement, etc. 76 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState; 77 78 /// GlobalValue - If we are tracking any values for the contents of a global 79 /// variable, we keep a mapping from the constant accessor to the element of 80 /// the global, to the currently known value. If the value becomes 81 /// overdefined, it's entry is simply removed from this map. 82 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals; 83 84 /// TrackedRetVals - If we are tracking arguments into and the return 85 /// value out of a function, it will have an entry in this map, indicating 86 /// what the known return value for the function is. 87 MapVector<Function *, ValueLatticeElement> TrackedRetVals; 88 89 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions 90 /// that return multiple values. 91 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement> 92 TrackedMultipleRetVals; 93 94 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is 95 /// represented here for efficient lookup. 96 SmallPtrSet<Function *, 16> MRVFunctionsTracked; 97 98 /// A list of functions whose return cannot be modified. 99 SmallPtrSet<Function *, 16> MustPreserveReturnsInFunctions; 100 101 /// TrackingIncomingArguments - This is the set of functions for whose 102 /// arguments we make optimistic assumptions about and try to prove as 103 /// constants. 104 SmallPtrSet<Function *, 16> TrackingIncomingArguments; 105 106 /// The reason for two worklists is that overdefined is the lowest state 107 /// on the lattice, and moving things to overdefined as fast as possible 108 /// makes SCCP converge much faster. 109 /// 110 /// By having a separate worklist, we accomplish this because everything 111 /// possibly overdefined will become overdefined at the soonest possible 112 /// point. 113 SmallVector<Value *, 64> OverdefinedInstWorkList; 114 SmallVector<Value *, 64> InstWorkList; 115 116 // The BasicBlock work list 117 SmallVector<BasicBlock *, 64> BBWorkList; 118 119 /// KnownFeasibleEdges - Entries in this set are edges which have already had 120 /// PHI nodes retriggered. 121 using Edge = std::pair<BasicBlock *, BasicBlock *>; 122 DenseSet<Edge> KnownFeasibleEdges; 123 124 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults; 125 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers; 126 127 LLVMContext &Ctx; 128 129 private: 130 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const { 131 return dyn_cast_or_null<ConstantInt>(getConstant(IV)); 132 } 133 134 // pushToWorkList - Helper for markConstant/markOverdefined 135 void pushToWorkList(ValueLatticeElement &IV, Value *V); 136 137 // Helper to push \p V to the worklist, after updating it to \p IV. Also 138 // prints a debug message with the updated value. 139 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V); 140 141 // markConstant - Make a value be marked as "constant". If the value 142 // is not already a constant, add it to the instruction work list so that 143 // the users of the instruction are updated later. 144 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C, 145 bool MayIncludeUndef = false); 146 147 bool markConstant(Value *V, Constant *C) { 148 assert(!V->getType()->isStructTy() && "structs should use mergeInValue"); 149 return markConstant(ValueState[V], V, C); 150 } 151 152 // markOverdefined - Make a value be marked as "overdefined". If the 153 // value is not already overdefined, add it to the overdefined instruction 154 // work list so that the users of the instruction are updated later. 155 bool markOverdefined(ValueLatticeElement &IV, Value *V); 156 157 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV 158 /// changes. 159 bool mergeInValue(ValueLatticeElement &IV, Value *V, 160 ValueLatticeElement MergeWithV, 161 ValueLatticeElement::MergeOptions Opts = { 162 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}); 163 164 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV, 165 ValueLatticeElement::MergeOptions Opts = { 166 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) { 167 assert(!V->getType()->isStructTy() && 168 "non-structs should use markConstant"); 169 return mergeInValue(ValueState[V], V, MergeWithV, Opts); 170 } 171 172 /// getValueState - Return the ValueLatticeElement object that corresponds to 173 /// the value. This function handles the case when the value hasn't been seen 174 /// yet by properly seeding constants etc. 175 ValueLatticeElement &getValueState(Value *V) { 176 assert(!V->getType()->isStructTy() && "Should use getStructValueState"); 177 178 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement())); 179 ValueLatticeElement &LV = I.first->second; 180 181 if (!I.second) 182 return LV; // Common case, already in the map. 183 184 if (auto *C = dyn_cast<Constant>(V)) 185 LV.markConstant(C); // Constants are constant 186 187 // All others are unknown by default. 188 return LV; 189 } 190 191 /// getStructValueState - Return the ValueLatticeElement object that 192 /// corresponds to the value/field pair. This function handles the case when 193 /// the value hasn't been seen yet by properly seeding constants etc. 194 ValueLatticeElement &getStructValueState(Value *V, unsigned i) { 195 assert(V->getType()->isStructTy() && "Should use getValueState"); 196 assert(i < cast<StructType>(V->getType())->getNumElements() && 197 "Invalid element #"); 198 199 auto I = StructValueState.insert( 200 std::make_pair(std::make_pair(V, i), ValueLatticeElement())); 201 ValueLatticeElement &LV = I.first->second; 202 203 if (!I.second) 204 return LV; // Common case, already in the map. 205 206 if (auto *C = dyn_cast<Constant>(V)) { 207 Constant *Elt = C->getAggregateElement(i); 208 209 if (!Elt) 210 LV.markOverdefined(); // Unknown sort of constant. 211 else if (isa<UndefValue>(Elt)) 212 ; // Undef values remain unknown. 213 else 214 LV.markConstant(Elt); // Constants are constant. 215 } 216 217 // All others are underdefined by default. 218 return LV; 219 } 220 221 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB 222 /// work list if it is not already executable. 223 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest); 224 225 // getFeasibleSuccessors - Return a vector of booleans to indicate which 226 // successors are reachable from a given terminator instruction. 227 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs); 228 229 // OperandChangedState - This method is invoked on all of the users of an 230 // instruction that was just changed state somehow. Based on this 231 // information, we need to update the specified user of this instruction. 232 void operandChangedState(Instruction *I) { 233 if (BBExecutable.count(I->getParent())) // Inst is executable? 234 visit(*I); 235 } 236 237 // Add U as additional user of V. 238 void addAdditionalUser(Value *V, User *U) { 239 auto Iter = AdditionalUsers.insert({V, {}}); 240 Iter.first->second.insert(U); 241 } 242 243 // Mark I's users as changed, including AdditionalUsers. 244 void markUsersAsChanged(Value *I) { 245 // Functions include their arguments in the use-list. Changed function 246 // values mean that the result of the function changed. We only need to 247 // update the call sites with the new function result and do not have to 248 // propagate the call arguments. 249 if (isa<Function>(I)) { 250 for (User *U : I->users()) { 251 if (auto *CB = dyn_cast<CallBase>(U)) 252 handleCallResult(*CB); 253 } 254 } else { 255 for (User *U : I->users()) 256 if (auto *UI = dyn_cast<Instruction>(U)) 257 operandChangedState(UI); 258 } 259 260 auto Iter = AdditionalUsers.find(I); 261 if (Iter != AdditionalUsers.end()) { 262 // Copy additional users before notifying them of changes, because new 263 // users may be added, potentially invalidating the iterator. 