1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 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 // Peephole optimize the CFG. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/APInt.h" 14 #include "llvm/ADT/ArrayRef.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/MapVector.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/Sequence.h" 19 #include "llvm/ADT/SetOperations.h" 20 #include "llvm/ADT/SetVector.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/ADT/StringRef.h" 25 #include "llvm/Analysis/AssumptionCache.h" 26 #include "llvm/Analysis/CaptureTracking.h" 27 #include "llvm/Analysis/ConstantFolding.h" 28 #include "llvm/Analysis/DomTreeUpdater.h" 29 #include "llvm/Analysis/GuardUtils.h" 30 #include "llvm/Analysis/InstructionSimplify.h" 31 #include "llvm/Analysis/MemorySSA.h" 32 #include "llvm/Analysis/MemorySSAUpdater.h" 33 #include "llvm/Analysis/TargetTransformInfo.h" 34 #include "llvm/Analysis/ValueTracking.h" 35 #include "llvm/IR/Attributes.h" 36 #include "llvm/IR/BasicBlock.h" 37 #include "llvm/IR/CFG.h" 38 #include "llvm/IR/Constant.h" 39 #include "llvm/IR/ConstantRange.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/DebugInfo.h" 43 #include "llvm/IR/DerivedTypes.h" 44 #include "llvm/IR/Function.h" 45 #include "llvm/IR/GlobalValue.h" 46 #include "llvm/IR/GlobalVariable.h" 47 #include "llvm/IR/IRBuilder.h" 48 #include "llvm/IR/InstrTypes.h" 49 #include "llvm/IR/Instruction.h" 50 #include "llvm/IR/Instructions.h" 51 #include "llvm/IR/IntrinsicInst.h" 52 #include "llvm/IR/LLVMContext.h" 53 #include "llvm/IR/MDBuilder.h" 54 #include "llvm/IR/Metadata.h" 55 #include "llvm/IR/Module.h" 56 #include "llvm/IR/NoFolder.h" 57 #include "llvm/IR/Operator.h" 58 #include "llvm/IR/PatternMatch.h" 59 #include "llvm/IR/ProfDataUtils.h" 60 #include "llvm/IR/Type.h" 61 #include "llvm/IR/Use.h" 62 #include "llvm/IR/User.h" 63 #include "llvm/IR/Value.h" 64 #include "llvm/IR/ValueHandle.h" 65 #include "llvm/Support/BranchProbability.h" 66 #include "llvm/Support/Casting.h" 67 #include "llvm/Support/CommandLine.h" 68 #include "llvm/Support/Debug.h" 69 #include "llvm/Support/ErrorHandling.h" 70 #include "llvm/Support/KnownBits.h" 71 #include "llvm/Support/MathExtras.h" 72 #include "llvm/Support/raw_ostream.h" 73 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 74 #include "llvm/Transforms/Utils/Local.h" 75 #include "llvm/Transforms/Utils/ValueMapper.h" 76 #include <algorithm> 77 #include <cassert> 78 #include <climits> 79 #include <cstddef> 80 #include <cstdint> 81 #include <iterator> 82 #include <map> 83 #include <optional> 84 #include <set> 85 #include <tuple> 86 #include <utility> 87 #include <vector> 88 89 using namespace llvm; 90 using namespace PatternMatch; 91 92 #define DEBUG_TYPE "simplifycfg" 93 94 cl::opt<bool> llvm::RequireAndPreserveDomTree( 95 "simplifycfg-require-and-preserve-domtree", cl::Hidden, 96 97 cl::desc("Temorary development switch used to gradually uplift SimplifyCFG " 98 "into preserving DomTree,")); 99 100 // Chosen as 2 so as to be cheap, but still to have enough power to fold 101 // a select, so the "clamp" idiom (of a min followed by a max) will be caught. 102 // To catch this, we need to fold a compare and a select, hence '2' being the 103 // minimum reasonable default. 104 static cl::opt<unsigned> PHINodeFoldingThreshold( 105 "phi-node-folding-threshold", cl::Hidden, cl::init(2), 106 cl::desc( 107 "Control the amount of phi node folding to perform (default = 2)")); 108 109 static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold( 110 "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4), 111 cl::desc("Control the maximal total instruction cost that we are willing " 112 "to speculatively execute to fold a 2-entry PHI node into a " 113 "select (default = 4)")); 114 115 static cl::opt<bool> 116 HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true), 117 cl::desc("Hoist common instructions up to the parent block")); 118 119 static cl::opt<unsigned> 120 HoistCommonSkipLimit("simplifycfg-hoist-common-skip-limit", cl::Hidden, 121 cl::init(20), 122 cl::desc("Allow reordering across at most this many " 123 "instructions when hoisting")); 124 125 static cl::opt<bool> 126 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 127 cl::desc("Sink common instructions down to the end block")); 128 129 static cl::opt<bool> HoistCondStores( 130 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true), 131 cl::desc("Hoist conditional stores if an unconditional store precedes")); 132 133 static cl::opt<bool> MergeCondStores( 134 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true), 135 cl::desc("Hoist conditional stores even if an unconditional store does not " 136 "precede - hoist multiple conditional stores into a single " 137 "predicated store")); 138 139 static cl::opt<bool> MergeCondStoresAggressively( 140 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false), 141 cl::desc("When merging conditional stores, do so even if the resultant " 142 "basic blocks are unlikely to be if-converted as a result")); 143 144 static cl::opt<bool> SpeculateOneExpensiveInst( 145 "speculate-one-expensive-inst", cl::Hidden, cl::init(true), 146 cl::desc("Allow exactly one expensive instruction to be speculatively " 147 "executed")); 148 149 static cl::opt<unsigned> MaxSpeculationDepth( 150 "max-speculation-depth", cl::Hidden, cl::init(10), 151 cl::desc("Limit maximum recursion depth when calculating costs of " 152 "speculatively executed instructions")); 153 154 static cl::opt<int> 155 MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden, 156 cl::init(10), 157 cl::desc("Max size of a block which is still considered " 158 "small enough to thread through")); 159 160 // Two is chosen to allow one negation and a logical combine. 161 static cl::opt<unsigned> 162 BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden, 163 cl::init(2), 164 cl::desc("Maximum cost of combining conditions when " 165 "folding branches")); 166 167 static cl::opt<unsigned> BranchFoldToCommonDestVectorMultiplier( 168 "simplifycfg-branch-fold-common-dest-vector-multiplier", cl::Hidden, 169 cl::init(2), 170 cl::desc("Multiplier to apply to threshold when determining whether or not " 171 "to fold branch to common destination when vector operations are " 172 "present")); 173 174 static cl::opt<bool> EnableMergeCompatibleInvokes( 175 "simplifycfg-merge-compatible-invokes", cl::Hidden, cl::init(true), 176 cl::desc("Allow SimplifyCFG to merge invokes together when appropriate")); 177 178 static cl::opt<unsigned> MaxSwitchCasesPerResult( 179 "max-switch-cases-per-result", cl::Hidden, cl::init(16), 180 cl::desc("Limit cases to analyze when converting a switch to select")); 181 182 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 183 STATISTIC(NumLinearMaps, 184 "Number of switch instructions turned into linear mapping"); 185 STATISTIC(NumLookupTables, 186 "Number of switch instructions turned into lookup tables"); 187 STATISTIC( 188 NumLookupTablesHoles, 189 "Number of switch instructions turned into lookup tables (holes checked)"); 190 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares"); 191 STATISTIC(NumFoldValueComparisonIntoPredecessors, 192 "Number of value comparisons folded into predecessor basic blocks"); 193 STATISTIC(NumFoldBranchToCommonDest, 194 "Number of branches folded into predecessor basic block"); 195 STATISTIC( 196 NumHoistCommonCode, 197 "Number of common instruction 'blocks' hoisted up to the begin block"); 198 STATISTIC(NumHoistCommonInstrs, 199 "Number of common instructions hoisted up to the begin block"); 200 STATISTIC(NumSinkCommonCode, 201 "Number of common instruction 'blocks' sunk down to the end block"); 202 STATISTIC(NumSinkCommonInstrs, 203 "Number of common instructions sunk down to the end block"); 204 STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 205 STATISTIC(NumInvokes, 206 "Number of invokes with empty resume blocks simplified into calls"); 207 STATISTIC(NumInvokesMerged, "Number of invokes that were merged together"); 208 STATISTIC(NumInvokeSetsFormed, "Number of invoke sets that were formed"); 209 210 namespace { 211 212 // The first field contains the value that the switch produces when a certain 213 // case group is selected, and the second field is a vector containing the 214 // cases composing the case group. 215 using SwitchCaseResultVectorTy = 216 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>; 217 218 // The first field contains the phi node that generates a result of the switch 219 // and the second field contains the value generated for a certain case in the 220 // switch for that PHI. 221 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>; 222 223 /// ValueEqualityComparisonCase - Represents a case of a switch. 224 struct ValueEqualityComparisonCase { 225 ConstantInt *Value; 226 BasicBlock *Dest; 227 228 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 229 : Value(Value), Dest(Dest) {} 230 231 bool operator<(ValueEqualityComparisonCase RHS) const { 232 // Comparing pointers is ok as we only rely on the order for uniquing. 233 return Value < RHS.Value; 234 } 235 236 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 237 }; 238 239 class SimplifyCFGOpt { 240 const TargetTransformInfo &TTI; 241 DomTreeUpdater *DTU; 242 const DataLayout &DL; 243 ArrayRef<WeakVH> LoopHeaders; 244 const SimplifyCFGOptions &Options; 245 bool Resimplify; 246 247 Value *isValueEqualityComparison(Instruction *TI); 248 BasicBlock *GetValueEqualityComparisonCases( 249 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases); 250 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI, 251 BasicBlock *Pred, 252 IRBuilder<> &Builder); 253 bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV, 254 Instruction *PTI, 255 IRBuilder<> &Builder); 256 bool FoldValueComparisonIntoPredecessors(Instruction *TI, 257 IRBuilder<> &Builder); 258 259 bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 260 bool simplifySingleResume(ResumeInst *RI); 261 bool simplifyCommonResume(ResumeInst *RI); 262 bool simplifyCleanupReturn(CleanupReturnInst *RI); 263 bool simplifyUnreachable(UnreachableInst *UI); 264 bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 265 bool simplifyIndirectBr(IndirectBrInst *IBI); 266 bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder); 267 bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder); 268 bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder); 269 270 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 271 IRBuilder<> &Builder); 272 273 bool hoistCommonCodeFromSuccessors(BasicBlock *BB, bool EqTermsOnly); 274 bool hoistSuccIdenticalTerminatorToSwitchOrIf( 275 Instruction *TI, Instruction *I1, 276 SmallVectorImpl<Instruction *> &OtherSuccTIs); 277 bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB); 278 bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond, 279 BasicBlock *TrueBB, BasicBlock *FalseBB, 280 uint32_t TrueWeight, uint32_t FalseWeight); 281 bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder, 282 const DataLayout &DL); 283 bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select); 284 bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI); 285 bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder); 286 287 public: 288 SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU, 289 const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders, 290 const SimplifyCFGOptions &Opts) 291 : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) { 292 assert((!DTU || !DTU->hasPostDomTree()) && 293 "SimplifyCFG is not yet capable of maintaining validity of a " 294 "PostDomTree, so don't ask for it."); 295 } 296 297 bool simplifyOnce(BasicBlock *BB); 298 bool run(BasicBlock *BB); 299 300 // Helper to set Resimplify and return change indication. 301 bool requestResimplify() { 302 Resimplify = true; 303 return true; 304 } 305 }; 306 307 } // end anonymous namespace 308 309 /// Return true if all the PHI nodes in the basic block \p BB 310 /// receive compatible (identical) incoming values when coming from 311 /// all of the predecessor blocks that are specified in \p IncomingBlocks. 312 /// 313 /// Note that if the values aren't exactly identical, but \p EquivalenceSet 314 /// is provided, and *both* of the values are present in the set, 315 /// then they are considered equal. 316 static bool IncomingValuesAreCompatible( 317 BasicBlock *BB, ArrayRef<BasicBlock *> IncomingBlocks, 318 SmallPtrSetImpl<Value *> *EquivalenceSet = nullptr) { 319 assert(IncomingBlocks.size() == 2 && 320 "Only for a pair of incoming blocks at the time!"); 321 322 // FIXME: it is okay if one of the incoming values is an `undef` value, 323 // iff the other incoming value is guaranteed to be a non-poison value. 324 // FIXME: it is okay if one of the incoming values is a `poison` value. 325 return all_of(BB->phis(), [IncomingBlocks, EquivalenceSet](PHINode &PN) { 326 Value *IV0 = PN.getIncomingValueForBlock(IncomingBlocks[0]); 327 Value *IV1 = PN.getIncomingValueForBlock(IncomingBlocks[1]); 328 if (IV0 == IV1) 329 return true; 330 if (EquivalenceSet && EquivalenceSet->contains(IV0) && 331 EquivalenceSet->contains(IV1)) 332 return true; 333 return false; 334 }); 335 } 336 337 /// Return true if it is safe to merge these two 338 /// terminator instructions together. 339 static bool 340 SafeToMergeTerminators(Instruction *SI1, Instruction *SI2, 341 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) { 342 if (SI1 == SI2) 343 return false; // Can't merge with self! 344 345 // It is not safe to merge these two switch instructions if they have a common 346 // successor, and if that successor has a PHI node, and if *that* PHI node has 347 // conflicting incoming values from the two switch blocks. 348 BasicBlock *SI1BB = SI1->getParent(); 349 BasicBlock *SI2BB = SI2->getParent(); 350 351 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 352 bool Fail = false; 353 for (BasicBlock *Succ : successors(SI2BB)) { 354 if (!SI1Succs.count(Succ)) 355 continue; 356 if (IncomingValuesAreCompatible(Succ, {SI1BB, SI2BB})) 357 continue; 358 Fail = true; 359 if (FailBlocks) 360 FailBlocks->insert(Succ); 361 else 362 break; 363 } 364 365 return !Fail; 366 } 367 368 /// Update PHI nodes in Succ to indicate that there will now be entries in it 369 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes 370 /// will be the same as those coming in from ExistPred, an existing predecessor 371 /// of Succ. 372 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 373 BasicBlock *ExistPred, 374 MemorySSAUpdater *MSSAU = nullptr) { 375 for (PHINode &PN : Succ->phis()) 376 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred); 377 if (MSSAU) 378 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ)) 379 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred); 380 } 381 382 /// Compute an abstract "cost" of speculating the given instruction, 383 /// which is assumed to be safe to speculate. TCC_Free means cheap, 384 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively 385 /// expensive. 386 static InstructionCost computeSpeculationCost(const User *I, 387 const TargetTransformInfo &TTI) { 388 assert((!isa<Instruction>(I) || 389 isSafeToSpeculativelyExecute(cast<Instruction>(I))) && 390 "Instruction is not safe to speculatively execute!"); 391 return TTI.getInstructionCost(I, TargetTransformInfo::TCK_SizeAndLatency); 392 } 393 394 /// If we have a merge point of an "if condition" as accepted above, 395 /// return true if the specified value dominates the block. We 396 /// don't handle the true generality of domination here, just a special case 397 /// which works well enough for us. 398 /// 399 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 400 /// see if V (which must be an instruction) and its recursive operands 401 /// that do not dominate BB have a combined cost lower than Budget and 402 /// are non-trapping. If both are true, the instruction is inserted into the 403 /// set and true is returned. 404 /// 405 /// The cost for most non-trapping instructions is defined as 1 except for 406 /// Select whose cost is 2. 407 /// 408 /// After this function returns, Cost is increased by the cost of 409 /// V plus its non-dominating operands. If that cost is greater than 410 /// Budget, false is returned and Cost is undefined. 411 static bool dominatesMergePoint(Value *V, BasicBlock *BB, 412 SmallPtrSetImpl<Instruction *> &AggressiveInsts, 413 InstructionCost &Cost, 414 InstructionCost Budget, 415 const TargetTransformInfo &TTI, 416 unsigned Depth = 0) { 417 // It is possible to hit a zero-cost cycle (phi/gep instructions for example), 418 // so limit the recursion depth. 419 // TODO: While this recursion limit does prevent pathological behavior, it 420 // would be better to track visited instructions to avoid cycles. 421 if (Depth == MaxSpeculationDepth) 422 return false; 423 424 Instruction *I = dyn_cast<Instruction>(V); 425 if (!I) { 426 // Non-instructions dominate all instructions and can be executed 427 // unconditionally. 428 return true; 429 } 430 BasicBlock *PBB = I->getParent(); 431 432 // We don't want to allow weird loops that might have the "if condition" in 433 // the bottom of this block. 434 if (PBB == BB) 435 return false; 436 437 // If this instruction is defined in a block that contains an unconditional 438 // branch to BB, then it must be in the 'conditional' part of the "if 439 // statement". If not, it definitely dominates the region. 440 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 441 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB) 442 return true; 443 444 // If we have seen this instruction before, don't count it again. 445 if (AggressiveInsts.count(I)) 446 return true; 447 448 // Okay, it looks like the instruction IS in the "condition". Check to 449 // see if it's a cheap instruction to unconditionally compute, and if it 450 // only uses stuff defined outside of the condition. If so, hoist it out. 451 if (!isSafeToSpeculativelyExecute(I)) 452 return false; 453 454 Cost += computeSpeculationCost(I, TTI); 455 456 // Allow exactly one instruction to be speculated regardless of its cost 457 // (as long as it is safe to do so). 458 // This is intended to flatten the CFG even if the instruction is a division 459 // or other expensive operation. The speculation of an expensive instruction 460 // is expected to be undone in CodeGenPrepare if the speculation has not 461 // enabled further IR optimizations. 462 if (Cost > Budget && 463 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 || 464 !Cost.isValid())) 465 return false; 466 467 // Okay, we can only really hoist these out if their operands do 468 // not take us over the cost threshold. 469 for (Use &Op : I->operands()) 470 if (!dominatesMergePoint(Op, BB, AggressiveInsts, Cost, Budget, TTI, 471 Depth + 1)) 472 return false; 473 // Okay, it's safe to do this! Remember this instruction. 474 AggressiveInsts.insert(I); 475 return true; 476 } 477 478 /// Extract ConstantInt from value, looking through IntToPtr 479 /// and PointerNullValue. Return NULL if value is not a constant int. 480 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) { 481 // Normal constant int. 482 ConstantInt *CI = dyn_cast<ConstantInt>(V); 483 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy() || 484 DL.isNonIntegralPointerType(V->getType())) 485 return CI; 486 487 // This is some kind of pointer constant. Turn it into a pointer-sized 488 // ConstantInt if possible. 489 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType())); 490 491 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 492 if (isa<ConstantPointerNull>(V)) 493 return ConstantInt::get(PtrTy, 0); 494 495 // IntToPtr const int. 496 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 497 if (CE->getOpcode() == Instruction::IntToPtr) 498 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 499 // The constant is very likely to have the right type already. 500 if (CI->getType() == PtrTy) 501 return CI; 502 else 503 return cast<ConstantInt>( 504 ConstantFoldIntegerCast(CI, PtrTy, /*isSigned=*/false, DL)); 505 } 506 return nullptr; 507 } 508 509 namespace { 510 511 /// Given a chain of or (||) or and (&&) comparison of a value against a 512 /// constant, this will try to recover the information required for a switch 513 /// structure. 514 /// It will depth-first traverse the chain of comparison, seeking for patterns 515 /// like %a == 12 or %a < 4 and combine them to produce a set of integer 516 /// representing the different cases for the switch. 517 /// Note that if the chain is composed of '||' it will build the set of elements 518 /// that matches the comparisons (i.e. any of this value validate the chain) 519 /// while for a chain of '&&' it will build the set elements that make the test 520 /// fail. 521 struct ConstantComparesGatherer { 522 const DataLayout &DL; 523 524 /// Value found for the switch comparison 525 Value *CompValue = nullptr; 526 527 /// Extra clause to be checked before the switch 528 Value *Extra = nullptr; 529 530 /// Set of integers to match in switch 531 SmallVector<ConstantInt *, 8> Vals; 532 533 /// Number of comparisons matched in the and/or chain 534 unsigned UsedICmps = 0; 535 536 /// Construct and compute the result for the comparison instruction Cond 537 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) { 538 gather(Cond); 539 } 540 541 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete; 542 ConstantComparesGatherer & 543 operator=(const ConstantComparesGatherer &) = delete; 544 545 private: 546 /// Try to set the current value used for the comparison, it succeeds only if 547 /// it wasn't set before or if the new value is the same as the old one 548 bool setValueOnce(Value *NewVal) { 549 if (CompValue && CompValue != NewVal) 550 return false; 551 CompValue = NewVal; 552 return (CompValue != nullptr); 553 } 554 555 /// Try to match Instruction "I" as a comparison against a constant and 556 /// populates the array Vals with the set of values that match (or do not 557 /// match depending on isEQ). 558 /// Return false on failure. On success, the Value the comparison matched 559 /// against is placed in CompValue. 560 /// If CompValue is already set, the function is expected to fail if a match 561 /// is found but the value compared to is different. 562 bool matchInstruction(Instruction *I, bool isEQ) { 563 // If this is an icmp against a constant, handle this as one of the cases. 564 ICmpInst *ICI; 565 ConstantInt *C; 566 if (!((ICI = dyn_cast<ICmpInst>(I)) && 567 (C = GetConstantInt(I->getOperand(1), DL)))) { 568 return false; 569 } 570 571 Value *RHSVal; 572 const APInt *RHSC; 573 574 // Pattern match a special case 575 // (x & ~2^z) == y --> x == y || x == y|2^z 576 // This undoes a transformation done by instcombine to fuse 2 compares. 577 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) { 578 // It's a little bit hard to see why the following transformations are 579 // correct. Here is a CVC3 program to verify them for 64-bit values: 580 581 /* 582 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63); 583 x : BITVECTOR(64); 584 y : BITVECTOR(64); 585 z : BITVECTOR(64); 586 mask : BITVECTOR(64) = BVSHL(ONE, z); 587 QUERY( (y & ~mask = y) => 588 ((x & ~mask = y) <=> (x = y OR x = (y | mask))) 589 ); 590 QUERY( (y | mask = y) => 591 ((x | mask = y) <=> (x = y OR x = (y & ~mask))) 592 ); 593 */ 594 595 // Please note that each pattern must be a dual implication (<--> or 596 // iff). One directional implication can create spurious matches. If the 597 // implication is only one-way, an unsatisfiable condition on the left 598 // side can imply a satisfiable condition on the right side. Dual 599 // implication ensures that satisfiable conditions are transformed to 600 // other satisfiable conditions and unsatisfiable conditions are 601 // transformed to other unsatisfiable conditions. 602 603 // Here is a concrete example of a unsatisfiable condition on the left 604 // implying a satisfiable condition on the right: 605 // 606 // mask = (1 << z) 607 // (x & ~mask) == y --> (x == y || x == (y | mask)) 608 // 609 // Substituting y = 3, z = 0 yields: 610 // (x & -2) == 3 --> (x == 3 || x == 2) 611 612 // Pattern match a special case: 613 /* 614 QUERY( (y & ~mask = y) => 615 ((x & ~mask = y) <=> (x = y OR x = (y | mask))) 616 ); 617 */ 618 if (match(ICI->getOperand(0), 619 m_And(m_Value(RHSVal), m_APInt(RHSC)))) { 620 APInt Mask = ~*RHSC; 621 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) { 622 // If we already have a value for the switch, it has to match! 623 if (!setValueOnce(RHSVal)) 624 return false; 625 626 Vals.push_back(C); 627 Vals.push_back( 628 ConstantInt::get(C->getContext(), 629 C->getValue() | Mask)); 630 UsedICmps++; 631 return true; 632 } 633 } 634 635 // Pattern match a special case: 636 /* 637 QUERY( (y | mask = y) => 638 ((x | mask = y) <=> (x = y OR x = (y & ~mask))) 639 ); 640 */ 641 if (match(ICI->getOperand(0), 642 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) { 643 APInt Mask = *RHSC; 644 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) { 645 // If we already have a value for the switch, it has to match! 646 if (!setValueOnce(RHSVal)) 647 return false; 648 649 Vals.push_back(C); 650 Vals.push_back(ConstantInt::get(C->getContext(), 651 C->getValue() & ~Mask)); 652 UsedICmps++; 653 return true; 654 } 655 } 656 657 // If we already have a value for the switch, it has to match! 658 if (!setValueOnce(ICI->getOperand(0))) 659 return false; 660 661 UsedICmps++; 662 Vals.push_back(C); 663 return ICI->getOperand(0); 664 } 665 666 // If we have "x ult 3", for example, then we can add 0,1,2 to the set. 667 ConstantRange Span = 668 ConstantRange::makeExactICmpRegion(ICI->getPredicate(), C->getValue()); 669 670 // Shift the range if the compare is fed by an add. This is the range 671 // compare idiom as emitted by instcombine. 672 Value *CandidateVal = I->getOperand(0); 673 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) { 674 Span = Span.subtract(*RHSC); 675 CandidateVal = RHSVal; 676 } 677 678 // If this is an and/!= check, then we are looking to build the set of 679 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into 680 // x != 0 && x != 1. 681 if (!isEQ) 682 Span = Span.inverse(); 683 684 // If there are a ton of values, we don't want to make a ginormous switch. 685 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) { 686 return false; 687 } 688 689 // If we already have a value for the switch, it has to match! 690 if (!setValueOnce(CandidateVal)) 691 return false; 692 693 // Add all values from the range to the set 694 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 695 Vals.push_back(ConstantInt::get(I->getContext(), Tmp)); 696 697 UsedICmps++; 698 return true; 699 } 700 701 /// Given a potentially 'or'd or 'and'd together collection of icmp 702 /// eq/ne/lt/gt instructions that compare a value against a constant, extract 703 /// the value being compared, and stick the list constants into the Vals 704 /// vector. 705 /// One "Extra" case is allowed to differ from the other. 706 void gather(Value *V) { 707 bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value())); 708 709 // Keep a stack (SmallVector for efficiency) for depth-first traversal 710 SmallVector<Value *, 8> DFT; 711 SmallPtrSet<Value *, 8> Visited; 712 713 // Initialize 714 Visited.insert(V); 715 DFT.push_back(V); 716 717 while (!DFT.empty()) { 718 V = DFT.pop_back_val(); 719 720 if (Instruction *I = dyn_cast<Instruction>(V)) { 721 // If it is a || (or && depending on isEQ), process the operands. 722 Value *Op0, *Op1; 723 if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) 724 : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) { 725 if (Visited.insert(Op1).second) 726 DFT.push_back(Op1); 727 if (Visited.insert(Op0).second) 728 DFT.push_back(Op0); 729 730 continue; 731 } 732 733 // Try to match the current instruction 734 if (matchInstruction(I, isEQ)) 735 // Match succeed, continue the loop 736 continue; 737 } 738 739 // One element of the sequence of || (or &&) could not be match as a 740 // comparison against the same value as the others. 741 // We allow only one "Extra" case to be checked before the switch 742 if (!Extra) { 743 Extra = V; 744 continue; 745 } 746 // Failed to parse a proper sequence, abort now 747 CompValue = nullptr; 748 break; 749 } 750 } 751 }; 752 753 } // end anonymous namespace 754 755 static void EraseTerminatorAndDCECond(Instruction *TI, 756 MemorySSAUpdater *MSSAU = nullptr) { 757 Instruction *Cond = nullptr; 758 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 759 Cond = dyn_cast<Instruction>(SI->getCondition()); 760 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 761 if (BI->isConditional()) 762 Cond = dyn_cast<Instruction>(BI->getCondition()); 763 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 764 Cond = dyn_cast<Instruction>(IBI->getAddress()); 765 } 766 767 TI->eraseFromParent(); 768 if (Cond) 769 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU); 770 } 771 772 /// Return true if the specified terminator checks 773 /// to see if a value is equal to constant integer value. 774 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) { 775 Value *CV = nullptr; 776 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 777 // Do not permit merging of large switch instructions into their 778 // predecessors unless there is only one predecessor. 779 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors())) 780 CV = SI->getCondition(); 781 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 782 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 783 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) { 784 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL)) 785 CV = ICI->getOperand(0); 786 } 787 788 // Unwrap any lossless ptrtoint cast. 789 if (CV) { 790 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) { 791 Value *Ptr = PTII->getPointerOperand(); 792 if (PTII->getType() == DL.getIntPtrType(Ptr->getType())) 793 CV = Ptr; 794 } 795 } 796 return CV; 797 } 798 799 /// Given a value comparison instruction, 800 /// decode all of the 'cases' that it represents and return the 'default' block. 801 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases( 802 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) { 803 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 804 Cases.reserve(SI->getNumCases()); 805 for (auto Case : SI->cases()) 806 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(), 807 Case.getCaseSuccessor())); 808 return SI->getDefaultDest(); 809 } 810 811 BranchInst *BI = cast<BranchInst>(TI); 812 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 813 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 814 Cases.push_back(ValueEqualityComparisonCase( 815 GetConstantInt(ICI->getOperand(1), DL), Succ)); 816 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 817 } 818 819 /// Given a vector of bb/value pairs, remove any entries 820 /// in the list that match the specified block. 821 static void 822 EliminateBlockCases(BasicBlock *BB, 823 std::vector<ValueEqualityComparisonCase> &Cases) { 824 llvm::erase(Cases, BB); 825 } 826 827 /// Return true if there are any keys in C1 that exist in C2 as well. 828 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 829 std::vector<ValueEqualityComparisonCase> &C2) { 830 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 831 832 // Make V1 be smaller than V2. 833 if (V1->size() > V2->size()) 834 std::swap(V1, V2); 835 836 if (V1->empty()) 837 return false; 838 if (V1->size() == 1) { 839 // Just scan V2. 840 ConstantInt *TheVal = (*V1)[0].Value; 841 for (const ValueEqualityComparisonCase &VECC : *V2) 842 if (TheVal == VECC.Value) 843 return true; 844 } 845 846 // Otherwise, just sort both lists and compare element by element. 847 array_pod_sort(V1->begin(), V1->end()); 848 array_pod_sort(V2->begin(), V2->end()); 849 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 850 while (i1 != e1 && i2 != e2) { 851 if ((*V1)[i1].Value == (*V2)[i2].Value) 852 return true; 853 if ((*V1)[i1].Value < (*V2)[i2].Value) 854 ++i1; 855 else 856 ++i2; 857 } 858 return false; 859 } 860 861 // Set branch weights on SwitchInst. This sets the metadata if there is at 862 // least one non-zero weight. 863 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) { 864 // Check that there is at least one non-zero weight. Otherwise, pass 865 // nullptr to setMetadata which will erase the existing metadata. 866 MDNode *N = nullptr; 867 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; })) 868 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights); 869 SI->setMetadata(LLVMContext::MD_prof, N); 870 } 871 872 // Similar to the above, but for branch and select instructions that take 873 // exactly 2 weights. 874 static void setBranchWeights(Instruction *I, uint32_t TrueWeight, 875 uint32_t FalseWeight) { 876 assert(isa<BranchInst>(I) || isa<SelectInst>(I)); 877 // Check that there is at least one non-zero weight. Otherwise, pass 878 // nullptr to setMetadata which will erase the existing metadata. 879 MDNode *N = nullptr; 880 if (TrueWeight || FalseWeight) 881 N = MDBuilder(I->getParent()->getContext()) 882 .createBranchWeights(TrueWeight, FalseWeight); 883 I->setMetadata(LLVMContext::MD_prof, N); 884 } 885 886 /// If TI is known to be a terminator instruction and its block is known to 887 /// only have a single predecessor block, check to see if that predecessor is 888 /// also a value comparison with the same value, and if that comparison 889 /// determines the outcome of this comparison. If so, simplify TI. This does a 890 /// very limited form of jump threading. 891 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor( 892 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) { 893 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 894 if (!PredVal) 895 return false; // Not a value comparison in predecessor. 896 897 Value *ThisVal = isValueEqualityComparison(TI); 898 assert(ThisVal && "This isn't a value comparison!!"); 899 if (ThisVal != PredVal) 900 return false; // Different predicates. 901 902 // TODO: Preserve branch weight metadata, similarly to how 903 // FoldValueComparisonIntoPredecessors preserves it. 904 905 // Find out information about when control will move from Pred to TI's block. 906 std::vector<ValueEqualityComparisonCase> PredCases; 907 BasicBlock *PredDef = 908 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases); 909 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 910 911 // Find information about how control leaves this block. 912 std::vector<ValueEqualityComparisonCase> ThisCases; 913 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 914 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 915 916 // If TI's block is the default block from Pred's comparison, potentially 917 // simplify TI based on this knowledge. 918 if (PredDef == TI->getParent()) { 919 // If we are here, we know that the value is none of those cases listed in 920 // PredCases. If there are any cases in ThisCases that are in PredCases, we 921 // can simplify TI. 922 if (!ValuesOverlap(PredCases, ThisCases)) 923 return false; 924 925 if (isa<BranchInst>(TI)) { 926 // Okay, one of the successors of this condbr is dead. Convert it to a 927 // uncond br. 928 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 929 // Insert the new branch. 930 Instruction *NI = Builder.CreateBr(ThisDef); 931 (void)NI; 932 933 // Remove PHI node entries for the dead edge. 934 ThisCases[0].Dest->removePredecessor(PredDef); 935 936 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 937 << "Through successor TI: " << *TI << "Leaving: " << *NI 938 << "\n"); 939 940 EraseTerminatorAndDCECond(TI); 941 942 if (DTU) 943 DTU->applyUpdates( 944 {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}}); 945 946 return true; 947 } 948 949 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI); 950 // Okay, TI has cases that are statically dead, prune them away. 951 SmallPtrSet<Constant *, 16> DeadCases; 952 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 953 DeadCases.insert(PredCases[i].Value); 954 955 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 956 << "Through successor TI: " << *TI); 957 958 SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases; 959 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 960 --i; 961 auto *Successor = i->getCaseSuccessor(); 962 if (DTU) 963 ++NumPerSuccessorCases[Successor]; 964 if (DeadCases.count(i->getCaseValue())) { 965 Successor->removePredecessor(PredDef); 966 SI.removeCase(i); 967 if (DTU) 968 --NumPerSuccessorCases[Successor]; 969 } 970 } 971 972 if (DTU) { 973 std::vector<DominatorTree::UpdateType> Updates; 974 for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases) 975 if (I.second == 0) 976 Updates.push_back({DominatorTree::Delete, PredDef, I.first}); 977 DTU->applyUpdates(Updates); 978 } 979 980 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 981 return true; 982 } 983 984 // Otherwise, TI's block must correspond to some matched value. Find out 985 // which value (or set of values) this is. 986 ConstantInt *TIV = nullptr; 987 BasicBlock *TIBB = TI->getParent(); 988 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 989 if (PredCases[i].Dest == TIBB) { 990 if (TIV) 991 return false; // Cannot handle multiple values coming to this block. 