brintos / llvm-project-archived public Read only

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-exception6//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/Loads.h"32#include "llvm/Analysis/MemorySSA.h"33#include "llvm/Analysis/MemorySSAUpdater.h"34#include "llvm/Analysis/TargetTransformInfo.h"35#include "llvm/Analysis/ValueTracking.h"36#include "llvm/IR/Attributes.h"37#include "llvm/IR/BasicBlock.h"38#include "llvm/IR/CFG.h"39#include "llvm/IR/Constant.h"40#include "llvm/IR/ConstantRange.h"41#include "llvm/IR/Constants.h"42#include "llvm/IR/DataLayout.h"43#include "llvm/IR/DebugInfo.h"44#include "llvm/IR/DerivedTypes.h"45#include "llvm/IR/Function.h"46#include "llvm/IR/GlobalValue.h"47#include "llvm/IR/GlobalVariable.h"48#include "llvm/IR/IRBuilder.h"49#include "llvm/IR/InstrTypes.h"50#include "llvm/IR/Instruction.h"51#include "llvm/IR/Instructions.h"52#include "llvm/IR/IntrinsicInst.h"53#include "llvm/IR/LLVMContext.h"54#include "llvm/IR/MDBuilder.h"55#include "llvm/IR/MemoryModelRelaxationAnnotations.h"56#include "llvm/IR/Metadata.h"57#include "llvm/IR/Module.h"58#include "llvm/IR/NoFolder.h"59#include "llvm/IR/Operator.h"60#include "llvm/IR/PatternMatch.h"61#include "llvm/IR/ProfDataUtils.h"62#include "llvm/IR/Type.h"63#include "llvm/IR/Use.h"64#include "llvm/IR/User.h"65#include "llvm/IR/Value.h"66#include "llvm/IR/ValueHandle.h"67#include "llvm/Support/BranchProbability.h"68#include "llvm/Support/Casting.h"69#include "llvm/Support/CommandLine.h"70#include "llvm/Support/Debug.h"71#include "llvm/Support/ErrorHandling.h"72#include "llvm/Support/KnownBits.h"73#include "llvm/Support/MathExtras.h"74#include "llvm/Support/raw_ostream.h"75#include "llvm/Transforms/Utils/BasicBlockUtils.h"76#include "llvm/Transforms/Utils/Cloning.h"77#include "llvm/Transforms/Utils/Local.h"78#include "llvm/Transforms/Utils/LockstepReverseIterator.h"79#include "llvm/Transforms/Utils/ValueMapper.h"80#include <algorithm>81#include <cassert>82#include <climits>83#include <cmath>84#include <cstddef>85#include <cstdint>86#include <iterator>87#include <map>88#include <optional>89#include <set>90#include <tuple>91#include <utility>92#include <vector>93 94using namespace llvm;95using namespace PatternMatch;96 97#define DEBUG_TYPE "simplifycfg"98 99namespace llvm {100 101cl::opt<bool> RequireAndPreserveDomTree(102    "simplifycfg-require-and-preserve-domtree", cl::Hidden,103 104    cl::desc(105        "Temporary development switch used to gradually uplift SimplifyCFG "106        "into preserving DomTree,"));107 108// Chosen as 2 so as to be cheap, but still to have enough power to fold109// a select, so the "clamp" idiom (of a min followed by a max) will be caught.110// To catch this, we need to fold a compare and a select, hence '2' being the111// minimum reasonable default.112static cl::opt<unsigned> PHINodeFoldingThreshold(113    "phi-node-folding-threshold", cl::Hidden, cl::init(2),114    cl::desc(115        "Control the amount of phi node folding to perform (default = 2)"));116 117static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold(118    "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),119    cl::desc("Control the maximal total instruction cost that we are willing "120             "to speculatively execute to fold a 2-entry PHI node into a "121             "select (default = 4)"));122 123static cl::opt<bool>124    HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true),125                cl::desc("Hoist common instructions up to the parent block"));126 127static cl::opt<bool> HoistLoadsWithCondFaulting(128    "simplifycfg-hoist-loads-with-cond-faulting", cl::Hidden, cl::init(true),129    cl::desc("Hoist loads if the target supports conditional faulting"));130 131static cl::opt<bool> HoistStoresWithCondFaulting(132    "simplifycfg-hoist-stores-with-cond-faulting", cl::Hidden, cl::init(true),133    cl::desc("Hoist stores if the target supports conditional faulting"));134 135static cl::opt<unsigned> HoistLoadsStoresWithCondFaultingThreshold(136    "hoist-loads-stores-with-cond-faulting-threshold", cl::Hidden, cl::init(6),137    cl::desc("Control the maximal conditional load/store that we are willing "138             "to speculatively execute to eliminate conditional branch "139             "(default = 6)"));140 141static cl::opt<unsigned>142    HoistCommonSkipLimit("simplifycfg-hoist-common-skip-limit", cl::Hidden,143                         cl::init(20),144                         cl::desc("Allow reordering across at most this many "145                                  "instructions when hoisting"));146 147static cl::opt<bool>148    SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),149               cl::desc("Sink common instructions down to the end block"));150 151static cl::opt<bool> HoistCondStores(152    "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),153    cl::desc("Hoist conditional stores if an unconditional store precedes"));154 155static cl::opt<bool> MergeCondStores(156    "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),157    cl::desc("Hoist conditional stores even if an unconditional store does not "158             "precede - hoist multiple conditional stores into a single "159             "predicated store"));160 161static cl::opt<bool> MergeCondStoresAggressively(162    "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),163    cl::desc("When merging conditional stores, do so even if the resultant "164             "basic blocks are unlikely to be if-converted as a result"));165 166static cl::opt<bool> SpeculateOneExpensiveInst(167    "speculate-one-expensive-inst", cl::Hidden, cl::init(true),168    cl::desc("Allow exactly one expensive instruction to be speculatively "169             "executed"));170 171static cl::opt<unsigned> MaxSpeculationDepth(172    "max-speculation-depth", cl::Hidden, cl::init(10),173    cl::desc("Limit maximum recursion depth when calculating costs of "174             "speculatively executed instructions"));175 176static cl::opt<int>177    MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden,178                      cl::init(10),179                      cl::desc("Max size of a block which is still considered "180                               "small enough to thread through"));181 182// Two is chosen to allow one negation and a logical combine.183static cl::opt<unsigned>184    BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden,185                        cl::init(2),186                        cl::desc("Maximum cost of combining conditions when "187                                 "folding branches"));188 189static cl::opt<unsigned> BranchFoldToCommonDestVectorMultiplier(190    "simplifycfg-branch-fold-common-dest-vector-multiplier", cl::Hidden,191    cl::init(2),192    cl::desc("Multiplier to apply to threshold when determining whether or not "193             "to fold branch to common destination when vector operations are "194             "present"));195 196static cl::opt<bool> EnableMergeCompatibleInvokes(197    "simplifycfg-merge-compatible-invokes", cl::Hidden, cl::init(true),198    cl::desc("Allow SimplifyCFG to merge invokes together when appropriate"));199 200static cl::opt<unsigned> MaxSwitchCasesPerResult(201    "max-switch-cases-per-result", cl::Hidden, cl::init(16),202    cl::desc("Limit cases to analyze when converting a switch to select"));203 204static cl::opt<unsigned> MaxJumpThreadingLiveBlocks(205    "max-jump-threading-live-blocks", cl::Hidden, cl::init(24),206    cl::desc("Limit number of blocks a define in a threaded block is allowed "207             "to be live in"));208 209extern cl::opt<bool> ProfcheckDisableMetadataFixes;210 211} // end namespace llvm212 213STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");214STATISTIC(NumLinearMaps,215          "Number of switch instructions turned into linear mapping");216STATISTIC(NumLookupTables,217          "Number of switch instructions turned into lookup tables");218STATISTIC(219    NumLookupTablesHoles,220    "Number of switch instructions turned into lookup tables (holes checked)");221STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");222STATISTIC(NumFoldValueComparisonIntoPredecessors,223          "Number of value comparisons folded into predecessor basic blocks");224STATISTIC(NumFoldBranchToCommonDest,225          "Number of branches folded into predecessor basic block");226STATISTIC(227    NumHoistCommonCode,228    "Number of common instruction 'blocks' hoisted up to the begin block");229STATISTIC(NumHoistCommonInstrs,230          "Number of common instructions hoisted up to the begin block");231STATISTIC(NumSinkCommonCode,232          "Number of common instruction 'blocks' sunk down to the end block");233STATISTIC(NumSinkCommonInstrs,234          "Number of common instructions sunk down to the end block");235STATISTIC(NumSpeculations, "Number of speculative executed instructions");236STATISTIC(NumInvokes,237          "Number of invokes with empty resume blocks simplified into calls");238STATISTIC(NumInvokesMerged, "Number of invokes that were merged together");239STATISTIC(NumInvokeSetsFormed, "Number of invoke sets that were formed");240 241namespace {242 243// The first field contains the value that the switch produces when a certain244// case group is selected, and the second field is a vector containing the245// cases composing the case group.246using SwitchCaseResultVectorTy =247    SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;248 249// The first field contains the phi node that generates a result of the switch250// and the second field contains the value generated for a certain case in the251// switch for that PHI.252using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;253 254/// ValueEqualityComparisonCase - Represents a case of a switch.255struct ValueEqualityComparisonCase {256  ConstantInt *Value;257  BasicBlock *Dest;258 259  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)260      : Value(Value), Dest(Dest) {}261 262  bool operator<(ValueEqualityComparisonCase RHS) const {263    // Comparing pointers is ok as we only rely on the order for uniquing.264    return Value < RHS.Value;265  }266 267  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }268};269 270class SimplifyCFGOpt {271  const TargetTransformInfo &TTI;272  DomTreeUpdater *DTU;273  const DataLayout &DL;274  ArrayRef<WeakVH> LoopHeaders;275  const SimplifyCFGOptions &Options;276  bool Resimplify;277 278  Value *isValueEqualityComparison(Instruction *TI);279  BasicBlock *getValueEqualityComparisonCases(280      Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);281  bool simplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,282                                                     BasicBlock *Pred,283                                                     IRBuilder<> &Builder);284  bool performValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV,285                                                    Instruction *PTI,286                                                    IRBuilder<> &Builder);287  bool foldValueComparisonIntoPredecessors(Instruction *TI,288                                           IRBuilder<> &Builder);289 290  bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);291  bool simplifySingleResume(ResumeInst *RI);292  bool simplifyCommonResume(ResumeInst *RI);293  bool simplifyCleanupReturn(CleanupReturnInst *RI);294  bool simplifyUnreachable(UnreachableInst *UI);295  bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);296  bool simplifyDuplicateSwitchArms(SwitchInst *SI, DomTreeUpdater *DTU);297  bool simplifyIndirectBr(IndirectBrInst *IBI);298  bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);299  bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);300  bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);301  bool foldCondBranchOnValueKnownInPredecessor(BranchInst *BI);302 303  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,304                                             IRBuilder<> &Builder);305  bool tryToSimplifyUncondBranchWithICmpSelectInIt(ICmpInst *ICI,306                                                   SelectInst *Select,307                                                   IRBuilder<> &Builder);308  bool hoistCommonCodeFromSuccessors(Instruction *TI, bool AllInstsEqOnly);309  bool hoistSuccIdenticalTerminatorToSwitchOrIf(310      Instruction *TI, Instruction *I1,311      SmallVectorImpl<Instruction *> &OtherSuccTIs);312  bool speculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB);313  bool simplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,314                                  BasicBlock *TrueBB, BasicBlock *FalseBB,315                                  uint32_t TrueWeight, uint32_t FalseWeight);316  bool simplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,317                                 const DataLayout &DL);318  bool simplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);319  bool simplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);320  bool turnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);321 322public:323  SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU,324                 const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders,325                 const SimplifyCFGOptions &Opts)326      : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {327    assert((!DTU || !DTU->hasPostDomTree()) &&328           "SimplifyCFG is not yet capable of maintaining validity of a "329           "PostDomTree, so don't ask for it.");330  }331 332  bool simplifyOnce(BasicBlock *BB);333  bool run(BasicBlock *BB);334 335  // Helper to set Resimplify and return change indication.336  bool requestResimplify() {337    Resimplify = true;338    return true;339  }340};341 342// we synthesize a || b as select a, true, b343// we synthesize a && b as select a, b, false344// this function determines if SI is playing one of those roles.345[[maybe_unused]] bool346isSelectInRoleOfConjunctionOrDisjunction(const SelectInst *SI) {347  return ((isa<ConstantInt>(SI->getTrueValue()) &&348           (dyn_cast<ConstantInt>(SI->getTrueValue())->isOne())) ||349          (isa<ConstantInt>(SI->getFalseValue()) &&350           (dyn_cast<ConstantInt>(SI->getFalseValue())->isNullValue())));351}352 353} // end anonymous namespace354 355/// Return true if all the PHI nodes in the basic block \p BB356/// receive compatible (identical) incoming values when coming from357/// all of the predecessor blocks that are specified in \p IncomingBlocks.358///359/// Note that if the values aren't exactly identical, but \p EquivalenceSet360/// is provided, and *both* of the values are present in the set,361/// then they are considered equal.362static bool incomingValuesAreCompatible(363    BasicBlock *BB, ArrayRef<BasicBlock *> IncomingBlocks,364    SmallPtrSetImpl<Value *> *EquivalenceSet = nullptr) {365  assert(IncomingBlocks.size() == 2 &&366         "Only for a pair of incoming blocks at the time!");367 368  // FIXME: it is okay if one of the incoming values is an `undef` value,369  //        iff the other incoming value is guaranteed to be a non-poison value.370  // FIXME: it is okay if one of the incoming values is a `poison` value.371  return all_of(BB->phis(), [IncomingBlocks, EquivalenceSet](PHINode &PN) {372    Value *IV0 = PN.getIncomingValueForBlock(IncomingBlocks[0]);373    Value *IV1 = PN.getIncomingValueForBlock(IncomingBlocks[1]);374    if (IV0 == IV1)375      return true;376    if (EquivalenceSet && EquivalenceSet->contains(IV0) &&377        EquivalenceSet->contains(IV1))378      return true;379    return false;380  });381}382 383/// Return true if it is safe to merge these two384/// terminator instructions together.385static bool386safeToMergeTerminators(Instruction *SI1, Instruction *SI2,387                       SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {388  if (SI1 == SI2)389    return false; // Can't merge with self!390 391  // It is not safe to merge these two switch instructions if they have a common392  // successor, and if that successor has a PHI node, and if *that* PHI node has393  // conflicting incoming values from the two switch blocks.394  BasicBlock *SI1BB = SI1->getParent();395  BasicBlock *SI2BB = SI2->getParent();396 397  SmallPtrSet<BasicBlock *, 16> SI1Succs(llvm::from_range, successors(SI1BB));398  bool Fail = false;399  for (BasicBlock *Succ : successors(SI2BB)) {400    if (!SI1Succs.count(Succ))401      continue;402    if (incomingValuesAreCompatible(Succ, {SI1BB, SI2BB}))403      continue;404    Fail = true;405    if (FailBlocks)406      FailBlocks->insert(Succ);407    else408      break;409  }410 411  return !Fail;412}413 414/// Update PHI nodes in Succ to indicate that there will now be entries in it415/// from the 'NewPred' block. The values that will be flowing into the PHI nodes416/// will be the same as those coming in from ExistPred, an existing predecessor417/// of Succ.418static void addPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,419                                  BasicBlock *ExistPred,420                                  MemorySSAUpdater *MSSAU = nullptr) {421  for (PHINode &PN : Succ->phis())422    PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);423  if (MSSAU)424    if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))425      MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);426}427 428/// Compute an abstract "cost" of speculating the given instruction,429/// which is assumed to be safe to speculate. TCC_Free means cheap,430/// TCC_Basic means less cheap, and TCC_Expensive means prohibitively431/// expensive.432static InstructionCost computeSpeculationCost(const User *I,433                                              const TargetTransformInfo &TTI) {434  return TTI.getInstructionCost(I, TargetTransformInfo::TCK_SizeAndLatency);435}436 437/// If we have a merge point of an "if condition" as accepted above,438/// return true if the specified value dominates the block.  We don't handle439/// the true generality of domination here, just a special case which works440/// well enough for us.441///442/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to443/// see if V (which must be an instruction) and its recursive operands444/// that do not dominate BB have a combined cost lower than Budget and445/// are non-trapping.  If both are true, the instruction is inserted into the446/// set and true is returned.447///448/// The cost for most non-trapping instructions is defined as 1 except for449/// Select whose cost is 2.450///451/// After this function returns, Cost is increased by the cost of452/// V plus its non-dominating operands.  If that cost is greater than453/// Budget, false is returned and Cost is undefined.454static bool dominatesMergePoint(455    Value *V, BasicBlock *BB, Instruction *InsertPt,456    SmallPtrSetImpl<Instruction *> &AggressiveInsts, InstructionCost &Cost,457    InstructionCost Budget, const TargetTransformInfo &TTI, AssumptionCache *AC,458    SmallPtrSetImpl<Instruction *> &ZeroCostInstructions, unsigned Depth = 0) {459  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),460  // so limit the recursion depth.461  // TODO: While this recursion limit does prevent pathological behavior, it462  // would be better to track visited instructions to avoid cycles.463  if (Depth == MaxSpeculationDepth)464    return false;465 466  Instruction *I = dyn_cast<Instruction>(V);467  if (!I) {468    // Non-instructions dominate all instructions and can be executed469    // unconditionally.470    return true;471  }472  BasicBlock *PBB = I->getParent();473 474  // We don't want to allow weird loops that might have the "if condition" in475  // the bottom of this block.476  if (PBB == BB)477    return false;478 479  // If this instruction is defined in a block that contains an unconditional480  // branch to BB, then it must be in the 'conditional' part of the "if481  // statement".  If not, it definitely dominates the region.482  BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());483  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)484    return true;485 486  // If we have seen this instruction before, don't count it again.487  if (AggressiveInsts.count(I))488    return true;489 490  // Okay, it looks like the instruction IS in the "condition".  Check to491  // see if it's a cheap instruction to unconditionally compute, and if it492  // only uses stuff defined outside of the condition.  If so, hoist it out.493  if (!isSafeToSpeculativelyExecute(I, InsertPt, AC))494    return false;495 496  // Overflow arithmetic instruction plus extract value are usually generated497  // when a division is being replaced. But, in this case, the zero check may498  // still be kept in the code. In that case it would be worth to hoist these499  // two instruction out of the basic block. Let's treat this pattern as one500  // single cheap instruction here!501  WithOverflowInst *OverflowInst;502  if (match(I, m_ExtractValue<1>(m_OneUse(m_WithOverflowInst(OverflowInst))))) {503    ZeroCostInstructions.insert(OverflowInst);504    Cost += 1;505  } else if (!ZeroCostInstructions.contains(I))506    Cost += computeSpeculationCost(I, TTI);507 508  // Allow exactly one instruction to be speculated regardless of its cost509  // (as long as it is safe to do so).510  // This is intended to flatten the CFG even if the instruction is a division511  // or other expensive operation. The speculation of an expensive instruction512  // is expected to be undone in CodeGenPrepare if the speculation has not513  // enabled further IR optimizations.514  if (Cost > Budget &&515      (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 ||516       !Cost.isValid()))517    return false;518 519  // Okay, we can only really hoist these out if their operands do520  // not take us over the cost threshold.521  for (Use &Op : I->operands())522    if (!dominatesMergePoint(Op, BB, InsertPt, AggressiveInsts, Cost, Budget,523                             TTI, AC, ZeroCostInstructions, Depth + 1))524      return false;525  // Okay, it's safe to do this!  Remember this instruction.526  AggressiveInsts.insert(I);527  return true;528}529 530/// Extract ConstantInt from value, looking through IntToPtr531/// and PointerNullValue. Return NULL if value is not a constant int.532static ConstantInt *getConstantInt(Value *V, const DataLayout &DL) {533  // Normal constant int.534  ConstantInt *CI = dyn_cast<ConstantInt>(V);535  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())536    return CI;537 538  // It is not safe to look through inttoptr or ptrtoint when using unstable539  // pointer types.540  if (DL.hasUnstableRepresentation(V->getType()))541    return nullptr;542 543  // This is some kind of pointer constant. Turn it into a pointer-sized544  // ConstantInt if possible.545  IntegerType *IntPtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));546 547  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).548  if (isa<ConstantPointerNull>(V))549    return ConstantInt::get(IntPtrTy, 0);550 551  // IntToPtr const int, we can look through this if the semantics of552  // inttoptr for this address space are a simple (truncating) bitcast.553  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))554    if (CE->getOpcode() == Instruction::IntToPtr)555      if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {556        // The constant is very likely to have the right type already.557        if (CI->getType() == IntPtrTy)558          return CI;559        else560          return cast<ConstantInt>(561              ConstantFoldIntegerCast(CI, IntPtrTy, /*isSigned=*/false, DL));562      }563  return nullptr;564}565 566namespace {567 568/// Given a chain of or (||) or and (&&) comparison of a value against a569/// constant, this will try to recover the information required for a switch570/// structure.571/// It will depth-first traverse the chain of comparison, seeking for patterns572/// like %a == 12 or %a < 4 and combine them to produce a set of integer573/// representing the different cases for the switch.574/// Note that if the chain is composed of '||' it will build the set of elements575/// that matches the comparisons (i.e. any of this value validate the chain)576/// while for a chain of '&&' it will build the set elements that make the test577/// fail.578struct ConstantComparesGatherer {579  const DataLayout &DL;580 581  /// Value found for the switch comparison582  Value *CompValue = nullptr;583 584  /// Extra clause to be checked before the switch585  Value *Extra = nullptr;586 587  /// Set of integers to match in switch588  SmallVector<ConstantInt *, 8> Vals;589 590  /// Number of comparisons matched in the and/or chain591  unsigned UsedICmps = 0;592 593  /// If the elements in Vals matches the comparisons594  bool IsEq = false;595 596  // Used to check if the first matched CompValue shall be the Extra check.597  bool IgnoreFirstMatch = false;598  bool MultipleMatches = false;599 600  /// Construct and compute the result for the comparison instruction Cond601  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {602    gather(Cond);603    if (CompValue || !MultipleMatches)604      return;605    Extra = nullptr;606    Vals.clear();607    UsedICmps = 0;608    IgnoreFirstMatch = true;609    gather(Cond);610  }611 612  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;613  ConstantComparesGatherer &614  operator=(const ConstantComparesGatherer &) = delete;615 616private:617  /// Try to set the current value used for the comparison, it succeeds only if618  /// it wasn't set before or if the new value is the same as the old one619  bool setValueOnce(Value *NewVal) {620    if (IgnoreFirstMatch) {621      IgnoreFirstMatch = false;622      return false;623    }624    if (CompValue && CompValue != NewVal) {625      MultipleMatches = true;626      return false;627    }628    CompValue = NewVal;629    return true;630  }631 632  /// Try to match Instruction "I" as a comparison against a constant and633  /// populates the array Vals with the set of values that match (or do not634  /// match depending on isEQ).635  /// Return false on failure. On success, the Value the comparison matched636  /// against is placed in CompValue.637  /// If CompValue is already set, the function is expected to fail if a match638  /// is found but the value compared to is different.639  bool matchInstruction(Instruction *I, bool isEQ) {640    if (match(I, m_Not(m_Instruction(I))))641      isEQ = !isEQ;642 643    Value *Val;644    if (match(I, m_NUWTrunc(m_Value(Val)))) {645      // If we already have a value for the switch, it has to match!646      if (!setValueOnce(Val))647        return false;648      UsedICmps++;649      Vals.push_back(ConstantInt::get(cast<IntegerType>(Val->getType()), isEQ));650      return true;651    }652    // If this is an icmp against a constant, handle this as one of the cases.653    ICmpInst *ICI;654    ConstantInt *C;655    if (!((ICI = dyn_cast<ICmpInst>(I)) &&656          (C = getConstantInt(I->getOperand(1), DL)))) {657      return false;658    }659 660    Value *RHSVal;661    const APInt *RHSC;662 663    // Pattern match a special case664    // (x & ~2^z) == y --> x == y || x == y|2^z665    // This undoes a transformation done by instcombine to fuse 2 compares.666    if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {667      // It's a little bit hard to see why the following transformations are668      // correct. Here is a CVC3 program to verify them for 64-bit values:669 670      /*671         ONE  : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);672         x    : BITVECTOR(64);673         y    : BITVECTOR(64);674         z    : BITVECTOR(64);675         mask : BITVECTOR(64) = BVSHL(ONE, z);676         QUERY( (y & ~mask = y) =>677                ((x & ~mask = y) <=> (x = y OR x = (y |  mask)))678         );679         QUERY( (y |  mask = y) =>680                ((x |  mask = y) <=> (x = y OR x = (y & ~mask)))681         );682      */683 684      // Please note that each pattern must be a dual implication (<--> or685      // iff). One directional implication can create spurious matches. If the686      // implication is only one-way, an unsatisfiable condition on the left687      // side can imply a satisfiable condition on the right side. Dual688      // implication ensures that satisfiable conditions are transformed to689      // other satisfiable conditions and unsatisfiable conditions are690      // transformed to other unsatisfiable conditions.691 692      // Here is a concrete example of a unsatisfiable condition on the left693      // implying a satisfiable condition on the right:694      //695      // mask = (1 << z)696      // (x & ~mask) == y  --> (x == y || x == (y | mask))697      //698      // Substituting y = 3, z = 0 yields:699      // (x & -2) == 3 --> (x == 3 || x == 2)700 701      // Pattern match a special case:702      /*703        QUERY( (y & ~mask = y) =>704               ((x & ~mask = y) <=> (x = y OR x = (y |  mask)))705        );706      */707      if (match(ICI->getOperand(0),708                m_And(m_Value(RHSVal), m_APInt(RHSC)))) {709        APInt Mask = ~*RHSC;710        if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {711          // If we already have a value for the switch, it has to match!712          if (!setValueOnce(RHSVal))713            return false;714 715          Vals.push_back(C);716          Vals.push_back(717              ConstantInt::get(C->getContext(),718                               C->getValue() | Mask));719          UsedICmps++;720          return true;721        }722      }723 724      // Pattern match a special case:725      /*726        QUERY( (y |  mask = y) =>727               ((x |  mask = y) <=> (x = y OR x = (y & ~mask)))728        );729      */730      if (match(ICI->getOperand(0),731                m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {732        APInt Mask = *RHSC;733        if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {734          // If we already have a value for the switch, it has to match!735          if (!setValueOnce(RHSVal))736            return false;737 738          Vals.push_back(C);739          Vals.push_back(ConstantInt::get(C->getContext(),740                                          C->getValue() & ~Mask));741          UsedICmps++;742          return true;743        }744      }745 746      // If we already have a value for the switch, it has to match!747      if (!setValueOnce(ICI->getOperand(0)))748        return false;749 750      UsedICmps++;751      Vals.push_back(C);752      return true;753    }754 755    // If we have "x ult 3", for example, then we can add 0,1,2 to the set.756    ConstantRange Span =757        ConstantRange::makeExactICmpRegion(ICI->getPredicate(), C->getValue());758 759    // Shift the range if the compare is fed by an add. This is the range760    // compare idiom as emitted by instcombine.761    Value *CandidateVal = I->getOperand(0);762    if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {763      Span = Span.subtract(*RHSC);764      CandidateVal = RHSVal;765    }766 767    // If this is an and/!= check, then we are looking to build the set of768    // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into769    // x != 0 && x != 1.770    if (!isEQ)771      Span = Span.inverse();772 773    // If there are a ton of values, we don't want to make a ginormous switch.774    if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {775      return false;776    }777 778    // If we already have a value for the switch, it has to match!779    if (!setValueOnce(CandidateVal))780      return false;781 782    // Add all values from the range to the set783    APInt Tmp = Span.getLower();784    do785      Vals.push_back(ConstantInt::get(I->getContext(), Tmp));786    while (++Tmp != Span.getUpper());787 788    UsedICmps++;789    return true;790  }791 792  /// Given a potentially 'or'd or 'and'd together collection of icmp793  /// eq/ne/lt/gt instructions that compare a value against a constant, extract794  /// the value being compared, and stick the list constants into the Vals795  /// vector.796  /// One "Extra" case is allowed to differ from the other.797  void gather(Value *V) {798    Value *Op0, *Op1;799    if (match(V, m_LogicalOr(m_Value(Op0), m_Value(Op1))))800      IsEq = true;801    else if (match(V, m_LogicalAnd(m_Value(Op0), m_Value(Op1))))802      IsEq = false;803    else804      return;805    // Keep a stack (SmallVector for efficiency) for depth-first traversal806    SmallVector<Value *, 8> DFT{Op0, Op1};807    SmallPtrSet<Value *, 8> Visited{V, Op0, Op1};808 809    while (!DFT.empty()) {810      V = DFT.pop_back_val();811 812      if (Instruction *I = dyn_cast<Instruction>(V)) {813        // If it is a || (or && depending on isEQ), process the operands.814        if (IsEq ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))815                 : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {816          if (Visited.insert(Op1).second)817            DFT.push_back(Op1);818          if (Visited.insert(Op0).second)819            DFT.push_back(Op0);820 821          continue;822        }823 824        // Try to match the current instruction825        if (matchInstruction(I, IsEq))826          // Match succeed, continue the loop827          continue;828      }829 830      // One element of the sequence of || (or &&) could not be match as a831      // comparison against the same value as the others.832      // We allow only one "Extra" case to be checked before the switch833      if (!Extra) {834        Extra = V;835        continue;836      }837      // Failed to parse a proper sequence, abort now838      CompValue = nullptr;839      break;840    }841  }842};843 844} // end anonymous namespace845 846static void eraseTerminatorAndDCECond(Instruction *TI,847                                      MemorySSAUpdater *MSSAU = nullptr) {848  Instruction *Cond = nullptr;849  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {850    Cond = dyn_cast<Instruction>(SI->getCondition());851  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {852    if (BI->isConditional())853      Cond = dyn_cast<Instruction>(BI->getCondition());854  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {855    Cond = dyn_cast<Instruction>(IBI->getAddress());856  }857 858  TI->eraseFromParent();859  if (Cond)860    RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);861}862 863/// Return true if the specified terminator checks864/// to see if a value is equal to constant integer value.865Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {866  Value *CV = nullptr;867  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {868    // Do not permit merging of large switch instructions into their869    // predecessors unless there is only one predecessor.870    if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))871      CV = SI->getCondition();872  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))873    if (BI->isConditional() && BI->getCondition()->hasOneUse()) {874      if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {875        if (ICI->isEquality() && getConstantInt(ICI->getOperand(1), DL))876          CV = ICI->getOperand(0);877      } else if (auto *Trunc = dyn_cast<TruncInst>(BI->getCondition())) {878        if (Trunc->hasNoUnsignedWrap())879          CV = Trunc->getOperand(0);880      }881    }882 883  // Unwrap any lossless ptrtoint cast (except for unstable pointers).884  if (CV) {885    if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {886      Value *Ptr = PTII->getPointerOperand();887      if (DL.hasUnstableRepresentation(Ptr->getType()))888        return CV;889      if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))890        CV = Ptr;891    }892  }893  return CV;894}895 896/// Given a value comparison instruction,897/// decode all of the 'cases' that it represents and return the 'default' block.898BasicBlock *SimplifyCFGOpt::getValueEqualityComparisonCases(899    Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {900  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {901    Cases.reserve(SI->getNumCases());902    for (auto Case : SI->cases())903      Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),904                                                  Case.getCaseSuccessor()));905    return SI->getDefaultDest();906  }907 908  BranchInst *BI = cast<BranchInst>(TI);909  Value *Cond = BI->getCondition();910  ICmpInst::Predicate Pred;911  ConstantInt *C;912  if (auto *ICI = dyn_cast<ICmpInst>(Cond)) {913    Pred = ICI->getPredicate();914    C = getConstantInt(ICI->getOperand(1), DL);915  } else {916    Pred = ICmpInst::ICMP_NE;917    auto *Trunc = cast<TruncInst>(Cond);918    C = ConstantInt::get(cast<IntegerType>(Trunc->getOperand(0)->getType()), 0);919  }920  BasicBlock *Succ = BI->getSuccessor(Pred == ICmpInst::ICMP_NE);921  Cases.push_back(ValueEqualityComparisonCase(C, Succ));922  return BI->getSuccessor(Pred == ICmpInst::ICMP_EQ);923}924 925/// Given a vector of bb/value pairs, remove any entries926/// in the list that match the specified block.927static void928eliminateBlockCases(BasicBlock *BB,929                    std::vector<ValueEqualityComparisonCase> &Cases) {930  llvm::erase(Cases, BB);931}932 933/// Return true if there are any keys in C1 that exist in C2 as well.934static bool valuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,935                          std::vector<ValueEqualityComparisonCase> &C2) {936  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;937 938  // Make V1 be smaller than V2.939  if (V1->size() > V2->size())940    std::swap(V1, V2);941 942  if (V1->empty())943    return false;944  if (V1->size() == 1) {945    // Just scan V2.946    ConstantInt *TheVal = (*V1)[0].Value;947    for (const ValueEqualityComparisonCase &VECC : *V2)948      if (TheVal == VECC.Value)949        return true;950  }951 952  // Otherwise, just sort both lists and compare element by element.953  array_pod_sort(V1->begin(), V1->end());954  array_pod_sort(V2->begin(), V2->end());955  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();956  while (i1 != e1 && i2 != e2) {957    if ((*V1)[i1].Value == (*V2)[i2].Value)958      return true;959    if ((*V1)[i1].Value < (*V2)[i2].Value)960      ++i1;961    else962      ++i2;963  }964  return false;965}966 967/// If TI is known to be a terminator instruction and its block is known to968/// only have a single predecessor block, check to see if that predecessor is969/// also a value comparison with the same value, and if that comparison970/// determines the outcome of this comparison. If so, simplify TI. This does a971/// very limited form of jump threading.972bool SimplifyCFGOpt::simplifyEqualityComparisonWithOnlyPredecessor(973    Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {974  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());975  if (!PredVal)976    return false; // Not a value comparison in predecessor.977 978  Value *ThisVal = isValueEqualityComparison(TI);979  assert(ThisVal && "This isn't a value comparison!!");980  if (ThisVal != PredVal)981    return false; // Different predicates.982 983  // TODO: Preserve branch weight metadata, similarly to how984  // foldValueComparisonIntoPredecessors preserves it.985 986  // Find out information about when control will move from Pred to TI's block.987  std::vector<ValueEqualityComparisonCase> PredCases;988  BasicBlock *PredDef =989      getValueEqualityComparisonCases(Pred->getTerminator(), PredCases);990  eliminateBlockCases(PredDef, PredCases); // Remove default from cases.991 992  // Find information about how control leaves this block.993  std::vector<ValueEqualityComparisonCase> ThisCases;994  BasicBlock *ThisDef = getValueEqualityComparisonCases(TI, ThisCases);995  eliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.996 997  // If TI's block is the default block from Pred's comparison, potentially998  // simplify TI based on this knowledge.999  if (PredDef == TI->getParent()) {1000    // If we are here, we know that the value is none of those cases listed in1001    // PredCases.  If there are any cases in ThisCases that are in PredCases, we1002    // can simplify TI.1003    if (!valuesOverlap(PredCases, ThisCases))1004      return false;1005 1006    if (isa<BranchInst>(TI)) {1007      // Okay, one of the successors of this condbr is dead.  Convert it to a1008      // uncond br.1009      assert(ThisCases.size() == 1 && "Branch can only have one case!");1010      // Insert the new branch.1011      Instruction *NI = Builder.CreateBr(ThisDef);1012      (void)NI;1013 1014      // Remove PHI node entries for the dead edge.1015      ThisCases[0].Dest->removePredecessor(PredDef);1016 1017      LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()1018                        << "Through successor TI: " << *TI << "Leaving: " << *NI1019                        << "\n");1020 1021      eraseTerminatorAndDCECond(TI);1022 1023      if (DTU)1024        DTU->applyUpdates(1025            {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});1026 1027      return true;1028    }1029 1030    SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);1031    // Okay, TI has cases that are statically dead, prune them away.1032    SmallPtrSet<Constant *, 16> DeadCases;1033    for (const ValueEqualityComparisonCase &Case : PredCases)1034      DeadCases.insert(Case.Value);1035 1036    LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()1037                      << "Through successor TI: " << *TI);1038 1039    SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;1040    for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {1041      --i;1042      auto *Successor = i->getCaseSuccessor();1043      if (DTU)1044        ++NumPerSuccessorCases[Successor];1045      if (DeadCases.count(i->getCaseValue())) {1046        Successor->removePredecessor(PredDef);1047        SI.removeCase(i);1048        if (DTU)1049          --NumPerSuccessorCases[Successor];1050      }1051    }1052 1053    if (DTU) {1054      std::vector<DominatorTree::UpdateType> Updates;1055      for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)1056        if (I.second == 0)1057          Updates.push_back({DominatorTree::Delete, PredDef, I.first});1058      DTU->applyUpdates(Updates);1059    }1060 1061    LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");1062    return true;1063  }1064 1065  // Otherwise, TI's block must correspond to some matched value.  Find out1066  // which value (or set of values) this is.1067  ConstantInt *TIV = nullptr;1068  BasicBlock *TIBB = TI->getParent();1069  for (const auto &[Value, Dest] : PredCases)1070    if (Dest == TIBB) {1071      if (TIV)1072        return false; // Cannot handle multiple values coming to this block.1073      TIV = Value;1074    }1075  assert(TIV && "No edge from pred to succ?");1076 1077  // Okay, we found the one constant that our value can be if we get into TI's1078  // BB.  