264 SmallVector<Instruction *, 2> ToNotify; 265 for (User *U : Iter->second) 266 if (auto *UI = dyn_cast<Instruction>(U)) 267 ToNotify.push_back(UI); 268 for (Instruction *UI : ToNotify) 269 operandChangedState(UI); 270 } 271 } 272 void handleCallOverdefined(CallBase &CB); 273 void handleCallResult(CallBase &CB); 274 void handleCallArguments(CallBase &CB); 275 276 private: 277 friend class InstVisitor<SCCPInstVisitor>; 278 279 // visit implementations - Something changed in this instruction. Either an 280 // operand made a transition, or the instruction is newly executable. Change 281 // the value type of I to reflect these changes if appropriate. 282 void visitPHINode(PHINode &I); 283 284 // Terminators 285 286 void visitReturnInst(ReturnInst &I); 287 void visitTerminator(Instruction &TI); 288 289 void visitCastInst(CastInst &I); 290 void visitSelectInst(SelectInst &I); 291 void visitUnaryOperator(Instruction &I); 292 void visitBinaryOperator(Instruction &I); 293 void visitCmpInst(CmpInst &I); 294 void visitExtractValueInst(ExtractValueInst &EVI); 295 void visitInsertValueInst(InsertValueInst &IVI); 296 297 void visitCatchSwitchInst(CatchSwitchInst &CPI) { 298 markOverdefined(&CPI); 299 visitTerminator(CPI); 300 } 301 302 // Instructions that cannot be folded away. 303 304 void visitStoreInst(StoreInst &I); 305 void visitLoadInst(LoadInst &I); 306 void visitGetElementPtrInst(GetElementPtrInst &I); 307 308 void visitInvokeInst(InvokeInst &II) { 309 visitCallBase(II); 310 visitTerminator(II); 311 } 312 313 void visitCallBrInst(CallBrInst &CBI) { 314 visitCallBase(CBI); 315 visitTerminator(CBI); 316 } 317 318 void visitCallBase(CallBase &CB); 319 void visitResumeInst(ResumeInst &I) { /*returns void*/ 320 } 321 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ 322 } 323 void visitFenceInst(FenceInst &I) { /*returns void*/ 324 } 325 326 void visitInstruction(Instruction &I); 327 328 public: 329 void addAnalysis(Function &F, AnalysisResultsForFn A) { 330 AnalysisResults.insert({&F, std::move(A)}); 331 } 332 333 void visitCallInst(CallInst &I) { visitCallBase(I); } 334 335 bool markBlockExecutable(BasicBlock *BB); 336 337 const PredicateBase *getPredicateInfoFor(Instruction *I) { 338 auto A = AnalysisResults.find(I->getParent()->getParent()); 339 if (A == AnalysisResults.end()) 340 return nullptr; 341 return A->second.PredInfo->getPredicateInfoFor(I); 342 } 343 344 DomTreeUpdater getDTU(Function &F) { 345 auto A = AnalysisResults.find(&F); 346 assert(A != AnalysisResults.end() && "Need analysis results for function."); 347 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy}; 348 } 349 350 SCCPInstVisitor(const DataLayout &DL, 351 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 352 LLVMContext &Ctx) 353 : DL(DL), GetTLI(GetTLI), Ctx(Ctx) {} 354 355 void trackValueOfGlobalVariable(GlobalVariable *GV) { 356 // We only track the contents of scalar globals. 357 if (GV->getValueType()->isSingleValueType()) { 358 ValueLatticeElement &IV = TrackedGlobals[GV]; 359 if (!isa<UndefValue>(GV->getInitializer())) 360 IV.markConstant(GV->getInitializer()); 361 } 362 } 363 364 void addTrackedFunction(Function *F) { 365 // Add an entry, F -> undef. 366 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 367 MRVFunctionsTracked.insert(F); 368 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 369 TrackedMultipleRetVals.insert( 370 std::make_pair(std::make_pair(F, i), ValueLatticeElement())); 371 } else if (!F->getReturnType()->isVoidTy()) 372 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement())); 373 } 374 375 void addToMustPreserveReturnsInFunctions(Function *F) { 376 MustPreserveReturnsInFunctions.insert(F); 377 } 378 379 bool mustPreserveReturn(Function *F) { 380 return MustPreserveReturnsInFunctions.count(F); 381 } 382 383 void addArgumentTrackedFunction(Function *F) { 384 TrackingIncomingArguments.insert(F); 385 } 386 387 bool isArgumentTrackedFunction(Function *F) { 388 return TrackingIncomingArguments.count(F); 389 } 390 391 void solve(); 392 393 bool resolvedUndefsIn(Function &F); 394 395 bool isBlockExecutable(BasicBlock *BB) const { 396 return BBExecutable.count(BB); 397 } 398 399 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const; 400 401 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const { 402 std::vector<ValueLatticeElement> StructValues; 403 auto *STy = dyn_cast<StructType>(V->getType()); 404 assert(STy && "getStructLatticeValueFor() can be called only on structs"); 405 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 406 auto I = StructValueState.find(std::make_pair(V, i)); 407 assert(I != StructValueState.end() && "Value not in valuemap!"); 408 StructValues.push_back(I->second); 409 } 410 return StructValues; 411 } 412 413 void removeLatticeValueFor(Value *V) { ValueState.erase(V); } 414 415 const ValueLatticeElement &getLatticeValueFor(Value *V) const { 416 assert(!V->getType()->isStructTy() && 417 "Should use getStructLatticeValueFor"); 418 DenseMap<Value *, ValueLatticeElement>::const_iterator I = 419 ValueState.find(V); 420 assert(I != ValueState.end() && 421 "V not found in ValueState nor Paramstate map!"); 422 return I->second; 423 } 424 425 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() { 426 return TrackedRetVals; 427 } 428 429 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() { 430 return TrackedGlobals; 431 } 432 433 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() { 434 return MRVFunctionsTracked; 435 } 436 437 void markOverdefined(Value *V) { 438 if (auto *STy = dyn_cast<StructType>(V->getType())) 439 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 440 markOverdefined(getStructValueState(V, i), V); 441 else 442 markOverdefined(ValueState[V], V); 443 } 444 445 bool isStructLatticeConstant(Function *F, StructType *STy); 446 447 Constant *getConstant(const ValueLatticeElement &LV) const; 448 449 SmallPtrSetImpl<Function *> &getArgumentTrackedFunctions() { 450 return TrackingIncomingArguments; 451 } 452 453 void markArgInFuncSpecialization(Function *F, const ArgInfo &Arg); 454 455 void markFunctionUnreachable(Function *F) { 456 for (auto &BB : *F) 457 BBExecutable.erase(&BB); 458 } 459 }; 460 461 } // namespace llvm 462 463 bool SCCPInstVisitor::markBlockExecutable(BasicBlock *BB) { 464 if (!BBExecutable.insert(BB).second) 465 return false; 466 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n'); 467 BBWorkList.push_back(BB); // Add the block to the work list! 468 return true; 469 } 470 471 void SCCPInstVisitor::pushToWorkList(ValueLatticeElement &IV, Value *V) { 472 if (IV.isOverdefined()) 473 return OverdefinedInstWorkList.push_back(V); 474 InstWorkList.push_back(V); 475 } 476 477 void SCCPInstVisitor::pushToWorkListMsg(ValueLatticeElement &IV, Value *V) { 478 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n'); 479 pushToWorkList(IV, V); 480 } 481 482 bool SCCPInstVisitor::markConstant(ValueLatticeElement &IV, Value *V, 483 Constant *C, bool MayIncludeUndef) { 484 if (!IV.markConstant(C, MayIncludeUndef)) 485 return false; 486 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n'); 487 pushToWorkList(IV, V); 488 return true; 489 } 490 491 bool SCCPInstVisitor::markOverdefined(ValueLatticeElement &IV, Value *V) { 492 if (!IV.markOverdefined()) 493 return false; 494 495 LLVM_DEBUG(dbgs() << "markOverdefined: "; 496 if (auto *F = dyn_cast<Function>(V)) dbgs() 497 << "Function '" << F->getName() << "'\n"; 498 else dbgs() << *V << '\n'); 499 // Only instructions go on the work list 500 pushToWorkList(IV, V); 501 return true; 502 } 503 504 bool SCCPInstVisitor::isStructLatticeConstant(Function *F, StructType *STy) { 505 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 506 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i)); 507 assert(It != TrackedMultipleRetVals.end()); 508 ValueLatticeElement LV = It->second; 509 if (!isConstant(LV)) 510 return false; 511 } 512 return true; 513 } 514 515 Constant *SCCPInstVisitor::getConstant(const ValueLatticeElement &LV) const { 516 if (LV.isConstant()) 517 return LV.getConstant(); 518 519 if (LV.isConstantRange()) { 520 const auto &CR = LV.getConstantRange(); 521 if (CR.getSingleElement()) 522 return ConstantInt::get(Ctx, *CR.getSingleElement()); 523 } 524 return nullptr; 525 } 526 527 void SCCPInstVisitor::markArgInFuncSpecialization(Function *F, 528 const ArgInfo &Arg) { 529 assert(F->arg_size() == Arg.Formal->getParent()->arg_size() && 530 "Functions should have the same number of arguments"); 531 532 Argument *NewArg = F->arg_begin(); 533 Argument *OldArg = Arg.Formal->getParent()->arg_begin(); 534 for (auto End = F->arg_end(); NewArg != End; ++NewArg, ++OldArg) { 535 536 LLVM_DEBUG(dbgs() << "SCCP: Marking argument " 537 << NewArg->getNameOrAsOperand() << "\n"); 538 539 if (OldArg == Arg.Formal) { 540 // Mark the argument constants in the new function. 541 markConstant(NewArg, Arg.Actual); 542 } else if (ValueState.count(OldArg)) { 543 // For the remaining arguments in the new function, copy the lattice state 544 // over from the old function. 545 // 546 // Note: This previously looked like this: 547 // ValueState[NewArg] = ValueState[OldArg]; 548 // This is incorrect because the DenseMap class may resize the underlying 549 // memory when inserting `NewArg`, which will invalidate the reference to 550 // `OldArg`. Instead, we make sure `NewArg` exists before setting it. 551 auto &NewValue = ValueState[NewArg]; 552 NewValue = ValueState[OldArg]; 553 pushToWorkList(NewValue, NewArg); 554 } 555 } 556 } 557 558 void SCCPInstVisitor::visitInstruction(Instruction &I) { 559 // All the instructions we don't do any special handling for just 560 // go to overdefined. 561 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n'); 562 markOverdefined(&I); 563 } 564 565 bool SCCPInstVisitor::mergeInValue(ValueLatticeElement &IV, Value *V, 566 ValueLatticeElement MergeWithV, 567 ValueLatticeElement::MergeOptions Opts) { 568 if (IV.mergeIn(MergeWithV, Opts)) { 569 pushToWorkList(IV, V); 570 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : " 571 << IV << "\n"); 572 return true; 573 } 574 return false; 575 } 576 577 bool SCCPInstVisitor::markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) { 578 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second) 579 return false; // This edge is already known to be executable! 580 581 if (!markBlockExecutable(Dest)) { 582 // If the destination is already executable, we just made an *edge* 583 // feasible that wasn't before. Revisit the PHI nodes in the block 584 // because they have potentially new operands. 585 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName() 586 << " -> " << Dest->getName() << '\n'); 587 588 for (PHINode &PN : Dest->phis()) 589 visitPHINode(PN); 590 } 591 return true; 592 } 593 594 // getFeasibleSuccessors - Return a vector of booleans to indicate which 595 // successors are reachable from a given terminator instruction. 596 void SCCPInstVisitor::getFeasibleSuccessors(Instruction &TI, 597 SmallVectorImpl<bool> &Succs) { 598 Succs.resize(TI.getNumSuccessors()); 599 if (auto *BI = dyn_cast<BranchInst>(&TI)) { 600 if (BI->isUnconditional()) { 601 Succs[0] = true; 602 return; 603 } 604 605 ValueLatticeElement BCValue = getValueState(BI->getCondition()); 606 ConstantInt *CI = getConstantInt(BCValue); 607 if (!CI) { 608 // Overdefined condition variables, and branches on unfoldable constant 609 // conditions, mean the branch could go either way. 610 if (!BCValue.isUnknownOrUndef()) 611 Succs[0] = Succs[1] = true; 612 return; 613 } 614 615 // Constant condition variables mean the branch can only go a single way. 616 Succs[CI->isZero()] = true; 617 return; 618 } 619 620 // Unwinding instructions successors are always executable. 621 if (TI.isExceptionalTerminator()) { 622 Succs.assign(TI.getNumSuccessors(), true); 623 return; 624 } 625 626 if (auto *SI = dyn_cast<SwitchInst>(&TI)) { 627 if (!SI->getNumCases()) { 628 Succs[0] = true; 629 return; 630 } 631 const ValueLatticeElement &SCValue = getValueState(SI->getCondition()); 632 if (ConstantInt *CI = getConstantInt(SCValue)) { 633 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true; 634 return; 635 } 636 637 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM 638 // is ready. 639 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) { 640 const ConstantRange &Range = SCValue.getConstantRange(); 641 for (const auto &Case : SI->cases()) { 642 const APInt &CaseValue = Case.getCaseValue()->getValue(); 643 if (Range.contains(CaseValue)) 644 Succs[Case.getSuccessorIndex()] = true; 645 } 646 647 // TODO: Determine whether default case is reachable. 648 Succs[SI->case_default()->getSuccessorIndex()] = true; 649 return; 650 } 651 652 // Overdefined or unknown condition? All destinations are executable! 653 if (!SCValue.isUnknownOrUndef()) 654 Succs.assign(TI.getNumSuccessors(), true); 655 return; 656 } 657 658 // In case of indirect branch and its address is a blockaddress, we mark 659 // the target as executable. 660 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) { 661 // Casts are folded by visitCastInst. 662 ValueLatticeElement IBRValue = getValueState(IBR->getAddress()); 663 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue)); 664 if (!Addr) { // Overdefined or unknown condition? 665 // All destinations are executable! 666 if (!IBRValue.isUnknownOrUndef()) 667 Succs.assign(TI.getNumSuccessors(), true); 668 return; 669 } 670 671 BasicBlock *T = Addr->getBasicBlock(); 672 assert(Addr->getFunction() == T->getParent() && 673 "Block address of a different function ?"); 674 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) { 675 // This is the target. 676 if (IBR->getDestination(i) == T) { 677 Succs[i] = true; 678 return; 679 } 680 } 681 682 // If we didn't find our destination in the IBR successor list, then we 683 // have undefined behavior. Its ok to assume no successor is executable. 684 return; 685 } 686 687 // In case of callbr, we pessimistically assume that all successors are 688 // feasible. 689 if (isa<CallBrInst>(&TI)) { 690 Succs.assign(TI.getNumSuccessors(), true); 691 return; 692 } 693 694 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n'); 695 llvm_unreachable("SCCP: Don't know how to handle this terminator!"); 696 } 697 698 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic 699 // block to the 'To' basic block is currently feasible. 700 bool SCCPInstVisitor::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 701 // Check if we've called markEdgeExecutable on the edge yet. (We could 702 // be more aggressive and try to consider edges which haven't been marked 703 // yet, but there isn't any need.) 704 return KnownFeasibleEdges.count(Edge(From, To)); 705 } 706 707 // visit Implementations - Something changed in this instruction, either an 708 // operand made a transition, or the instruction is newly executable. Change 709 // the value type of I to reflect these changes if appropriate. This method 710 // makes sure to do the following actions: 711 // 712 // 1. If a phi node merges two constants in, and has conflicting value coming 713 // from different branches, or if the PHI node merges in an overdefined 714 // value, then the PHI node becomes overdefined. 