992 TIV = PredCases[i].Value; 993 } 994 assert(TIV && "No edge from pred to succ?"); 995 996 // Okay, we found the one constant that our value can be if we get into TI's 997 // BB. Find out which successor will unconditionally be branched to. 998 BasicBlock *TheRealDest = nullptr; 999 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 1000 if (ThisCases[i].Value == TIV) { 1001 TheRealDest = ThisCases[i].Dest; 1002 break; 1003 } 1004 1005 // If not handled by any explicit cases, it is handled by the default case. 1006 if (!TheRealDest) 1007 TheRealDest = ThisDef; 1008 1009 SmallPtrSet<BasicBlock *, 2> RemovedSuccs; 1010 1011 // Remove PHI node entries for dead edges. 1012 BasicBlock *CheckEdge = TheRealDest; 1013 for (BasicBlock *Succ : successors(TIBB)) 1014 if (Succ != CheckEdge) { 1015 if (Succ != TheRealDest) 1016 RemovedSuccs.insert(Succ); 1017 Succ->removePredecessor(TIBB); 1018 } else 1019 CheckEdge = nullptr; 1020 1021 // Insert the new branch. 1022 Instruction *NI = Builder.CreateBr(TheRealDest); 1023 (void)NI; 1024 1025 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 1026 << "Through successor TI: " << *TI << "Leaving: " << *NI 1027 << "\n"); 1028 1029 EraseTerminatorAndDCECond(TI); 1030 if (DTU) { 1031 SmallVector<DominatorTree::UpdateType, 2> Updates; 1032 Updates.reserve(RemovedSuccs.size()); 1033 for (auto *RemovedSucc : RemovedSuccs) 1034 Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc}); 1035 DTU->applyUpdates(Updates); 1036 } 1037 return true; 1038 } 1039 1040 namespace { 1041 1042 /// This class implements a stable ordering of constant 1043 /// integers that does not depend on their address. This is important for 1044 /// applications that sort ConstantInt's to ensure uniqueness. 1045 struct ConstantIntOrdering { 1046 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 1047 return LHS->getValue().ult(RHS->getValue()); 1048 } 1049 }; 1050 1051 } // end anonymous namespace 1052 1053 static int ConstantIntSortPredicate(ConstantInt *const *P1, 1054 ConstantInt *const *P2) { 1055 const ConstantInt *LHS = *P1; 1056 const ConstantInt *RHS = *P2; 1057 if (LHS == RHS) 1058 return 0; 1059 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1; 1060 } 1061 1062 /// Get Weights of a given terminator, the default weight is at the front 1063 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight 1064 /// metadata. 1065 static void GetBranchWeights(Instruction *TI, 1066 SmallVectorImpl<uint64_t> &Weights) { 1067 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof); 1068 assert(MD); 1069 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 1070 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i)); 1071 Weights.push_back(CI->getValue().getZExtValue()); 1072 } 1073 1074 // If TI is a conditional eq, the default case is the false case, 1075 // and the corresponding branch-weight data is at index 2. We swap the 1076 // default weight to be the first entry. 1077 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1078 assert(Weights.size() == 2); 1079 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 1080 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1081 std::swap(Weights.front(), Weights.back()); 1082 } 1083 } 1084 1085 /// Keep halving the weights until all can fit in uint32_t. 1086 static void FitWeights(MutableArrayRef<uint64_t> Weights) { 1087 uint64_t Max = *std::max_element(Weights.begin(), Weights.end()); 1088 if (Max > UINT_MAX) { 1089 unsigned Offset = 32 - llvm::countl_zero(Max); 1090 for (uint64_t &I : Weights) 1091 I >>= Offset; 1092 } 1093 } 1094 1095 static void CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses( 1096 BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) { 1097 Instruction *PTI = PredBlock->getTerminator(); 1098 1099 // If we have bonus instructions, clone them into the predecessor block. 1100 // Note that there may be multiple predecessor blocks, so we cannot move 1101 // bonus instructions to a predecessor block. 1102 for (Instruction &BonusInst : *BB) { 1103 if (BonusInst.isTerminator()) 1104 continue; 1105 1106 Instruction *NewBonusInst = BonusInst.clone(); 1107 1108 if (!isa<DbgInfoIntrinsic>(BonusInst) && 1109 PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) { 1110 // Unless the instruction has the same !dbg location as the original 1111 // branch, drop it. When we fold the bonus instructions we want to make 1112 // sure we reset their debug locations in order to avoid stepping on 1113 // dead code caused by folding dead branches. 1114 NewBonusInst->setDebugLoc(DebugLoc()); 1115 } 1116 1117 RemapInstruction(NewBonusInst, VMap, 1118 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1119 1120 // If we speculated an instruction, we need to drop any metadata that may 1121 // result in undefined behavior, as the metadata might have been valid 1122 // only given the branch precondition. 1123 // Similarly strip attributes on call parameters that may cause UB in 1124 // location the call is moved to. 1125 NewBonusInst->dropUBImplyingAttrsAndMetadata(); 1126 1127 NewBonusInst->insertInto(PredBlock, PTI->getIterator()); 1128 auto Range = NewBonusInst->cloneDebugInfoFrom(&BonusInst); 1129 RemapDPValueRange(NewBonusInst->getModule(), Range, VMap, 1130 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1131 1132 if (isa<DbgInfoIntrinsic>(BonusInst)) 1133 continue; 1134 1135 NewBonusInst->takeName(&BonusInst); 1136 BonusInst.setName(NewBonusInst->getName() + ".old"); 1137 VMap[&BonusInst] = NewBonusInst; 1138 1139 // Update (liveout) uses of bonus instructions, 1140 // now that the bonus instruction has been cloned into predecessor. 1141 // Note that we expect to be in a block-closed SSA form for this to work! 1142 for (Use &U : make_early_inc_range(BonusInst.uses())) { 1143 auto *UI = cast<Instruction>(U.getUser()); 1144 auto *PN = dyn_cast<PHINode>(UI); 1145 if (!PN) { 1146 assert(UI->getParent() == BB && BonusInst.comesBefore(UI) && 1147 "If the user is not a PHI node, then it should be in the same " 1148 "block as, and come after, the original bonus instruction."); 1149 continue; // Keep using the original bonus instruction. 1150 } 1151 // Is this the block-closed SSA form PHI node? 1152 if (PN->getIncomingBlock(U) == BB) 1153 continue; // Great, keep using the original bonus instruction. 1154 // The only other alternative is an "use" when coming from 1155 // the predecessor block - here we should refer to the cloned bonus instr. 1156 assert(PN->getIncomingBlock(U) == PredBlock && 1157 "Not in block-closed SSA form?"); 1158 U.set(NewBonusInst); 1159 } 1160 } 1161 } 1162 1163 bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding( 1164 Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) { 1165 BasicBlock *BB = TI->getParent(); 1166 BasicBlock *Pred = PTI->getParent(); 1167 1168 SmallVector<DominatorTree::UpdateType, 32> Updates; 1169 1170 // Figure out which 'cases' to copy from SI to PSI. 1171 std::vector<ValueEqualityComparisonCase> BBCases; 1172 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 1173 1174 std::vector<ValueEqualityComparisonCase> PredCases; 1175 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 1176 1177 // Based on whether the default edge from PTI goes to BB or not, fill in 1178 // PredCases and PredDefault with the new switch cases we would like to 1179 // build. 1180 SmallMapVector<BasicBlock *, int, 8> NewSuccessors; 1181 1182 // Update the branch weight metadata along the way 1183 SmallVector<uint64_t, 8> Weights; 1184 bool PredHasWeights = hasBranchWeightMD(*PTI); 1185 bool SuccHasWeights = hasBranchWeightMD(*TI); 1186 1187 if (PredHasWeights) { 1188 GetBranchWeights(PTI, Weights); 1189 // branch-weight metadata is inconsistent here. 1190 if (Weights.size() != 1 + PredCases.size()) 1191 PredHasWeights = SuccHasWeights = false; 1192 } else if (SuccHasWeights) 1193 // If there are no predecessor weights but there are successor weights, 1194 // populate Weights with 1, which will later be scaled to the sum of 1195 // successor's weights 1196 Weights.assign(1 + PredCases.size(), 1); 1197 1198 SmallVector<uint64_t, 8> SuccWeights; 1199 if (SuccHasWeights) { 1200 GetBranchWeights(TI, SuccWeights); 1201 // branch-weight metadata is inconsistent here. 1202 if (SuccWeights.size() != 1 + BBCases.size()) 1203 PredHasWeights = SuccHasWeights = false; 1204 } else if (PredHasWeights) 1205 SuccWeights.assign(1 + BBCases.size(), 1); 1206 1207 if (PredDefault == BB) { 1208 // If this is the default destination from PTI, only the edges in TI 1209 // that don't occur in PTI, or that branch to BB will be activated. 1210 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled; 1211 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 1212 if (PredCases[i].Dest != BB) 1213 PTIHandled.insert(PredCases[i].Value); 1214 else { 1215 // The default destination is BB, we don't need explicit targets. 1216 std::swap(PredCases[i], PredCases.back()); 1217 1218 if (PredHasWeights || SuccHasWeights) { 1219 // Increase weight for the default case. 1220 Weights[0] += Weights[i + 1]; 1221 std::swap(Weights[i + 1], Weights.back()); 1222 Weights.pop_back(); 1223 } 1224 1225 PredCases.pop_back(); 1226 --i; 1227 --e; 1228 } 1229 1230 // Reconstruct the new switch statement we will be building. 1231 if (PredDefault != BBDefault) { 1232 PredDefault->removePredecessor(Pred); 1233 if (DTU && PredDefault != BB) 1234 Updates.push_back({DominatorTree::Delete, Pred, PredDefault}); 1235 PredDefault = BBDefault; 1236 ++NewSuccessors[BBDefault]; 1237 } 1238 1239 unsigned CasesFromPred = Weights.size(); 1240 uint64_t ValidTotalSuccWeight = 0; 1241 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 1242 if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) { 1243 PredCases.push_back(BBCases[i]); 1244 ++NewSuccessors[BBCases[i].Dest]; 1245 if (SuccHasWeights || PredHasWeights) { 1246 // The default weight is at index 0, so weight for the ith case 1247 // should be at index i+1. Scale the cases from successor by 1248 // PredDefaultWeight (Weights[0]). 1249 Weights.push_back(Weights[0] * SuccWeights[i + 1]); 1250 ValidTotalSuccWeight += SuccWeights[i + 1]; 1251 } 1252 } 1253 1254 if (SuccHasWeights || PredHasWeights) { 1255 ValidTotalSuccWeight += SuccWeights[0]; 1256 // Scale the cases from predecessor by ValidTotalSuccWeight. 1257 for (unsigned i = 1; i < CasesFromPred; ++i) 1258 Weights[i] *= ValidTotalSuccWeight; 1259 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 1260 Weights[0] *= SuccWeights[0]; 1261 } 1262 } else { 1263 // If this is not the default destination from PSI, only the edges 1264 // in SI that occur in PSI with a destination of BB will be 1265 // activated. 1266 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled; 1267 std::map<ConstantInt *, uint64_t> WeightsForHandled; 1268 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 1269 if (PredCases[i].Dest == BB) { 1270 PTIHandled.insert(PredCases[i].Value); 1271 1272 if (PredHasWeights || SuccHasWeights) { 1273 WeightsForHandled[PredCases[i].Value] = Weights[i + 1]; 1274 std::swap(Weights[i + 1], Weights.back()); 1275 Weights.pop_back(); 1276 } 1277 1278 std::swap(PredCases[i], PredCases.back()); 1279 PredCases.pop_back(); 1280 --i; 1281 --e; 1282 } 1283 1284 // Okay, now we know which constants were sent to BB from the 1285 // predecessor. Figure out where they will all go now. 1286 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 1287 if (PTIHandled.count(BBCases[i].Value)) { 1288 // If this is one we are capable of getting... 1289 if (PredHasWeights || SuccHasWeights) 1290 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 1291 PredCases.push_back(BBCases[i]); 1292 ++NewSuccessors[BBCases[i].Dest]; 1293 PTIHandled.erase(BBCases[i].Value); // This constant is taken care of 1294 } 1295 1296 // If there are any constants vectored to BB that TI doesn't handle, 1297 // they must go to the default destination of TI. 1298 for (ConstantInt *I : PTIHandled) { 1299 if (PredHasWeights || SuccHasWeights) 1300 Weights.push_back(WeightsForHandled[I]); 1301 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault)); 1302 ++NewSuccessors[BBDefault]; 1303 } 1304 } 1305 1306 // Okay, at this point, we know which new successor Pred will get. Make 1307 // sure we update the number of entries in the PHI nodes for these 1308 // successors. 1309 SmallPtrSet<BasicBlock *, 2> SuccsOfPred; 1310 if (DTU) { 1311 SuccsOfPred = {succ_begin(Pred), succ_end(Pred)}; 1312 Updates.reserve(Updates.size() + NewSuccessors.size()); 1313 } 1314 for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor : 1315 NewSuccessors) { 1316 for (auto I : seq(NewSuccessor.second)) { 1317 (void)I; 1318 AddPredecessorToBlock(NewSuccessor.first, Pred, BB); 1319 } 1320 if (DTU && !SuccsOfPred.contains(NewSuccessor.first)) 1321 Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first}); 1322 } 1323 1324 Builder.SetInsertPoint(PTI); 1325 // Convert pointer to int before we switch. 1326 if (CV->getType()->isPointerTy()) { 1327 CV = 1328 Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr"); 1329 } 1330 1331 // Now that the successors are updated, create the new Switch instruction. 1332 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size()); 1333 NewSI->setDebugLoc(PTI->getDebugLoc()); 1334 for (ValueEqualityComparisonCase &V : PredCases) 1335 NewSI->addCase(V.Value, V.Dest); 1336 1337 if (PredHasWeights || SuccHasWeights) { 1338 // Halve the weights if any of them cannot fit in an uint32_t 1339 FitWeights(Weights); 1340 1341 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1342 1343 setBranchWeights(NewSI, MDWeights); 1344 } 1345 1346 EraseTerminatorAndDCECond(PTI); 1347 1348 // Okay, last check. If BB is still a successor of PSI, then we must 1349 // have an infinite loop case. If so, add an infinitely looping block 1350 // to handle the case to preserve the behavior of the code. 1351 BasicBlock *InfLoopBlock = nullptr; 1352 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1353 if (NewSI->getSuccessor(i) == BB) { 1354 if (!InfLoopBlock) { 1355 // Insert it at the end of the function, because it's either code, 1356 // or it won't matter if it's hot. :) 1357 InfLoopBlock = 1358 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent()); 1359 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1360 if (DTU) 1361 Updates.push_back( 1362 {DominatorTree::Insert, InfLoopBlock, InfLoopBlock}); 1363 } 1364 NewSI->setSuccessor(i, InfLoopBlock); 1365 } 1366 1367 if (DTU) { 1368 if (InfLoopBlock) 1369 Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock}); 1370 1371 Updates.push_back({DominatorTree::Delete, Pred, BB}); 1372 1373 DTU->applyUpdates(Updates); 1374 } 1375 1376 ++NumFoldValueComparisonIntoPredecessors; 1377 return true; 1378 } 1379 1380 /// The specified terminator is a value equality comparison instruction 1381 /// (either a switch or a branch on "X == c"). 1382 /// See if any of the predecessors of the terminator block are value comparisons 1383 /// on the same value. If so, and if safe to do so, fold them together. 1384 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI, 1385 IRBuilder<> &Builder) { 1386 BasicBlock *BB = TI->getParent(); 1387 Value *CV = isValueEqualityComparison(TI); // CondVal 1388 assert(CV && "Not a comparison?"); 1389 1390 bool Changed = false; 1391 1392 SmallSetVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB)); 1393 while (!Preds.empty()) { 1394 BasicBlock *Pred = Preds.pop_back_val(); 1395 Instruction *PTI = Pred->getTerminator(); 1396 1397 // Don't try to fold into itself. 1398 if (Pred == BB) 1399 continue; 1400 1401 // See if the predecessor is a comparison with the same value. 1402 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 1403 if (PCV != CV) 1404 continue; 1405 1406 SmallSetVector<BasicBlock *, 4> FailBlocks; 1407 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) { 1408 for (auto *Succ : FailBlocks) { 1409 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU)) 1410 return false; 1411 } 1412 } 1413 1414 PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder); 1415 Changed = true; 1416 } 1417 return Changed; 1418 } 1419 1420 // If we would need to insert a select that uses the value of this invoke 1421 // (comments in hoistSuccIdenticalTerminatorToSwitchOrIf explain why we would 1422 // need to do this), we can't hoist the invoke, as there is nowhere to put the 1423 // select in this case. 1424 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1425 Instruction *I1, Instruction *I2) { 1426 for (BasicBlock *Succ : successors(BB1)) { 1427 for (const PHINode &PN : Succ->phis()) { 1428 Value *BB1V = PN.getIncomingValueForBlock(BB1); 1429 Value *BB2V = PN.getIncomingValueForBlock(BB2); 1430 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) { 1431 return false; 1432 } 1433 } 1434 } 1435 return true; 1436 } 1437 1438 // Get interesting characteristics of instructions that 1439 // `hoistCommonCodeFromSuccessors` didn't hoist. They restrict what kind of 1440 // instructions can be reordered across. 1441 enum SkipFlags { 1442 SkipReadMem = 1, 1443 SkipSideEffect = 2, 1444 SkipImplicitControlFlow = 4 1445 }; 1446 1447 static unsigned skippedInstrFlags(Instruction *I) { 1448 unsigned Flags = 0; 1449 if (I->mayReadFromMemory()) 1450 Flags |= SkipReadMem; 1451 // We can't arbitrarily move around allocas, e.g. moving allocas (especially 1452 // inalloca) across stacksave/stackrestore boundaries. 1453 if (I->mayHaveSideEffects() || isa<AllocaInst>(I)) 1454 Flags |= SkipSideEffect; 1455 if (!isGuaranteedToTransferExecutionToSuccessor(I)) 1456 Flags |= SkipImplicitControlFlow; 1457 return Flags; 1458 } 1459 1460 // Returns true if it is safe to reorder an instruction across preceding 1461 // instructions in a basic block. 1462 static bool isSafeToHoistInstr(Instruction *I, unsigned Flags) { 1463 // Don't reorder a store over a load. 1464 if ((Flags & SkipReadMem) && I->mayWriteToMemory()) 1465 return false; 1466 1467 // If we have seen an instruction with side effects, it's unsafe to reorder an 1468 // instruction which reads memory or itself has side effects. 1469 if ((Flags & SkipSideEffect) && 1470 (I->mayReadFromMemory() || I->mayHaveSideEffects() || isa<AllocaInst>(I))) 1471 return false; 1472 1473 // Reordering across an instruction which does not necessarily transfer 1474 // control to the next instruction is speculation. 1475 if ((Flags & SkipImplicitControlFlow) && !isSafeToSpeculativelyExecute(I)) 1476 return false; 1477 1478 // Hoisting of llvm.deoptimize is only legal together with the next return 1479 // instruction, which this pass is not always able to do. 1480 if (auto *CB = dyn_cast<CallBase>(I)) 1481 if (CB->getIntrinsicID() == Intrinsic::experimental_deoptimize) 1482 return false; 1483 1484 // It's also unsafe/illegal to hoist an instruction above its instruction 1485 // operands 1486 BasicBlock *BB = I->getParent(); 1487 for (Value *Op : I->operands()) { 1488 if (auto *J = dyn_cast<Instruction>(Op)) 1489 if (J->getParent() == BB) 1490 return false; 1491 } 1492 1493 return true; 1494 } 1495 1496 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false); 1497 1498 /// Helper function for hoistCommonCodeFromSuccessors. Return true if identical 1499 /// instructions \p I1 and \p I2 can and should be hoisted. 1500 static bool shouldHoistCommonInstructions(Instruction *I1, Instruction *I2, 1501 const TargetTransformInfo &TTI) { 1502 // If we're going to hoist a call, make sure that the two instructions 1503 // we're commoning/hoisting are both marked with musttail, or neither of 1504 // them is marked as such. Otherwise, we might end up in a situation where 1505 // we hoist from a block where the terminator is a `ret` to a block where 1506 // the terminator is a `br`, and `musttail` calls expect to be followed by 1507 // a return. 1508 auto *C1 = dyn_cast<CallInst>(I1); 1509 auto *C2 = dyn_cast<CallInst>(I2); 1510 if (C1 && C2) 1511 if (C1->isMustTailCall() != C2->isMustTailCall()) 1512 return false; 1513 1514 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2)) 1515 return false; 1516 1517 // If any of the two call sites has nomerge or convergent attribute, stop 1518 // hoisting. 1519 if (const auto *CB1 = dyn_cast<CallBase>(I1)) 1520 if (CB1->cannotMerge() || CB1->isConvergent()) 1521 return false; 1522 if (const auto *CB2 = dyn_cast<CallBase>(I2)) 1523 if (CB2->cannotMerge() || CB2->isConvergent()) 1524 return false; 1525 1526 return true; 1527 } 1528 1529 /// Hoist any common code in the successor blocks up into the block. This 1530 /// function guarantees that BB dominates all successors. If EqTermsOnly is 1531 /// given, only perform hoisting in case both blocks only contain a terminator. 1532 /// In that case, only the original BI will be replaced and selects for PHIs are 1533 /// added. 1534 bool SimplifyCFGOpt::hoistCommonCodeFromSuccessors(BasicBlock *BB, 1535 bool EqTermsOnly) { 1536 // This does very trivial matching, with limited scanning, to find identical 1537 // instructions in the two blocks. In particular, we don't want to get into 1538 // O(N1*N2*...) situations here where Ni are the sizes of these successors. As 1539 // such, we currently just scan for obviously identical instructions in an 1540 // identical order, possibly separated by the same number of non-identical 1541 // instructions. 1542 unsigned int SuccSize = succ_size(BB); 1543 if (SuccSize < 2) 1544 return false; 1545 1546 // If either of the blocks has it's address taken, then we can't do this fold, 1547 // because the code we'd hoist would no longer run when we jump into the block 1548 // by it's address. 1549 for (auto *Succ : successors(BB)) 1550 if (Succ->hasAddressTaken() || !Succ->getSinglePredecessor()) 1551 return false; 1552 1553 auto *TI = BB->getTerminator(); 1554 1555 // The second of pair is a SkipFlags bitmask. 1556 using SuccIterPair = std::pair<BasicBlock::iterator, unsigned>; 1557 SmallVector<SuccIterPair, 8> SuccIterPairs; 1558 for (auto *Succ : successors(BB)) { 1559 BasicBlock::iterator SuccItr = Succ->begin(); 1560 if (isa<PHINode>(*SuccItr)) 1561 return false; 1562 SuccIterPairs.push_back(SuccIterPair(SuccItr, 0)); 1563 } 1564 1565 // Check if only hoisting terminators is allowed. This does not add new 1566 // instructions to the hoist location. 1567 if (EqTermsOnly) { 1568 // Skip any debug intrinsics, as they are free to hoist. 1569 for (auto &SuccIter : make_first_range(SuccIterPairs)) { 1570 auto *INonDbg = &*skipDebugIntrinsics(SuccIter); 1571 if (!INonDbg->isTerminator()) 1572 return false; 1573 } 1574 // Now we know that we only need to hoist debug intrinsics and the 1575 // terminator. Let the loop below handle those 2 cases. 1576 } 1577 1578 // Count how many instructions were not hoisted so far. There's a limit on how 1579 // many instructions we skip, serving as a compilation time control as well as 1580 // preventing excessive increase of life ranges. 1581 unsigned NumSkipped = 0; 1582 // If we find an unreachable instruction at the beginning of a basic block, we 1583 // can still hoist instructions from the rest of the basic blocks. 1584 if (SuccIterPairs.size() > 2) { 1585 erase_if(SuccIterPairs, 1586 [](const auto &Pair) { return isa<UnreachableInst>(Pair.first); }); 1587 if (SuccIterPairs.size() < 2) 1588 return false; 1589 } 1590 1591 bool Changed = false; 1592 1593 for (;;) { 1594 auto *SuccIterPairBegin = SuccIterPairs.begin(); 1595 auto &BB1ItrPair = *SuccIterPairBegin++; 1596 auto OtherSuccIterPairRange = 1597 iterator_range(SuccIterPairBegin, SuccIterPairs.end()); 1598 auto OtherSuccIterRange = make_first_range(OtherSuccIterPairRange); 1599 1600 Instruction *I1 = &*BB1ItrPair.first; 1601 auto *BB1 = I1->getParent(); 1602 1603 // Skip debug info if it is not identical. 1604 bool AllDbgInstsAreIdentical = all_of(OtherSuccIterRange, [I1](auto &Iter) { 1605 Instruction *I2 = &*Iter; 1606 return I1->isIdenticalToWhenDefined(I2); 1607 }); 1608 if (!AllDbgInstsAreIdentical) { 1609 while (isa<DbgInfoIntrinsic>(I1)) 1610 I1 = &*++BB1ItrPair.first; 1611 for (auto &SuccIter : OtherSuccIterRange) { 1612 Instruction *I2 = &*SuccIter; 1613 while (isa<DbgInfoIntrinsic>(I2)) 1614 I2 = &*++SuccIter; 1615 } 1616 } 1617 1618 bool AllInstsAreIdentical = true; 1619 bool HasTerminator = I1->isTerminator(); 1620 for (auto &SuccIter : OtherSuccIterRange) { 1621 Instruction *I2 = &*SuccIter; 1622 HasTerminator |= I2->isTerminator(); 1623 if (AllInstsAreIdentical && !I1->isIdenticalToWhenDefined(I2)) 1624 AllInstsAreIdentical = false; 1625 } 1626 1627 // If we are hoisting the terminator instruction, don't move one (making a 1628 // broken BB), instead clone it, and remove BI. 1629 if (HasTerminator) { 1630 // Even if BB, which contains only one unreachable instruction, is ignored 1631 // at the beginning of the loop, we can hoist the terminator instruction. 1632 // If any instructions remain in the block, we cannot hoist terminators. 1633 if (NumSkipped || !AllInstsAreIdentical) 1634 return Changed; 1635 SmallVector<Instruction *, 8> Insts; 1636 for (auto &SuccIter : OtherSuccIterRange) 1637 Insts.push_back(&*SuccIter); 1638 return hoistSuccIdenticalTerminatorToSwitchOrIf(TI, I1, Insts) || Changed; 1639 } 1640 1641 if (AllInstsAreIdentical) { 1642 unsigned SkipFlagsBB1 = BB1ItrPair.second; 1643 AllInstsAreIdentical = 1644 isSafeToHoistInstr(I1, SkipFlagsBB1) && 1645 all_of(OtherSuccIterPairRange, [=](const auto &Pair) { 1646 Instruction *I2 = &*Pair.first; 1647 unsigned SkipFlagsBB2 = Pair.second; 1648 // Even if the instructions are identical, it may not 1649 // be safe to hoist them if we have skipped over 1650 // instructions with side effects or their operands 1651 // weren't hoisted. 1652 return isSafeToHoistInstr(I2, SkipFlagsBB2) && 1653 shouldHoistCommonInstructions(I1, I2, TTI); 1654 }); 1655 } 1656 1657 if (AllInstsAreIdentical) { 1658 BB1ItrPair.first++; 1659 if (isa<DbgInfoIntrinsic>(I1)) { 1660 // The debug location is an integral part of a debug info intrinsic 1661 // and can't be separated from it or replaced. Instead of attempting 1662 // to merge locations, simply hoist both copies of the intrinsic. 1663 I1->moveBeforePreserving(TI); 1664 for (auto &SuccIter : OtherSuccIterRange) { 1665 auto *I2 = &*SuccIter++; 1666 assert(isa<DbgInfoIntrinsic>(I2)); 1667 I2->moveBeforePreserving(TI); 1668 } 1669 } else { 1670 // For a normal instruction, we just move one to right before the 1671 // branch, then replace all uses of the other with the first. Finally, 1672 // we remove the now redundant second instruction. 1673 I1->moveBeforePreserving(TI); 1674 BB->splice(TI->getIterator(), BB1, I1->getIterator()); 1675 for (auto &SuccIter : OtherSuccIterRange) { 1676 Instruction *I2 = &*SuccIter++; 1677 assert(I2 != I1); 1678 if (!I2->use_empty()) 1679 I2->replaceAllUsesWith(I1); 1680 I1->andIRFlags(I2); 1681 combineMetadataForCSE(I1, I2, true); 1682 // I1 and I2 are being combined into a single instruction. Its debug 1683 // location is the merged locations of the original instructions. 1684 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()); 1685 I2->eraseFromParent(); 1686 } 1687 } 1688 if (!Changed) 1689 NumHoistCommonCode += SuccIterPairs.size(); 1690 Changed = true; 1691 NumHoistCommonInstrs += SuccIterPairs.size(); 1692 } else { 1693 if (NumSkipped >= HoistCommonSkipLimit) 1694 return Changed; 1695 // We are about to skip over a pair of non-identical instructions. Record 1696 // if any have characteristics that would prevent reordering instructions 1697 // across them. 1698 for (auto &SuccIterPair : SuccIterPairs) { 1699 Instruction *I = &*SuccIterPair.first++; 1700 SuccIterPair.second |= skippedInstrFlags(I); 1701 } 1702 ++NumSkipped; 1703 } 1704 } 1705 } 1706 1707 bool SimplifyCFGOpt::hoistSuccIdenticalTerminatorToSwitchOrIf( 1708 Instruction *TI, Instruction *I1, 1709 SmallVectorImpl<Instruction *> &OtherSuccTIs) { 1710 1711 auto *BI = dyn_cast<BranchInst>(TI); 1712 1713 bool Changed = false; 1714 BasicBlock *TIParent = TI->getParent(); 1715 BasicBlock *BB1 = I1->getParent(); 1716 1717 // Use only for an if statement. 1718 auto *I2 = *OtherSuccTIs.begin(); 1719 auto *BB2 = I2->getParent(); 1720 if (BI) { 1721 assert(OtherSuccTIs.size() == 1); 1722 assert(BI->getSuccessor(0) == I1->getParent()); 1723 assert(BI->getSuccessor(1) == I2->getParent()); 1724 } 1725 1726 // In the case of an if statement, we try to hoist an invoke. 1727 // FIXME: Can we define a safety predicate for CallBr? 1728 // FIXME: Test case llvm/test/Transforms/SimplifyCFG/2009-06-15-InvokeCrash.ll 1729 // removed in 4c923b3b3fd0ac1edebf0603265ca3ba51724937 commit? 1730 if (isa<InvokeInst>(I1) && (!BI || !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1731 return false; 1732 1733 // TODO: callbr hoisting currently disabled pending further study. 1734 if (isa<CallBrInst>(I1)) 1735 return false; 1736 1737 for (BasicBlock *Succ : successors(BB1)) { 1738 for (PHINode &PN : Succ->phis()) { 1739 Value *BB1V = PN.getIncomingValueForBlock(BB1); 1740 for (Instruction *OtherSuccTI : OtherSuccTIs) { 1741 Value *BB2V = PN.getIncomingValueForBlock(OtherSuccTI->getParent()); 1742 if (BB1V == BB2V) 1743 continue; 1744 1745 // In the case of an if statement, check for 1746 // passingValueIsAlwaysUndefined here because we would rather eliminate 1747 // undefined control flow then converting it to a select. 1748 if (!BI || passingValueIsAlwaysUndefined(BB1V, &PN) || 1749 passingValueIsAlwaysUndefined(BB2V, &PN)) 1750 return false; 1751 } 1752 } 1753 } 1754 1755 // Okay, it is safe to hoist the terminator. 1756 Instruction *NT = I1->clone(); 1757 NT->insertInto(TIParent, TI->getIterator()); 1758 if (!NT->getType()->isVoidTy()) { 1759 I1->replaceAllUsesWith(NT); 1760 for (Instruction *OtherSuccTI : OtherSuccTIs) 1761 OtherSuccTI->replaceAllUsesWith(NT); 1762 NT->takeName(I1); 1763 } 1764 Changed = true; 1765 NumHoistCommonInstrs += OtherSuccTIs.size() + 1; 1766 1767 // Ensure terminator gets a debug location, even an unknown one, in case 1768 // it involves inlinable calls. 1769 SmallVector<DILocation *, 4> Locs; 1770 Locs.push_back(I1->getDebugLoc()); 1771 for (auto *OtherSuccTI : OtherSuccTIs) 1772 Locs.push_back(OtherSuccTI->getDebugLoc()); 1773 NT->setDebugLoc(DILocation::getMergedLocations(Locs)); 1774 1775 // PHIs created below will adopt NT's merged DebugLoc. 1776 IRBuilder<NoFolder> Builder(NT); 1777 1778 // In the case of an if statement, hoisting one of the terminators from our 1779 // successor is a great thing. Unfortunately, the successors of the if/else 1780 // blocks may have PHI nodes in them. If they do, all PHI entries for BB1/BB2 1781 // must agree for all PHI nodes, so we insert select instruction to compute 1782 // the final result. 1783 if (BI) { 1784 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects; 1785 for (BasicBlock *Succ : successors(BB1)) { 1786 for (PHINode &PN : Succ->phis()) { 1787 Value *BB1V = PN.getIncomingValueForBlock(BB1); 1788 Value *BB2V = PN.getIncomingValueForBlock(BB2); 1789 if (BB1V == BB2V) 1790 continue; 1791 1792 // These values do not agree. Insert a select instruction before NT 1793 // that determines the right value. 1794 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1795 if (!SI) { 1796 // Propagate fast-math-flags from phi node to its replacement select. 1797 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 1798 if (isa<FPMathOperator>(PN)) 1799 Builder.setFastMathFlags(PN.getFastMathFlags()); 1800 1801 SI = cast<SelectInst>(Builder.CreateSelect( 1802 BI->getCondition(), BB1V, BB2V, 1803 BB1V->getName() + "." + BB2V->getName(), BI)); 1804 } 1805 1806 // Make the PHI node use the select for all incoming values for BB1/BB2 1807 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) 1808 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2) 1809 PN.setIncomingValue(i, SI); 1810 } 1811 } 1812 } 1813 1814 SmallVector<DominatorTree::UpdateType, 4> Updates; 1815 1816 // Update any PHI nodes in our new successors. 1817 for (BasicBlock *Succ : successors(BB1)) { 1818 AddPredecessorToBlock(Succ, TIParent, BB1); 1819 if (DTU) 1820 Updates.push_back({DominatorTree::Insert, TIParent, Succ}); 1821 } 1822 1823 if (DTU) 1824 for (BasicBlock *Succ : successors(TI)) 1825 Updates.push_back({DominatorTree::Delete, TIParent, Succ}); 1826 1827 EraseTerminatorAndDCECond(TI); 1828 if (DTU) 1829 DTU->applyUpdates(Updates); 1830 return Changed; 1831 } 1832 1833 // Check lifetime markers. 1834 static bool isLifeTimeMarker(const Instruction *I) { 1835 if (auto II = dyn_cast<IntrinsicInst>(I)) { 1836 switch (II->getIntrinsicID()) { 1837 default: 1838 break; 1839 case Intrinsic::lifetime_start: 1840 case Intrinsic::lifetime_end: 1841 return true; 1842 } 1843 } 1844 return false; 1845 } 1846 1847 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes 1848 // into variables. 1849 static bool replacingOperandWithVariableIsCheap(const Instruction *I, 1850 int OpIdx) { 1851 return !isa<IntrinsicInst>(I); 1852 } 1853 1854 // All instructions in Insts belong to different blocks that all unconditionally 1855 // branch to a common successor. Analyze each instruction and return true if it 1856 // would be possible to sink them into their successor, creating one common 1857 // instruction instead. For every value that would be required to be provided by 1858 // PHI node (because an operand varies in each input block), add to PHIOperands. 1859 static bool canSinkInstructions( 1860 ArrayRef<Instruction *> Insts, 1861 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) { 1862 // Prune out obviously bad instructions to move. Each instruction must have 1863 // exactly zero or one use, and we check later that use is by a single, common 1864 // PHI instruction in the successor. 1865 bool HasUse = !Insts.front()->user_empty(); 1866 for (auto *I : Insts) { 1867 // These instructions may change or break semantics if moved. 1868 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) || 1869 I->getType()->isTokenTy()) 1870 return false; 1871 1872 // Do not try to sink an instruction in an infinite loop - it can cause 1873 // this algorithm to infinite loop. 1874 if (I->getParent()->getSingleSuccessor() == I->getParent()) 1875 return false; 1876 1877 // Conservatively return false if I is an inline-asm instruction. Sinking 1878 // and merging inline-asm instructions can potentially create arguments 1879 // that cannot satisfy the inline-asm constraints. 1880 // If the instruction has nomerge or convergent attribute, return false. 1881 if (const auto *C = dyn_cast<CallBase>(I)) 1882 if (C->isInlineAsm() || C->cannotMerge() || C->isConvergent()) 1883 return false; 1884 1885 // Each instruction must have zero or one use. 1886 if (HasUse && !I->hasOneUse()) 1887 return false; 1888 if (!HasUse && !I->user_empty()) 1889 return false; 1890 } 1891 1892 const Instruction *I0 = Insts.front(); 1893 for (auto *I : Insts) { 1894 if (!I->isSameOperationAs(I0)) 1895 return false; 1896 1897 // swifterror pointers can only be used by a load or store; sinking a load 1898 // or store would require introducing a select for the pointer operand, 1899 // which isn't allowed for swifterror pointers. 1900 if (isa<StoreInst>(I) && I->getOperand(1)->isSwiftError()) 1901 return false; 1902 if (isa<LoadInst>(I) && I->getOperand(0)->isSwiftError()) 1903 return false; 1904 } 1905 1906 // All instructions in Insts are known to be the same opcode. If they have a 1907 // use, check that the only user is a PHI or in the same block as the 1908 // instruction, because if a user is in the same block as an instruction we're 1909 // contemplating sinking, it must already be determined to be sinkable. 