Find out which successor will unconditionally be branched to.1079  BasicBlock *TheRealDest = nullptr;1080  for (const auto &[Value, Dest] : ThisCases)1081    if (Value == TIV) {1082      TheRealDest = Dest;1083      break;1084    }1085 1086  // If not handled by any explicit cases, it is handled by the default case.1087  if (!TheRealDest)1088    TheRealDest = ThisDef;1089 1090  SmallPtrSet<BasicBlock *, 2> RemovedSuccs;1091 1092  // Remove PHI node entries for dead edges.1093  BasicBlock *CheckEdge = TheRealDest;1094  for (BasicBlock *Succ : successors(TIBB))1095    if (Succ != CheckEdge) {1096      if (Succ != TheRealDest)1097        RemovedSuccs.insert(Succ);1098      Succ->removePredecessor(TIBB);1099    } else1100      CheckEdge = nullptr;1101 1102  // Insert the new branch.1103  Instruction *NI = Builder.CreateBr(TheRealDest);1104  (void)NI;1105 1106  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()1107                    << "Through successor TI: " << *TI << "Leaving: " << *NI1108                    << "\n");1109 1110  eraseTerminatorAndDCECond(TI);1111  if (DTU) {1112    SmallVector<DominatorTree::UpdateType, 2> Updates;1113    Updates.reserve(RemovedSuccs.size());1114    for (auto *RemovedSucc : RemovedSuccs)1115      Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});1116    DTU->applyUpdates(Updates);1117  }1118  return true;1119}1120 1121namespace {1122 1123/// This class implements a stable ordering of constant1124/// integers that does not depend on their address.  This is important for1125/// applications that sort ConstantInt's to ensure uniqueness.1126struct ConstantIntOrdering {1127  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {1128    return LHS->getValue().ult(RHS->getValue());1129  }1130};1131 1132} // end anonymous namespace1133 1134static int constantIntSortPredicate(ConstantInt *const *P1,1135                                    ConstantInt *const *P2) {1136  const ConstantInt *LHS = *P1;1137  const ConstantInt *RHS = *P2;1138  if (LHS == RHS)1139    return 0;1140  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;1141}1142 1143/// Get Weights of a given terminator, the default weight is at the front1144/// of the vector. If TI is a conditional eq, we need to swap the branch-weight1145/// metadata.1146static void getBranchWeights(Instruction *TI,1147                             SmallVectorImpl<uint64_t> &Weights) {1148  MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);1149  assert(MD && "Invalid branch-weight metadata");1150  extractFromBranchWeightMD64(MD, Weights);1151 1152  // If TI is a conditional eq, the default case is the false case,1153  // and the corresponding branch-weight data is at index 2. We swap the1154  // default weight to be the first entry.1155  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {1156    assert(Weights.size() == 2);1157    auto *ICI = dyn_cast<ICmpInst>(BI->getCondition());1158    if (!ICI)1159      return;1160 1161    if (ICI->getPredicate() == ICmpInst::ICMP_EQ)1162      std::swap(Weights.front(), Weights.back());1163  }1164}1165 1166static void cloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(1167    BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {1168  Instruction *PTI = PredBlock->getTerminator();1169 1170  // If we have bonus instructions, clone them into the predecessor block.1171  // Note that there may be multiple predecessor blocks, so we cannot move1172  // bonus instructions to a predecessor block.1173  for (Instruction &BonusInst : *BB) {1174    if (BonusInst.isTerminator())1175      continue;1176 1177    Instruction *NewBonusInst = BonusInst.clone();1178 1179    if (!NewBonusInst->getDebugLoc().isSameSourceLocation(PTI->getDebugLoc())) {1180      // Unless the instruction has the same !dbg location as the original1181      // branch, drop it. When we fold the bonus instructions we want to make1182      // sure we reset their debug locations in order to avoid stepping on1183      // dead code caused by folding dead branches.1184      NewBonusInst->setDebugLoc(DebugLoc::getDropped());1185    } else if (const DebugLoc &DL = NewBonusInst->getDebugLoc()) {1186      mapAtomInstance(DL, VMap);1187    }1188 1189    RemapInstruction(NewBonusInst, VMap,1190                     RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);1191 1192    // If we speculated an instruction, we need to drop any metadata that may1193    // result in undefined behavior, as the metadata might have been valid1194    // only given the branch precondition.1195    // Similarly strip attributes on call parameters that may cause UB in1196    // location the call is moved to.1197    NewBonusInst->dropUBImplyingAttrsAndMetadata();1198 1199    NewBonusInst->insertInto(PredBlock, PTI->getIterator());1200    auto Range = NewBonusInst->cloneDebugInfoFrom(&BonusInst);1201    RemapDbgRecordRange(NewBonusInst->getModule(), Range, VMap,1202                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);1203 1204    NewBonusInst->takeName(&BonusInst);1205    BonusInst.setName(NewBonusInst->getName() + ".old");1206    VMap[&BonusInst] = NewBonusInst;1207 1208    // Update (liveout) uses of bonus instructions,1209    // now that the bonus instruction has been cloned into predecessor.1210    // Note that we expect to be in a block-closed SSA form for this to work!1211    for (Use &U : make_early_inc_range(BonusInst.uses())) {1212      auto *UI = cast<Instruction>(U.getUser());1213      auto *PN = dyn_cast<PHINode>(UI);1214      if (!PN) {1215        assert(UI->getParent() == BB && BonusInst.comesBefore(UI) &&1216               "If the user is not a PHI node, then it should be in the same "1217               "block as, and come after, the original bonus instruction.");1218        continue; // Keep using the original bonus instruction.1219      }1220      // Is this the block-closed SSA form PHI node?1221      if (PN->getIncomingBlock(U) == BB)1222        continue; // Great, keep using the original bonus instruction.1223      // The only other alternative is an "use" when coming from1224      // the predecessor block - here we should refer to the cloned bonus instr.1225      assert(PN->getIncomingBlock(U) == PredBlock &&1226             "Not in block-closed SSA form?");1227      U.set(NewBonusInst);1228    }1229  }1230 1231  // Key Instructions: We may have propagated atom info into the pred. If the1232  // pred's terminator already has atom info do nothing as merging would drop1233  // one atom group anyway. If it doesn't, propagte the remapped atom group1234  // from BB's terminator.1235  if (auto &PredDL = PTI->getDebugLoc()) {1236    auto &DL = BB->getTerminator()->getDebugLoc();1237    if (!PredDL->getAtomGroup() && DL && DL->getAtomGroup() &&1238        PredDL.isSameSourceLocation(DL)) {1239      PTI->setDebugLoc(DL);1240      RemapSourceAtom(PTI, VMap);1241    }1242  }1243}1244 1245bool SimplifyCFGOpt::performValueComparisonIntoPredecessorFolding(1246    Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {1247  BasicBlock *BB = TI->getParent();1248  BasicBlock *Pred = PTI->getParent();1249 1250  SmallVector<DominatorTree::UpdateType, 32> Updates;1251 1252  // Figure out which 'cases' to copy from SI to PSI.1253  std::vector<ValueEqualityComparisonCase> BBCases;1254  BasicBlock *BBDefault = getValueEqualityComparisonCases(TI, BBCases);1255 1256  std::vector<ValueEqualityComparisonCase> PredCases;1257  BasicBlock *PredDefault = getValueEqualityComparisonCases(PTI, PredCases);1258 1259  // Based on whether the default edge from PTI goes to BB or not, fill in1260  // PredCases and PredDefault with the new switch cases we would like to1261  // build.1262  SmallMapVector<BasicBlock *, int, 8> NewSuccessors;1263 1264  // Update the branch weight metadata along the way1265  SmallVector<uint64_t, 8> Weights;1266  bool PredHasWeights = hasBranchWeightMD(*PTI);1267  bool SuccHasWeights = hasBranchWeightMD(*TI);1268 1269  if (PredHasWeights) {1270    getBranchWeights(PTI, Weights);1271    // branch-weight metadata is inconsistent here.1272    if (Weights.size() != 1 + PredCases.size())1273      PredHasWeights = SuccHasWeights = false;1274  } else if (SuccHasWeights)1275    // If there are no predecessor weights but there are successor weights,1276    // populate Weights with 1, which will later be scaled to the sum of1277    // successor's weights1278    Weights.assign(1 + PredCases.size(), 1);1279 1280  SmallVector<uint64_t, 8> SuccWeights;1281  if (SuccHasWeights) {1282    getBranchWeights(TI, SuccWeights);1283    // branch-weight metadata is inconsistent here.1284    if (SuccWeights.size() != 1 + BBCases.size())1285      PredHasWeights = SuccHasWeights = false;1286  } else if (PredHasWeights)1287    SuccWeights.assign(1 + BBCases.size(), 1);1288 1289  if (PredDefault == BB) {1290    // If this is the default destination from PTI, only the edges in TI1291    // that don't occur in PTI, or that branch to BB will be activated.1292    std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;1293    for (unsigned i = 0, e = PredCases.size(); i != e; ++i)1294      if (PredCases[i].Dest != BB)1295        PTIHandled.insert(PredCases[i].Value);1296      else {1297        // The default destination is BB, we don't need explicit targets.1298        std::swap(PredCases[i], PredCases.back());1299 1300        if (PredHasWeights || SuccHasWeights) {1301          // Increase weight for the default case.1302          Weights[0] += Weights[i + 1];1303          std::swap(Weights[i + 1], Weights.back());1304          Weights.pop_back();1305        }1306 1307        PredCases.pop_back();1308        --i;1309        --e;1310      }1311 1312    // Reconstruct the new switch statement we will be building.1313    if (PredDefault != BBDefault) {1314      PredDefault->removePredecessor(Pred);1315      if (DTU && PredDefault != BB)1316        Updates.push_back({DominatorTree::Delete, Pred, PredDefault});1317      PredDefault = BBDefault;1318      ++NewSuccessors[BBDefault];1319    }1320 1321    unsigned CasesFromPred = Weights.size();1322    uint64_t ValidTotalSuccWeight = 0;1323    for (unsigned i = 0, e = BBCases.size(); i != e; ++i)1324      if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {1325        PredCases.push_back(BBCases[i]);1326        ++NewSuccessors[BBCases[i].Dest];1327        if (SuccHasWeights || PredHasWeights) {1328          // The default weight is at index 0, so weight for the ith case1329          // should be at index i+1. Scale the cases from successor by1330          // PredDefaultWeight (Weights[0]).1331          Weights.push_back(Weights[0] * SuccWeights[i + 1]);1332          ValidTotalSuccWeight += SuccWeights[i + 1];1333        }1334      }1335 1336    if (SuccHasWeights || PredHasWeights) {1337      ValidTotalSuccWeight += SuccWeights[0];1338      // Scale the cases from predecessor by ValidTotalSuccWeight.1339      for (unsigned i = 1; i < CasesFromPred; ++i)1340        Weights[i] *= ValidTotalSuccWeight;1341      // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).1342      Weights[0] *= SuccWeights[0];1343    }1344  } else {1345    // If this is not the default destination from PSI, only the edges1346    // in SI that occur in PSI with a destination of BB will be1347    // activated.1348    std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;1349    std::map<ConstantInt *, uint64_t> WeightsForHandled;1350    for (unsigned i = 0, e = PredCases.size(); i != e; ++i)1351      if (PredCases[i].Dest == BB) {1352        PTIHandled.insert(PredCases[i].Value);1353 1354        if (PredHasWeights || SuccHasWeights) {1355          WeightsForHandled[PredCases[i].Value] = Weights[i + 1];1356          std::swap(Weights[i + 1], Weights.back());1357          Weights.pop_back();1358        }1359 1360        std::swap(PredCases[i], PredCases.back());1361        PredCases.pop_back();1362        --i;1363        --e;1364      }1365 1366    // Okay, now we know which constants were sent to BB from the1367    // predecessor.  Figure out where they will all go now.1368    for (const ValueEqualityComparisonCase &Case : BBCases)1369      if (PTIHandled.count(Case.Value)) {1370        // If this is one we are capable of getting...1371        if (PredHasWeights || SuccHasWeights)1372          Weights.push_back(WeightsForHandled[Case.Value]);1373        PredCases.push_back(Case);1374        ++NewSuccessors[Case.Dest];1375        PTIHandled.erase(Case.Value); // This constant is taken care of1376      }1377 1378    // If there are any constants vectored to BB that TI doesn't handle,1379    // they must go to the default destination of TI.1380    for (ConstantInt *I : PTIHandled) {1381      if (PredHasWeights || SuccHasWeights)1382        Weights.push_back(WeightsForHandled[I]);1383      PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));1384      ++NewSuccessors[BBDefault];1385    }1386  }1387 1388  // Okay, at this point, we know which new successor Pred will get.  Make1389  // sure we update the number of entries in the PHI nodes for these1390  // successors.1391  SmallPtrSet<BasicBlock *, 2> SuccsOfPred;1392  if (DTU) {1393    SuccsOfPred = {llvm::from_range, successors(Pred)};1394    Updates.reserve(Updates.size() + NewSuccessors.size());1395  }1396  for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :1397       NewSuccessors) {1398    for (auto I : seq(NewSuccessor.second)) {1399      (void)I;1400      addPredecessorToBlock(NewSuccessor.first, Pred, BB);1401    }1402    if (DTU && !SuccsOfPred.contains(NewSuccessor.first))1403      Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});1404  }1405 1406  Builder.SetInsertPoint(PTI);1407  // Convert pointer to int before we switch.1408  if (CV->getType()->isPointerTy()) {1409    assert(!DL.hasUnstableRepresentation(CV->getType()) &&1410           "Should not end up here with unstable pointers");1411    CV =1412        Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");1413  }1414 1415  // Now that the successors are updated, create the new Switch instruction.1416  SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());1417  NewSI->setDebugLoc(PTI->getDebugLoc());1418  for (ValueEqualityComparisonCase &V : PredCases)1419    NewSI->addCase(V.Value, V.Dest);1420 1421  if (PredHasWeights || SuccHasWeights)1422    setFittedBranchWeights(*NewSI, Weights, /*IsExpected=*/false,1423                           /*ElideAllZero=*/true);1424 1425  eraseTerminatorAndDCECond(PTI);1426 1427  // Okay, last check.  If BB is still a successor of PSI, then we must1428  // have an infinite loop case.  If so, add an infinitely looping block1429  // to handle the case to preserve the behavior of the code.1430  BasicBlock *InfLoopBlock = nullptr;1431  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)1432    if (NewSI->getSuccessor(i) == BB) {1433      if (!InfLoopBlock) {1434        // Insert it at the end of the function, because it's either code,1435        // or it won't matter if it's hot. :)1436        InfLoopBlock =1437            BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());1438        BranchInst::Create(InfLoopBlock, InfLoopBlock);1439        if (DTU)1440          Updates.push_back(1441              {DominatorTree::Insert, InfLoopBlock, InfLoopBlock});1442      }1443      NewSI->setSuccessor(i, InfLoopBlock);1444    }1445 1446  if (DTU) {1447    if (InfLoopBlock)1448      Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});1449 1450    Updates.push_back({DominatorTree::Delete, Pred, BB});1451 1452    DTU->applyUpdates(Updates);1453  }1454 1455  ++NumFoldValueComparisonIntoPredecessors;1456  return true;1457}1458 1459/// The specified terminator is a value equality comparison instruction1460/// (either a switch or a branch on "X == c").1461/// See if any of the predecessors of the terminator block are value comparisons1462/// on the same value.  If so, and if safe to do so, fold them together.1463bool SimplifyCFGOpt::foldValueComparisonIntoPredecessors(Instruction *TI,1464                                                         IRBuilder<> &Builder) {1465  BasicBlock *BB = TI->getParent();1466  Value *CV = isValueEqualityComparison(TI); // CondVal1467  assert(CV && "Not a comparison?");1468 1469  bool Changed = false;1470 1471  SmallSetVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));1472  while (!Preds.empty()) {1473    BasicBlock *Pred = Preds.pop_back_val();1474    Instruction *PTI = Pred->getTerminator();1475 1476    // Don't try to fold into itself.1477    if (Pred == BB)1478      continue;1479 1480    // See if the predecessor is a comparison with the same value.1481    Value *PCV = isValueEqualityComparison(PTI); // PredCondVal1482    if (PCV != CV)1483      continue;1484 1485    SmallSetVector<BasicBlock *, 4> FailBlocks;1486    if (!safeToMergeTerminators(TI, PTI, &FailBlocks)) {1487      for (auto *Succ : FailBlocks) {1488        if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))1489          return false;1490      }1491    }1492 1493    performValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);1494    Changed = true;1495  }1496  return Changed;1497}1498 1499// If we would need to insert a select that uses the value of this invoke1500// (comments in hoistSuccIdenticalTerminatorToSwitchOrIf explain why we would1501// need to do this), we can't hoist the invoke, as there is nowhere to put the1502// select in this case.1503static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,1504                                Instruction *I1, Instruction *I2) {1505  for (BasicBlock *Succ : successors(BB1)) {1506    for (const PHINode &PN : Succ->phis()) {1507      Value *BB1V = PN.getIncomingValueForBlock(BB1);1508      Value *BB2V = PN.getIncomingValueForBlock(BB2);1509      if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {1510        return false;1511      }1512    }1513  }1514  return true;1515}1516 1517// Get interesting characteristics of instructions that1518// `hoistCommonCodeFromSuccessors` didn't hoist. They restrict what kind of1519// instructions can be reordered across.1520enum SkipFlags {1521  SkipReadMem = 1,1522  SkipSideEffect = 2,1523  SkipImplicitControlFlow = 41524};1525 1526static unsigned skippedInstrFlags(Instruction *I) {1527  unsigned Flags = 0;1528  if (I->mayReadFromMemory())1529    Flags |= SkipReadMem;1530  // We can't arbitrarily move around allocas, e.g. moving allocas (especially1531  // inalloca) across stacksave/stackrestore boundaries.1532  if (I->mayHaveSideEffects() || isa<AllocaInst>(I))1533    Flags |= SkipSideEffect;1534  if (!isGuaranteedToTransferExecutionToSuccessor(I))1535    Flags |= SkipImplicitControlFlow;1536  return Flags;1537}1538 1539// Returns true if it is safe to reorder an instruction across preceding1540// instructions in a basic block.1541static bool isSafeToHoistInstr(Instruction *I, unsigned Flags) {1542  // Don't reorder a store over a load.1543  if ((Flags & SkipReadMem) && I->mayWriteToMemory())1544    return false;1545 1546  // If we have seen an instruction with side effects, it's unsafe to reorder an1547  // instruction which reads memory or itself has side effects.1548  if ((Flags & SkipSideEffect) &&1549      (I->mayReadFromMemory() || I->mayHaveSideEffects() || isa<AllocaInst>(I)))1550    return false;1551 1552  // Reordering across an instruction which does not necessarily transfer1553  // control to the next instruction is speculation.1554  if ((Flags & SkipImplicitControlFlow) && !isSafeToSpeculativelyExecute(I))1555    return false;1556 1557  // Hoisting of llvm.deoptimize is only legal together with the next return1558  // instruction, which this pass is not always able to do.1559  if (auto *CB = dyn_cast<CallBase>(I))1560    if (CB->getIntrinsicID() == Intrinsic::experimental_deoptimize)1561      return false;1562 1563  // It's also unsafe/illegal to hoist an instruction above its instruction1564  // operands1565  BasicBlock *BB = I->getParent();1566  for (Value *Op : I->operands()) {1567    if (auto *J = dyn_cast<Instruction>(Op))1568      if (J->getParent() == BB)1569        return false;1570  }1571 1572  return true;1573}1574 1575static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);1576 1577/// Helper function for hoistCommonCodeFromSuccessors. Return true if identical1578/// instructions \p I1 and \p I2 can and should be hoisted.1579static bool shouldHoistCommonInstructions(Instruction *I1, Instruction *I2,1580                                          const TargetTransformInfo &TTI) {1581  // If we're going to hoist a call, make sure that the two instructions1582  // we're commoning/hoisting are both marked with musttail, or neither of1583  // them is marked as such. Otherwise, we might end up in a situation where1584  // we hoist from a block where the terminator is a `ret` to a block where1585  // the terminator is a `br`, and `musttail` calls expect to be followed by1586  // a return.1587  auto *C1 = dyn_cast<CallInst>(I1);1588  auto *C2 = dyn_cast<CallInst>(I2);1589  if (C1 && C2)1590    if (C1->isMustTailCall() != C2->isMustTailCall())1591      return false;1592 1593  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))1594    return false;1595 1596  // If any of the two call sites has nomerge or convergent attribute, stop1597  // hoisting.1598  if (const auto *CB1 = dyn_cast<CallBase>(I1))1599    if (CB1->cannotMerge() || CB1->isConvergent())1600      return false;1601  if (const auto *CB2 = dyn_cast<CallBase>(I2))1602    if (CB2->cannotMerge() || CB2->isConvergent())1603      return false;1604 1605  return true;1606}1607 1608/// Hoists DbgVariableRecords from \p I1 and \p OtherInstrs that are identical1609/// in lock-step to \p TI. This matches how dbg.* intrinsics are hoisting in1610/// hoistCommonCodeFromSuccessors. e.g. The input:1611///    I1                DVRs: { x, z },1612///    OtherInsts: { I2  DVRs: { x, y, z } }1613/// would result in hoisting only DbgVariableRecord x.1614static void hoistLockstepIdenticalDbgVariableRecords(1615    Instruction *TI, Instruction *I1,1616    SmallVectorImpl<Instruction *> &OtherInsts) {1617  if (!I1->hasDbgRecords())1618    return;1619  using CurrentAndEndIt =1620      std::pair<DbgRecord::self_iterator, DbgRecord::self_iterator>;1621  // Vector of {Current, End} iterators.1622  SmallVector<CurrentAndEndIt> Itrs;1623  Itrs.reserve(OtherInsts.size() + 1);1624  // Helper lambdas for lock-step checks:1625  // Return true if this Current == End.1626  auto atEnd = [](const CurrentAndEndIt &Pair) {1627    return Pair.first == Pair.second;1628  };1629  // Return true if all Current are identical.1630  auto allIdentical = [](const SmallVector<CurrentAndEndIt> &Itrs) {1631    return all_of(make_first_range(ArrayRef(Itrs).drop_front()),1632                  [&](DbgRecord::self_iterator I) {1633                    return Itrs[0].first->isIdenticalToWhenDefined(*I);1634                  });1635  };1636 1637  // Collect the iterators.1638  Itrs.push_back(1639      {I1->getDbgRecordRange().begin(), I1->getDbgRecordRange().end()});1640  for (Instruction *Other : OtherInsts) {1641    if (!Other->hasDbgRecords())1642      return;1643    Itrs.push_back(1644        {Other->getDbgRecordRange().begin(), Other->getDbgRecordRange().end()});1645  }1646 1647  // Iterate in lock-step until any of the DbgRecord lists are exausted. If1648  // the lock-step DbgRecord are identical, hoist all of them to TI.1649  // This replicates the dbg.* intrinsic behaviour in1650  // hoistCommonCodeFromSuccessors.1651  while (none_of(Itrs, atEnd)) {1652    bool HoistDVRs = allIdentical(Itrs);1653    for (CurrentAndEndIt &Pair : Itrs) {1654      // Increment Current iterator now as we may be about to move the1655      // DbgRecord.1656      DbgRecord &DR = *Pair.first++;1657      if (HoistDVRs) {1658        DR.removeFromParent();1659        TI->getParent()->insertDbgRecordBefore(&DR, TI->getIterator());1660      }1661    }1662  }1663}1664 1665static bool areIdenticalUpToCommutativity(const Instruction *I1,1666                                          const Instruction *I2) {1667  if (I1->isIdenticalToWhenDefined(I2, /*IntersectAttrs=*/true))1668    return true;1669 1670  if (auto *Cmp1 = dyn_cast<CmpInst>(I1))1671    if (auto *Cmp2 = dyn_cast<CmpInst>(I2))1672      return Cmp1->getPredicate() == Cmp2->getSwappedPredicate() &&1673             Cmp1->getOperand(0) == Cmp2->getOperand(1) &&1674             Cmp1->getOperand(1) == Cmp2->getOperand(0);1675 1676  if (I1->isCommutative() && I1->isSameOperationAs(I2)) {1677    return I1->getOperand(0) == I2->getOperand(1) &&1678           I1->getOperand(1) == I2->getOperand(0) &&1679           equal(drop_begin(I1->operands(), 2), drop_begin(I2->operands(), 2));1680  }1681 1682  return false;1683}1684 1685/// If the target supports conditional faulting,1686/// we look for the following pattern:1687/// \code1688///   BB:1689///     ...1690///     %cond = icmp ult %x, %y1691///     br i1 %cond, label %TrueBB, label %FalseBB1692///   FalseBB:1693///     store i32 1, ptr %q, align 41694///     ...1695///   TrueBB:1696///     %maskedloadstore = load i32, ptr %b, align 41697///     store i32 %maskedloadstore, ptr %p, align 41698///     ...1699/// \endcode1700///1701/// and transform it into:1702///1703/// \code1704///   BB:1705///     ...1706///     %cond = icmp ult %x, %y1707///     %maskedloadstore = cload i32, ptr %b, %cond1708///     cstore i32 %maskedloadstore, ptr %p, %cond1709///     cstore i32 1, ptr %q, ~%cond1710///     br i1 %cond, label %TrueBB, label %FalseBB1711///   FalseBB:1712///     ...1713///   TrueBB:1714///     ...1715/// \endcode1716///1717/// where cload/cstore are represented by llvm.masked.load/store intrinsics,1718/// e.g.1719///1720/// \code1721///   %vcond = bitcast i1 %cond to <1 x i1>1722///   %v0 = call <1 x i32> @llvm.masked.load.v1i32.p01723///                         (ptr %b, i32 4, <1 x i1> %vcond, <1 x i32> poison)1724///   %maskedloadstore = bitcast <1 x i32> %v0 to i321725///   call void @llvm.masked.store.v1i32.p01726///                          (<1 x i32> %v0, ptr %p, i32 4, <1 x i1> %vcond)1727///   %cond.not = xor i1 %cond, true1728///   %vcond.not = bitcast i1 %cond.not to <1 x i>1729///   call void @llvm.masked.store.v1i32.p01730///              (<1 x i32> <i32 1>, ptr %q, i32 4, <1x i1> %vcond.not)1731/// \endcode1732///1733/// So we need to turn hoisted load/store into cload/cstore.1734///1735/// \param BI The branch instruction.1736/// \param SpeculatedConditionalLoadsStores The load/store instructions that1737///                                         will be speculated.1738/// \param Invert indicates if speculates FalseBB. Only used in triangle CFG.1739static void hoistConditionalLoadsStores(1740    BranchInst *BI,1741    SmallVectorImpl<Instruction *> &SpeculatedConditionalLoadsStores,1742    std::optional<bool> Invert, Instruction *Sel) {1743  auto &Context = BI->getParent()->getContext();1744  auto *VCondTy = FixedVectorType::get(Type::getInt1Ty(Context), 1);1745  auto *Cond = BI->getOperand(0);1746  // Construct the condition if needed.1747  BasicBlock *BB = BI->getParent();1748  Value *Mask = nullptr;1749  Value *MaskFalse = nullptr;1750  Value *MaskTrue = nullptr;1751  if (Invert.has_value()) {1752    IRBuilder<> Builder(Sel ? Sel : SpeculatedConditionalLoadsStores.back());1753    Mask = Builder.CreateBitCast(1754        *Invert ? Builder.CreateXor(Cond, ConstantInt::getTrue(Context)) : Cond,1755        VCondTy);1756  } else {1757    IRBuilder<> Builder(BI);1758    MaskFalse = Builder.CreateBitCast(1759        Builder.CreateXor(Cond, ConstantInt::getTrue(Context)), VCondTy);1760    MaskTrue = Builder.CreateBitCast(Cond, VCondTy);1761  }1762  auto PeekThroughBitcasts = [](Value *V) {1763    while (auto *BitCast = dyn_cast<BitCastInst>(V))1764      V = BitCast->getOperand(0);1765    return V;1766  };1767  for (auto *I : SpeculatedConditionalLoadsStores) {1768    IRBuilder<> Builder(Invert.has_value() ? I : BI);1769    if (!Invert.has_value())1770      Mask = I->getParent() == BI->getSuccessor(0) ? MaskTrue : MaskFalse;1771    // We currently assume conditional faulting load/store is supported for1772    // scalar types only when creating new instructions. This can be easily1773    // extended for vector types in the future.1774    assert(!getLoadStoreType(I)->isVectorTy() && "not implemented");1775    auto *Op0 = I->getOperand(0);1776    CallInst *MaskedLoadStore = nullptr;1777    if (auto *LI = dyn_cast<LoadInst>(I)) {1778      // Handle Load.1779      auto *Ty = I->getType();1780      PHINode *PN = nullptr;1781      Value *PassThru = nullptr;1782      if (Invert.has_value())1783        for (User *U : I->users()) {1784          if ((PN = dyn_cast<PHINode>(U))) {1785            PassThru = Builder.CreateBitCast(1786                PeekThroughBitcasts(PN->getIncomingValueForBlock(BB)),1787                FixedVectorType::get(Ty, 1));1788          } else if (auto *Ins = cast<Instruction>(U);1789                     Sel && Ins->getParent() == BB) {1790            // This happens when store or/and a speculative instruction between1791            // load and store were hoisted to the BB. Make sure the masked load1792            // inserted before its use.1793            // We assume there's one of such use.1794            Builder.SetInsertPoint(Ins);1795          }1796        }1797      MaskedLoadStore = Builder.CreateMaskedLoad(1798          FixedVectorType::get(Ty, 1), Op0, LI->getAlign(), Mask, PassThru);1799      Value *NewLoadStore = Builder.CreateBitCast(MaskedLoadStore, Ty);1800      if (PN)1801        PN->setIncomingValue(PN->getBasicBlockIndex(BB), NewLoadStore);1802      I->replaceAllUsesWith(NewLoadStore);1803    } else {1804      // Handle Store.1805      auto *StoredVal = Builder.CreateBitCast(1806          PeekThroughBitcasts(Op0), FixedVectorType::get(Op0->getType(), 1));1807      MaskedLoadStore = Builder.CreateMaskedStore(1808          StoredVal, I->getOperand(1), cast<StoreInst>(I)->getAlign(), Mask);1809    }1810    // For non-debug metadata, only !annotation, !range, !nonnull and !align are1811    // kept when hoisting (see Instruction::dropUBImplyingAttrsAndMetadata).1812    //1813    // !nonnull, !align : Not support pointer type, no need to keep.1814    // !range: Load type is changed from scalar to vector, but the metadata on1815    //         vector specifies a per-element range, so the semantics stay the1816    //         same. Keep it.1817    // !annotation: Not impact semantics. Keep it.1818    if (const MDNode *Ranges = I->getMetadata(LLVMContext::MD_range))1819      MaskedLoadStore->addRangeRetAttr(getConstantRangeFromMetadata(*Ranges));1820    I->dropUBImplyingAttrsAndUnknownMetadata({LLVMContext::MD_annotation});1821    // FIXME: DIAssignID is not supported for masked store yet.1822    // (Verifier::visitDIAssignIDMetadata)1823    at::deleteAssignmentMarkers(I);1824    I->eraseMetadataIf([](unsigned MDKind, MDNode *Node) {1825      return Node->getMetadataID() == Metadata::DIAssignIDKind;1826    });1827    MaskedLoadStore->copyMetadata(*I);1828    I->eraseFromParent();1829  }1830}1831 1832static bool isSafeCheapLoadStore(const Instruction *I,1833                                 const TargetTransformInfo &TTI) {1834  // Not handle volatile or atomic.1835  bool IsStore = false;1836  if (auto *L = dyn_cast<LoadInst>(I)) {1837    if (!L->isSimple() || !HoistLoadsWithCondFaulting)1838      return false;1839  } else if (auto *S = dyn_cast<StoreInst>(I)) {1840    if (!S->isSimple() || !HoistStoresWithCondFaulting)1841      return false;1842    IsStore = true;1843  } else1844    return false;1845 1846  // llvm.masked.load/store use i32 for alignment while load/store use i64.1847  // That's why we have the alignment limitation.1848  // FIXME: Update the prototype of the intrinsics?1849  return TTI.hasConditionalLoadStoreForType(getLoadStoreType(I), IsStore) &&1850         getLoadStoreAlignment(I) < Value::MaximumAlignment;1851}1852 1853/// Hoist any common code in the successor blocks up into the block. This1854/// function guarantees that BB dominates all successors. If AllInstsEqOnly is1855/// given, only perform hoisting in case all successors blocks contain matching1856/// instructions only. In that case, all instructions can be hoisted and the1857/// original branch will be replaced and selects for PHIs are added.1858bool SimplifyCFGOpt::hoistCommonCodeFromSuccessors(Instruction *TI,1859                                                   bool AllInstsEqOnly) {1860  // This does very trivial matching, with limited scanning, to find identical1861  // instructions in the two blocks. In particular, we don't want to get into1862  // O(N1*N2*...) situations here where Ni are the sizes of these successors. As1863  // such, we currently just scan for obviously identical instructions in an1864  // identical order, possibly separated by the same number of non-identical1865  // instructions.1866  BasicBlock *BB = TI->getParent();1867  unsigned int SuccSize = succ_size(BB);1868  if (SuccSize < 2)1869    return false;1870 1871  // If either of the blocks has it's address taken, then we can't do this fold,1872  // because the code we'd hoist would no longer run when we jump into the block1873  // by it's address.1874  for (auto *Succ : successors(BB)) {1875    if (Succ->hasAddressTaken())1876      return false;1877    if (Succ->getSinglePredecessor())1878      continue;1879    // If Succ has >1 predecessors, continue to check if the Succ contains only1880    // one `unreachable` inst. Since executing `unreachable` inst is an UB, we1881    // can relax the condition based on the assumptiom that the program would1882    // never enter Succ and trigger such an UB.1883    if (isa<UnreachableInst>(*Succ->begin()))1884      continue;1885    return false;1886  }1887  // The second of pair is a SkipFlags bitmask.1888  using SuccIterPair = std::pair<BasicBlock::iterator, unsigned>;1889  SmallVector<SuccIterPair, 8> SuccIterPairs;1890  for (auto *Succ : successors(BB)) {1891    BasicBlock::iterator SuccItr = Succ->begin();1892    if (isa<PHINode>(*SuccItr))1893      return false;1894    SuccIterPairs.push_back(SuccIterPair(SuccItr, 0));1895  }1896 1897  if (AllInstsEqOnly) {1898    // Check if all instructions in the successor blocks match. This allows1899    // hoisting all instructions and removing the blocks we are hoisting from,1900    // so does not add any new instructions.1901    SmallVector<BasicBlock *> Succs = to_vector(successors(BB));1902    // Check if sizes and terminators of all successors match.1903    bool AllSame = none_of(Succs, [&Succs](BasicBlock *Succ) {1904      Instruction *Term0 = Succs[0]->getTerminator();1905      Instruction *Term = Succ->getTerminator();1906      return !Term->isSameOperationAs(Term0) ||1907             !equal(Term->operands(), Term0->operands()) ||1908             Succs[0]->size() != Succ->size();1909    });1910    if (!AllSame)1911      return false;1912    if (AllSame) {1913      LockstepReverseIterator<true> LRI(Succs);1914      while (LRI.isValid()) {1915        Instruction *I0 = (*LRI)[0];1916        if (any_of(*LRI, [I0](Instruction *I) {1917              return !areIdenticalUpToCommutativity(I0, I);1918            })) {1919          return false;1920        }1921        --LRI;1922      }1923    }1924    // Now we know that all instructions in all successors can be hoisted. Let1925    // the loop below handle the hoisting.1926  }1927 1928  // Count how many instructions were not hoisted so far. There's a limit on how1929  // many instructions we skip, serving as a compilation time control as well as1930  // preventing excessive increase of life ranges.1931  unsigned NumSkipped = 0;1932  // If we find an unreachable instruction at the beginning of a basic block, we1933  // can still hoist instructions from the rest of the basic blocks.1934  if (SuccIterPairs.size() > 2) {1935    erase_if(SuccIterPairs,1936             [](const auto &Pair) { return isa<UnreachableInst>(Pair.first); });1937    if (SuccIterPairs.size() < 2)1938      return false;1939  }1940 1941  bool Changed = false;1942 1943  for (;;) {1944    auto *SuccIterPairBegin = SuccIterPairs.begin();1945    auto &BB1ItrPair = *SuccIterPairBegin++;1946    auto OtherSuccIterPairRange =1947        iterator_range(SuccIterPairBegin, SuccIterPairs.end());1948    auto OtherSuccIterRange = make_first_range(OtherSuccIterPairRange);1949 1950    Instruction *I1 = &*BB1ItrPair.first;1951 1952    bool AllInstsAreIdentical = true;1953    bool HasTerminator = I1->isTerminator();1954    for (auto &SuccIter : OtherSuccIterRange) {1955      Instruction *I2 = &*SuccIter;1956      HasTerminator |= I2->isTerminator();1957      if (AllInstsAreIdentical && (!areIdenticalUpToCommutativity(I1, I2) ||1958                                   MMRAMetadata(*I1) != MMRAMetadata(*I2)))1959        AllInstsAreIdentical = false;1960    }1961 1962    SmallVector<Instruction *, 8> OtherInsts;1963    for (auto &SuccIter : OtherSuccIterRange)1964      OtherInsts.push_back(&*SuccIter);1965 1966    // If we are hoisting the terminator instruction, don't move one (making a1967    // broken BB), instead clone it, and remove BI.1968    if (HasTerminator) {1969      // Even if BB, which contains only one unreachable instruction, is ignored1970      // at the beginning of the loop, we can hoist the terminator instruction.1971      // If any instructions remain in the block, we cannot hoist terminators.1972      if (NumSkipped || !AllInstsAreIdentical) {1973        hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts);1974        return Changed;1975      }1976 1977      return hoistSuccIdenticalTerminatorToSwitchOrIf(TI, I1, OtherInsts) ||1978             Changed;1979    }1980 1981    if (AllInstsAreIdentical) {1982      unsigned SkipFlagsBB1 = BB1ItrPair.second;1983      AllInstsAreIdentical =1984          isSafeToHoistInstr(I1, SkipFlagsBB1) &&1985          all_of(OtherSuccIterPairRange, [=](const auto &Pair) {1986            Instruction *I2 = &*Pair.first;1987            unsigned SkipFlagsBB2 = Pair.second;1988            // Even if the instructions are identical, it may not1989            // be safe to hoist them if we have skipped over1990            // instructions with side effects or their operands1991            // weren't hoisted.1992            return isSafeToHoistInstr(I2, SkipFlagsBB2) &&1993                   shouldHoistCommonInstructions(I1, I2, TTI);1994          });1995    }1996 1997    if (AllInstsAreIdentical) {1998      BB1ItrPair.first++;1999      // For a normal instruction, we just move one to right before the2000      // branch, then replace all uses of the other with the first.  Finally,2001      // we remove the now redundant second instruction.2002      hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts);2003      // We've just hoisted DbgVariableRecords; move I1 after them (before TI)2004      // and leave any that were not hoisted behind (by calling moveBefore2005      // rather than moveBeforePreserving).2006      I1->moveBefore(TI->getIterator());2007      for (auto &SuccIter : OtherSuccIterRange) {2008        Instruction *I2 = &*SuccIter++;2009        assert(I2 != I1);2010        if (!I2->use_empty())2011          I2->replaceAllUsesWith(I1);2012        I1->andIRFlags(I2);2013        if (auto *CB = dyn_cast<CallBase>(I1)) {2014          bool Success = CB->tryIntersectAttributes(cast<CallBase>(I2));2015          assert(Success && "We should not be trying to hoist callbases "2016                            "with non-intersectable attributes");2017          // For NDEBUG Compile.2018          (void)Success;2019        }2020 2021        combineMetadataForCSE(I1, I2, true);2022        // I1 and I2 are being combined into a single instruction.  Its debug2023        // location is the merged locations of the original instructions.2024        I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());2025        I2->eraseFromParent();2026      }2027      if (!Changed)2028        NumHoistCommonCode += SuccIterPairs.size();2029      Changed = true;2030      NumHoistCommonInstrs += SuccIterPairs.size();2031    } else {2032      if (NumSkipped >= HoistCommonSkipLimit) {2033        hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts);2034        return Changed;2035      }2036      // We are about to skip over a pair of non-identical instructions. Record2037      // if any have characteristics that would prevent reordering instructions2038      // across them.2039      for (auto &SuccIterPair : SuccIterPairs) {2040        Instruction *I = &*SuccIterPair.first++;2041        SuccIterPair.second |= skippedInstrFlags(I);2042      }2043      ++NumSkipped;2044    }2045  }2046}2047 2048bool SimplifyCFGOpt::hoistSuccIdenticalTerminatorToSwitchOrIf(2049    Instruction *TI, Instruction *I1,2050    SmallVectorImpl<Instruction *> &OtherSuccTIs) {2051 2052  auto *BI = dyn_cast<BranchInst>(TI);2053 2054  bool Changed = false;2055  BasicBlock *TIParent = TI->getParent();2056  BasicBlock *BB1 = I1->getParent();2057 2058  // Use only for an if statement.2059  auto *I2 = *OtherSuccTIs.begin();2060  auto *BB2 = I2->getParent();2061  if (BI) {2062    assert(OtherSuccTIs.size() == 1);2063    assert(BI->getSuccessor(0) == I1->getParent());2064    assert(BI->getSuccessor(1) == I2->getParent());2065  }2066 2067  // In the case of an if statement, we try to hoist an invoke.2068  // FIXME: Can we define a safety predicate for CallBr?2069  // FIXME: Test case llvm/test/Transforms/SimplifyCFG/2009-06-15-InvokeCrash.ll2070  // removed in 4c923b3b3fd0ac1edebf0603265ca3ba51724937 commit?2071  if (isa<InvokeInst>(I1) && (!BI || !isSafeToHoistInvoke(BB1, BB2, I1, I2)))2072    return false;2073 2074  // TODO: callbr hoisting currently disabled pending further study.2075  if (isa<CallBrInst>(I1))2076    return false;2077 2078  for (BasicBlock *Succ : successors(BB1)) {2079    for (PHINode &PN : Succ->phis()) {2080      Value *BB1V = PN.getIncomingValueForBlock(BB1);2081      for (Instruction *OtherSuccTI : OtherSuccTIs) {2082        Value *BB2V = PN.getIncomingValueForBlock(OtherSuccTI->getParent());2083        if (BB1V == BB2V)2084          continue;2085 2086        // In the case of an if statement, check for2087        // passingValueIsAlwaysUndefined here because we would rather eliminate2088        // undefined control flow then converting it to a select.2089        if (!BI || passingValueIsAlwaysUndefined(BB1V, &PN) ||2090            passingValueIsAlwaysUndefined(BB2V, &PN))2091          return false;2092      }2093    }2094  }2095 2096  // Hoist DbgVariableRecords attached to the terminator to match dbg.