715 // 2. If a phi node merges only constants in, and they all agree on value, the 716 // PHI node becomes a constant value equal to that. 717 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant 718 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined 719 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined 720 // 6. If a conditional branch has a value that is constant, make the selected 721 // destination executable 722 // 7. If a conditional branch has a value that is overdefined, make all 723 // successors executable. 724 void SCCPInstVisitor::visitPHINode(PHINode &PN) { 725 // If this PN returns a struct, just mark the result overdefined. 726 // TODO: We could do a lot better than this if code actually uses this. 727 if (PN.getType()->isStructTy()) 728 return (void)markOverdefined(&PN); 729 730 if (getValueState(&PN).isOverdefined()) 731 return; // Quick exit 732 733 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant, 734 // and slow us down a lot. Just mark them overdefined. 735 if (PN.getNumIncomingValues() > 64) 736 return (void)markOverdefined(&PN); 737 738 unsigned NumActiveIncoming = 0; 739 740 // Look at all of the executable operands of the PHI node. If any of them 741 // are overdefined, the PHI becomes overdefined as well. If they are all 742 // constant, and they agree with each other, the PHI becomes the identical 743 // constant. If they are constant and don't agree, the PHI is a constant 744 // range. If there are no executable operands, the PHI remains unknown. 745 ValueLatticeElement PhiState = getValueState(&PN); 746 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 747 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) 748 continue; 749 750 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i)); 751 PhiState.mergeIn(IV); 752 NumActiveIncoming++; 753 if (PhiState.isOverdefined()) 754 break; 755 } 756 757 // We allow up to 1 range extension per active incoming value and one 758 // additional extension. Note that we manually adjust the number of range 759 // extensions to match the number of active incoming values. This helps to 760 // limit multiple extensions caused by the same incoming value, if other 761 // incoming values are equal. 762 mergeInValue(&PN, PhiState, 763 ValueLatticeElement::MergeOptions().setMaxWidenSteps( 764 NumActiveIncoming + 1)); 765 ValueLatticeElement &PhiStateRef = getValueState(&PN); 766 PhiStateRef.setNumRangeExtensions( 767 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions())); 768 } 769 770 void SCCPInstVisitor::visitReturnInst(ReturnInst &I) { 771 if (I.getNumOperands() == 0) 772 return; // ret void 773 774 Function *F = I.getParent()->getParent(); 775 Value *ResultOp = I.getOperand(0); 776 777 // If we are tracking the return value of this function, merge it in. 778 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) { 779 auto TFRVI = TrackedRetVals.find(F); 780 if (TFRVI != TrackedRetVals.end()) { 781 mergeInValue(TFRVI->second, F, getValueState(ResultOp)); 782 return; 783 } 784 } 785 786 // Handle functions that return multiple values. 787 if (!TrackedMultipleRetVals.empty()) { 788 if (auto *STy = dyn_cast<StructType>(ResultOp->getType())) 789 if (MRVFunctionsTracked.count(F)) 790 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 791 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F, 792 getStructValueState(ResultOp, i)); 793 } 794 } 795 796 void SCCPInstVisitor::visitTerminator(Instruction &TI) { 797 SmallVector<bool, 16> SuccFeasible; 798 getFeasibleSuccessors(TI, SuccFeasible); 799 800 BasicBlock *BB = TI.getParent(); 801 802 // Mark all feasible successors executable. 803 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i) 804 if (SuccFeasible[i]) 805 markEdgeExecutable(BB, TI.getSuccessor(i)); 806 } 807 808 void SCCPInstVisitor::visitCastInst(CastInst &I) { 809 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would 810 // discover a concrete value later. 811 if (ValueState[&I].isOverdefined()) 812 return; 813 814 ValueLatticeElement OpSt = getValueState(I.getOperand(0)); 815 if (OpSt.isUnknownOrUndef()) 816 return; 817 818 if (Constant *OpC = getConstant(OpSt)) { 819 // Fold the constant as we build. 820 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL); 821 if (isa<UndefValue>(C)) 822 return; 823 // Propagate constant value 824 markConstant(&I, C); 825 } else if (I.getDestTy()->isIntegerTy()) { 826 auto &LV = getValueState(&I); 827 ConstantRange OpRange = 828 OpSt.isConstantRange() 829 ? OpSt.getConstantRange() 830 : ConstantRange::getFull( 831 I.getOperand(0)->getType()->getScalarSizeInBits()); 832 833 Type *DestTy = I.getDestTy(); 834 // Vectors where all elements have the same known constant range are treated 835 // as a single constant range in the lattice. When bitcasting such vectors, 836 // there is a mis-match between the width of the lattice value (single 837 // constant range) and the original operands (vector). Go to overdefined in 838 // that case. 839 if (I.getOpcode() == Instruction::BitCast && 840 I.getOperand(0)->getType()->isVectorTy() && 841 OpRange.getBitWidth() < DL.getTypeSizeInBits(DestTy)) 842 return (void)markOverdefined(&I); 843 844 ConstantRange Res = 845 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy)); 846 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res)); 847 } else 848 markOverdefined(&I); 849 } 850 851 void SCCPInstVisitor::visitExtractValueInst(ExtractValueInst &EVI) { 852 // If this returns a struct, mark all elements over defined, we don't track 853 // structs in structs. 854 if (EVI.getType()->isStructTy()) 855 return (void)markOverdefined(&EVI); 856 857 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 858 // discover a concrete value later. 859 if (ValueState[&EVI].isOverdefined()) 860 return (void)markOverdefined(&EVI); 861 862 // If this is extracting from more than one level of struct, we don't know. 863 if (EVI.getNumIndices() != 1) 864 return (void)markOverdefined(&EVI); 865 866 Value *AggVal = EVI.getAggregateOperand(); 867 if (AggVal->getType()->isStructTy()) { 868 unsigned i = *EVI.idx_begin(); 869 ValueLatticeElement EltVal = getStructValueState(AggVal, i); 870 mergeInValue(getValueState(&EVI), &EVI, EltVal); 871 } else { 872 // Otherwise, must be extracting from an array. 873 return (void)markOverdefined(&EVI); 874 } 875 } 876 877 void SCCPInstVisitor::visitInsertValueInst(InsertValueInst &IVI) { 878 auto *STy = dyn_cast<StructType>(IVI.getType()); 879 if (!STy) 880 return (void)markOverdefined(&IVI); 881 882 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 883 // discover a concrete value later. 884 if (isOverdefined(ValueState[&IVI])) 885 return (void)markOverdefined(&IVI); 886 887 // If this has more than one index, we can't handle it, drive all results to 888 // undef. 889 if (IVI.getNumIndices() != 1) 890 return (void)markOverdefined(&IVI); 891 892 Value *Aggr = IVI.getAggregateOperand(); 893 unsigned Idx = *IVI.idx_begin(); 894 895 // Compute the result based on what we're inserting. 896 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 897 // This passes through all values that aren't the inserted element. 