1910 if (HasUse) { 1911 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin()); 1912 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0); 1913 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool { 1914 auto *U = cast<Instruction>(*I->user_begin()); 1915 return (PNUse && 1916 PNUse->getParent() == Succ && 1917 PNUse->getIncomingValueForBlock(I->getParent()) == I) || 1918 U->getParent() == I->getParent(); 1919 })) 1920 return false; 1921 } 1922 1923 // Because SROA can't handle speculating stores of selects, try not to sink 1924 // loads, stores or lifetime markers of allocas when we'd have to create a 1925 // PHI for the address operand. Also, because it is likely that loads or 1926 // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink 1927 // them. 1928 // This can cause code churn which can have unintended consequences down 1929 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244. 1930 // FIXME: This is a workaround for a deficiency in SROA - see 1931 // https://llvm.org/bugs/show_bug.cgi?id=30188 1932 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) { 1933 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts()); 1934 })) 1935 return false; 1936 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) { 1937 return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts()); 1938 })) 1939 return false; 1940 if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) { 1941 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts()); 1942 })) 1943 return false; 1944 1945 // For calls to be sinkable, they must all be indirect, or have same callee. 1946 // I.e. if we have two direct calls to different callees, we don't want to 1947 // turn that into an indirect call. Likewise, if we have an indirect call, 1948 // and a direct call, we don't actually want to have a single indirect call. 1949 if (isa<CallBase>(I0)) { 1950 auto IsIndirectCall = [](const Instruction *I) { 1951 return cast<CallBase>(I)->isIndirectCall(); 1952 }; 1953 bool HaveIndirectCalls = any_of(Insts, IsIndirectCall); 1954 bool AllCallsAreIndirect = all_of(Insts, IsIndirectCall); 1955 if (HaveIndirectCalls) { 1956 if (!AllCallsAreIndirect) 1957 return false; 1958 } else { 1959 // All callees must be identical. 1960 Value *Callee = nullptr; 1961 for (const Instruction *I : Insts) { 1962 Value *CurrCallee = cast<CallBase>(I)->getCalledOperand(); 1963 if (!Callee) 1964 Callee = CurrCallee; 1965 else if (Callee != CurrCallee) 1966 return false; 1967 } 1968 } 1969 } 1970 1971 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) { 1972 Value *Op = I0->getOperand(OI); 1973 if (Op->getType()->isTokenTy()) 1974 // Don't touch any operand of token type. 1975 return false; 1976 1977 auto SameAsI0 = [&I0, OI](const Instruction *I) { 1978 assert(I->getNumOperands() == I0->getNumOperands()); 1979 return I->getOperand(OI) == I0->getOperand(OI); 1980 }; 1981 if (!all_of(Insts, SameAsI0)) { 1982 if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) || 1983 !canReplaceOperandWithVariable(I0, OI)) 1984 // We can't create a PHI from this GEP. 1985 return false; 1986 for (auto *I : Insts) 1987 PHIOperands[I].push_back(I->getOperand(OI)); 1988 } 1989 } 1990 return true; 1991 } 1992 1993 // Assuming canSinkInstructions(Blocks) has returned true, sink the last 1994 // instruction of every block in Blocks to their common successor, commoning 1995 // into one instruction. 1996 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) { 1997 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0); 1998 1999 // canSinkInstructions returning true guarantees that every block has at 2000 // least one non-terminator instruction. 2001 SmallVector<Instruction*,4> Insts; 2002 for (auto *BB : Blocks) { 2003 Instruction *I = BB->getTerminator(); 2004 do { 2005 I = I->getPrevNode(); 2006 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front()); 2007 if (!isa<DbgInfoIntrinsic>(I)) 2008 Insts.push_back(I); 2009 } 2010 2011 // The only checking we need to do now is that all users of all instructions 2012 // are the same PHI node. canSinkInstructions should have checked this but 2013 // it is slightly over-aggressive - it gets confused by commutative 2014 // instructions so double-check it here. 2015 Instruction *I0 = Insts.front(); 2016 if (!I0->user_empty()) { 2017 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin()); 2018 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool { 2019 auto *U = cast<Instruction>(*I->user_begin()); 2020 return U == PNUse; 2021 })) 2022 return false; 2023 } 2024 2025 // We don't need to do any more checking here; canSinkInstructions should 2026 // have done it all for us. 2027 SmallVector<Value*, 4> NewOperands; 2028 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) { 2029 // This check is different to that in canSinkInstructions. There, we 2030 // cared about the global view once simplifycfg (and instcombine) have 2031 // completed - it takes into account PHIs that become trivially 2032 // simplifiable. However here we need a more local view; if an operand 2033 // differs we create a PHI and rely on instcombine to clean up the very 2034 // small mess we may make. 2035 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) { 2036 return I->getOperand(O) != I0->getOperand(O); 2037 }); 2038 if (!NeedPHI) { 2039 NewOperands.push_back(I0->getOperand(O)); 2040 continue; 2041 } 2042 2043 // Create a new PHI in the successor block and populate it. 2044 auto *Op = I0->getOperand(O); 2045 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!"); 2046 auto *PN = 2047 PHINode::Create(Op->getType(), Insts.size(), Op->getName() + ".sink"); 2048 PN->insertBefore(BBEnd->begin()); 2049 for (auto *I : Insts) 2050 PN->addIncoming(I->getOperand(O), I->getParent()); 2051 NewOperands.push_back(PN); 2052 } 2053 2054 // Arbitrarily use I0 as the new "common" instruction; remap its operands 2055 // and move it to the start of the successor block. 2056 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) 2057 I0->getOperandUse(O).set(NewOperands[O]); 2058 2059 I0->moveBefore(*BBEnd, BBEnd->getFirstInsertionPt()); 2060 2061 // Update metadata and IR flags, and merge debug locations. 2062 for (auto *I : Insts) 2063 if (I != I0) { 2064 // The debug location for the "common" instruction is the merged locations 2065 // of all the commoned instructions. We start with the original location 2066 // of the "common" instruction and iteratively merge each location in the 2067 // loop below. 2068 // This is an N-way merge, which will be inefficient if I0 is a CallInst. 2069 // However, as N-way merge for CallInst is rare, so we use simplified API 2070 // instead of using complex API for N-way merge. 2071 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc()); 2072 combineMetadataForCSE(I0, I, true); 2073 I0->andIRFlags(I); 2074 } 2075 2076 if (!I0->user_empty()) { 2077 // canSinkLastInstruction checked that all instructions were used by 2078 // one and only one PHI node. Find that now, RAUW it to our common 2079 // instruction and nuke it. 2080 auto *PN = cast<PHINode>(*I0->user_begin()); 2081 PN->replaceAllUsesWith(I0); 2082 PN->eraseFromParent(); 2083 } 2084 2085 // Finally nuke all instructions apart from the common instruction. 2086 for (auto *I : Insts) { 2087 if (I == I0) 2088 continue; 2089 // The remaining uses are debug users, replace those with the common inst. 2090 // In most (all?) cases this just introduces a use-before-def. 2091 assert(I->user_empty() && "Inst unexpectedly still has non-dbg users"); 2092 I->replaceAllUsesWith(I0); 2093 I->eraseFromParent(); 2094 } 2095 2096 return true; 2097 } 2098 2099 namespace { 2100 2101 // LockstepReverseIterator - Iterates through instructions 2102 // in a set of blocks in reverse order from the first non-terminator. 2103 // For example (assume all blocks have size n): 2104 // LockstepReverseIterator I([B1, B2, B3]); 2105 // *I-- = [B1[n], B2[n], B3[n]]; 2106 // *I-- = [B1[n-1], B2[n-1], B3[n-1]]; 2107 // *I-- = [B1[n-2], B2[n-2], B3[n-2]]; 2108 // ... 2109 class LockstepReverseIterator { 2110 ArrayRef<BasicBlock*> Blocks; 2111 SmallVector<Instruction*,4> Insts; 2112 bool Fail; 2113 2114 public: 2115 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) { 2116 reset(); 2117 } 2118 2119 void reset() { 2120 Fail = false; 2121 Insts.clear(); 2122 for (auto *BB : Blocks) { 2123 Instruction *Inst = BB->getTerminator(); 2124 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);) 2125 Inst = Inst->getPrevNode(); 2126 if (!Inst) { 2127 // Block wasn't big enough. 2128 Fail = true; 2129 return; 2130 } 2131 Insts.push_back(Inst); 2132 } 2133 } 2134 2135 bool isValid() const { 2136 return !Fail; 2137 } 2138 2139 void operator--() { 2140 if (Fail) 2141 return; 2142 for (auto *&Inst : Insts) { 2143 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);) 2144 Inst = Inst->getPrevNode(); 2145 // Already at beginning of block. 2146 if (!Inst) { 2147 Fail = true; 2148 return; 2149 } 2150 } 2151 } 2152 2153 void operator++() { 2154 if (Fail) 2155 return; 2156 for (auto *&Inst : Insts) { 2157 for (Inst = Inst->getNextNode(); Inst && isa<DbgInfoIntrinsic>(Inst);) 2158 Inst = Inst->getNextNode(); 2159 // Already at end of block. 2160 if (!Inst) { 2161 Fail = true; 2162 return; 2163 } 2164 } 2165 } 2166 2167 ArrayRef<Instruction*> operator * () const { 2168 return Insts; 2169 } 2170 }; 2171 2172 } // end anonymous namespace 2173 2174 /// Check whether BB's predecessors end with unconditional branches. If it is 2175 /// true, sink any common code from the predecessors to BB. 2176 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB, 2177 DomTreeUpdater *DTU) { 2178 // We support two situations: 2179 // (1) all incoming arcs are unconditional 2180 // (2) there are non-unconditional incoming arcs 2181 // 2182 // (2) is very common in switch defaults and 2183 // else-if patterns; 2184 // 2185 // if (a) f(1); 2186 // else if (b) f(2); 2187 // 2188 // produces: 2189 // 2190 // [if] 2191 // / \ 2192 // [f(1)] [if] 2193 // | | \ 2194 // | | | 2195 // | [f(2)]| 2196 // \ | / 2197 // [ end ] 2198 // 2199 // [end] has two unconditional predecessor arcs and one conditional. The 2200 // conditional refers to the implicit empty 'else' arc. This conditional 2201 // arc can also be caused by an empty default block in a switch. 2202 // 2203 // In this case, we attempt to sink code from all *unconditional* arcs. 2204 // If we can sink instructions from these arcs (determined during the scan 2205 // phase below) we insert a common successor for all unconditional arcs and 2206 // connect that to [end], to enable sinking: 2207 // 2208 // [if] 2209 // / \ 2210 // [x(1)] [if] 2211 // | | \ 2212 // | | \ 2213 // | [x(2)] | 2214 // \ / | 2215 // [sink.split] | 2216 // \ / 2217 // [ end ] 2218 // 2219 SmallVector<BasicBlock*,4> UnconditionalPreds; 2220 bool HaveNonUnconditionalPredecessors = false; 2221 for (auto *PredBB : predecessors(BB)) { 2222 auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()); 2223 if (PredBr && PredBr->isUnconditional()) 2224 UnconditionalPreds.push_back(PredBB); 2225 else 2226 HaveNonUnconditionalPredecessors = true; 2227 } 2228 if (UnconditionalPreds.size() < 2) 2229 return false; 2230 2231 // We take a two-step approach to tail sinking. First we scan from the end of 2232 // each block upwards in lockstep. If the n'th instruction from the end of each 2233 // block can be sunk, those instructions are added to ValuesToSink and we 2234 // carry on. If we can sink an instruction but need to PHI-merge some operands 2235 // (because they're not identical in each instruction) we add these to 2236 // PHIOperands. 2237 int ScanIdx = 0; 2238 SmallPtrSet<Value*,4> InstructionsToSink; 2239 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands; 2240 LockstepReverseIterator LRI(UnconditionalPreds); 2241 while (LRI.isValid() && 2242 canSinkInstructions(*LRI, PHIOperands)) { 2243 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] 2244 << "\n"); 2245 InstructionsToSink.insert((*LRI).begin(), (*LRI).end()); 2246 ++ScanIdx; 2247 --LRI; 2248 } 2249 2250 // If no instructions can be sunk, early-return. 2251 if (ScanIdx == 0) 2252 return false; 2253 2254 bool followedByDeoptOrUnreachable = IsBlockFollowedByDeoptOrUnreachable(BB); 2255 2256 if (!followedByDeoptOrUnreachable) { 2257 // Okay, we *could* sink last ScanIdx instructions. But how many can we 2258 // actually sink before encountering instruction that is unprofitable to 2259 // sink? 2260 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) { 2261 unsigned NumPHIdValues = 0; 2262 for (auto *I : *LRI) 2263 for (auto *V : PHIOperands[I]) { 2264 if (!InstructionsToSink.contains(V)) 2265 ++NumPHIdValues; 2266 // FIXME: this check is overly optimistic. We may end up not sinking 2267 // said instruction, due to the very same profitability check. 2268 // See @creating_too_many_phis in sink-common-code.ll. 2269 } 2270 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n"); 2271 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size(); 2272 if ((NumPHIdValues % UnconditionalPreds.size()) != 0) 2273 NumPHIInsts++; 2274 2275 return NumPHIInsts <= 1; 2276 }; 2277 2278 // We've determined that we are going to sink last ScanIdx instructions, 2279 // and recorded them in InstructionsToSink. Now, some instructions may be 2280 // unprofitable to sink. But that determination depends on the instructions 2281 // that we are going to sink. 2282 2283 // First, forward scan: find the first instruction unprofitable to sink, 2284 // recording all the ones that are profitable to sink. 2285 // FIXME: would it be better, after we detect that not all are profitable. 2286 // to either record the profitable ones, or erase the unprofitable ones? 2287 // Maybe we need to choose (at runtime) the one that will touch least 2288 // instrs? 2289 LRI.reset(); 2290 int Idx = 0; 2291 SmallPtrSet<Value *, 4> InstructionsProfitableToSink; 2292 while (Idx < ScanIdx) { 2293 if (!ProfitableToSinkInstruction(LRI)) { 2294 // Too many PHIs would be created. 2295 LLVM_DEBUG( 2296 dbgs() << "SINK: stopping here, too many PHIs would be created!\n"); 2297 break; 2298 } 2299 InstructionsProfitableToSink.insert((*LRI).begin(), (*LRI).end()); 2300 --LRI; 2301 ++Idx; 2302 } 2303 2304 // If no instructions can be sunk, early-return. 2305 if (Idx == 0) 2306 return false; 2307 2308 // Did we determine that (only) some instructions are unprofitable to sink? 2309 if (Idx < ScanIdx) { 2310 // Okay, some instructions are unprofitable. 2311 ScanIdx = Idx; 2312 InstructionsToSink = InstructionsProfitableToSink; 2313 2314 // But, that may make other instructions unprofitable, too. 2315 // So, do a backward scan, do any earlier instructions become 2316 // unprofitable? 2317 assert( 2318 !ProfitableToSinkInstruction(LRI) && 2319 "We already know that the last instruction is unprofitable to sink"); 2320 ++LRI; 2321 --Idx; 2322 while (Idx >= 0) { 2323 // If we detect that an instruction becomes unprofitable to sink, 2324 // all earlier instructions won't be sunk either, 2325 // so preemptively keep InstructionsProfitableToSink in sync. 2326 // FIXME: is this the most performant approach? 2327 for (auto *I : *LRI) 2328 InstructionsProfitableToSink.erase(I); 2329 if (!ProfitableToSinkInstruction(LRI)) { 2330 // Everything starting with this instruction won't be sunk. 2331 ScanIdx = Idx; 2332 InstructionsToSink = InstructionsProfitableToSink; 2333 } 2334 ++LRI; 2335 --Idx; 2336 } 2337 } 2338 2339 // If no instructions can be sunk, early-return. 2340 if (ScanIdx == 0) 2341 return false; 2342 } 2343 2344 bool Changed = false; 2345 2346 if (HaveNonUnconditionalPredecessors) { 2347 if (!followedByDeoptOrUnreachable) { 2348 // It is always legal to sink common instructions from unconditional 2349 // predecessors. However, if not all predecessors are unconditional, 2350 // this transformation might be pessimizing. So as a rule of thumb, 2351 // don't do it unless we'd sink at least one non-speculatable instruction. 2352 // See https://bugs.llvm.org/show_bug.cgi?id=30244 2353 LRI.reset(); 2354 int Idx = 0; 2355 bool Profitable = false; 2356 while (Idx < ScanIdx) { 2357 if (!isSafeToSpeculativelyExecute((*LRI)[0])) { 2358 Profitable = true; 2359 break; 2360 } 2361 --LRI; 2362 ++Idx; 2363 } 2364 if (!Profitable) 2365 return false; 2366 } 2367 2368 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n"); 2369 // We have a conditional edge and we're going to sink some instructions. 2370 // Insert a new block postdominating all blocks we're going to sink from. 2371 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU)) 2372 // Edges couldn't be split. 2373 return false; 2374 Changed = true; 2375 } 2376 2377 // Now that we've analyzed all potential sinking candidates, perform the 2378 // actual sink. We iteratively sink the last non-terminator of the source 2379 // blocks into their common successor unless doing so would require too 2380 // many PHI instructions to be generated (currently only one PHI is allowed 2381 // per sunk instruction). 2382 // 2383 // We can use InstructionsToSink to discount values needing PHI-merging that will 2384 // actually be sunk in a later iteration. This allows us to be more 2385 // aggressive in what we sink. This does allow a false positive where we 2386 // sink presuming a later value will also be sunk, but stop half way through 2387 // and never actually sink it which means we produce more PHIs than intended. 2388 // This is unlikely in practice though. 2389 int SinkIdx = 0; 2390 for (; SinkIdx != ScanIdx; ++SinkIdx) { 2391 LLVM_DEBUG(dbgs() << "SINK: Sink: " 2392 << *UnconditionalPreds[0]->getTerminator()->getPrevNode() 2393 << "\n"); 2394 2395 // Because we've sunk every instruction in turn, the current instruction to 2396 // sink is always at index 0. 2397 LRI.reset(); 2398 2399 if (!sinkLastInstruction(UnconditionalPreds)) { 2400 LLVM_DEBUG( 2401 dbgs() 2402 << "SINK: stopping here, failed to actually sink instruction!\n"); 2403 break; 2404 } 2405 2406 NumSinkCommonInstrs++; 2407 Changed = true; 2408 } 2409 if (SinkIdx != 0) 2410 ++NumSinkCommonCode; 2411 return Changed; 2412 } 2413 2414 namespace { 2415 2416 struct CompatibleSets { 2417 using SetTy = SmallVector<InvokeInst *, 2>; 2418 2419 SmallVector<SetTy, 1> Sets; 2420 2421 static bool shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes); 2422 2423 SetTy &getCompatibleSet(InvokeInst *II); 2424 2425 void insert(InvokeInst *II); 2426 }; 2427 2428 CompatibleSets::SetTy &CompatibleSets::getCompatibleSet(InvokeInst *II) { 2429 // Perform a linear scan over all the existing sets, see if the new `invoke` 2430 // is compatible with any particular set. Since we know that all the `invokes` 2431 // within a set are compatible, only check the first `invoke` in each set. 2432 // WARNING: at worst, this has quadratic complexity. 2433 for (CompatibleSets::SetTy &Set : Sets) { 2434 if (CompatibleSets::shouldBelongToSameSet({Set.front(), II})) 2435 return Set; 2436 } 2437 2438 // Otherwise, we either had no sets yet, or this invoke forms a new set. 2439 return Sets.emplace_back(); 2440 } 2441 2442 void CompatibleSets::insert(InvokeInst *II) { 2443 getCompatibleSet(II).emplace_back(II); 2444 } 2445 2446 bool CompatibleSets::shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes) { 2447 assert(Invokes.size() == 2 && "Always called with exactly two candidates."); 2448 2449 // Can we theoretically merge these `invoke`s? 2450 auto IsIllegalToMerge = [](InvokeInst *II) { 2451 return II->cannotMerge() || II->isInlineAsm(); 2452 }; 2453 if (any_of(Invokes, IsIllegalToMerge)) 2454 return false; 2455 2456 // Either both `invoke`s must be direct, 2457 // or both `invoke`s must be indirect. 2458 auto IsIndirectCall = [](InvokeInst *II) { return II->isIndirectCall(); }; 2459 bool HaveIndirectCalls = any_of(Invokes, IsIndirectCall); 2460 bool AllCallsAreIndirect = all_of(Invokes, IsIndirectCall); 2461 if (HaveIndirectCalls) { 2462 if (!AllCallsAreIndirect) 2463 return false; 2464 } else { 2465 // All callees must be identical. 2466 Value *Callee = nullptr; 2467 for (InvokeInst *II : Invokes) { 2468 Value *CurrCallee = II->getCalledOperand(); 2469 assert(CurrCallee && "There is always a called operand."); 2470 if (!Callee) 2471 Callee = CurrCallee; 2472 else if (Callee != CurrCallee) 2473 return false; 2474 } 2475 } 2476 2477 // Either both `invoke`s must not have a normal destination, 2478 // or both `invoke`s must have a normal destination, 2479 auto HasNormalDest = [](InvokeInst *II) { 2480 return !isa<UnreachableInst>(II->getNormalDest()->getFirstNonPHIOrDbg()); 2481 }; 2482 if (any_of(Invokes, HasNormalDest)) { 2483 // Do not merge `invoke` that does not have a normal destination with one 2484 // that does have a normal destination, even though doing so would be legal. 2485 if (!all_of(Invokes, HasNormalDest)) 2486 return false; 2487 2488 // All normal destinations must be identical. 2489 BasicBlock *NormalBB = nullptr; 2490 for (InvokeInst *II : Invokes) { 2491 BasicBlock *CurrNormalBB = II->getNormalDest(); 2492 assert(CurrNormalBB && "There is always a 'continue to' basic block."); 2493 if (!NormalBB) 2494 NormalBB = CurrNormalBB; 2495 else if (NormalBB != CurrNormalBB) 2496 return false; 2497 } 2498 2499 // In the normal destination, the incoming values for these two `invoke`s 2500 // must be compatible. 2501 SmallPtrSet<Value *, 16> EquivalenceSet(Invokes.begin(), Invokes.end()); 2502 if (!IncomingValuesAreCompatible( 2503 NormalBB, {Invokes[0]->getParent(), Invokes[1]->getParent()}, 2504 &EquivalenceSet)) 2505 return false; 2506 } 2507 2508 #ifndef NDEBUG 2509 // All unwind destinations must be identical. 2510 // We know that because we have started from said unwind destination. 2511 BasicBlock *UnwindBB = nullptr; 2512 for (InvokeInst *II : Invokes) { 2513 BasicBlock *CurrUnwindBB = II->getUnwindDest(); 2514 assert(CurrUnwindBB && "There is always an 'unwind to' basic block."); 2515 if (!UnwindBB) 2516 UnwindBB = CurrUnwindBB; 2517 else 2518 assert(UnwindBB == CurrUnwindBB && "Unexpected unwind destination."); 2519 } 2520 #endif 2521 2522 // In the unwind destination, the incoming values for these two `invoke`s 2523 // must be compatible. 2524 if (!IncomingValuesAreCompatible( 2525 Invokes.front()->getUnwindDest(), 2526 {Invokes[0]->getParent(), Invokes[1]->getParent()})) 2527 return false; 2528 2529 // Ignoring arguments, these `invoke`s must be identical, 2530 // including operand bundles. 2531 const InvokeInst *II0 = Invokes.front(); 2532 for (auto *II : Invokes.drop_front()) 2533 if (!II->isSameOperationAs(II0)) 2534 return false; 2535 2536 // Can we theoretically form the data operands for the merged `invoke`? 2537 auto IsIllegalToMergeArguments = [](auto Ops) { 2538 Use &U0 = std::get<0>(Ops); 2539 Use &U1 = std::get<1>(Ops); 2540 if (U0 == U1) 2541 return false; 2542 return U0->getType()->isTokenTy() || 2543 !canReplaceOperandWithVariable(cast<Instruction>(U0.getUser()), 2544 U0.getOperandNo()); 2545 }; 2546 assert(Invokes.size() == 2 && "Always called with exactly two candidates."); 2547 if (any_of(zip(Invokes[0]->data_ops(), Invokes[1]->data_ops()), 2548 IsIllegalToMergeArguments)) 2549 return false; 2550 2551 return true; 2552 } 2553 2554 } // namespace 2555 2556 // Merge all invokes in the provided set, all of which are compatible 2557 // as per the `CompatibleSets::shouldBelongToSameSet()`. 2558 static void MergeCompatibleInvokesImpl(ArrayRef<InvokeInst *> Invokes, 2559 DomTreeUpdater *DTU) { 2560 assert(Invokes.size() >= 2 && "Must have at least two invokes to merge."); 2561 2562 SmallVector<DominatorTree::UpdateType, 8> Updates; 2563 if (DTU) 2564 Updates.reserve(2 + 3 * Invokes.size()); 2565 2566 bool HasNormalDest = 2567 !isa<UnreachableInst>(Invokes[0]->getNormalDest()->getFirstNonPHIOrDbg()); 2568 2569 // Clone one of the invokes into a new basic block. 2570 // Since they are all compatible, it doesn't matter which invoke is cloned. 2571 InvokeInst *MergedInvoke = [&Invokes, HasNormalDest]() { 2572 InvokeInst *II0 = Invokes.front(); 2573 BasicBlock *II0BB = II0->getParent(); 2574 BasicBlock *InsertBeforeBlock = 2575 II0->getParent()->getIterator()->getNextNode(); 2576 Function *Func = II0BB->getParent(); 2577 LLVMContext &Ctx = II0->getContext(); 2578 2579 BasicBlock *MergedInvokeBB = BasicBlock::Create( 2580 Ctx, II0BB->getName() + ".invoke", Func, InsertBeforeBlock); 2581 2582 auto *MergedInvoke = cast<InvokeInst>(II0->clone()); 2583 // NOTE: all invokes have the same attributes, so no handling needed. 2584 MergedInvoke->insertInto(MergedInvokeBB, MergedInvokeBB->end()); 2585 2586 if (!HasNormalDest) { 2587 // This set does not have a normal destination, 2588 // so just form a new block with unreachable terminator. 2589 BasicBlock *MergedNormalDest = BasicBlock::Create( 2590 Ctx, II0BB->getName() + ".cont", Func, InsertBeforeBlock); 2591 new UnreachableInst(Ctx, MergedNormalDest); 2592 MergedInvoke->setNormalDest(MergedNormalDest); 2593 } 2594 2595 // The unwind destination, however, remainds identical for all invokes here. 2596 2597 return MergedInvoke; 2598 }(); 2599 2600 if (DTU) { 2601 // Predecessor blocks that contained these invokes will now branch to 2602 // the new block that contains the merged invoke, ... 2603 for (InvokeInst *II : Invokes) 2604 Updates.push_back( 2605 {DominatorTree::Insert, II->getParent(), MergedInvoke->getParent()}); 2606 2607 // ... which has the new `unreachable` block as normal destination, 2608 // or unwinds to the (same for all `invoke`s in this set) `landingpad`, 2609 for (BasicBlock *SuccBBOfMergedInvoke : successors(MergedInvoke)) 2610 Updates.push_back({DominatorTree::Insert, MergedInvoke->getParent(), 2611 SuccBBOfMergedInvoke}); 2612 2613 // Since predecessor blocks now unconditionally branch to a new block, 2614 // they no longer branch to their original successors. 2615 for (InvokeInst *II : Invokes) 2616 for (BasicBlock *SuccOfPredBB : successors(II->getParent())) 2617 Updates.push_back( 2618 {DominatorTree::Delete, II->getParent(), SuccOfPredBB}); 2619 } 2620 2621 bool IsIndirectCall = Invokes[0]->isIndirectCall(); 2622 2623 // Form the merged operands for the merged invoke. 2624 for (Use &U : MergedInvoke->operands()) { 2625 // Only PHI together the indirect callees and data operands. 2626 if (MergedInvoke->isCallee(&U)) { 2627 if (!IsIndirectCall) 2628 continue; 2629 } else if (!MergedInvoke->isDataOperand(&U)) 2630 continue; 2631 2632 // Don't create trivial PHI's with all-identical incoming values. 2633 bool NeedPHI = any_of(Invokes, [&U](InvokeInst *II) { 2634 return II->getOperand(U.getOperandNo()) != U.get(); 2635 }); 2636 if (!NeedPHI) 2637 continue; 2638 2639 // Form a PHI out of all the data ops under this index. 2640 PHINode *PN = PHINode::Create( 2641 U->getType(), /*NumReservedValues=*/Invokes.size(), "", MergedInvoke); 2642 for (InvokeInst *II : Invokes) 2643 PN->addIncoming(II->getOperand(U.getOperandNo()), II->getParent()); 2644 2645 U.set(PN); 2646 } 2647 2648 // We've ensured that each PHI node has compatible (identical) incoming values 2649 // when coming from each of the `invoke`s in the current merge set, 2650 // so update the PHI nodes accordingly. 2651 for (BasicBlock *Succ : successors(MergedInvoke)) 2652 AddPredecessorToBlock(Succ, /*NewPred=*/MergedInvoke->getParent(), 2653 /*ExistPred=*/Invokes.front()->getParent()); 2654 2655 // And finally, replace the original `invoke`s with an unconditional branch 2656 // to the block with the merged `invoke`. Also, give that merged `invoke` 2657 // the merged debugloc of all the original `invoke`s. 2658 DILocation *MergedDebugLoc = nullptr; 2659 for (InvokeInst *II : Invokes) { 2660 // Compute the debug location common to all the original `invoke`s. 2661 if (!MergedDebugLoc) 2662 MergedDebugLoc = II->getDebugLoc(); 2663 else 2664 MergedDebugLoc = 2665 DILocation::getMergedLocation(MergedDebugLoc, II->getDebugLoc()); 2666 2667 // And replace the old `invoke` with an unconditionally branch 2668 // to the block with the merged `invoke`. 2669 for (BasicBlock *OrigSuccBB : successors(II->getParent())) 2670 OrigSuccBB->removePredecessor(II->getParent()); 2671 BranchInst::Create(MergedInvoke->getParent(), II->getParent()); 2672 II->replaceAllUsesWith(MergedInvoke); 2673 II->eraseFromParent(); 2674 ++NumInvokesMerged; 2675 } 2676 MergedInvoke->setDebugLoc(MergedDebugLoc); 2677 ++NumInvokeSetsFormed; 2678 2679 if (DTU) 2680 DTU->applyUpdates(Updates); 2681 } 2682 2683 /// If this block is a `landingpad` exception handling block, categorize all 2684 /// the predecessor `invoke`s into sets, with all `invoke`s in each set 2685 /// being "mergeable" together, and then merge invokes in each set together. 2686 /// 2687 /// This is a weird mix of hoisting and sinking. Visually, it goes from: 2688 /// [...] [...] 2689 /// | | 2690 /// [invoke0] [invoke1] 2691 /// / \ / \ 2692 /// [cont0] [landingpad] [cont1] 2693 /// to: 2694 /// [...] [...] 2695 /// \ / 2696 /// [invoke] 2697 /// / \ 2698 /// [cont] [landingpad] 2699 /// 2700 /// But of course we can only do that if the invokes share the `landingpad`, 2701 /// edges invoke0->cont0 and invoke1->cont1 are "compatible", 2702 /// and the invoked functions are "compatible". 2703 static bool MergeCompatibleInvokes(BasicBlock *BB, DomTreeUpdater *DTU) { 2704 if (!EnableMergeCompatibleInvokes) 2705 return false; 2706 2707 bool Changed = false; 2708 2709 // FIXME: generalize to all exception handling blocks? 2710 if (!BB->isLandingPad()) 2711 return Changed; 2712 2713 CompatibleSets Grouper; 2714 2715 // Record all the predecessors of this `landingpad`. As per verifier, 2716 // the only allowed predecessor is the unwind edge of an `invoke`. 2717 // We want to group "compatible" `invokes` into the same set to be merged. 2718 for (BasicBlock *PredBB : predecessors(BB)) 2719 Grouper.insert(cast<InvokeInst>(PredBB->getTerminator())); 2720 2721 // And now, merge `invoke`s that were grouped togeter. 2722 for (ArrayRef<InvokeInst *> Invokes : Grouper.Sets) { 2723 if (Invokes.size() < 2) 2724 continue; 2725 Changed = true; 2726 MergeCompatibleInvokesImpl(Invokes, DTU); 2727 } 2728 2729 return Changed; 2730 } 2731 2732 namespace { 2733 /// Track ephemeral values, which should be ignored for cost-modelling 2734 /// purposes. Requires walking instructions in reverse order. 2735 class EphemeralValueTracker { 2736 SmallPtrSet<const Instruction *, 32> EphValues; 2737 2738 bool isEphemeral(const Instruction *I) { 2739 if (isa<AssumeInst>(I)) 2740 return true; 2741 return !I->mayHaveSideEffects() && !I->isTerminator() && 2742 all_of(I->users(), [&](const User *U) { 2743 return EphValues.count(cast<Instruction>(U)); 2744 }); 2745 } 2746 2747 public: 2748 bool track(const Instruction *I) { 2749 if (isEphemeral(I)) { 2750 EphValues.insert(I); 2751 return true; 2752 } 2753 return false; 2754 } 2755 2756 bool contains(const Instruction *I) const { return EphValues.contains(I); } 2757 }; 2758 } // namespace 2759 2760 /// Determine if we can hoist sink a sole store instruction out of a 2761 /// conditional block. 2762 /// 2763 /// We are looking for code like the following: 2764 /// BrBB: 2765 /// store i32 %add, i32* %arrayidx2 2766 /// ... // No other stores or function calls (we could be calling a memory 2767 /// ... // function). 2768 /// %cmp = icmp ult %x, %y 2769 /// br i1 %cmp, label %EndBB, label %ThenBB 2770 /// ThenBB: 2771 /// store i32 %add5, i32* %arrayidx2 2772 /// br label EndBB 2773 /// EndBB: 2774 /// ... 2775 /// We are going to transform this into: 2776 /// BrBB: 2777 /// store i32 %add, i32* %arrayidx2 2778 /// ... // 2779 /// %cmp = icmp ult %x, %y 2780 /// %add.add5 = select i1 %cmp, i32 %add, %add5 2781 /// store i32 %add.add5, i32* %arrayidx2 2782 /// ... 2783 /// 2784 /// \return The pointer to the value of the previous store if the store can be 2785 /// hoisted into the predecessor block. 0 otherwise. 2786 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, 2787 BasicBlock *StoreBB, BasicBlock *EndBB) { 2788 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); 2789 if (!StoreToHoist) 2790 return nullptr; 2791 2792 // Volatile or atomic. 2793 if (!StoreToHoist->isSimple()) 2794 return nullptr; 2795 2796 Value *StorePtr = StoreToHoist->getPointerOperand(); 2797 Type *StoreTy = StoreToHoist->getValueOperand()->getType(); 2798 2799 // Look for a store to the same pointer in BrBB. 2800 unsigned MaxNumInstToLookAt = 9; 2801 // Skip pseudo probe intrinsic calls which are not really killing any memory 2802 // accesses. 2803 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) { 2804 if (!MaxNumInstToLookAt) 2805 break; 2806 --MaxNumInstToLookAt; 2807 2808 // Could be calling an instruction that affects memory like free(). 2809 if (CurI.mayWriteToMemory() && !isa<StoreInst>(CurI)) 2810 return nullptr; 2811 2812 if (auto *SI = dyn_cast<StoreInst>(&CurI)) { 2813 // Found the previous store to same location and type. Make sure it is 2814 // simple, to avoid introducing a spurious non-atomic write after an 2815 // atomic write. 2816 if (SI->getPointerOperand() == StorePtr && 2817 SI->getValueOperand()->getType() == StoreTy && SI->isSimple()) 2818 // Found the previous store, return its value operand. 2819 return SI->getValueOperand(); 2820 return nullptr; // Unknown store. 2821 } 2822 2823 if (auto *LI = dyn_cast<LoadInst>(&CurI)) { 2824 if (LI->getPointerOperand() == StorePtr && LI->getType() == StoreTy && 2825 LI->isSimple()) { 2826 // Local objects (created by an `alloca` instruction) are always 2827 // writable, so once we are past a read from a location it is valid to 2828 // also write to that same location. 2829 // If the address of the local object never escapes the function, that 2830 // means it's never concurrently read or written, hence moving the store 2831 // from under the condition will not introduce a data race. 2832 auto *AI = dyn_cast<AllocaInst>(getUnderlyingObject(StorePtr)); 2833 if (AI && !PointerMayBeCaptured(AI, false, true)) 2834 // Found a previous load, return it. 2835 return LI; 2836 } 2837 // The load didn't work out, but we may still find a store. 2838 } 2839 } 2840 2841 return nullptr; 2842 } 2843 2844 /// Estimate the cost of the insertion(s) and check that the PHI nodes can be 2845 /// converted to selects. 2846 static bool validateAndCostRequiredSelects(BasicBlock *BB, BasicBlock *ThenBB, 2847 BasicBlock *EndBB, 2848 unsigned &SpeculatedInstructions, 2849 InstructionCost &Cost, 2850 const TargetTransformInfo &TTI) { 2851 TargetTransformInfo::TargetCostKind CostKind = 2852 BB->getParent()->hasMinSize() 2853 ? TargetTransformInfo::TCK_CodeSize 2854 : TargetTransformInfo::TCK_SizeAndLatency; 2855 2856 bool HaveRewritablePHIs = false; 2857 for (PHINode &PN : EndBB->phis()) { 2858 Value *OrigV = PN.getIncomingValueForBlock(BB); 2859 Value *ThenV = PN.getIncomingValueForBlock(ThenBB); 2860 2861 // FIXME: Try to remove some of the duplication with 2862 // hoistCommonCodeFromSuccessors. Skip PHIs which are trivial. 2863 if (ThenV == OrigV) 2864 continue; 2865 2866 Cost += TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr, 2867 CmpInst::BAD_ICMP_PREDICATE, CostKind); 2868 2869 // Don't convert to selects if we could remove undefined behavior instead. 2870 if (passingValueIsAlwaysUndefined(OrigV, &PN) || 2871 passingValueIsAlwaysUndefined(ThenV, &PN)) 2872 return false; 2873 2874 HaveRewritablePHIs = true; 2875 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV); 2876 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV); 2877 if (!OrigCE && !ThenCE) 2878 continue; // Known cheap (FIXME: Maybe not true for aggregates). 2879 2880 InstructionCost OrigCost = OrigCE ? computeSpeculationCost(OrigCE, TTI) : 0; 2881 InstructionCost ThenCost = ThenCE ? computeSpeculationCost(ThenCE, TTI) : 0; 2882 InstructionCost MaxCost = 2883 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; 2884 if (OrigCost + ThenCost > MaxCost) 2885 return false; 2886 2887 // Account for the cost of an unfolded ConstantExpr which could end up 2888 // getting expanded into Instructions. 2889 // FIXME: This doesn't account for how many operations are combined in the 2890 // constant expression. 2891 ++SpeculatedInstructions; 2892 if (SpeculatedInstructions > 1) 2893 return false; 2894 } 2895 2896 return HaveRewritablePHIs; 2897 } 2898 2899 /// Speculate a conditional basic block flattening the CFG. 2900 /// 2901 /// Note that this is a very risky transform currently. Speculating 2902 /// instructions like this is most often not desirable. Instead, there is an MI 2903 /// pass which can do it with full awareness of the resource constraints. 2904 /// However, some cases are "obvious" and we should do directly. An example of 2905 /// this is speculating a single, reasonably cheap instruction. 2906 /// 2907 /// There is only one distinct advantage to flattening the CFG at the IR level: 2908 /// it makes very common but simplistic optimizations such as are common in 2909 /// instcombine and the DAG combiner more powerful by removing CFG edges and 2910 /// modeling their effects with easier to reason about SSA value graphs. 