*2097  // intrinsic hoisting behaviour in hoistCommonCodeFromSuccessors.2098  hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherSuccTIs);2099  // Clone the terminator and hoist it into the pred, without any debug info.2100  Instruction *NT = I1->clone();2101  NT->insertInto(TIParent, TI->getIterator());2102  if (!NT->getType()->isVoidTy()) {2103    I1->replaceAllUsesWith(NT);2104    for (Instruction *OtherSuccTI : OtherSuccTIs)2105      OtherSuccTI->replaceAllUsesWith(NT);2106    NT->takeName(I1);2107  }2108  Changed = true;2109  NumHoistCommonInstrs += OtherSuccTIs.size() + 1;2110 2111  // Ensure terminator gets a debug location, even an unknown one, in case2112  // it involves inlinable calls.2113  SmallVector<DebugLoc, 4> Locs;2114  Locs.push_back(I1->getDebugLoc());2115  for (auto *OtherSuccTI : OtherSuccTIs)2116    Locs.push_back(OtherSuccTI->getDebugLoc());2117  NT->setDebugLoc(DebugLoc::getMergedLocations(Locs));2118 2119  // PHIs created below will adopt NT's merged DebugLoc.2120  IRBuilder<NoFolder> Builder(NT);2121 2122  // In the case of an if statement, hoisting one of the terminators from our2123  // successor is a great thing. Unfortunately, the successors of the if/else2124  // blocks may have PHI nodes in them.  If they do, all PHI entries for BB1/BB22125  // must agree for all PHI nodes, so we insert select instruction to compute2126  // the final result.2127  if (BI) {2128    std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;2129    for (BasicBlock *Succ : successors(BB1)) {2130      for (PHINode &PN : Succ->phis()) {2131        Value *BB1V = PN.getIncomingValueForBlock(BB1);2132        Value *BB2V = PN.getIncomingValueForBlock(BB2);2133        if (BB1V == BB2V)2134          continue;2135 2136        // These values do not agree.  Insert a select instruction before NT2137        // that determines the right value.2138        SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];2139        if (!SI) {2140          // Propagate fast-math-flags from phi node to its replacement select.2141          SI = cast<SelectInst>(Builder.CreateSelectFMF(2142              BI->getCondition(), BB1V, BB2V,2143              isa<FPMathOperator>(PN) ? &PN : nullptr,2144              BB1V->getName() + "." + BB2V->getName(), BI));2145        }2146 2147        // Make the PHI node use the select for all incoming values for BB1/BB22148        for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)2149          if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)2150            PN.setIncomingValue(i, SI);2151      }2152    }2153  }2154 2155  SmallVector<DominatorTree::UpdateType, 4> Updates;2156 2157  // Update any PHI nodes in our new successors.2158  for (BasicBlock *Succ : successors(BB1)) {2159    addPredecessorToBlock(Succ, TIParent, BB1);2160    if (DTU)2161      Updates.push_back({DominatorTree::Insert, TIParent, Succ});2162  }2163 2164  if (DTU)2165    for (BasicBlock *Succ : successors(TI))2166      Updates.push_back({DominatorTree::Delete, TIParent, Succ});2167 2168  eraseTerminatorAndDCECond(TI);2169  if (DTU)2170    DTU->applyUpdates(Updates);2171  return Changed;2172}2173 2174// TODO: Refine this. This should avoid cases like turning constant memcpy sizes2175// into variables.2176static bool replacingOperandWithVariableIsCheap(const Instruction *I,2177                                                int OpIdx) {2178  // Divide/Remainder by constant is typically much cheaper than by variable.2179  if (I->isIntDivRem())2180    return OpIdx != 1;2181  return !isa<IntrinsicInst>(I);2182}2183 2184// All instructions in Insts belong to different blocks that all unconditionally2185// branch to a common successor. Analyze each instruction and return true if it2186// would be possible to sink them into their successor, creating one common2187// instruction instead. For every value that would be required to be provided by2188// PHI node (because an operand varies in each input block), add to PHIOperands.2189static bool canSinkInstructions(2190    ArrayRef<Instruction *> Insts,2191    DenseMap<const Use *, SmallVector<Value *, 4>> &PHIOperands) {2192  // Prune out obviously bad instructions to move. Each instruction must have2193  // the same number of uses, and we check later that the uses are consistent.2194  std::optional<unsigned> NumUses;2195  for (auto *I : Insts) {2196    // These instructions may change or break semantics if moved.2197    if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||2198        I->getType()->isTokenTy())2199      return false;2200 2201    // Do not try to sink an instruction in an infinite loop - it can cause2202    // this algorithm to infinite loop.2203    if (I->getParent()->getSingleSuccessor() == I->getParent())2204      return false;2205 2206    // Conservatively return false if I is an inline-asm instruction. Sinking2207    // and merging inline-asm instructions can potentially create arguments2208    // that cannot satisfy the inline-asm constraints.2209    // If the instruction has nomerge or convergent attribute, return false.2210    if (const auto *C = dyn_cast<CallBase>(I))2211      if (C->isInlineAsm() || C->cannotMerge() || C->isConvergent())2212        return false;2213 2214    if (!NumUses)2215      NumUses = I->getNumUses();2216    else if (NumUses != I->getNumUses())2217      return false;2218  }2219 2220  const Instruction *I0 = Insts.front();2221  const auto I0MMRA = MMRAMetadata(*I0);2222  for (auto *I : Insts) {2223    if (!I->isSameOperationAs(I0, Instruction::CompareUsingIntersectedAttrs))2224      return false;2225 2226    // Treat MMRAs conservatively. This pass can be quite aggressive and2227    // could drop a lot of MMRAs otherwise.2228    if (MMRAMetadata(*I) != I0MMRA)2229      return false;2230  }2231 2232  // Uses must be consistent: If I0 is used in a phi node in the sink target,2233  // then the other phi operands must match the instructions from Insts. This2234  // also has to hold true for any phi nodes that would be created as a result2235  // of sinking. Both of these cases are represented by PhiOperands.2236  for (const Use &U : I0->uses()) {2237    auto It = PHIOperands.find(&U);2238    if (It == PHIOperands.end())2239      // There may be uses in other blocks when sinking into a loop header.2240      return false;2241    if (!equal(Insts, It->second))2242      return false;2243  }2244 2245  // For calls to be sinkable, they must all be indirect, or have same callee.2246  // I.e. if we have two direct calls to different callees, we don't want to2247  // turn that into an indirect call. Likewise, if we have an indirect call,2248  // and a direct call, we don't actually want to have a single indirect call.2249  if (isa<CallBase>(I0)) {2250    auto IsIndirectCall = [](const Instruction *I) {2251      return cast<CallBase>(I)->isIndirectCall();2252    };2253    bool HaveIndirectCalls = any_of(Insts, IsIndirectCall);2254    bool AllCallsAreIndirect = all_of(Insts, IsIndirectCall);2255    if (HaveIndirectCalls) {2256      if (!AllCallsAreIndirect)2257        return false;2258    } else {2259      // All callees must be identical.2260      Value *Callee = nullptr;2261      for (const Instruction *I : Insts) {2262        Value *CurrCallee = cast<CallBase>(I)->getCalledOperand();2263        if (!Callee)2264          Callee = CurrCallee;2265        else if (Callee != CurrCallee)2266          return false;2267      }2268    }2269  }2270 2271  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {2272    Value *Op = I0->getOperand(OI);2273    auto SameAsI0 = [&I0, OI](const Instruction *I) {2274      assert(I->getNumOperands() == I0->getNumOperands());2275      return I->getOperand(OI) == I0->getOperand(OI);2276    };2277    if (!all_of(Insts, SameAsI0)) {2278      if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||2279          !canReplaceOperandWithVariable(I0, OI))2280        // We can't create a PHI from this GEP.2281        return false;2282      auto &Ops = PHIOperands[&I0->getOperandUse(OI)];2283      for (auto *I : Insts)2284        Ops.push_back(I->getOperand(OI));2285    }2286  }2287  return true;2288}2289 2290// Assuming canSinkInstructions(Blocks) has returned true, sink the last2291// instruction of every block in Blocks to their common successor, commoning2292// into one instruction.2293static void sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {2294  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);2295 2296  // canSinkInstructions returning true guarantees that every block has at2297  // least one non-terminator instruction.2298  SmallVector<Instruction*,4> Insts;2299  for (auto *BB : Blocks) {2300    Instruction *I = BB->getTerminator();2301    I = I->getPrevNode();2302    Insts.push_back(I);2303  }2304 2305  // We don't need to do any more checking here; canSinkInstructions should2306  // have done it all for us.2307  SmallVector<Value*, 4> NewOperands;2308  Instruction *I0 = Insts.front();2309  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {2310    // This check is different to that in canSinkInstructions. There, we2311    // cared about the global view once simplifycfg (and instcombine) have2312    // completed - it takes into account PHIs that become trivially2313    // simplifiable.  However here we need a more local view; if an operand2314    // differs we create a PHI and rely on instcombine to clean up the very2315    // small mess we may make.2316    bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {2317      return I->getOperand(O) != I0->getOperand(O);2318    });2319    if (!NeedPHI) {2320      NewOperands.push_back(I0->getOperand(O));2321      continue;2322    }2323 2324    // Create a new PHI in the successor block and populate it.2325    auto *Op = I0->getOperand(O);2326    assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");2327    auto *PN =2328        PHINode::Create(Op->getType(), Insts.size(), Op->getName() + ".sink");2329    PN->insertBefore(BBEnd->begin());2330    for (auto *I : Insts)2331      PN->addIncoming(I->getOperand(O), I->getParent());2332    NewOperands.push_back(PN);2333  }2334 2335  // Arbitrarily use I0 as the new "common" instruction; remap its operands2336  // and move it to the start of the successor block.2337  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)2338    I0->getOperandUse(O).set(NewOperands[O]);2339 2340  I0->moveBefore(*BBEnd, BBEnd->getFirstInsertionPt());2341 2342  // Update metadata and IR flags, and merge debug locations.2343  for (auto *I : Insts)2344    if (I != I0) {2345      // The debug location for the "common" instruction is the merged locations2346      // of all the commoned instructions.  We start with the original location2347      // of the "common" instruction and iteratively merge each location in the2348      // loop below.2349      // This is an N-way merge, which will be inefficient if I0 is a CallInst.2350      // However, as N-way merge for CallInst is rare, so we use simplified API2351      // instead of using complex API for N-way merge.2352      I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());2353      combineMetadataForCSE(I0, I, true);2354      I0->andIRFlags(I);2355      if (auto *CB = dyn_cast<CallBase>(I0)) {2356        bool Success = CB->tryIntersectAttributes(cast<CallBase>(I));2357        assert(Success && "We should not be trying to sink callbases "2358                          "with non-intersectable attributes");2359        // For NDEBUG Compile.2360        (void)Success;2361      }2362    }2363 2364  for (User *U : make_early_inc_range(I0->users())) {2365    // canSinkLastInstruction checked that all instructions are only used by2366    // phi nodes in a way that allows replacing the phi node with the common2367    // instruction.2368    auto *PN = cast<PHINode>(U);2369    PN->replaceAllUsesWith(I0);2370    PN->eraseFromParent();2371  }2372 2373  // Finally nuke all instructions apart from the common instruction.2374  for (auto *I : Insts) {2375    if (I == I0)2376      continue;2377    // The remaining uses are debug users, replace those with the common inst.2378    // In most (all?) cases this just introduces a use-before-def.2379    assert(I->user_empty() && "Inst unexpectedly still has non-dbg users");2380    I->replaceAllUsesWith(I0);2381    I->eraseFromParent();2382  }2383}2384 2385/// Check whether BB's predecessors end with unconditional branches. If it is2386/// true, sink any common code from the predecessors to BB.2387static bool sinkCommonCodeFromPredecessors(BasicBlock *BB,2388                                           DomTreeUpdater *DTU) {2389  // We support two situations:2390  //   (1) all incoming arcs are unconditional2391  //   (2) there are non-unconditional incoming arcs2392  //2393  // (2) is very common in switch defaults and2394  // else-if patterns;2395  //2396  //   if (a) f(1);2397  //   else if (b) f(2);2398  //2399  // produces:2400  //2401  //       [if]2402  //      /    \2403  //    [f(1)] [if]2404  //      |     | \2405  //      |     |  |2406  //      |  [f(2)]|2407  //       \    | /2408  //        [ end ]2409  //2410  // [end] has two unconditional predecessor arcs and one conditional. The2411  // conditional refers to the implicit empty 'else' arc. This conditional2412  // arc can also be caused by an empty default block in a switch.2413  //2414  // In this case, we attempt to sink code from all *unconditional* arcs.2415  // If we can sink instructions from these arcs (determined during the scan2416  // phase below) we insert a common successor for all unconditional arcs and2417  // connect that to [end], to enable sinking:2418  //2419  //       [if]2420  //      /    \2421  //    [x(1)] [if]2422  //      |     | \2423  //      |     |  \2424  //      |  [x(2)] |2425  //       \   /    |2426  //   [sink.split] |2427  //         \     /2428  //         [ end ]2429  //2430  SmallVector<BasicBlock*,4> UnconditionalPreds;2431  bool HaveNonUnconditionalPredecessors = false;2432  for (auto *PredBB : predecessors(BB)) {2433    auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());2434    if (PredBr && PredBr->isUnconditional())2435      UnconditionalPreds.push_back(PredBB);2436    else2437      HaveNonUnconditionalPredecessors = true;2438  }2439  if (UnconditionalPreds.size() < 2)2440    return false;2441 2442  // We take a two-step approach to tail sinking. First we scan from the end of2443  // each block upwards in lockstep. If the n'th instruction from the end of each2444  // block can be sunk, those instructions are added to ValuesToSink and we2445  // carry on. If we can sink an instruction but need to PHI-merge some operands2446  // (because they're not identical in each instruction) we add these to2447  // PHIOperands.2448  // We prepopulate PHIOperands with the phis that already exist in BB.2449  DenseMap<const Use *, SmallVector<Value *, 4>> PHIOperands;2450  for (PHINode &PN : BB->phis()) {2451    SmallDenseMap<BasicBlock *, const Use *, 4> IncomingVals;2452    for (const Use &U : PN.incoming_values())2453      IncomingVals.insert({PN.getIncomingBlock(U), &U});2454    auto &Ops = PHIOperands[IncomingVals[UnconditionalPreds[0]]];2455    for (BasicBlock *Pred : UnconditionalPreds)2456      Ops.push_back(*IncomingVals[Pred]);2457  }2458 2459  int ScanIdx = 0;2460  SmallPtrSet<Value*,4> InstructionsToSink;2461  LockstepReverseIterator<true> LRI(UnconditionalPreds);2462  while (LRI.isValid() &&2463         canSinkInstructions(*LRI, PHIOperands)) {2464    LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]2465                      << "\n");2466    InstructionsToSink.insert_range(*LRI);2467    ++ScanIdx;2468    --LRI;2469  }2470 2471  // If no instructions can be sunk, early-return.2472  if (ScanIdx == 0)2473    return false;2474 2475  bool followedByDeoptOrUnreachable = IsBlockFollowedByDeoptOrUnreachable(BB);2476 2477  if (!followedByDeoptOrUnreachable) {2478    // Check whether this is the pointer operand of a load/store.2479    auto IsMemOperand = [](Use &U) {2480      auto *I = cast<Instruction>(U.getUser());2481      if (isa<LoadInst>(I))2482        return U.getOperandNo() == LoadInst::getPointerOperandIndex();2483      if (isa<StoreInst>(I))2484        return U.getOperandNo() == StoreInst::getPointerOperandIndex();2485      return false;2486    };2487 2488    // Okay, we *could* sink last ScanIdx instructions. But how many can we2489    // actually sink before encountering instruction that is unprofitable to2490    // sink?2491    auto ProfitableToSinkInstruction = [&](LockstepReverseIterator<true> &LRI) {2492      unsigned NumPHIInsts = 0;2493      for (Use &U : (*LRI)[0]->operands()) {2494        auto It = PHIOperands.find(&U);2495        if (It != PHIOperands.end() && !all_of(It->second, [&](Value *V) {2496              return InstructionsToSink.contains(V);2497            })) {2498          ++NumPHIInsts;2499          // Do not separate a load/store from the gep producing the address.2500          // The gep can likely be folded into the load/store as an addressing2501          // mode. Additionally, a load of a gep is easier to analyze than a2502          // load of a phi.2503          if (IsMemOperand(U) &&2504              any_of(It->second, [](Value *V) { return isa<GEPOperator>(V); }))2505            return false;2506          // FIXME: this check is overly optimistic. We may end up not sinking2507          // said instruction, due to the very same profitability check.2508          // See @creating_too_many_phis in sink-common-code.ll.2509        }2510      }2511      LLVM_DEBUG(dbgs() << "SINK: #phi insts: " << NumPHIInsts << "\n");2512      return NumPHIInsts <= 1;2513    };2514 2515    // We've determined that we are going to sink last ScanIdx instructions,2516    // and recorded them in InstructionsToSink. Now, some instructions may be2517    // unprofitable to sink. But that determination depends on the instructions2518    // that we are going to sink.2519 2520    // First, forward scan: find the first instruction unprofitable to sink,2521    // recording all the ones that are profitable to sink.2522    // FIXME: would it be better, after we detect that not all are profitable.2523    // to either record the profitable ones, or erase the unprofitable ones?2524    // Maybe we need to choose (at runtime) the one that will touch least2525    // instrs?2526    LRI.reset();2527    int Idx = 0;2528    SmallPtrSet<Value *, 4> InstructionsProfitableToSink;2529    while (Idx < ScanIdx) {2530      if (!ProfitableToSinkInstruction(LRI)) {2531        // Too many PHIs would be created.2532        LLVM_DEBUG(2533            dbgs() << "SINK: stopping here, too many PHIs would be created!\n");2534        break;2535      }2536      InstructionsProfitableToSink.insert_range(*LRI);2537      --LRI;2538      ++Idx;2539    }2540 2541    // If no instructions can be sunk, early-return.2542    if (Idx == 0)2543      return false;2544 2545    // Did we determine that (only) some instructions are unprofitable to sink?2546    if (Idx < ScanIdx) {2547      // Okay, some instructions are unprofitable.2548      ScanIdx = Idx;2549      InstructionsToSink = InstructionsProfitableToSink;2550 2551      // But, that may make other instructions unprofitable, too.2552      // So, do a backward scan, do any earlier instructions become2553      // unprofitable?2554      assert(2555          !ProfitableToSinkInstruction(LRI) &&2556          "We already know that the last instruction is unprofitable to sink");2557      ++LRI;2558      --Idx;2559      while (Idx >= 0) {2560        // If we detect that an instruction becomes unprofitable to sink,2561        // all earlier instructions won't be sunk either,2562        // so preemptively keep InstructionsProfitableToSink in sync.2563        // FIXME: is this the most performant approach?2564        for (auto *I : *LRI)2565          InstructionsProfitableToSink.erase(I);2566        if (!ProfitableToSinkInstruction(LRI)) {2567          // Everything starting with this instruction won't be sunk.2568          ScanIdx = Idx;2569          InstructionsToSink = InstructionsProfitableToSink;2570        }2571        ++LRI;2572        --Idx;2573      }2574    }2575 2576    // If no instructions can be sunk, early-return.2577    if (ScanIdx == 0)2578      return false;2579  }2580 2581  bool Changed = false;2582 2583  if (HaveNonUnconditionalPredecessors) {2584    if (!followedByDeoptOrUnreachable) {2585      // It is always legal to sink common instructions from unconditional2586      // predecessors. However, if not all predecessors are unconditional,2587      // this transformation might be pessimizing. So as a rule of thumb,2588      // don't do it unless we'd sink at least one non-speculatable instruction.2589      // See https://bugs.llvm.org/show_bug.cgi?id=302442590      LRI.reset();2591      int Idx = 0;2592      bool Profitable = false;2593      while (Idx < ScanIdx) {2594        if (!isSafeToSpeculativelyExecute((*LRI)[0])) {2595          Profitable = true;2596          break;2597        }2598        --LRI;2599        ++Idx;2600      }2601      if (!Profitable)2602        return false;2603    }2604 2605    LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");2606    // We have a conditional edge and we're going to sink some instructions.2607    // Insert a new block postdominating all blocks we're going to sink from.2608    if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))2609      // Edges couldn't be split.2610      return false;2611    Changed = true;2612  }2613 2614  // Now that we've analyzed all potential sinking candidates, perform the2615  // actual sink. We iteratively sink the last non-terminator of the source2616  // blocks into their common successor unless doing so would require too2617  // many PHI instructions to be generated (currently only one PHI is allowed2618  // per sunk instruction).2619  //2620  // We can use InstructionsToSink to discount values needing PHI-merging that will2621  // actually be sunk in a later iteration. This allows us to be more2622  // aggressive in what we sink. This does allow a false positive where we2623  // sink presuming a later value will also be sunk, but stop half way through2624  // and never actually sink it which means we produce more PHIs than intended.2625  // This is unlikely in practice though.2626  int SinkIdx = 0;2627  for (; SinkIdx != ScanIdx; ++SinkIdx) {2628    LLVM_DEBUG(dbgs() << "SINK: Sink: "2629                      << *UnconditionalPreds[0]->getTerminator()->getPrevNode()2630                      << "\n");2631 2632    // Because we've sunk every instruction in turn, the current instruction to2633    // sink is always at index 0.2634    LRI.reset();2635 2636    sinkLastInstruction(UnconditionalPreds);2637    NumSinkCommonInstrs++;2638    Changed = true;2639  }2640  if (SinkIdx != 0)2641    ++NumSinkCommonCode;2642  return Changed;2643}2644 2645namespace {2646 2647struct CompatibleSets {2648  using SetTy = SmallVector<InvokeInst *, 2>;2649 2650  SmallVector<SetTy, 1> Sets;2651 2652  static bool shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes);2653 2654  SetTy &getCompatibleSet(InvokeInst *II);2655 2656  void insert(InvokeInst *II);2657};2658 2659CompatibleSets::SetTy &CompatibleSets::getCompatibleSet(InvokeInst *II) {2660  // Perform a linear scan over all the existing sets, see if the new `invoke`2661  // is compatible with any particular set. Since we know that all the `invokes`2662  // within a set are compatible, only check the first `invoke` in each set.2663  // WARNING: at worst, this has quadratic complexity.2664  for (CompatibleSets::SetTy &Set : Sets) {2665    if (CompatibleSets::shouldBelongToSameSet({Set.front(), II}))2666      return Set;2667  }2668 2669  // Otherwise, we either had no sets yet, or this invoke forms a new set.2670  return Sets.emplace_back();2671}2672 2673void CompatibleSets::insert(InvokeInst *II) {2674  getCompatibleSet(II).emplace_back(II);2675}2676 2677bool CompatibleSets::shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes) {2678  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");2679 2680  // Can we theoretically merge these `invoke`s?2681  auto IsIllegalToMerge = [](InvokeInst *II) {2682    return II->cannotMerge() || II->isInlineAsm();2683  };2684  if (any_of(Invokes, IsIllegalToMerge))2685    return false;2686 2687  // Either both `invoke`s must be   direct,2688  // or     both `invoke`s must be indirect.2689  auto IsIndirectCall = [](InvokeInst *II) { return II->isIndirectCall(); };2690  bool HaveIndirectCalls = any_of(Invokes, IsIndirectCall);2691  bool AllCallsAreIndirect = all_of(Invokes, IsIndirectCall);2692  if (HaveIndirectCalls) {2693    if (!AllCallsAreIndirect)2694      return false;2695  } else {2696    // All callees must be identical.2697    Value *Callee = nullptr;2698    for (InvokeInst *II : Invokes) {2699      Value *CurrCallee = II->getCalledOperand();2700      assert(CurrCallee && "There is always a called operand.");2701      if (!Callee)2702        Callee = CurrCallee;2703      else if (Callee != CurrCallee)2704        return false;2705    }2706  }2707 2708  // Either both `invoke`s must not have a normal destination,2709  // or     both `invoke`s must     have a normal destination,2710  auto HasNormalDest = [](InvokeInst *II) {2711    return !isa<UnreachableInst>(II->getNormalDest()->getFirstNonPHIOrDbg());2712  };2713  if (any_of(Invokes, HasNormalDest)) {2714    // Do not merge `invoke` that does not have a normal destination with one2715    // that does have a normal destination, even though doing so would be legal.2716    if (!all_of(Invokes, HasNormalDest))2717      return false;2718 2719    // All normal destinations must be identical.2720    BasicBlock *NormalBB = nullptr;2721    for (InvokeInst *II : Invokes) {2722      BasicBlock *CurrNormalBB = II->getNormalDest();2723      assert(CurrNormalBB && "There is always a 'continue to' basic block.");2724      if (!NormalBB)2725        NormalBB = CurrNormalBB;2726      else if (NormalBB != CurrNormalBB)2727        return false;2728    }2729 2730    // In the normal destination, the incoming values for these two `invoke`s2731    // must be compatible.2732    SmallPtrSet<Value *, 16> EquivalenceSet(llvm::from_range, Invokes);2733    if (!incomingValuesAreCompatible(2734            NormalBB, {Invokes[0]->getParent(), Invokes[1]->getParent()},2735            &EquivalenceSet))2736      return false;2737  }2738 2739#ifndef NDEBUG2740  // All unwind destinations must be identical.2741  // We know that because we have started from said unwind destination.2742  BasicBlock *UnwindBB = nullptr;2743  for (InvokeInst *II : Invokes) {2744    BasicBlock *CurrUnwindBB = II->getUnwindDest();2745    assert(CurrUnwindBB && "There is always an 'unwind to' basic block.");2746    if (!UnwindBB)2747      UnwindBB = CurrUnwindBB;2748    else2749      assert(UnwindBB == CurrUnwindBB && "Unexpected unwind destination.");2750  }2751#endif2752 2753  // In the unwind destination, the incoming values for these two `invoke`s2754  // must be compatible.2755  if (!incomingValuesAreCompatible(2756          Invokes.front()->getUnwindDest(),2757          {Invokes[0]->getParent(), Invokes[1]->getParent()}))2758    return false;2759 2760  // Ignoring arguments, these `invoke`s must be identical,2761  // including operand bundles.2762  const InvokeInst *II0 = Invokes.front();2763  for (auto *II : Invokes.drop_front())2764    if (!II->isSameOperationAs(II0, Instruction::CompareUsingIntersectedAttrs))2765      return false;2766 2767  // Can we theoretically form the data operands for the merged `invoke`?2768  auto IsIllegalToMergeArguments = [](auto Ops) {2769    Use &U0 = std::get<0>(Ops);2770    Use &U1 = std::get<1>(Ops);2771    if (U0 == U1)2772      return false;2773    return !canReplaceOperandWithVariable(cast<Instruction>(U0.getUser()),2774                                          U0.getOperandNo());2775  };2776  assert(Invokes.size() == 2 && "Always called with exactly two candidates.");2777  if (any_of(zip(Invokes[0]->data_ops(), Invokes[1]->data_ops()),2778             IsIllegalToMergeArguments))2779    return false;2780 2781  return true;2782}2783 2784} // namespace2785 2786// Merge all invokes in the provided set, all of which are compatible2787// as per the `CompatibleSets::shouldBelongToSameSet()`.2788static void mergeCompatibleInvokesImpl(ArrayRef<InvokeInst *> Invokes,2789                                       DomTreeUpdater *DTU) {2790  assert(Invokes.size() >= 2 && "Must have at least two invokes to merge.");2791 2792  SmallVector<DominatorTree::UpdateType, 8> Updates;2793  if (DTU)2794    Updates.reserve(2 + 3 * Invokes.size());2795 2796  bool HasNormalDest =2797      !isa<UnreachableInst>(Invokes[0]->getNormalDest()->getFirstNonPHIOrDbg());2798 2799  // Clone one of the invokes into a new basic block.2800  // Since they are all compatible, it doesn't matter which invoke is cloned.2801  InvokeInst *MergedInvoke = [&Invokes, HasNormalDest]() {2802    InvokeInst *II0 = Invokes.front();2803    BasicBlock *II0BB = II0->getParent();2804    BasicBlock *InsertBeforeBlock =2805        II0->getParent()->getIterator()->getNextNode();2806    Function *Func = II0BB->getParent();2807    LLVMContext &Ctx = II0->getContext();2808 2809    BasicBlock *MergedInvokeBB = BasicBlock::Create(2810        Ctx, II0BB->getName() + ".invoke", Func, InsertBeforeBlock);2811 2812    auto *MergedInvoke = cast<InvokeInst>(II0->clone());2813    // NOTE: all invokes have the same attributes, so no handling needed.2814    MergedInvoke->insertInto(MergedInvokeBB, MergedInvokeBB->end());2815 2816    if (!HasNormalDest) {2817      // This set does not have a normal destination,2818      // so just form a new block with unreachable terminator.2819      BasicBlock *MergedNormalDest = BasicBlock::Create(2820          Ctx, II0BB->getName() + ".cont", Func, InsertBeforeBlock);2821      auto *UI = new UnreachableInst(Ctx, MergedNormalDest);2822      UI->setDebugLoc(DebugLoc::getTemporary());2823      MergedInvoke->setNormalDest(MergedNormalDest);2824    }2825 2826    // The unwind destination, however, remainds identical for all invokes here.2827 2828    return MergedInvoke;2829  }();2830 2831  if (DTU) {2832    // Predecessor blocks that contained these invokes will now branch to2833    // the new block that contains the merged invoke, ...2834    for (InvokeInst *II : Invokes)2835      Updates.push_back(2836          {DominatorTree::Insert, II->getParent(), MergedInvoke->getParent()});2837 2838    // ... which has the new `unreachable` block as normal destination,2839    // or unwinds to the (same for all `invoke`s in this set) `landingpad`,2840    for (BasicBlock *SuccBBOfMergedInvoke : successors(MergedInvoke))2841      Updates.push_back({DominatorTree::Insert, MergedInvoke->getParent(),2842                         SuccBBOfMergedInvoke});2843 2844    // Since predecessor blocks now unconditionally branch to a new block,2845    // they no longer branch to their original successors.2846    for (InvokeInst *II : Invokes)2847      for (BasicBlock *SuccOfPredBB : successors(II->getParent()))2848        Updates.push_back(2849            {DominatorTree::Delete, II->getParent(), SuccOfPredBB});2850  }2851 2852  bool IsIndirectCall = Invokes[0]->isIndirectCall();2853 2854  // Form the merged operands for the merged invoke.2855  for (Use &U : MergedInvoke->operands()) {2856    // Only PHI together the indirect callees and data operands.2857    if (MergedInvoke->isCallee(&U)) {2858      if (!IsIndirectCall)2859        continue;2860    } else if (!MergedInvoke->isDataOperand(&U))2861      continue;2862 2863    // Don't create trivial PHI's with all-identical incoming values.2864    bool NeedPHI = any_of(Invokes, [&U](InvokeInst *II) {2865      return II->getOperand(U.getOperandNo()) != U.get();2866    });2867    if (!NeedPHI)2868      continue;2869 2870    // Form a PHI out of all the data ops under this index.2871    PHINode *PN = PHINode::Create(2872        U->getType(), /*NumReservedValues=*/Invokes.size(), "", MergedInvoke->getIterator());2873    for (InvokeInst *II : Invokes)2874      PN->addIncoming(II->getOperand(U.getOperandNo()), II->getParent());2875 2876    U.set(PN);2877  }2878 2879  // We've ensured that each PHI node has compatible (identical) incoming values2880  // when coming from each of the `invoke`s in the current merge set,2881  // so update the PHI nodes accordingly.2882  for (BasicBlock *Succ : successors(MergedInvoke))2883    addPredecessorToBlock(Succ, /*NewPred=*/MergedInvoke->getParent(),2884                          /*ExistPred=*/Invokes.front()->getParent());2885 2886  // And finally, replace the original `invoke`s with an unconditional branch2887  // to the block with the merged `invoke`. Also, give that merged `invoke`2888  // the merged debugloc of all the original `invoke`s.2889  DILocation *MergedDebugLoc = nullptr;2890  for (InvokeInst *II : Invokes) {2891    // Compute the debug location common to all the original `invoke`s.2892    if (!MergedDebugLoc)2893      MergedDebugLoc = II->getDebugLoc();2894    else2895      MergedDebugLoc =2896          DebugLoc::getMergedLocation(MergedDebugLoc, II->getDebugLoc());2897 2898    // And replace the old `invoke` with an unconditionally branch2899    // to the block with the merged `invoke`.2900    for (BasicBlock *OrigSuccBB : successors(II->getParent()))2901      OrigSuccBB->removePredecessor(II->getParent());2902    auto *BI = BranchInst::Create(MergedInvoke->getParent(), II->getParent());2903    // The unconditional branch is part of the replacement for the original2904    // invoke, so should use its DebugLoc.2905    BI->setDebugLoc(II->getDebugLoc());2906    bool Success = MergedInvoke->tryIntersectAttributes(II);2907    assert(Success && "Merged invokes with incompatible attributes");2908    // For NDEBUG Compile2909    (void)Success;2910    II->replaceAllUsesWith(MergedInvoke);2911    II->eraseFromParent();2912    ++NumInvokesMerged;2913  }2914  MergedInvoke->setDebugLoc(MergedDebugLoc);2915  ++NumInvokeSetsFormed;2916 2917  if (DTU)2918    DTU->applyUpdates(Updates);2919}2920 2921/// If this block is a `landingpad` exception handling block, categorize all2922/// the predecessor `invoke`s into sets, with all `invoke`s in each set2923/// being "mergeable" together, and then merge invokes in each set together.2924///2925/// This is a weird mix of hoisting and sinking. Visually, it goes from:2926///          [...]        [...]2927///            |            |2928///        [invoke0]    [invoke1]2929///           / \          / \2930///     [cont0] [landingpad] [cont1]2931/// to:2932///      [...] [...]2933///          \ /2934///       [invoke]2935///          / \2936///     [cont] [landingpad]2937///2938/// But of course we can only do that if the invokes share the `landingpad`,2939/// edges invoke0->cont0 and invoke1->cont1 are "compatible",2940/// and the invoked functions are "compatible".2941static bool mergeCompatibleInvokes(BasicBlock *BB, DomTreeUpdater *DTU) {2942  if (!EnableMergeCompatibleInvokes)2943    return false;2944 2945  bool Changed = false;2946 2947  // FIXME: generalize to all exception handling blocks?2948  if (!BB->isLandingPad())2949    return Changed;2950 2951  CompatibleSets Grouper;2952 2953  // Record all the predecessors of this `landingpad`. As per verifier,2954  // the only allowed predecessor is the unwind edge of an `invoke`.2955  // We want to group "compatible" `invokes` into the same set to be merged.2956  for (BasicBlock *PredBB : predecessors(BB))2957    Grouper.insert(cast<InvokeInst>(PredBB->getTerminator()));2958 2959  // And now, merge `invoke`s that were grouped togeter.2960  for (ArrayRef<InvokeInst *> Invokes : Grouper.Sets) {2961    if (Invokes.size() < 2)2962      continue;2963    Changed = true;2964    mergeCompatibleInvokesImpl(Invokes, DTU);2965  }2966 2967  return Changed;2968}2969 2970namespace {2971/// Track ephemeral values, which should be ignored for cost-modelling2972/// purposes. Requires walking instructions in reverse order.2973class EphemeralValueTracker {2974  SmallPtrSet<const Instruction *, 32> EphValues;2975 2976  bool isEphemeral(const Instruction *I) {2977    if (isa<AssumeInst>(I))2978      return true;2979    return !I->mayHaveSideEffects() && !I->isTerminator() &&2980           all_of(I->users(), [&](const User *U) {2981             return EphValues.count(cast<Instruction>(U));2982           });2983  }2984 2985public:2986  bool track(const Instruction *I) {2987    if (isEphemeral(I)) {2988      EphValues.insert(I);2989      return true;2990    }2991    return false;2992  }2993 2994  bool contains(const Instruction *I) const { return EphValues.contains(I); }2995};2996} // namespace2997 2998/// Determine if we can hoist sink a sole store instruction out of a2999/// conditional block.3000///3001/// We are looking for code like the following:3002///   BrBB:3003///     store i32 %add, i32* %arrayidx23004///     ... // No other stores or function calls (we could be calling a memory3005///     ... // function).3006///     %cmp = icmp ult %x, %y3007///     br i1 %cmp, label %EndBB, label %ThenBB3008///   ThenBB:3009///     store i32 %add5, i32* %arrayidx23010///     br label EndBB3011///   EndBB:3012///     ...3013///   We are going to transform this into:3014///   BrBB:3015///     store i32 %add, i32* %arrayidx23016///     ... //3017///     %cmp = icmp ult %x, %y3018///     %add.add5 = select i1 %cmp, i32 %add, %add53019///     store i32 %add.add5, i32* %arrayidx23020///     ...3021///3022/// \return The pointer to the value of the previous store if the store can be3023///         hoisted into the predecessor block. 0 otherwise.3024static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,3025                                     BasicBlock *StoreBB, BasicBlock *EndBB) {3026  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);3027  if (!StoreToHoist)3028    return nullptr;3029 3030  // Volatile or atomic.3031  if (!StoreToHoist->isSimple())3032    return nullptr;3033 3034  Value *StorePtr = StoreToHoist->getPointerOperand();3035  Type *StoreTy = StoreToHoist->getValueOperand()->getType();3036 3037  // Look for a store to the same pointer in BrBB.3038  unsigned MaxNumInstToLookAt = 9;3039  // Skip pseudo probe intrinsic calls which are not really killing any memory3040  // accesses.3041  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {3042    if (!MaxNumInstToLookAt)3043      break;3044    --MaxNumInstToLookAt;3045 3046    // Could be calling an instruction that affects memory like free().3047    if (CurI.mayWriteToMemory() && !isa<StoreInst>(CurI))3048      return nullptr;3049 3050    if (auto *SI = dyn_cast<StoreInst>(&CurI)) {3051      // Found the previous store to same location and type. Make sure it is3052      // simple, to avoid introducing a spurious non-atomic write after an3053      // atomic write.3054      if (SI->getPointerOperand() == StorePtr &&3055          SI->getValueOperand()->getType() == StoreTy && SI->isSimple() &&3056          SI->getAlign() >= StoreToHoist->getAlign())3057        // Found the previous store, return its value operand.3058        return SI->getValueOperand();3059      return nullptr; // Unknown store.3060    }3061 3062    if (auto *LI = dyn_cast<LoadInst>(&CurI)) {3063      if (LI->getPointerOperand() == StorePtr && LI->getType() == StoreTy &&3064          LI->isSimple() && LI->getAlign() >= StoreToHoist->getAlign()) {3065        Value *Obj = getUnderlyingObject(StorePtr);3066        bool ExplicitlyDereferenceableOnly;3067        if (isWritableObject(Obj, ExplicitlyDereferenceableOnly) &&3068            capturesNothing(3069                PointerMayBeCaptured(Obj, /*ReturnCaptures=*/false,3070                                     CaptureComponents::Provenance)) &&3071            (!ExplicitlyDereferenceableOnly ||3072             isDereferenceablePointer(StorePtr, StoreTy,3073                                      LI->getDataLayout()))) {3074          // Found a previous load, return it.3075          return LI;3076        }3077      }3078      // The load didn't work out, but we may still find a store.3079    }3080  }3081 3082  return nullptr;3083}3084 3085/// Estimate the cost of the insertion(s) and check that the PHI nodes can be3086/// converted to selects.3087static bool validateAndCostRequiredSelects(BasicBlock *BB, BasicBlock *ThenBB,3088                                           BasicBlock *EndBB,3089                                           unsigned &SpeculatedInstructions,3090                                           InstructionCost &Cost,3091                                           const TargetTransformInfo &TTI) {3092  TargetTransformInfo::TargetCostKind CostKind =3093    BB->getParent()->hasMinSize()3094    ? TargetTransformInfo::TCK_CodeSize3095    : TargetTransformInfo::TCK_SizeAndLatency;3096 3097  bool HaveRewritablePHIs = false;3098  for (PHINode &PN : EndBB->phis()) {3099    Value *OrigV = PN.getIncomingValueForBlock(BB);3100    Value *ThenV = PN.getIncomingValueForBlock(ThenBB);3101 3102    // FIXME: Try to remove some of the duplication with3103    // hoistCommonCodeFromSuccessors. Skip PHIs which are trivial.3104    if (ThenV == OrigV)3105      continue;3106 3107    Cost += TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(),3108                                   CmpInst::makeCmpResultType(PN.getType()),3109                                   CmpInst::BAD_ICMP_PREDICATE, CostKind);3110 3111    // Don't convert to selects if we could remove undefined behavior instead.3112    if (passingValueIsAlwaysUndefined(OrigV, &PN) ||3113        passingValueIsAlwaysUndefined(ThenV, &PN))3114      return false;3115 3116    HaveRewritablePHIs = true;3117    ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);3118    ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);3119    if (!OrigCE && !ThenCE)3120      continue; // Known cheap (FIXME: Maybe not true for aggregates).3121 3122    InstructionCost OrigCost = OrigCE ? computeSpeculationCost(OrigCE, TTI) : 0;3123    InstructionCost ThenCost = ThenCE ? computeSpeculationCost(ThenCE, TTI) : 0;3124    InstructionCost MaxCost =3125        2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;3126    if (OrigCost + ThenCost > MaxCost)3127      return false;3128 3129    // Account for the cost of an unfolded ConstantExpr which could end up3130    // getting expanded into Instructions.3131    // FIXME: This doesn't account for how many operations are combined in the3132    // constant expression.3133    ++SpeculatedInstructions;3134    if (SpeculatedInstructions > 1)3135      return false;3136  }3137 3138  return HaveRewritablePHIs;3139}3140 3141static bool isProfitableToSpeculate(const BranchInst *BI,3142                                    std::optional<bool> Invert,3143                                    const TargetTransformInfo &TTI) {3144  // If the branch is non-unpredictable, and is predicted to *not* branch to3145  // the `then` block, then avoid speculating it.3146  if (BI->getMetadata(LLVMContext::MD_unpredictable))3147    return true;3148 3149  uint64_t TWeight, FWeight;3150  if (!extractBranchWeights(*BI, TWeight, FWeight) || (TWeight + FWeight) == 0)3151    return true;3152 3153  if (!Invert.has_value())3154    return false;3155 3156  uint64_t EndWeight = *Invert ? TWeight : FWeight;3157  BranchProbability BIEndProb =3158      BranchProbability::getBranchProbability(EndWeight, TWeight + FWeight);3159  BranchProbability Likely = TTI.getPredictableBranchThreshold();3160  return BIEndProb < Likely;3161}3162 3163/// Speculate a conditional basic block flattening the CFG.3164///3165/// Note that this is a very risky transform currently. Speculating3166/// instructions like this is most often not desirable. Instead, there is an MI3167/// pass which can do it with full awareness of the resource constraints.3168/// However, some cases are "obvious" and we should do directly. An example of3169/// this is speculating a single, reasonably cheap instruction.3170///3171/// There is only one distinct advantage to flattening the CFG at the IR level:3172/// it makes very common but simplistic optimizations such as are common in3173/// instcombine and the DAG combiner more powerful by removing CFG edges and3174/// modeling their effects with easier to reason about SSA value graphs.3175///3176///3177/// An illustration of this transform is turning this IR:3178/// \code3179///   BB:3180///     %cmp = icmp ult %x, %y3181///     br i1 %cmp, label %EndBB, label %ThenBB3182///   ThenBB:3183///     %sub = sub %x, %y3184///     br label BB23185///   EndBB:3186///     %phi = phi [ %sub, %ThenBB ], [ 0, %BB ]3187///     ...3188/// \endcode3189///3190/// Into this IR:3191/// \code3192///   BB:3193///     %cmp = icmp ult %x, %y3194///     %sub = sub %x, %y3195///     %cond = select i1 %cmp, 0, %sub3196///     ...3197/// \endcode3198///3199/// \returns true if the conditional block is removed.3200bool SimplifyCFGOpt::speculativelyExecuteBB(BranchInst *BI,3201                                            BasicBlock *ThenBB) {3202  if (!Options.SpeculateBlocks)3203    return false;3204 3205  // Be conservative for now. FP select instruction can often be expensive.3206  Value *BrCond = BI->getCondition();3207  if (isa<FCmpInst>(BrCond))3208    return false;3209 3210  BasicBlock *BB = BI->getParent();3211  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);3212  InstructionCost Budget =3213      PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;3214 3215  // If ThenBB is actually on the false edge of the conditional branch, remember3216  // to swap the select operands later.3217  bool Invert = false;3218  if (ThenBB != BI->getSuccessor(0)) {3219    assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");3220    Invert = true;3221  }3222  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");3223 3224  if (!isProfitableToSpeculate(BI, Invert, TTI))3225    return false;3226 3227  // Keep a count of how many times instructions are used within ThenBB when3228  // they are candidates for sinking into ThenBB. Specifically:3229  // - They are defined in BB, and3230  // - They have no side effects, and3231  // - All of their uses are in ThenBB.3232  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;3233 3234  SmallVector<Instruction *, 4> SpeculatedPseudoProbes;3235 3236  unsigned SpeculatedInstructions = 0;3237  bool HoistLoadsStores = Options.HoistLoadsStoresWithCondFaulting;3238  SmallVector<Instruction *, 2> SpeculatedConditionalLoadsStores;3239  Value *SpeculatedStoreValue = nullptr;3240  StoreInst *SpeculatedStore = nullptr;3241  EphemeralValueTracker EphTracker;3242  for (Instruction &I : reverse(drop_end(*ThenBB))) {3243    // Skip pseudo probes. The consequence is we lose track of the branch3244    // probability for ThenBB, which is fine since the optimization here takes3245    // place regardless of the branch probability.3246    if (isa<PseudoProbeInst>(I)) {3247      // The probe should be deleted so that it will not be over-counted when3248      // the samples collected on the non-conditional path are counted towards3249      // the conditional path. We leave it for the counts inference algorithm to3250      // figure out a proper count for an unknown probe.3251      SpeculatedPseudoProbes.push_back(&I);3252      continue;3253    }3254 3255    // Ignore ephemeral values, they will be dropped by the transform.3256    if (EphTracker.track(&I))3257      continue;3258 3259    // Only speculatively execute a single instruction (not counting the3260    // terminator) for now.3261    bool IsSafeCheapLoadStore = HoistLoadsStores &&3262                                isSafeCheapLoadStore(&I, TTI) &&3263                                SpeculatedConditionalLoadsStores.size() <3264                                    HoistLoadsStoresWithCondFaultingThreshold;3265    // Not count load/store into cost if target supports conditional faulting3266    // b/c it's cheap to speculate it.3267    if (IsSafeCheapLoadStore)3268      SpeculatedConditionalLoadsStores.push_back(&I);3269    else3270      ++SpeculatedInstructions;3271 3272    if (SpeculatedInstructions > 1)3273      return false;3274 3275    // Don't hoist the instruction if it's unsafe or expensive.3276    if (!IsSafeCheapLoadStore &&3277        !isSafeToSpeculativelyExecute(&I, BI, Options.AC) &&3278        !