898 if (i != Idx) { 899 ValueLatticeElement EltVal = getStructValueState(Aggr, i); 900 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal); 901 continue; 902 } 903 904 Value *Val = IVI.getInsertedValueOperand(); 905 if (Val->getType()->isStructTy()) 906 // We don't track structs in structs. 907 markOverdefined(getStructValueState(&IVI, i), &IVI); 908 else { 909 ValueLatticeElement InVal = getValueState(Val); 910 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal); 911 } 912 } 913 } 914 915 void SCCPInstVisitor::visitSelectInst(SelectInst &I) { 916 // If this select returns a struct, just mark the result overdefined. 917 // TODO: We could do a lot better than this if code actually uses this. 918 if (I.getType()->isStructTy()) 919 return (void)markOverdefined(&I); 920 921 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 922 // discover a concrete value later. 923 if (ValueState[&I].isOverdefined()) 924 return (void)markOverdefined(&I); 925 926 ValueLatticeElement CondValue = getValueState(I.getCondition()); 927 if (CondValue.isUnknownOrUndef()) 928 return; 929 930 if (ConstantInt *CondCB = getConstantInt(CondValue)) { 931 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue(); 932 mergeInValue(&I, getValueState(OpVal)); 933 return; 934 } 935 936 // Otherwise, the condition is overdefined or a constant we can't evaluate. 937 // See if we can produce something better than overdefined based on the T/F 938 // value. 939 ValueLatticeElement TVal = getValueState(I.getTrueValue()); 940 ValueLatticeElement FVal = getValueState(I.getFalseValue()); 941 942 bool Changed = ValueState[&I].mergeIn(TVal); 943 Changed |= ValueState[&I].mergeIn(FVal); 944 if (Changed) 945 pushToWorkListMsg(ValueState[&I], &I); 946 } 947 948 // Handle Unary Operators. 949 void SCCPInstVisitor::visitUnaryOperator(Instruction &I) { 950 ValueLatticeElement V0State = getValueState(I.getOperand(0)); 951 952 ValueLatticeElement &IV = ValueState[&I]; 953 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 954 // discover a concrete value later. 955 if (isOverdefined(IV)) 956 return (void)markOverdefined(&I); 957 958 if (isConstant(V0State)) { 959 Constant *C = ConstantExpr::get(I.getOpcode(), getConstant(V0State)); 960 961 // op Y -> undef. 962 if (isa<UndefValue>(C)) 963 return; 964 return (void)markConstant(IV, &I, C); 965 } 966 967 // If something is undef, wait for it to resolve. 968 if (!isOverdefined(V0State)) 969 return; 970 971 markOverdefined(&I); 972 } 973 974 // Handle Binary Operators. 975 void SCCPInstVisitor::visitBinaryOperator(Instruction &I) { 976 ValueLatticeElement V1State = getValueState(I.getOperand(0)); 977 ValueLatticeElement V2State = getValueState(I.getOperand(1)); 978 979 ValueLatticeElement &IV = ValueState[&I]; 980 if (IV.isOverdefined()) 981 return; 982 983 // If something is undef, wait for it to resolve. 984 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) 985 return; 986 987 if (V1State.isOverdefined() && V2State.isOverdefined()) 988 return (void)markOverdefined(&I); 989 990 // If either of the operands is a constant, try to fold it to a constant. 991 // TODO: Use information from notconstant better. 992 if ((V1State.isConstant() || V2State.isConstant())) { 993 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0); 994 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1); 995 Value *R = SimplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL)); 996 auto *C = dyn_cast_or_null<Constant>(R); 997 if (C) { 998 // X op Y -> undef. 999 if (isa<UndefValue>(C)) 1000 return; 1001 // Conservatively assume that the result may be based on operands that may 1002 // be undef. Note that we use mergeInValue to combine the constant with 1003 // the existing lattice value for I, as different constants might be found 1004 // after one of the operands go to overdefined, e.g. due to one operand 1005 // being a special floating value. 1006 ValueLatticeElement NewV; 1007 NewV.markConstant(C, /*MayIncludeUndef=*/true); 1008 return (void)mergeInValue(&I, NewV); 1009 } 1010 } 1011 1012 // Only use ranges for binary operators on integers. 1013 if (!I.getType()->isIntegerTy()) 1014 return markOverdefined(&I); 1015 1016 // Try to simplify to a constant range. 1017 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1018 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits()); 1019 if (V1State.isConstantRange()) 1020 A = V1State.getConstantRange(); 1021 if (V2State.isConstantRange()) 1022 B = V2State.getConstantRange(); 1023 1024 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B); 1025 mergeInValue(&I, ValueLatticeElement::getRange(R)); 1026 1027 // TODO: Currently we do not exploit special values that produce something 1028 // better than overdefined with an overdefined operand for vector or floating 1029 // point types, like and <4 x i32> overdefined, zeroinitializer. 1030 } 1031 1032 // Handle ICmpInst instruction. 1033 void SCCPInstVisitor::visitCmpInst(CmpInst &I) { 1034 // Do not cache this lookup, getValueState calls later in the function might 1035 // invalidate the reference. 1036 if (isOverdefined(ValueState[&I])) 1037 return (void)markOverdefined(&I); 1038 1039 Value *Op1 = I.getOperand(0); 1040 Value *Op2 = I.getOperand(1); 1041 1042 // For parameters, use ParamState which includes constant range info if 1043 // available. 1044 auto V1State = getValueState(Op1); 1045 auto V2State = getValueState(Op2); 1046 1047 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State); 1048 if (C) { 1049 if (isa<UndefValue>(C)) 1050 return; 1051 ValueLatticeElement CV; 1052 CV.markConstant(C); 1053 mergeInValue(&I, CV); 1054 return; 1055 } 1056 1057 // If operands are still unknown, wait for it to resolve. 1058 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) && 1059 !isConstant(ValueState[&I])) 1060 return; 1061 1062 markOverdefined(&I); 1063 } 1064 1065 // Handle getelementptr instructions. If all operands are constants then we 1066 // can turn this into a getelementptr ConstantExpr. 1067 void SCCPInstVisitor::visitGetElementPtrInst(GetElementPtrInst &I) { 1068 if (isOverdefined(ValueState[&I])) 1069 return (void)markOverdefined(&I); 1070 1071 SmallVector<Constant *, 8> Operands; 1072 Operands.reserve(I.getNumOperands()); 1073 1074 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 1075 ValueLatticeElement State = getValueState(I.getOperand(i)); 1076 if (State.isUnknownOrUndef()) 1077 return; // Operands are not resolved yet. 1078 1079 if (isOverdefined(State)) 1080 return (void)markOverdefined(&I); 1081 1082 if (Constant *C = getConstant(State)) { 1083 Operands.push_back(C); 1084 continue; 1085 } 1086 1087 return (void)markOverdefined(&I); 1088 } 1089 1090 Constant *Ptr = Operands[0]; 1091 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end()); 1092 Constant *C = 1093 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices); 1094 if (isa<UndefValue>(C)) 1095 return; 1096 markConstant(&I, C); 1097 } 1098 1099 void SCCPInstVisitor::visitStoreInst(StoreInst &SI) { 1100 // If this store is of a struct, ignore it. 1101 if (SI.getOperand(0)->getType()->isStructTy()) 1102 return; 1103 1104 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1))) 1105 return; 1106 1107 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1)); 1108 auto I = TrackedGlobals.find(GV); 1109 if (I == TrackedGlobals.end()) 1110 return; 1111 1112 // Get the value we are storing into the global, then merge it. 1113 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)), 1114 ValueLatticeElement::MergeOptions().setCheckWiden(false)); 1115 if (I->second.isOverdefined()) 1116 TrackedGlobals.erase(I); // No need to keep tracking this! 1117 } 1118 1119 static ValueLatticeElement getValueFromMetadata(const Instruction *I) { 1120 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) 1121 if (I->getType()->isIntegerTy()) 1122 return ValueLatticeElement::getRange( 1123 getConstantRangeFromMetadata(*Ranges)); 1124 if (I->hasMetadata(LLVMContext::MD_nonnull)) 1125 return ValueLatticeElement::getNot( 1126 ConstantPointerNull::get(cast<PointerType>(I->getType()))); 1127 return ValueLatticeElement::getOverdefined(); 1128 } 1129 1130 // Handle load instructions. If the operand is a constant pointer to a constant 1131 // global, we can replace the load with the loaded constant value! 