2911 /// 2912 /// 2913 /// An illustration of this transform is turning this IR: 2914 /// \code 2915 /// BB: 2916 /// %cmp = icmp ult %x, %y 2917 /// br i1 %cmp, label %EndBB, label %ThenBB 2918 /// ThenBB: 2919 /// %sub = sub %x, %y 2920 /// br label BB2 2921 /// EndBB: 2922 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 2923 /// ... 2924 /// \endcode 2925 /// 2926 /// Into this IR: 2927 /// \code 2928 /// BB: 2929 /// %cmp = icmp ult %x, %y 2930 /// %sub = sub %x, %y 2931 /// %cond = select i1 %cmp, 0, %sub 2932 /// ... 2933 /// \endcode 2934 /// 2935 /// \returns true if the conditional block is removed. 2936 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, 2937 BasicBlock *ThenBB) { 2938 if (!Options.SpeculateBlocks) 2939 return false; 2940 2941 // Be conservative for now. FP select instruction can often be expensive. 2942 Value *BrCond = BI->getCondition(); 2943 if (isa<FCmpInst>(BrCond)) 2944 return false; 2945 2946 BasicBlock *BB = BI->getParent(); 2947 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 2948 InstructionCost Budget = 2949 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; 2950 2951 // If ThenBB is actually on the false edge of the conditional branch, remember 2952 // to swap the select operands later. 2953 bool Invert = false; 2954 if (ThenBB != BI->getSuccessor(0)) { 2955 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 2956 Invert = true; 2957 } 2958 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 2959 2960 // If the branch is non-unpredictable, and is predicted to *not* branch to 2961 // the `then` block, then avoid speculating it. 2962 if (!BI->getMetadata(LLVMContext::MD_unpredictable)) { 2963 uint64_t TWeight, FWeight; 2964 if (extractBranchWeights(*BI, TWeight, FWeight) && 2965 (TWeight + FWeight) != 0) { 2966 uint64_t EndWeight = Invert ? TWeight : FWeight; 2967 BranchProbability BIEndProb = 2968 BranchProbability::getBranchProbability(EndWeight, TWeight + FWeight); 2969 BranchProbability Likely = TTI.getPredictableBranchThreshold(); 2970 if (BIEndProb >= Likely) 2971 return false; 2972 } 2973 } 2974 2975 // Keep a count of how many times instructions are used within ThenBB when 2976 // they are candidates for sinking into ThenBB. Specifically: 2977 // - They are defined in BB, and 2978 // - They have no side effects, and 2979 // - All of their uses are in ThenBB. 2980 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 2981 2982 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics; 2983 2984 unsigned SpeculatedInstructions = 0; 2985 Value *SpeculatedStoreValue = nullptr; 2986 StoreInst *SpeculatedStore = nullptr; 2987 EphemeralValueTracker EphTracker; 2988 for (Instruction &I : reverse(drop_end(*ThenBB))) { 2989 // Skip debug info. 2990 if (isa<DbgInfoIntrinsic>(I)) { 2991 SpeculatedDbgIntrinsics.push_back(&I); 2992 continue; 2993 } 2994 2995 // Skip pseudo probes. The consequence is we lose track of the branch 2996 // probability for ThenBB, which is fine since the optimization here takes 2997 // place regardless of the branch probability. 2998 if (isa<PseudoProbeInst>(I)) { 2999 // The probe should be deleted so that it will not be over-counted when 3000 // the samples collected on the non-conditional path are counted towards 3001 // the conditional path. We leave it for the counts inference algorithm to 3002 // figure out a proper count for an unknown probe. 3003 SpeculatedDbgIntrinsics.push_back(&I); 3004 continue; 3005 } 3006 3007 // Ignore ephemeral values, they will be dropped by the transform. 3008 if (EphTracker.track(&I)) 3009 continue; 3010 3011 // Only speculatively execute a single instruction (not counting the 3012 // terminator) for now. 3013 ++SpeculatedInstructions; 3014 if (SpeculatedInstructions > 1) 3015 return false; 3016 3017 // Don't hoist the instruction if it's unsafe or expensive. 3018 if (!isSafeToSpeculativelyExecute(&I) && 3019 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore( 3020 &I, BB, ThenBB, EndBB)))) 3021 return false; 3022 if (!SpeculatedStoreValue && 3023 computeSpeculationCost(&I, TTI) > 3024 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic) 3025 return false; 3026 3027 // Store the store speculation candidate. 3028 if (SpeculatedStoreValue) 3029 SpeculatedStore = cast<StoreInst>(&I); 3030 3031 // Do not hoist the instruction if any of its operands are defined but not 3032 // used in BB. The transformation will prevent the operand from 3033 // being sunk into the use block. 3034 for (Use &Op : I.operands()) { 3035 Instruction *OpI = dyn_cast<Instruction>(Op); 3036 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects()) 3037 continue; // Not a candidate for sinking. 3038 3039 ++SinkCandidateUseCounts[OpI]; 3040 } 3041 } 3042 3043 // Consider any sink candidates which are only used in ThenBB as costs for 3044 // speculation. Note, while we iterate over a DenseMap here, we are summing 3045 // and so iteration order isn't significant. 3046 for (const auto &[Inst, Count] : SinkCandidateUseCounts) 3047 if (Inst->hasNUses(Count)) { 3048 ++SpeculatedInstructions; 3049 if (SpeculatedInstructions > 1) 3050 return false; 3051 } 3052 3053 // Check that we can insert the selects and that it's not too expensive to do 3054 // so. 3055 bool Convert = SpeculatedStore != nullptr; 3056 InstructionCost Cost = 0; 3057 Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB, 3058 SpeculatedInstructions, 3059 Cost, TTI); 3060 if (!Convert || Cost > Budget) 3061 return false; 3062 3063 // If we get here, we can hoist the instruction and if-convert. 3064 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 3065 3066 // Insert a select of the value of the speculated store. 3067 if (SpeculatedStoreValue) { 3068 IRBuilder<NoFolder> Builder(BI); 3069 Value *OrigV = SpeculatedStore->getValueOperand(); 3070 Value *TrueV = SpeculatedStore->getValueOperand(); 3071 Value *FalseV = SpeculatedStoreValue; 3072 if (Invert) 3073 std::swap(TrueV, FalseV); 3074 Value *S = Builder.CreateSelect( 3075 BrCond, TrueV, FalseV, "spec.store.select", BI); 3076 SpeculatedStore->setOperand(0, S); 3077 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(), 3078 SpeculatedStore->getDebugLoc()); 3079 // The value stored is still conditional, but the store itself is now 3080 // unconditonally executed, so we must be sure that any linked dbg.assign 3081 // intrinsics are tracking the new stored value (the result of the 3082 // select). If we don't, and the store were to be removed by another pass 3083 // (e.g. DSE), then we'd eventually end up emitting a location describing 3084 // the conditional value, unconditionally. 3085 // 3086 // === Before this transformation === 3087 // pred: 3088 // store %one, %x.dest, !DIAssignID !1 3089 // dbg.assign %one, "x", ..., !1, ... 3090 // br %cond if.then 3091 // 3092 // if.then: 3093 // store %two, %x.dest, !DIAssignID !2 3094 // dbg.assign %two, "x", ..., !2, ... 3095 // 3096 // === After this transformation === 3097 // pred: 3098 // store %one, %x.dest, !DIAssignID !1 3099 // dbg.assign %one, "x", ..., !1 3100 /// ... 3101 // %merge = select %cond, %two, %one 3102 // store %merge, %x.dest, !DIAssignID !2 3103 // dbg.assign %merge, "x", ..., !2 3104 for (auto *DAI : at::getAssignmentMarkers(SpeculatedStore)) { 3105 if (llvm::is_contained(DAI->location_ops(), OrigV)) 3106 DAI->replaceVariableLocationOp(OrigV, S); 3107 } 3108 } 3109 3110 // Metadata can be dependent on the condition we are hoisting above. 3111 // Strip all UB-implying metadata on the instruction. Drop the debug loc 3112 // to avoid making it appear as if the condition is a constant, which would 3113 // be misleading while debugging. 3114 // Similarly strip attributes that maybe dependent on condition we are 3115 // hoisting above. 3116 for (auto &I : make_early_inc_range(*ThenBB)) { 3117 if (!SpeculatedStoreValue || &I != SpeculatedStore) { 3118 // Don't update the DILocation of dbg.assign intrinsics. 3119 if (!isa<DbgAssignIntrinsic>(&I)) 3120 I.setDebugLoc(DebugLoc()); 3121 } 3122 I.dropUBImplyingAttrsAndMetadata(); 3123 3124 // Drop ephemeral values. 3125 if (EphTracker.contains(&I)) { 3126 I.replaceAllUsesWith(PoisonValue::get(I.getType())); 3127 I.eraseFromParent(); 3128 } 3129 } 3130 3131 // Hoist the instructions. 3132 // In "RemoveDIs" non-instr debug-info mode, drop DPValues attached to these 3133 // instructions, in the same way that dbg.value intrinsics are dropped at the 3134 // end of this block. 3135 for (auto &It : make_range(ThenBB->begin(), ThenBB->end())) 3136 It.dropDbgValues(); 3137 BB->splice(BI->getIterator(), ThenBB, ThenBB->begin(), 3138 std::prev(ThenBB->end())); 3139 3140 // Insert selects and rewrite the PHI operands. 3141 IRBuilder<NoFolder> Builder(BI); 3142 for (PHINode &PN : EndBB->phis()) { 3143 unsigned OrigI = PN.getBasicBlockIndex(BB); 3144 unsigned ThenI = PN.getBasicBlockIndex(ThenBB); 3145 Value *OrigV = PN.getIncomingValue(OrigI); 3146 Value *ThenV = PN.getIncomingValue(ThenI); 3147 3148 // Skip PHIs which are trivial. 3149 if (OrigV == ThenV) 3150 continue; 3151 3152 // Create a select whose true value is the speculatively executed value and 3153 // false value is the pre-existing value. Swap them if the branch 3154 // destinations were inverted. 3155 Value *TrueV = ThenV, *FalseV = OrigV; 3156 if (Invert) 3157 std::swap(TrueV, FalseV); 3158 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI); 3159 PN.setIncomingValue(OrigI, V); 3160 PN.setIncomingValue(ThenI, V); 3161 } 3162 3163 // Remove speculated dbg intrinsics. 3164 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the 3165 // dbg value for the different flows and inserting it after the select. 3166 for (Instruction *I : SpeculatedDbgIntrinsics) { 3167 // We still want to know that an assignment took place so don't remove 3168 // dbg.assign intrinsics. 3169 if (!isa<DbgAssignIntrinsic>(I)) 3170 I->eraseFromParent(); 3171 } 3172 3173 ++NumSpeculations; 3174 return true; 3175 } 3176 3177 /// Return true if we can thread a branch across this block. 3178 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 3179 int Size = 0; 3180 EphemeralValueTracker EphTracker; 3181 3182 // Walk the loop in reverse so that we can identify ephemeral values properly 3183 // (values only feeding assumes). 3184 for (Instruction &I : reverse(BB->instructionsWithoutDebug(false))) { 3185 // Can't fold blocks that contain noduplicate or convergent calls. 3186 if (CallInst *CI = dyn_cast<CallInst>(&I)) 3187 if (CI->cannotDuplicate() || CI->isConvergent()) 3188 return false; 3189 3190 // Ignore ephemeral values which are deleted during codegen. 3191 // We will delete Phis while threading, so Phis should not be accounted in 3192 // block's size. 3193 if (!EphTracker.track(&I) && !isa<PHINode>(I)) { 3194 if (Size++ > MaxSmallBlockSize) 3195 return false; // Don't clone large BB's. 3196 } 3197 3198 // We can only support instructions that do not define values that are 3199 // live outside of the current basic block. 3200 for (User *U : I.users()) { 3201 Instruction *UI = cast<Instruction>(U); 3202 if (UI->getParent() != BB || isa<PHINode>(UI)) 3203 return false; 3204 } 3205 3206 // Looks ok, continue checking. 3207 } 3208 3209 return true; 3210 } 3211 3212 static ConstantInt *getKnownValueOnEdge(Value *V, BasicBlock *From, 3213 BasicBlock *To) { 3214 // Don't look past the block defining the value, we might get the value from 3215 // a previous loop iteration. 3216 auto *I = dyn_cast<Instruction>(V); 3217 if (I && I->getParent() == To) 3218 return nullptr; 3219 3220 // We know the value if the From block branches on it. 3221 auto *BI = dyn_cast<BranchInst>(From->getTerminator()); 3222 if (BI && BI->isConditional() && BI->getCondition() == V && 3223 BI->getSuccessor(0) != BI->getSuccessor(1)) 3224 return BI->getSuccessor(0) == To ? ConstantInt::getTrue(BI->getContext()) 3225 : ConstantInt::getFalse(BI->getContext()); 3226 3227 return nullptr; 3228 } 3229 3230 /// If we have a conditional branch on something for which we know the constant 3231 /// value in predecessors (e.g. a phi node in the current block), thread edges 3232 /// from the predecessor to their ultimate destination. 3233 static std::optional<bool> 3234 FoldCondBranchOnValueKnownInPredecessorImpl(BranchInst *BI, DomTreeUpdater *DTU, 3235 const DataLayout &DL, 3236 AssumptionCache *AC) { 3237 SmallMapVector<ConstantInt *, SmallSetVector<BasicBlock *, 2>, 2> KnownValues; 3238 BasicBlock *BB = BI->getParent(); 3239 Value *Cond = BI->getCondition(); 3240 PHINode *PN = dyn_cast<PHINode>(Cond); 3241 if (PN && PN->getParent() == BB) { 3242 // Degenerate case of a single entry PHI. 3243 if (PN->getNumIncomingValues() == 1) { 3244 FoldSingleEntryPHINodes(PN->getParent()); 3245 return true; 3246 } 3247 3248 for (Use &U : PN->incoming_values()) 3249 if (auto *CB = dyn_cast<ConstantInt>(U)) 3250 KnownValues[CB].insert(PN->getIncomingBlock(U)); 3251 } else { 3252 for (BasicBlock *Pred : predecessors(BB)) { 3253 if (ConstantInt *CB = getKnownValueOnEdge(Cond, Pred, BB)) 3254 KnownValues[CB].insert(Pred); 3255 } 3256 } 3257 3258 if (KnownValues.empty()) 3259 return false; 3260 3261 // Now we know that this block has multiple preds and two succs. 3262 // Check that the block is small enough and values defined in the block are 3263 // not used outside of it. 3264 if (!BlockIsSimpleEnoughToThreadThrough(BB)) 3265 return false; 3266 3267 for (const auto &Pair : KnownValues) { 3268 // Okay, we now know that all edges from PredBB should be revectored to 3269 // branch to RealDest. 3270 ConstantInt *CB = Pair.first; 3271 ArrayRef<BasicBlock *> PredBBs = Pair.second.getArrayRef(); 3272 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 3273 3274 if (RealDest == BB) 3275 continue; // Skip self loops. 3276 3277 // Skip if the predecessor's terminator is an indirect branch. 3278 if (any_of(PredBBs, [](BasicBlock *PredBB) { 3279 return isa<IndirectBrInst>(PredBB->getTerminator()); 3280 })) 3281 continue; 3282 3283 LLVM_DEBUG({ 3284 dbgs() << "Condition " << *Cond << " in " << BB->getName() 3285 << " has value " << *Pair.first << " in predecessors:\n"; 3286 for (const BasicBlock *PredBB : Pair.second) 3287 dbgs() << " " << PredBB->getName() << "\n"; 3288 dbgs() << "Threading to destination " << RealDest->getName() << ".\n"; 3289 }); 3290 3291 // Split the predecessors we are threading into a new edge block. We'll 3292 // clone the instructions into this block, and then redirect it to RealDest. 3293 BasicBlock *EdgeBB = SplitBlockPredecessors(BB, PredBBs, ".critedge", DTU); 3294 3295 // TODO: These just exist to reduce test diff, we can drop them if we like. 3296 EdgeBB->setName(RealDest->getName() + ".critedge"); 3297 EdgeBB->moveBefore(RealDest); 3298 3299 // Update PHI nodes. 3300 AddPredecessorToBlock(RealDest, EdgeBB, BB); 3301 3302 // BB may have instructions that are being threaded over. Clone these 3303 // instructions into EdgeBB. We know that there will be no uses of the 3304 // cloned instructions outside of EdgeBB. 3305 BasicBlock::iterator InsertPt = EdgeBB->getFirstInsertionPt(); 3306 DenseMap<Value *, Value *> TranslateMap; // Track translated values. 3307 TranslateMap[Cond] = CB; 3308 3309 // RemoveDIs: track instructions that we optimise away while folding, so 3310 // that we can copy DPValues from them later. 3311 BasicBlock::iterator SrcDbgCursor = BB->begin(); 3312 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 3313 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 3314 TranslateMap[PN] = PN->getIncomingValueForBlock(EdgeBB); 3315 continue; 3316 } 3317 // Clone the instruction. 3318 Instruction *N = BBI->clone(); 3319 // Insert the new instruction into its new home. 3320 N->insertInto(EdgeBB, InsertPt); 3321 3322 if (BBI->hasName()) 3323 N->setName(BBI->getName() + ".c"); 3324 3325 // Update operands due to translation. 3326 for (Use &Op : N->operands()) { 3327 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(Op); 3328 if (PI != TranslateMap.end()) 3329 Op = PI->second; 3330 } 3331 3332 // Check for trivial simplification. 3333 if (Value *V = simplifyInstruction(N, {DL, nullptr, nullptr, AC})) { 3334 if (!BBI->use_empty()) 3335 TranslateMap[&*BBI] = V; 3336 if (!N->mayHaveSideEffects()) { 3337 N->eraseFromParent(); // Instruction folded away, don't need actual 3338 // inst 3339 N = nullptr; 3340 } 3341 } else { 3342 if (!BBI->use_empty()) 3343 TranslateMap[&*BBI] = N; 3344 } 3345 if (N) { 3346 // Copy all debug-info attached to instructions from the last we 3347 // successfully clone, up to this instruction (they might have been 3348 // folded away). 3349 for (; SrcDbgCursor != BBI; ++SrcDbgCursor) 3350 N->cloneDebugInfoFrom(&*SrcDbgCursor); 3351 SrcDbgCursor = std::next(BBI); 3352 // Clone debug-info on this instruction too. 3353 N->cloneDebugInfoFrom(&*BBI); 3354 3355 // Register the new instruction with the assumption cache if necessary. 3356 if (auto *Assume = dyn_cast<AssumeInst>(N)) 3357 if (AC) 3358 AC->registerAssumption(Assume); 3359 } 3360 } 3361 3362 for (; &*SrcDbgCursor != BI; ++SrcDbgCursor) 3363 InsertPt->cloneDebugInfoFrom(&*SrcDbgCursor); 3364 InsertPt->cloneDebugInfoFrom(BI); 3365 3366 BB->removePredecessor(EdgeBB); 3367 BranchInst *EdgeBI = cast<BranchInst>(EdgeBB->getTerminator()); 3368 EdgeBI->setSuccessor(0, RealDest); 3369 EdgeBI->setDebugLoc(BI->getDebugLoc()); 3370 3371 if (DTU) { 3372 SmallVector<DominatorTree::UpdateType, 2> Updates; 3373 Updates.push_back({DominatorTree::Delete, EdgeBB, BB}); 3374 Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest}); 3375 DTU->applyUpdates(Updates); 3376 } 3377 3378 // For simplicity, we created a separate basic block for the edge. Merge 3379 // it back into the predecessor if possible. This not only avoids 3380 // unnecessary SimplifyCFG iterations, but also makes sure that we don't 3381 // bypass the check for trivial cycles above. 3382 MergeBlockIntoPredecessor(EdgeBB, DTU); 3383 3384 // Signal repeat, simplifying any other constants. 3385 return std::nullopt; 3386 } 3387 3388 return false; 3389 } 3390 3391 static bool FoldCondBranchOnValueKnownInPredecessor(BranchInst *BI, 3392 DomTreeUpdater *DTU, 3393 const DataLayout &DL, 3394 AssumptionCache *AC) { 3395 std::optional<bool> Result; 3396 bool EverChanged = false; 3397 do { 3398 // Note that None means "we changed things, but recurse further." 3399 Result = FoldCondBranchOnValueKnownInPredecessorImpl(BI, DTU, DL, AC); 3400 EverChanged |= Result == std::nullopt || *Result; 3401 } while (Result == std::nullopt); 3402 return EverChanged; 3403 } 3404 3405 /// Given a BB that starts with the specified two-entry PHI node, 3406 /// see if we can eliminate it. 3407 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI, 3408 DomTreeUpdater *DTU, const DataLayout &DL) { 3409 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 3410 // statement", which has a very simple dominance structure. Basically, we 3411 // are trying to find the condition that is being branched on, which 3412 // subsequently causes this merge to happen. We really want control 3413 // dependence information for this check, but simplifycfg can't keep it up 3414 // to date, and this catches most of the cases we care about anyway. 3415 BasicBlock *BB = PN->getParent(); 3416 3417 BasicBlock *IfTrue, *IfFalse; 3418 BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse); 3419 if (!DomBI) 3420 return false; 3421 Value *IfCond = DomBI->getCondition(); 3422 // Don't bother if the branch will be constant folded trivially. 3423 if (isa<ConstantInt>(IfCond)) 3424 return false; 3425 3426 BasicBlock *DomBlock = DomBI->getParent(); 3427 SmallVector<BasicBlock *, 2> IfBlocks; 3428 llvm::copy_if( 3429 PN->blocks(), std::back_inserter(IfBlocks), [](BasicBlock *IfBlock) { 3430 return cast<BranchInst>(IfBlock->getTerminator())->isUnconditional(); 3431 }); 3432 assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) && 3433 "Will have either one or two blocks to speculate."); 3434 3435 // If the branch is non-unpredictable, see if we either predictably jump to 3436 // the merge bb (if we have only a single 'then' block), or if we predictably 3437 // jump to one specific 'then' block (if we have two of them). 3438 // It isn't beneficial to speculatively execute the code 3439 // from the block that we know is predictably not entered. 3440 if (!DomBI->getMetadata(LLVMContext::MD_unpredictable)) { 3441 uint64_t TWeight, FWeight; 3442 if (extractBranchWeights(*DomBI, TWeight, FWeight) && 3443 (TWeight + FWeight) != 0) { 3444 BranchProbability BITrueProb = 3445 BranchProbability::getBranchProbability(TWeight, TWeight + FWeight); 3446 BranchProbability Likely = TTI.getPredictableBranchThreshold(); 3447 BranchProbability BIFalseProb = BITrueProb.getCompl(); 3448 if (IfBlocks.size() == 1) { 3449 BranchProbability BIBBProb = 3450 DomBI->getSuccessor(0) == BB ? BITrueProb : BIFalseProb; 3451 if (BIBBProb >= Likely) 3452 return false; 3453 } else { 3454 if (BITrueProb >= Likely || BIFalseProb >= Likely) 3455 return false; 3456 } 3457 } 3458 } 3459 3460 // Don't try to fold an unreachable block. For example, the phi node itself 3461 // can't be the candidate if-condition for a select that we want to form. 3462 if (auto *IfCondPhiInst = dyn_cast<PHINode>(IfCond)) 3463 if (IfCondPhiInst->getParent() == BB) 3464 return false; 3465 3466 // Okay, we found that we can merge this two-entry phi node into a select. 3467 // Doing so would require us to fold *all* two entry phi nodes in this block. 3468 // At some point this becomes non-profitable (particularly if the target 3469 // doesn't support cmov's). Only do this transformation if there are two or 3470 // fewer PHI nodes in this block. 3471 unsigned NumPhis = 0; 3472 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 3473 if (NumPhis > 2) 3474 return false; 3475 3476 // Loop over the PHI's seeing if we can promote them all to select 3477 // instructions. While we are at it, keep track of the instructions 3478 // that need to be moved to the dominating block. 3479 SmallPtrSet<Instruction *, 4> AggressiveInsts; 3480 InstructionCost Cost = 0; 3481 InstructionCost Budget = 3482 TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; 3483 3484 bool Changed = false; 3485 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 3486 PHINode *PN = cast<PHINode>(II++); 3487 if (Value *V = simplifyInstruction(PN, {DL, PN})) { 3488 PN->replaceAllUsesWith(V); 3489 PN->eraseFromParent(); 3490 Changed = true; 3491 continue; 3492 } 3493 3494 if (!dominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts, 3495 Cost, Budget, TTI) || 3496 !dominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts, 3497 Cost, Budget, TTI)) 3498 return Changed; 3499 } 3500 3501 // If we folded the first phi, PN dangles at this point. Refresh it. If 3502 // we ran out of PHIs then we simplified them all. 3503 PN = dyn_cast<PHINode>(BB->begin()); 3504 if (!PN) 3505 return true; 3506 3507 // Return true if at least one of these is a 'not', and another is either 3508 // a 'not' too, or a constant. 3509 auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) { 3510 if (!match(V0, m_Not(m_Value()))) 3511 std::swap(V0, V1); 3512 auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant()); 3513 return match(V0, m_Not(m_Value())) && match(V1, Invertible); 3514 }; 3515 3516 // Don't fold i1 branches on PHIs which contain binary operators or 3517 // (possibly inverted) select form of or/ands, unless one of 3518 // the incoming values is an 'not' and another one is freely invertible. 3519 // These can often be turned into switches and other things. 3520 auto IsBinOpOrAnd = [](Value *V) { 3521 return match( 3522 V, m_CombineOr( 3523 m_BinOp(), 3524 m_CombineOr(m_Select(m_Value(), m_ImmConstant(), m_Value()), 3525 m_Select(m_Value(), m_Value(), m_ImmConstant())))); 3526 }; 3527 if (PN->getType()->isIntegerTy(1) && 3528 (IsBinOpOrAnd(PN->getIncomingValue(0)) || 3529 IsBinOpOrAnd(PN->getIncomingValue(1)) || IsBinOpOrAnd(IfCond)) && 3530 !CanHoistNotFromBothValues(PN->getIncomingValue(0), 3531 PN->getIncomingValue(1))) 3532 return Changed; 3533 3534 // If all PHI nodes are promotable, check to make sure that all instructions 3535 // in the predecessor blocks can be promoted as well. If not, we won't be able 3536 // to get rid of the control flow, so it's not worth promoting to select 3537 // instructions. 3538 for (BasicBlock *IfBlock : IfBlocks) 3539 for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I) 3540 if (!AggressiveInsts.count(&*I) && !I->isDebugOrPseudoInst()) { 3541 // This is not an aggressive instruction that we can promote. 3542 // Because of this, we won't be able to get rid of the control flow, so 3543 // the xform is not worth it. 3544 return Changed; 3545 } 3546 3547 // If either of the blocks has it's address taken, we can't do this fold. 3548 if (any_of(IfBlocks, 3549 [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); })) 3550 return Changed; 3551 3552 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond 3553 << " T: " << IfTrue->getName() 3554 << " F: " << IfFalse->getName() << "\n"); 3555 3556 // If we can still promote the PHI nodes after this gauntlet of tests, 3557 // do all of the PHI's now. 3558 3559 // Move all 'aggressive' instructions, which are defined in the 3560 // conditional parts of the if's up to the dominating block. 3561 for (BasicBlock *IfBlock : IfBlocks) 3562 hoistAllInstructionsInto(DomBlock, DomBI, IfBlock); 3563 3564 IRBuilder<NoFolder> Builder(DomBI); 3565 // Propagate fast-math-flags from phi nodes to replacement selects. 3566 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); 3567 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 3568 if (isa<FPMathOperator>(PN)) 3569 Builder.setFastMathFlags(PN->getFastMathFlags()); 3570 3571 // Change the PHI node into a select instruction. 3572 Value *TrueVal = PN->getIncomingValueForBlock(IfTrue); 3573 Value *FalseVal = PN->getIncomingValueForBlock(IfFalse); 3574 3575 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", DomBI); 3576 PN->replaceAllUsesWith(Sel); 3577 Sel->takeName(PN); 3578 PN->eraseFromParent(); 3579 } 3580 3581 // At this point, all IfBlocks are empty, so our if statement 3582 // has been flattened. Change DomBlock to jump directly to our new block to 3583 // avoid other simplifycfg's kicking in on the diamond. 3584 Builder.CreateBr(BB); 3585 3586 SmallVector<DominatorTree::UpdateType, 3> Updates; 3587 if (DTU) { 3588 Updates.push_back({DominatorTree::Insert, DomBlock, BB}); 3589 for (auto *Successor : successors(DomBlock)) 3590 Updates.push_back({DominatorTree::Delete, DomBlock, Successor}); 3591 } 3592 3593 DomBI->eraseFromParent(); 3594 if (DTU) 3595 DTU->applyUpdates(Updates); 3596 3597 return true; 3598 } 3599 3600 static Value *createLogicalOp(IRBuilderBase &Builder, 3601 Instruction::BinaryOps Opc, Value *LHS, 3602 Value *RHS, const Twine &Name = "") { 3603 // Try to relax logical op to binary op. 3604 if (impliesPoison(RHS, LHS)) 3605 return Builder.CreateBinOp(Opc, LHS, RHS, Name); 3606 if (Opc == Instruction::And) 3607 return Builder.CreateLogicalAnd(LHS, RHS, Name); 3608 if (Opc == Instruction::Or) 3609 return Builder.CreateLogicalOr(LHS, RHS, Name); 3610 llvm_unreachable("Invalid logical opcode"); 3611 } 3612 3613 /// Return true if either PBI or BI has branch weight available, and store 3614 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does 3615 /// not have branch weight, use 1:1 as its weight. 3616 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI, 3617 uint64_t &PredTrueWeight, 3618 uint64_t &PredFalseWeight, 3619 uint64_t &SuccTrueWeight, 3620 uint64_t &SuccFalseWeight) { 3621 bool PredHasWeights = 3622 extractBranchWeights(*PBI, PredTrueWeight, PredFalseWeight); 3623 bool SuccHasWeights = 3624 extractBranchWeights(*BI, SuccTrueWeight, SuccFalseWeight); 3625 if (PredHasWeights || SuccHasWeights) { 3626 if (!PredHasWeights) 3627 PredTrueWeight = PredFalseWeight = 1; 3628 if (!SuccHasWeights) 3629 SuccTrueWeight = SuccFalseWeight = 1; 3630 return true; 3631 } else { 3632 return false; 3633 } 3634 } 3635 3636 /// Determine if the two branches share a common destination and deduce a glue 3637 /// that joins the branches' conditions to arrive at the common destination if 3638 /// that would be profitable. 3639 static std::optional<std::tuple<BasicBlock *, Instruction::BinaryOps, bool>> 3640 shouldFoldCondBranchesToCommonDestination(BranchInst *BI, BranchInst *PBI, 3641 const TargetTransformInfo *TTI) { 3642 assert(BI && PBI && BI->isConditional() && PBI->isConditional() && 3643 "Both blocks must end with a conditional branches."); 3644 assert(is_contained(predecessors(BI->getParent()), PBI->getParent()) && 3645 "PredBB must be a predecessor of BB."); 3646 3647 // We have the potential to fold the conditions together, but if the 3648 // predecessor branch is predictable, we may not want to merge them. 3649 uint64_t PTWeight, PFWeight; 3650 BranchProbability PBITrueProb, Likely; 3651 if (TTI && !PBI->getMetadata(LLVMContext::MD_unpredictable) && 3652 extractBranchWeights(*PBI, PTWeight, PFWeight) && 3653 (PTWeight + PFWeight) != 0) { 3654 PBITrueProb = 3655 BranchProbability::getBranchProbability(PTWeight, PTWeight + PFWeight); 3656 Likely = TTI->getPredictableBranchThreshold(); 3657 } 3658 3659 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) { 3660 // Speculate the 2nd condition unless the 1st is probably true. 3661 if (PBITrueProb.isUnknown() || PBITrueProb < Likely) 3662 return {{BI->getSuccessor(0), Instruction::Or, false}}; 3663 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) { 3664 // Speculate the 2nd condition unless the 1st is probably false. 3665 if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely) 3666 return {{BI->getSuccessor(1), Instruction::And, false}}; 3667 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) { 3668 // Speculate the 2nd condition unless the 1st is probably true. 3669 if (PBITrueProb.isUnknown() || PBITrueProb < Likely) 3670 return {{BI->getSuccessor(1), Instruction::And, true}}; 3671 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) { 3672 // Speculate the 2nd condition unless the 1st is probably false. 3673 if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely) 3674 return {{BI->getSuccessor(0), Instruction::Or, true}}; 3675 } 3676 return std::nullopt; 3677 } 3678 3679 static bool performBranchToCommonDestFolding(BranchInst *BI, BranchInst *PBI, 3680 DomTreeUpdater *DTU, 3681 MemorySSAUpdater *MSSAU, 3682 const TargetTransformInfo *TTI) { 3683 BasicBlock *BB = BI->getParent(); 3684 BasicBlock *PredBlock = PBI->getParent(); 3685 3686 // Determine if the two branches share a common destination. 3687 BasicBlock *CommonSucc; 3688 Instruction::BinaryOps Opc; 3689 bool InvertPredCond; 3690 std::tie(CommonSucc, Opc, InvertPredCond) = 3691 *shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI); 3692 3693 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 3694 3695 IRBuilder<> Builder(PBI); 3696 // The builder is used to create instructions to eliminate the branch in BB. 3697 // If BB's terminator has !annotation metadata, add it to the new 3698 // instructions. 3699 Builder.CollectMetadataToCopy(BB->getTerminator(), 3700 {LLVMContext::MD_annotation}); 3701 3702 // If we need to invert the condition in the pred block to match, do so now. 3703 if (InvertPredCond) { 3704 InvertBranch(PBI, Builder); 3705 } 3706 3707 BasicBlock *UniqueSucc = 3708 PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1); 3709 3710 // Before cloning instructions, notify the successor basic block that it 3711 // is about to have a new predecessor. This will update PHI nodes, 3712 // which will allow us to update live-out uses of bonus instructions. 3713 AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU); 3714 3715 // Try to update branch weights. 3716 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 3717 if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight, 3718 SuccTrueWeight, SuccFalseWeight)) { 3719 SmallVector<uint64_t, 8> NewWeights; 3720 3721 if (PBI->getSuccessor(0) == BB) { 3722 // PBI: br i1 %x, BB, FalseDest 3723 // BI: br i1 %y, UniqueSucc, FalseDest 3724 // TrueWeight is TrueWeight for PBI * TrueWeight for BI. 3725 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 3726 // FalseWeight is FalseWeight for PBI * TotalWeight for BI + 3727 // TrueWeight for PBI * FalseWeight for BI. 3728 // We assume that total weights of a BranchInst can fit into 32 bits. 3729 // Therefore, we will not have overflow using 64-bit arithmetic. 3730 NewWeights.push_back(PredFalseWeight * 3731 (SuccFalseWeight + SuccTrueWeight) + 3732 PredTrueWeight * SuccFalseWeight); 3733 } else { 3734 // PBI: br i1 %x, TrueDest, BB 3735 // BI: br i1 %y, TrueDest, UniqueSucc 3736 // TrueWeight is TrueWeight for PBI * TotalWeight for BI + 3737 // FalseWeight for PBI * TrueWeight for BI. 3738 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) + 3739 PredFalseWeight * SuccTrueWeight); 3740 // FalseWeight is FalseWeight for PBI * FalseWeight for BI. 3741 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 3742 } 3743 3744 // Halve the weights if any of them cannot fit in an uint32_t 3745 FitWeights(NewWeights); 3746 3747 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end()); 3748 setBranchWeights(PBI, MDWeights[0], MDWeights[1]); 3749 3750 // TODO: If BB is reachable from all paths through PredBlock, then we 3751 // could replace PBI's branch probabilities with BI's. 3752 } else 3753 PBI->setMetadata(LLVMContext::MD_prof, nullptr); 3754 3755 // Now, update the CFG. 3756 PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc); 3757 3758 if (DTU) 3759 DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc}, 3760 {DominatorTree::Delete, PredBlock, BB}}); 3761 3762 // If BI was a loop latch, it may have had associated loop metadata. 3763 // We need to copy it to the new latch, that is, PBI. 3764 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop)) 3765 PBI->setMetadata(LLVMContext::MD_loop, LoopMD); 3766 3767 ValueToValueMapTy VMap; // maps original values to cloned values 3768 CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BB, PredBlock, VMap); 3769 3770 Module *M = BB->getModule(); 3771 3772 if (PredBlock->IsNewDbgInfoFormat) { 3773 PredBlock->getTerminator()->cloneDebugInfoFrom(BB->getTerminator()); 3774 for (DPValue &DPV : PredBlock->getTerminator()->getDbgValueRange()) { 3775 RemapDPValue(M, &DPV, VMap, 3776 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 3777 } 3778 } 3779 3780 // Now that the Cond was cloned into the predecessor basic block, 3781 // or/and the two conditions together. 3782 Value *BICond = VMap[BI->getCondition()]; 3783 PBI->setCondition( 3784 createLogicalOp(Builder, Opc, PBI->getCondition(), BICond, "or.cond")); 3785 3786 ++NumFoldBranchToCommonDest; 3787 return true; 3788 } 3789 3790 /// Return if an instruction's type or any of its operands' types are a vector 3791 /// type. 3792 static bool isVectorOp(Instruction &I) { 3793 return I.getType()->isVectorTy() || any_of(I.operands(), [](Use &U) { 3794 return U->getType()->isVectorTy(); 3795 }); 3796 } 3797 3798 /// If this basic block is simple enough, and if a predecessor branches to us 3799 /// and one of our successors, fold the block into the predecessor and use 3800 /// logical operations to pick the right destination. 3801 bool llvm::FoldBranchToCommonDest(BranchInst *BI, DomTreeUpdater *DTU, 3802 MemorySSAUpdater *MSSAU, 3803 const TargetTransformInfo *TTI, 3804 unsigned BonusInstThreshold) { 3805 // If this block ends with an unconditional branch, 3806 // let SpeculativelyExecuteBB() deal with it. 3807 if (!BI->isConditional()) 3808 return false; 3809 3810 BasicBlock *BB = BI->getParent(); 3811 TargetTransformInfo::TargetCostKind CostKind = 3812 BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize 3813 : TargetTransformInfo::TCK_SizeAndLatency; 3814 3815 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 3816 3817 if (!Cond || 3818 (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond) && 3819 !isa<SelectInst>(Cond)) || 3820 Cond->getParent() != BB || !Cond->hasOneUse()) 3821 return false; 3822 3823 // Finally, don't infinitely unroll conditional loops. 3824 if (is_contained(successors(BB), BB)) 3825 return false; 3826 3827 // With which predecessors will we want to deal with? 3828 SmallVector<BasicBlock *, 8> Preds; 3829 for (BasicBlock *PredBlock : predecessors(BB)) { 3830 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 3831 3832 // Check that we have two conditional branches. If there is a PHI node in 3833 // the common successor, verify that the same value flows in from both 3834 // blocks. 