(HoistCondStores && !SpeculatedStoreValue &&3279          (SpeculatedStoreValue =3280               isSafeToSpeculateStore(&I, BB, ThenBB, EndBB))))3281      return false;3282    if (!IsSafeCheapLoadStore && !SpeculatedStoreValue &&3283        computeSpeculationCost(&I, TTI) >3284            PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)3285      return false;3286 3287    // Store the store speculation candidate.3288    if (!SpeculatedStore && SpeculatedStoreValue)3289      SpeculatedStore = cast<StoreInst>(&I);3290 3291    // Do not hoist the instruction if any of its operands are defined but not3292    // used in BB. The transformation will prevent the operand from3293    // being sunk into the use block.3294    for (Use &Op : I.operands()) {3295      Instruction *OpI = dyn_cast<Instruction>(Op);3296      if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())3297        continue; // Not a candidate for sinking.3298 3299      ++SinkCandidateUseCounts[OpI];3300    }3301  }3302 3303  // Consider any sink candidates which are only used in ThenBB as costs for3304  // speculation. Note, while we iterate over a DenseMap here, we are summing3305  // and so iteration order isn't significant.3306  for (const auto &[Inst, Count] : SinkCandidateUseCounts)3307    if (Inst->hasNUses(Count)) {3308      ++SpeculatedInstructions;3309      if (SpeculatedInstructions > 1)3310        return false;3311    }3312 3313  // Check that we can insert the selects and that it's not too expensive to do3314  // so.3315  bool Convert =3316      SpeculatedStore != nullptr || !SpeculatedConditionalLoadsStores.empty();3317  InstructionCost Cost = 0;3318  Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,3319                                            SpeculatedInstructions, Cost, TTI);3320  if (!Convert || Cost > Budget)3321    return false;3322 3323  // If we get here, we can hoist the instruction and if-convert.3324  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);3325 3326  Instruction *Sel = nullptr;3327  // Insert a select of the value of the speculated store.3328  if (SpeculatedStoreValue) {3329    IRBuilder<NoFolder> Builder(BI);3330    Value *OrigV = SpeculatedStore->getValueOperand();3331    Value *TrueV = SpeculatedStore->getValueOperand();3332    Value *FalseV = SpeculatedStoreValue;3333    if (Invert)3334      std::swap(TrueV, FalseV);3335    Value *S = Builder.CreateSelect(3336        BrCond, TrueV, FalseV, "spec.store.select", BI);3337    Sel = cast<Instruction>(S);3338    SpeculatedStore->setOperand(0, S);3339    SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),3340                                         SpeculatedStore->getDebugLoc());3341    // The value stored is still conditional, but the store itself is now3342    // unconditonally executed, so we must be sure that any linked dbg.assign3343    // intrinsics are tracking the new stored value (the result of the3344    // select). If we don't, and the store were to be removed by another pass3345    // (e.g. DSE), then we'd eventually end up emitting a location describing3346    // the conditional value, unconditionally.3347    //3348    // === Before this transformation ===3349    // pred:3350    //   store %one, %x.dest, !DIAssignID !13351    //   dbg.assign %one, "x", ..., !1, ...3352    //   br %cond if.then3353    //3354    // if.then:3355    //   store %two, %x.dest, !DIAssignID !23356    //   dbg.assign %two, "x", ..., !2, ...3357    //3358    // === After this transformation ===3359    // pred:3360    //   store %one, %x.dest, !DIAssignID !13361    //   dbg.assign %one, "x", ..., !13362    ///  ...3363    //   %merge = select %cond, %two, %one3364    //   store %merge, %x.dest, !DIAssignID !23365    //   dbg.assign %merge, "x", ..., !23366    for (DbgVariableRecord *DbgAssign :3367         at::getDVRAssignmentMarkers(SpeculatedStore))3368      if (llvm::is_contained(DbgAssign->location_ops(), OrigV))3369        DbgAssign->replaceVariableLocationOp(OrigV, S);3370  }3371 3372  // Metadata can be dependent on the condition we are hoisting above.3373  // Strip all UB-implying metadata on the instruction. Drop the debug loc3374  // to avoid making it appear as if the condition is a constant, which would3375  // be misleading while debugging.3376  // Similarly strip attributes that maybe dependent on condition we are3377  // hoisting above.3378  for (auto &I : make_early_inc_range(*ThenBB)) {3379    if (!SpeculatedStoreValue || &I != SpeculatedStore) {3380      I.dropLocation();3381    }3382    I.dropUBImplyingAttrsAndMetadata();3383 3384    // Drop ephemeral values.3385    if (EphTracker.contains(&I)) {3386      I.replaceAllUsesWith(PoisonValue::get(I.getType()));3387      I.eraseFromParent();3388    }3389  }3390 3391  // Hoist the instructions.3392  // Drop DbgVariableRecords attached to these instructions.3393  for (auto &It : *ThenBB)3394    for (DbgRecord &DR : make_early_inc_range(It.getDbgRecordRange()))3395      // Drop all records except assign-kind DbgVariableRecords (dbg.assign3396      // equivalent).3397      if (DbgVariableRecord *DVR = dyn_cast<DbgVariableRecord>(&DR);3398          !DVR || !DVR->isDbgAssign())3399        It.dropOneDbgRecord(&DR);3400  BB->splice(BI->getIterator(), ThenBB, ThenBB->begin(),3401             std::prev(ThenBB->end()));3402 3403  if (!SpeculatedConditionalLoadsStores.empty())3404    hoistConditionalLoadsStores(BI, SpeculatedConditionalLoadsStores, Invert,3405                                Sel);3406 3407  // Insert selects and rewrite the PHI operands.3408  IRBuilder<NoFolder> Builder(BI);3409  for (PHINode &PN : EndBB->phis()) {3410    unsigned OrigI = PN.getBasicBlockIndex(BB);3411    unsigned ThenI = PN.getBasicBlockIndex(ThenBB);3412    Value *OrigV = PN.getIncomingValue(OrigI);3413    Value *ThenV = PN.getIncomingValue(ThenI);3414 3415    // Skip PHIs which are trivial.3416    if (OrigV == ThenV)3417      continue;3418 3419    // Create a select whose true value is the speculatively executed value and3420    // false value is the pre-existing value. Swap them if the branch3421    // destinations were inverted.3422    Value *TrueV = ThenV, *FalseV = OrigV;3423    if (Invert)3424      std::swap(TrueV, FalseV);3425    Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);3426    PN.setIncomingValue(OrigI, V);3427    PN.setIncomingValue(ThenI, V);3428  }3429 3430  // Remove speculated pseudo probes.3431  for (Instruction *I : SpeculatedPseudoProbes)3432    I->eraseFromParent();3433 3434  ++NumSpeculations;3435  return true;3436}3437 3438using BlocksSet = SmallPtrSet<BasicBlock *, 8>;3439 3440// Return false if number of blocks searched is too much.3441static bool findReaching(BasicBlock *BB, BasicBlock *DefBB,3442                         BlocksSet &ReachesNonLocalUses) {3443  if (BB == DefBB)3444    return true;3445  if (!ReachesNonLocalUses.insert(BB).second)3446    return true;3447 3448  if (ReachesNonLocalUses.size() > MaxJumpThreadingLiveBlocks)3449    return false;3450  for (BasicBlock *Pred : predecessors(BB))3451    if (!findReaching(Pred, DefBB, ReachesNonLocalUses))3452      return false;3453  return true;3454}3455 3456/// Return true if we can thread a branch across this block.3457static bool blockIsSimpleEnoughToThreadThrough(BasicBlock *BB,3458                                               BlocksSet &NonLocalUseBlocks) {3459  int Size = 0;3460  EphemeralValueTracker EphTracker;3461 3462  // Walk the loop in reverse so that we can identify ephemeral values properly3463  // (values only feeding assumes).3464  for (Instruction &I : reverse(BB->instructionsWithoutDebug(false))) {3465    // Can't fold blocks that contain noduplicate or convergent calls.3466    if (CallInst *CI = dyn_cast<CallInst>(&I))3467      if (CI->cannotDuplicate() || CI->isConvergent())3468        return false;3469 3470    // Ignore ephemeral values which are deleted during codegen.3471    // We will delete Phis while threading, so Phis should not be accounted in3472    // block's size.3473    if (!EphTracker.track(&I) && !isa<PHINode>(I)) {3474      if (Size++ > MaxSmallBlockSize)3475        return false; // Don't clone large BB's.3476    }3477 3478    // Record blocks with non-local uses of values defined in the current basic3479    // block.3480    for (User *U : I.users()) {3481      Instruction *UI = cast<Instruction>(U);3482      BasicBlock *UsedInBB = UI->getParent();3483      if (UsedInBB == BB) {3484        if (isa<PHINode>(UI))3485          return false;3486      } else3487        NonLocalUseBlocks.insert(UsedInBB);3488    }3489 3490    // Looks ok, continue checking.3491  }3492 3493  return true;3494}3495 3496static ConstantInt *getKnownValueOnEdge(Value *V, BasicBlock *From,3497                                        BasicBlock *To) {3498  // Don't look past the block defining the value, we might get the value from3499  // a previous loop iteration.3500  auto *I = dyn_cast<Instruction>(V);3501  if (I && I->getParent() == To)3502    return nullptr;3503 3504  // We know the value if the From block branches on it.3505  auto *BI = dyn_cast<BranchInst>(From->getTerminator());3506  if (BI && BI->isConditional() && BI->getCondition() == V &&3507      BI->getSuccessor(0) != BI->getSuccessor(1))3508    return BI->getSuccessor(0) == To ? ConstantInt::getTrue(BI->getContext())3509                                     : ConstantInt::getFalse(BI->getContext());3510 3511  return nullptr;3512}3513 3514/// If we have a conditional branch on something for which we know the constant3515/// value in predecessors (e.g. a phi node in the current block), thread edges3516/// from the predecessor to their ultimate destination.3517static std::optional<bool>3518foldCondBranchOnValueKnownInPredecessorImpl(BranchInst *BI, DomTreeUpdater *DTU,3519                                            const DataLayout &DL,3520                                            AssumptionCache *AC) {3521  SmallMapVector<ConstantInt *, SmallSetVector<BasicBlock *, 2>, 2> KnownValues;3522  BasicBlock *BB = BI->getParent();3523  Value *Cond = BI->getCondition();3524  PHINode *PN = dyn_cast<PHINode>(Cond);3525  if (PN && PN->getParent() == BB) {3526    // Degenerate case of a single entry PHI.3527    if (PN->getNumIncomingValues() == 1) {3528      FoldSingleEntryPHINodes(PN->getParent());3529      return true;3530    }3531 3532    for (Use &U : PN->incoming_values())3533      if (auto *CB = dyn_cast<ConstantInt>(U))3534        KnownValues[CB].insert(PN->getIncomingBlock(U));3535  } else {3536    for (BasicBlock *Pred : predecessors(BB)) {3537      if (ConstantInt *CB = getKnownValueOnEdge(Cond, Pred, BB))3538        KnownValues[CB].insert(Pred);3539    }3540  }3541 3542  if (KnownValues.empty())3543    return false;3544 3545  // Now we know that this block has multiple preds and two succs.3546  // Check that the block is small enough and record which non-local blocks use3547  // values defined in the block.3548 3549  BlocksSet NonLocalUseBlocks;3550  BlocksSet ReachesNonLocalUseBlocks;3551  if (!blockIsSimpleEnoughToThreadThrough(BB, NonLocalUseBlocks))3552    return false;3553 3554  // Jump-threading can only be done to destinations where no values defined3555  // in BB are live.3556 3557  // Quickly check if both destinations have uses.  If so, jump-threading cannot3558  // be done.3559  if (NonLocalUseBlocks.contains(BI->getSuccessor(0)) &&3560      NonLocalUseBlocks.contains(BI->getSuccessor(1)))3561    return false;3562 3563  // Search backward from NonLocalUseBlocks to find which blocks3564  // reach non-local uses.3565  for (BasicBlock *UseBB : NonLocalUseBlocks)3566    // Give up if too many blocks are searched.3567    if (!findReaching(UseBB, BB, ReachesNonLocalUseBlocks))3568      return false;3569 3570  for (const auto &Pair : KnownValues) {3571    ConstantInt *CB = Pair.first;3572    ArrayRef<BasicBlock *> PredBBs = Pair.second.getArrayRef();3573    BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());3574 3575    // Okay, we now know that all edges from PredBB should be revectored to3576    // branch to RealDest.3577    if (RealDest == BB)3578      continue; // Skip self loops.3579 3580    // Skip if the predecessor's terminator is an indirect branch.3581    if (any_of(PredBBs, [](BasicBlock *PredBB) {3582          return isa<IndirectBrInst>(PredBB->getTerminator());3583        }))3584      continue;3585 3586    // Only revector to RealDest if no values defined in BB are live.3587    if (ReachesNonLocalUseBlocks.contains(RealDest))3588      continue;3589 3590    LLVM_DEBUG({3591      dbgs() << "Condition " << *Cond << " in " << BB->getName()3592             << " has value " << *Pair.first << " in predecessors:\n";3593      for (const BasicBlock *PredBB : Pair.second)3594        dbgs() << "  " << PredBB->getName() << "\n";3595      dbgs() << "Threading to destination " << RealDest->getName() << ".\n";3596    });3597 3598    // Split the predecessors we are threading into a new edge block. We'll3599    // clone the instructions into this block, and then redirect it to RealDest.3600    BasicBlock *EdgeBB = SplitBlockPredecessors(BB, PredBBs, ".critedge", DTU);3601 3602    // TODO: These just exist to reduce test diff, we can drop them if we like.3603    EdgeBB->setName(RealDest->getName() + ".critedge");3604    EdgeBB->moveBefore(RealDest);3605 3606    // Update PHI nodes.3607    addPredecessorToBlock(RealDest, EdgeBB, BB);3608 3609    // BB may have instructions that are being threaded over.  Clone these3610    // instructions into EdgeBB.  We know that there will be no uses of the3611    // cloned instructions outside of EdgeBB.3612    BasicBlock::iterator InsertPt = EdgeBB->getFirstInsertionPt();3613    ValueToValueMapTy TranslateMap; // Track translated values.3614    TranslateMap[Cond] = CB;3615 3616    // RemoveDIs: track instructions that we optimise away while folding, so3617    // that we can copy DbgVariableRecords from them later.3618    BasicBlock::iterator SrcDbgCursor = BB->begin();3619    for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {3620      if (PHINode *PN = dyn_cast<PHINode>(BBI)) {3621        TranslateMap[PN] = PN->getIncomingValueForBlock(EdgeBB);3622        continue;3623      }3624      // Clone the instruction.3625      Instruction *N = BBI->clone();3626      // Insert the new instruction into its new home.3627      N->insertInto(EdgeBB, InsertPt);3628 3629      if (BBI->hasName())3630        N->setName(BBI->getName() + ".c");3631 3632      // Update operands due to translation.3633      // Key Instructions: Remap all the atom groups.3634      if (const DebugLoc &DL = BBI->getDebugLoc())3635        mapAtomInstance(DL, TranslateMap);3636      RemapInstruction(N, TranslateMap,3637                       RF_IgnoreMissingLocals | RF_NoModuleLevelChanges);3638 3639      // Check for trivial simplification.3640      if (Value *V = simplifyInstruction(N, {DL, nullptr, nullptr, AC})) {3641        if (!BBI->use_empty())3642          TranslateMap[&*BBI] = V;3643        if (!N->mayHaveSideEffects()) {3644          N->eraseFromParent(); // Instruction folded away, don't need actual3645                                // inst3646          N = nullptr;3647        }3648      } else {3649        if (!BBI->use_empty())3650          TranslateMap[&*BBI] = N;3651      }3652      if (N) {3653        // Copy all debug-info attached to instructions from the last we3654        // successfully clone, up to this instruction (they might have been3655        // folded away).3656        for (; SrcDbgCursor != BBI; ++SrcDbgCursor)3657          N->cloneDebugInfoFrom(&*SrcDbgCursor);3658        SrcDbgCursor = std::next(BBI);3659        // Clone debug-info on this instruction too.3660        N->cloneDebugInfoFrom(&*BBI);3661 3662        // Register the new instruction with the assumption cache if necessary.3663        if (auto *Assume = dyn_cast<AssumeInst>(N))3664          if (AC)3665            AC->registerAssumption(Assume);3666      }3667    }3668 3669    for (; &*SrcDbgCursor != BI; ++SrcDbgCursor)3670      InsertPt->cloneDebugInfoFrom(&*SrcDbgCursor);3671    InsertPt->cloneDebugInfoFrom(BI);3672 3673    BB->removePredecessor(EdgeBB);3674    BranchInst *EdgeBI = cast<BranchInst>(EdgeBB->getTerminator());3675    EdgeBI->setSuccessor(0, RealDest);3676    EdgeBI->setDebugLoc(BI->getDebugLoc());3677 3678    if (DTU) {3679      SmallVector<DominatorTree::UpdateType, 2> Updates;3680      Updates.push_back({DominatorTree::Delete, EdgeBB, BB});3681      Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});3682      DTU->applyUpdates(Updates);3683    }3684 3685    // For simplicity, we created a separate basic block for the edge. Merge3686    // it back into the predecessor if possible. This not only avoids3687    // unnecessary SimplifyCFG iterations, but also makes sure that we don't3688    // bypass the check for trivial cycles above.3689    MergeBlockIntoPredecessor(EdgeBB, DTU);3690 3691    // Signal repeat, simplifying any other constants.3692    return std::nullopt;3693  }3694 3695  return false;3696}3697 3698bool SimplifyCFGOpt::foldCondBranchOnValueKnownInPredecessor(BranchInst *BI) {3699  // Note: If BB is a loop header then there is a risk that threading introduces3700  // a non-canonical loop by moving a back edge. So we avoid this optimization3701  // for loop headers if NeedCanonicalLoop is set.3702  if (Options.NeedCanonicalLoop && is_contained(LoopHeaders, BI->getParent()))3703    return false;3704 3705  std::optional<bool> Result;3706  bool EverChanged = false;3707  do {3708    // Note that None means "we changed things, but recurse further."3709    Result =3710        foldCondBranchOnValueKnownInPredecessorImpl(BI, DTU, DL, Options.AC);3711    EverChanged |= Result == std::nullopt || *Result;3712  } while (Result == std::nullopt);3713  return EverChanged;3714}3715 3716/// Given a BB that starts with the specified two-entry PHI node,3717/// see if we can eliminate it.3718static bool foldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,3719                                DomTreeUpdater *DTU, AssumptionCache *AC,3720                                const DataLayout &DL,3721                                bool SpeculateUnpredictables) {3722  // Ok, this is a two entry PHI node.  Check to see if this is a simple "if3723  // statement", which has a very simple dominance structure.  Basically, we3724  // are trying to find the condition that is being branched on, which3725  // subsequently causes this merge to happen.  We really want control3726  // dependence information for this check, but simplifycfg can't keep it up3727  // to date, and this catches most of the cases we care about anyway.3728  BasicBlock *BB = PN->getParent();3729 3730  BasicBlock *IfTrue, *IfFalse;3731  BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse);3732  if (!DomBI)3733    return false;3734  Value *IfCond = DomBI->getCondition();3735  // Don't bother if the branch will be constant folded trivially.3736  if (isa<ConstantInt>(IfCond))3737    return false;3738 3739  BasicBlock *DomBlock = DomBI->getParent();3740  SmallVector<BasicBlock *, 2> IfBlocks;3741  llvm::copy_if(3742      PN->blocks(), std::back_inserter(IfBlocks), [](BasicBlock *IfBlock) {3743        return cast<BranchInst>(IfBlock->getTerminator())->isUnconditional();3744      });3745  assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) &&3746         "Will have either one or two blocks to speculate.");3747 3748  // If the branch is non-unpredictable, see if we either predictably jump to3749  // the merge bb (if we have only a single 'then' block), or if we predictably3750  // jump to one specific 'then' block (if we have two of them).3751  // It isn't beneficial to speculatively execute the code3752  // from the block that we know is predictably not entered.3753  bool IsUnpredictable = DomBI->getMetadata(LLVMContext::MD_unpredictable);3754  if (!IsUnpredictable) {3755    uint64_t TWeight, FWeight;3756    if (extractBranchWeights(*DomBI, TWeight, FWeight) &&3757        (TWeight + FWeight) != 0) {3758      BranchProbability BITrueProb =3759          BranchProbability::getBranchProbability(TWeight, TWeight + FWeight);3760      BranchProbability Likely = TTI.getPredictableBranchThreshold();3761      BranchProbability BIFalseProb = BITrueProb.getCompl();3762      if (IfBlocks.size() == 1) {3763        BranchProbability BIBBProb =3764            DomBI->getSuccessor(0) == BB ? BITrueProb : BIFalseProb;3765        if (BIBBProb >= Likely)3766          return false;3767      } else {3768        if (BITrueProb >= Likely || BIFalseProb >= Likely)3769          return false;3770      }3771    }3772  }3773 3774  // Don't try to fold an unreachable block. For example, the phi node itself3775  // can't be the candidate if-condition for a select that we want to form.3776  if (auto *IfCondPhiInst = dyn_cast<PHINode>(IfCond))3777    if (IfCondPhiInst->getParent() == BB)3778      return false;3779 3780  // Okay, we found that we can merge this two-entry phi node into a select.3781  // Doing so would require us to fold *all* two entry phi nodes in this block.3782  // At some point this becomes non-profitable (particularly if the target3783  // doesn't support cmov's).  Only do this transformation if there are two or3784  // fewer PHI nodes in this block.3785  unsigned NumPhis = 0;3786  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)3787    if (NumPhis > 2)3788      return false;3789 3790  // Loop over the PHI's seeing if we can promote them all to select3791  // instructions.  While we are at it, keep track of the instructions3792  // that need to be moved to the dominating block.3793  SmallPtrSet<Instruction *, 4> AggressiveInsts;3794  SmallPtrSet<Instruction *, 2> ZeroCostInstructions;3795  InstructionCost Cost = 0;3796  InstructionCost Budget =3797      TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;3798  if (SpeculateUnpredictables && IsUnpredictable)3799    Budget += TTI.getBranchMispredictPenalty();3800 3801  bool Changed = false;3802  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {3803    PHINode *PN = cast<PHINode>(II++);3804    if (Value *V = simplifyInstruction(PN, {DL, PN})) {3805      PN->replaceAllUsesWith(V);3806      PN->eraseFromParent();3807      Changed = true;3808      continue;3809    }3810 3811    if (!dominatesMergePoint(PN->getIncomingValue(0), BB, DomBI,3812                             AggressiveInsts, Cost, Budget, TTI, AC,3813                             ZeroCostInstructions) ||3814        !dominatesMergePoint(PN->getIncomingValue(1), BB, DomBI,3815                             AggressiveInsts, Cost, Budget, TTI, AC,3816                             ZeroCostInstructions))3817      return Changed;3818  }3819 3820  // If we folded the first phi, PN dangles at this point.  Refresh it.  If3821  // we ran out of PHIs then we simplified them all.3822  PN = dyn_cast<PHINode>(BB->begin());3823  if (!PN)3824    return true;3825 3826  // Return true if at least one of these is a 'not', and another is either3827  // a 'not' too, or a constant.3828  auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {3829    if (!match(V0, m_Not(m_Value())))3830      std::swap(V0, V1);3831    auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());3832    return match(V0, m_Not(m_Value())) && match(V1, Invertible);3833  };3834 3835  // Don't fold i1 branches on PHIs which contain binary operators or3836  // (possibly inverted) select form of or/ands,  unless one of3837  // the incoming values is an 'not' and another one is freely invertible.3838  // These can often be turned into switches and other things.3839  auto IsBinOpOrAnd = [](Value *V) {3840    return match(3841        V, m_CombineOr(m_BinOp(), m_c_Select(m_ImmConstant(), m_Value())));3842  };3843  if (PN->getType()->isIntegerTy(1) &&3844      (IsBinOpOrAnd(PN->getIncomingValue(0)) ||3845       IsBinOpOrAnd(PN->getIncomingValue(1)) || IsBinOpOrAnd(IfCond)) &&3846      !CanHoistNotFromBothValues(PN->getIncomingValue(0),3847                                 PN->getIncomingValue(1)))3848    return Changed;3849 3850  // If all PHI nodes are promotable, check to make sure that all instructions3851  // in the predecessor blocks can be promoted as well. If not, we won't be able3852  // to get rid of the control flow, so it's not worth promoting to select3853  // instructions.3854  for (BasicBlock *IfBlock : IfBlocks)3855    for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I)3856      if (!AggressiveInsts.count(&*I) && !I->isDebugOrPseudoInst()) {3857        // This is not an aggressive instruction that we can promote.3858        // Because of this, we won't be able to get rid of the control flow, so3859        // the xform is not worth it.3860        return Changed;3861      }3862 3863  // If either of the blocks has it's address taken, we can't do this fold.3864  if (any_of(IfBlocks,3865             [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); }))3866    return Changed;3867 3868  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION!  " << *IfCond;3869             if (IsUnpredictable) dbgs() << " (unpredictable)";3870             dbgs() << "  T: " << IfTrue->getName()3871                    << "  F: " << IfFalse->getName() << "\n");3872 3873  // If we can still promote the PHI nodes after this gauntlet of tests,3874  // do all of the PHI's now.3875 3876  // Move all 'aggressive' instructions, which are defined in the3877  // conditional parts of the if's up to the dominating block.3878  for (BasicBlock *IfBlock : IfBlocks)3879      hoistAllInstructionsInto(DomBlock, DomBI, IfBlock);3880 3881  IRBuilder<NoFolder> Builder(DomBI);3882  // Propagate fast-math-flags from phi nodes to replacement selects.3883  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {3884    // Change the PHI node into a select instruction.3885    Value *TrueVal = PN->getIncomingValueForBlock(IfTrue);3886    Value *FalseVal = PN->getIncomingValueForBlock(IfFalse);3887 3888    Value *Sel = Builder.CreateSelectFMF(IfCond, TrueVal, FalseVal,3889                                         isa<FPMathOperator>(PN) ? PN : nullptr,3890                                         "", DomBI);3891    PN->replaceAllUsesWith(Sel);3892    Sel->takeName(PN);3893    PN->eraseFromParent();3894  }3895 3896  // At this point, all IfBlocks are empty, so our if statement3897  // has been flattened.  Change DomBlock to jump directly to our new block to3898  // avoid other simplifycfg's kicking in on the diamond.3899  Builder.CreateBr(BB);3900 3901  SmallVector<DominatorTree::UpdateType, 3> Updates;3902  if (DTU) {3903    Updates.push_back({DominatorTree::Insert, DomBlock, BB});3904    for (auto *Successor : successors(DomBlock))3905      Updates.push_back({DominatorTree::Delete, DomBlock, Successor});3906  }3907 3908  DomBI->eraseFromParent();3909  if (DTU)3910    DTU->applyUpdates(Updates);3911 3912  return true;3913}3914 3915static Value *createLogicalOp(IRBuilderBase &Builder,3916                              Instruction::BinaryOps Opc, Value *LHS,3917                              Value *RHS, const Twine &Name = "") {3918  // Try to relax logical op to binary op.3919  if (impliesPoison(RHS, LHS))3920    return Builder.CreateBinOp(Opc, LHS, RHS, Name);3921  if (Opc == Instruction::And)3922    return Builder.CreateLogicalAnd(LHS, RHS, Name);3923  if (Opc == Instruction::Or)3924    return Builder.CreateLogicalOr(LHS, RHS, Name);3925  llvm_unreachable("Invalid logical opcode");3926}3927 3928/// Return true if either PBI or BI has branch weight available, and store3929/// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does3930/// not have branch weight, use 1:1 as its weight.3931static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,3932                                   uint64_t &PredTrueWeight,3933                                   uint64_t &PredFalseWeight,3934                                   uint64_t &SuccTrueWeight,3935                                   uint64_t &SuccFalseWeight) {3936  bool PredHasWeights =3937      extractBranchWeights(*PBI, PredTrueWeight, PredFalseWeight);3938  bool SuccHasWeights =3939      extractBranchWeights(*BI, SuccTrueWeight, SuccFalseWeight);3940  if (PredHasWeights || SuccHasWeights) {3941    if (!PredHasWeights)3942      PredTrueWeight = PredFalseWeight = 1;3943    if (!SuccHasWeights)3944      SuccTrueWeight = SuccFalseWeight = 1;3945    return true;3946  } else {3947    return false;3948  }3949}3950 3951/// Determine if the two branches share a common destination and deduce a glue3952/// that joins the branches' conditions to arrive at the common destination if3953/// that would be profitable.3954static std::optional<std::tuple<BasicBlock *, Instruction::BinaryOps, bool>>3955shouldFoldCondBranchesToCommonDestination(BranchInst *BI, BranchInst *PBI,3956                                          const TargetTransformInfo *TTI) {3957  assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&3958         "Both blocks must end with a conditional branches.");3959  assert(is_contained(predecessors(BI->getParent()), PBI->getParent()) &&3960         "PredBB must be a predecessor of BB.");3961 3962  // We have the potential to fold the conditions together, but if the3963  // predecessor branch is predictable, we may not want to merge them.3964  uint64_t PTWeight, PFWeight;3965  BranchProbability PBITrueProb, Likely;3966  if (TTI && !PBI->getMetadata(LLVMContext::MD_unpredictable) &&3967      extractBranchWeights(*PBI, PTWeight, PFWeight) &&3968      (PTWeight + PFWeight) != 0) {3969    PBITrueProb =3970        BranchProbability::getBranchProbability(PTWeight, PTWeight + PFWeight);3971    Likely = TTI->getPredictableBranchThreshold();3972  }3973 3974  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {3975    // Speculate the 2nd condition unless the 1st is probably true.3976    if (PBITrueProb.isUnknown() || PBITrueProb < Likely)3977      return {{BI->getSuccessor(0), Instruction::Or, false}};3978  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {3979    // Speculate the 2nd condition unless the 1st is probably false.3980    if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)3981      return {{BI->getSuccessor(1), Instruction::And, false}};3982  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {3983    // Speculate the 2nd condition unless the 1st is probably true.3984    if (PBITrueProb.isUnknown() || PBITrueProb < Likely)3985      return {{BI->getSuccessor(1), Instruction::And, true}};3986  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {3987    // Speculate the 2nd condition unless the 1st is probably false.3988    if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely)3989      return {{BI->getSuccessor(0), Instruction::Or, true}};3990  }3991  return std::nullopt;3992}3993 3994static bool performBranchToCommonDestFolding(BranchInst *BI, BranchInst *PBI,3995                                             DomTreeUpdater *DTU,3996                                             MemorySSAUpdater *MSSAU,3997                                             const TargetTransformInfo *TTI) {3998  BasicBlock *BB = BI->getParent();3999  BasicBlock *PredBlock = PBI->getParent();4000 4001  // Determine if the two branches share a common destination.4002  BasicBlock *CommonSucc;4003  Instruction::BinaryOps Opc;4004  bool InvertPredCond;4005  std::tie(CommonSucc, Opc, InvertPredCond) =4006      *shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI);4007 4008  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);4009 4010  IRBuilder<> Builder(PBI);4011  // The builder is used to create instructions to eliminate the branch in BB.4012  // If BB's terminator has !annotation metadata, add it to the new4013  // instructions.4014  Builder.CollectMetadataToCopy(BB->getTerminator(),4015                                {LLVMContext::MD_annotation});4016 4017  // If we need to invert the condition in the pred block to match, do so now.4018  if (InvertPredCond) {4019    InvertBranch(PBI, Builder);4020  }4021 4022  BasicBlock *UniqueSucc =4023      PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);4024 4025  // Before cloning instructions, notify the successor basic block that it4026  // is about to have a new predecessor. This will update PHI nodes,4027  // which will allow us to update live-out uses of bonus instructions.4028  addPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);4029 4030  // Try to update branch weights.4031  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;4032  SmallVector<uint64_t, 2> MDWeights;4033  if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,4034                             SuccTrueWeight, SuccFalseWeight)) {4035 4036    if (PBI->getSuccessor(0) == BB) {4037      // PBI: br i1 %x, BB, FalseDest4038      // BI:  br i1 %y, UniqueSucc, FalseDest4039      // TrueWeight is TrueWeight for PBI * TrueWeight for BI.4040      MDWeights.push_back(PredTrueWeight * SuccTrueWeight);4041      // FalseWeight is FalseWeight for PBI * TotalWeight for BI +4042      //               TrueWeight for PBI * FalseWeight for BI.4043      // We assume that total weights of a BranchInst can fit into 32 bits.4044      // Therefore, we will not have overflow using 64-bit arithmetic.4045      MDWeights.push_back(PredFalseWeight * (SuccFalseWeight + SuccTrueWeight) +4046                          PredTrueWeight * SuccFalseWeight);4047    } else {4048      // PBI: br i1 %x, TrueDest, BB4049      // BI:  br i1 %y, TrueDest, UniqueSucc4050      // TrueWeight is TrueWeight for PBI * TotalWeight for BI +4051      //              FalseWeight for PBI * TrueWeight for BI.4052      MDWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +4053                          PredFalseWeight * SuccTrueWeight);4054      // FalseWeight is FalseWeight for PBI * FalseWeight for BI.4055      MDWeights.push_back(PredFalseWeight * SuccFalseWeight);4056    }4057 4058    setFittedBranchWeights(*PBI, MDWeights, /*IsExpected=*/false,4059                           /*ElideAllZero=*/true);4060 4061    // TODO: If BB is reachable from all paths through PredBlock, then we4062    // could replace PBI's branch probabilities with BI's.4063  } else4064    PBI->setMetadata(LLVMContext::MD_prof, nullptr);4065 4066  // Now, update the CFG.4067  PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);4068 4069  if (DTU)4070    DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},4071                       {DominatorTree::Delete, PredBlock, BB}});4072 4073  // If BI was a loop latch, it may have had associated loop metadata.4074  // We need to copy it to the new latch, that is, PBI.4075  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))4076    PBI->setMetadata(LLVMContext::MD_loop, LoopMD);4077 4078  ValueToValueMapTy VMap; // maps original values to cloned values4079  cloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BB, PredBlock, VMap);4080 4081  Module *M = BB->getModule();4082 4083  PredBlock->getTerminator()->cloneDebugInfoFrom(BB->getTerminator());4084  for (DbgVariableRecord &DVR :4085       filterDbgVars(PredBlock->getTerminator()->getDbgRecordRange())) {4086    RemapDbgRecord(M, &DVR, VMap,4087                   RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);4088  }4089 4090  // Now that the Cond was cloned into the predecessor basic block,4091  // or/and the two conditions together.4092  Value *BICond = VMap[BI->getCondition()];4093  PBI->setCondition(4094      createLogicalOp(Builder, Opc, PBI->getCondition(), BICond, "or.cond"));4095  if (!ProfcheckDisableMetadataFixes)4096    if (auto *SI = dyn_cast<SelectInst>(PBI->getCondition()))4097      if (!MDWeights.empty()) {4098        assert(isSelectInRoleOfConjunctionOrDisjunction(SI));4099        setFittedBranchWeights(*SI, {MDWeights[0], MDWeights[1]},4100                               /*IsExpected=*/false, /*ElideAllZero=*/true);4101      }4102 4103  ++NumFoldBranchToCommonDest;4104  return true;4105}4106 4107/// Return if an instruction's type or any of its operands' types are a vector4108/// type.4109static bool isVectorOp(Instruction &I) {4110  return I.getType()->isVectorTy() || any_of(I.operands(), [](Use &U) {4111           return U->getType()->isVectorTy();4112         });4113}4114 4115/// If this basic block is simple enough, and if a predecessor branches to us4116/// and one of our successors, fold the block into the predecessor and use4117/// logical operations to pick the right destination.4118bool llvm::foldBranchToCommonDest(BranchInst *BI, DomTreeUpdater *DTU,4119                                  MemorySSAUpdater *MSSAU,4120                                  const TargetTransformInfo *TTI,4121                                  unsigned BonusInstThreshold) {4122  // If this block ends with an unconditional branch,4123  // let speculativelyExecuteBB() deal with it.4124  if (!BI->isConditional())4125    return false;4126 4127  BasicBlock *BB = BI->getParent();4128  TargetTransformInfo::TargetCostKind CostKind =4129    BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize4130                                  : TargetTransformInfo::TCK_SizeAndLatency;4131 4132  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());4133 4134  if (!Cond || !isa<CmpInst, BinaryOperator, SelectInst, TruncInst>(Cond) ||4135      Cond->getParent() != BB || !Cond->hasOneUse())4136    return false;4137 4138  // Finally, don't infinitely unroll conditional loops.4139  if (is_contained(successors(BB), BB))4140    return false;4141 4142  // With which predecessors will we want to deal with?4143  SmallVector<BasicBlock *, 8> Preds;4144  for (BasicBlock *PredBlock : predecessors(BB)) {4145    BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());4146 4147    // Check that we have two conditional branches.  