1132 void SCCPInstVisitor::visitLoadInst(LoadInst &I) { 1133 // If this load is of a struct or the load is volatile, just mark the result 1134 // as overdefined. 1135 if (I.getType()->isStructTy() || I.isVolatile()) 1136 return (void)markOverdefined(&I); 1137 1138 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would 1139 // discover a concrete value later. 1140 if (ValueState[&I].isOverdefined()) 1141 return (void)markOverdefined(&I); 1142 1143 ValueLatticeElement PtrVal = getValueState(I.getOperand(0)); 1144 if (PtrVal.isUnknownOrUndef()) 1145 return; // The pointer is not resolved yet! 1146 1147 ValueLatticeElement &IV = ValueState[&I]; 1148 1149 if (isConstant(PtrVal)) { 1150 Constant *Ptr = getConstant(PtrVal); 1151 1152 // load null is undefined. 1153 if (isa<ConstantPointerNull>(Ptr)) { 1154 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace())) 1155 return (void)markOverdefined(IV, &I); 1156 else 1157 return; 1158 } 1159 1160 // Transform load (constant global) into the value loaded. 1161 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) { 1162 if (!TrackedGlobals.empty()) { 1163 // If we are tracking this global, merge in the known value for it. 1164 auto It = TrackedGlobals.find(GV); 1165 if (It != TrackedGlobals.end()) { 1166 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts()); 1167 return; 1168 } 1169 } 1170 } 1171 1172 // Transform load from a constant into a constant if possible. 1173 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) { 1174 if (isa<UndefValue>(C)) 1175 return; 1176 return (void)markConstant(IV, &I, C); 1177 } 1178 } 1179 1180 // Fall back to metadata. 1181 mergeInValue(&I, getValueFromMetadata(&I)); 1182 } 1183 1184 void SCCPInstVisitor::visitCallBase(CallBase &CB) { 1185 handleCallResult(CB); 1186 handleCallArguments(CB); 1187 } 1188 1189 void SCCPInstVisitor::handleCallOverdefined(CallBase &CB) { 1190 Function *F = CB.getCalledFunction(); 1191 1192 // Void return and not tracking callee, just bail. 1193 if (CB.getType()->isVoidTy()) 1194 return; 1195 1196 // Always mark struct return as overdefined. 1197 if (CB.getType()->isStructTy()) 1198 return (void)markOverdefined(&CB); 1199 1200 // Otherwise, if we have a single return value case, and if the function is 1201 // a declaration, maybe we can constant fold it. 1202 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) { 1203 SmallVector<Constant *, 8> Operands; 1204 for (const Use &A : CB.args()) { 1205 if (A.get()->getType()->isStructTy()) 1206 return markOverdefined(&CB); // Can't handle struct args. 1207 ValueLatticeElement State = getValueState(A); 1208 1209 if (State.isUnknownOrUndef()) 1210 return; // Operands are not resolved yet. 1211 if (isOverdefined(State)) 1212 return (void)markOverdefined(&CB); 1213 assert(isConstant(State) && "Unknown state!"); 1214 Operands.push_back(getConstant(State)); 1215 } 1216 1217 if (isOverdefined(getValueState(&CB))) 1218 return (void)markOverdefined(&CB); 1219 1220 // If we can constant fold this, mark the result of the call as a 1221 // constant. 1222 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F))) { 1223 // call -> undef. 1224 if (isa<UndefValue>(C)) 1225 return; 1226 return (void)markConstant(&CB, C); 1227 } 1228 } 1229 1230 // Fall back to metadata. 1231 mergeInValue(&CB, getValueFromMetadata(&CB)); 1232 } 1233 1234 void SCCPInstVisitor::handleCallArguments(CallBase &CB) { 1235 Function *F = CB.getCalledFunction(); 1236 // If this is a local function that doesn't have its address taken, mark its 1237 // entry block executable and merge in the actual arguments to the call into 1238 // the formal arguments of the function. 1239 if (!TrackingIncomingArguments.empty() && 1240 TrackingIncomingArguments.count(F)) { 1241 markBlockExecutable(&F->front()); 1242 1243 // Propagate information from this call site into the callee. 1244 auto CAI = CB.arg_begin(); 1245 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 1246 ++AI, ++CAI) { 1247 // If this argument is byval, and if the function is not readonly, there 1248 // will be an implicit copy formed of the input aggregate. 1249 if (AI->hasByValAttr() && !F->onlyReadsMemory()) { 1250 markOverdefined(&*AI); 1251 continue; 1252 } 1253 1254 if (auto *STy = dyn_cast<StructType>(AI->getType())) { 1255 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1256 ValueLatticeElement CallArg = getStructValueState(*CAI, i); 1257 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg, 1258 getMaxWidenStepsOpts()); 1259 } 1260 } else 1261 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts()); 1262 } 1263 } 1264 } 1265 1266 void SCCPInstVisitor::handleCallResult(CallBase &CB) { 1267 Function *F = CB.getCalledFunction(); 1268 1269 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) { 1270 if (II->getIntrinsicID() == Intrinsic::ssa_copy) { 1271 if (ValueState[&CB].isOverdefined()) 1272 return; 1273 1274 Value *CopyOf = CB.getOperand(0); 1275 ValueLatticeElement CopyOfVal = getValueState(CopyOf); 1276 const auto *PI = getPredicateInfoFor(&CB); 1277 assert(PI && "Missing predicate info for ssa.copy"); 1278 1279 const Optional<PredicateConstraint> &Constraint = PI->getConstraint(); 1280 if (!Constraint) { 1281 mergeInValue(ValueState[&CB], &CB, CopyOfVal); 1282 return; 1283 } 1284 1285 CmpInst::Predicate Pred = Constraint->Predicate; 1286 Value *OtherOp = Constraint->OtherOp; 1287 1288 // Wait until OtherOp is resolved. 1289 if (getValueState(OtherOp).isUnknown()) { 1290 addAdditionalUser(OtherOp, &CB); 1291 return; 1292 } 1293 1294 // TODO: Actually filp MayIncludeUndef for the created range to false, 1295 // once most places in the optimizer respect the branches on 1296 // undef/poison are UB rule. The reason why the new range cannot be 1297 // undef is as follows below: 1298 // The new range is based on a branch condition. That guarantees that 1299 // neither of the compare operands can be undef in the branch targets, 1300 // unless we have conditions that are always true/false (e.g. icmp ule 1301 // i32, %a, i32_max). For the latter overdefined/empty range will be 1302 // inferred, but the branch will get folded accordingly anyways. 1303 bool MayIncludeUndef = !isa<PredicateAssume>(PI); 1304 1305 ValueLatticeElement CondVal = getValueState(OtherOp); 1306 ValueLatticeElement &IV = ValueState[&CB]; 1307 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) { 1308 auto ImposedCR = 1309 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType())); 1310 1311 // Get the range imposed by the condition. 1312 if (CondVal.isConstantRange()) 1313 ImposedCR = ConstantRange::makeAllowedICmpRegion( 1314 Pred, CondVal.getConstantRange()); 1315 1316 // Combine range info for the original value with the new range from the 1317 // condition. 