3835 if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI)) 3836 continue; 3837 3838 // Determine if the two branches share a common destination. 3839 BasicBlock *CommonSucc; 3840 Instruction::BinaryOps Opc; 3841 bool InvertPredCond; 3842 if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI)) 3843 std::tie(CommonSucc, Opc, InvertPredCond) = *Recipe; 3844 else 3845 continue; 3846 3847 // Check the cost of inserting the necessary logic before performing the 3848 // transformation. 3849 if (TTI) { 3850 Type *Ty = BI->getCondition()->getType(); 3851 InstructionCost Cost = TTI->getArithmeticInstrCost(Opc, Ty, CostKind); 3852 if (InvertPredCond && (!PBI->getCondition()->hasOneUse() || 3853 !isa<CmpInst>(PBI->getCondition()))) 3854 Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind); 3855 3856 if (Cost > BranchFoldThreshold) 3857 continue; 3858 } 3859 3860 // Ok, we do want to deal with this predecessor. Record it. 3861 Preds.emplace_back(PredBlock); 3862 } 3863 3864 // If there aren't any predecessors into which we can fold, 3865 // don't bother checking the cost. 3866 if (Preds.empty()) 3867 return false; 3868 3869 // Only allow this transformation if computing the condition doesn't involve 3870 // too many instructions and these involved instructions can be executed 3871 // unconditionally. We denote all involved instructions except the condition 3872 // as "bonus instructions", and only allow this transformation when the 3873 // number of the bonus instructions we'll need to create when cloning into 3874 // each predecessor does not exceed a certain threshold. 3875 unsigned NumBonusInsts = 0; 3876 bool SawVectorOp = false; 3877 const unsigned PredCount = Preds.size(); 3878 for (Instruction &I : *BB) { 3879 // Don't check the branch condition comparison itself. 3880 if (&I == Cond) 3881 continue; 3882 // Ignore dbg intrinsics, and the terminator. 3883 if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I)) 3884 continue; 3885 // I must be safe to execute unconditionally. 3886 if (!isSafeToSpeculativelyExecute(&I)) 3887 return false; 3888 SawVectorOp |= isVectorOp(I); 3889 3890 // Account for the cost of duplicating this instruction into each 3891 // predecessor. Ignore free instructions. 3892 if (!TTI || TTI->getInstructionCost(&I, CostKind) != 3893 TargetTransformInfo::TCC_Free) { 3894 NumBonusInsts += PredCount; 3895 3896 // Early exits once we reach the limit. 3897 if (NumBonusInsts > 3898 BonusInstThreshold * BranchFoldToCommonDestVectorMultiplier) 3899 return false; 3900 } 3901 3902 auto IsBCSSAUse = [BB, &I](Use &U) { 3903 auto *UI = cast<Instruction>(U.getUser()); 3904 if (auto *PN = dyn_cast<PHINode>(UI)) 3905 return PN->getIncomingBlock(U) == BB; 3906 return UI->getParent() == BB && I.comesBefore(UI); 3907 }; 3908 3909 // Does this instruction require rewriting of uses? 3910 if (!all_of(I.uses(), IsBCSSAUse)) 3911 return false; 3912 } 3913 if (NumBonusInsts > 3914 BonusInstThreshold * 3915 (SawVectorOp ? BranchFoldToCommonDestVectorMultiplier : 1)) 3916 return false; 3917 3918 // Ok, we have the budget. Perform the transformation. 3919 for (BasicBlock *PredBlock : Preds) { 3920 auto *PBI = cast<BranchInst>(PredBlock->getTerminator()); 3921 return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI); 3922 } 3923 return false; 3924 } 3925 3926 // If there is only one store in BB1 and BB2, return it, otherwise return 3927 // nullptr. 3928 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) { 3929 StoreInst *S = nullptr; 3930 for (auto *BB : {BB1, BB2}) { 3931 if (!BB) 3932 continue; 3933 for (auto &I : *BB) 3934 if (auto *SI = dyn_cast<StoreInst>(&I)) { 3935 if (S) 3936 // Multiple stores seen. 3937 return nullptr; 3938 else 3939 S = SI; 3940 } 3941 } 3942 return S; 3943 } 3944 3945 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB, 3946 Value *AlternativeV = nullptr) { 3947 // PHI is going to be a PHI node that allows the value V that is defined in 3948 // BB to be referenced in BB's only successor. 3949 // 3950 // If AlternativeV is nullptr, the only value we care about in PHI is V. It 3951 // doesn't matter to us what the other operand is (it'll never get used). We 3952 // could just create a new PHI with an undef incoming value, but that could 3953 // increase register pressure if EarlyCSE/InstCombine can't fold it with some 3954 // other PHI. So here we directly look for some PHI in BB's successor with V 3955 // as an incoming operand. If we find one, we use it, else we create a new 3956 // one. 3957 // 3958 // If AlternativeV is not nullptr, we care about both incoming values in PHI. 3959 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV] 3960 // where OtherBB is the single other predecessor of BB's only successor. 3961 PHINode *PHI = nullptr; 3962 BasicBlock *Succ = BB->getSingleSuccessor(); 3963 3964 for (auto I = Succ->begin(); isa<PHINode>(I); ++I) 3965 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) { 3966 PHI = cast<PHINode>(I); 3967 if (!AlternativeV) 3968 break; 3969 3970 assert(Succ->hasNPredecessors(2)); 3971 auto PredI = pred_begin(Succ); 3972 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI; 3973 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV) 3974 break; 3975 PHI = nullptr; 3976 } 3977 if (PHI) 3978 return PHI; 3979 3980 // If V is not an instruction defined in BB, just return it. 3981 if (!AlternativeV && 3982 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB)) 3983 return V; 3984 3985 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge"); 3986 PHI->insertBefore(Succ->begin()); 3987 PHI->addIncoming(V, BB); 3988 for (BasicBlock *PredBB : predecessors(Succ)) 3989 if (PredBB != BB) 3990 PHI->addIncoming( 3991 AlternativeV ? AlternativeV : PoisonValue::get(V->getType()), PredBB); 3992 return PHI; 3993 } 3994 3995 static bool mergeConditionalStoreToAddress( 3996 BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB, 3997 BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond, 3998 DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) { 3999 // For every pointer, there must be exactly two stores, one coming from 4000 // PTB or PFB, and the other from QTB or QFB. We don't support more than one 4001 // store (to any address) in PTB,PFB or QTB,QFB. 4002 // FIXME: We could relax this restriction with a bit more work and performance 4003 // testing. 4004 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB); 4005 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB); 4006 if (!PStore || !QStore) 4007 return false; 4008 4009 // Now check the stores are compatible. 4010 if (!QStore->isUnordered() || !PStore->isUnordered() || 4011 PStore->getValueOperand()->getType() != 4012 QStore->getValueOperand()->getType()) 4013 return false; 4014 4015 // Check that sinking the store won't cause program behavior changes. Sinking 4016 // the store out of the Q blocks won't change any behavior as we're sinking 4017 // from a block to its unconditional successor. But we're moving a store from 4018 // the P blocks down through the middle block (QBI) and past both QFB and QTB. 4019 // So we need to check that there are no aliasing loads or stores in 4020 // QBI, QTB and QFB. We also need to check there are no conflicting memory 4021 // operations between PStore and the end of its parent block. 4022 // 4023 // The ideal way to do this is to query AliasAnalysis, but we don't 4024 // preserve AA currently so that is dangerous. Be super safe and just 4025 // check there are no other memory operations at all. 4026 for (auto &I : *QFB->getSinglePredecessor()) 4027 if (I.mayReadOrWriteMemory()) 4028 return false; 4029 for (auto &I : *QFB) 4030 if (&I != QStore && I.mayReadOrWriteMemory()) 4031 return false; 4032 if (QTB) 4033 for (auto &I : *QTB) 4034 if (&I != QStore && I.mayReadOrWriteMemory()) 4035 return false; 4036 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end(); 4037 I != E; ++I) 4038 if (&*I != PStore && I->mayReadOrWriteMemory()) 4039 return false; 4040 4041 // If we're not in aggressive mode, we only optimize if we have some 4042 // confidence that by optimizing we'll allow P and/or Q to be if-converted. 4043 auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) { 4044 if (!BB) 4045 return true; 4046 // Heuristic: if the block can be if-converted/phi-folded and the 4047 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to 4048 // thread this store. 4049 InstructionCost Cost = 0; 4050 InstructionCost Budget = 4051 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; 4052 for (auto &I : BB->instructionsWithoutDebug(false)) { 4053 // Consider terminator instruction to be free. 4054 if (I.isTerminator()) 4055 continue; 4056 // If this is one the stores that we want to speculate out of this BB, 4057 // then don't count it's cost, consider it to be free. 4058 if (auto *S = dyn_cast<StoreInst>(&I)) 4059 if (llvm::find(FreeStores, S)) 4060 continue; 4061 // Else, we have a white-list of instructions that we are ak speculating. 4062 if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I)) 4063 return false; // Not in white-list - not worthwhile folding. 4064 // And finally, if this is a non-free instruction that we are okay 4065 // speculating, ensure that we consider the speculation budget. 4066 Cost += 4067 TTI.getInstructionCost(&I, TargetTransformInfo::TCK_SizeAndLatency); 4068 if (Cost > Budget) 4069 return false; // Eagerly refuse to fold as soon as we're out of budget. 4070 } 4071 assert(Cost <= Budget && 4072 "When we run out of budget we will eagerly return from within the " 4073 "per-instruction loop."); 4074 return true; 4075 }; 4076 4077 const std::array<StoreInst *, 2> FreeStores = {PStore, QStore}; 4078 if (!MergeCondStoresAggressively && 4079 (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) || 4080 !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores))) 4081 return false; 4082 4083 // If PostBB has more than two predecessors, we need to split it so we can 4084 // sink the store. 4085 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) { 4086 // We know that QFB's only successor is PostBB. And QFB has a single 4087 // predecessor. If QTB exists, then its only successor is also PostBB. 4088 // If QTB does not exist, then QFB's only predecessor has a conditional 4089 // branch to QFB and PostBB. 4090 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor(); 4091 BasicBlock *NewBB = 4092 SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU); 4093 if (!NewBB) 4094 return false; 4095 PostBB = NewBB; 4096 } 4097 4098 // OK, we're going to sink the stores to PostBB. The store has to be 4099 // conditional though, so first create the predicate. 4100 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator()) 4101 ->getCondition(); 4102 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator()) 4103 ->getCondition(); 4104 4105 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(), 4106 PStore->getParent()); 4107 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(), 4108 QStore->getParent(), PPHI); 4109 4110 BasicBlock::iterator PostBBFirst = PostBB->getFirstInsertionPt(); 4111 IRBuilder<> QB(PostBB, PostBBFirst); 4112 QB.SetCurrentDebugLocation(PostBBFirst->getStableDebugLoc()); 4113 4114 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond); 4115 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond); 4116 4117 if (InvertPCond) 4118 PPred = QB.CreateNot(PPred); 4119 if (InvertQCond) 4120 QPred = QB.CreateNot(QPred); 4121 Value *CombinedPred = QB.CreateOr(PPred, QPred); 4122 4123 BasicBlock::iterator InsertPt = QB.GetInsertPoint(); 4124 auto *T = SplitBlockAndInsertIfThen(CombinedPred, InsertPt, 4125 /*Unreachable=*/false, 4126 /*BranchWeights=*/nullptr, DTU); 4127 4128 QB.SetInsertPoint(T); 4129 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address)); 4130 SI->setAAMetadata(PStore->getAAMetadata().merge(QStore->getAAMetadata())); 4131 // Choose the minimum alignment. If we could prove both stores execute, we 4132 // could use biggest one. In this case, though, we only know that one of the 4133 // stores executes. And we don't know it's safe to take the alignment from a 4134 // store that doesn't execute. 4135 SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign())); 4136 4137 QStore->eraseFromParent(); 4138 PStore->eraseFromParent(); 4139 4140 return true; 4141 } 4142 4143 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI, 4144 DomTreeUpdater *DTU, const DataLayout &DL, 4145 const TargetTransformInfo &TTI) { 4146 // The intention here is to find diamonds or triangles (see below) where each 4147 // conditional block contains a store to the same address. Both of these 4148 // stores are conditional, so they can't be unconditionally sunk. But it may 4149 // be profitable to speculatively sink the stores into one merged store at the 4150 // end, and predicate the merged store on the union of the two conditions of 4151 // PBI and QBI. 4152 // 4153 // This can reduce the number of stores executed if both of the conditions are 4154 // true, and can allow the blocks to become small enough to be if-converted. 4155 // This optimization will also chain, so that ladders of test-and-set 4156 // sequences can be if-converted away. 4157 // 4158 // We only deal with simple diamonds or triangles: 4159 // 4160 // PBI or PBI or a combination of the two 4161 // / \ | \ 4162 // PTB PFB | PFB 4163 // \ / | / 4164 // QBI QBI 4165 // / \ | \ 4166 // QTB QFB | QFB 4167 // \ / | / 4168 // PostBB PostBB 4169 // 4170 // We model triangles as a type of diamond with a nullptr "true" block. 4171 // Triangles are canonicalized so that the fallthrough edge is represented by 4172 // a true condition, as in the diagram above. 4173 BasicBlock *PTB = PBI->getSuccessor(0); 4174 BasicBlock *PFB = PBI->getSuccessor(1); 4175 BasicBlock *QTB = QBI->getSuccessor(0); 4176 BasicBlock *QFB = QBI->getSuccessor(1); 4177 BasicBlock *PostBB = QFB->getSingleSuccessor(); 4178 4179 // Make sure we have a good guess for PostBB. If QTB's only successor is 4180 // QFB, then QFB is a better PostBB. 4181 if (QTB->getSingleSuccessor() == QFB) 4182 PostBB = QFB; 4183 4184 // If we couldn't find a good PostBB, stop. 4185 if (!PostBB) 4186 return false; 4187 4188 bool InvertPCond = false, InvertQCond = false; 4189 // Canonicalize fallthroughs to the true branches. 4190 if (PFB == QBI->getParent()) { 4191 std::swap(PFB, PTB); 4192 InvertPCond = true; 4193 } 4194 if (QFB == PostBB) { 4195 std::swap(QFB, QTB); 4196 InvertQCond = true; 4197 } 4198 4199 // From this point on we can assume PTB or QTB may be fallthroughs but PFB 4200 // and QFB may not. Model fallthroughs as a nullptr block. 4201 if (PTB == QBI->getParent()) 4202 PTB = nullptr; 4203 if (QTB == PostBB) 4204 QTB = nullptr; 4205 4206 // Legality bailouts. We must have at least the non-fallthrough blocks and 4207 // the post-dominating block, and the non-fallthroughs must only have one 4208 // predecessor. 4209 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) { 4210 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S; 4211 }; 4212 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) || 4213 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB)) 4214 return false; 4215 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) || 4216 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB))) 4217 return false; 4218 if (!QBI->getParent()->hasNUses(2)) 4219 return false; 4220 4221 // OK, this is a sequence of two diamonds or triangles. 4222 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB. 4223 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses; 4224 for (auto *BB : {PTB, PFB}) { 4225 if (!BB) 4226 continue; 4227 for (auto &I : *BB) 4228 if (StoreInst *SI = dyn_cast<StoreInst>(&I)) 4229 PStoreAddresses.insert(SI->getPointerOperand()); 4230 } 4231 for (auto *BB : {QTB, QFB}) { 4232 if (!BB) 4233 continue; 4234 for (auto &I : *BB) 4235 if (StoreInst *SI = dyn_cast<StoreInst>(&I)) 4236 QStoreAddresses.insert(SI->getPointerOperand()); 4237 } 4238 4239 set_intersect(PStoreAddresses, QStoreAddresses); 4240 // set_intersect mutates PStoreAddresses in place. Rename it here to make it 4241 // clear what it contains. 4242 auto &CommonAddresses = PStoreAddresses; 4243 4244 bool Changed = false; 4245 for (auto *Address : CommonAddresses) 4246 Changed |= 4247 mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address, 4248 InvertPCond, InvertQCond, DTU, DL, TTI); 4249 return Changed; 4250 } 4251 4252 /// If the previous block ended with a widenable branch, determine if reusing 4253 /// the target block is profitable and legal. This will have the effect of 4254 /// "widening" PBI, but doesn't require us to reason about hosting safety. 4255 static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI, 4256 DomTreeUpdater *DTU) { 4257 // TODO: This can be generalized in two important ways: 4258 // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input 4259 // values from the PBI edge. 4260 // 2) We can sink side effecting instructions into BI's fallthrough 4261 // successor provided they doesn't contribute to computation of 4262 // BI's condition. 4263 BasicBlock *IfTrueBB = PBI->getSuccessor(0); 4264 BasicBlock *IfFalseBB = PBI->getSuccessor(1); 4265 if (!isWidenableBranch(PBI) || IfTrueBB != BI->getParent() || 4266 !BI->getParent()->getSinglePredecessor()) 4267 return false; 4268 if (!IfFalseBB->phis().empty()) 4269 return false; // TODO 4270 // This helps avoid infinite loop with SimplifyCondBranchToCondBranch which 4271 // may undo the transform done here. 4272 // TODO: There might be a more fine-grained solution to this. 4273 if (!llvm::succ_empty(IfFalseBB)) 4274 return false; 4275 // Use lambda to lazily compute expensive condition after cheap ones. 4276 auto NoSideEffects = [](BasicBlock &BB) { 4277 return llvm::none_of(BB, [](const Instruction &I) { 4278 return I.mayWriteToMemory() || I.mayHaveSideEffects(); 4279 }); 4280 }; 4281 if (BI->getSuccessor(1) != IfFalseBB && // no inf looping 4282 BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability 4283 NoSideEffects(*BI->getParent())) { 4284 auto *OldSuccessor = BI->getSuccessor(1); 4285 OldSuccessor->removePredecessor(BI->getParent()); 4286 BI->setSuccessor(1, IfFalseBB); 4287 if (DTU) 4288 DTU->applyUpdates( 4289 {{DominatorTree::Insert, BI->getParent(), IfFalseBB}, 4290 {DominatorTree::Delete, BI->getParent(), OldSuccessor}}); 4291 return true; 4292 } 4293 if (BI->getSuccessor(0) != IfFalseBB && // no inf looping 4294 BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability 4295 NoSideEffects(*BI->getParent())) { 4296 auto *OldSuccessor = BI->getSuccessor(0); 4297 OldSuccessor->removePredecessor(BI->getParent()); 4298 BI->setSuccessor(0, IfFalseBB); 4299 if (DTU) 4300 DTU->applyUpdates( 4301 {{DominatorTree::Insert, BI->getParent(), IfFalseBB}, 4302 {DominatorTree::Delete, BI->getParent(), OldSuccessor}}); 4303 return true; 4304 } 4305 return false; 4306 } 4307 4308 /// If we have a conditional branch as a predecessor of another block, 4309 /// this function tries to simplify it. We know 4310 /// that PBI and BI are both conditional branches, and BI is in one of the 4311 /// successor blocks of PBI - PBI branches to BI. 4312 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI, 4313 DomTreeUpdater *DTU, 4314 const DataLayout &DL, 4315 const TargetTransformInfo &TTI) { 4316 assert(PBI->isConditional() && BI->isConditional()); 4317 BasicBlock *BB = BI->getParent(); 4318 4319 // If this block ends with a branch instruction, and if there is a 4320 // predecessor that ends on a branch of the same condition, make 4321 // this conditional branch redundant. 4322 if (PBI->getCondition() == BI->getCondition() && 4323 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 4324 // Okay, the outcome of this conditional branch is statically 4325 // knowable. If this block had a single pred, handle specially, otherwise 4326 // FoldCondBranchOnValueKnownInPredecessor() will handle it. 4327 if (BB->getSinglePredecessor()) { 4328 // Turn this into a branch on constant. 4329 bool CondIsTrue = PBI->getSuccessor(0) == BB; 4330 BI->setCondition( 4331 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue)); 4332 return true; // Nuke the branch on constant. 4333 } 4334 } 4335 4336 // If the previous block ended with a widenable branch, determine if reusing 4337 // the target block is profitable and legal. This will have the effect of 4338 // "widening" PBI, but doesn't require us to reason about hosting safety. 4339 if (tryWidenCondBranchToCondBranch(PBI, BI, DTU)) 4340 return true; 4341 4342 // If both branches are conditional and both contain stores to the same 4343 // address, remove the stores from the conditionals and create a conditional 4344 // merged store at the end. 4345 if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI)) 4346 return true; 4347 4348 // If this is a conditional branch in an empty block, and if any 4349 // predecessors are a conditional branch to one of our destinations, 4350 // fold the conditions into logical ops and one cond br. 4351 4352 // Ignore dbg intrinsics. 4353 if (&*BB->instructionsWithoutDebug(false).begin() != BI) 4354 return false; 4355 4356 int PBIOp, BIOp; 4357 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) { 4358 PBIOp = 0; 4359 BIOp = 0; 4360 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) { 4361 PBIOp = 0; 4362 BIOp = 1; 4363 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) { 4364 PBIOp = 1; 4365 BIOp = 0; 4366 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) { 4367 PBIOp = 1; 4368 BIOp = 1; 4369 } else { 4370 return false; 4371 } 4372 4373 // Check to make sure that the other destination of this branch 4374 // isn't BB itself. If so, this is an infinite loop that will 4375 // keep getting unwound. 4376 if (PBI->getSuccessor(PBIOp) == BB) 4377 return false; 4378 4379 // If predecessor's branch probability to BB is too low don't merge branches. 4380 SmallVector<uint32_t, 2> PredWeights; 4381 if (!PBI->getMetadata(LLVMContext::MD_unpredictable) && 4382 extractBranchWeights(*PBI, PredWeights) && 4383 (static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]) != 0) { 4384 4385 BranchProbability CommonDestProb = BranchProbability::getBranchProbability( 4386 PredWeights[PBIOp], 4387 static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]); 4388 4389 BranchProbability Likely = TTI.getPredictableBranchThreshold(); 4390 if (CommonDestProb >= Likely) 4391 return false; 4392 } 4393 4394 // Do not perform this transformation if it would require 4395 // insertion of a large number of select instructions. For targets 4396 // without predication/cmovs, this is a big pessimization. 4397 4398 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 4399 BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1); 4400 unsigned NumPhis = 0; 4401 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II); 4402 ++II, ++NumPhis) { 4403 if (NumPhis > 2) // Disable this xform. 4404 return false; 4405 } 4406 4407 // Finally, if everything is ok, fold the branches to logical ops. 4408 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 4409 4410 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 4411 << "AND: " << *BI->getParent()); 4412 4413 SmallVector<DominatorTree::UpdateType, 5> Updates; 4414 4415 // If OtherDest *is* BB, then BB is a basic block with a single conditional 4416 // branch in it, where one edge (OtherDest) goes back to itself but the other 4417 // exits. We don't *know* that the program avoids the infinite loop 4418 // (even though that seems likely). If we do this xform naively, we'll end up 4419 // recursively unpeeling the loop. Since we know that (after the xform is 4420 // done) that the block *is* infinite if reached, we just make it an obviously 4421 // infinite loop with no cond branch. 4422 if (OtherDest == BB) { 4423 // Insert it at the end of the function, because it's either code, 4424 // or it won't matter if it's hot. :) 4425 BasicBlock *InfLoopBlock = 4426 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent()); 4427 BranchInst::Create(InfLoopBlock, InfLoopBlock); 4428 if (DTU) 4429 Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock}); 4430 OtherDest = InfLoopBlock; 4431 } 4432 4433 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent()); 4434 4435 // BI may have other predecessors. Because of this, we leave 4436 // it alone, but modify PBI. 4437 4438 // Make sure we get to CommonDest on True&True directions. 4439 Value *PBICond = PBI->getCondition(); 4440 IRBuilder<NoFolder> Builder(PBI); 4441 if (PBIOp) 4442 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not"); 4443 4444 Value *BICond = BI->getCondition(); 4445 if (BIOp) 4446 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not"); 4447 4448 // Merge the conditions. 4449 Value *Cond = 4450 createLogicalOp(Builder, Instruction::Or, PBICond, BICond, "brmerge"); 4451 4452 // Modify PBI to branch on the new condition to the new dests. 4453 PBI->setCondition(Cond); 4454 PBI->setSuccessor(0, CommonDest); 4455 PBI->setSuccessor(1, OtherDest); 4456 4457 if (DTU) { 4458 Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest}); 4459 Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest}); 4460 4461 DTU->applyUpdates(Updates); 4462 } 4463 4464 // Update branch weight for PBI. 4465 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 4466 uint64_t PredCommon, PredOther, SuccCommon, SuccOther; 4467 bool HasWeights = 4468 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight, 4469 SuccTrueWeight, SuccFalseWeight); 4470 if (HasWeights) { 4471 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 4472 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight; 4473 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 4474 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 4475 // The weight to CommonDest should be PredCommon * SuccTotal + 4476 // PredOther * SuccCommon. 4477 // The weight to OtherDest should be PredOther * SuccOther. 4478 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) + 4479 PredOther * SuccCommon, 4480 PredOther * SuccOther}; 4481 // Halve the weights if any of them cannot fit in an uint32_t 4482 FitWeights(NewWeights); 4483 4484 setBranchWeights(PBI, NewWeights[0], NewWeights[1]); 4485 } 4486 4487 // OtherDest may have phi nodes. If so, add an entry from PBI's 4488 // block that are identical to the entries for BI's block. 4489 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 4490 4491 // We know that the CommonDest already had an edge from PBI to 4492 // it. If it has PHIs though, the PHIs may have different 4493 // entries for BB and PBI's BB. If so, insert a select to make 4494 // them agree. 4495 for (PHINode &PN : CommonDest->phis()) { 4496 Value *BIV = PN.getIncomingValueForBlock(BB); 4497 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent()); 4498 Value *PBIV = PN.getIncomingValue(PBBIdx); 4499 if (BIV != PBIV) { 4500 // Insert a select in PBI to pick the right value. 4501 SelectInst *NV = cast<SelectInst>( 4502 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux")); 4503 PN.setIncomingValue(PBBIdx, NV); 4504 // Although the select has the same condition as PBI, the original branch 4505 // weights for PBI do not apply to the new select because the select's 4506 // 'logical' edges are incoming edges of the phi that is eliminated, not 4507 // the outgoing edges of PBI. 4508 if (HasWeights) { 4509 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 4510 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight; 4511 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 4512 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 4513 // The weight to PredCommonDest should be PredCommon * SuccTotal. 4514 // The weight to PredOtherDest should be PredOther * SuccCommon. 4515 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther), 4516 PredOther * SuccCommon}; 4517 4518 FitWeights(NewWeights); 4519 4520 setBranchWeights(NV, NewWeights[0], NewWeights[1]); 4521 } 4522 } 4523 } 4524 4525 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 4526 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent()); 4527 4528 // This basic block is probably dead. We know it has at least 4529 // one fewer predecessor. 4530 return true; 4531 } 4532 4533 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is 4534 // true or to FalseBB if Cond is false. 4535 // Takes care of updating the successors and removing the old terminator. 4536 // Also makes sure not to introduce new successors by assuming that edges to 4537 // non-successor TrueBBs and FalseBBs aren't reachable. 4538 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm, 4539 Value *Cond, BasicBlock *TrueBB, 4540 BasicBlock *FalseBB, 4541 uint32_t TrueWeight, 4542 uint32_t FalseWeight) { 4543 auto *BB = OldTerm->getParent(); 4544 // Remove any superfluous successor edges from the CFG. 4545 // First, figure out which successors to preserve. 4546 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 4547 // successor. 4548 BasicBlock *KeepEdge1 = TrueBB; 4549 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr; 4550 4551 SmallSetVector<BasicBlock *, 2> RemovedSuccessors; 4552 4553 // Then remove the rest. 4554 for (BasicBlock *Succ : successors(OldTerm)) { 4555 // Make sure only to keep exactly one copy of each edge. 4556 if (Succ == KeepEdge1) 4557 KeepEdge1 = nullptr; 4558 else if (Succ == KeepEdge2) 4559 KeepEdge2 = nullptr; 4560 else { 4561 Succ->removePredecessor(BB, 4562 /*KeepOneInputPHIs=*/true); 4563 4564 if (Succ != TrueBB && Succ != FalseBB) 4565 RemovedSuccessors.insert(Succ); 4566 } 4567 } 4568 4569 IRBuilder<> Builder(OldTerm); 4570 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 4571 4572 // Insert an appropriate new terminator. 4573 if (!KeepEdge1 && !KeepEdge2) { 4574 if (TrueBB == FalseBB) { 4575 // We were only looking for one successor, and it was present. 4576 // Create an unconditional branch to it. 4577 Builder.CreateBr(TrueBB); 4578 } else { 4579 // We found both of the successors we were looking for. 4580 // Create a conditional branch sharing the condition of the select. 4581 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 4582 if (TrueWeight != FalseWeight) 4583 setBranchWeights(NewBI, TrueWeight, FalseWeight); 4584 } 4585 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 4586 // Neither of the selected blocks were successors, so this 4587 // terminator must be unreachable. 4588 new UnreachableInst(OldTerm->getContext(), OldTerm); 4589 } else { 4590 // One of the selected values was a successor, but the other wasn't. 4591 // Insert an unconditional branch to the one that was found; 4592 // the edge to the one that wasn't must be unreachable. 4593 if (!KeepEdge1) { 4594 // Only TrueBB was found. 4595 Builder.CreateBr(TrueBB); 4596 } else { 4597 // Only FalseBB was found. 4598 Builder.CreateBr(FalseBB); 4599 } 4600 } 4601 4602 EraseTerminatorAndDCECond(OldTerm); 4603 4604 if (DTU) { 4605 SmallVector<DominatorTree::UpdateType, 2> Updates; 4606 Updates.reserve(RemovedSuccessors.size()); 4607 for (auto *RemovedSuccessor : RemovedSuccessors) 4608 Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor}); 4609 DTU->applyUpdates(Updates); 4610 } 4611 4612 return true; 4613 } 4614 4615 // Replaces 4616 // (switch (select cond, X, Y)) on constant X, Y 4617 // with a branch - conditional if X and Y lead to distinct BBs, 4618 // unconditional otherwise. 4619 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI, 4620 SelectInst *Select) { 4621 // Check for constant integer values in the select. 4622 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 4623 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 4624 if (!TrueVal || !FalseVal) 4625 return false; 4626 4627 // Find the relevant condition and destinations. 4628 Value *Condition = Select->getCondition(); 4629 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor(); 4630 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor(); 4631 4632 // Get weight for TrueBB and FalseBB. 4633 uint32_t TrueWeight = 0, FalseWeight = 0; 4634 SmallVector<uint64_t, 8> Weights; 4635 bool HasWeights = hasBranchWeightMD(*SI); 4636 if (HasWeights) { 4637 GetBranchWeights(SI, Weights); 4638 if (Weights.size() == 1 + SI->getNumCases()) { 4639 TrueWeight = 4640 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()]; 4641 FalseWeight = 4642 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()]; 4643 } 4644 } 4645 4646 // Perform the actual simplification. 4647 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight, 4648 FalseWeight); 4649 } 4650 4651 // Replaces 4652 // (indirectbr (select cond, blockaddress(@fn, BlockA), 4653 // blockaddress(@fn, BlockB))) 4654 // with 4655 // (br cond, BlockA, BlockB). 4656 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, 4657 SelectInst *SI) { 4658 // Check that both operands of the select are block addresses. 4659 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 4660 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 4661 if (!TBA || !FBA) 4662 return false; 4663 4664 // Extract the actual blocks. 4665 BasicBlock *TrueBB = TBA->getBasicBlock(); 4666 BasicBlock *FalseBB = FBA->getBasicBlock(); 4667 4668 // Perform the actual simplification. 4669 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0, 4670 0); 4671 } 4672 4673 /// This is called when we find an icmp instruction 4674 /// (a seteq/setne with a constant) as the only instruction in a 4675 /// block that ends with an uncond branch. We are looking for a very specific 4676 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 4677 /// this case, we merge the first two "or's of icmp" into a switch, but then the 4678 /// default value goes to an uncond block with a seteq in it, we get something 4679 /// like: 4680 /// 4681 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 4682 /// DEFAULT: 4683 /// %tmp = icmp eq i8 %A, 92 4684 /// br label %end 4685 /// end: 4686 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 4687 /// 4688 /// We prefer to split the edge to 'end' so that there is a true/false entry to 4689 /// the PHI, merging the third icmp into the switch. 4690 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt( 4691 ICmpInst *ICI, IRBuilder<> &Builder) { 4692 BasicBlock *BB = ICI->getParent(); 4693 4694 // If the block has any PHIs in it or the icmp has multiple uses, it is too 4695 // complex. 4696 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) 4697 return false; 4698 4699 Value *V = ICI->getOperand(0); 4700 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 4701 4702 // The pattern we're looking for is where our only predecessor is a switch on 4703 // 'V' and this block is the default case for the switch. In this case we can 4704 // fold the compared value into the switch to simplify things. 4705 BasicBlock *Pred = BB->getSinglePredecessor(); 4706 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) 4707 return false; 4708 4709 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 4710 if (SI->getCondition() != V) 4711 return false; 4712 4713 // If BB is reachable on a non-default case, then we simply know the value of 4714 // V in this block. Substitute it and constant fold the icmp instruction 4715 // away. 4716 if (SI->getDefaultDest() != BB) { 4717 ConstantInt *VVal = SI->findCaseDest(BB); 4718 assert(VVal && "Should have a unique destination value"); 4719 ICI->setOperand(0, VVal); 4720 4721 if (Value *V = simplifyInstruction(ICI, {DL, ICI})) { 4722 ICI->replaceAllUsesWith(V); 4723 ICI->eraseFromParent(); 4724 } 4725 // BB is now empty, so it is likely to simplify away. 4726 return requestResimplify(); 4727 } 4728 4729 // Ok, the block is reachable from the default dest. If the constant we're 4730 // comparing exists in one of the other edges, then we can constant fold ICI 4731 // and zap it. 4732 if (SI->findCaseValue(Cst) != SI->case_default()) { 4733 Value *V; 4734 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 4735 V = ConstantInt::getFalse(BB->getContext()); 4736 else 4737 V = ConstantInt::getTrue(BB->getContext()); 4738 4739 ICI->replaceAllUsesWith(V); 4740 ICI->eraseFromParent(); 4741 // BB is now empty, so it is likely to simplify away. 4742 return requestResimplify(); 4743 } 4744 4745 // The use of the icmp has to be in the 'end' block, by the only PHI node in 4746 // the block. 