If there is a PHI node in4148    // the common successor, verify that the same value flows in from both4149    // blocks.4150    if (!PBI || PBI->isUnconditional() || !safeToMergeTerminators(BI, PBI))4151      continue;4152 4153    // Determine if the two branches share a common destination.4154    BasicBlock *CommonSucc;4155    Instruction::BinaryOps Opc;4156    bool InvertPredCond;4157    if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI))4158      std::tie(CommonSucc, Opc, InvertPredCond) = *Recipe;4159    else4160      continue;4161 4162    // Check the cost of inserting the necessary logic before performing the4163    // transformation.4164    if (TTI) {4165      Type *Ty = BI->getCondition()->getType();4166      InstructionCost Cost = TTI->getArithmeticInstrCost(Opc, Ty, CostKind);4167      if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||4168                             !isa<CmpInst>(PBI->getCondition())))4169        Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);4170 4171      if (Cost > BranchFoldThreshold)4172        continue;4173    }4174 4175    // Ok, we do want to deal with this predecessor. Record it.4176    Preds.emplace_back(PredBlock);4177  }4178 4179  // If there aren't any predecessors into which we can fold,4180  // don't bother checking the cost.4181  if (Preds.empty())4182    return false;4183 4184  // Only allow this transformation if computing the condition doesn't involve4185  // too many instructions and these involved instructions can be executed4186  // unconditionally. We denote all involved instructions except the condition4187  // as "bonus instructions", and only allow this transformation when the4188  // number of the bonus instructions we'll need to create when cloning into4189  // each predecessor does not exceed a certain threshold.4190  unsigned NumBonusInsts = 0;4191  bool SawVectorOp = false;4192  const unsigned PredCount = Preds.size();4193  for (Instruction &I : *BB) {4194    // Don't check the branch condition comparison itself.4195    if (&I == Cond)4196      continue;4197    // Ignore the terminator.4198    if (isa<BranchInst>(I))4199      continue;4200    // I must be safe to execute unconditionally.4201    if (!isSafeToSpeculativelyExecute(&I))4202      return false;4203    SawVectorOp |= isVectorOp(I);4204 4205    // Account for the cost of duplicating this instruction into each4206    // predecessor. Ignore free instructions.4207    if (!TTI || TTI->getInstructionCost(&I, CostKind) !=4208                    TargetTransformInfo::TCC_Free) {4209      NumBonusInsts += PredCount;4210 4211      // Early exits once we reach the limit.4212      if (NumBonusInsts >4213          BonusInstThreshold * BranchFoldToCommonDestVectorMultiplier)4214        return false;4215    }4216 4217    auto IsBCSSAUse = [BB, &I](Use &U) {4218      auto *UI = cast<Instruction>(U.getUser());4219      if (auto *PN = dyn_cast<PHINode>(UI))4220        return PN->getIncomingBlock(U) == BB;4221      return UI->getParent() == BB && I.comesBefore(UI);4222    };4223 4224    // Does this instruction require rewriting of uses?4225    if (!all_of(I.uses(), IsBCSSAUse))4226      return false;4227  }4228  if (NumBonusInsts >4229      BonusInstThreshold *4230          (SawVectorOp ? BranchFoldToCommonDestVectorMultiplier : 1))4231    return false;4232 4233  // Ok, we have the budget. Perform the transformation.4234  for (BasicBlock *PredBlock : Preds) {4235    auto *PBI = cast<BranchInst>(PredBlock->getTerminator());4236    return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI);4237  }4238  return false;4239}4240 4241// If there is only one store in BB1 and BB2, return it, otherwise return4242// nullptr.4243static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {4244  StoreInst *S = nullptr;4245  for (auto *BB : {BB1, BB2}) {4246    if (!BB)4247      continue;4248    for (auto &I : *BB)4249      if (auto *SI = dyn_cast<StoreInst>(&I)) {4250        if (S)4251          // Multiple stores seen.4252          return nullptr;4253        else4254          S = SI;4255      }4256  }4257  return S;4258}4259 4260static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,4261                                              Value *AlternativeV = nullptr) {4262  // PHI is going to be a PHI node that allows the value V that is defined in4263  // BB to be referenced in BB's only successor.4264  //4265  // If AlternativeV is nullptr, the only value we care about in PHI is V. It4266  // doesn't matter to us what the other operand is (it'll never get used). We4267  // could just create a new PHI with an undef incoming value, but that could4268  // increase register pressure if EarlyCSE/InstCombine can't fold it with some4269  // other PHI. So here we directly look for some PHI in BB's successor with V4270  // as an incoming operand. If we find one, we use it, else we create a new4271  // one.4272  //4273  // If AlternativeV is not nullptr, we care about both incoming values in PHI.4274  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]4275  // where OtherBB is the single other predecessor of BB's only successor.4276  PHINode *PHI = nullptr;4277  BasicBlock *Succ = BB->getSingleSuccessor();4278 4279  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)4280    if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {4281      PHI = cast<PHINode>(I);4282      if (!AlternativeV)4283        break;4284 4285      assert(Succ->hasNPredecessors(2));4286      auto PredI = pred_begin(Succ);4287      BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;4288      if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)4289        break;4290      PHI = nullptr;4291    }4292  if (PHI)4293    return PHI;4294 4295  // If V is not an instruction defined in BB, just return it.4296  if (!AlternativeV &&4297      (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))4298    return V;4299 4300  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge");4301  PHI->insertBefore(Succ->begin());4302  PHI->addIncoming(V, BB);4303  for (BasicBlock *PredBB : predecessors(Succ))4304    if (PredBB != BB)4305      PHI->addIncoming(4306          AlternativeV ? AlternativeV : PoisonValue::get(V->getType()), PredBB);4307  return PHI;4308}4309 4310static bool mergeConditionalStoreToAddress(4311    BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,4312    BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,4313    DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {4314  // For every pointer, there must be exactly two stores, one coming from4315  // PTB or PFB, and the other from QTB or QFB. We don't support more than one4316  // store (to any address) in PTB,PFB or QTB,QFB.4317  // FIXME: We could relax this restriction with a bit more work and performance4318  // testing.4319  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);4320  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);4321  if (!PStore || !QStore)4322    return false;4323 4324  // Now check the stores are compatible.4325  if (!QStore->isUnordered() || !PStore->isUnordered() ||4326      PStore->getValueOperand()->getType() !=4327          QStore->getValueOperand()->getType())4328    return false;4329 4330  // Check that sinking the store won't cause program behavior changes. Sinking4331  // the store out of the Q blocks won't change any behavior as we're sinking4332  // from a block to its unconditional successor. But we're moving a store from4333  // the P blocks down through the middle block (QBI) and past both QFB and QTB.4334  // So we need to check that there are no aliasing loads or stores in4335  // QBI, QTB and QFB. We also need to check there are no conflicting memory4336  // operations between PStore and the end of its parent block.4337  //4338  // The ideal way to do this is to query AliasAnalysis, but we don't4339  // preserve AA currently so that is dangerous. Be super safe and just4340  // check there are no other memory operations at all.4341  for (auto &I : *QFB->getSinglePredecessor())4342    if (I.mayReadOrWriteMemory())4343      return false;4344  for (auto &I : *QFB)4345    if (&I != QStore && I.mayReadOrWriteMemory())4346      return false;4347  if (QTB)4348    for (auto &I : *QTB)4349      if (&I != QStore && I.mayReadOrWriteMemory())4350        return false;4351  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();4352       I != E; ++I)4353    if (&*I != PStore && I->mayReadOrWriteMemory())4354      return false;4355 4356  // If we're not in aggressive mode, we only optimize if we have some4357  // confidence that by optimizing we'll allow P and/or Q to be if-converted.4358  auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {4359    if (!BB)4360      return true;4361    // Heuristic: if the block can be if-converted/phi-folded and the4362    // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to4363    // thread this store.4364    InstructionCost Cost = 0;4365    InstructionCost Budget =4366        PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;4367    for (auto &I : BB->instructionsWithoutDebug(false)) {4368      // Consider terminator instruction to be free.4369      if (I.isTerminator())4370        continue;4371      // If this is one the stores that we want to speculate out of this BB,4372      // then don't count it's cost, consider it to be free.4373      if (auto *S = dyn_cast<StoreInst>(&I))4374        if (llvm::find(FreeStores, S))4375          continue;4376      // Else, we have a white-list of instructions that we are ak speculating.4377      if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))4378        return false; // Not in white-list - not worthwhile folding.4379      // And finally, if this is a non-free instruction that we are okay4380      // speculating, ensure that we consider the speculation budget.4381      Cost +=4382          TTI.getInstructionCost(&I, TargetTransformInfo::TCK_SizeAndLatency);4383      if (Cost > Budget)4384        return false; // Eagerly refuse to fold as soon as we're out of budget.4385    }4386    assert(Cost <= Budget &&4387           "When we run out of budget we will eagerly return from within the "4388           "per-instruction loop.");4389    return true;4390  };4391 4392  const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};4393  if (!MergeCondStoresAggressively &&4394      (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||4395       !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))4396    return false;4397 4398  // If PostBB has more than two predecessors, we need to split it so we can4399  // sink the store.4400  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {4401    // We know that QFB's only successor is PostBB. And QFB has a single4402    // predecessor. If QTB exists, then its only successor is also PostBB.4403    // If QTB does not exist, then QFB's only predecessor has a conditional4404    // branch to QFB and PostBB.4405    BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();4406    BasicBlock *NewBB =4407        SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);4408    if (!NewBB)4409      return false;4410    PostBB = NewBB;4411  }4412 4413  // OK, we're going to sink the stores to PostBB. The store has to be4414  // conditional though, so first create the predicate.4415  BranchInst *PBranch =4416      cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator());4417  BranchInst *QBranch =4418      cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator());4419  Value *PCond = PBranch->getCondition();4420  Value *QCond = QBranch->getCondition();4421 4422  Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),4423                                                PStore->getParent());4424  Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),4425                                                QStore->getParent(), PPHI);4426 4427  BasicBlock::iterator PostBBFirst = PostBB->getFirstInsertionPt();4428  IRBuilder<> QB(PostBB, PostBBFirst);4429  QB.SetCurrentDebugLocation(PostBBFirst->getStableDebugLoc());4430 4431  InvertPCond ^= (PStore->getParent() != PTB);4432  InvertQCond ^= (QStore->getParent() != QTB);4433  Value *PPred = InvertPCond ? QB.CreateNot(PCond) : PCond;4434  Value *QPred = InvertQCond ? QB.CreateNot(QCond) : QCond;4435 4436  Value *CombinedPred = QB.CreateOr(PPred, QPred);4437 4438  BasicBlock::iterator InsertPt = QB.GetInsertPoint();4439  auto *T = SplitBlockAndInsertIfThen(CombinedPred, InsertPt,4440                                      /*Unreachable=*/false,4441                                      /*BranchWeights=*/nullptr, DTU);4442  if (hasBranchWeightMD(*PBranch) && hasBranchWeightMD(*QBranch) &&4443      !ProfcheckDisableMetadataFixes) {4444    SmallVector<uint32_t, 2> PWeights, QWeights;4445    extractBranchWeights(*PBranch, PWeights);4446    extractBranchWeights(*QBranch, QWeights);4447    if (InvertPCond)4448      std::swap(PWeights[0], PWeights[1]);4449    if (InvertQCond)4450      std::swap(QWeights[0], QWeights[1]);4451    auto CombinedWeights = getDisjunctionWeights(PWeights, QWeights);4452    setFittedBranchWeights(*PostBB->getTerminator(),4453                           {CombinedWeights[0], CombinedWeights[1]},4454                           /*IsExpected=*/false, /*ElideAllZero=*/true);4455  }4456 4457  QB.SetInsertPoint(T);4458  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));4459  SI->setAAMetadata(PStore->getAAMetadata().merge(QStore->getAAMetadata()));4460  // Choose the minimum alignment. If we could prove both stores execute, we4461  // could use biggest one.  In this case, though, we only know that one of the4462  // stores executes.  And we don't know it's safe to take the alignment from a4463  // store that doesn't execute.4464  SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));4465 4466  QStore->eraseFromParent();4467  PStore->eraseFromParent();4468 4469  return true;4470}4471 4472static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,4473                                   DomTreeUpdater *DTU, const DataLayout &DL,4474                                   const TargetTransformInfo &TTI) {4475  // The intention here is to find diamonds or triangles (see below) where each4476  // conditional block contains a store to the same address. Both of these4477  // stores are conditional, so they can't be unconditionally sunk. But it may4478  // be profitable to speculatively sink the stores into one merged store at the4479  // end, and predicate the merged store on the union of the two conditions of4480  // PBI and QBI.4481  //4482  // This can reduce the number of stores executed if both of the conditions are4483  // true, and can allow the blocks to become small enough to be if-converted.4484  // This optimization will also chain, so that ladders of test-and-set4485  // sequences can be if-converted away.4486  //4487  // We only deal with simple diamonds or triangles:4488  //4489  //     PBI       or      PBI        or a combination of the two4490  //    /   \               | \4491  //   PTB  PFB             |  PFB4492  //    \   /               | /4493  //     QBI                QBI4494  //    /  \                | \4495  //   QTB  QFB             |  QFB4496  //    \  /                | /4497  //    PostBB            PostBB4498  //4499  // We model triangles as a type of diamond with a nullptr "true" block.4500  // Triangles are canonicalized so that the fallthrough edge is represented by4501  // a true condition, as in the diagram above.4502  BasicBlock *PTB = PBI->getSuccessor(0);4503  BasicBlock *PFB = PBI->getSuccessor(1);4504  BasicBlock *QTB = QBI->getSuccessor(0);4505  BasicBlock *QFB = QBI->getSuccessor(1);4506  BasicBlock *PostBB = QFB->getSingleSuccessor();4507 4508  // Make sure we have a good guess for PostBB. If QTB's only successor is4509  // QFB, then QFB is a better PostBB.4510  if (QTB->getSingleSuccessor() == QFB)4511    PostBB = QFB;4512 4513  // If we couldn't find a good PostBB, stop.4514  if (!PostBB)4515    return false;4516 4517  bool InvertPCond = false, InvertQCond = false;4518  // Canonicalize fallthroughs to the true branches.4519  if (PFB == QBI->getParent()) {4520    std::swap(PFB, PTB);4521    InvertPCond = true;4522  }4523  if (QFB == PostBB) {4524    std::swap(QFB, QTB);4525    InvertQCond = true;4526  }4527 4528  // From this point on we can assume PTB or QTB may be fallthroughs but PFB4529  // and QFB may not. Model fallthroughs as a nullptr block.4530  if (PTB == QBI->getParent())4531    PTB = nullptr;4532  if (QTB == PostBB)4533    QTB = nullptr;4534 4535  // Legality bailouts. We must have at least the non-fallthrough blocks and4536  // the post-dominating block, and the non-fallthroughs must only have one4537  // predecessor.4538  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {4539    return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;4540  };4541  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||4542      !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))4543    return false;4544  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||4545      (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))4546    return false;4547  if (!QBI->getParent()->hasNUses(2))4548    return false;4549 4550  // OK, this is a sequence of two diamonds or triangles.4551  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.4552  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;4553  for (auto *BB : {PTB, PFB}) {4554    if (!BB)4555      continue;4556    for (auto &I : *BB)4557      if (StoreInst *SI = dyn_cast<StoreInst>(&I))4558        PStoreAddresses.insert(SI->getPointerOperand());4559  }4560  for (auto *BB : {QTB, QFB}) {4561    if (!BB)4562      continue;4563    for (auto &I : *BB)4564      if (StoreInst *SI = dyn_cast<StoreInst>(&I))4565        QStoreAddresses.insert(SI->getPointerOperand());4566  }4567 4568  set_intersect(PStoreAddresses, QStoreAddresses);4569  // set_intersect mutates PStoreAddresses in place. Rename it here to make it4570  // clear what it contains.4571  auto &CommonAddresses = PStoreAddresses;4572 4573  bool Changed = false;4574  for (auto *Address : CommonAddresses)4575    Changed |=4576        mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,4577                                       InvertPCond, InvertQCond, DTU, DL, TTI);4578  return Changed;4579}4580 4581/// If the previous block ended with a widenable branch, determine if reusing4582/// the target block is profitable and legal.  This will have the effect of4583/// "widening" PBI, but doesn't require us to reason about hosting safety.4584static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,4585                                           DomTreeUpdater *DTU) {4586  // TODO: This can be generalized in two important ways:4587  // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input4588  //    values from the PBI edge.4589  // 2) We can sink side effecting instructions into BI's fallthrough4590  //    successor provided they doesn't contribute to computation of4591  //    BI's condition.4592  BasicBlock *IfTrueBB = PBI->getSuccessor(0);4593  BasicBlock *IfFalseBB = PBI->getSuccessor(1);4594  if (!isWidenableBranch(PBI) || IfTrueBB != BI->getParent() ||4595      !BI->getParent()->getSinglePredecessor())4596    return false;4597  if (!IfFalseBB->phis().empty())4598    return false; // TODO4599  // This helps avoid infinite loop with SimplifyCondBranchToCondBranch which4600  // may undo the transform done here.4601  // TODO: There might be a more fine-grained solution to this.4602  if (!llvm::succ_empty(IfFalseBB))4603    return false;4604  // Use lambda to lazily compute expensive condition after cheap ones.4605  auto NoSideEffects = [](BasicBlock &BB) {4606    return llvm::none_of(BB, [](const Instruction &I) {4607        return I.mayWriteToMemory() || I.mayHaveSideEffects();4608      });4609  };4610  if (BI->getSuccessor(1) != IfFalseBB && // no inf looping4611      BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability4612      NoSideEffects(*BI->getParent())) {4613    auto *OldSuccessor = BI->getSuccessor(1);4614    OldSuccessor->removePredecessor(BI->getParent());4615    BI->setSuccessor(1, IfFalseBB);4616    if (DTU)4617      DTU->applyUpdates(4618          {{DominatorTree::Insert, BI->getParent(), IfFalseBB},4619           {DominatorTree::Delete, BI->getParent(), OldSuccessor}});4620    return true;4621  }4622  if (BI->getSuccessor(0) != IfFalseBB && // no inf looping4623      BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability4624      NoSideEffects(*BI->getParent())) {4625    auto *OldSuccessor = BI->getSuccessor(0);4626    OldSuccessor->removePredecessor(BI->getParent());4627    BI->setSuccessor(0, IfFalseBB);4628    if (DTU)4629      DTU->applyUpdates(4630          {{DominatorTree::Insert, BI->getParent(), IfFalseBB},4631           {DominatorTree::Delete, BI->getParent(), OldSuccessor}});4632    return true;4633  }4634  return false;4635}4636 4637/// If we have a conditional branch as a predecessor of another block,4638/// this function tries to simplify it.  We know4639/// that PBI and BI are both conditional branches, and BI is in one of the4640/// successor blocks of PBI - PBI branches to BI.4641static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,4642                                           DomTreeUpdater *DTU,4643                                           const DataLayout &DL,4644                                           const TargetTransformInfo &TTI) {4645  assert(PBI->isConditional() && BI->isConditional());4646  BasicBlock *BB = BI->getParent();4647 4648  // If this block ends with a branch instruction, and if there is a4649  // predecessor that ends on a branch of the same condition, make4650  // this conditional branch redundant.4651  if (PBI->getCondition() == BI->getCondition() &&4652      PBI->getSuccessor(0) != PBI->getSuccessor(1)) {4653    // Okay, the outcome of this conditional branch is statically4654    // knowable.  If this block had a single pred, handle specially, otherwise4655    // foldCondBranchOnValueKnownInPredecessor() will handle it.4656    if (BB->getSinglePredecessor()) {4657      // Turn this into a branch on constant.4658      bool CondIsTrue = PBI->getSuccessor(0) == BB;4659      BI->setCondition(4660          ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));4661      return true; // Nuke the branch on constant.4662    }4663  }4664 4665  // If the previous block ended with a widenable branch, determine if reusing4666  // the target block is profitable and legal.  This will have the effect of4667  // "widening" PBI, but doesn't require us to reason about hosting safety.4668  if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))4669    return true;4670 4671  // If both branches are conditional and both contain stores to the same4672  // address, remove the stores from the conditionals and create a conditional4673  // merged store at the end.4674  if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))4675    return true;4676 4677  // If this is a conditional branch in an empty block, and if any4678  // predecessors are a conditional branch to one of our destinations,4679  // fold the conditions into logical ops and one cond br.4680 4681  // Ignore dbg intrinsics.4682  if (&*BB->instructionsWithoutDebug(false).begin() != BI)4683    return false;4684 4685  int PBIOp, BIOp;4686  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {4687    PBIOp = 0;4688    BIOp = 0;4689  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {4690    PBIOp = 0;4691    BIOp = 1;4692  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {4693    PBIOp = 1;4694    BIOp = 0;4695  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {4696    PBIOp = 1;4697    BIOp = 1;4698  } else {4699    return false;4700  }4701 4702  // Check to make sure that the other destination of this branch4703  // isn't BB itself.  If so, this is an infinite loop that will4704  // keep getting unwound.4705  if (PBI->getSuccessor(PBIOp) == BB)4706    return false;4707 4708  // If predecessor's branch probability to BB is too low don't merge branches.4709  SmallVector<uint32_t, 2> PredWeights;4710  if (!PBI->getMetadata(LLVMContext::MD_unpredictable) &&4711      extractBranchWeights(*PBI, PredWeights) &&4712      (static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]) != 0) {4713 4714    BranchProbability CommonDestProb = BranchProbability::getBranchProbability(4715        PredWeights[PBIOp],4716        static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]);4717 4718    BranchProbability Likely = TTI.getPredictableBranchThreshold();4719    if (CommonDestProb >= Likely)4720      return false;4721  }4722 4723  // Do not perform this transformation if it would require4724  // insertion of a large number of select instructions. For targets4725  // without predication/cmovs, this is a big pessimization.4726 4727  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);4728  BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);4729  unsigned NumPhis = 0;4730  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);4731       ++II, ++NumPhis) {4732    if (NumPhis > 2) // Disable this xform.4733      return false;4734  }4735 4736  // Finally, if everything is ok, fold the branches to logical ops.4737  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);4738 4739  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()4740                    << "AND: " << *BI->getParent());4741 4742  SmallVector<DominatorTree::UpdateType, 5> Updates;4743 4744  // If OtherDest *is* BB, then BB is a basic block with a single conditional4745  // branch in it, where one edge (OtherDest) goes back to itself but the other4746  // exits.  We don't *know* that the program avoids the infinite loop4747  // (even though that seems likely).  If we do this xform naively, we'll end up4748  // recursively unpeeling the loop.  Since we know that (after the xform is4749  // done) that the block *is* infinite if reached, we just make it an obviously4750  // infinite loop with no cond branch.4751  if (OtherDest == BB) {4752    // Insert it at the end of the function, because it's either code,4753    // or it won't matter if it's hot. :)4754    BasicBlock *InfLoopBlock =4755        BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());4756    BranchInst::Create(InfLoopBlock, InfLoopBlock);4757    if (DTU)4758      Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});4759    OtherDest = InfLoopBlock;4760  }4761 4762  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());4763 4764  // BI may have other predecessors.  Because of this, we leave4765  // it alone, but modify PBI.4766 4767  // Make sure we get to CommonDest on True&True directions.4768  Value *PBICond = PBI->getCondition();4769  IRBuilder<NoFolder> Builder(PBI);4770  if (PBIOp)4771    PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");4772 4773  Value *BICond = BI->getCondition();4774  if (BIOp)4775    BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");4776 4777  // Merge the conditions.4778  Value *Cond =4779      createLogicalOp(Builder, Instruction::Or, PBICond, BICond, "brmerge");4780 4781  // Modify PBI to branch on the new condition to the new dests.4782  PBI->setCondition(Cond);4783  PBI->setSuccessor(0, CommonDest);4784  PBI->setSuccessor(1, OtherDest);4785 4786  if (DTU) {4787    Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});4788    Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});4789 4790    DTU->applyUpdates(Updates);4791  }4792 4793  // Update branch weight for PBI.4794  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;4795  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;4796  bool HasWeights =4797      extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,4798                             SuccTrueWeight, SuccFalseWeight);4799  if (HasWeights) {4800    PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;4801    PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;4802    SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;4803    SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;4804    // The weight to CommonDest should be PredCommon * SuccTotal +4805    //                                    PredOther * SuccCommon.4806    // The weight to OtherDest should be PredOther * SuccOther.4807    uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +4808                                  PredOther * SuccCommon,4809                              PredOther * SuccOther};4810 4811    setFittedBranchWeights(*PBI, NewWeights, /*IsExpected=*/false,4812                           /*ElideAllZero=*/true);4813    // Cond may be a select instruction with the first operand set to "true", or4814    // the second to "false" (see how createLogicalOp works for `and` and `or`)4815    if (!ProfcheckDisableMetadataFixes)4816      if (auto *SI = dyn_cast<SelectInst>(Cond)) {4817        assert(isSelectInRoleOfConjunctionOrDisjunction(SI));4818        // The select is predicated on PBICond4819        assert(dyn_cast<SelectInst>(SI)->getCondition() == PBICond);4820        // The corresponding probabilities are what was referred to above as4821        // PredCommon and PredOther.4822        setFittedBranchWeights(*SI, {PredCommon, PredOther},4823                               /*IsExpected=*/false, /*ElideAllZero=*/true);4824      }4825  }4826 4827  // OtherDest may have phi nodes.  If so, add an entry from PBI's4828  // block that are identical to the entries for BI's block.4829  addPredecessorToBlock(OtherDest, PBI->getParent(), BB);4830 4831  // We know that the CommonDest already had an edge from PBI to4832  // it.  If it has PHIs though, the PHIs may have different4833  // entries for BB and PBI's BB.  If so, insert a select to make4834  // them agree.4835  for (PHINode &PN : CommonDest->phis()) {4836    Value *BIV = PN.getIncomingValueForBlock(BB);4837    unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());4838    Value *PBIV = PN.getIncomingValue(PBBIdx);4839    if (BIV != PBIV) {4840      // Insert a select in PBI to pick the right value.4841      SelectInst *NV = cast<SelectInst>(4842          Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));4843      PN.setIncomingValue(PBBIdx, NV);4844      // The select has the same condition as PBI, in the same BB. The4845      // probabilities don't change.4846      if (HasWeights) {4847        uint64_t TrueWeight = PBIOp ? PredFalseWeight : PredTrueWeight;4848        uint64_t FalseWeight = PBIOp ? PredTrueWeight : PredFalseWeight;4849        setFittedBranchWeights(*NV, {TrueWeight, FalseWeight},4850                               /*IsExpected=*/false, /*ElideAllZero=*/true);4851      }4852    }4853  }4854 4855  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());4856  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());4857 4858  // This basic block is probably dead.  We know it has at least4859  // one fewer predecessor.4860  return true;4861}4862 4863// Simplifies a terminator by replacing it with a branch to TrueBB if Cond is4864// true or to FalseBB if Cond is false.4865// Takes care of updating the successors and removing the old terminator.4866// Also makes sure not to introduce new successors by assuming that edges to4867// non-successor TrueBBs and FalseBBs aren't reachable.4868bool SimplifyCFGOpt::simplifyTerminatorOnSelect(Instruction *OldTerm,4869                                                Value *Cond, BasicBlock *TrueBB,4870                                                BasicBlock *FalseBB,4871                                                uint32_t TrueWeight,4872                                                uint32_t FalseWeight) {4873  auto *BB = OldTerm->getParent();4874  // Remove any superfluous successor edges from the CFG.4875  // First, figure out which successors to preserve.4876  // If TrueBB and FalseBB are equal, only try to preserve one copy of that4877  // successor.4878  BasicBlock *KeepEdge1 = TrueBB;4879  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;4880 4881  SmallSetVector<BasicBlock *, 2> RemovedSuccessors;4882 4883  // Then remove the rest.4884  for (BasicBlock *Succ : successors(OldTerm)) {4885    // Make sure only to keep exactly one copy of each edge.4886    if (Succ == KeepEdge1)4887      KeepEdge1 = nullptr;4888    else if (Succ == KeepEdge2)4889      KeepEdge2 = nullptr;4890    else {4891      Succ->removePredecessor(BB,4892                              /*KeepOneInputPHIs=*/true);4893 4894      if (Succ != TrueBB && Succ != FalseBB)4895        RemovedSuccessors.insert(Succ);4896    }4897  }4898 4899  IRBuilder<> Builder(OldTerm);4900  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());4901 4902  // Insert an appropriate new terminator.4903  if (!KeepEdge1 && !KeepEdge2) {4904    if (TrueBB == FalseBB) {4905      // We were only looking for one successor, and it was present.4906      // Create an unconditional branch to it.4907      Builder.CreateBr(TrueBB);4908    } else {4909      // We found both of the successors we were looking for.4910      // Create a conditional branch sharing the condition of the select.4911      BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);4912      setBranchWeights(*NewBI, {TrueWeight, FalseWeight},4913                       /*IsExpected=*/false, /*ElideAllZero=*/true);4914    }4915  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {4916    // Neither of the selected blocks were successors, so this4917    // terminator must be unreachable.4918    new UnreachableInst(OldTerm->getContext(), OldTerm->getIterator());4919  } else {4920    // One of the selected values was a successor, but the other wasn't.4921    // Insert an unconditional branch to the one that was found;4922    // the edge to the one that wasn't must be unreachable.4923    if (!KeepEdge1) {4924      // Only TrueBB was found.4925      Builder.CreateBr(TrueBB);4926    } else {4927      // Only FalseBB was found.4928      Builder.CreateBr(FalseBB);4929    }4930  }4931 4932  eraseTerminatorAndDCECond(OldTerm);4933 4934  if (DTU) {4935    SmallVector<DominatorTree::UpdateType, 2> Updates;4936    Updates.reserve(RemovedSuccessors.size());4937    for (auto *RemovedSuccessor : RemovedSuccessors)4938      Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});4939    DTU->applyUpdates(Updates);4940  }4941 4942  return true;4943}4944 4945// Replaces4946//   (switch (select cond, X, Y)) on constant X, Y4947// with a branch - conditional if X and Y lead to distinct BBs,4948// unconditional otherwise.4949bool SimplifyCFGOpt::simplifySwitchOnSelect(SwitchInst *SI,4950                                            SelectInst *Select) {4951  // Check for constant integer values in the select.4952  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());4953  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());4954  if (!TrueVal || !FalseVal)4955    return false;4956 4957  // Find the relevant condition and destinations.4958  Value *Condition = Select->getCondition();4959  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();4960  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();4961 4962  // Get weight for TrueBB and FalseBB.4963  uint32_t TrueWeight = 0, FalseWeight = 0;4964  SmallVector<uint64_t, 8> Weights;4965  bool HasWeights = hasBranchWeightMD(*SI);4966  if (HasWeights) {4967    getBranchWeights(SI, Weights);4968    if (Weights.size() == 1 + SI->getNumCases()) {4969      TrueWeight =4970          (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];4971      FalseWeight =4972          (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];4973    }4974  }4975 4976  // Perform the actual simplification.4977  return simplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,4978                                    FalseWeight);4979}4980 4981// Replaces4982//   (indirectbr (select cond, blockaddress(@fn, BlockA),4983//                             blockaddress(@fn, BlockB)))4984// with4985//   (br cond, BlockA, BlockB).4986bool SimplifyCFGOpt::simplifyIndirectBrOnSelect(IndirectBrInst *IBI,4987                                                SelectInst *SI) {4988  // Check that both operands of the select are block addresses.4989  BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());4990  BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());4991  if (!TBA || !FBA)4992    return false;4993 4994  // Extract the actual blocks.4995  BasicBlock *TrueBB = TBA->getBasicBlock();4996  BasicBlock *FalseBB = FBA->getBasicBlock();4997 4998  // The select's profile becomes the profile of the conditional branch that4999  // replaces the indirect branch.5000  SmallVector<uint32_t> SelectBranchWeights(2);5001  if (!ProfcheckDisableMetadataFixes)5002    extractBranchWeights(*SI, SelectBranchWeights);5003  // Perform the actual simplification.5004  return simplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,5005                                    SelectBranchWeights[0],5006                                    SelectBranchWeights[1]);5007}5008 5009/// This is called when we find an icmp instruction5010/// (a seteq/setne with a constant) as the only instruction in a5011/// block that ends with an uncond branch.  We are looking for a very specific5012/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified.  In5013/// this case, we merge the first two "or's of icmp" into a switch, but then the5014/// default value goes to an uncond block with a seteq in it, we get something5015/// like:5016///5017///   switch i8 %A, label %DEFAULT [ i8 1, label %end    i8 2, label %end ]5018/// DEFAULT:5019///   %tmp = icmp eq i8 %A, 925020///   br label %end5021/// end:5022///   ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]5023///5024/// We prefer to split the edge to 'end' so that there is a true/false entry to5025/// the PHI, merging the third icmp into the switch.5026bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(5027    ICmpInst *ICI, IRBuilder<> &Builder) {5028  // Select == nullptr means we assume that there is a hidden no-op select5029  // instruction of `_ = select %icmp, true, false` after `%icmp = icmp ...`5030  return tryToSimplifyUncondBranchWithICmpSelectInIt(ICI, nullptr, Builder);5031}5032 5033/// Similar to tryToSimplifyUncondBranchWithICmpInIt, but handle a more generic5034/// case. This is called when we find an icmp instruction (a seteq/setne with a5035/// constant) and its following select instruction as the only TWO instructions5036/// in a block that ends with an uncond branch.  We are looking for a very5037/// specific pattern that occurs when "5038///    if (A == 1) return C1;5039///    if (A == 2) return C2;5040///    if (A < 3) return C3;5041///    return C4;5042/// " gets simplified.  In this case, we merge the first two "branches of icmp"5043/// into a switch, but then the default value goes to an uncond block with a lt5044/// icmp and select in it, as InstCombine can not simplify "A < 3" as "A == 2".5045/// After SimplifyCFG and other subsequent optimizations (e.g., SCCP), we might5046/// get something like:5047///5048/// case1:5049///   switch i8 %A, label %DEFAULT [ i8 0, label %end    i8 1, label %case2 ]5050/// case2:5051///   br label %end5052/// DEFAULT:5053///   %tmp = icmp eq i8 %A, 25054///   %val = select i1 %tmp, i8 C3, i8 C45055///   br label %end5056/// end:5057///   _ = phi i8 [ C1, %case1 ], [ C2, %case2 ], [ %val, %DEFAULT ]5058///5059/// We prefer to split the edge to 'end' so that there are TWO entries of V3/V45060/// to the PHI, merging the icmp & select into the switch, as follows:5061///5062/// case1:5063///   switch i8 %A, label %DEFAULT [5064///     i8 0, label %end5065///     i8 1, label %case25066///     i8 2, label %case35067///   ]5068/// case2:5069///   br label %end5070/// case3:5071///   br label %end5072/// DEFAULT:5073///   br label %end5074/// end:5075///   _ = phi i8 [ C1, %case1 ], [ C2, %case2 ], [ C3, %case2 ], [ C4, %DEFAULT]5076bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpSelectInIt(5077    ICmpInst *ICI, SelectInst *Select, IRBuilder<> &Builder) {5078  BasicBlock *BB = ICI->getParent();5079 5080  // If the block has any PHIs in it or the icmp/select has multiple uses, it is5081  // too complex.5082  /// TODO: support multi-phis in succ BB of select's BB.5083  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse() ||5084      (Select && !Select->hasOneUse()))5085    return false;5086 5087  // The pattern we're looking for is where our only predecessor is a switch on5088  // 'V' and this block is the default case for the switch.  In this case we can5089  // fold the compared value into the switch to simplify things.5090  BasicBlock *Pred = BB->getSinglePredecessor();5091  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))5092    return false;5093 5094  Value *IcmpCond;5095  ConstantInt *NewCaseVal;5096  CmpPredicate Predicate;5097 5098  // Match icmp X, C5099  if (!match(ICI,5100             m_ICmp(Predicate, m_Value(IcmpCond), m_ConstantInt(NewCaseVal))))5101    return false;5102 5103  Value *SelectCond, *SelectTrueVal, *SelectFalseVal;5104  Instruction *User;5105  if (!Select) {5106    // If Select == nullptr, we can assume that there is a hidden no-op select5107    // just after icmp5108    SelectCond = ICI;5109    SelectTrueVal = Builder.getTrue();5110    SelectFalseVal = Builder.getFalse();5111    User = ICI->user_back();5112  } else {5113    SelectCond = Select->getCondition();5114    // Check if the select condition is the same as the icmp condition.5115    if (SelectCond != ICI)5116      return false;5117    SelectTrueVal = Select->getTrueValue();5118    SelectFalseVal = Select->getFalseValue();5119    User = Select->user_back();5120  }5121 5122  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());5123  if (SI->getCondition() != IcmpCond)5124    return false;5125 5126  // If BB is reachable on a non-default case, then we simply know the value of5127  // V in this block.  Substitute it and constant fold the icmp instruction5128  // away.5129  if (SI->getDefaultDest() != BB) {5130    ConstantInt *VVal = SI->findCaseDest(BB);5131    assert(VVal && "Should have a unique destination value");5132    ICI->setOperand(0, VVal);5133 5134    if (Value *V = simplifyInstruction(ICI, {DL, ICI})) {5135      ICI->replaceAllUsesWith(V);5136      ICI->eraseFromParent();5137    }5138    // BB is now empty, so it is likely to simplify away.5139    return requestResimplify();5140  }5141 5142  // Ok, the block is reachable from the default dest.  