1318 auto CopyOfCR = CopyOfVal.isConstantRange() 1319 ? CopyOfVal.getConstantRange() 1320 : ConstantRange::getFull( 1321 DL.getTypeSizeInBits(CopyOf->getType())); 1322 auto NewCR = ImposedCR.intersectWith(CopyOfCR); 1323 // If the existing information is != x, do not use the information from 1324 // a chained predicate, as the != x information is more likely to be 1325 // helpful in practice. 1326 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement()) 1327 NewCR = CopyOfCR; 1328 1329 addAdditionalUser(OtherOp, &CB); 1330 mergeInValue(IV, &CB, 1331 ValueLatticeElement::getRange(NewCR, MayIncludeUndef)); 1332 return; 1333 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) { 1334 // For non-integer values or integer constant expressions, only 1335 // propagate equal constants. 1336 addAdditionalUser(OtherOp, &CB); 1337 mergeInValue(IV, &CB, CondVal); 1338 return; 1339 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant() && 1340 !MayIncludeUndef) { 1341 // Propagate inequalities. 1342 addAdditionalUser(OtherOp, &CB); 1343 mergeInValue(IV, &CB, 1344 ValueLatticeElement::getNot(CondVal.getConstant())); 1345 return; 1346 } 1347 1348 return (void)mergeInValue(IV, &CB, CopyOfVal); 1349 } 1350 1351 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { 1352 // Compute result range for intrinsics supported by ConstantRange. 1353 // Do this even if we don't know a range for all operands, as we may 1354 // still know something about the result range, e.g. of abs(x). 1355 SmallVector<ConstantRange, 2> OpRanges; 1356 for (Value *Op : II->args()) { 1357 const ValueLatticeElement &State = getValueState(Op); 1358 if (State.isConstantRange()) 1359 OpRanges.push_back(State.getConstantRange()); 1360 else 1361 OpRanges.push_back( 1362 ConstantRange::getFull(Op->getType()->getScalarSizeInBits())); 1363 } 1364 1365 ConstantRange Result = 1366 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges); 1367 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result)); 1368 } 1369 } 1370 1371 // The common case is that we aren't tracking the callee, either because we 1372 // are not doing interprocedural analysis or the callee is indirect, or is 1373 // external. Handle these cases first. 1374 if (!F || F->isDeclaration()) 1375 return handleCallOverdefined(CB); 1376 1377 // If this is a single/zero retval case, see if we're tracking the function. 1378 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) { 1379 if (!MRVFunctionsTracked.count(F)) 1380 return handleCallOverdefined(CB); // Not tracking this callee. 1381 1382 // If we are tracking this callee, propagate the result of the function 1383 // into this call site. 1384 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1385 mergeInValue(getStructValueState(&CB, i), &CB, 1386 TrackedMultipleRetVals[std::make_pair(F, i)], 1387 getMaxWidenStepsOpts()); 1388 } else { 1389 auto TFRVI = TrackedRetVals.find(F); 1390 if (TFRVI == TrackedRetVals.end()) 1391 return handleCallOverdefined(CB); // Not tracking this callee. 1392 1393 // If so, propagate the return value of the callee into this call result. 1394 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts()); 1395 } 1396 } 1397 1398 void SCCPInstVisitor::solve() { 1399 // Process the work lists until they are empty! 1400 while (!BBWorkList.empty() || !InstWorkList.empty() || 1401 !OverdefinedInstWorkList.empty()) { 1402 // Process the overdefined instruction's work list first, which drives other 1403 // things to overdefined more quickly. 1404 while (!OverdefinedInstWorkList.empty()) { 1405 Value *I = OverdefinedInstWorkList.pop_back_val(); 1406 1407 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n'); 1408 1409 // "I" got into the work list because it either made the transition from 1410 // bottom to constant, or to overdefined. 1411 // 1412 // Anything on this worklist that is overdefined need not be visited 1413 // since all of its users will have already been marked as overdefined 1414 // Update all of the users of this instruction's value. 1415 // 1416 markUsersAsChanged(I); 1417 } 1418 1419 // Process the instruction work list. 1420 while (!InstWorkList.empty()) { 1421 Value *I = InstWorkList.pop_back_val(); 1422 1423 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n'); 1424 1425 // "I" got into the work list because it made the transition from undef to 1426 // constant. 1427 // 1428 // Anything on this worklist that is overdefined need not be visited 1429 // since all of its users will have already been marked as overdefined. 1430 // Update all of the users of this instruction's value. 1431 // 1432 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined()) 1433 markUsersAsChanged(I); 1434 } 1435 1436 // Process the basic block work list. 1437 while (!BBWorkList.empty()) { 1438 BasicBlock *BB = BBWorkList.pop_back_val(); 1439 1440 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n'); 1441 1442 // Notify all instructions in this basic block that they are newly 1443 // executable. 1444 visit(BB); 1445 } 1446 } 1447 } 1448 1449 /// resolvedUndefsIn - While solving the dataflow for a function, we assume 1450 /// that branches on undef values cannot reach any of their successors. 1451 /// However, this is not a safe assumption. After we solve dataflow, this 1452 /// method should be use to handle this. If this returns true, the solver 1453 /// should be rerun. 1454 /// 1455 /// This method handles this by finding an unresolved branch and marking it one 1456 /// of the edges from the block as being feasible, even though the condition 1457 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the 1458 /// CFG and only slightly pessimizes the analysis results (by marking one, 1459 /// potentially infeasible, edge feasible). This cannot usefully modify the 1460 /// constraints on the condition of the branch, as that would impact other users 1461 /// of the value. 1462 /// 1463 /// This scan also checks for values that use undefs. It conservatively marks 1464 /// them as overdefined. 1465 bool SCCPInstVisitor::resolvedUndefsIn(Function &F) { 1466 bool MadeChange = false; 1467 for (BasicBlock &BB : F) { 1468 if (!BBExecutable.count(&BB)) 1469 continue; 1470 1471 for (Instruction &I : BB) { 1472 // Look for instructions which produce undef values. 1473 if (I.getType()->isVoidTy()) 1474 continue; 1475 1476 if (auto *STy = dyn_cast<StructType>(I.getType())) { 1477 // Only a few things that can be structs matter for undef. 1478 1479 // Tracked calls must never be marked overdefined in resolvedUndefsIn. 1480 if (auto *CB = dyn_cast<CallBase>(&I)) 1481 if (Function *F = CB->getCalledFunction()) 1482 if (MRVFunctionsTracked.count(F)) 1483 continue; 1484 1485 // extractvalue and insertvalue don't need to be marked; they are 1486 // tracked as precisely as their operands. 1487 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I)) 1488 continue; 1489 // Send the results of everything else to overdefined. We could be 1490 // more precise than this but it isn't worth bothering. 1491 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1492 ValueLatticeElement &LV = getStructValueState(&I, i); 1493 if (LV.isUnknownOrUndef()) { 1494 markOverdefined(LV, &I); 1495 MadeChange = true; 1496 } 1497 } 1498 continue; 1499 } 1500 1501 ValueLatticeElement &LV = getValueState(&I); 1502 if (!LV.isUnknownOrUndef()) 1503 continue; 1504 1505 // There are two reasons a call can have an undef result 1506 // 1. It could be tracked. 1507 // 2. It could be constant-foldable. 1508 // Because of the way we solve return values, tracked calls must 1509 // never be marked overdefined in resolvedUndefsIn. 