4747 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 4748 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back()); 4749 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() || 4750 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 4751 return false; 4752 4753 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 4754 // true in the PHI. 4755 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 4756 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 4757 4758 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 4759 std::swap(DefaultCst, NewCst); 4760 4761 // Replace ICI (which is used by the PHI for the default value) with true or 4762 // false depending on if it is EQ or NE. 4763 ICI->replaceAllUsesWith(DefaultCst); 4764 ICI->eraseFromParent(); 4765 4766 SmallVector<DominatorTree::UpdateType, 2> Updates; 4767 4768 // Okay, the switch goes to this block on a default value. Add an edge from 4769 // the switch to the merge point on the compared value. 4770 BasicBlock *NewBB = 4771 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB); 4772 { 4773 SwitchInstProfUpdateWrapper SIW(*SI); 4774 auto W0 = SIW.getSuccessorWeight(0); 4775 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW; 4776 if (W0) { 4777 NewW = ((uint64_t(*W0) + 1) >> 1); 4778 SIW.setSuccessorWeight(0, *NewW); 4779 } 4780 SIW.addCase(Cst, NewBB, NewW); 4781 if (DTU) 4782 Updates.push_back({DominatorTree::Insert, Pred, NewBB}); 4783 } 4784 4785 // NewBB branches to the phi block, add the uncond branch and the phi entry. 4786 Builder.SetInsertPoint(NewBB); 4787 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 4788 Builder.CreateBr(SuccBlock); 4789 PHIUse->addIncoming(NewCst, NewBB); 4790 if (DTU) { 4791 Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock}); 4792 DTU->applyUpdates(Updates); 4793 } 4794 return true; 4795 } 4796 4797 /// The specified branch is a conditional branch. 4798 /// Check to see if it is branching on an or/and chain of icmp instructions, and 4799 /// fold it into a switch instruction if so. 4800 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI, 4801 IRBuilder<> &Builder, 4802 const DataLayout &DL) { 4803 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 4804 if (!Cond) 4805 return false; 4806 4807 // Change br (X == 0 | X == 1), T, F into a switch instruction. 4808 // If this is a bunch of seteq's or'd together, or if it's a bunch of 4809 // 'setne's and'ed together, collect them. 4810 4811 // Try to gather values from a chain of and/or to be turned into a switch 4812 ConstantComparesGatherer ConstantCompare(Cond, DL); 4813 // Unpack the result 4814 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals; 4815 Value *CompVal = ConstantCompare.CompValue; 4816 unsigned UsedICmps = ConstantCompare.UsedICmps; 4817 Value *ExtraCase = ConstantCompare.Extra; 4818 4819 // If we didn't have a multiply compared value, fail. 4820 if (!CompVal) 4821 return false; 4822 4823 // Avoid turning single icmps into a switch. 4824 if (UsedICmps <= 1) 4825 return false; 4826 4827 bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value())); 4828 4829 // There might be duplicate constants in the list, which the switch 4830 // instruction can't handle, remove them now. 4831 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 4832 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 4833 4834 // If Extra was used, we require at least two switch values to do the 4835 // transformation. A switch with one value is just a conditional branch. 4836 if (ExtraCase && Values.size() < 2) 4837 return false; 4838 4839 // TODO: Preserve branch weight metadata, similarly to how 4840 // FoldValueComparisonIntoPredecessors preserves it. 4841 4842 // Figure out which block is which destination. 4843 BasicBlock *DefaultBB = BI->getSuccessor(1); 4844 BasicBlock *EdgeBB = BI->getSuccessor(0); 4845 if (!TrueWhenEqual) 4846 std::swap(DefaultBB, EdgeBB); 4847 4848 BasicBlock *BB = BI->getParent(); 4849 4850 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 4851 << " cases into SWITCH. BB is:\n" 4852 << *BB); 4853 4854 SmallVector<DominatorTree::UpdateType, 2> Updates; 4855 4856 // If there are any extra values that couldn't be folded into the switch 4857 // then we evaluate them with an explicit branch first. Split the block 4858 // right before the condbr to handle it. 4859 if (ExtraCase) { 4860 BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr, 4861 /*MSSAU=*/nullptr, "switch.early.test"); 4862 4863 // Remove the uncond branch added to the old block. 4864 Instruction *OldTI = BB->getTerminator(); 4865 Builder.SetInsertPoint(OldTI); 4866 4867 // There can be an unintended UB if extra values are Poison. Before the 4868 // transformation, extra values may not be evaluated according to the 4869 // condition, and it will not raise UB. But after transformation, we are 4870 // evaluating extra values before checking the condition, and it will raise 4871 // UB. It can be solved by adding freeze instruction to extra values. 4872 AssumptionCache *AC = Options.AC; 4873 4874 if (!isGuaranteedNotToBeUndefOrPoison(ExtraCase, AC, BI, nullptr)) 4875 ExtraCase = Builder.CreateFreeze(ExtraCase); 4876 4877 if (TrueWhenEqual) 4878 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 4879 else 4880 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 4881 4882 OldTI->eraseFromParent(); 4883 4884 if (DTU) 4885 Updates.push_back({DominatorTree::Insert, BB, EdgeBB}); 4886 4887 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 4888 // for the edge we just added. 4889 AddPredecessorToBlock(EdgeBB, BB, NewBB); 4890 4891 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 4892 << "\nEXTRABB = " << *BB); 4893 BB = NewBB; 4894 } 4895 4896 Builder.SetInsertPoint(BI); 4897 // Convert pointer to int before we switch. 4898 if (CompVal->getType()->isPointerTy()) { 4899 CompVal = Builder.CreatePtrToInt( 4900 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr"); 4901 } 4902 4903 // Create the new switch instruction now. 4904 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 4905 4906 // Add all of the 'cases' to the switch instruction. 4907 for (unsigned i = 0, e = Values.size(); i != e; ++i) 4908 New->addCase(Values[i], EdgeBB); 4909 4910 // We added edges from PI to the EdgeBB. As such, if there were any 4911 // PHI nodes in EdgeBB, they need entries to be added corresponding to 4912 // the number of edges added. 4913 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) { 4914 PHINode *PN = cast<PHINode>(BBI); 4915 Value *InVal = PN->getIncomingValueForBlock(BB); 4916 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i) 4917 PN->addIncoming(InVal, BB); 4918 } 4919 4920 // Erase the old branch instruction. 4921 EraseTerminatorAndDCECond(BI); 4922 if (DTU) 4923 DTU->applyUpdates(Updates); 4924 4925 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 4926 return true; 4927 } 4928 4929 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 4930 if (isa<PHINode>(RI->getValue())) 4931 return simplifyCommonResume(RI); 4932 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) && 4933 RI->getValue() == RI->getParent()->getFirstNonPHI()) 4934 // The resume must unwind the exception that caused control to branch here. 4935 return simplifySingleResume(RI); 4936 4937 return false; 4938 } 4939 4940 // Check if cleanup block is empty 4941 static bool isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R) { 4942 for (Instruction &I : R) { 4943 auto *II = dyn_cast<IntrinsicInst>(&I); 4944 if (!II) 4945 return false; 4946 4947 Intrinsic::ID IntrinsicID = II->getIntrinsicID(); 4948 switch (IntrinsicID) { 4949 case Intrinsic::dbg_declare: 4950 case Intrinsic::dbg_value: 4951 case Intrinsic::dbg_label: 4952 case Intrinsic::lifetime_end: 4953 break; 4954 default: 4955 return false; 4956 } 4957 } 4958 return true; 4959 } 4960 4961 // Simplify resume that is shared by several landing pads (phi of landing pad). 4962 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) { 4963 BasicBlock *BB = RI->getParent(); 4964 4965 // Check that there are no other instructions except for debug and lifetime 4966 // intrinsics between the phi's and resume instruction. 4967 if (!isCleanupBlockEmpty( 4968 make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator()))) 4969 return false; 4970 4971 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks; 4972 auto *PhiLPInst = cast<PHINode>(RI->getValue()); 4973 4974 // Check incoming blocks to see if any of them are trivial. 4975 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End; 4976 Idx++) { 4977 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx); 4978 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx); 4979 4980 // If the block has other successors, we can not delete it because 4981 // it has other dependents. 4982 if (IncomingBB->getUniqueSuccessor() != BB) 4983 continue; 4984 4985 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI()); 4986 // Not the landing pad that caused the control to branch here. 4987 if (IncomingValue != LandingPad) 4988 continue; 4989 4990 if (isCleanupBlockEmpty( 4991 make_range(LandingPad->getNextNode(), IncomingBB->getTerminator()))) 4992 TrivialUnwindBlocks.insert(IncomingBB); 4993 } 4994 4995 // If no trivial unwind blocks, don't do any simplifications. 4996 if (TrivialUnwindBlocks.empty()) 4997 return false; 4998 4999 // Turn all invokes that unwind here into calls. 5000 for (auto *TrivialBB : TrivialUnwindBlocks) { 5001 // Blocks that will be simplified should be removed from the phi node. 5002 // Note there could be multiple edges to the resume block, and we need 5003 // to remove them all. 5004 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1) 5005 BB->removePredecessor(TrivialBB, true); 5006 5007 for (BasicBlock *Pred : 5008 llvm::make_early_inc_range(predecessors(TrivialBB))) { 5009 removeUnwindEdge(Pred, DTU); 5010 ++NumInvokes; 5011 } 5012 5013 // In each SimplifyCFG run, only the current processed block can be erased. 5014 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead 5015 // of erasing TrivialBB, we only remove the branch to the common resume 5016 // block so that we can later erase the resume block since it has no 5017 // predecessors. 5018 TrivialBB->getTerminator()->eraseFromParent(); 5019 new UnreachableInst(RI->getContext(), TrivialBB); 5020 if (DTU) 5021 DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}}); 5022 } 5023 5024 // Delete the resume block if all its predecessors have been removed. 5025 if (pred_empty(BB)) 5026 DeleteDeadBlock(BB, DTU); 5027 5028 return !TrivialUnwindBlocks.empty(); 5029 } 5030 5031 // Simplify resume that is only used by a single (non-phi) landing pad. 5032 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) { 5033 BasicBlock *BB = RI->getParent(); 5034 auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI()); 5035 assert(RI->getValue() == LPInst && 5036 "Resume must unwind the exception that caused control to here"); 5037 5038 // Check that there are no other instructions except for debug intrinsics. 5039 if (!isCleanupBlockEmpty( 5040 make_range<Instruction *>(LPInst->getNextNode(), RI))) 5041 return false; 5042 5043 // Turn all invokes that unwind here into calls and delete the basic block. 5044 for (BasicBlock *Pred : llvm::make_early_inc_range(predecessors(BB))) { 5045 removeUnwindEdge(Pred, DTU); 5046 ++NumInvokes; 5047 } 5048 5049 // The landingpad is now unreachable. Zap it. 5050 DeleteDeadBlock(BB, DTU); 5051 return true; 5052 } 5053 5054 static bool removeEmptyCleanup(CleanupReturnInst *RI, DomTreeUpdater *DTU) { 5055 // If this is a trivial cleanup pad that executes no instructions, it can be 5056 // eliminated. If the cleanup pad continues to the caller, any predecessor 5057 // that is an EH pad will be updated to continue to the caller and any 5058 // predecessor that terminates with an invoke instruction will have its invoke 5059 // instruction converted to a call instruction. If the cleanup pad being 5060 // simplified does not continue to the caller, each predecessor will be 5061 // updated to continue to the unwind destination of the cleanup pad being 5062 // simplified. 5063 BasicBlock *BB = RI->getParent(); 5064 CleanupPadInst *CPInst = RI->getCleanupPad(); 5065 if (CPInst->getParent() != BB) 5066 // This isn't an empty cleanup. 5067 return false; 5068 5069 // We cannot kill the pad if it has multiple uses. This typically arises 5070 // from unreachable basic blocks. 5071 if (!CPInst->hasOneUse()) 5072 return false; 5073 5074 // Check that there are no other instructions except for benign intrinsics. 5075 if (!isCleanupBlockEmpty( 5076 make_range<Instruction *>(CPInst->getNextNode(), RI))) 5077 return false; 5078 5079 // If the cleanup return we are simplifying unwinds to the caller, this will 5080 // set UnwindDest to nullptr. 5081 BasicBlock *UnwindDest = RI->getUnwindDest(); 5082 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr; 5083 5084 // We're about to remove BB from the control flow. Before we do, sink any 5085 // PHINodes into the unwind destination. Doing this before changing the 5086 // control flow avoids some potentially slow checks, since we can currently 5087 // be certain that UnwindDest and BB have no common predecessors (since they 5088 // are both EH pads). 5089 if (UnwindDest) { 5090 // First, go through the PHI nodes in UnwindDest and update any nodes that 5091 // reference the block we are removing 5092 for (PHINode &DestPN : UnwindDest->phis()) { 5093 int Idx = DestPN.getBasicBlockIndex(BB); 5094 // Since BB unwinds to UnwindDest, it has to be in the PHI node. 5095 assert(Idx != -1); 5096 // This PHI node has an incoming value that corresponds to a control 5097 // path through the cleanup pad we are removing. If the incoming 5098 // value is in the cleanup pad, it must be a PHINode (because we 5099 // verified above that the block is otherwise empty). Otherwise, the 5100 // value is either a constant or a value that dominates the cleanup 5101 // pad being removed. 5102 // 5103 // Because BB and UnwindDest are both EH pads, all of their 5104 // predecessors must unwind to these blocks, and since no instruction 5105 // can have multiple unwind destinations, there will be no overlap in 5106 // incoming blocks between SrcPN and DestPN. 5107 Value *SrcVal = DestPN.getIncomingValue(Idx); 5108 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal); 5109 5110 bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB; 5111 for (auto *Pred : predecessors(BB)) { 5112 Value *Incoming = 5113 NeedPHITranslation ? SrcPN->getIncomingValueForBlock(Pred) : SrcVal; 5114 DestPN.addIncoming(Incoming, Pred); 5115 } 5116 } 5117 5118 // Sink any remaining PHI nodes directly into UnwindDest. 5119 Instruction *InsertPt = DestEHPad; 5120 for (PHINode &PN : make_early_inc_range(BB->phis())) { 5121 if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB)) 5122 // If the PHI node has no uses or all of its uses are in this basic 5123 // block (meaning they are debug or lifetime intrinsics), just leave 5124 // it. It will be erased when we erase BB below. 5125 continue; 5126 5127 // Otherwise, sink this PHI node into UnwindDest. 5128 // Any predecessors to UnwindDest which are not already represented 5129 // must be back edges which inherit the value from the path through 5130 // BB. In this case, the PHI value must reference itself. 5131 for (auto *pred : predecessors(UnwindDest)) 5132 if (pred != BB) 5133 PN.addIncoming(&PN, pred); 5134 PN.moveBefore(InsertPt); 5135 // Also, add a dummy incoming value for the original BB itself, 5136 // so that the PHI is well-formed until we drop said predecessor. 5137 PN.addIncoming(PoisonValue::get(PN.getType()), BB); 5138 } 5139 } 5140 5141 std::vector<DominatorTree::UpdateType> Updates; 5142 5143 // We use make_early_inc_range here because we will remove all predecessors. 5144 for (BasicBlock *PredBB : llvm::make_early_inc_range(predecessors(BB))) { 5145 if (UnwindDest == nullptr) { 5146 if (DTU) { 5147 DTU->applyUpdates(Updates); 5148 Updates.clear(); 5149 } 5150 removeUnwindEdge(PredBB, DTU); 5151 ++NumInvokes; 5152 } else { 5153 BB->removePredecessor(PredBB); 5154 Instruction *TI = PredBB->getTerminator(); 5155 TI->replaceUsesOfWith(BB, UnwindDest); 5156 if (DTU) { 5157 Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest}); 5158 Updates.push_back({DominatorTree::Delete, PredBB, BB}); 5159 } 5160 } 5161 } 5162 5163 if (DTU) 5164 DTU->applyUpdates(Updates); 5165 5166 DeleteDeadBlock(BB, DTU); 5167 5168 return true; 5169 } 5170 5171 // Try to merge two cleanuppads together. 5172 static bool mergeCleanupPad(CleanupReturnInst *RI) { 5173 // Skip any cleanuprets which unwind to caller, there is nothing to merge 5174 // with. 5175 BasicBlock *UnwindDest = RI->getUnwindDest(); 5176 if (!UnwindDest) 5177 return false; 5178 5179 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't 5180 // be safe to merge without code duplication. 5181 if (UnwindDest->getSinglePredecessor() != RI->getParent()) 5182 return false; 5183 5184 // Verify that our cleanuppad's unwind destination is another cleanuppad. 5185 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front()); 5186 if (!SuccessorCleanupPad) 5187 return false; 5188 5189 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad(); 5190 // Replace any uses of the successor cleanupad with the predecessor pad 5191 // The only cleanuppad uses should be this cleanupret, it's cleanupret and 5192 // funclet bundle operands. 5193 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad); 5194 // Remove the old cleanuppad. 5195 SuccessorCleanupPad->eraseFromParent(); 5196 // Now, we simply replace the cleanupret with a branch to the unwind 5197 // destination. 5198 BranchInst::Create(UnwindDest, RI->getParent()); 5199 RI->eraseFromParent(); 5200 5201 return true; 5202 } 5203 5204 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) { 5205 // It is possible to transiantly have an undef cleanuppad operand because we 5206 // have deleted some, but not all, dead blocks. 5207 // Eventually, this block will be deleted. 5208 if (isa<UndefValue>(RI->getOperand(0))) 5209 return false; 5210 5211 if (mergeCleanupPad(RI)) 5212 return true; 5213 5214 if (removeEmptyCleanup(RI, DTU)) 5215 return true; 5216 5217 return false; 5218 } 5219 5220 // WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()! 5221 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) { 5222 BasicBlock *BB = UI->getParent(); 5223 5224 bool Changed = false; 5225 5226 // Ensure that any debug-info records that used to occur after the Unreachable 5227 // are moved to in front of it -- otherwise they'll "dangle" at the end of 5228 // the block. 5229 BB->flushTerminatorDbgValues(); 5230 5231 // Debug-info records on the unreachable inst itself should be deleted, as 5232 // below we delete everything past the final executable instruction. 5233 UI->dropDbgValues(); 5234 5235 // If there are any instructions immediately before the unreachable that can 5236 // be removed, do so. 5237 while (UI->getIterator() != BB->begin()) { 5238 BasicBlock::iterator BBI = UI->getIterator(); 5239 --BBI; 5240 5241 if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI)) 5242 break; // Can not drop any more instructions. We're done here. 5243 // Otherwise, this instruction can be freely erased, 5244 // even if it is not side-effect free. 5245 5246 // Note that deleting EH's here is in fact okay, although it involves a bit 5247 // of subtle reasoning. If this inst is an EH, all the predecessors of this 5248 // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn, 5249 // and we can therefore guarantee this block will be erased. 5250 5251 // If we're deleting this, we're deleting any subsequent dbg.values, so 5252 // delete DPValue records of variable information. 5253 BBI->dropDbgValues(); 5254 5255 // Delete this instruction (any uses are guaranteed to be dead) 5256 BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType())); 5257 BBI->eraseFromParent(); 5258 Changed = true; 5259 } 5260 5261 // If the unreachable instruction is the first in the block, take a gander 5262 // at all of the predecessors of this instruction, and simplify them. 5263 if (&BB->front() != UI) 5264 return Changed; 5265 5266 std::vector<DominatorTree::UpdateType> Updates; 5267 5268 SmallSetVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB)); 5269 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 5270 auto *Predecessor = Preds[i]; 5271 Instruction *TI = Predecessor->getTerminator(); 5272 IRBuilder<> Builder(TI); 5273 if (auto *BI = dyn_cast<BranchInst>(TI)) { 5274 // We could either have a proper unconditional branch, 5275 // or a degenerate conditional branch with matching destinations. 5276 if (all_of(BI->successors(), 5277 [BB](auto *Successor) { return Successor == BB; })) { 5278 new UnreachableInst(TI->getContext(), TI); 5279 TI->eraseFromParent(); 5280 Changed = true; 5281 } else { 5282 assert(BI->isConditional() && "Can't get here with an uncond branch."); 5283 Value* Cond = BI->getCondition(); 5284 assert(BI->getSuccessor(0) != BI->getSuccessor(1) && 5285 "The destinations are guaranteed to be different here."); 5286 CallInst *Assumption; 5287 if (BI->getSuccessor(0) == BB) { 5288 Assumption = Builder.CreateAssumption(Builder.CreateNot(Cond)); 5289 Builder.CreateBr(BI->getSuccessor(1)); 5290 } else { 5291 assert(BI->getSuccessor(1) == BB && "Incorrect CFG"); 5292 Assumption = Builder.CreateAssumption(Cond); 5293 Builder.CreateBr(BI->getSuccessor(0)); 5294 } 5295 if (Options.AC) 5296 Options.AC->registerAssumption(cast<AssumeInst>(Assumption)); 5297 5298 EraseTerminatorAndDCECond(BI); 5299 Changed = true; 5300 } 5301 if (DTU) 5302 Updates.push_back({DominatorTree::Delete, Predecessor, BB}); 5303 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) { 5304 SwitchInstProfUpdateWrapper SU(*SI); 5305 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) { 5306 if (i->getCaseSuccessor() != BB) { 5307 ++i; 5308 continue; 5309 } 5310 BB->removePredecessor(SU->getParent()); 5311 i = SU.removeCase(i); 5312 e = SU->case_end(); 5313 Changed = true; 5314 } 5315 // Note that the default destination can't be removed! 5316 if (DTU && SI->getDefaultDest() != BB) 5317 Updates.push_back({DominatorTree::Delete, Predecessor, BB}); 5318 } else if (auto *II = dyn_cast<InvokeInst>(TI)) { 5319 if (II->getUnwindDest() == BB) { 5320 if (DTU) { 5321 DTU->applyUpdates(Updates); 5322 Updates.clear(); 5323 } 5324 auto *CI = cast<CallInst>(removeUnwindEdge(TI->getParent(), DTU)); 5325 if (!CI->doesNotThrow()) 5326 CI->setDoesNotThrow(); 5327 Changed = true; 5328 } 5329 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) { 5330 if (CSI->getUnwindDest() == BB) { 5331 if (DTU) { 5332 DTU->applyUpdates(Updates); 5333 Updates.clear(); 5334 } 5335 removeUnwindEdge(TI->getParent(), DTU); 5336 Changed = true; 5337 continue; 5338 } 5339 5340 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(), 5341 E = CSI->handler_end(); 5342 I != E; ++I) { 5343 if (*I == BB) { 5344 CSI->removeHandler(I); 5345 --I; 5346 --E; 5347 Changed = true; 5348 } 5349 } 5350 if (DTU) 5351 Updates.push_back({DominatorTree::Delete, Predecessor, BB}); 5352 if (CSI->getNumHandlers() == 0) { 5353 if (CSI->hasUnwindDest()) { 5354 // Redirect all predecessors of the block containing CatchSwitchInst 5355 // to instead branch to the CatchSwitchInst's unwind destination. 5356 if (DTU) { 5357 for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) { 5358 Updates.push_back({DominatorTree::Insert, 5359 PredecessorOfPredecessor, 5360 CSI->getUnwindDest()}); 5361 Updates.push_back({DominatorTree::Delete, 5362 PredecessorOfPredecessor, Predecessor}); 5363 } 5364 } 5365 Predecessor->replaceAllUsesWith(CSI->getUnwindDest()); 5366 } else { 5367 // Rewrite all preds to unwind to caller (or from invoke to call). 5368 if (DTU) { 5369 DTU->applyUpdates(Updates); 5370 Updates.clear(); 5371 } 5372 SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor)); 5373 for (BasicBlock *EHPred : EHPreds) 5374 removeUnwindEdge(EHPred, DTU); 5375 } 5376 // The catchswitch is no longer reachable. 5377 new UnreachableInst(CSI->getContext(), CSI); 5378 CSI->eraseFromParent(); 5379 Changed = true; 5380 } 5381 } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) { 5382 (void)CRI; 5383 assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB && 5384 "Expected to always have an unwind to BB."); 5385 if (DTU) 5386 Updates.push_back({DominatorTree::Delete, Predecessor, BB}); 5387 new UnreachableInst(TI->getContext(), TI); 5388 TI->eraseFromParent(); 5389 Changed = true; 5390 } 5391 } 5392 5393 if (DTU) 5394 DTU->applyUpdates(Updates); 5395 5396 // If this block is now dead, remove it. 5397 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) { 5398 DeleteDeadBlock(BB, DTU); 5399 return true; 5400 } 5401 5402 return Changed; 5403 } 5404 5405 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) { 5406 assert(Cases.size() >= 1); 5407 5408 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 5409 for (size_t I = 1, E = Cases.size(); I != E; ++I) { 5410 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1) 5411 return false; 5412 } 5413 return true; 5414 } 5415 5416 static void createUnreachableSwitchDefault(SwitchInst *Switch, 5417 DomTreeUpdater *DTU) { 5418 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n"); 5419 auto *BB = Switch->getParent(); 5420 auto *OrigDefaultBlock = Switch->getDefaultDest(); 5421 OrigDefaultBlock->removePredecessor(BB); 5422 BasicBlock *NewDefaultBlock = BasicBlock::Create( 5423 BB->getContext(), BB->getName() + ".unreachabledefault", BB->getParent(), 5424 OrigDefaultBlock); 5425 new UnreachableInst(Switch->getContext(), NewDefaultBlock); 5426 Switch->setDefaultDest(&*NewDefaultBlock); 5427 if (DTU) { 5428 SmallVector<DominatorTree::UpdateType, 2> Updates; 5429 Updates.push_back({DominatorTree::Insert, BB, &*NewDefaultBlock}); 5430 if (!is_contained(successors(BB), OrigDefaultBlock)) 5431 Updates.push_back({DominatorTree::Delete, BB, &*OrigDefaultBlock}); 5432 DTU->applyUpdates(Updates); 5433 } 5434 } 5435 5436 /// Turn a switch into an integer range comparison and branch. 5437 /// Switches with more than 2 destinations are ignored. 5438 /// Switches with 1 destination are also ignored. 5439 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI, 5440 IRBuilder<> &Builder) { 5441 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 5442 5443 bool HasDefault = 5444 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg()); 5445 5446 auto *BB = SI->getParent(); 5447 5448 // Partition the cases into two sets with different destinations. 5449 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr; 5450 BasicBlock *DestB = nullptr; 5451 SmallVector<ConstantInt *, 16> CasesA; 5452 SmallVector<ConstantInt *, 16> CasesB; 5453 5454 for (auto Case : SI->cases()) { 5455 BasicBlock *Dest = Case.getCaseSuccessor(); 5456 if (!DestA) 5457 DestA = Dest; 5458 if (Dest == DestA) { 5459 CasesA.push_back(Case.getCaseValue()); 5460 continue; 5461 } 5462 if (!DestB) 5463 DestB = Dest; 5464 if (Dest == DestB) { 5465 CasesB.push_back(Case.getCaseValue()); 5466 continue; 5467 } 5468 return false; // More than two destinations. 5469 } 5470 if (!DestB) 5471 return false; // All destinations are the same and the default is unreachable 5472 5473 assert(DestA && DestB && 5474 "Single-destination switch should have been folded."); 5475 assert(DestA != DestB); 5476 assert(DestB != SI->getDefaultDest()); 5477 assert(!CasesB.empty() && "There must be non-default cases."); 5478 assert(!CasesA.empty() || HasDefault); 5479 5480 // Figure out if one of the sets of cases form a contiguous range. 5481 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr; 5482 BasicBlock *ContiguousDest = nullptr; 5483 BasicBlock *OtherDest = nullptr; 5484 if (!CasesA.empty() && CasesAreContiguous(CasesA)) { 5485 ContiguousCases = &CasesA; 5486 ContiguousDest = DestA; 5487 OtherDest = DestB; 5488 } else if (CasesAreContiguous(CasesB)) { 5489 ContiguousCases = &CasesB; 5490 ContiguousDest = DestB; 5491 OtherDest = DestA; 5492 } else 5493 return false; 5494 5495 // Start building the compare and branch. 5496 5497 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back()); 5498 Constant *NumCases = 5499 ConstantInt::get(Offset->getType(), ContiguousCases->size()); 5500 5501 Value *Sub = SI->getCondition(); 5502 if (!Offset->isNullValue()) 5503 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off"); 5504 5505 Value *Cmp; 5506 // If NumCases overflowed, then all possible values jump to the successor. 5507 if (NumCases->isNullValue() && !ContiguousCases->empty()) 5508 Cmp = ConstantInt::getTrue(SI->getContext()); 5509 else 5510 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 5511 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest); 5512 5513 // Update weight for the newly-created conditional branch. 5514 if (hasBranchWeightMD(*SI)) { 5515 SmallVector<uint64_t, 8> Weights; 5516 GetBranchWeights(SI, Weights); 5517 if (Weights.size() == 1 + SI->getNumCases()) { 5518 uint64_t TrueWeight = 0; 5519 uint64_t FalseWeight = 0; 5520 for (size_t I = 0, E = Weights.size(); I != E; ++I) { 5521 if (SI->getSuccessor(I) == ContiguousDest) 5522 TrueWeight += Weights[I]; 5523 else 5524 FalseWeight += Weights[I]; 5525 } 5526 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) { 5527 TrueWeight /= 2; 5528 FalseWeight /= 2; 5529 } 5530 setBranchWeights(NewBI, TrueWeight, FalseWeight); 5531 } 5532 } 5533 5534 // Prune obsolete incoming values off the successors' PHI nodes. 5535 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) { 5536 unsigned PreviousEdges = ContiguousCases->size(); 5537 if (ContiguousDest == SI->getDefaultDest()) 5538 ++PreviousEdges; 5539 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I) 5540 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 5541 } 5542 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) { 5543 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size(); 5544 if (OtherDest == SI->getDefaultDest()) 5545 ++PreviousEdges; 5546 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I) 5547 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 5548 } 5549 5550 // Clean up the default block - it may have phis or other instructions before 5551 // the unreachable terminator. 5552 if (!HasDefault) 5553 createUnreachableSwitchDefault(SI, DTU); 5554 5555 auto *UnreachableDefault = SI->getDefaultDest(); 5556 5557 // Drop the switch. 5558 SI->eraseFromParent(); 5559 5560 if (!HasDefault && DTU) 5561 DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}}); 5562 5563 return true; 5564 } 5565 5566 /// Compute masked bits for the condition of a switch 5567 /// and use it to remove dead cases. 5568 static bool eliminateDeadSwitchCases(SwitchInst *SI, DomTreeUpdater *DTU, 5569 AssumptionCache *AC, 5570 const DataLayout &DL) { 5571 Value *Cond = SI->getCondition(); 5572 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI); 5573 5574 // We can also eliminate cases by determining that their values are outside of 5575 // the limited range of the condition based on how many significant (non-sign) 5576 // bits are in the condition value. 5577 unsigned MaxSignificantBitsInCond = 5578 ComputeMaxSignificantBits(Cond, DL, 0, AC, SI); 5579 5580 // Gather dead cases. 5581 SmallVector<ConstantInt *, 8> DeadCases; 5582 SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases; 5583 SmallVector<BasicBlock *, 8> UniqueSuccessors; 5584 for (const auto &Case : SI->cases()) { 5585 auto *Successor = Case.getCaseSuccessor(); 5586 if (DTU) { 5587 if (!NumPerSuccessorCases.count(Successor)) 5588 UniqueSuccessors.push_back(Successor); 5589 ++NumPerSuccessorCases[Successor]; 5590 } 5591 const APInt &CaseVal = Case.getCaseValue()->getValue(); 5592 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) || 5593 (CaseVal.getSignificantBits() > MaxSignificantBitsInCond)) { 5594 DeadCases.push_back(Case.getCaseValue()); 5595 if (DTU) 5596 --NumPerSuccessorCases[Successor]; 5597 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal 5598 << " is dead.\n"); 5599 } 5600 } 5601 5602 // If we can prove that the cases must cover all possible values, the 5603 // default destination becomes dead and we can remove it. If we know some 5604 // of the bits in the value, we can use that to more precisely compute the 5605 // number of possible unique case values. 5606 bool HasDefault = 5607 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg()); 5608 const unsigned NumUnknownBits = 5609 Known.getBitWidth() - (Known.Zero | Known.One).popcount(); 5610 assert(NumUnknownBits <= Known.getBitWidth()); 5611 if (HasDefault && DeadCases.empty() && 5612 NumUnknownBits < 64 /* avoid overflow */ && 5613 SI->getNumCases() == (1ULL << NumUnknownBits)) { 5614 createUnreachableSwitchDefault(SI, DTU); 5615 return true; 5616 } 5617 5618 if (DeadCases.empty()) 5619 return false; 5620 5621 SwitchInstProfUpdateWrapper SIW(*SI); 5622 for (ConstantInt *DeadCase : DeadCases) { 5623 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase); 5624 assert(CaseI != SI->case_default() && 5625 "Case was not found. Probably mistake in DeadCases forming."); 5626 // Prune unused values from PHI nodes. 5627 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent()); 5628 SIW.removeCase(CaseI); 5629 } 5630 5631 if (DTU) { 5632 std::vector<DominatorTree::UpdateType> Updates; 5633 for (auto *Successor : UniqueSuccessors) 5634 if (NumPerSuccessorCases[Successor] == 0) 5635 Updates.push_back({DominatorTree::Delete, SI->getParent(), Successor}); 5636 DTU->applyUpdates(Updates); 5637 } 5638 5639 return true; 5640 } 5641 5642 /// If BB would be eligible for simplification by 5643 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 5644 /// by an unconditional branch), look at the phi node for BB in the successor 5645 /// block and see if the incoming value is equal to CaseValue. If so, return 5646 /// the phi node, and set PhiIndex to BB's index in the phi node. 5647 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 5648 BasicBlock *BB, int *PhiIndex) { 5649 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 5650 return nullptr; // BB must be empty to be a candidate for simplification. 5651 if (!BB->getSinglePredecessor()) 5652 return nullptr; // BB must be dominated by the switch. 5653 5654 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 5655 if (!Branch || !Branch->isUnconditional()) 5656 return nullptr; // Terminator must be unconditional branch. 5657 5658 BasicBlock *Succ = Branch->getSuccessor(0); 5659 5660 for (PHINode &PHI : Succ->phis()) { 5661 int Idx = PHI.getBasicBlockIndex(BB); 5662 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 5663 5664 Value *InValue = PHI.getIncomingValue(Idx); 5665 if (InValue != CaseValue) 5666 continue; 5667 5668 *PhiIndex = Idx; 5669 return &PHI; 5670 } 5671 5672 return nullptr; 5673 } 5674 5675 /// Try to forward the condition of a switch instruction to a phi node 5676 /// dominated by the switch, if that would mean that some of the destination 5677 /// blocks of the switch can be folded away. Return true if a change is made. 5678 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 5679 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>; 5680 5681 ForwardingNodesMap ForwardingNodes; 5682 BasicBlock *SwitchBlock = SI->getParent(); 5683 bool Changed = false; 5684 for (const auto &Case : SI->cases()) { 5685 ConstantInt *CaseValue = Case.getCaseValue(); 5686 BasicBlock *CaseDest = Case.getCaseSuccessor(); 5687 5688 // Replace phi operands in successor blocks that are using the constant case 5689 // value rather than the switch condition variable: 5690 // switchbb: 5691 // switch i32 %x, label %default [ 5692 // i32 17, label %succ 5693 // ... 5694 // succ: 5695 // %r = phi i32 ... [ 17, %switchbb ] ... 5696 // --> 5697 // %r = phi i32 ... [ %x, %switchbb ] ... 5698 5699 for (PHINode &Phi : CaseDest->phis()) { 5700 // This only works if there is exactly 1 incoming edge from the switch to 5701 // a phi. If there is >1, that means multiple cases of the switch map to 1 5702 // value in the phi, and that phi value is not the switch condition. Thus, 5703 // this transform would not make sense (the phi would be invalid because 5704 // a phi can't have different incoming values from the same block). 5705 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock); 5706 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue && 5707 count(Phi.blocks(), SwitchBlock) == 1) { 5708 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition()); 5709 Changed = true; 5710 } 5711 } 5712 5713 // Collect phi nodes that are indirectly using this switch's case constants. 