If the constant we're5143  // comparing exists in one of the other edges, then we can constant fold ICI5144  // and zap it.5145  if (SI->findCaseValue(NewCaseVal) != SI->case_default()) {5146    Value *V;5147    if (Predicate == ICmpInst::ICMP_EQ)5148      V = ConstantInt::getFalse(BB->getContext());5149    else5150      V = ConstantInt::getTrue(BB->getContext());5151 5152    ICI->replaceAllUsesWith(V);5153    ICI->eraseFromParent();5154    // BB is now empty, so it is likely to simplify away.5155    return requestResimplify();5156  }5157 5158  // The use of the select has to be in the 'end' block, by the only PHI node in5159  // the block.5160  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);5161  PHINode *PHIUse = dyn_cast<PHINode>(User);5162  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||5163      isa<PHINode>(++BasicBlock::iterator(PHIUse)))5164    return false;5165 5166  // If the icmp is a SETEQ, then the default dest gets SelectFalseVal, the new5167  // edge gets SelectTrueVal in the PHI.5168  Value *DefaultCst = SelectFalseVal;5169  Value *NewCst = SelectTrueVal;5170 5171  if (ICI->getPredicate() == ICmpInst::ICMP_NE)5172    std::swap(DefaultCst, NewCst);5173 5174  // Replace Select (which is used by the PHI for the default value) with5175  // SelectFalseVal or SelectTrueVal depending on if ICI is EQ or NE.5176  if (Select) {5177    Select->replaceAllUsesWith(DefaultCst);5178    Select->eraseFromParent();5179  } else {5180    ICI->replaceAllUsesWith(DefaultCst);5181  }5182  ICI->eraseFromParent();5183 5184  SmallVector<DominatorTree::UpdateType, 2> Updates;5185 5186  // Okay, the switch goes to this block on a default value.  Add an edge from5187  // the switch to the merge point on the compared value.5188  BasicBlock *NewBB =5189      BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);5190  {5191    SwitchInstProfUpdateWrapper SIW(*SI);5192    auto W0 = SIW.getSuccessorWeight(0);5193    SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;5194    if (W0) {5195      NewW = ((uint64_t(*W0) + 1) >> 1);5196      SIW.setSuccessorWeight(0, *NewW);5197    }5198    SIW.addCase(NewCaseVal, NewBB, NewW);5199    if (DTU)5200      Updates.push_back({DominatorTree::Insert, Pred, NewBB});5201  }5202 5203  // NewBB branches to the phi block, add the uncond branch and the phi entry.5204  Builder.SetInsertPoint(NewBB);5205  Builder.SetCurrentDebugLocation(SI->getDebugLoc());5206  Builder.CreateBr(SuccBlock);5207  PHIUse->addIncoming(NewCst, NewBB);5208  if (DTU) {5209    Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});5210    DTU->applyUpdates(Updates);5211  }5212  return true;5213}5214 5215/// The specified branch is a conditional branch.5216/// Check to see if it is branching on an or/and chain of icmp instructions, and5217/// fold it into a switch instruction if so.5218bool SimplifyCFGOpt::simplifyBranchOnICmpChain(BranchInst *BI,5219                                               IRBuilder<> &Builder,5220                                               const DataLayout &DL) {5221  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());5222  if (!Cond)5223    return false;5224 5225  // Change br (X == 0 | X == 1), T, F into a switch instruction.5226  // If this is a bunch of seteq's or'd together, or if it's a bunch of5227  // 'setne's and'ed together, collect them.5228 5229  // Try to gather values from a chain of and/or to be turned into a switch5230  ConstantComparesGatherer ConstantCompare(Cond, DL);5231  // Unpack the result5232  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;5233  Value *CompVal = ConstantCompare.CompValue;5234  unsigned UsedICmps = ConstantCompare.UsedICmps;5235  Value *ExtraCase = ConstantCompare.Extra;5236  bool TrueWhenEqual = ConstantCompare.IsEq;5237 5238  // If we didn't have a multiply compared value, fail.5239  if (!CompVal)5240    return false;5241 5242  // Avoid turning single icmps into a switch.5243  if (UsedICmps <= 1)5244    return false;5245 5246  // There might be duplicate constants in the list, which the switch5247  // instruction can't handle, remove them now.5248  array_pod_sort(Values.begin(), Values.end(), constantIntSortPredicate);5249  Values.erase(llvm::unique(Values), Values.end());5250 5251  // If Extra was used, we require at least two switch values to do the5252  // transformation.  A switch with one value is just a conditional branch.5253  if (ExtraCase && Values.size() < 2)5254    return false;5255 5256  SmallVector<uint32_t> BranchWeights;5257  const bool HasProfile = !ProfcheckDisableMetadataFixes &&5258                          extractBranchWeights(*BI, BranchWeights);5259 5260  // Figure out which block is which destination.5261  BasicBlock *DefaultBB = BI->getSuccessor(1);5262  BasicBlock *EdgeBB = BI->getSuccessor(0);5263  if (!TrueWhenEqual) {5264    std::swap(DefaultBB, EdgeBB);5265    if (HasProfile)5266      std::swap(BranchWeights[0], BranchWeights[1]);5267  }5268 5269  BasicBlock *BB = BI->getParent();5270 5271  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()5272                    << " cases into SWITCH.  BB is:\n"5273                    << *BB);5274 5275  SmallVector<DominatorTree::UpdateType, 2> Updates;5276 5277  // If there are any extra values that couldn't be folded into the switch5278  // then we evaluate them with an explicit branch first. Split the block5279  // right before the condbr to handle it.5280  if (ExtraCase) {5281    BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,5282                                   /*MSSAU=*/nullptr, "switch.early.test");5283 5284    // Remove the uncond branch added to the old block.5285    Instruction *OldTI = BB->getTerminator();5286    Builder.SetInsertPoint(OldTI);5287 5288    // There can be an unintended UB if extra values are Poison. Before the5289    // transformation, extra values may not be evaluated according to the5290    // condition, and it will not raise UB. But after transformation, we are5291    // evaluating extra values before checking the condition, and it will raise5292    // UB. It can be solved by adding freeze instruction to extra values.5293    AssumptionCache *AC = Options.AC;5294 5295    if (!isGuaranteedNotToBeUndefOrPoison(ExtraCase, AC, BI, nullptr))5296      ExtraCase = Builder.CreateFreeze(ExtraCase);5297 5298    // We don't have any info about this condition.5299    auto *Br = TrueWhenEqual ? Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB)5300                             : Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);5301    setExplicitlyUnknownBranchWeightsIfProfiled(*Br, DEBUG_TYPE);5302 5303    OldTI->eraseFromParent();5304 5305    if (DTU)5306      Updates.push_back({DominatorTree::Insert, BB, EdgeBB});5307 5308    // If there are PHI nodes in EdgeBB, then we need to add a new entry to them5309    // for the edge we just added.5310    addPredecessorToBlock(EdgeBB, BB, NewBB);5311 5312    LLVM_DEBUG(dbgs() << "  ** 'icmp' chain unhandled condition: " << *ExtraCase5313                      << "\nEXTRABB = " << *BB);5314    BB = NewBB;5315  }5316 5317  Builder.SetInsertPoint(BI);5318  // Convert pointer to int before we switch.5319  if (CompVal->getType()->isPointerTy()) {5320    assert(!DL.hasUnstableRepresentation(CompVal->getType()) &&5321           "Should not end up here with unstable pointers");5322    CompVal = Builder.CreatePtrToInt(5323        CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");5324  }5325 5326  // Check if we can represent the values as a contiguous range. If so, we use a5327  // range check + conditional branch instead of a switch.5328  if (Values.front()->getValue() - Values.back()->getValue() ==5329      Values.size() - 1) {5330    ConstantRange RangeToCheck = ConstantRange::getNonEmpty(5331        Values.back()->getValue(), Values.front()->getValue() + 1);5332    APInt Offset, RHS;5333    ICmpInst::Predicate Pred;5334    RangeToCheck.getEquivalentICmp(Pred, RHS, Offset);5335    Value *X = CompVal;5336    if (!Offset.isZero())5337      X = Builder.CreateAdd(X, ConstantInt::get(CompVal->getType(), Offset));5338    Value *Cond =5339        Builder.CreateICmp(Pred, X, ConstantInt::get(CompVal->getType(), RHS));5340    BranchInst *NewBI = Builder.CreateCondBr(Cond, EdgeBB, DefaultBB);5341    if (HasProfile)5342      setBranchWeights(*NewBI, BranchWeights, /*IsExpected=*/false);5343    // We don't need to update PHI nodes since we don't add any new edges.5344  } else {5345    // Create the new switch instruction now.5346    SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());5347    if (HasProfile) {5348      // We know the weight of the default case. We don't know the weight of the5349      // other cases, but rather than completely lose profiling info, we split5350      // the remaining probability equally over them.5351      SmallVector<uint32_t> NewWeights(Values.size() + 1);5352      NewWeights[0] = BranchWeights[1]; // this is the default, and we swapped5353                                        // if TrueWhenEqual.5354      for (auto &V : drop_begin(NewWeights))5355        V = BranchWeights[0] / Values.size();5356      setBranchWeights(*New, NewWeights, /*IsExpected=*/false);5357    }5358 5359    // Add all of the 'cases' to the switch instruction.5360    for (ConstantInt *Val : Values)5361      New->addCase(Val, EdgeBB);5362 5363    // We added edges from PI to the EdgeBB.  As such, if there were any5364    // PHI nodes in EdgeBB, they need entries to be added corresponding to5365    // the number of edges added.5366    for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {5367      PHINode *PN = cast<PHINode>(BBI);5368      Value *InVal = PN->getIncomingValueForBlock(BB);5369      for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)5370        PN->addIncoming(InVal, BB);5371    }5372  }5373 5374  // Erase the old branch instruction.5375  eraseTerminatorAndDCECond(BI);5376  if (DTU)5377    DTU->applyUpdates(Updates);5378 5379  LLVM_DEBUG(dbgs() << "  ** 'icmp' chain result is:\n" << *BB << '\n');5380  return true;5381}5382 5383bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {5384  if (isa<PHINode>(RI->getValue()))5385    return simplifyCommonResume(RI);5386  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHIIt()) &&5387           RI->getValue() == &*RI->getParent()->getFirstNonPHIIt())5388    // The resume must unwind the exception that caused control to branch here.5389    return simplifySingleResume(RI);5390 5391  return false;5392}5393 5394// Check if cleanup block is empty5395static bool isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R) {5396  for (Instruction &I : R) {5397    auto *II = dyn_cast<IntrinsicInst>(&I);5398    if (!II)5399      return false;5400 5401    Intrinsic::ID IntrinsicID = II->getIntrinsicID();5402    switch (IntrinsicID) {5403    case Intrinsic::dbg_declare:5404    case Intrinsic::dbg_value:5405    case Intrinsic::dbg_label:5406    case Intrinsic::lifetime_end:5407      break;5408    default:5409      return false;5410    }5411  }5412  return true;5413}5414 5415// Simplify resume that is shared by several landing pads (phi of landing pad).5416bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {5417  BasicBlock *BB = RI->getParent();5418 5419  // Check that there are no other instructions except for debug and lifetime5420  // intrinsics between the phi's and resume instruction.5421  if (!isCleanupBlockEmpty(make_range(RI->getParent()->getFirstNonPHIIt(),5422                                      BB->getTerminator()->getIterator())))5423    return false;5424 5425  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;5426  auto *PhiLPInst = cast<PHINode>(RI->getValue());5427 5428  // Check incoming blocks to see if any of them are trivial.5429  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;5430       Idx++) {5431    auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);5432    auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);5433 5434    // If the block has other successors, we can not delete it because5435    // it has other dependents.5436    if (IncomingBB->getUniqueSuccessor() != BB)5437      continue;5438 5439    auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHIIt());5440    // Not the landing pad that caused the control to branch here.5441    if (IncomingValue != LandingPad)5442      continue;5443 5444    if (isCleanupBlockEmpty(5445            make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))5446      TrivialUnwindBlocks.insert(IncomingBB);5447  }5448 5449  // If no trivial unwind blocks, don't do any simplifications.5450  if (TrivialUnwindBlocks.empty())5451    return false;5452 5453  // Turn all invokes that unwind here into calls.5454  for (auto *TrivialBB : TrivialUnwindBlocks) {5455    // Blocks that will be simplified should be removed from the phi node.5456    // Note there could be multiple edges to the resume block, and we need5457    // to remove them all.5458    while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)5459      BB->removePredecessor(TrivialBB, true);5460 5461    for (BasicBlock *Pred :5462         llvm::make_early_inc_range(predecessors(TrivialBB))) {5463      removeUnwindEdge(Pred, DTU);5464      ++NumInvokes;5465    }5466 5467    // In each SimplifyCFG run, only the current processed block can be erased.5468    // Otherwise, it will break the iteration of SimplifyCFG pass. So instead5469    // of erasing TrivialBB, we only remove the branch to the common resume5470    // block so that we can later erase the resume block since it has no5471    // predecessors.5472    TrivialBB->getTerminator()->eraseFromParent();5473    new UnreachableInst(RI->getContext(), TrivialBB);5474    if (DTU)5475      DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});5476  }5477 5478  // Delete the resume block if all its predecessors have been removed.5479  if (pred_empty(BB))5480    DeleteDeadBlock(BB, DTU);5481 5482  return !TrivialUnwindBlocks.empty();5483}5484 5485// Simplify resume that is only used by a single (non-phi) landing pad.5486bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {5487  BasicBlock *BB = RI->getParent();5488  auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHIIt());5489  assert(RI->getValue() == LPInst &&5490         "Resume must unwind the exception that caused control to here");5491 5492  // Check that there are no other instructions except for debug intrinsics.5493  if (!isCleanupBlockEmpty(5494          make_range<Instruction *>(LPInst->getNextNode(), RI)))5495    return false;5496 5497  // Turn all invokes that unwind here into calls and delete the basic block.5498  for (BasicBlock *Pred : llvm::make_early_inc_range(predecessors(BB))) {5499    removeUnwindEdge(Pred, DTU);5500    ++NumInvokes;5501  }5502 5503  // The landingpad is now unreachable.  Zap it.5504  DeleteDeadBlock(BB, DTU);5505  return true;5506}5507 5508static bool removeEmptyCleanup(CleanupReturnInst *RI, DomTreeUpdater *DTU) {5509  // If this is a trivial cleanup pad that executes no instructions, it can be5510  // eliminated.  If the cleanup pad continues to the caller, any predecessor5511  // that is an EH pad will be updated to continue to the caller and any5512  // predecessor that terminates with an invoke instruction will have its invoke5513  // instruction converted to a call instruction.  If the cleanup pad being5514  // simplified does not continue to the caller, each predecessor will be5515  // updated to continue to the unwind destination of the cleanup pad being5516  // simplified.5517  BasicBlock *BB = RI->getParent();5518  CleanupPadInst *CPInst = RI->getCleanupPad();5519  if (CPInst->getParent() != BB)5520    // This isn't an empty cleanup.5521    return false;5522 5523  // We cannot kill the pad if it has multiple uses.  This typically arises5524  // from unreachable basic blocks.5525  if (!CPInst->hasOneUse())5526    return false;5527 5528  // Check that there are no other instructions except for benign intrinsics.5529  if (!isCleanupBlockEmpty(5530          make_range<Instruction *>(CPInst->getNextNode(), RI)))5531    return false;5532 5533  // If the cleanup return we are simplifying unwinds to the caller, this will5534  // set UnwindDest to nullptr.5535  BasicBlock *UnwindDest = RI->getUnwindDest();5536 5537  // We're about to remove BB from the control flow.  Before we do, sink any5538  // PHINodes into the unwind destination.  Doing this before changing the5539  // control flow avoids some potentially slow checks, since we can currently5540  // be certain that UnwindDest and BB have no common predecessors (since they5541  // are both EH pads).5542  if (UnwindDest) {5543    // First, go through the PHI nodes in UnwindDest and update any nodes that5544    // reference the block we are removing5545    for (PHINode &DestPN : UnwindDest->phis()) {5546      int Idx = DestPN.getBasicBlockIndex(BB);5547      // Since BB unwinds to UnwindDest, it has to be in the PHI node.5548      assert(Idx != -1);5549      // This PHI node has an incoming value that corresponds to a control5550      // path through the cleanup pad we are removing.  If the incoming5551      // value is in the cleanup pad, it must be a PHINode (because we5552      // verified above that the block is otherwise empty).  Otherwise, the5553      // value is either a constant or a value that dominates the cleanup5554      // pad being removed.5555      //5556      // Because BB and UnwindDest are both EH pads, all of their5557      // predecessors must unwind to these blocks, and since no instruction5558      // can have multiple unwind destinations, there will be no overlap in5559      // incoming blocks between SrcPN and DestPN.5560      Value *SrcVal = DestPN.getIncomingValue(Idx);5561      PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);5562 5563      bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB;5564      for (auto *Pred : predecessors(BB)) {5565        Value *Incoming =5566            NeedPHITranslation ? SrcPN->getIncomingValueForBlock(Pred) : SrcVal;5567        DestPN.addIncoming(Incoming, Pred);5568      }5569    }5570 5571    // Sink any remaining PHI nodes directly into UnwindDest.5572    BasicBlock::iterator InsertPt = UnwindDest->getFirstNonPHIIt();5573    for (PHINode &PN : make_early_inc_range(BB->phis())) {5574      if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB))5575        // If the PHI node has no uses or all of its uses are in this basic5576        // block (meaning they are debug or lifetime intrinsics), just leave5577        // it.  It will be erased when we erase BB below.5578        continue;5579 5580      // Otherwise, sink this PHI node into UnwindDest.5581      // Any predecessors to UnwindDest which are not already represented5582      // must be back edges which inherit the value from the path through5583      // BB.  In this case, the PHI value must reference itself.5584      for (auto *pred : predecessors(UnwindDest))5585        if (pred != BB)5586          PN.addIncoming(&PN, pred);5587      PN.moveBefore(InsertPt);5588      // Also, add a dummy incoming value for the original BB itself,5589      // so that the PHI is well-formed until we drop said predecessor.5590      PN.addIncoming(PoisonValue::get(PN.getType()), BB);5591    }5592  }5593 5594  std::vector<DominatorTree::UpdateType> Updates;5595 5596  // We use make_early_inc_range here because we will remove all predecessors.5597  for (BasicBlock *PredBB : llvm::make_early_inc_range(predecessors(BB))) {5598    if (UnwindDest == nullptr) {5599      if (DTU) {5600        DTU->applyUpdates(Updates);5601        Updates.clear();5602      }5603      removeUnwindEdge(PredBB, DTU);5604      ++NumInvokes;5605    } else {5606      BB->removePredecessor(PredBB);5607      Instruction *TI = PredBB->getTerminator();5608      TI->replaceUsesOfWith(BB, UnwindDest);5609      if (DTU) {5610        Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});5611        Updates.push_back({DominatorTree::Delete, PredBB, BB});5612      }5613    }5614  }5615 5616  if (DTU)5617    DTU->applyUpdates(Updates);5618 5619  DeleteDeadBlock(BB, DTU);5620 5621  return true;5622}5623 5624// Try to merge two cleanuppads together.5625static bool mergeCleanupPad(CleanupReturnInst *RI) {5626  // Skip any cleanuprets which unwind to caller, there is nothing to merge5627  // with.5628  BasicBlock *UnwindDest = RI->getUnwindDest();5629  if (!UnwindDest)5630    return false;5631 5632  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't5633  // be safe to merge without code duplication.5634  if (UnwindDest->getSinglePredecessor() != RI->getParent())5635    return false;5636 5637  // Verify that our cleanuppad's unwind destination is another cleanuppad.5638  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());5639  if (!SuccessorCleanupPad)5640    return false;5641 5642  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();5643  // Replace any uses of the successor cleanupad with the predecessor pad5644  // The only cleanuppad uses should be this cleanupret, it's cleanupret and5645  // funclet bundle operands.5646  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);5647  // Remove the old cleanuppad.5648  SuccessorCleanupPad->eraseFromParent();5649  // Now, we simply replace the cleanupret with a branch to the unwind5650  // destination.5651  BranchInst::Create(UnwindDest, RI->getParent());5652  RI->eraseFromParent();5653 5654  return true;5655}5656 5657bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {5658  // It is possible to transiantly have an undef cleanuppad operand because we5659  // have deleted some, but not all, dead blocks.5660  // Eventually, this block will be deleted.5661  if (isa<UndefValue>(RI->getOperand(0)))5662    return false;5663 5664  if (mergeCleanupPad(RI))5665    return true;5666 5667  if (removeEmptyCleanup(RI, DTU))5668    return true;5669 5670  return false;5671}5672 5673// WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()!5674bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {5675  BasicBlock *BB = UI->getParent();5676 5677  bool Changed = false;5678 5679  // Ensure that any debug-info records that used to occur after the Unreachable5680  // are moved to in front of it -- otherwise they'll "dangle" at the end of5681  // the block.5682  BB->flushTerminatorDbgRecords();5683 5684  // Debug-info records on the unreachable inst itself should be deleted, as5685  // below we delete everything past the final executable instruction.5686  UI->dropDbgRecords();5687 5688  // If there are any instructions immediately before the unreachable that can5689  // be removed, do so.5690  while (UI->getIterator() != BB->begin()) {5691    BasicBlock::iterator BBI = UI->getIterator();5692    --BBI;5693 5694    if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI))5695      break; // Can not drop any more instructions. We're done here.5696    // Otherwise, this instruction can be freely erased,5697    // even if it is not side-effect free.5698 5699    // Note that deleting EH's here is in fact okay, although it involves a bit5700    // of subtle reasoning. If this inst is an EH, all the predecessors of this5701    // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn,5702    // and we can therefore guarantee this block will be erased.5703 5704    // If we're deleting this, we're deleting any subsequent debug info, so5705    // delete DbgRecords.5706    BBI->dropDbgRecords();5707 5708    // Delete this instruction (any uses are guaranteed to be dead)5709    BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType()));5710    BBI->eraseFromParent();5711    Changed = true;5712  }5713 5714  // If the unreachable instruction is the first in the block, take a gander5715  // at all of the predecessors of this instruction, and simplify them.5716  if (&BB->front() != UI)5717    return Changed;5718 5719  std::vector<DominatorTree::UpdateType> Updates;5720 5721  SmallSetVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));5722  for (BasicBlock *Predecessor : Preds) {5723    Instruction *TI = Predecessor->getTerminator();5724    IRBuilder<> Builder(TI);5725    if (auto *BI = dyn_cast<BranchInst>(TI)) {5726      // We could either have a proper unconditional branch,5727      // or a degenerate conditional branch with matching destinations.5728      if (all_of(BI->successors(),5729                 [BB](auto *Successor) { return Successor == BB; })) {5730        new UnreachableInst(TI->getContext(), TI->getIterator());5731        TI->eraseFromParent();5732        Changed = true;5733      } else {5734        assert(BI->isConditional() && "Can't get here with an uncond branch.");5735        Value* Cond = BI->getCondition();5736        assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&5737               "The destinations are guaranteed to be different here.");5738        CallInst *Assumption;5739        if (BI->getSuccessor(0) == BB) {5740          Assumption = Builder.CreateAssumption(Builder.CreateNot(Cond));5741          Builder.CreateBr(BI->getSuccessor(1));5742        } else {5743          assert(BI->getSuccessor(1) == BB && "Incorrect CFG");5744          Assumption = Builder.CreateAssumption(Cond);5745          Builder.CreateBr(BI->getSuccessor(0));5746        }5747        if (Options.AC)5748          Options.AC->registerAssumption(cast<AssumeInst>(Assumption));5749 5750        eraseTerminatorAndDCECond(BI);5751        Changed = true;5752      }5753      if (DTU)5754        Updates.push_back({DominatorTree::Delete, Predecessor, BB});5755    } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {5756      SwitchInstProfUpdateWrapper SU(*SI);5757      for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {5758        if (i->getCaseSuccessor() != BB) {5759          ++i;5760          continue;5761        }5762        BB->removePredecessor(SU->getParent());5763        i = SU.removeCase(i);5764        e = SU->case_end();5765        Changed = true;5766      }5767      // Note that the default destination can't be removed!5768      if (DTU && SI->getDefaultDest() != BB)5769        Updates.push_back({DominatorTree::Delete, Predecessor, BB});5770    } else if (auto *II = dyn_cast<InvokeInst>(TI)) {5771      if (II->getUnwindDest() == BB) {5772        if (DTU) {5773          DTU->applyUpdates(Updates);5774          Updates.clear();5775        }5776        auto *CI = cast<CallInst>(removeUnwindEdge(TI->getParent(), DTU));5777        if (!CI->doesNotThrow())5778          CI->setDoesNotThrow();5779        Changed = true;5780      }5781    } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {5782      if (CSI->getUnwindDest() == BB) {5783        if (DTU) {5784          DTU->applyUpdates(Updates);5785          Updates.clear();5786        }5787        removeUnwindEdge(TI->getParent(), DTU);5788        Changed = true;5789        continue;5790      }5791 5792      for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),5793                                             E = CSI->handler_end();5794           I != E; ++I) {5795        if (*I == BB) {5796          CSI->removeHandler(I);5797          --I;5798          --E;5799          Changed = true;5800        }5801      }5802      if (DTU)5803        Updates.push_back({DominatorTree::Delete, Predecessor, BB});5804      if (CSI->getNumHandlers() == 0) {5805        if (CSI->hasUnwindDest()) {5806          // Redirect all predecessors of the block containing CatchSwitchInst5807          // to instead branch to the CatchSwitchInst's unwind destination.5808          if (DTU) {5809            for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {5810              Updates.push_back({DominatorTree::Insert,5811                                 PredecessorOfPredecessor,5812                                 CSI->getUnwindDest()});5813              Updates.push_back({DominatorTree::Delete,5814                                 PredecessorOfPredecessor, Predecessor});5815            }5816          }5817          Predecessor->replaceAllUsesWith(CSI->getUnwindDest());5818        } else {5819          // Rewrite all preds to unwind to caller (or from invoke to call).5820          if (DTU) {5821            DTU->applyUpdates(Updates);5822            Updates.clear();5823          }5824          SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));5825          for (BasicBlock *EHPred : EHPreds)5826            removeUnwindEdge(EHPred, DTU);5827        }5828        // The catchswitch is no longer reachable.5829        new UnreachableInst(CSI->getContext(), CSI->getIterator());5830        CSI->eraseFromParent();5831        Changed = true;5832      }5833    } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {5834      (void)CRI;5835      assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&5836             "Expected to always have an unwind to BB.");5837      if (DTU)5838        Updates.push_back({DominatorTree::Delete, Predecessor, BB});5839      new UnreachableInst(TI->getContext(), TI->getIterator());5840      TI->eraseFromParent();5841      Changed = true;5842    }5843  }5844 5845  if (DTU)5846    DTU->applyUpdates(Updates);5847 5848  // If this block is now dead, remove it.5849  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {5850    DeleteDeadBlock(BB, DTU);5851    return true;5852  }5853 5854  return Changed;5855}5856 5857struct ContiguousCasesResult {5858  ConstantInt *Min;5859  ConstantInt *Max;5860  BasicBlock *Dest;5861  BasicBlock *OtherDest;5862  SmallVectorImpl<ConstantInt *> *Cases;5863  SmallVectorImpl<ConstantInt *> *OtherCases;5864};5865 5866static std::optional<ContiguousCasesResult>5867findContiguousCases(Value *Condition, SmallVectorImpl<ConstantInt *> &Cases,5868                    SmallVectorImpl<ConstantInt *> &OtherCases,5869                    BasicBlock *Dest, BasicBlock *OtherDest) {5870  assert(Cases.size() >= 1);5871 5872  array_pod_sort(Cases.begin(), Cases.end(), constantIntSortPredicate);5873  const APInt &Min = Cases.back()->getValue();5874  const APInt &Max = Cases.front()->getValue();5875  APInt Offset = Max - Min;5876  size_t ContiguousOffset = Cases.size() - 1;5877  if (Offset == ContiguousOffset) {5878    return ContiguousCasesResult{5879        /*Min=*/Cases.back(),5880        /*Max=*/Cases.front(),5881        /*Dest=*/Dest,5882        /*OtherDest=*/OtherDest,5883        /*Cases=*/&Cases,5884        /*OtherCases=*/&OtherCases,5885    };5886  }5887  ConstantRange CR = computeConstantRange(Condition, /*ForSigned=*/false);5888  // If this is a wrapping contiguous range, that is, [Min, OtherMin] +5889  // [OtherMax, Max] (also [OtherMax, OtherMin]), [OtherMin+1, OtherMax-1] is a5890  // contiguous range for the other destination. N.B. If CR is not a full range,5891  // Max+1 is not equal to Min. It's not continuous in arithmetic.5892  if (Max == CR.getUnsignedMax() && Min == CR.getUnsignedMin()) {5893    assert(Cases.size() >= 2);5894    auto *It =5895        std::adjacent_find(Cases.begin(), Cases.end(), [](auto L, auto R) {5896          return L->getValue() != R->getValue() + 1;5897        });5898    if (It == Cases.end())5899      return std::nullopt;5900    auto [OtherMax, OtherMin] = std::make_pair(*It, *std::next(It));5901    if ((Max - OtherMax->getValue()) + (OtherMin->getValue() - Min) ==5902        Cases.size() - 2) {5903      return ContiguousCasesResult{5904          /*Min=*/cast<ConstantInt>(5905              ConstantInt::get(OtherMin->getType(), OtherMin->getValue() + 1)),5906          /*Max=*/5907          cast<ConstantInt>(5908              ConstantInt::get(OtherMax->getType(), OtherMax->getValue() - 1)),5909          /*Dest=*/OtherDest,5910          /*OtherDest=*/Dest,5911          /*Cases=*/&OtherCases,5912          /*OtherCases=*/&Cases,5913      };5914    }5915  }5916  return std::nullopt;5917}5918 5919static void createUnreachableSwitchDefault(SwitchInst *Switch,5920                                           DomTreeUpdater *DTU,5921                                           bool RemoveOrigDefaultBlock = true) {5922  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");5923  auto *BB = Switch->getParent();5924  auto *OrigDefaultBlock = Switch->getDefaultDest();5925  if (RemoveOrigDefaultBlock)5926    OrigDefaultBlock->removePredecessor(BB);5927  BasicBlock *NewDefaultBlock = BasicBlock::Create(5928      BB->getContext(), BB->getName() + ".unreachabledefault", BB->getParent(),5929      OrigDefaultBlock);5930  auto *UI = new UnreachableInst(Switch->getContext(), NewDefaultBlock);5931  UI->setDebugLoc(DebugLoc::getTemporary());5932  Switch->setDefaultDest(&*NewDefaultBlock);5933  if (DTU) {5934    SmallVector<DominatorTree::UpdateType, 2> Updates;5935    Updates.push_back({DominatorTree::Insert, BB, &*NewDefaultBlock});5936    if (RemoveOrigDefaultBlock &&5937        !is_contained(successors(BB), OrigDefaultBlock))5938      Updates.push_back({DominatorTree::Delete, BB, &*OrigDefaultBlock});5939    DTU->applyUpdates(Updates);5940  }5941}5942 5943/// Turn a switch into an integer range comparison and branch.5944/// Switches with more than 2 destinations are ignored.5945/// Switches with 1 destination are also ignored.5946bool SimplifyCFGOpt::turnSwitchRangeIntoICmp(SwitchInst *SI,5947                                             IRBuilder<> &Builder) {5948  assert(SI->getNumCases() > 1 && "Degenerate switch?");5949 5950  bool HasDefault = !SI->defaultDestUnreachable();5951 5952  auto *BB = SI->getParent();5953  // Partition the cases into two sets with different destinations.5954  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;5955  BasicBlock *DestB = nullptr;5956  SmallVector<ConstantInt *, 16> CasesA;5957  SmallVector<ConstantInt *, 16> CasesB;5958 5959  for (auto Case : SI->cases()) {5960    BasicBlock *Dest = Case.getCaseSuccessor();5961    if (!DestA)5962      DestA = Dest;5963    if (Dest == DestA) {5964      CasesA.push_back(Case.getCaseValue());5965      continue;5966    }5967    if (!DestB)5968      DestB = Dest;5969    if (Dest == DestB) {5970      CasesB.push_back(Case.getCaseValue());5971      continue;5972    }5973    return false; // More than two destinations.5974  }5975  if (!DestB)5976    return false; // All destinations are the same and the default is unreachable5977 5978  assert(DestA && DestB &&5979         "Single-destination switch should have been folded.");5980  assert(DestA != DestB);5981  assert(DestB != SI->getDefaultDest());5982  assert(!CasesB.empty() && "There must be non-default cases.");5983  assert(!CasesA.empty() || HasDefault);5984 5985  // Figure out if one of the sets of cases form a contiguous range.5986  std::optional<ContiguousCasesResult> ContiguousCases;5987 5988  // Only one icmp is needed when there is only one case.5989  if (!HasDefault && CasesA.size() == 1)5990    ContiguousCases = ContiguousCasesResult{5991        /*Min=*/CasesA[0],5992        /*Max=*/CasesA[0],5993        /*Dest=*/DestA,5994        /*OtherDest=*/DestB,5995        /*Cases=*/&CasesA,5996        /*OtherCases=*/&CasesB,5997    };5998  else if (CasesB.size() == 1)5999    ContiguousCases = ContiguousCasesResult{6000        /*Min=*/CasesB[0],6001        /*Max=*/CasesB[0],6002        /*Dest=*/DestB,6003        /*OtherDest=*/DestA,6004        /*Cases=*/&CasesB,6005        /*OtherCases=*/&CasesA,6006    };6007  // Correctness: Cases to the default destination cannot be contiguous cases.6008  else if (!HasDefault)6009    ContiguousCases =6010        findContiguousCases(SI->getCondition(), CasesA, CasesB, DestA, DestB);6011 6012  if (!ContiguousCases)6013    ContiguousCases =6014        findContiguousCases(SI->getCondition(), CasesB, CasesA, DestB, DestA);6015 6016  if (!ContiguousCases)6017    return false;6018 6019  auto [Min, Max, Dest, OtherDest, Cases, OtherCases] = *ContiguousCases;6020 6021  // Start building the compare and branch.6022 6023  Constant *Offset = ConstantExpr::getNeg(Min);6024  Constant *NumCases = ConstantInt::get(Offset->getType(),6025                                        Max->getValue() - Min->getValue() + 1);6026  BranchInst *NewBI;6027  if (NumCases->isOneValue()) {6028    assert(Max->getValue() == Min->getValue());6029    Value *Cmp = Builder.CreateICmpEQ(SI->getCondition(), Min);6030    NewBI = Builder.CreateCondBr(Cmp, Dest, OtherDest);6031  }6032  // If NumCases overflowed, then all possible values jump to the successor.6033  else if (NumCases->isNullValue() && !Cases->empty()) {6034    NewBI = Builder.CreateBr(Dest);6035  } else {6036    Value *Sub = SI->getCondition();6037    if (!Offset->isNullValue())6038      Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");6039    Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");6040    NewBI = Builder.CreateCondBr(Cmp, Dest, OtherDest);6041  }6042 6043  // Update weight for the newly-created conditional branch.6044  if (hasBranchWeightMD(*SI) && NewBI->isConditional()) {6045    SmallVector<uint64_t, 8> Weights;6046    getBranchWeights(SI, Weights);6047    if (Weights.size() == 1 + SI->getNumCases()) {6048      uint64_t TrueWeight = 0;6049      uint64_t FalseWeight = 0;6050      for (size_t I = 0, E = Weights.size(); I != E; ++I) {6051        if (SI->getSuccessor(I) == Dest)6052          TrueWeight += Weights[I];6053        else6054          FalseWeight += Weights[I];6055      }6056      while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {6057        TrueWeight /= 2;6058        FalseWeight /= 2;6059      }6060      setFittedBranchWeights(*NewBI, {TrueWeight, FalseWeight},6061                             /*IsExpected=*/false, /*ElideAllZero=*/true);6062    }6063  }6064 6065  // Prune obsolete incoming values off the successors' PHI nodes.6066  for (auto &PHI : make_early_inc_range(Dest->phis())) {6067    unsigned PreviousEdges = Cases->size();6068    if (Dest == SI->getDefaultDest())6069      ++PreviousEdges;6070    for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)6071      PHI.removeIncomingValue(SI->getParent());6072  }6073  for (auto &PHI : make_early_inc_range(OtherDest->phis())) {6074    unsigned PreviousEdges = OtherCases->size();6075    if (OtherDest == SI->getDefaultDest())6076      ++PreviousEdges;6077    unsigned E = PreviousEdges - 1;6078    // Remove all incoming values from OtherDest if OtherDest is unreachable.6079    if (NewBI->isUnconditional())6080      ++E;6081    for (unsigned I = 0; I != E; ++I)6082      PHI.removeIncomingValue(SI->getParent());6083  }6084 6085  // Clean up the default block - it may have phis or other instructions before6086  // the unreachable terminator.6087  if (!HasDefault)6088    createUnreachableSwitchDefault(SI, DTU);6089 6090  auto *UnreachableDefault = SI->getDefaultDest();6091 6092  // Drop the switch.6093  SI->eraseFromParent();6094 6095  if (!HasDefault && DTU)6096    DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});6097 6098  return true;6099}6100 6101/// Compute masked bits for the condition of a switch6102/// and use it to remove dead cases.6103static bool eliminateDeadSwitchCases(SwitchInst *SI, DomTreeUpdater *DTU,6104                                     AssumptionCache *AC,6105                                     const DataLayout &DL) {6106  Value *Cond = SI->getCondition();6107  KnownBits Known = computeKnownBits(Cond, DL, AC, SI);6108  SmallPtrSet<const Constant *, 4> KnownValues;6109  bool IsKnownValuesValid = collectPossibleValues(Cond, KnownValues, 4);6110 6111  // We can also eliminate cases by determining that their values are outside of6112  // the limited range of the condition based on how many significant (non-sign)6113  // bits are in the condition value.6114  unsigned MaxSignificantBitsInCond =6115      ComputeMaxSignificantBits(Cond, DL, AC, SI);6116 6117  // Gather dead cases.6118  SmallVector<ConstantInt *, 8> DeadCases;6119  SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;6120  SmallVector<BasicBlock *, 8> UniqueSuccessors;6121  for (const auto &Case : SI->cases()) {6122    auto *Successor = Case.getCaseSuccessor();6123    if (DTU) {6124      auto [It, Inserted] = NumPerSuccessorCases.try_emplace(Successor);6125      if (Inserted)6126        UniqueSuccessors.push_back(Successor);6127      ++It->second;6128    }6129    ConstantInt *CaseC = Case.getCaseValue();6130    const APInt &CaseVal = CaseC->getValue();6131    if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||6132        (CaseVal.getSignificantBits() > MaxSignificantBitsInCond) ||6133        (IsKnownValuesValid && !KnownValues.contains(CaseC))) {6134      DeadCases.push_back(CaseC);6135      if (DTU)6136        --NumPerSuccessorCases[Successor];6137      LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal6138                        << " is dead.\n");6139    } else if (IsKnownValuesValid)6140      KnownValues.erase(CaseC);6141  }6142 6143  // If we can prove that the cases must cover all possible values, the6144  // default destination becomes dead and we can remove it.  If we know some6145  // of the bits in the value, we can use that to more precisely compute the6146  // number of possible unique case values.6147  bool HasDefault = !SI->defaultDestUnreachable();6148  const unsigned NumUnknownBits =6149      Known.getBitWidth() - (Known.Zero | Known.One).popcount();6150  assert(NumUnknownBits <= Known.getBitWidth());6151  if (HasDefault && DeadCases.empty()) {6152    if (IsKnownValuesValid && all_of(KnownValues, IsaPred<UndefValue>)) {6153      createUnreachableSwitchDefault(SI, DTU);6154      return true;6155    }6156 6157    if (NumUnknownBits < 64 /* avoid overflow */) {6158      uint64_t AllNumCases = 1ULL << NumUnknownBits;6159      if (SI->getNumCases() == AllNumCases) {6160        createUnreachableSwitchDefault(SI, DTU);6161        return true;6162      }6163      // When only one case value is missing, replace default with that case.6164      // Eliminating the default branch will provide more opportunities for6165      // optimization, such as lookup tables.6166      if (SI->getNumCases() == AllNumCases - 1) {6167        assert(NumUnknownBits > 1 && "Should be canonicalized to a branch");6168        IntegerType *CondTy = cast<IntegerType>(Cond->getType());6169        if (CondTy->getIntegerBitWidth() > 64 ||6170            !