1510 if (auto *CB = dyn_cast<CallBase>(&I)) 1511 if (Function *F = CB->getCalledFunction()) 1512 if (TrackedRetVals.count(F)) 1513 continue; 1514 1515 if (isa<LoadInst>(I)) { 1516 // A load here means one of two things: a load of undef from a global, 1517 // a load from an unknown pointer. Either way, having it return undef 1518 // is okay. 1519 continue; 1520 } 1521 1522 markOverdefined(&I); 1523 MadeChange = true; 1524 } 1525 1526 // Check to see if we have a branch or switch on an undefined value. If so 1527 // we force the branch to go one way or the other to make the successor 1528 // values live. It doesn't really matter which way we force it. 1529 Instruction *TI = BB.getTerminator(); 1530 if (auto *BI = dyn_cast<BranchInst>(TI)) { 1531 if (!BI->isConditional()) 1532 continue; 1533 if (!getValueState(BI->getCondition()).isUnknownOrUndef()) 1534 continue; 1535 1536 // If the input to SCCP is actually branch on undef, fix the undef to 1537 // false. 1538 if (isa<UndefValue>(BI->getCondition())) { 1539 BI->setCondition(ConstantInt::getFalse(BI->getContext())); 1540 markEdgeExecutable(&BB, TI->getSuccessor(1)); 1541 MadeChange = true; 1542 continue; 1543 } 1544 1545 // Otherwise, it is a branch on a symbolic value which is currently 1546 // considered to be undef. Make sure some edge is executable, so a 1547 // branch on "undef" always flows somewhere. 1548 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1549 BasicBlock *DefaultSuccessor = TI->getSuccessor(1); 1550 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1551 MadeChange = true; 1552 1553 continue; 1554 } 1555 1556 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) { 1557 // Indirect branch with no successor ?. Its ok to assume it branches 1558 // to no target. 1559 if (IBR->getNumSuccessors() < 1) 1560 continue; 1561 1562 if (!getValueState(IBR->getAddress()).isUnknownOrUndef()) 1563 continue; 1564 1565 // If the input to SCCP is actually branch on undef, fix the undef to 1566 // the first successor of the indirect branch. 1567 if (isa<UndefValue>(IBR->getAddress())) { 1568 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0))); 1569 markEdgeExecutable(&BB, IBR->getSuccessor(0)); 1570 MadeChange = true; 1571 continue; 1572 } 1573 1574 // Otherwise, it is a branch on a symbolic value which is currently 1575 // considered to be undef. Make sure some edge is executable, so a 1576 // branch on "undef" always flows somewhere. 1577 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere: 1578 // we can assume the branch has undefined behavior instead. 1579 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0); 1580 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1581 MadeChange = true; 1582 1583 continue; 1584 } 1585 1586 if (auto *SI = dyn_cast<SwitchInst>(TI)) { 1587 if (!SI->getNumCases() || 1588 !getValueState(SI->getCondition()).isUnknownOrUndef()) 1589 continue; 1590 1591 // If the input to SCCP is actually switch on undef, fix the undef to 1592 // the first constant. 1593 if (isa<UndefValue>(SI->getCondition())) { 1594 SI->setCondition(SI->case_begin()->getCaseValue()); 1595 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor()); 1596 MadeChange = true; 1597 continue; 1598 } 1599 1600 // Otherwise, it is a branch on a symbolic value which is currently 1601 // considered to be undef. Make sure some edge is executable, so a 1602 // branch on "undef" always flows somewhere. 1603 // FIXME: Distinguish between dead code and an LLVM "undef" value. 1604 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor(); 1605 if (markEdgeExecutable(&BB, DefaultSuccessor)) 1606 MadeChange = true; 1607 1608 continue; 1609 } 1610 } 1611 1612 return MadeChange; 1613 } 1614 1615 //===----------------------------------------------------------------------===// 1616 // 1617 // SCCPSolver implementations 1618 // 1619 SCCPSolver::SCCPSolver( 1620 const DataLayout &DL, 1621 std::function<const TargetLibraryInfo &(Function &)> GetTLI, 1622 LLVMContext &Ctx) 1623 : Visitor(new SCCPInstVisitor(DL, std::move(GetTLI), Ctx)) {} 1624 1625 SCCPSolver::~SCCPSolver() = default; 1626 1627 void SCCPSolver::addAnalysis(Function &F, AnalysisResultsForFn A) { 1628 return Visitor->addAnalysis(F, std::move(A)); 1629 } 1630 1631 bool SCCPSolver::markBlockExecutable(BasicBlock *BB) { 1632 return Visitor->markBlockExecutable(BB); 1633 } 1634 1635 const PredicateBase *SCCPSolver::getPredicateInfoFor(Instruction *I) { 1636 return Visitor->getPredicateInfoFor(I); 1637 } 1638 1639 DomTreeUpdater SCCPSolver::getDTU(Function &F) { return Visitor->getDTU(F); } 1640 1641 void SCCPSolver::trackValueOfGlobalVariable(GlobalVariable *GV) { 1642 Visitor->trackValueOfGlobalVariable(GV); 1643 } 1644 1645 void SCCPSolver::addTrackedFunction(Function *F) { 1646 Visitor->addTrackedFunction(F); 1647 } 1648 1649 void SCCPSolver::addToMustPreserveReturnsInFunctions(Function *F) { 1650 Visitor->addToMustPreserveReturnsInFunctions(F); 1651 } 1652 1653 bool SCCPSolver::mustPreserveReturn(Function *F) { 1654 return Visitor->mustPreserveReturn(F); 1655 } 1656 1657 void SCCPSolver::addArgumentTrackedFunction(Function *F) { 1658 Visitor->addArgumentTrackedFunction(F); 1659 } 1660 1661 bool SCCPSolver::isArgumentTrackedFunction(Function *F) { 1662 return Visitor->isArgumentTrackedFunction(F); 1663 } 1664 1665 void SCCPSolver::solve() { Visitor->solve(); } 1666 1667 bool SCCPSolver::resolvedUndefsIn(Function &F) { 1668 return Visitor->resolvedUndefsIn(F); 1669 } 1670 1671 bool SCCPSolver::isBlockExecutable(BasicBlock *BB) const { 1672 return Visitor->isBlockExecutable(BB); 1673 } 1674 1675 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const { 1676 return Visitor->isEdgeFeasible(From, To); 1677 } 1678 1679 std::vector<ValueLatticeElement> 1680 SCCPSolver::getStructLatticeValueFor(Value *V) const { 1681 return Visitor->getStructLatticeValueFor(V); 1682 } 1683 1684 void SCCPSolver::removeLatticeValueFor(Value *V) { 1685 return Visitor->removeLatticeValueFor(V); 1686 } 1687 1688 const ValueLatticeElement &SCCPSolver::getLatticeValueFor(Value *V) const { 1689 return Visitor->getLatticeValueFor(V); 1690 } 1691 1692 const MapVector<Function *, ValueLatticeElement> & 1693 SCCPSolver::getTrackedRetVals() { 1694 return Visitor->getTrackedRetVals(); 1695 } 1696 1697 const DenseMap<GlobalVariable *, ValueLatticeElement> & 1698 SCCPSolver::getTrackedGlobals() { 1699 return Visitor->getTrackedGlobals(); 1700 } 1701 1702 const SmallPtrSet<Function *, 16> SCCPSolver::getMRVFunctionsTracked() { 1703 return Visitor->getMRVFunctionsTracked(); 1704 } 1705 1706 void SCCPSolver::markOverdefined(Value *V) { Visitor->markOverdefined(V); } 1707 1708 bool SCCPSolver::isStructLatticeConstant(Function *F, StructType *STy) { 1709 return Visitor->isStructLatticeConstant(F, STy); 1710 } 1711 1712 Constant *SCCPSolver::getConstant(const ValueLatticeElement &LV) const { 1713 return Visitor->getConstant(LV); 1714 } 1715 1716 SmallPtrSetImpl<Function *> &SCCPSolver::getArgumentTrackedFunctions() { 1717 return Visitor->getArgumentTrackedFunctions(); 1718 } 1719 1720 void SCCPSolver::markArgInFuncSpecialization(Function *F, const ArgInfo &Arg) { 1721 Visitor->markArgInFuncSpecialization(F, Arg); 1722 } 1723 1724 void SCCPSolver::markFunctionUnreachable(Function *F) { 1725 Visitor->markFunctionUnreachable(F); 1726 } 1727 1728 void SCCPSolver::visit(Instruction *I) { Visitor->visit(I); } 1729 1730 void SCCPSolver::visitCall(CallInst &I) { Visitor->visitCall(I); } 1731