5714 int PhiIdx; 5715 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx)) 5716 ForwardingNodes[Phi].push_back(PhiIdx); 5717 } 5718 5719 for (auto &ForwardingNode : ForwardingNodes) { 5720 PHINode *Phi = ForwardingNode.first; 5721 SmallVectorImpl<int> &Indexes = ForwardingNode.second; 5722 if (Indexes.size() < 2) 5723 continue; 5724 5725 for (int Index : Indexes) 5726 Phi->setIncomingValue(Index, SI->getCondition()); 5727 Changed = true; 5728 } 5729 5730 return Changed; 5731 } 5732 5733 /// Return true if the backend will be able to handle 5734 /// initializing an array of constants like C. 5735 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) { 5736 if (C->isThreadDependent()) 5737 return false; 5738 if (C->isDLLImportDependent()) 5739 return false; 5740 5741 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) && 5742 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) && 5743 !isa<UndefValue>(C) && !isa<ConstantExpr>(C)) 5744 return false; 5745 5746 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 5747 // Pointer casts and in-bounds GEPs will not prohibit the backend from 5748 // materializing the array of constants. 5749 Constant *StrippedC = cast<Constant>(CE->stripInBoundsConstantOffsets()); 5750 if (StrippedC == C || !ValidLookupTableConstant(StrippedC, TTI)) 5751 return false; 5752 } 5753 5754 if (!TTI.shouldBuildLookupTablesForConstant(C)) 5755 return false; 5756 5757 return true; 5758 } 5759 5760 /// If V is a Constant, return it. Otherwise, try to look up 5761 /// its constant value in ConstantPool, returning 0 if it's not there. 5762 static Constant * 5763 LookupConstant(Value *V, 5764 const SmallDenseMap<Value *, Constant *> &ConstantPool) { 5765 if (Constant *C = dyn_cast<Constant>(V)) 5766 return C; 5767 return ConstantPool.lookup(V); 5768 } 5769 5770 /// Try to fold instruction I into a constant. This works for 5771 /// simple instructions such as binary operations where both operands are 5772 /// constant or can be replaced by constants from the ConstantPool. Returns the 5773 /// resulting constant on success, 0 otherwise. 5774 static Constant * 5775 ConstantFold(Instruction *I, const DataLayout &DL, 5776 const SmallDenseMap<Value *, Constant *> &ConstantPool) { 5777 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 5778 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 5779 if (!A) 5780 return nullptr; 5781 if (A->isAllOnesValue()) 5782 return LookupConstant(Select->getTrueValue(), ConstantPool); 5783 if (A->isNullValue()) 5784 return LookupConstant(Select->getFalseValue(), ConstantPool); 5785 return nullptr; 5786 } 5787 5788 SmallVector<Constant *, 4> COps; 5789 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) { 5790 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool)) 5791 COps.push_back(A); 5792 else 5793 return nullptr; 5794 } 5795 5796 return ConstantFoldInstOperands(I, COps, DL); 5797 } 5798 5799 /// Try to determine the resulting constant values in phi nodes 5800 /// at the common destination basic block, *CommonDest, for one of the case 5801 /// destionations CaseDest corresponding to value CaseVal (0 for the default 5802 /// case), of a switch instruction SI. 5803 static bool 5804 getCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest, 5805 BasicBlock **CommonDest, 5806 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res, 5807 const DataLayout &DL, const TargetTransformInfo &TTI) { 5808 // The block from which we enter the common destination. 5809 BasicBlock *Pred = SI->getParent(); 5810 5811 // If CaseDest is empty except for some side-effect free instructions through 5812 // which we can constant-propagate the CaseVal, continue to its successor. 5813 SmallDenseMap<Value *, Constant *> ConstantPool; 5814 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 5815 for (Instruction &I : CaseDest->instructionsWithoutDebug(false)) { 5816 if (I.isTerminator()) { 5817 // If the terminator is a simple branch, continue to the next block. 5818 if (I.getNumSuccessors() != 1 || I.isSpecialTerminator()) 5819 return false; 5820 Pred = CaseDest; 5821 CaseDest = I.getSuccessor(0); 5822 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) { 5823 // Instruction is side-effect free and constant. 5824 5825 // If the instruction has uses outside this block or a phi node slot for 5826 // the block, it is not safe to bypass the instruction since it would then 5827 // no longer dominate all its uses. 5828 for (auto &Use : I.uses()) { 5829 User *User = Use.getUser(); 5830 if (Instruction *I = dyn_cast<Instruction>(User)) 5831 if (I->getParent() == CaseDest) 5832 continue; 5833 if (PHINode *Phi = dyn_cast<PHINode>(User)) 5834 if (Phi->getIncomingBlock(Use) == CaseDest) 5835 continue; 5836 return false; 5837 } 5838 5839 ConstantPool.insert(std::make_pair(&I, C)); 5840 } else { 5841 break; 5842 } 5843 } 5844 5845 // If we did not have a CommonDest before, use the current one. 5846 if (!*CommonDest) 5847 *CommonDest = CaseDest; 5848 // If the destination isn't the common one, abort. 5849 if (CaseDest != *CommonDest) 5850 return false; 5851 5852 // Get the values for this case from phi nodes in the destination block. 5853 for (PHINode &PHI : (*CommonDest)->phis()) { 5854 int Idx = PHI.getBasicBlockIndex(Pred); 5855 if (Idx == -1) 5856 continue; 5857 5858 Constant *ConstVal = 5859 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool); 5860 if (!ConstVal) 5861 return false; 5862 5863 // Be conservative about which kinds of constants we support. 5864 if (!ValidLookupTableConstant(ConstVal, TTI)) 5865 return false; 5866 5867 Res.push_back(std::make_pair(&PHI, ConstVal)); 5868 } 5869 5870 return Res.size() > 0; 5871 } 5872 5873 // Helper function used to add CaseVal to the list of cases that generate 5874 // Result. Returns the updated number of cases that generate this result. 5875 static size_t mapCaseToResult(ConstantInt *CaseVal, 5876 SwitchCaseResultVectorTy &UniqueResults, 5877 Constant *Result) { 5878 for (auto &I : UniqueResults) { 5879 if (I.first == Result) { 5880 I.second.push_back(CaseVal); 5881 return I.second.size(); 5882 } 5883 } 5884 UniqueResults.push_back( 5885 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal))); 5886 return 1; 5887 } 5888 5889 // Helper function that initializes a map containing 5890 // results for the PHI node of the common destination block for a switch 5891 // instruction. Returns false if multiple PHI nodes have been found or if 5892 // there is not a common destination block for the switch. 5893 static bool initializeUniqueCases(SwitchInst *SI, PHINode *&PHI, 5894 BasicBlock *&CommonDest, 5895 SwitchCaseResultVectorTy &UniqueResults, 5896 Constant *&DefaultResult, 5897 const DataLayout &DL, 5898 const TargetTransformInfo &TTI, 5899 uintptr_t MaxUniqueResults) { 5900 for (const auto &I : SI->cases()) { 5901 ConstantInt *CaseVal = I.getCaseValue(); 5902 5903 // Resulting value at phi nodes for this case value. 5904 SwitchCaseResultsTy Results; 5905 if (!getCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results, 5906 DL, TTI)) 5907 return false; 5908 5909 // Only one value per case is permitted. 5910 if (Results.size() > 1) 5911 return false; 5912 5913 // Add the case->result mapping to UniqueResults. 5914 const size_t NumCasesForResult = 5915 mapCaseToResult(CaseVal, UniqueResults, Results.begin()->second); 5916 5917 // Early out if there are too many cases for this result. 5918 if (NumCasesForResult > MaxSwitchCasesPerResult) 5919 return false; 5920 5921 // Early out if there are too many unique results. 5922 if (UniqueResults.size() > MaxUniqueResults) 5923 return false; 5924 5925 // Check the PHI consistency. 5926 if (!PHI) 5927 PHI = Results[0].first; 5928 else if (PHI != Results[0].first) 5929 return false; 5930 } 5931 // Find the default result value. 5932 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults; 5933 BasicBlock *DefaultDest = SI->getDefaultDest(); 5934 getCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults, 5935 DL, TTI); 5936 // If the default value is not found abort unless the default destination 5937 // is unreachable. 5938 DefaultResult = 5939 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr; 5940 if ((!DefaultResult && 5941 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()))) 5942 return false; 5943 5944 return true; 5945 } 5946 5947 // Helper function that checks if it is possible to transform a switch with only 5948 // two cases (or two cases + default) that produces a result into a select. 5949 // TODO: Handle switches with more than 2 cases that map to the same result. 5950 static Value *foldSwitchToSelect(const SwitchCaseResultVectorTy &ResultVector, 5951 Constant *DefaultResult, Value *Condition, 5952 IRBuilder<> &Builder) { 5953 // If we are selecting between only two cases transform into a simple 5954 // select or a two-way select if default is possible. 5955 // Example: 5956 // switch (a) { %0 = icmp eq i32 %a, 10 5957 // case 10: return 42; %1 = select i1 %0, i32 42, i32 4 5958 // case 20: return 2; ----> %2 = icmp eq i32 %a, 20 5959 // default: return 4; %3 = select i1 %2, i32 2, i32 %1 5960 // } 5961 if (ResultVector.size() == 2 && ResultVector[0].second.size() == 1 && 5962 ResultVector[1].second.size() == 1) { 5963 ConstantInt *FirstCase = ResultVector[0].second[0]; 5964 ConstantInt *SecondCase = ResultVector[1].second[0]; 5965 Value *SelectValue = ResultVector[1].first; 5966 if (DefaultResult) { 5967 Value *ValueCompare = 5968 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp"); 5969 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first, 5970 DefaultResult, "switch.select"); 5971 } 5972 Value *ValueCompare = 5973 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp"); 5974 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, 5975 SelectValue, "switch.select"); 5976 } 5977 5978 // Handle the degenerate case where two cases have the same result value. 5979 if (ResultVector.size() == 1 && DefaultResult) { 5980 ArrayRef<ConstantInt *> CaseValues = ResultVector[0].second; 5981 unsigned CaseCount = CaseValues.size(); 5982 // n bits group cases map to the same result: 5983 // case 0,4 -> Cond & 0b1..1011 == 0 ? result : default 5984 // case 0,2,4,6 -> Cond & 0b1..1001 == 0 ? result : default 5985 // case 0,2,8,10 -> Cond & 0b1..0101 == 0 ? result : default 5986 if (isPowerOf2_32(CaseCount)) { 5987 ConstantInt *MinCaseVal = CaseValues[0]; 5988 // Find mininal value. 5989 for (auto *Case : CaseValues) 5990 if (Case->getValue().slt(MinCaseVal->getValue())) 5991 MinCaseVal = Case; 5992 5993 // Mark the bits case number touched. 5994 APInt BitMask = APInt::getZero(MinCaseVal->getBitWidth()); 5995 for (auto *Case : CaseValues) 5996 BitMask |= (Case->getValue() - MinCaseVal->getValue()); 5997 5998 // Check if cases with the same result can cover all number 5999 // in touched bits. 6000 if (BitMask.popcount() == Log2_32(CaseCount)) { 6001 if (!MinCaseVal->isNullValue()) 6002 Condition = Builder.CreateSub(Condition, MinCaseVal); 6003 Value *And = Builder.CreateAnd(Condition, ~BitMask, "switch.and"); 6004 Value *Cmp = Builder.CreateICmpEQ( 6005 And, Constant::getNullValue(And->getType()), "switch.selectcmp"); 6006 return Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult); 6007 } 6008 } 6009 6010 // Handle the degenerate case where two cases have the same value. 6011 if (CaseValues.size() == 2) { 6012 Value *Cmp1 = Builder.CreateICmpEQ(Condition, CaseValues[0], 6013 "switch.selectcmp.case1"); 6014 Value *Cmp2 = Builder.CreateICmpEQ(Condition, CaseValues[1], 6015 "switch.selectcmp.case2"); 6016 Value *Cmp = Builder.CreateOr(Cmp1, Cmp2, "switch.selectcmp"); 6017 return Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult); 6018 } 6019 } 6020 6021 return nullptr; 6022 } 6023 6024 // Helper function to cleanup a switch instruction that has been converted into 6025 // a select, fixing up PHI nodes and basic blocks. 6026 static void removeSwitchAfterSelectFold(SwitchInst *SI, PHINode *PHI, 6027 Value *SelectValue, 6028 IRBuilder<> &Builder, 6029 DomTreeUpdater *DTU) { 6030 std::vector<DominatorTree::UpdateType> Updates; 6031 6032 BasicBlock *SelectBB = SI->getParent(); 6033 BasicBlock *DestBB = PHI->getParent(); 6034 6035 if (DTU && !is_contained(predecessors(DestBB), SelectBB)) 6036 Updates.push_back({DominatorTree::Insert, SelectBB, DestBB}); 6037 Builder.CreateBr(DestBB); 6038 6039 // Remove the switch. 6040 6041 PHI->removeIncomingValueIf( 6042 [&](unsigned Idx) { return PHI->getIncomingBlock(Idx) == SelectBB; }); 6043 PHI->addIncoming(SelectValue, SelectBB); 6044 6045 SmallPtrSet<BasicBlock *, 4> RemovedSuccessors; 6046 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) { 6047 BasicBlock *Succ = SI->getSuccessor(i); 6048 6049 if (Succ == DestBB) 6050 continue; 6051 Succ->removePredecessor(SelectBB); 6052 if (DTU && RemovedSuccessors.insert(Succ).second) 6053 Updates.push_back({DominatorTree::Delete, SelectBB, Succ}); 6054 } 6055 SI->eraseFromParent(); 6056 if (DTU) 6057 DTU->applyUpdates(Updates); 6058 } 6059 6060 /// If a switch is only used to initialize one or more phi nodes in a common 6061 /// successor block with only two different constant values, try to replace the 6062 /// switch with a select. Returns true if the fold was made. 6063 static bool trySwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder, 6064 DomTreeUpdater *DTU, const DataLayout &DL, 6065 const TargetTransformInfo &TTI) { 6066 Value *const Cond = SI->getCondition(); 6067 PHINode *PHI = nullptr; 6068 BasicBlock *CommonDest = nullptr; 6069 Constant *DefaultResult; 6070 SwitchCaseResultVectorTy UniqueResults; 6071 // Collect all the cases that will deliver the same value from the switch. 6072 if (!initializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult, 6073 DL, TTI, /*MaxUniqueResults*/ 2)) 6074 return false; 6075 6076 assert(PHI != nullptr && "PHI for value select not found"); 6077 Builder.SetInsertPoint(SI); 6078 Value *SelectValue = 6079 foldSwitchToSelect(UniqueResults, DefaultResult, Cond, Builder); 6080 if (!SelectValue) 6081 return false; 6082 6083 removeSwitchAfterSelectFold(SI, PHI, SelectValue, Builder, DTU); 6084 return true; 6085 } 6086 6087 namespace { 6088 6089 /// This class represents a lookup table that can be used to replace a switch. 6090 class SwitchLookupTable { 6091 public: 6092 /// Create a lookup table to use as a switch replacement with the contents 6093 /// of Values, using DefaultValue to fill any holes in the table. 6094 SwitchLookupTable( 6095 Module &M, uint64_t TableSize, ConstantInt *Offset, 6096 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values, 6097 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName); 6098 6099 /// Build instructions with Builder to retrieve the value at 6100 /// the position given by Index in the lookup table. 6101 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 6102 6103 /// Return true if a table with TableSize elements of 6104 /// type ElementType would fit in a target-legal register. 6105 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize, 6106 Type *ElementType); 6107 6108 private: 6109 // Depending on the contents of the table, it can be represented in 6110 // different ways. 6111 enum { 6112 // For tables where each element contains the same value, we just have to 6113 // store that single value and return it for each lookup. 6114 SingleValueKind, 6115 6116 // For tables where there is a linear relationship between table index 6117 // and values. We calculate the result with a simple multiplication 6118 // and addition instead of a table lookup. 6119 LinearMapKind, 6120 6121 // For small tables with integer elements, we can pack them into a bitmap 6122 // that fits into a target-legal register. Values are retrieved by 6123 // shift and mask operations. 6124 BitMapKind, 6125 6126 // The table is stored as an array of values. Values are retrieved by load 6127 // instructions from the table. 6128 ArrayKind 6129 } Kind; 6130 6131 // For SingleValueKind, this is the single value. 6132 Constant *SingleValue = nullptr; 6133 6134 // For BitMapKind, this is the bitmap. 6135 ConstantInt *BitMap = nullptr; 6136 IntegerType *BitMapElementTy = nullptr; 6137 6138 // For LinearMapKind, these are the constants used to derive the value. 6139 ConstantInt *LinearOffset = nullptr; 6140 ConstantInt *LinearMultiplier = nullptr; 6141 bool LinearMapValWrapped = false; 6142 6143 // For ArrayKind, this is the array. 6144 GlobalVariable *Array = nullptr; 6145 }; 6146 6147 } // end anonymous namespace 6148 6149 SwitchLookupTable::SwitchLookupTable( 6150 Module &M, uint64_t TableSize, ConstantInt *Offset, 6151 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values, 6152 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) { 6153 assert(Values.size() && "Can't build lookup table without values!"); 6154 assert(TableSize >= Values.size() && "Can't fit values in table!"); 6155 6156 // If all values in the table are equal, this is that value. 6157 SingleValue = Values.begin()->second; 6158 6159 Type *ValueType = Values.begin()->second->getType(); 6160 6161 // Build up the table contents. 6162 SmallVector<Constant *, 64> TableContents(TableSize); 6163 for (size_t I = 0, E = Values.size(); I != E; ++I) { 6164 ConstantInt *CaseVal = Values[I].first; 6165 Constant *CaseRes = Values[I].second; 6166 assert(CaseRes->getType() == ValueType); 6167 6168 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue(); 6169 TableContents[Idx] = CaseRes; 6170 6171 if (CaseRes != SingleValue) 6172 SingleValue = nullptr; 6173 } 6174 6175 // Fill in any holes in the table with the default result. 6176 if (Values.size() < TableSize) { 6177 assert(DefaultValue && 6178 "Need a default value to fill the lookup table holes."); 6179 assert(DefaultValue->getType() == ValueType); 6180 for (uint64_t I = 0; I < TableSize; ++I) { 6181 if (!TableContents[I]) 6182 TableContents[I] = DefaultValue; 6183 } 6184 6185 if (DefaultValue != SingleValue) 6186 SingleValue = nullptr; 6187 } 6188 6189 // If each element in the table contains the same value, we only need to store 6190 // that single value. 6191 if (SingleValue) { 6192 Kind = SingleValueKind; 6193 return; 6194 } 6195 6196 // Check if we can derive the value with a linear transformation from the 6197 // table index. 6198 if (isa<IntegerType>(ValueType)) { 6199 bool LinearMappingPossible = true; 6200 APInt PrevVal; 6201 APInt DistToPrev; 6202 // When linear map is monotonic and signed overflow doesn't happen on 6203 // maximum index, we can attach nsw on Add and Mul. 6204 bool NonMonotonic = false; 6205 assert(TableSize >= 2 && "Should be a SingleValue table."); 6206 // Check if there is the same distance between two consecutive values. 6207 for (uint64_t I = 0; I < TableSize; ++I) { 6208 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]); 6209 if (!ConstVal) { 6210 // This is an undef. We could deal with it, but undefs in lookup tables 6211 // are very seldom. It's probably not worth the additional complexity. 6212 LinearMappingPossible = false; 6213 break; 6214 } 6215 const APInt &Val = ConstVal->getValue(); 6216 if (I != 0) { 6217 APInt Dist = Val - PrevVal; 6218 if (I == 1) { 6219 DistToPrev = Dist; 6220 } else if (Dist != DistToPrev) { 6221 LinearMappingPossible = false; 6222 break; 6223 } 6224 NonMonotonic |= 6225 Dist.isStrictlyPositive() ? Val.sle(PrevVal) : Val.sgt(PrevVal); 6226 } 6227 PrevVal = Val; 6228 } 6229 if (LinearMappingPossible) { 6230 LinearOffset = cast<ConstantInt>(TableContents[0]); 6231 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev); 6232 bool MayWrap = false; 6233 APInt M = LinearMultiplier->getValue(); 6234 (void)M.smul_ov(APInt(M.getBitWidth(), TableSize - 1), MayWrap); 6235 LinearMapValWrapped = NonMonotonic || MayWrap; 6236 Kind = LinearMapKind; 6237 ++NumLinearMaps; 6238 return; 6239 } 6240 } 6241 6242 // If the type is integer and the table fits in a register, build a bitmap. 6243 if (WouldFitInRegister(DL, TableSize, ValueType)) { 6244 IntegerType *IT = cast<IntegerType>(ValueType); 6245 APInt TableInt(TableSize * IT->getBitWidth(), 0); 6246 for (uint64_t I = TableSize; I > 0; --I) { 6247 TableInt <<= IT->getBitWidth(); 6248 // Insert values into the bitmap. Undef values are set to zero. 6249 if (!isa<UndefValue>(TableContents[I - 1])) { 6250 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 6251 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 6252 } 6253 } 6254 BitMap = ConstantInt::get(M.getContext(), TableInt); 6255 BitMapElementTy = IT; 6256 Kind = BitMapKind; 6257 ++NumBitMaps; 6258 return; 6259 } 6260 6261 // Store the table in an array. 6262 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize); 6263 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 6264 6265 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true, 6266 GlobalVariable::PrivateLinkage, Initializer, 6267 "switch.table." + FuncName); 6268 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 6269 // Set the alignment to that of an array items. We will be only loading one 6270 // value out of it. 6271 Array->setAlignment(DL.getPrefTypeAlign(ValueType)); 6272 Kind = ArrayKind; 6273 } 6274 6275 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 6276 switch (Kind) { 6277 case SingleValueKind: 6278 return SingleValue; 6279 case LinearMapKind: { 6280 // Derive the result value from the input value. 6281 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(), 6282 false, "switch.idx.cast"); 6283 if (!LinearMultiplier->isOne()) 6284 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult", 6285 /*HasNUW = */ false, 6286 /*HasNSW = */ !LinearMapValWrapped); 6287 6288 if (!LinearOffset->isZero()) 6289 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset", 6290 /*HasNUW = */ false, 6291 /*HasNSW = */ !LinearMapValWrapped); 6292 return Result; 6293 } 6294 case BitMapKind: { 6295 // Type of the bitmap (e.g. i59). 6296 IntegerType *MapTy = BitMap->getType(); 6297 6298 // Cast Index to the same type as the bitmap. 6299 // Note: The Index is <= the number of elements in the table, so 6300 // truncating it to the width of the bitmask is safe. 6301 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 6302 6303 // Multiply the shift amount by the element width. NUW/NSW can always be 6304 // set, because WouldFitInRegister guarantees Index * ShiftAmt is in 6305 // BitMap's bit width. 6306 ShiftAmt = Builder.CreateMul( 6307 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 6308 "switch.shiftamt",/*HasNUW =*/true,/*HasNSW =*/true); 6309 6310 // Shift down. 6311 Value *DownShifted = 6312 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift"); 6313 // Mask off. 6314 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked"); 6315 } 6316 case ArrayKind: { 6317 // Make sure the table index will not overflow when treated as signed. 6318 IntegerType *IT = cast<IntegerType>(Index->getType()); 6319 uint64_t TableSize = 6320 Array->getInitializer()->getType()->getArrayNumElements(); 6321 if (TableSize > (1ULL << std::min(IT->getBitWidth() - 1, 63u))) 6322 Index = Builder.CreateZExt( 6323 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1), 6324 "switch.tableidx.zext"); 6325 6326 Value *GEPIndices[] = {Builder.getInt32(0), Index}; 6327 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array, 6328 GEPIndices, "switch.gep"); 6329 return Builder.CreateLoad( 6330 cast<ArrayType>(Array->getValueType())->getElementType(), GEP, 6331 "switch.load"); 6332 } 6333 } 6334 llvm_unreachable("Unknown lookup table kind!"); 6335 } 6336 6337 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL, 6338 uint64_t TableSize, 6339 Type *ElementType) { 6340 auto *IT = dyn_cast<IntegerType>(ElementType); 6341 if (!IT) 6342 return false; 6343 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 6344 // are <= 15, we could try to narrow the type. 6345 6346 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 6347 if (TableSize >= UINT_MAX / IT->getBitWidth()) 6348 return false; 6349 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth()); 6350 } 6351 6352 static bool isTypeLegalForLookupTable(Type *Ty, const TargetTransformInfo &TTI, 6353 const DataLayout &DL) { 6354 // Allow any legal type. 6355 if (TTI.isTypeLegal(Ty)) 6356 return true; 6357 6358 auto *IT = dyn_cast<IntegerType>(Ty); 6359 if (!IT) 6360 return false; 6361 6362 // Also allow power of 2 integer types that have at least 8 bits and fit in 6363 // a register. These types are common in frontend languages and targets 6364 // usually support loads of these types. 6365 // TODO: We could relax this to any integer that fits in a register and rely 6366 // on ABI alignment and padding in the table to allow the load to be widened. 6367 // Or we could widen the constants and truncate the load. 6368 unsigned BitWidth = IT->getBitWidth(); 6369 return BitWidth >= 8 && isPowerOf2_32(BitWidth) && 6370 DL.fitsInLegalInteger(IT->getBitWidth()); 6371 } 6372 6373 static bool isSwitchDense(uint64_t NumCases, uint64_t CaseRange) { 6374 // 40% is the default density for building a jump table in optsize/minsize 6375 // mode. See also TargetLoweringBase::isSuitableForJumpTable(), which this 6376 // function was based on. 6377 const uint64_t MinDensity = 40; 6378 6379 if (CaseRange >= UINT64_MAX / 100) 6380 return false; // Avoid multiplication overflows below. 6381 6382 return NumCases * 100 >= CaseRange * MinDensity; 6383 } 6384 6385 static bool isSwitchDense(ArrayRef<int64_t> Values) { 6386 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front(); 6387 uint64_t Range = Diff + 1; 6388 if (Range < Diff) 6389 return false; // Overflow. 6390 6391 return isSwitchDense(Values.size(), Range); 6392 } 6393 6394 /// Determine whether a lookup table should be built for this switch, based on 6395 /// the number of cases, size of the table, and the types of the results. 6396 // TODO: We could support larger than legal types by limiting based on the 6397 // number of loads required and/or table size. If the constants are small we 6398 // could use smaller table entries and extend after the load. 6399 static bool 6400 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize, 6401 const TargetTransformInfo &TTI, const DataLayout &DL, 6402 const SmallDenseMap<PHINode *, Type *> &ResultTypes) { 6403 if (SI->getNumCases() > TableSize) 6404 return false; // TableSize overflowed. 6405 6406 bool AllTablesFitInRegister = true; 6407 bool HasIllegalType = false; 6408 for (const auto &I : ResultTypes) { 6409 Type *Ty = I.second; 6410 6411 // Saturate this flag to true. 6412 HasIllegalType = HasIllegalType || !isTypeLegalForLookupTable(Ty, TTI, DL); 6413 6414 // Saturate this flag to false. 6415 AllTablesFitInRegister = 6416 AllTablesFitInRegister && 6417 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty); 6418 6419 // If both flags saturate, we're done. NOTE: This *only* works with 6420 // saturating flags, and all flags have to saturate first due to the 6421 // non-deterministic behavior of iterating over a dense map. 6422 if (HasIllegalType && !AllTablesFitInRegister) 6423 break; 6424 } 6425 6426 // If each table would fit in a register, we should build it anyway. 6427 if (AllTablesFitInRegister) 6428 return true; 6429 6430 // Don't build a table that doesn't fit in-register if it has illegal types. 6431 if (HasIllegalType) 6432 return false; 6433 6434 return isSwitchDense(SI->getNumCases(), TableSize); 6435 } 6436 6437 static bool ShouldUseSwitchConditionAsTableIndex( 6438 ConstantInt &MinCaseVal, const ConstantInt &MaxCaseVal, 6439 bool HasDefaultResults, const SmallDenseMap<PHINode *, Type *> &ResultTypes, 6440 const DataLayout &DL, const TargetTransformInfo &TTI) { 6441 if (MinCaseVal.isNullValue()) 6442 return true; 6443 if (MinCaseVal.isNegative() || 6444 MaxCaseVal.getLimitedValue() == std::numeric_limits<uint64_t>::max() || 6445 !HasDefaultResults) 6446 return false; 6447 return all_of(ResultTypes, [&](const auto &KV) { 6448 return SwitchLookupTable::WouldFitInRegister( 6449 DL, MaxCaseVal.getLimitedValue() + 1 /* TableSize */, 6450 KV.second /* ResultType */); 6451 }); 6452 } 6453 6454 /// Try to reuse the switch table index compare. Following pattern: 6455 /// \code 6456 /// if (idx < tablesize) 6457 /// r = table[idx]; // table does not contain default_value 6458 /// else 6459 /// r = default_value; 6460 /// if (r != default_value) 6461 /// ... 6462 /// \endcode 6463 /// Is optimized to: 6464 /// \code 6465 /// cond = idx < tablesize; 6466 /// if (cond) 6467 /// r = table[idx]; 6468 /// else 6469 /// r = default_value; 6470 /// if (cond) 6471 /// ... 6472 /// \endcode 6473 /// Jump threading will then eliminate the second if(cond). 6474 static void reuseTableCompare( 6475 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch, 6476 Constant *DefaultValue, 6477 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) { 6478 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser); 6479 if (!CmpInst) 6480 return; 6481 6482 // We require that the compare is in the same block as the phi so that jump 6483 // threading can do its work afterwards. 6484 if (CmpInst->getParent() != PhiBlock) 6485 return; 6486 6487 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1)); 6488 if (!CmpOp1) 6489 return; 6490 6491 Value *RangeCmp = RangeCheckBranch->getCondition(); 6492 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType()); 6493 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType()); 6494 6495 // Check if the compare with the default value is constant true or false. 6496 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(), 6497 DefaultValue, CmpOp1, true); 6498 if (DefaultConst != TrueConst && DefaultConst != FalseConst) 6499 return; 6500 6501 // Check if the compare with the case values is distinct from the default 6502 // compare result. 6503 for (auto ValuePair : Values) { 6504 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(), 6505 ValuePair.second, CmpOp1, true); 6506 if (!CaseConst || CaseConst == DefaultConst || 6507 (CaseConst != TrueConst && CaseConst != FalseConst)) 6508 return; 6509 } 6510 6511 // Check if the branch instruction dominates the phi node. It's a simple 6512 // dominance check, but sufficient for our needs. 6513 // Although this check is invariant in the calling loops, it's better to do it 6514 // at this late stage. Practically we do it at most once for a switch. 6515 BasicBlock *BranchBlock = RangeCheckBranch->getParent(); 6516 for (BasicBlock *Pred : predecessors(PhiBlock)) { 6517 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock) 6518 return; 6519 } 6520 6521 if (DefaultConst == FalseConst) { 6522 // The compare yields the same result. We can replace it. 6523 CmpInst->replaceAllUsesWith(RangeCmp); 6524 ++NumTableCmpReuses; 6525 } else { 6526 // The compare yields the same result, just inverted. We can replace it. 6527 Value *InvertedTableCmp = BinaryOperator::CreateXor( 6528 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp", 6529 RangeCheckBranch); 6530 CmpInst->replaceAllUsesWith(InvertedTableCmp); 6531 ++NumTableCmpReuses; 6532 } 6533 } 6534 6535 /// If the switch is only used to initialize one or more phi nodes in a common 6536 /// successor block with different constant values, replace the switch with 6537 /// lookup tables. 6538 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder, 6539 DomTreeUpdater *DTU, const DataLayout &DL, 6540 const TargetTransformInfo &TTI) { 6541 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 6542 6543 BasicBlock *BB = SI->getParent(); 6544 Function *Fn = BB->getParent(); 6545 // Only build lookup table when we have a target that supports it or the 6546 // attribute is not set. 6547 if (!TTI.shouldBuildLookupTables() || 6548 (Fn->getFnAttribute("no-jump-tables").getValueAsBool())) 6549 return false; 6550 6551 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 6552 // split off a dense part and build a lookup table for that. 6553 6554 // FIXME: This creates arrays of GEPs to constant strings, which means each 6555 // GEP needs a runtime relocation in PIC code. We should just build one big 6556 // string and lookup indices into that. 6557 6558 // Ignore switches with less than three cases. Lookup tables will not make 6559 // them faster, so we don't analyze them. 6560 if (SI->getNumCases() < 3) 6561 return false; 6562 6563 // Figure out the corresponding result for each case value and phi node in the 6564 // common destination, as well as the min and max case values. 6565 assert(!SI->cases().empty()); 6566 SwitchInst::CaseIt CI = SI->case_begin(); 6567 ConstantInt *MinCaseVal = CI->getCaseValue(); 6568 ConstantInt *MaxCaseVal = CI->getCaseValue(); 6569 6570 BasicBlock *CommonDest = nullptr; 6571 6572 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>; 6573 SmallDenseMap<PHINode *, ResultListTy> ResultLists; 6574 6575 SmallDenseMap<PHINode *, Constant *> DefaultResults; 6576 SmallDenseMap<PHINode *, Type *> ResultTypes; 6577 SmallVector<PHINode *, 4> PHIs; 6578 6579 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 6580 ConstantInt *CaseVal = CI->getCaseValue(); 6581 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 6582 MinCaseVal = CaseVal; 6583 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 6584 MaxCaseVal = CaseVal; 6585 6586 // Resulting value at phi nodes for this case value. 6587 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>; 6588 ResultsTy Results; 6589 if (!getCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest, 6590 Results, DL, TTI)) 6591 return false; 6592 6593 // Append the result from this case to the list for each phi. 6594 for (const auto &I : Results) { 6595 PHINode *PHI = I.first; 6596 Constant *Value = I.second; 6597 if (!ResultLists.count(PHI)) 6598 PHIs.push_back(PHI); 6599 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value)); 6600 } 6601 } 6602 6603 // Keep track of the result types. 6604 for (PHINode *PHI : PHIs) { 6605 ResultTypes[PHI] = ResultLists[PHI][0].second->getType(); 6606 } 6607 6608 uint64_t NumResults = ResultLists[PHIs[0]].size(); 6609 6610 // If the table has holes, we need a constant result for the default case 6611 // or a bitmask that fits in a register. 6612 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList; 6613 bool HasDefaultResults = 6614 getCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, 6615 DefaultResultsList, DL, TTI); 6616 6617 for (const auto &I : DefaultResultsList) { 6618 PHINode *PHI = I.first; 6619 Constant *Result = I.second; 6620 DefaultResults[PHI] = Result; 6621 } 6622 6623 bool UseSwitchConditionAsTableIndex = ShouldUseSwitchConditionAsTableIndex( 6624 *MinCaseVal, *MaxCaseVal, HasDefaultResults, ResultTypes, DL, TTI); 6625 uint64_t TableSize; 6626 if (UseSwitchConditionAsTableIndex) 6627 TableSize = MaxCaseVal->getLimitedValue() + 1; 6628 else 6629 TableSize = 6630 (MaxCaseVal->getValue() - MinCaseVal->getValue()).getLimitedValue() + 1; 6631 6632 bool TableHasHoles = (NumResults < TableSize); 6633 bool NeedMask = (TableHasHoles && !HasDefaultResults); 6634 if (NeedMask) { 6635 // As an extra penalty for the validity test we require more cases. 6636 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark). 6637 return false; 6638 if (!DL.fitsInLegalInteger(TableSize)) 6639 return false; 6640 } 6641 6642 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes)) 6643 return false; 6644 6645 std::vector<DominatorTree::UpdateType> Updates; 6646 6647 // Compute the maximum table size representable by the integer type we are 6648 // switching upon. 6649 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits(); 6650 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize; 6651 assert(MaxTableSize >= TableSize && 6652 "It is impossible for a switch to have more entries than the max " 6653 "representable value of its input integer type's size."); 6654 6655 // If the default destination is unreachable, or if the lookup table covers 6656 // all values of the conditional variable, branch directly to the lookup table 6657 // BB. Otherwise, check that the condition is within the case range. 6658 bool DefaultIsReachable = 6659 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg()); 6660 6661 // Create the BB that does the lookups. 6662 Module &Mod = *CommonDest->getParent()->getParent(); 6663 BasicBlock *LookupBB = BasicBlock::Create( 6664 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest); 6665 6666 // Compute the table index value. 6667 Builder.SetInsertPoint(SI); 6668 Value *TableIndex; 6669 ConstantInt *TableIndexOffset; 6670 if (UseSwitchConditionAsTableIndex) { 6671 TableIndexOffset = ConstantInt::get(MaxCaseVal->getType(), 0); 6672 TableIndex = SI->getCondition(); 6673 } else { 6674 TableIndexOffset = MinCaseVal; 6675 // If the default is unreachable, all case values are s>= MinCaseVal. Then 6676 // we can try to attach nsw. 6677 bool MayWrap = true; 6678 if (!DefaultIsReachable) { 6679 APInt Res = MaxCaseVal->getValue().ssub_ov(MinCaseVal->getValue(), MayWrap); 6680 (void)Res; 6681 } 6682 6683 TableIndex = Builder.CreateSub(SI->getCondition(), TableIndexOffset, 6684 "switch.tableidx", /*HasNUW =*/false, 6685 /*HasNSW =*/!MayWrap); 6686 } 6687 6688 BranchInst *RangeCheckBranch = nullptr; 6689 6690 // Grow the table to cover all possible index values to avoid the range check. 6691 // It will use the default result to fill in the table hole later, so make 6692 // sure it exist. 6693 if (UseSwitchConditionAsTableIndex && HasDefaultResults) { 6694 ConstantRange CR = computeConstantRange(TableIndex, /* ForSigned */ false); 6695 // Grow the table shouldn't have any size impact by checking 6696 // WouldFitInRegister. 6697 // TODO: Consider growing the table also when it doesn't fit in a register 6698 // if no optsize is specified. 6699 const uint64_t UpperBound = CR.getUpper().getLimitedValue(); 6700 if (!CR.isUpperWrapped() && all_of(ResultTypes, [&](const auto &KV) { 6701 return SwitchLookupTable::WouldFitInRegister( 6702 DL, UpperBound, KV.second /* ResultType */); 6703 })) { 6704 // The default branch is unreachable after we enlarge the lookup table. 6705 // Adjust DefaultIsReachable to reuse code path. 6706 TableSize = UpperBound; 6707 DefaultIsReachable = false; 6708 } 6709 } 6710 6711 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize); 6712 if (!DefaultIsReachable || GeneratingCoveredLookupTable) { 6713 Builder.CreateBr(LookupBB); 6714 if (DTU) 6715 Updates.push_back({DominatorTree::Insert, BB, LookupBB}); 6716 // Note: We call removeProdecessor later since we need to be able to get the 6717 // PHI value for the default case in case we're using a bit mask. 6718 } else { 6719 Value *Cmp = Builder.CreateICmpULT( 6720 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize)); 6721 RangeCheckBranch = 6722 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 6723 if (DTU) 6724 Updates.push_back({DominatorTree::Insert, BB, LookupBB}); 6725 } 6726 6727 // Populate the BB that does the lookups. 6728 Builder.SetInsertPoint(LookupBB); 6729 6730 if (NeedMask) { 6731 // Before doing the lookup, we do the hole check. The LookupBB is therefore 6732 // re-purposed to do the hole check, and we create a new LookupBB. 6733 BasicBlock *MaskBB = LookupBB; 6734 MaskBB->setName("switch.hole_check"); 6735 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup", 6736 CommonDest->getParent(), CommonDest); 6737 6738 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid 6739 // unnecessary illegal types. 6740 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL)); 6741 APInt MaskInt(TableSizePowOf2, 0); 6742 APInt One(TableSizePowOf2, 1); 6743 // Build bitmask; fill in a 1 bit for every case. 6744 const ResultListTy &ResultList = ResultLists[PHIs[0]]; 6745 for (size_t I = 0, E = ResultList.size(); I != E; ++I) { 6746 uint64_t Idx = (ResultList[I].first->getValue() - TableIndexOffset->getValue()) 6747 .getLimitedValue(); 6748 MaskInt |= One << Idx; 6749 } 6750 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt); 6751 6752 // Get the TableIndex'th bit of the bitmask. 6753 // If this bit is 0 (meaning hole) jump to the default destination, 6754 // else continue with table lookup. 6755 IntegerType *MapTy = TableMask->getType(); 6756 Value *MaskIndex = 6757 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex"); 6758 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted"); 6759 Value *LoBit = Builder.CreateTrunc( 6760 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit"); 6761 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest()); 6762 if (DTU) { 6763 Updates.push_back({DominatorTree::Insert, MaskBB, LookupBB}); 6764 Updates.push_back({DominatorTree::Insert, MaskBB, SI->getDefaultDest()}); 6765 } 6766 Builder.SetInsertPoint(LookupBB); 6767 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, BB); 6768 } 6769 6770 if (!DefaultIsReachable || GeneratingCoveredLookupTable) { 6771 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later, 6772 // do not delete PHINodes here. 6773 SI->getDefaultDest()->removePredecessor(BB, 6774 /*KeepOneInputPHIs=*/true); 6775 if (DTU) 6776 Updates.push_back({DominatorTree::Delete, BB, SI->getDefaultDest()}); 6777 } 6778 6779 for (PHINode *PHI : PHIs) { 6780 const ResultListTy &ResultList = ResultLists[PHI]; 6781 6782 // If using a bitmask, use any value to fill the lookup table holes. 6783 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI]; 6784 StringRef FuncName = Fn->getName(); 6785 SwitchLookupTable Table(Mod, TableSize, TableIndexOffset, ResultList, DV, 6786 DL, FuncName); 6787 6788 Value *Result = Table.BuildLookup(TableIndex, Builder); 6789 6790 // Do a small peephole optimization: re-use the switch table compare if 6791 // possible. 6792 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) { 6793 BasicBlock *PhiBlock = PHI->getParent(); 6794 // Search for compare instructions which use the phi. 6795 for (auto *User : PHI->users()) { 6796 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList); 6797 } 6798 } 6799 6800 PHI->addIncoming(Result, LookupBB); 6801 } 6802 6803 Builder.CreateBr(CommonDest); 6804 if (DTU) 6805 Updates.push_back({DominatorTree::Insert, LookupBB, CommonDest}); 6806 6807 // Remove the switch. 6808 SmallPtrSet<BasicBlock *, 8> RemovedSuccessors; 6809 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) { 6810 BasicBlock *Succ = SI->getSuccessor(i); 6811 6812 if (Succ == SI->getDefaultDest()) 6813 continue; 6814 Succ->removePredecessor(BB); 6815 if (DTU && RemovedSuccessors.insert(Succ).second) 6816 Updates.push_back({DominatorTree::Delete, BB, Succ}); 6817 } 6818 SI->eraseFromParent(); 6819 6820 if (DTU) 6821 DTU->applyUpdates(Updates); 6822 6823 ++NumLookupTables; 6824 if (NeedMask) 6825 ++NumLookupTablesHoles; 6826 return true; 6827 } 6828 6829 /// Try to transform a switch that has "holes" in it to a contiguous sequence 6830 /// of cases. 6831 /// 6832 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be 6833 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}. 6834 /// 6835 /// This converts a sparse switch into a dense switch which allows better 6836 /// lowering and could also allow transforming into a lookup table. 6837 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder, 6838 const DataLayout &DL, 6839 const TargetTransformInfo &TTI) { 6840 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType()); 6841 if (CondTy->getIntegerBitWidth() > 64 || 6842 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth())) 6843 return false; 6844 // Only bother with this optimization if there are more than 3 switch cases; 6845 // SDAG will only bother creating jump tables for 4 or more cases. 6846 if (SI->getNumCases() < 4) 6847 return false; 6848 6849 // This transform is agnostic to the signedness of the input or case values. We 6850 // can treat the case values as signed or unsigned. We can optimize more common 6851 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values 6852 // as signed. 6853 SmallVector<int64_t,4> Values; 6854 for (const auto &C : SI->cases()) 6855 Values.push_back(C.getCaseValue()->getValue().getSExtValue()); 6856 llvm::sort(Values); 6857 6858 // If the switch is already dense, there's nothing useful to do here. 6859 if (isSwitchDense(Values)) 6860 return false; 6861 6862 // First, transform the values such that they start at zero and ascend. 6863 int64_t Base = Values[0]; 6864 for (auto &V : Values) 6865 V -= (uint64_t)(Base); 6866 6867 // Now we have signed numbers that have been shifted so that, given enough 6868 // precision, there are no negative values. Since the rest of the transform 6869 // is bitwise only, we switch now to an unsigned representation. 6870 6871 // This transform can be done speculatively because it is so cheap - it 6872 // results in a single rotate operation being inserted. 6873 6874 // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than 6875 // one element and LLVM disallows duplicate cases, Shift is guaranteed to be 6876 // less than 64. 6877 unsigned Shift = 64; 6878 for (auto &V : Values) 6879 Shift = std::min(Shift, (unsigned)llvm::countr_zero((uint64_t)V)); 6880 assert(Shift < 64); 6881 if (Shift > 0) 6882 for (auto &V : Values) 6883 V = (int64_t)((uint64_t)V >> Shift); 6884 6885 if (!isSwitchDense(Values)) 6886 // Transform didn't create a dense switch. 6887 return false; 6888 6889 // The obvious transform is to shift the switch condition right and emit a 6890 // check that the condition actually cleanly divided by GCD, i.e. 6891 // C & (1 << Shift - 1) == 0 6892 // inserting a new CFG edge to handle the case where it didn't divide cleanly. 6893 // 6894 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the 6895 // shift and puts the shifted-off bits in the uppermost bits. If any of these 6896 // are nonzero then the switch condition will be very large and will hit the 6897 // default case. 6898 6899 auto *Ty = cast<IntegerType>(SI->getCondition()->getType()); 6900 Builder.SetInsertPoint(SI); 6901 auto *ShiftC = ConstantInt::get(Ty, Shift); 6902 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base)); 6903 auto *LShr = Builder.CreateLShr(Sub, ShiftC); 6904 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift); 6905 auto *Rot = Builder.CreateOr(LShr, Shl); 6906 SI->replaceUsesOfWith(SI->getCondition(), Rot); 6907 6908 for (auto Case : SI->cases()) { 6909 auto *Orig = Case.getCaseValue(); 6910 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base); 6911 Case.setValue( 6912 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue())))); 6913 } 6914 return true; 6915 } 6916 6917 /// Tries to transform switch of powers of two to reduce switch range. 6918 /// For example, switch like: 6919 /// switch (C) { case 1: case 2: case 64: case 128: } 6920 /// will be transformed to: 6921 /// switch (count_trailing_zeros(C)) { case 0: case 1: case 6: case 7: } 6922 /// 6923 /// This transformation allows better lowering and could allow transforming into 6924 /// a lookup table. 6925 static bool simplifySwitchOfPowersOfTwo(SwitchInst *SI, IRBuilder<> &Builder, 6926 const DataLayout &DL, 6927 const TargetTransformInfo &TTI) { 6928 Value *Condition = SI->getCondition(); 6929 LLVMContext &Context = SI->getContext(); 6930 auto *CondTy = cast<IntegerType>(Condition->getType()); 6931 6932 if (CondTy->getIntegerBitWidth() > 64 || 6933 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth())) 6934 return false; 6935 6936 const auto CttzIntrinsicCost = TTI.getIntrinsicInstrCost( 6937 IntrinsicCostAttributes(Intrinsic::cttz, CondTy, 6938 {Condition, ConstantInt::getTrue(Context)}), 6939 TTI::TCK_SizeAndLatency); 6940 6941 if (CttzIntrinsicCost > TTI::TCC_Basic) 6942 // Inserting intrinsic is too expensive. 6943 return false; 6944 6945 // Only bother with this optimization if there are more than 3 switch cases. 6946 // SDAG will only bother creating jump tables for 4 or more cases. 6947 if (SI->getNumCases() < 4) 6948 return false; 6949 6950 // We perform this optimization only for switches with 6951 // unreachable default case. 6952 // This assumtion will save us from checking if `Condition` is a power of two. 6953 if (!isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg())) 6954 return false; 6955 6956 // Check that switch cases are powers of two. 6957 SmallVector<uint64_t, 4> Values; 6958 for (const auto &Case : SI->cases()) { 6959 uint64_t CaseValue = Case.getCaseValue()->getValue().getZExtValue(); 6960 if (llvm::has_single_bit(CaseValue)) 6961 Values.push_back(CaseValue); 6962 else 6963 return false; 6964 } 6965 6966 // isSwichDense requires case values to be sorted. 6967 llvm::sort(Values); 6968 if (!isSwitchDense(Values.size(), llvm::countr_zero(Values.back()) - 6969 llvm::countr_zero(Values.front()) + 1)) 6970 // Transform is unable to generate dense switch. 6971 return false; 6972 6973 Builder.SetInsertPoint(SI); 6974 6975 // Replace each case with its trailing zeros number. 6976 for (auto &Case : SI->cases()) { 6977 auto *OrigValue = Case.getCaseValue(); 6978 Case.setValue(ConstantInt::get(OrigValue->getType(), 6979 OrigValue->getValue().countr_zero())); 6980 } 6981 6982 // Replace condition with its trailing zeros number. 6983 auto *ConditionTrailingZeros = Builder.CreateIntrinsic( 6984 Intrinsic::cttz, {CondTy}, {Condition, ConstantInt::getTrue(Context)}); 6985 6986 SI->setCondition(ConditionTrailingZeros); 6987 6988 return true; 6989 } 6990 6991 bool SimplifyCFGOpt::simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 6992 BasicBlock *BB = SI->getParent(); 6993 6994 if (isValueEqualityComparison(SI)) { 6995 // If we only have one predecessor, and if it is a branch on this value, 6996 // see if that predecessor totally determines the outcome of this switch. 6997 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 6998 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 6999 return requestResimplify(); 7000 7001 Value *Cond = SI->getCondition(); 7002 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 7003 if (SimplifySwitchOnSelect(SI, Select)) 7004 return requestResimplify(); 7005 7006 // If the block only contains the switch, see if we can fold the block 7007 // away into any preds. 7008 if (SI == &*BB->instructionsWithoutDebug(false).begin()) 7009 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 7010 return requestResimplify(); 7011 } 7012 7013 // Try to transform the switch into an icmp and a branch. 7014 // The conversion from switch to comparison may lose information on 7015 // impossible switch values, so disable it early in the pipeline. 7016 if (Options.ConvertSwitchRangeToICmp && TurnSwitchRangeIntoICmp(SI, Builder)) 7017 return requestResimplify(); 7018 7019 // Remove unreachable cases. 7020 if (eliminateDeadSwitchCases(SI, DTU, Options.AC, DL)) 7021 return requestResimplify(); 7022 7023 if (trySwitchToSelect(SI, Builder, DTU, DL, TTI)) 7024 return requestResimplify(); 7025 7026 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI)) 7027 return requestResimplify(); 7028 7029 // The conversion from switch to lookup tables results in difficult-to-analyze 7030 // code and makes pruning branches much harder. This is a problem if the 7031 // switch expression itself can still be restricted as a result of inlining or 7032 // CVP. Therefore, only apply this transformation during late stages of the 7033 // optimisation pipeline. 7034 if (Options.ConvertSwitchToLookupTable && 7035 SwitchToLookupTable(SI, Builder, DTU, DL, TTI)) 7036 return requestResimplify(); 7037 7038 if (simplifySwitchOfPowersOfTwo(SI, Builder, DL, TTI)) 7039 return requestResimplify(); 7040 7041 if (ReduceSwitchRange(SI, Builder, DL, TTI)) 7042 return requestResimplify(); 7043 7044 if (HoistCommon && 7045 hoistCommonCodeFromSuccessors(SI->getParent(), !Options.HoistCommonInsts)) 7046 return requestResimplify(); 7047 7048 return false; 7049 } 7050 7051 bool SimplifyCFGOpt::simplifyIndirectBr(IndirectBrInst *IBI) { 7052 BasicBlock *BB = IBI->getParent(); 7053 bool Changed = false; 7054 7055 // Eliminate redundant destinations. 7056 SmallPtrSet<Value *, 8> Succs; 7057 SmallSetVector<BasicBlock *, 8> RemovedSuccs; 7058 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 7059 BasicBlock *Dest = IBI->getDestination(i); 7060 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) { 7061 if (!Dest->hasAddressTaken()) 7062 RemovedSuccs.insert(Dest); 7063 Dest->removePredecessor(BB); 7064 IBI->removeDestination(i); 7065 --i; 7066 --e; 7067 Changed = true; 7068 } 7069 } 7070 7071 if (DTU) { 7072 std::vector<DominatorTree::UpdateType> Updates; 7073 Updates.reserve(RemovedSuccs.size()); 7074 for (auto *RemovedSucc : RemovedSuccs) 7075 Updates.push_back({DominatorTree::Delete, BB, RemovedSucc}); 7076 DTU->applyUpdates(Updates); 7077 } 7078 7079 if (IBI->getNumDestinations() == 0) { 7080 // If the indirectbr has no successors, change it to unreachable. 7081 new UnreachableInst(IBI->getContext(), IBI); 7082 EraseTerminatorAndDCECond(IBI); 7083 return true; 7084 } 7085 7086 if (IBI->getNumDestinations() == 1) { 7087 // If the indirectbr has one successor, change it to a direct branch. 7088 BranchInst::Create(IBI->getDestination(0), IBI); 7089 EraseTerminatorAndDCECond(IBI); 7090 return true; 7091 } 7092 7093 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 7094 if (SimplifyIndirectBrOnSelect(IBI, SI)) 7095 return requestResimplify(); 7096 } 7097 return Changed; 7098 } 7099 7100 /// Given an block with only a single landing pad and a unconditional branch 7101 /// try to find another basic block which this one can be merged with. This 7102 /// handles cases where we have multiple invokes with unique landing pads, but 7103 /// a shared handler. 7104 /// 7105 /// We specifically choose to not worry about merging non-empty blocks 7106 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In 7107 /// practice, the optimizer produces empty landing pad blocks quite frequently 7108 /// when dealing with exception dense code. (see: instcombine, gvn, if-else 7109 /// sinking in this file) 7110 /// 7111 /// This is primarily a code size optimization. We need to avoid performing 7112 /// any transform which might inhibit optimization (such as our ability to 7113 /// specialize a particular handler via tail commoning). We do this by not 7114 /// merging any blocks which require us to introduce a phi. Since the same 7115 /// values are flowing through both blocks, we don't lose any ability to 7116 /// specialize. If anything, we make such specialization more likely. 7117 /// 7118 /// TODO - This transformation could remove entries from a phi in the target 7119 /// block when the inputs in the phi are the same for the two blocks being 7120 /// merged. In some cases, this could result in removal of the PHI entirely. 7121 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI, 7122 BasicBlock *BB, DomTreeUpdater *DTU) { 7123 auto Succ = BB->getUniqueSuccessor(); 7124 assert(Succ); 7125 // If there's a phi in the successor block, we'd likely have to introduce 7126 // a phi into the merged landing pad block. 7127 if (isa<PHINode>(*Succ->begin())) 7128 return false; 7129 7130 for (BasicBlock *OtherPred : predecessors(Succ)) { 7131 if (BB == OtherPred) 7132 continue; 7133 BasicBlock::iterator I = OtherPred->begin(); 7134 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I); 7135 if (!LPad2 || !LPad2->isIdenticalTo(LPad)) 7136 continue; 7137 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 7138 ; 7139 BranchInst *BI2 = dyn_cast<BranchInst>(I); 7140 if (!BI2 || !BI2->isIdenticalTo(BI)) 7141 continue; 7142 7143 std::vector<DominatorTree::UpdateType> Updates; 7144 7145 // We've found an identical block. Update our predecessors to take that 7146 // path instead and make ourselves dead. 7147 SmallSetVector<BasicBlock *, 16> UniquePreds(pred_begin(BB), pred_end(BB)); 7148 for (BasicBlock *Pred : UniquePreds) { 7149 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator()); 7150 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB && 7151 "unexpected successor"); 7152 II->setUnwindDest(OtherPred); 7153 if (DTU) { 7154 Updates.push_back({DominatorTree::Insert, Pred, OtherPred}); 7155 Updates.push_back({DominatorTree::Delete, Pred, BB}); 7156 } 7157 } 7158 7159 // The debug info in OtherPred doesn't cover the merged control flow that 7160 // used to go through BB. We need to delete it or update it. 7161 for (Instruction &Inst : llvm::make_early_inc_range(*OtherPred)) 7162 if (isa<DbgInfoIntrinsic>(Inst)) 7163 Inst.eraseFromParent(); 7164 7165 SmallSetVector<BasicBlock *, 16> UniqueSuccs(succ_begin(BB), succ_end(BB)); 7166 for (BasicBlock *Succ : UniqueSuccs) { 7167 Succ->removePredecessor(BB); 7168 if (DTU) 7169 Updates.push_back({DominatorTree::Delete, BB, Succ}); 7170 } 7171 7172 IRBuilder<> Builder(BI); 7173 Builder.CreateUnreachable(); 7174 BI->eraseFromParent(); 7175 if (DTU) 7176 DTU->applyUpdates(Updates); 7177 return true; 7178 } 7179 return false; 7180 } 7181 7182 bool SimplifyCFGOpt::simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder) { 7183 return Branch->isUnconditional() ? simplifyUncondBranch(Branch, Builder) 7184 : simplifyCondBranch(Branch, Builder); 7185 } 7186 7187 bool SimplifyCFGOpt::simplifyUncondBranch(BranchInst *BI, 7188 IRBuilder<> &Builder) { 7189 BasicBlock *BB = BI->getParent(); 7190 BasicBlock *Succ = BI->getSuccessor(0); 7191 7192 // If the Terminator is the only non-phi instruction, simplify the block. 7193 // If LoopHeader is provided, check if the block or its successor is a loop 7194 // header. (This is for early invocations before loop simplify and 7195 // vectorization to keep canonical loop forms for nested loops. These blocks 7196 // can be eliminated when the pass is invoked later in the back-end.) 7197 // Note that if BB has only one predecessor then we do not introduce new 7198 // backedge, so we can eliminate BB. 7199 bool NeedCanonicalLoop = 7200 Options.NeedCanonicalLoop && 7201 (!LoopHeaders.empty() && BB->hasNPredecessorsOrMore(2) && 7202 (is_contained(LoopHeaders, BB) || is_contained(LoopHeaders, Succ))); 7203 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg(true)->getIterator(); 7204 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 7205 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB, DTU)) 7206 return true; 7207 7208 // If the only instruction in the block is a seteq/setne comparison against a 7209 // constant, try to simplify the block. 7210 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 7211 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 7212 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 7213 ; 7214 if (I->isTerminator() && 7215 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder)) 7216 return true; 7217 } 7218 7219 // See if we can merge an empty landing pad block with another which is 7220 // equivalent. 7221 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) { 7222 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 7223 ; 7224 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB, DTU)) 7225 return true; 7226 } 7227 7228 // If this basic block is ONLY a compare and a branch, and if a predecessor 7229 // branches to us and our successor, fold the comparison into the 7230 // predecessor and use logical operations to update the incoming value 7231 // for PHI nodes in common successor. 7232 if (Options.SpeculateBlocks && 7233 FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI, 7234 Options.BonusInstThreshold)) 7235 return requestResimplify(); 7236 return false; 7237 } 7238 7239 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) { 7240 BasicBlock *PredPred = nullptr; 7241 for (auto *P : predecessors(BB)) { 7242 BasicBlock *PPred = P->getSinglePredecessor(); 7243 if (!PPred || (PredPred && PredPred != PPred)) 7244 return nullptr; 7245 PredPred = PPred; 7246 } 7247 return PredPred; 7248 } 7249 7250 bool SimplifyCFGOpt::simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 7251 assert( 7252 !isa<ConstantInt>(BI->getCondition()) && 7253 BI->getSuccessor(0) != BI->getSuccessor(1) && 7254 "Tautological conditional branch should have been eliminated already."); 7255 7256 BasicBlock *BB = BI->getParent(); 7257 if (!Options.SimplifyCondBranch || 7258 BI->getFunction()->hasFnAttribute(Attribute::OptForFuzzing)) 7259 return false; 7260 7261 // Conditional branch 7262 if (isValueEqualityComparison(BI)) { 7263 // If we only have one predecessor, and if it is a branch on this value, 7264 // see if that predecessor totally determines the outcome of this 7265 // switch. 7266 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 7267 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 7268 return requestResimplify(); 7269 7270 // This block must be empty, except for the setcond inst, if it exists. 7271 // Ignore dbg and pseudo intrinsics. 7272 auto I = BB->instructionsWithoutDebug(true).begin(); 7273 if (&*I == BI) { 7274 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 7275 return requestResimplify(); 7276 } else if (&*I == cast<Instruction>(BI->getCondition())) { 7277 ++I; 7278 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 7279 return requestResimplify(); 7280 } 7281 } 7282 7283 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 7284 if (SimplifyBranchOnICmpChain(BI, Builder, DL)) 7285 return true; 7286 7287 // If this basic block has dominating predecessor blocks and the dominating 7288 // blocks' conditions imply BI's condition, we know the direction of BI. 7289 std::optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL); 7290 if (Imp) { 7291 // Turn this into a branch on constant. 7292 auto *OldCond = BI->getCondition(); 7293 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext()) 7294 : ConstantInt::getFalse(BB->getContext()); 7295 BI->setCondition(TorF); 7296 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 7297 return requestResimplify(); 7298 } 7299 7300 // If this basic block is ONLY a compare and a branch, and if a predecessor 7301 // branches to us and one of our successors, fold the comparison into the 7302 // predecessor and use logical operations to pick the right destination. 7303 if (Options.SpeculateBlocks && 7304 FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI, 7305 Options.BonusInstThreshold)) 7306 return requestResimplify(); 7307 7308 // We have a conditional branch to two blocks that are only reachable 7309 // from BI. We know that the condbr dominates the two blocks, so see if 7310 // there is any identical code in the "then" and "else" blocks. If so, we 7311 // can hoist it up to the branching block. 7312 if (BI->getSuccessor(0)->getSinglePredecessor()) { 7313 if (BI->getSuccessor(1)->getSinglePredecessor()) { 7314 if (HoistCommon && hoistCommonCodeFromSuccessors( 7315 BI->getParent(), !Options.HoistCommonInsts)) 7316 return requestResimplify(); 7317 } else { 7318 // If Successor #1 has multiple preds, we may be able to conditionally 7319 // execute Successor #0 if it branches to Successor #1. 7320 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator(); 7321 if (Succ0TI->getNumSuccessors() == 1 && 7322 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 7323 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 7324 return requestResimplify(); 7325 } 7326 } else if (BI->getSuccessor(1)->getSinglePredecessor()) { 7327 // If Successor #0 has multiple preds, we may be able to conditionally 7328 // execute Successor #1 if it branches to Successor #0. 7329 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator(); 7330 if (Succ1TI->getNumSuccessors() == 1 && 7331 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 7332 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 7333 return requestResimplify(); 7334 } 7335 7336 // If this is a branch on something for which we know the constant value in 7337 // predecessors (e.g. a phi node in the current block), thread control 7338 // through this block. 7339 if (FoldCondBranchOnValueKnownInPredecessor(BI, DTU, DL, Options.AC)) 7340 return requestResimplify(); 7341 7342 // Scan predecessor blocks for conditional branches. 7343 for (BasicBlock *Pred : predecessors(BB)) 7344 if (BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator())) 7345 if (PBI != BI && PBI->isConditional()) 7346 if (SimplifyCondBranchToCondBranch(PBI, BI, DTU, DL, TTI)) 7347 return requestResimplify(); 7348 7349 // Look for diamond patterns. 7350 if (MergeCondStores) 7351 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB)) 7352 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator())) 7353 if (PBI != BI && PBI->isConditional()) 7354 if (mergeConditionalStores(PBI, BI, DTU, DL, TTI)) 7355 return requestResimplify(); 7356 7357 return false; 7358 } 7359 7360 /// Check if passing a value to an instruction will cause undefined behavior. 7361 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified) { 7362 Constant *C = dyn_cast<Constant>(V); 7363 if (!C) 7364 return false; 7365 7366 if (I->use_empty()) 7367 return false; 7368 7369 if (C->isNullValue() || isa<UndefValue>(C)) { 7370 // Only look at the first use, avoid hurting compile time with long uselists 7371 auto *Use = cast<Instruction>(*I->user_begin()); 7372 // Bail out if Use is not in the same BB as I or Use == I or Use comes 7373 // before I in the block. The latter two can be the case if Use is a PHI 7374 // node. 7375 if (Use->getParent() != I->getParent() || Use == I || Use->comesBefore(I)) 7376 return false; 7377 7378 // Now make sure that there are no instructions in between that can alter 7379 // control flow (eg. calls) 7380 auto InstrRange = 7381 make_range(std::next(I->getIterator()), Use->getIterator()); 7382 if (any_of(InstrRange, [](Instruction &I) { 7383 return !isGuaranteedToTransferExecutionToSuccessor(&I); 7384 })) 7385 return false; 7386 7387 // Look through GEPs. A load from a GEP derived from NULL is still undefined 7388 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 7389 if (GEP->getPointerOperand() == I) { 7390 if (!GEP->isInBounds() || !GEP->hasAllZeroIndices()) 7391 PtrValueMayBeModified = true; 7392 return passingValueIsAlwaysUndefined(V, GEP, PtrValueMayBeModified); 7393 } 7394 7395 // Look through bitcasts. 7396 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 7397 return passingValueIsAlwaysUndefined(V, BC, PtrValueMayBeModified); 7398 7399 // Load from null is undefined. 7400 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 7401 if (!LI->isVolatile()) 7402 return !NullPointerIsDefined(LI->getFunction(), 7403 LI->getPointerAddressSpace()); 7404 7405 // Store to null is undefined. 7406 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 7407 if (!SI->isVolatile()) 7408 return (!NullPointerIsDefined(SI->getFunction(), 7409 SI->getPointerAddressSpace())) && 7410 SI->getPointerOperand() == I; 7411 7412 if (auto *CB = dyn_cast<CallBase>(Use)) { 7413 if (C->isNullValue() && NullPointerIsDefined(CB->getFunction())) 7414 return false; 7415 // A call to null is undefined. 7416 if (CB->getCalledOperand() == I) 7417 return true; 7418 7419 if (C->isNullValue()) { 7420 for (const llvm::Use &Arg : CB->args()) 7421 if (Arg == I) { 7422 unsigned ArgIdx = CB->getArgOperandNo(&Arg); 7423 if (CB->isPassingUndefUB(ArgIdx) && 7424 CB->paramHasAttr(ArgIdx, Attribute::NonNull)) { 7425 // Passing null to a nonnnull+noundef argument is undefined. 7426 return !PtrValueMayBeModified; 7427 } 7428 } 7429 } else if (isa<UndefValue>(C)) { 7430 // Passing undef to a noundef argument is undefined. 7431 for (const llvm::Use &Arg : CB->args()) 7432 if (Arg == I) { 7433 unsigned ArgIdx = CB->getArgOperandNo(&Arg); 7434 if (CB->isPassingUndefUB(ArgIdx)) { 7435 // Passing undef to a noundef argument is undefined. 7436 return true; 7437 } 7438 } 7439 } 7440 } 7441 } 7442 return false; 7443 } 7444 7445 /// If BB has an incoming value that will always trigger undefined behavior 7446 /// (eg. null pointer dereference), remove the branch leading here. 7447 static bool removeUndefIntroducingPredecessor(BasicBlock *BB, 7448 DomTreeUpdater *DTU, 7449 AssumptionCache *AC) { 7450 for (PHINode &PHI : BB->phis()) 7451 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) 7452 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) { 7453 BasicBlock *Predecessor = PHI.getIncomingBlock(i); 7454 Instruction *T = Predecessor->getTerminator(); 7455 IRBuilder<> Builder(T); 7456 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 7457 BB->removePredecessor(Predecessor); 7458 // Turn unconditional branches into unreachables and remove the dead 7459 // destination from conditional branches. 7460 if (BI->isUnconditional()) 7461 Builder.CreateUnreachable(); 7462 else { 7463 // Preserve guarding condition in assume, because it might not be 7464 // inferrable from any dominating condition. 7465 Value *Cond = BI->getCondition(); 7466 CallInst *Assumption; 7467 if (BI->getSuccessor(0) == BB) 7468 Assumption = Builder.CreateAssumption(Builder.CreateNot(Cond)); 7469 else 7470 Assumption = Builder.CreateAssumption(Cond); 7471 if (AC) 7472 AC->registerAssumption(cast<AssumeInst>(Assumption)); 7473 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) 7474 : BI->getSuccessor(0)); 7475 } 7476 BI->eraseFromParent(); 7477 if (DTU) 7478 DTU->applyUpdates({{DominatorTree::Delete, Predecessor, BB}}); 7479 return true; 7480 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) { 7481 // Redirect all branches leading to UB into 7482 // a newly created unreachable block. 7483 BasicBlock *Unreachable = BasicBlock::Create( 7484 Predecessor->getContext(), "unreachable", BB->getParent(), BB); 7485 Builder.SetInsertPoint(Unreachable); 7486 // The new block contains only one instruction: Unreachable 7487 Builder.CreateUnreachable(); 7488 for (const auto &Case : SI->cases()) 7489 if (Case.getCaseSuccessor() == BB) { 7490 BB->removePredecessor(Predecessor); 7491 Case.setSuccessor(Unreachable); 7492 } 7493 if (SI->getDefaultDest() == BB) { 7494 BB->removePredecessor(Predecessor); 7495 SI->setDefaultDest(Unreachable); 7496 } 7497 7498 if (DTU) 7499 DTU->applyUpdates( 7500 { { DominatorTree::Insert, Predecessor, Unreachable }, 7501 { DominatorTree::Delete, Predecessor, BB } }); 7502 return true; 7503 } 7504 } 7505 7506 return false; 7507 } 7508 7509 bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) { 7510 bool Changed = false; 7511 7512 assert(BB && BB->getParent() && "Block not embedded in function!"); 7513 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 7514 7515 // Remove basic blocks that have no predecessors (except the entry block)... 7516 // or that just have themself as a predecessor. These are unreachable. 7517 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) || 7518 BB->getSinglePredecessor() == BB) { 7519 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB); 7520 DeleteDeadBlock(BB, DTU); 7521 return true; 7522 } 7523 7524 // Check to see if we can constant propagate this terminator instruction 7525 // away... 7526 Changed |= ConstantFoldTerminator(BB, /*DeleteDeadConditions=*/true, 7527 /*TLI=*/nullptr, DTU); 7528 7529 // Check for and eliminate duplicate PHI nodes in this block. 7530 Changed |= EliminateDuplicatePHINodes(BB); 7531 7532 // Check for and remove branches that will always cause undefined behavior. 7533 if (removeUndefIntroducingPredecessor(BB, DTU, Options.AC)) 7534 return requestResimplify(); 7535 7536 // Merge basic blocks into their predecessor if there is only one distinct 7537 // pred, and if there is only one distinct successor of the predecessor, and 7538 // if there are no PHI nodes. 7539 if (MergeBlockIntoPredecessor(BB, DTU)) 7540 return true; 7541 7542 if (SinkCommon && Options.SinkCommonInsts) 7543 if (SinkCommonCodeFromPredecessors(BB, DTU) || 7544 MergeCompatibleInvokes(BB, DTU)) { 7545 // SinkCommonCodeFromPredecessors() does not automatically CSE PHI's, 7546 // so we may now how duplicate PHI's. 7547 // Let's rerun EliminateDuplicatePHINodes() first, 7548 // before FoldTwoEntryPHINode() potentially converts them into select's, 7549 // after which we'd need a whole EarlyCSE pass run to cleanup them. 7550 return true; 7551 } 7552 7553 IRBuilder<> Builder(BB); 7554 7555 if (Options.SpeculateBlocks && 7556 !BB->getParent()->hasFnAttribute(Attribute::OptForFuzzing)) { 7557 // If there is a trivial two-entry PHI node in this basic block, and we can 7558 // eliminate it, do so now. 7559 if (auto *PN = dyn_cast<PHINode>(BB->begin())) 7560 if (PN->getNumIncomingValues() == 2) 7561 if (FoldTwoEntryPHINode(PN, TTI, DTU, DL)) 7562 return true; 7563 } 7564 7565 Instruction *Terminator = BB->getTerminator(); 7566 Builder.SetInsertPoint(Terminator); 7567 switch (Terminator->getOpcode()) { 7568 case Instruction::Br: 7569 Changed |= simplifyBranch(cast<BranchInst>(Terminator), Builder); 7570 break; 7571 case Instruction::Resume: 7572 Changed |= simplifyResume(cast<ResumeInst>(Terminator), Builder); 7573 break; 7574 case Instruction::CleanupRet: 7575 Changed |= simplifyCleanupReturn(cast<CleanupReturnInst>(Terminator)); 7576 break; 7577 case Instruction::Switch: 7578 Changed |= simplifySwitch(cast<SwitchInst>(Terminator), Builder); 7579 break; 7580 case Instruction::Unreachable: 7581 Changed |= simplifyUnreachable(cast<UnreachableInst>(Terminator)); 7582 break; 7583 case Instruction::IndirectBr: 7584 Changed |= simplifyIndirectBr(cast<IndirectBrInst>(Terminator)); 7585 break; 7586 } 7587 7588 return Changed; 7589 } 7590 7591 bool SimplifyCFGOpt::run(BasicBlock *BB) { 7592 bool Changed = false; 7593 7594 // Repeated simplify BB as long as resimplification is requested. 7595 do { 7596 Resimplify = false; 7597 7598 // Perform one round of simplifcation. Resimplify flag will be set if 7599 // another iteration is requested. 7600 Changed |= simplifyOnce(BB); 7601 } while (Resimplify); 7602 7603 return Changed; 7604 } 7605 7606 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 7607 DomTreeUpdater *DTU, const SimplifyCFGOptions &Options, 7608 ArrayRef<WeakVH> LoopHeaders) { 7609 return SimplifyCFGOpt(TTI, DTU, BB->getModule()->getDataLayout(), LoopHeaders, 7610 Options) 7611 .run(BB); 7612 } 7613