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))6171          return false;6172 6173        uint64_t MissingCaseVal = 0;6174        for (const auto &Case : SI->cases())6175          MissingCaseVal ^= Case.getCaseValue()->getValue().getLimitedValue();6176        auto *MissingCase = cast<ConstantInt>(6177            ConstantInt::get(Cond->getType(), MissingCaseVal));6178        SwitchInstProfUpdateWrapper SIW(*SI);6179        SIW.addCase(MissingCase, SI->getDefaultDest(),6180                    SIW.getSuccessorWeight(0));6181        createUnreachableSwitchDefault(SI, DTU,6182                                       /*RemoveOrigDefaultBlock*/ false);6183        SIW.setSuccessorWeight(0, 0);6184        return true;6185      }6186    }6187  }6188 6189  if (DeadCases.empty())6190    return false;6191 6192  SwitchInstProfUpdateWrapper SIW(*SI);6193  for (ConstantInt *DeadCase : DeadCases) {6194    SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);6195    assert(CaseI != SI->case_default() &&6196           "Case was not found. Probably mistake in DeadCases forming.");6197    // Prune unused values from PHI nodes.6198    CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());6199    SIW.removeCase(CaseI);6200  }6201 6202  if (DTU) {6203    std::vector<DominatorTree::UpdateType> Updates;6204    for (auto *Successor : UniqueSuccessors)6205      if (NumPerSuccessorCases[Successor] == 0)6206        Updates.push_back({DominatorTree::Delete, SI->getParent(), Successor});6207    DTU->applyUpdates(Updates);6208  }6209 6210  return true;6211}6212 6213/// If BB would be eligible for simplification by6214/// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated6215/// by an unconditional branch), look at the phi node for BB in the successor6216/// block and see if the incoming value is equal to CaseValue. If so, return6217/// the phi node, and set PhiIndex to BB's index in the phi node.6218static PHINode *findPHIForConditionForwarding(ConstantInt *CaseValue,6219                                              BasicBlock *BB, int *PhiIndex) {6220  if (&*BB->getFirstNonPHIIt() != BB->getTerminator())6221    return nullptr; // BB must be empty to be a candidate for simplification.6222  if (!BB->getSinglePredecessor())6223    return nullptr; // BB must be dominated by the switch.6224 6225  BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());6226  if (!Branch || !Branch->isUnconditional())6227    return nullptr; // Terminator must be unconditional branch.6228 6229  BasicBlock *Succ = Branch->getSuccessor(0);6230 6231  for (PHINode &PHI : Succ->phis()) {6232    int Idx = PHI.getBasicBlockIndex(BB);6233    assert(Idx >= 0 && "PHI has no entry for predecessor?");6234 6235    Value *InValue = PHI.getIncomingValue(Idx);6236    if (InValue != CaseValue)6237      continue;6238 6239    *PhiIndex = Idx;6240    return &PHI;6241  }6242 6243  return nullptr;6244}6245 6246/// Try to forward the condition of a switch instruction to a phi node6247/// dominated by the switch, if that would mean that some of the destination6248/// blocks of the switch can be folded away. Return true if a change is made.6249static bool forwardSwitchConditionToPHI(SwitchInst *SI) {6250  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;6251 6252  ForwardingNodesMap ForwardingNodes;6253  BasicBlock *SwitchBlock = SI->getParent();6254  bool Changed = false;6255  for (const auto &Case : SI->cases()) {6256    ConstantInt *CaseValue = Case.getCaseValue();6257    BasicBlock *CaseDest = Case.getCaseSuccessor();6258 6259    // Replace phi operands in successor blocks that are using the constant case6260    // value rather than the switch condition variable:6261    //   switchbb:6262    //   switch i32 %x, label %default [6263    //     i32 17, label %succ6264    //   ...6265    //   succ:6266    //     %r = phi i32 ... [ 17, %switchbb ] ...6267    // -->6268    //     %r = phi i32 ... [ %x, %switchbb ] ...6269 6270    for (PHINode &Phi : CaseDest->phis()) {6271      // This only works if there is exactly 1 incoming edge from the switch to6272      // a phi. If there is >1, that means multiple cases of the switch map to 16273      // value in the phi, and that phi value is not the switch condition. Thus,6274      // this transform would not make sense (the phi would be invalid because6275      // a phi can't have different incoming values from the same block).6276      int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);6277      if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&6278          count(Phi.blocks(), SwitchBlock) == 1) {6279        Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());6280        Changed = true;6281      }6282    }6283 6284    // Collect phi nodes that are indirectly using this switch's case constants.6285    int PhiIdx;6286    if (auto *Phi = findPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))6287      ForwardingNodes[Phi].push_back(PhiIdx);6288  }6289 6290  for (auto &ForwardingNode : ForwardingNodes) {6291    PHINode *Phi = ForwardingNode.first;6292    SmallVectorImpl<int> &Indexes = ForwardingNode.second;6293    // Check if it helps to fold PHI.6294    if (Indexes.size() < 2 && !llvm::is_contained(Phi->incoming_values(), SI->getCondition()))6295      continue;6296 6297    for (int Index : Indexes)6298      Phi->setIncomingValue(Index, SI->getCondition());6299    Changed = true;6300  }6301 6302  return Changed;6303}6304 6305/// Return true if the backend will be able to handle6306/// initializing an array of constants like C.6307static bool validLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {6308  if (C->isThreadDependent())6309    return false;6310  if (C->isDLLImportDependent())6311    return false;6312 6313  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&6314      !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&6315      !isa<UndefValue>(C) && !isa<ConstantExpr>(C))6316    return false;6317 6318  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {6319    // Pointer casts and in-bounds GEPs will not prohibit the backend from6320    // materializing the array of constants.6321    Constant *StrippedC = cast<Constant>(CE->stripInBoundsConstantOffsets());6322    if (StrippedC == C || !validLookupTableConstant(StrippedC, TTI))6323      return false;6324  }6325 6326  if (!TTI.shouldBuildLookupTablesForConstant(C))6327    return false;6328 6329  return true;6330}6331 6332/// If V is a Constant, return it. Otherwise, try to look up6333/// its constant value in ConstantPool, returning 0 if it's not there.6334static Constant *6335lookupConstant(Value *V,6336               const SmallDenseMap<Value *, Constant *> &ConstantPool) {6337  if (Constant *C = dyn_cast<Constant>(V))6338    return C;6339  return ConstantPool.lookup(V);6340}6341 6342/// Try to fold instruction I into a constant. This works for6343/// simple instructions such as binary operations where both operands are6344/// constant or can be replaced by constants from the ConstantPool. Returns the6345/// resulting constant on success, 0 otherwise.6346static Constant *6347constantFold(Instruction *I, const DataLayout &DL,6348             const SmallDenseMap<Value *, Constant *> &ConstantPool) {6349  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {6350    Constant *A = lookupConstant(Select->getCondition(), ConstantPool);6351    if (!A)6352      return nullptr;6353    if (A->isAllOnesValue())6354      return lookupConstant(Select->getTrueValue(), ConstantPool);6355    if (A->isNullValue())6356      return lookupConstant(Select->getFalseValue(), ConstantPool);6357    return nullptr;6358  }6359 6360  SmallVector<Constant *, 4> COps;6361  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {6362    if (Constant *A = lookupConstant(I->getOperand(N), ConstantPool))6363      COps.push_back(A);6364    else6365      return nullptr;6366  }6367 6368  return ConstantFoldInstOperands(I, COps, DL);6369}6370 6371/// Try to determine the resulting constant values in phi nodes6372/// at the common destination basic block, *CommonDest, for one of the case6373/// destionations CaseDest corresponding to value CaseVal (0 for the default6374/// case), of a switch instruction SI.6375static bool6376getCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,6377               BasicBlock **CommonDest,6378               SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,6379               const DataLayout &DL, const TargetTransformInfo &TTI) {6380  // The block from which we enter the common destination.6381  BasicBlock *Pred = SI->getParent();6382 6383  // If CaseDest is empty except for some side-effect free instructions through6384  // which we can constant-propagate the CaseVal, continue to its successor.6385  SmallDenseMap<Value *, Constant *> ConstantPool;6386  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));6387  for (Instruction &I : CaseDest->instructionsWithoutDebug(false)) {6388    if (I.isTerminator()) {6389      // If the terminator is a simple branch, continue to the next block.6390      if (I.getNumSuccessors() != 1 || I.isSpecialTerminator())6391        return false;6392      Pred = CaseDest;6393      CaseDest = I.getSuccessor(0);6394    } else if (Constant *C = constantFold(&I, DL, ConstantPool)) {6395      // Instruction is side-effect free and constant.6396 6397      // If the instruction has uses outside this block or a phi node slot for6398      // the block, it is not safe to bypass the instruction since it would then6399      // no longer dominate all its uses.6400      for (auto &Use : I.uses()) {6401        User *User = Use.getUser();6402        if (Instruction *I = dyn_cast<Instruction>(User))6403          if (I->getParent() == CaseDest)6404            continue;6405        if (PHINode *Phi = dyn_cast<PHINode>(User))6406          if (Phi->getIncomingBlock(Use) == CaseDest)6407            continue;6408        return false;6409      }6410 6411      ConstantPool.insert(std::make_pair(&I, C));6412    } else {6413      break;6414    }6415  }6416 6417  // If we did not have a CommonDest before, use the current one.6418  if (!*CommonDest)6419    *CommonDest = CaseDest;6420  // If the destination isn't the common one, abort.6421  if (CaseDest != *CommonDest)6422    return false;6423 6424  // Get the values for this case from phi nodes in the destination block.6425  for (PHINode &PHI : (*CommonDest)->phis()) {6426    int Idx = PHI.getBasicBlockIndex(Pred);6427    if (Idx == -1)6428      continue;6429 6430    Constant *ConstVal =6431        lookupConstant(PHI.getIncomingValue(Idx), ConstantPool);6432    if (!ConstVal)6433      return false;6434 6435    // Be conservative about which kinds of constants we support.6436    if (!validLookupTableConstant(ConstVal, TTI))6437      return false;6438 6439    Res.push_back(std::make_pair(&PHI, ConstVal));6440  }6441 6442  return Res.size() > 0;6443}6444 6445// Helper function used to add CaseVal to the list of cases that generate6446// Result. Returns the updated number of cases that generate this result.6447static size_t mapCaseToResult(ConstantInt *CaseVal,6448                              SwitchCaseResultVectorTy &UniqueResults,6449                              Constant *Result) {6450  for (auto &I : UniqueResults) {6451    if (I.first == Result) {6452      I.second.push_back(CaseVal);6453      return I.second.size();6454    }6455  }6456  UniqueResults.push_back(6457      std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));6458  return 1;6459}6460 6461// Helper function that initializes a map containing6462// results for the PHI node of the common destination block for a switch6463// instruction. Returns false if multiple PHI nodes have been found or if6464// there is not a common destination block for the switch.6465static bool initializeUniqueCases(SwitchInst *SI, PHINode *&PHI,6466                                  BasicBlock *&CommonDest,6467                                  SwitchCaseResultVectorTy &UniqueResults,6468                                  Constant *&DefaultResult,6469                                  const DataLayout &DL,6470                                  const TargetTransformInfo &TTI,6471                                  uintptr_t MaxUniqueResults) {6472  for (const auto &I : SI->cases()) {6473    ConstantInt *CaseVal = I.getCaseValue();6474 6475    // Resulting value at phi nodes for this case value.6476    SwitchCaseResultsTy Results;6477    if (!getCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,6478                        DL, TTI))6479      return false;6480 6481    // Only one value per case is permitted.6482    if (Results.size() > 1)6483      return false;6484 6485    // Add the case->result mapping to UniqueResults.6486    const size_t NumCasesForResult =6487        mapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);6488 6489    // Early out if there are too many cases for this result.6490    if (NumCasesForResult > MaxSwitchCasesPerResult)6491      return false;6492 6493    // Early out if there are too many unique results.6494    if (UniqueResults.size() > MaxUniqueResults)6495      return false;6496 6497    // Check the PHI consistency.6498    if (!PHI)6499      PHI = Results[0].first;6500    else if (PHI != Results[0].first)6501      return false;6502  }6503  // Find the default result value.6504  SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;6505  getCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,6506                 DL, TTI);6507  // If the default value is not found abort unless the default destination6508  // is unreachable.6509  DefaultResult =6510      DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;6511 6512  return DefaultResult || SI->defaultDestUnreachable();6513}6514 6515// Helper function that checks if it is possible to transform a switch with only6516// two cases (or two cases + default) that produces a result into a select.6517// TODO: Handle switches with more than 2 cases that map to the same result.6518// The branch weights correspond to the provided Condition (i.e. if Condition is6519// modified from the original SwitchInst, the caller must adjust the weights)6520static Value *foldSwitchToSelect(const SwitchCaseResultVectorTy &ResultVector,6521                                 Constant *DefaultResult, Value *Condition,6522                                 IRBuilder<> &Builder, const DataLayout &DL,6523                                 ArrayRef<uint32_t> BranchWeights) {6524  // If we are selecting between only two cases transform into a simple6525  // select or a two-way select if default is possible.6526  // Example:6527  // switch (a) {                  %0 = icmp eq i32 %a, 106528  //   case 10: return 42;         %1 = select i1 %0, i32 42, i32 46529  //   case 20: return 2;   ---->  %2 = icmp eq i32 %a, 206530  //   default: return 4;          %3 = select i1 %2, i32 2, i32 %16531  // }6532 6533  const bool HasBranchWeights =6534      !BranchWeights.empty() && !ProfcheckDisableMetadataFixes;6535 6536  if (ResultVector.size() == 2 && ResultVector[0].second.size() == 1 &&6537      ResultVector[1].second.size() == 1) {6538    ConstantInt *FirstCase = ResultVector[0].second[0];6539    ConstantInt *SecondCase = ResultVector[1].second[0];6540    Value *SelectValue = ResultVector[1].first;6541    if (DefaultResult) {6542      Value *ValueCompare =6543          Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");6544      SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,6545                                         DefaultResult, "switch.select");6546      if (auto *SI = dyn_cast<SelectInst>(SelectValue);6547          SI && HasBranchWeights) {6548        // We start with 3 probabilities, where the numerator is the6549        // corresponding BranchWeights[i], and the denominator is the sum over6550        // BranchWeights. We want the probability and negative probability of6551        // Condition == SecondCase.6552        assert(BranchWeights.size() == 3);6553        setBranchWeights(6554            *SI, {BranchWeights[2], BranchWeights[0] + BranchWeights[1]},6555            /*IsExpected=*/false, /*ElideAllZero=*/true);6556      }6557    }6558    Value *ValueCompare =6559        Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");6560    Value *Ret = Builder.CreateSelect(ValueCompare, ResultVector[0].first,6561                                      SelectValue, "switch.select");6562    if (auto *SI = dyn_cast<SelectInst>(Ret); SI && HasBranchWeights) {6563      // We may have had a DefaultResult. Base the position of the first and6564      // second's branch weights accordingly. Also the proability that Condition6565      // != FirstCase needs to take that into account.6566      assert(BranchWeights.size() >= 2);6567      size_t FirstCasePos = (Condition != nullptr);6568      size_t SecondCasePos = FirstCasePos + 1;6569      uint32_t DefaultCase = (Condition != nullptr) ? BranchWeights[0] : 0;6570      setBranchWeights(*SI,6571                       {BranchWeights[FirstCasePos],6572                        DefaultCase + BranchWeights[SecondCasePos]},6573                       /*IsExpected=*/false, /*ElideAllZero=*/true);6574    }6575    return Ret;6576  }6577 6578  // Handle the degenerate case where two cases have the same result value.6579  if (ResultVector.size() == 1 && DefaultResult) {6580    ArrayRef<ConstantInt *> CaseValues = ResultVector[0].second;6581    unsigned CaseCount = CaseValues.size();6582    // n bits group cases map to the same result:6583    // case 0,4      -> Cond & 0b1..1011 == 0 ? result : default6584    // case 0,2,4,6  -> Cond & 0b1..1001 == 0 ? result : default6585    // case 0,2,8,10 -> Cond & 0b1..0101 == 0 ? result : default6586    if (isPowerOf2_32(CaseCount)) {6587      ConstantInt *MinCaseVal = CaseValues[0];6588      // If there are bits that are set exclusively by CaseValues, we6589      // can transform the switch into a select if the conjunction of6590      // all the values uniquely identify CaseValues.6591      APInt AndMask = APInt::getAllOnes(MinCaseVal->getBitWidth());6592 6593      // Find the minimum value and compute the and of all the case values.6594      for (auto *Case : CaseValues) {6595        if (Case->getValue().slt(MinCaseVal->getValue()))6596          MinCaseVal = Case;6597        AndMask &= Case->getValue();6598      }6599      KnownBits Known = computeKnownBits(Condition, DL);6600 6601      if (!AndMask.isZero() && Known.getMaxValue().uge(AndMask)) {6602        // Compute the number of bits that are free to vary.6603        unsigned FreeBits = Known.countMaxActiveBits() - AndMask.popcount();6604 6605        // Check if the number of values covered by the mask is equal6606        // to the number of cases.6607        if (FreeBits == Log2_32(CaseCount)) {6608          Value *And = Builder.CreateAnd(Condition, AndMask);6609          Value *Cmp = Builder.CreateICmpEQ(6610              And, Constant::getIntegerValue(And->getType(), AndMask));6611          Value *Ret =6612              Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult);6613          if (auto *SI = dyn_cast<SelectInst>(Ret); SI && HasBranchWeights) {6614            // We know there's a Default case. We base the resulting branch6615            // weights off its probability.6616            assert(BranchWeights.size() >= 2);6617            setBranchWeights(6618                *SI,6619                {accumulate(drop_begin(BranchWeights), 0U), BranchWeights[0]},6620                /*IsExpected=*/false, /*ElideAllZero=*/true);6621          }6622          return Ret;6623        }6624      }6625 6626      // Mark the bits case number touched.6627      APInt BitMask = APInt::getZero(MinCaseVal->getBitWidth());6628      for (auto *Case : CaseValues)6629        BitMask |= (Case->getValue() - MinCaseVal->getValue());6630 6631      // Check if cases with the same result can cover all number6632      // in touched bits.6633      if (BitMask.popcount() == Log2_32(CaseCount)) {6634        if (!MinCaseVal->isNullValue())6635          Condition = Builder.CreateSub(Condition, MinCaseVal);6636        Value *And = Builder.CreateAnd(Condition, ~BitMask, "switch.and");6637        Value *Cmp = Builder.CreateICmpEQ(6638            And, Constant::getNullValue(And->getType()), "switch.selectcmp");6639        Value *Ret =6640            Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult);6641        if (auto *SI = dyn_cast<SelectInst>(Ret); SI && HasBranchWeights) {6642          assert(BranchWeights.size() >= 2);6643          setBranchWeights(6644              *SI,6645              {accumulate(drop_begin(BranchWeights), 0U), BranchWeights[0]},6646              /*IsExpected=*/false, /*ElideAllZero=*/true);6647        }6648        return Ret;6649      }6650    }6651 6652    // Handle the degenerate case where two cases have the same value.6653    if (CaseValues.size() == 2) {6654      Value *Cmp1 = Builder.CreateICmpEQ(Condition, CaseValues[0],6655                                         "switch.selectcmp.case1");6656      Value *Cmp2 = Builder.CreateICmpEQ(Condition, CaseValues[1],6657                                         "switch.selectcmp.case2");6658      Value *Cmp = Builder.CreateOr(Cmp1, Cmp2, "switch.selectcmp");6659      Value *Ret =6660          Builder.CreateSelect(Cmp, ResultVector[0].first, DefaultResult);6661      if (auto *SI = dyn_cast<SelectInst>(Ret); SI && HasBranchWeights) {6662        assert(BranchWeights.size() >= 2);6663        setBranchWeights(6664            *SI, {accumulate(drop_begin(BranchWeights), 0U), BranchWeights[0]},6665            /*IsExpected=*/false, /*ElideAllZero=*/true);6666      }6667      return Ret;6668    }6669  }6670 6671  return nullptr;6672}6673 6674// Helper function to cleanup a switch instruction that has been converted into6675// a select, fixing up PHI nodes and basic blocks.6676static void removeSwitchAfterSelectFold(SwitchInst *SI, PHINode *PHI,6677                                        Value *SelectValue,6678                                        IRBuilder<> &Builder,6679                                        DomTreeUpdater *DTU) {6680  std::vector<DominatorTree::UpdateType> Updates;6681 6682  BasicBlock *SelectBB = SI->getParent();6683  BasicBlock *DestBB = PHI->getParent();6684 6685  if (DTU && !is_contained(predecessors(DestBB), SelectBB))6686    Updates.push_back({DominatorTree::Insert, SelectBB, DestBB});6687  Builder.CreateBr(DestBB);6688 6689  // Remove the switch.6690 6691  PHI->removeIncomingValueIf(6692      [&](unsigned Idx) { return PHI->getIncomingBlock(Idx) == SelectBB; });6693  PHI->addIncoming(SelectValue, SelectBB);6694 6695  SmallPtrSet<BasicBlock *, 4> RemovedSuccessors;6696  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {6697    BasicBlock *Succ = SI->getSuccessor(i);6698 6699    if (Succ == DestBB)6700      continue;6701    Succ->removePredecessor(SelectBB);6702    if (DTU && RemovedSuccessors.insert(Succ).second)6703      Updates.push_back({DominatorTree::Delete, SelectBB, Succ});6704  }6705  SI->eraseFromParent();6706  if (DTU)6707    DTU->applyUpdates(Updates);6708}6709 6710/// If a switch is only used to initialize one or more phi nodes in a common6711/// successor block with only two different constant values, try to replace the6712/// switch with a select. Returns true if the fold was made.6713static bool trySwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,6714                              DomTreeUpdater *DTU, const DataLayout &DL,6715                              const TargetTransformInfo &TTI) {6716  Value *const Cond = SI->getCondition();6717  PHINode *PHI = nullptr;6718  BasicBlock *CommonDest = nullptr;6719  Constant *DefaultResult;6720  SwitchCaseResultVectorTy UniqueResults;6721  // Collect all the cases that will deliver the same value from the switch.6722  if (!initializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,6723                             DL, TTI, /*MaxUniqueResults*/ 2))6724    return false;6725 6726  assert(PHI != nullptr && "PHI for value select not found");6727  Builder.SetInsertPoint(SI);6728  SmallVector<uint32_t, 4> BranchWeights;6729  if (!ProfcheckDisableMetadataFixes) {6730    [[maybe_unused]] auto HasWeights =6731        extractBranchWeights(getBranchWeightMDNode(*SI), BranchWeights);6732    assert(!HasWeights == (BranchWeights.empty()));6733  }6734  assert(BranchWeights.empty() ||6735         (BranchWeights.size() >=6736          UniqueResults.size() + (DefaultResult != nullptr)));6737 6738  Value *SelectValue = foldSwitchToSelect(UniqueResults, DefaultResult, Cond,6739                                          Builder, DL, BranchWeights);6740  if (!SelectValue)6741    return false;6742 6743  removeSwitchAfterSelectFold(SI, PHI, SelectValue, Builder, DTU);6744  return true;6745}6746 6747namespace {6748 6749/// This class finds alternatives for switches to ultimately6750/// replace the switch.6751class SwitchReplacement {6752public:6753  /// Create a helper for optimizations to use as a switch replacement.6754  /// Find a better representation for the content of Values,6755  /// using DefaultValue to fill any holes in the table.6756  SwitchReplacement(6757      Module &M, uint64_t TableSize, ConstantInt *Offset,6758      const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,6759      Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);6760 6761  /// Build instructions with Builder to retrieve values using Index6762  /// and replace the switch.6763  Value *replaceSwitch(Value *Index, IRBuilder<> &Builder, const DataLayout &DL,6764                       Function *Func);6765 6766  /// Return true if a table with TableSize elements of6767  /// type ElementType would fit in a target-legal register.6768  static bool wouldFitInRegister(const DataLayout &DL, uint64_t TableSize,6769                                 Type *ElementType);6770 6771  /// Return the default value of the switch.6772  Constant *getDefaultValue();6773 6774  /// Return true if the replacement is a lookup table.6775  bool isLookupTable();6776 6777  /// Return true if the replacement is a bit map.6778  bool isBitMap();6779 6780private:6781  // Depending on the switch, there are different alternatives.6782  enum {6783    // For switches where each case contains the same value, we just have to6784    // store that single value and return it for each lookup.6785    SingleValueKind,6786 6787    // For switches where there is a linear relationship between table index6788    // and values. We calculate the result with a simple multiplication6789    // and addition instead of a table lookup.6790    LinearMapKind,6791 6792    // For small tables with integer elements, we can pack them into a bitmap6793    // that fits into a target-legal register. Values are retrieved by6794    // shift and mask operations.6795    BitMapKind,6796 6797    // The table is stored as an array of values. Values are retrieved by load6798    // instructions from the table.6799    LookupTableKind6800  } Kind;6801 6802  // The default value of the switch.6803  Constant *DefaultValue;6804 6805  // The type of the output values.6806  Type *ValueType;6807 6808  // For SingleValueKind, this is the single value.6809  Constant *SingleValue = nullptr;6810 6811  // For BitMapKind, this is the bitmap.6812  ConstantInt *BitMap = nullptr;6813  IntegerType *BitMapElementTy = nullptr;6814 6815  // For LinearMapKind, these are the constants used to derive the value.6816  ConstantInt *LinearOffset = nullptr;6817  ConstantInt *LinearMultiplier = nullptr;6818  bool LinearMapValWrapped = false;6819 6820  // For LookupTableKind, this is the table.6821  Constant *Initializer = nullptr;6822};6823 6824} // end anonymous namespace6825 6826SwitchReplacement::SwitchReplacement(6827    Module &M, uint64_t TableSize, ConstantInt *Offset,6828    const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,6829    Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName)6830    : DefaultValue(DefaultValue) {6831  assert(Values.size() && "Can't build lookup table without values!");6832  assert(TableSize >= Values.size() && "Can't fit values in table!");6833 6834  // If all values in the table are equal, this is that value.6835  SingleValue = Values.begin()->second;6836 6837  ValueType = Values.begin()->second->getType();6838 6839  // Build up the table contents.6840  SmallVector<Constant *, 64> TableContents(TableSize);6841  for (const auto &[CaseVal, CaseRes] : Values) {6842    assert(CaseRes->getType() == ValueType);6843 6844    uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();6845    TableContents[Idx] = CaseRes;6846 6847    if (SingleValue && !isa<PoisonValue>(CaseRes) && CaseRes != SingleValue)6848      SingleValue = isa<PoisonValue>(SingleValue) ? CaseRes : nullptr;6849  }6850 6851  // Fill in any holes in the table with the default result.6852  if (Values.size() < TableSize) {6853    assert(DefaultValue &&6854           "Need a default value to fill the lookup table holes.");6855    assert(DefaultValue->getType() == ValueType);6856    for (uint64_t I = 0; I < TableSize; ++I) {6857      if (!TableContents[I])6858        TableContents[I] = DefaultValue;6859    }6860 6861    // If the default value is poison, all the holes are poison.6862    bool DefaultValueIsPoison = isa<PoisonValue>(DefaultValue);6863 6864    if (DefaultValue != SingleValue && !DefaultValueIsPoison)6865      SingleValue = nullptr;6866  }6867 6868  // If each element in the table contains the same value, we only need to store6869  // that single value.6870  if (SingleValue) {6871    Kind = SingleValueKind;6872    return;6873  }6874 6875  // Check if we can derive the value with a linear transformation from the6876  // table index.6877  if (isa<IntegerType>(ValueType)) {6878    bool LinearMappingPossible = true;6879    APInt PrevVal;6880    APInt DistToPrev;6881    // When linear map is monotonic and signed overflow doesn't happen on6882    // maximum index, we can attach nsw on Add and Mul.6883    bool NonMonotonic = false;6884    assert(TableSize >= 2 && "Should be a SingleValue table.");6885    // Check if there is the same distance between two consecutive values.6886    for (uint64_t I = 0; I < TableSize; ++I) {6887      ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);6888 6889      if (!ConstVal && isa<PoisonValue>(TableContents[I])) {6890        // This is an poison, so it's (probably) a lookup table hole.6891        // To prevent any regressions from before we switched to using poison as6892        // the default value, holes will fall back to using the first value.6893        // This can be removed once we add proper handling for poisons in lookup6894        // tables.6895        ConstVal = dyn_cast<ConstantInt>(Values[0].second);6896      }6897 6898      if (!ConstVal) {6899        // This is an undef. We could deal with it, but undefs in lookup tables6900        // are very seldom. It's probably not worth the additional complexity.6901        LinearMappingPossible = false;6902        break;6903      }6904      const APInt &Val = ConstVal->getValue();6905      if (I != 0) {6906        APInt Dist = Val - PrevVal;6907        if (I == 1) {6908          DistToPrev = Dist;6909        } else if (Dist != DistToPrev) {6910          LinearMappingPossible = false;6911          break;6912        }6913        NonMonotonic |=6914            Dist.isStrictlyPositive() ? Val.sle(PrevVal) : Val.sgt(PrevVal);6915      }6916      PrevVal = Val;6917    }6918    if (LinearMappingPossible) {6919      LinearOffset = cast<ConstantInt>(TableContents[0]);6920      LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);6921      APInt M = LinearMultiplier->getValue();6922      bool MayWrap = true;6923      if (isIntN(M.getBitWidth(), TableSize - 1))6924        (void)M.smul_ov(APInt(M.getBitWidth(), TableSize - 1), MayWrap);6925      LinearMapValWrapped = NonMonotonic || MayWrap;6926      Kind = LinearMapKind;6927      return;6928    }6929  }6930 6931  // If the type is integer and the table fits in a register, build a bitmap.6932  if (wouldFitInRegister(DL, TableSize, ValueType)) {6933    IntegerType *IT = cast<IntegerType>(ValueType);6934    APInt TableInt(TableSize * IT->getBitWidth(), 0);6935    for (uint64_t I = TableSize; I > 0; --I) {6936      TableInt <<= IT->getBitWidth();6937      // Insert values into the bitmap. Undef values are set to zero.6938      if (!isa<UndefValue>(TableContents[I - 1])) {6939        ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);6940        TableInt |= Val->getValue().zext(TableInt.getBitWidth());6941      }6942    }6943    BitMap = ConstantInt::get(M.getContext(), TableInt);6944    BitMapElementTy = IT;6945    Kind = BitMapKind;6946    return;6947  }6948 6949  // Store the table in an array.6950  auto *TableTy = ArrayType::get(ValueType, TableSize);6951  Initializer = ConstantArray::get(TableTy, TableContents);6952 6953  Kind = LookupTableKind;6954}6955 6956Value *SwitchReplacement::replaceSwitch(Value *Index, IRBuilder<> &Builder,6957                                        const DataLayout &DL, Function *Func) {6958  switch (Kind) {6959  case SingleValueKind:6960    return SingleValue;6961  case LinearMapKind: {6962    ++NumLinearMaps;6963    // Derive the result value from the input value.6964    Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),6965                                          false, "switch.idx.cast");6966    if (!LinearMultiplier->isOne())6967      Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult",6968                                 /*HasNUW = */ false,6969                                 /*HasNSW = */ !LinearMapValWrapped);6970 6971    if (!LinearOffset->isZero())6972      Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset",6973                                 /*HasNUW = */ false,6974                                 /*HasNSW = */ !LinearMapValWrapped);6975    return Result;6976  }6977  case BitMapKind: {6978    ++NumBitMaps;6979    // Type of the bitmap (e.g. i59).6980    IntegerType *MapTy = BitMap->getIntegerType();6981 6982    // Cast Index to the same type as the bitmap.6983    // Note: The Index is <= the number of elements in the table, so6984    // truncating it to the width of the bitmask is safe.6985    Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");6986 6987    // Multiply the shift amount by the element width. NUW/NSW can always be6988    // set, because wouldFitInRegister guarantees Index * ShiftAmt is in6989    // BitMap's bit width.6990    ShiftAmt = Builder.CreateMul(6991        ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),6992        "switch.shiftamt",/*HasNUW =*/true,/*HasNSW =*/true);6993 6994    // Shift down.6995    Value *DownShifted =6996        Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");6997    // Mask off.6998    return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");6999  }7000  case LookupTableKind: {7001    ++NumLookupTables;7002    auto *Table =7003        new GlobalVariable(*Func->getParent(), Initializer->getType(),7004                           /*isConstant=*/true, GlobalVariable::PrivateLinkage,7005                           Initializer, "switch.table." + Func->getName());7006    Table->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);7007    // Set the alignment to that of an array items. We will be only loading one7008    // value out of it.7009    Table->setAlignment(DL.getPrefTypeAlign(ValueType));7010    Type *IndexTy = DL.getIndexType(Table->getType());7011    auto *ArrayTy = cast<ArrayType>(Table->getValueType());7012 7013    if (Index->getType() != IndexTy) {7014      unsigned OldBitWidth = Index->getType()->getIntegerBitWidth();7015      Index = Builder.CreateZExtOrTrunc(Index, IndexTy);7016      if (auto *Zext = dyn_cast<ZExtInst>(Index))7017        Zext->setNonNeg(7018            isUIntN(OldBitWidth - 1, ArrayTy->getNumElements() - 1));7019    }7020 7021    Value *GEPIndices[] = {ConstantInt::get(IndexTy, 0), Index};7022    Value *GEP =7023        Builder.CreateInBoundsGEP(ArrayTy, Table, GEPIndices, "switch.gep");7024    return Builder.CreateLoad(ArrayTy->getElementType(), GEP, "switch.load");7025  }7026  }7027  llvm_unreachable("Unknown helper kind!");7028}7029 7030bool SwitchReplacement::wouldFitInRegister(const DataLayout &DL,7031                                           uint64_t TableSize,7032                                           Type *ElementType) {7033  auto *IT = dyn_cast<IntegerType>(ElementType);7034  if (!IT)7035    return false;7036  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values7037  // are <= 15, we could try to narrow the type.7038 7039  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.7040  if (TableSize >= UINT_MAX / IT->getBitWidth())7041    return false;7042  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());7043}7044 7045static bool isTypeLegalForLookupTable(Type *Ty, const TargetTransformInfo &TTI,7046                                      const DataLayout &DL) {7047  // Allow any legal type.7048  if (TTI.isTypeLegal(Ty))7049    return true;7050 7051  auto *IT = dyn_cast<IntegerType>(Ty);7052  if (!IT)7053    return false;7054 7055  // Also allow power of 2 integer types that have at least 8 bits and fit in7056  // a register. These types are common in frontend languages and targets7057  // usually support loads of these types.7058  // TODO: We could relax this to any integer that fits in a register and rely7059  // on ABI alignment and padding in the table to allow the load to be widened.7060  // Or we could widen the constants and truncate the load.7061  unsigned BitWidth = IT->getBitWidth();7062  return BitWidth >= 8 && isPowerOf2_32(BitWidth) &&7063         DL.fitsInLegalInteger(IT->getBitWidth());7064}7065 7066Constant *SwitchReplacement::getDefaultValue() { return DefaultValue; }7067 7068bool SwitchReplacement::isLookupTable() { return Kind == LookupTableKind; }7069 7070bool SwitchReplacement::isBitMap() { return Kind == BitMapKind; }7071 7072static bool isSwitchDense(uint64_t NumCases, uint64_t CaseRange) {7073  // 40% is the default density for building a jump table in optsize/minsize7074  // mode. See also TargetLoweringBase::isSuitableForJumpTable(), which this7075  // function was based on.7076  const uint64_t MinDensity = 40;7077 7078  if (CaseRange >= UINT64_MAX / 100)7079    return false; // Avoid multiplication overflows below.7080 7081  return NumCases * 100 >= CaseRange * MinDensity;7082}7083 7084static bool isSwitchDense(ArrayRef<int64_t> Values) {7085  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();7086  uint64_t Range = Diff + 1;7087  if (Range < Diff)7088    return false; // Overflow.7089 7090  return isSwitchDense(Values.size(), Range);7091}7092 7093/// Determine whether a lookup table should be built for this switch, based on7094/// the number of cases, size of the table, and the types of the results.7095// TODO: We could support larger than legal types by limiting based on the7096// number of loads required and/or table size. If the constants are small we7097// could use smaller table entries and extend after the load.7098static bool shouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,7099                                   const TargetTransformInfo &TTI,7100                                   const DataLayout &DL,7101                                   const SmallVector<Type *> &ResultTypes) {7102  if (SI->getNumCases() > TableSize)7103    return false; // TableSize overflowed.7104 7105  bool AllTablesFitInRegister = true;7106  bool HasIllegalType = false;7107  for (const auto &Ty : ResultTypes) {7108    // Saturate this flag to true.7109    HasIllegalType = HasIllegalType || !isTypeLegalForLookupTable(Ty, TTI, DL);7110 7111    // Saturate this flag to false.7112    AllTablesFitInRegister =7113        AllTablesFitInRegister &&7114        SwitchReplacement::wouldFitInRegister(DL, TableSize, Ty);7115 7116    // If both flags saturate, we're done. NOTE: This *only* works with7117    // saturating flags, and all flags have to saturate first due to the7118    // non-deterministic behavior of iterating over a dense map.7119    if (HasIllegalType && !AllTablesFitInRegister)7120      break;7121  }7122 7123  // If each table would fit in a register, we should build it anyway.7124  if (AllTablesFitInRegister)7125    return true;7126 7127  // Don't build a table that doesn't fit in-register if it has illegal types.7128  if (HasIllegalType)7129    return false;7130 7131  return isSwitchDense(SI->getNumCases(), TableSize);7132}7133 7134static bool shouldUseSwitchConditionAsTableIndex(7135    ConstantInt &MinCaseVal, const ConstantInt &MaxCaseVal,7136    bool HasDefaultResults, const SmallVector<Type *> &ResultTypes,7137    const DataLayout &DL, const TargetTransformInfo &TTI) {7138  if (MinCaseVal.isNullValue())7139    return true;7140  if (MinCaseVal.isNegative() ||7141      MaxCaseVal.getLimitedValue() == std::numeric_limits<uint64_t>::max() ||7142      !HasDefaultResults)7143    return false;7144  return all_of(ResultTypes, [&](const auto &ResultType) {7145    return SwitchReplacement::wouldFitInRegister(7146        DL, MaxCaseVal.getLimitedValue() + 1 /* TableSize */, ResultType);7147  });7148}7149 7150/// Try to reuse the switch table index compare. Following pattern:7151/// \code7152///     if (idx < tablesize)7153///        r = table[idx]; // table does not contain default_value7154///     else7155///        r = default_value;7156///     if (r != default_value)7157///        ...7158/// \endcode7159/// Is optimized to:7160/// \code7161///     cond = idx < tablesize;7162///     if (cond)7163///        r = table[idx];7164///     else7165///        r = default_value;7166///     if (cond)7167///        ...7168/// \endcode7169/// Jump threading will then eliminate the second if(cond).7170static void reuseTableCompare(7171    User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,7172    Constant *DefaultValue,7173    const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {7174  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);7175  if (!CmpInst)7176    return;7177 7178  // We require that the compare is in the same block as the phi so that jump7179  // threading can do its work afterwards.7180  if (CmpInst->getParent() != PhiBlock)7181    return;7182 7183  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));7184  if (!CmpOp1)7185    return;7186 7187  Value *RangeCmp = RangeCheckBranch->getCondition();7188  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());7189  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());7190 7191  // Check if the compare with the default value is constant true or false.7192  const DataLayout &DL = PhiBlock->getDataLayout();7193  Constant *DefaultConst = ConstantFoldCompareInstOperands(7194      CmpInst->getPredicate(), DefaultValue, CmpOp1, DL);7195  if (DefaultConst != TrueConst && DefaultConst != FalseConst)7196    return;7197 7198  // Check if the compare with the case values is distinct from the default7199  // compare result.7200  for (auto ValuePair : Values) {7201    Constant *CaseConst = ConstantFoldCompareInstOperands(7202        CmpInst->getPredicate(), ValuePair.second, CmpOp1, DL);7203    if (!CaseConst || CaseConst == DefaultConst ||7204        (CaseConst != TrueConst && CaseConst != FalseConst))7205      return;7206  }7207 7208  // Check if the branch instruction dominates the phi node. It's a simple7209  // dominance check, but sufficient for our needs.7210  // Although this check is invariant in the calling loops, it's better to do it7211  // at this late stage. Practically we do it at most once for a switch.7212  BasicBlock *BranchBlock = RangeCheckBranch->getParent();7213  for (BasicBlock *Pred : predecessors(PhiBlock)) {7214    if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)7215      return;7216  }7217 7218  if (DefaultConst == FalseConst) {7219    // The compare yields the same result. We can replace it.7220    CmpInst->replaceAllUsesWith(RangeCmp);7221    ++NumTableCmpReuses;7222  } else {7223    // The compare yields the same result, just inverted. We can replace it.7224    Value *InvertedTableCmp = BinaryOperator::CreateXor(7225        RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",7226        RangeCheckBranch->getIterator());7227    CmpInst->replaceAllUsesWith(InvertedTableCmp);7228    ++NumTableCmpReuses;7229  }7230}7231 7232/// If the switch is only used to initialize one or more phi nodes in a common7233/// successor block with different constant values, replace the switch with7234/// lookup tables.7235static bool simplifySwitchLookup(SwitchInst *SI, IRBuilder<> &Builder,7236                                 DomTreeUpdater *DTU, const DataLayout &DL,7237                                 const TargetTransformInfo &TTI,7238                                 bool ConvertSwitchToLookupTable) {7239  assert(SI->getNumCases() > 1 && "Degenerate switch?");7240 7241  BasicBlock *BB = SI->getParent();7242  Function *Fn = BB->getParent();7243 7244  // FIXME: If the switch is too sparse for a lookup table, perhaps we could7245  // split off a dense part and build a lookup table for that.7246 7247  // FIXME: This creates arrays of GEPs to constant strings, which means each7248  // GEP needs a runtime relocation in PIC code. We should just build one big7249  // string and lookup indices into that.7250 7251  // Ignore switches with less than three cases. Lookup tables will not make7252  // them faster, so we don't analyze them.7253  if (SI->getNumCases() < 3)7254    return false;7255 7256  // Figure out the corresponding result for each case value and phi node in the7257  // common destination, as well as the min and max case values.7258  assert(!SI->cases().empty());7259  SwitchInst::CaseIt CI = SI->case_begin();7260  ConstantInt *MinCaseVal = CI->getCaseValue();7261  ConstantInt *MaxCaseVal = CI->getCaseValue();7262 7263  BasicBlock *CommonDest = nullptr;7264 7265  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;7266  SmallDenseMap<PHINode *, ResultListTy> ResultLists;7267 7268  SmallDenseMap<PHINode *, Constant *> DefaultResults;7269  SmallVector<Type *> ResultTypes;7270  SmallVector<PHINode *, 4> PHIs;7271 7272  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {7273    ConstantInt *CaseVal = CI->getCaseValue();7274    if (CaseVal->getValue().slt(MinCaseVal->getValue()))7275      MinCaseVal = CaseVal;7276    if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))7277      MaxCaseVal = CaseVal;7278 7279    // Resulting value at phi nodes for this case value.7280    using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;7281    ResultsTy Results;7282    if (!getCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,7283                        Results, DL, TTI))7284      return false;7285 7286    // Append the result and result types from this case to the list for each7287    // phi.7288    for (const auto &I : Results) {7289      PHINode *PHI = I.first;7290      Constant *Value = I.second;7291      auto [It, Inserted] = ResultLists.try_emplace(PHI);7292      if (Inserted)7293        PHIs.push_back(PHI);7294      It->second.push_back(std::make_pair(CaseVal, Value));7295      ResultTypes.push_back(PHI->getType());7296    }7297  }7298 7299  // If the table has holes, we need a constant result for the default case7300  // or a bitmask that fits in a register.7301  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;7302  bool HasDefaultResults =7303      getCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,7304                     DefaultResultsList, DL, TTI);7305  for (const auto &I : DefaultResultsList) {7306    PHINode *PHI = I.first;7307    Constant *Result = I.second;7308    DefaultResults[PHI] = Result;7309  }7310 7311  bool UseSwitchConditionAsTableIndex = shouldUseSwitchConditionAsTableIndex(7312      *MinCaseVal, *MaxCaseVal, HasDefaultResults, ResultTypes, DL, TTI);7313  uint64_t TableSize;7314  ConstantInt *TableIndexOffset;7315  if (UseSwitchConditionAsTableIndex) {7316    TableSize = MaxCaseVal->getLimitedValue() + 1;7317    TableIndexOffset = ConstantInt::get(MaxCaseVal->getIntegerType(), 0);7318  } else {7319    TableSize =7320        (MaxCaseVal->getValue() - MinCaseVal->getValue()).getLimitedValue() + 1;7321 7322    TableIndexOffset = MinCaseVal;7323  }7324 7325  // If the default destination is unreachable, or if the lookup table covers7326  // all values of the conditional variable, branch directly to the lookup table7327  // BB. Otherwise, check that the condition is within the case range.7328  uint64_t NumResults = ResultLists[PHIs[0]].size();7329  bool DefaultIsReachable = !SI->defaultDestUnreachable();7330 7331  bool TableHasHoles = (NumResults < TableSize);7332 7333  // If the table has holes but the default destination doesn't produce any7334  // constant results, the lookup table entries corresponding to the holes will7335  // contain poison.7336  bool AllHolesArePoison = TableHasHoles && !HasDefaultResults;7337 7338  // If the default destination doesn't produce a constant result but is still7339  // reachable, and the lookup table has holes, we need to use a mask to7340  // determine if the current index should load from the lookup table or jump7341  // to the default case.7342  // The mask is unnecessary if the table has holes but the default destination7343  // is unreachable, as in that case the holes must also be unreachable.7344  bool NeedMask = AllHolesArePoison && DefaultIsReachable;7345  if (NeedMask) {7346    // As an extra penalty for the validity test we require more cases.7347    if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).7348      return false;7349    if (!DL.fitsInLegalInteger(TableSize))7350      return false;7351  }7352 7353  if (!shouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))7354    return false;7355 7356  // Compute the table index value.7357  Value *TableIndex;7358  if (UseSwitchConditionAsTableIndex) {7359    TableIndex = SI->getCondition();7360    if (HasDefaultResults) {7361      // Grow the table to cover all possible index values to avoid the range7362      // check. It will use the default result to fill in the table hole later,7363      // so make sure it exist.7364      ConstantRange CR =7365          computeConstantRange(TableIndex, /* ForSigned */ false);7366      // Grow the table shouldn't have any size impact by checking7367      // wouldFitInRegister.7368      // TODO: Consider growing the table also when it doesn't fit in a register7369      // if no optsize is specified.7370      const uint64_t UpperBound = CR.getUpper().getLimitedValue();7371      if (!CR.isUpperWrapped() &&7372          all_of(ResultTypes, [&](const auto &ResultType) {7373            return SwitchReplacement::wouldFitInRegister(DL, UpperBound,7374                                                         ResultType);7375          })) {7376        // There may be some case index larger than the UpperBound (unreachable7377        // case), so make sure the table size does not get smaller.7378        TableSize = std::max(UpperBound, TableSize);7379        // The default branch is unreachable after we enlarge the lookup table.7380        // Adjust DefaultIsReachable to reuse code path.7381        DefaultIsReachable = false;7382      }7383    }7384  }7385 7386  // Keep track of the switch replacement for each phi7387  SmallDenseMap<PHINode *, SwitchReplacement> PhiToReplacementMap;7388  for (PHINode *PHI : PHIs) {7389    const auto &ResultList = ResultLists[PHI];7390 7391    Type *ResultType = ResultList.begin()->second->getType();7392    // Use any value to fill the lookup table holes.7393    Constant *DefaultVal =7394        AllHolesArePoison ? PoisonValue::get(ResultType) : DefaultResults[PHI];7395    StringRef FuncName = Fn->getName();7396    SwitchReplacement Replacement(*Fn->getParent(), TableSize, TableIndexOffset,7397                                  ResultList, DefaultVal, DL, FuncName);7398    PhiToReplacementMap.insert({PHI, Replacement});7399  }7400 7401  bool AnyLookupTables = any_of(7402      PhiToReplacementMap, [](auto &KV) { return KV.second.isLookupTable(); });7403  bool AnyBitMaps = any_of(PhiToReplacementMap,7404                           [](auto &KV) { return KV.second.isBitMap(); });7405 7406  // A few conditions prevent the generation of lookup tables:7407  //     1. The target does not support lookup tables.7408  //     2. The "no-jump-tables" function attribute is set.7409  // However, these objections do not apply to other switch replacements, like7410  // the bitmap, so we only stop here if any of these conditions are met and we7411  // want to create a LUT. Otherwise, continue with the switch replacement.7412  if (AnyLookupTables &&7413      (!TTI.shouldBuildLookupTables() ||7414       Fn->getFnAttribute("no-jump-tables").getValueAsBool()))7415    return false;7416 7417  // In the early optimization pipeline, disable formation of lookup tables,7418  // bit maps and mask checks, as they may inhibit further optimization.7419  if (!ConvertSwitchToLookupTable &&7420      (AnyLookupTables || AnyBitMaps || NeedMask))7421    return false;7422 7423  Builder.SetInsertPoint(SI);7424  // TableIndex is the switch condition - TableIndexOffset if we don't7425  // use the condition directly7426  if (!UseSwitchConditionAsTableIndex) {7427    // If the default is unreachable, all case values are s>= MinCaseVal. Then7428    // we can try to attach nsw.7429    bool MayWrap = true;7430    if (!DefaultIsReachable) {7431      APInt Res =7432          MaxCaseVal->getValue().ssub_ov(MinCaseVal->getValue(), MayWrap);7433      (void)Res;7434    }7435    TableIndex = Builder.CreateSub(SI->getCondition(), TableIndexOffset,7436                                   "switch.tableidx", /*HasNUW =*/false,7437                                   /*HasNSW =*/!MayWrap);7438  }7439 7440  std::vector<DominatorTree::UpdateType> Updates;7441 7442  // Compute the maximum table size representable by the integer type we are7443  // switching upon.7444  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();7445  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;7446  assert(MaxTableSize >= TableSize &&7447         "It is impossible for a switch to have more entries than the max "7448         "representable value of its input integer type's size.");7449 7450  // Create the BB that does the lookups.7451  Module &Mod = *CommonDest->getParent()->getParent();7452  BasicBlock *LookupBB = BasicBlock::Create(7453      Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);7454 7455  BranchInst *RangeCheckBranch = nullptr;7456  BranchInst *CondBranch = nullptr;7457 7458  Builder.SetInsertPoint(SI);7459  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);7460  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {7461    Builder.CreateBr(LookupBB);7462    if (DTU)7463      Updates.push_back({DominatorTree::Insert, BB, LookupBB});7464    // Note: We call removeProdecessor later since we need to be able to get the7465    // PHI value for the default case in case we're using a bit mask.7466  } else {7467    Value *Cmp = Builder.CreateICmpULT(7468        TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));7469    RangeCheckBranch =7470        Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());7471    CondBranch = RangeCheckBranch;7472    if (DTU)7473      Updates.push_back({DominatorTree::Insert, BB, LookupBB});7474  }7475 7476  // Populate the BB that does the lookups.7477  Builder.SetInsertPoint(LookupBB);7478 7479  if (NeedMask) {7480    // Before doing the lookup, we do the hole check. The LookupBB is therefore7481    // re-purposed to do the hole check, and we create a new LookupBB.7482    BasicBlock *MaskBB = LookupBB;7483    MaskBB->setName("switch.hole_check");7484    LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",7485                                  CommonDest->getParent(), CommonDest);7486 7487    // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid7488    // unnecessary illegal types.7489    uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));7490    APInt MaskInt(TableSizePowOf2, 0);7491    APInt One(TableSizePowOf2, 1);7492    // Build bitmask; fill in a 1 bit for every case.7493    const ResultListTy &ResultList = ResultLists[PHIs[0]];7494    for (const auto &Result : ResultList) {7495      uint64_t Idx = (Result.first->getValue() - TableIndexOffset->getValue())7496                         .getLimitedValue();7497      MaskInt |= One << Idx;7498    }7499    ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);7500 7501    // Get the TableIndex'th bit of the bitmask.7502    // If this bit is 0 (meaning hole) jump to the default destination,7503    // else continue with table lookup.7504    IntegerType *MapTy = TableMask->getIntegerType();7505    Value *MaskIndex =7506        Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");7507    Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");7508    Value *LoBit = Builder.CreateTrunc(7509        Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");7510    CondBranch = Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());7511    if (DTU) {7512      Updates.push_back({DominatorTree::Insert, MaskBB, LookupBB});7513      Updates.push_back({DominatorTree::Insert, MaskBB, SI->getDefaultDest()});7514    }7515    Builder.SetInsertPoint(LookupBB);7516    addPredecessorToBlock(SI->getDefaultDest(), MaskBB, BB);7517  }7518 7519  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {7520    // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,7521    // do not delete PHINodes here.7522    SI->getDefaultDest()->removePredecessor(BB,7523                                            /*KeepOneInputPHIs=*/true);7524    if (DTU)7525      Updates.push_back({DominatorTree::Delete, BB, SI->getDefaultDest()});7526  }7527 7528  for (PHINode *PHI : PHIs) {7529    const ResultListTy &ResultList = ResultLists[PHI];7530    auto Replacement = PhiToReplacementMap.at(PHI);7531    auto *Result = Replacement.replaceSwitch(TableIndex, Builder, DL, Fn);7532    // Do a small peephole optimization: re-use the switch table compare if7533    // possible.7534    if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {7535      BasicBlock *PhiBlock = PHI->getParent();7536      // Search for compare instructions which use the phi.7537      for (auto *User : PHI->users()) {7538        reuseTableCompare(User, PhiBlock, RangeCheckBranch,7539                          Replacement.getDefaultValue(), ResultList);7540      }7541    }7542 7543    PHI->addIncoming(Result, LookupBB);7544  }7545 7546  Builder.CreateBr(CommonDest);7547  if (DTU)7548    Updates.push_back({DominatorTree::Insert, LookupBB, CommonDest});7549 7550  SmallVector<uint32_t> BranchWeights;7551  const bool HasBranchWeights = CondBranch && !ProfcheckDisableMetadataFixes &&7552                                extractBranchWeights(*SI, BranchWeights);7553  uint64_t ToLookupWeight = 0;7554  uint64_t ToDefaultWeight = 0;7555 7556  // Remove the switch.7557  SmallPtrSet<BasicBlock *, 8> RemovedSuccessors;7558  for (unsigned I = 0, E = SI->getNumSuccessors(); I < E; ++I) {7559    BasicBlock *Succ = SI->getSuccessor(I);7560 7561    if (Succ == SI->getDefaultDest()) {7562      if (HasBranchWeights)7563        ToDefaultWeight += BranchWeights[I];7564      continue;7565    }7566    Succ->removePredecessor(BB);7567    if (DTU && RemovedSuccessors.insert(Succ).second)7568      Updates.push_back({DominatorTree::Delete, BB, Succ});7569    if (HasBranchWeights)7570      ToLookupWeight += BranchWeights[I];7571  }7572  SI->eraseFromParent();7573  if (HasBranchWeights)7574    setFittedBranchWeights(*CondBranch, {ToLookupWeight, ToDefaultWeight},7575                           /*IsExpected=*/false);7576  if (DTU)7577    DTU->applyUpdates(Updates);7578 7579  if (NeedMask)7580    ++NumLookupTablesHoles;7581  return true;7582}7583 7584/// Try to transform a switch that has "holes" in it to a contiguous sequence7585/// of cases.7586///7587/// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be7588/// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.7589///7590/// This converts a sparse switch into a dense switch which allows better7591/// lowering and could also allow transforming into a lookup table.7592static bool reduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,7593                              const DataLayout &DL,7594                              const TargetTransformInfo &TTI) {7595  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());7596  if (CondTy->getIntegerBitWidth() > 64 ||7597      !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))7598    return false;7599  // Only bother with this optimization if there are more than 3 switch cases;7600  // SDAG will only bother creating jump tables for 4 or more cases.7601  if (SI->getNumCases() < 4)7602    return false;7603 7604  // This transform is agnostic to the signedness of the input or case values. We7605  // can treat the case values as signed or unsigned. We can optimize more common7606  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values7607  // as signed.7608  SmallVector<int64_t,4> Values;7609  for (const auto &C : SI->cases())7610    Values.push_back(C.getCaseValue()->getValue().getSExtValue());7611  llvm::sort(Values);7612 7613  // If the switch is already dense, there's nothing useful to do here.7614  if (isSwitchDense(Values))7615    return false;7616 7617  // First, transform the values such that they start at zero and ascend.7618  int64_t Base = Values[0];7619  for (auto &V : Values)7620    V -= (uint64_t)(Base);7621 7622  // Now we have signed numbers that have been shifted so that, given enough7623  // precision, there are no negative values. Since the rest of the transform7624  // is bitwise only, we switch now to an unsigned representation.7625 7626  // This transform can be done speculatively because it is so cheap - it7627  // results in a single rotate operation being inserted.7628 7629  // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than7630  // one element and LLVM disallows duplicate cases, Shift is guaranteed to be7631  // less than 64.7632  unsigned Shift = 64;7633  for (auto &V : Values)7634    Shift = std::min(Shift, (unsigned)llvm::countr_zero((uint64_t)V));7635  assert(Shift < 64);7636  if (Shift > 0)7637    for (auto &V : Values)7638      V = (int64_t)((uint64_t)V >> Shift);7639 7640  if (!isSwitchDense(Values))7641    // Transform didn't create a dense switch.7642    return false;7643 7644  // The obvious transform is to shift the switch condition right and emit a7645  // check that the condition actually cleanly divided by GCD, i.e.7646  //   C & (1 << Shift - 1) == 07647  // inserting a new CFG edge to handle the case where it didn't divide cleanly.7648  //7649  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the7650  // shift and puts the shifted-off bits in the uppermost bits. If any of these7651  // are nonzero then the switch condition will be very large and will hit the7652  // default case.7653 7654  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());7655  Builder.SetInsertPoint(SI);7656  Value *Sub =7657      Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));7658  Value *Rot = Builder.CreateIntrinsic(7659      Ty, Intrinsic::fshl,7660      {Sub, Sub, ConstantInt::get(Ty, Ty->getBitWidth() - Shift)});7661  SI->replaceUsesOfWith(SI->getCondition(), Rot);7662 7663  for (auto Case : SI->cases()) {7664    auto *Orig = Case.getCaseValue();7665    auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base, true);7666    Case.setValue(cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(Shift))));7667  }7668  return true;7669}7670 7671/// Tries to transform the switch when the condition is umin with a constant.7672/// In that case, the default branch can be replaced by the constant's branch.7673/// This method also removes dead cases when the simplification cannot replace7674/// the default branch.7675///7676/// For example:7677/// switch(umin(a, 3)) {7678/// case 0:7679/// case 1:7680/// case 2:7681/// case 3:7682/// case 4:7683///   // ...7684/// default:7685///   unreachable7686/// }7687///7688/// Transforms into:7689///7690/// switch(a) {7691/// case 0:7692/// case 1:7693/// case 2:7694/// default:7695///   // This is case 37696/// }7697static bool simplifySwitchWhenUMin(SwitchInst *SI, DomTreeUpdater *DTU) {7698  Value *A;7699  ConstantInt *Constant;7700 7701  if (!match(SI->getCondition(), m_UMin(m_Value(A), m_ConstantInt(Constant))))7702    return false;7703 7704  SmallVector<DominatorTree::UpdateType> Updates;7705  SwitchInstProfUpdateWrapper SIW(*SI);7706  BasicBlock *BB = SIW->getParent();7707 7708  // Dead cases are removed even when the simplification fails.7709  // A case is dead when its value is higher than the Constant.7710  for (auto I = SI->case_begin(), E = SI->case_end(); I != E;) {7711    if (!I->getCaseValue()->getValue().ugt(Constant->getValue())) {7712      ++I;7713      continue;7714    }7715    BasicBlock *DeadCaseBB = I->getCaseSuccessor();7716    DeadCaseBB->removePredecessor(BB);7717    Updates.push_back({DominatorTree::Delete, BB, DeadCaseBB});7718    I = SIW->removeCase(I);7719    E = SIW->case_end();7720  }7721 7722  auto Case = SI->findCaseValue(Constant);7723  // If the case value is not found, `findCaseValue` returns the default case.7724  // In this scenario, since there is no explicit `case 3:`, the simplification7725  // fails. The simplification also fails when the switch’s default destination7726  // is reachable.7727  if (!SI->defaultDestUnreachable() || Case == SI->case_default()) {7728    if (DTU)7729      DTU->applyUpdates(Updates);7730    return !Updates.empty();7731  }7732 7733  BasicBlock *Unreachable = SI->getDefaultDest();7734  SIW.replaceDefaultDest(Case);7735  SIW.removeCase(Case);7736  SIW->setCondition(A);7737 7738  Updates.push_back({DominatorTree::Delete, BB, Unreachable});7739 7740  if (DTU)7741    DTU->applyUpdates(Updates);7742 7743  return true;7744}7745 7746/// Tries to transform switch of powers of two to reduce switch range.7747/// For example, switch like:7748/// switch (C) { case 1: case 2: case 64: case 128: }7749/// will be transformed to:7750/// switch (count_trailing_zeros(C)) { case 0: case 1: case 6: case 7: }7751///7752/// This transformation allows better lowering and may transform the switch7753/// instruction into a sequence of bit manipulation and a smaller7754/// log2(C)-indexed value table (instead of traditionally emitting a load of the7755/// address of the jump target, and indirectly jump to it).7756static bool simplifySwitchOfPowersOfTwo(SwitchInst *SI, IRBuilder<> &Builder,7757                                        DomTreeUpdater *DTU,7758                                        const DataLayout &DL,7759                                        const TargetTransformInfo &TTI) {7760  Value *Condition = SI->getCondition();7761  LLVMContext &Context = SI->getContext();7762  auto *CondTy = cast<IntegerType>(Condition->getType());7763 7764  if (CondTy->getIntegerBitWidth() > 64 ||7765      !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))7766    return false;7767 7768  // Ensure trailing zeroes count intrinsic emission is not too expensive.7769  IntrinsicCostAttributes Attrs(Intrinsic::cttz, CondTy,7770                                {Condition, ConstantInt::getTrue(Context)});7771  if (TTI.getIntrinsicInstrCost(Attrs, TTI::TCK_SizeAndLatency) >7772      TTI::TCC_Basic * 2)7773    return false;7774 7775  // Only bother with this optimization if there are more than 3 switch cases.7776  // SDAG will start emitting jump tables for 4 or more cases.7777  if (SI->getNumCases() < 4)7778    return false;7779 7780  // Check that switch cases are powers of two.7781  SmallVector<uint64_t, 4> Values;7782  for (const auto &Case : SI->cases()) {7783    uint64_t CaseValue = Case.getCaseValue()->getValue().getZExtValue();7784    if (llvm::has_single_bit(CaseValue))7785      Values.push_back(CaseValue);7786    else7787      return false;7788  }7789 7790  // isSwichDense requires case values to be sorted.7791  llvm::sort(Values);7792  if (!isSwitchDense(Values.size(), llvm::countr_zero(Values.back()) -7793                                        llvm::countr_zero(Values.front()) + 1))7794    // Transform is unable to generate dense switch.7795    return false;7796 7797  Builder.SetInsertPoint(SI);7798 7799  if (!SI->defaultDestUnreachable()) {7800    // Let non-power-of-two inputs jump to the default case, when the latter is7801    // reachable.7802    auto *PopC = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, Condition);7803    auto *IsPow2 = Builder.CreateICmpEQ(PopC, ConstantInt::get(CondTy, 1));7804 7805    auto *OrigBB = SI->getParent();7806    auto *DefaultCaseBB = SI->getDefaultDest();7807    BasicBlock *SplitBB = SplitBlock(OrigBB, SI, DTU);7808    auto It = OrigBB->getTerminator()->getIterator();7809    SmallVector<uint32_t> Weights;7810    auto HasWeights =7811        !ProfcheckDisableMetadataFixes && extractBranchWeights(*SI, Weights);7812    auto *BI = BranchInst::Create(SplitBB, DefaultCaseBB, IsPow2, It);7813    if (HasWeights && any_of(Weights, [](const auto &V) { return V != 0; })) {7814      // IsPow2 covers a subset of the cases in which we'd go to the default7815      // label. The other is those powers of 2 that don't appear in the case7816      // statement. We don't know the distribution of the values coming in, so7817      // the safest is to split 50-50 the original probability to `default`.7818      uint64_t OrigDenominator =7819          sum_of(map_range(Weights, StaticCastTo<uint64_t>));7820      SmallVector<uint64_t> NewWeights(2);7821      NewWeights[1] = Weights[0] / 2;7822      NewWeights[0] = OrigDenominator - NewWeights[1];7823      setFittedBranchWeights(*BI, NewWeights, /*IsExpected=*/false);7824      // The probability of executing the default block stays constant. It was7825      //  p_d = Weights[0] / OrigDenominator7826      //  we rewrite as W/D7827      // We want to find the probability of the default branch of the switch7828      // statement. Let's call it X. We have W/D = W/2D + X * (1-W/2D)7829      // i.e. the original probability is the probability we go to the default7830      // branch from the BI branch, or we take the default branch on the SI.7831      // Meaning X = W / (2D - W), or (W/2) / (D - W/2)7832      // This matches using W/2 for the default branch probability numerator and7833      // D-W/2 as the denominator.7834      Weights[0] = NewWeights[1];7835      uint64_t CasesDenominator = OrigDenominator - Weights[0];7836      for (auto &W : drop_begin(Weights))7837        W = NewWeights[0] * static_cast<double>(W) / CasesDenominator;7838 7839      setBranchWeights(*SI, Weights, /*IsExpected=*/false);7840    }7841    // BI is handling the default case for SI, and so should share its DebugLoc.7842    BI->setDebugLoc(SI->getDebugLoc());7843    It->eraseFromParent();7844 7845    addPredecessorToBlock(DefaultCaseBB, OrigBB, SplitBB);7846    if (DTU)7847      DTU->applyUpdates({{DominatorTree::Insert, OrigBB, DefaultCaseBB}});7848  }7849 7850  // Replace each case with its trailing zeros number.7851  for (auto &Case : SI->cases()) {7852    auto *OrigValue = Case.getCaseValue();7853    Case.setValue(ConstantInt::get(OrigValue->getIntegerType(),7854                                   OrigValue->getValue().countr_zero()));7855  }7856 7857  // Replace condition with its trailing zeros number.7858  auto *ConditionTrailingZeros = Builder.CreateIntrinsic(7859      Intrinsic::cttz, {CondTy}, {Condition, ConstantInt::getTrue(Context)});7860 7861  SI->setCondition(ConditionTrailingZeros);7862 7863  return true;7864}7865 7866/// Fold switch over ucmp/scmp intrinsic to br if two of the switch arms have7867/// the same destination.7868static bool simplifySwitchOfCmpIntrinsic(SwitchInst *SI, IRBuilderBase &Builder,7869                                         DomTreeUpdater *DTU) {7870  auto *Cmp = dyn_cast<CmpIntrinsic>(SI->getCondition());7871  if (!Cmp || !Cmp->hasOneUse())7872    return false;7873 7874  SmallVector<uint32_t, 4> Weights;7875  bool HasWeights = extractBranchWeights(getBranchWeightMDNode(*SI), Weights);7876  if (!HasWeights)7877    Weights.resize(4); // Avoid checking HasWeights everywhere.7878 7879  // Normalize to [us]cmp == Res ? Succ : OtherSucc.7880  int64_t Res;7881  BasicBlock *Succ, *OtherSucc;7882  uint32_t SuccWeight = 0, OtherSuccWeight = 0;7883  BasicBlock *Unreachable = nullptr;7884 7885  if (SI->getNumCases() == 2) {7886    // Find which of 1, 0 or -1 is missing (handled by default dest).7887    SmallSet<int64_t, 3> Missing;7888    Missing.insert(1);7889    Missing.insert(0);7890    Missing.insert(-1);7891 7892    Succ = SI->getDefaultDest();7893    SuccWeight = Weights[0];7894    OtherSucc = nullptr;7895    for (auto &Case : SI->cases()) {7896      std::optional<int64_t> Val =7897          Case.getCaseValue()->getValue().trySExtValue();7898      if (!Val)7899        return false;7900      if (!Missing.erase(*Val))7901        return false;7902      if (OtherSucc && OtherSucc != Case.getCaseSuccessor())7903        return false;7904      OtherSucc = Case.getCaseSuccessor();7905      OtherSuccWeight += Weights[Case.getSuccessorIndex()];7906    }7907 7908    assert(Missing.size() == 1 && "Should have one case left");7909    Res = *Missing.begin();7910  } else if (SI->getNumCases() == 3 && SI->defaultDestUnreachable()) {7911    // Normalize so that Succ is taken once and OtherSucc twice.7912    Unreachable = SI->getDefaultDest();7913    Succ = OtherSucc = nullptr;7914    for (auto &Case : SI->cases()) {7915      BasicBlock *NewSucc = Case.getCaseSuccessor();7916      uint32_t Weight = Weights[Case.getSuccessorIndex()];7917      if (!OtherSucc || OtherSucc == NewSucc) {7918        OtherSucc = NewSucc;7919        OtherSuccWeight += Weight;7920      } else if (!Succ) {7921        Succ = NewSucc;7922        SuccWeight = Weight;7923      } else if (Succ == NewSucc) {7924        std::swap(Succ, OtherSucc);7925        std::swap(SuccWeight, OtherSuccWeight);7926      } else7927        return false;7928    }7929    for (auto &Case : SI->cases()) {7930      std::optional<int64_t> Val =7931          Case.getCaseValue()->getValue().trySExtValue();7932      if (!Val || (Val != 1 && Val != 0 && Val != -1))7933        return false;7934      if (Case.getCaseSuccessor() == Succ) {7935        Res = *Val;7936        break;7937      }7938    }7939  } else {7940    return false;7941  }7942 7943  // Determine predicate for the missing case.7944  ICmpInst::Predicate Pred;7945  switch (Res) {7946  case 1:7947    Pred = ICmpInst::ICMP_UGT;7948    break;7949  case 0:7950    Pred = ICmpInst::ICMP_EQ;7951    break;7952  case -1:7953    Pred = ICmpInst::ICMP_ULT;7954    break;7955  }7956  if (Cmp->isSigned())7957    Pred = ICmpInst::getSignedPredicate(Pred);7958 7959  MDNode *NewWeights = nullptr;7960  if (HasWeights)7961    NewWeights = MDBuilder(SI->getContext())7962                     .createBranchWeights(SuccWeight, OtherSuccWeight);7963 7964  BasicBlock *BB = SI->getParent();7965  Builder.SetInsertPoint(SI->getIterator());7966  Value *ICmp = Builder.CreateICmp(Pred, Cmp->getLHS(), Cmp->getRHS());7967  Builder.CreateCondBr(ICmp, Succ, OtherSucc, NewWeights,7968                       SI->getMetadata(LLVMContext::MD_unpredictable));7969  OtherSucc->removePredecessor(BB);7970  if (Unreachable)7971    Unreachable->removePredecessor(BB);7972  SI->eraseFromParent();7973  Cmp->eraseFromParent();7974  if (DTU && Unreachable)7975    DTU->applyUpdates({{DominatorTree::Delete, BB, Unreachable}});7976  return true;7977}7978 7979/// Checking whether two cases of SI are equal depends on the contents of the7980/// BasicBlock and the incoming values of their successor PHINodes.7981/// PHINode::getIncomingValueForBlock is O(|Preds|), so we'd like to avoid7982/// calling this function on each BasicBlock every time isEqual is called,7983/// especially since the same BasicBlock may be passed as an argument multiple7984/// times. To do this, we can precompute a map of PHINode -> Pred BasicBlock ->7985/// IncomingValue and add it in the Wrapper so isEqual can do O(1) checking7986/// of the incoming values.7987struct SwitchSuccWrapper {7988  BasicBlock *Dest;7989  DenseMap<PHINode *, SmallDenseMap<BasicBlock *, Value *, 8>> *PhiPredIVs;7990};7991 7992template <> struct llvm::DenseMapInfo<const SwitchSuccWrapper *> {7993  static const SwitchSuccWrapper *getEmptyKey() {7994    return static_cast<SwitchSuccWrapper *>(7995        DenseMapInfo<void *>::getEmptyKey());7996  }7997  static const SwitchSuccWrapper *getTombstoneKey() {7998    return static_cast<SwitchSuccWrapper *>(7999        DenseMapInfo<void *>::getTombstoneKey());8000  }8001  static unsigned getHashValue(const SwitchSuccWrapper *SSW) {8002    BasicBlock *Succ = SSW->Dest;8003    BranchInst *BI = cast<BranchInst>(Succ->getTerminator());8004    assert(BI->isUnconditional() &&8005           "Only supporting unconditional branches for now");8006    assert(BI->getNumSuccessors() == 1 &&8007           "Expected unconditional branches to have one successor");8008    assert(Succ->size() == 1 && "Expected just a single branch in the BB");8009 8010    // Since we assume the BB is just a single BranchInst with a single8011    // successor, we hash as the BB and the incoming Values of its successor8012    // PHIs. Initially, we tried to just use the successor BB as the hash, but8013    // including the incoming PHI values leads to better performance.8014    // We also tried to build a map from BB -> Succs.IncomingValues ahead of8015    // time and passing it in SwitchSuccWrapper, but this slowed down the8016    // average compile time without having any impact on the worst case compile8017    // time.8018    BasicBlock *BB = BI->getSuccessor(0);8019    SmallVector<Value *> PhiValsForBB;8020    for (PHINode &Phi : BB->phis())8021      PhiValsForBB.emplace_back((*SSW->PhiPredIVs)[&Phi][BB]);8022 8023    return hash_combine(BB, hash_combine_range(PhiValsForBB));8024  }8025  static bool isEqual(const SwitchSuccWrapper *LHS,8026                      const SwitchSuccWrapper *RHS) {8027    auto EKey = DenseMapInfo<SwitchSuccWrapper *>::getEmptyKey();8028    auto TKey = DenseMapInfo<SwitchSuccWrapper *>::getTombstoneKey();8029    if (LHS == EKey || RHS == EKey || LHS == TKey || RHS == TKey)8030      return LHS == RHS;8031 8032    BasicBlock *A = LHS->Dest;8033    BasicBlock *B = RHS->Dest;8034 8035    // FIXME: we checked that the size of A and B are both 1 in8036    // simplifyDuplicateSwitchArms to make the Case list smaller to8037    // improve performance. If we decide to support BasicBlocks with more8038    // than just a single instruction, we need to check that A.size() ==8039    // B.size() here, and we need to check more than just the BranchInsts8040    // for equality.8041 8042    BranchInst *ABI = cast<BranchInst>(A->getTerminator());8043    BranchInst *BBI = cast<BranchInst>(B->getTerminator());8044    assert(ABI->isUnconditional() && BBI->isUnconditional() &&8045           "Only supporting unconditional branches for now");8046    if (ABI->getSuccessor(0) != BBI->getSuccessor(0))8047      return false;8048 8049    // Need to check that PHIs in successor have matching values8050    BasicBlock *Succ = ABI->getSuccessor(0);8051    for (PHINode &Phi : Succ->phis()) {8052      auto &PredIVs = (*LHS->PhiPredIVs)[&Phi];8053      if (PredIVs[A] != PredIVs[B])8054        return false;8055    }8056 8057    return true;8058  }8059};8060 8061bool SimplifyCFGOpt::simplifyDuplicateSwitchArms(SwitchInst *SI,8062                                                 DomTreeUpdater *DTU) {8063  // Build Cases. Skip BBs that are not candidates for simplification. Mark8064  // PHINodes which need to be processed into PhiPredIVs. We decide to process8065  // an entire PHI at once after the loop, opposed to calling8066  // getIncomingValueForBlock inside this loop, since each call to8067  // getIncomingValueForBlock is O(|Preds|).8068  SmallPtrSet<PHINode *, 8> Phis;8069  SmallPtrSet<BasicBlock *, 8> Seen;8070  DenseMap<PHINode *, SmallDenseMap<BasicBlock *, Value *, 8>> PhiPredIVs;8071  DenseMap<BasicBlock *, SmallVector<unsigned, 32>> BBToSuccessorIndexes;8072  SmallVector<SwitchSuccWrapper> Cases;8073  Cases.reserve(SI->getNumSuccessors());8074 8075  for (unsigned I = 0; I < SI->getNumSuccessors(); ++I) {8076    BasicBlock *BB = SI->getSuccessor(I);8077 8078    // FIXME: Support more than just a single BranchInst. One way we could do8079    // this is by taking a hashing approach of all insts in BB.8080    if (BB->size() != 1)8081      continue;8082 8083    // FIXME: Relax that the terminator is a BranchInst by checking for equality8084    // on other kinds of terminators. We decide to only support unconditional8085    // branches for now for compile time reasons.8086    auto *BI = dyn_cast<BranchInst>(BB->getTerminator());8087    if (!BI || BI->isConditional())8088      continue;8089 8090    if (!Seen.insert(BB).second) {8091      auto It = BBToSuccessorIndexes.find(BB);8092      if (It != BBToSuccessorIndexes.end())8093        It->second.emplace_back(I);8094      continue;8095    }8096 8097    // FIXME: This case needs some extra care because the terminators other than8098    // SI need to be updated. For now, consider only backedges to the SI.8099    if (BB->getUniquePredecessor() != SI->getParent())8100      continue;8101 8102    // Keep track of which PHIs we need as keys in PhiPredIVs below.8103    for (BasicBlock *Succ : BI->successors())8104      Phis.insert_range(llvm::make_pointer_range(Succ->phis()));8105 8106    // Add the successor only if not previously visited.8107    Cases.emplace_back(SwitchSuccWrapper{BB, &PhiPredIVs});8108    BBToSuccessorIndexes[BB].emplace_back(I);8109  }8110 8111  // Precompute a data structure to improve performance of isEqual for8112  // SwitchSuccWrapper.8113  PhiPredIVs.reserve(Phis.size());8114  for (PHINode *Phi : Phis) {8115    auto &IVs =8116        PhiPredIVs.try_emplace(Phi, Phi->getNumIncomingValues()).first->second;8117    for (auto &IV : Phi->incoming_values())8118      IVs.insert({Phi->getIncomingBlock(IV), IV.get()});8119  }8120 8121  // Build a set such that if the SwitchSuccWrapper exists in the set and8122  // another SwitchSuccWrapper isEqual, then the equivalent SwitchSuccWrapper8123  // which is not in the set should be replaced with the one in the set. If the8124  // SwitchSuccWrapper is not in the set, then it should be added to the set so8125  // other SwitchSuccWrappers can check against it in the same manner. We use8126  // SwitchSuccWrapper instead of just BasicBlock because we'd like to pass8127  // around information to isEquality, getHashValue, and when doing the8128  // replacement with better performance.8129  DenseSet<const SwitchSuccWrapper *> ReplaceWith;8130  ReplaceWith.reserve(Cases.size());8131 8132  SmallVector<DominatorTree::UpdateType> Updates;8133  Updates.reserve(ReplaceWith.size());8134  bool MadeChange = false;8135  for (auto &SSW : Cases) {8136    // SSW is a candidate for simplification. If we find a duplicate BB,8137    // replace it.8138    const auto [It, Inserted] = ReplaceWith.insert(&SSW);8139    if (!Inserted) {8140      // We know that SI's parent BB no longer dominates the old case successor8141      // since we are making it dead.8142      Updates.push_back({DominatorTree::Delete, SI->getParent(), SSW.Dest});8143      const auto &Successors = BBToSuccessorIndexes.at(SSW.Dest);8144      for (unsigned Idx : Successors)8145        SI->setSuccessor(Idx, (*It)->Dest);8146      MadeChange = true;8147    }8148  }8149 8150  if (DTU)8151    DTU->applyUpdates(Updates);8152 8153  return MadeChange;8154}8155 8156bool SimplifyCFGOpt::simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {8157  BasicBlock *BB = SI->getParent();8158 8159  if (isValueEqualityComparison(SI)) {8160    // If we only have one predecessor, and if it is a branch on this value,8161    // see if that predecessor totally determines the outcome of this switch.8162    if (BasicBlock *OnlyPred = BB->getSinglePredecessor())8163      if (simplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))8164        return requestResimplify();8165 8166    Value *Cond = SI->getCondition();8167    if (SelectInst *Select = dyn_cast<SelectInst>(Cond))8168      if (simplifySwitchOnSelect(SI, Select))8169        return requestResimplify();8170 8171    // If the block only contains the switch, see if we can fold the block8172    // away into any preds.8173    if (SI == &*BB->instructionsWithoutDebug(false).begin())8174      if (foldValueComparisonIntoPredecessors(SI, Builder))8175        return requestResimplify();8176  }8177 8178  // Try to transform the switch into an icmp and a branch.8179  // The conversion from switch to comparison may lose information on8180  // impossible switch values, so disable it early in the pipeline.8181  if (Options.ConvertSwitchRangeToICmp && turnSwitchRangeIntoICmp(SI, Builder))8182    return requestResimplify();8183 8184  // Remove unreachable cases.8185  if (eliminateDeadSwitchCases(SI, DTU, Options.AC, DL))8186    return requestResimplify();8187 8188  if (simplifySwitchOfCmpIntrinsic(SI, Builder, DTU))8189    return requestResimplify();8190 8191  if (trySwitchToSelect(SI, Builder, DTU, DL, TTI))8192    return requestResimplify();8193 8194  if (Options.ForwardSwitchCondToPhi && forwardSwitchConditionToPHI(SI))8195    return requestResimplify();8196 8197  // The conversion of switches to arithmetic or lookup table is disabled in8198  // the early optimization pipeline, as it may lose information or make the8199  // resulting code harder to analyze.8200  if (Options.ConvertSwitchToArithmetic || Options.ConvertSwitchToLookupTable)8201    if (simplifySwitchLookup(SI, Builder, DTU, DL, TTI,8202                             Options.ConvertSwitchToLookupTable))8203      return requestResimplify();8204 8205  if (simplifySwitchOfPowersOfTwo(SI, Builder, DTU, DL, TTI))8206    return requestResimplify();8207 8208  if (reduceSwitchRange(SI, Builder, DL, TTI))8209    return requestResimplify();8210 8211  if (HoistCommon &&8212      hoistCommonCodeFromSuccessors(SI, !Options.HoistCommonInsts))8213    return requestResimplify();8214 8215  if (simplifyDuplicateSwitchArms(SI, DTU))8216    return requestResimplify();8217 8218  if (simplifySwitchWhenUMin(SI, DTU))8219    return requestResimplify();8220 8221  return false;8222}8223 8224bool SimplifyCFGOpt::simplifyIndirectBr(IndirectBrInst *IBI) {8225  BasicBlock *BB = IBI->getParent();8226  bool Changed = false;8227  SmallVector<uint32_t> BranchWeights;8228  const bool HasBranchWeights = !ProfcheckDisableMetadataFixes &&8229                                extractBranchWeights(*IBI, BranchWeights);8230 8231  DenseMap<const BasicBlock *, uint64_t> TargetWeight;8232  if (HasBranchWeights)8233    for (size_t I = 0, E = IBI->getNumDestinations(); I < E; ++I)8234      TargetWeight[IBI->getDestination(I)] += BranchWeights[I];8235 8236  // Eliminate redundant destinations.8237  SmallPtrSet<Value *, 8> Succs;8238  SmallSetVector<BasicBlock *, 8> RemovedSuccs;8239  for (unsigned I = 0, E = IBI->getNumDestinations(); I != E; ++I) {8240    BasicBlock *Dest = IBI->getDestination(I);8241    if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {8242      if (!Dest->hasAddressTaken())8243        RemovedSuccs.insert(Dest);8244      Dest->removePredecessor(BB);8245      IBI->removeDestination(I);8246      --I;8247      --E;8248      Changed = true;8249    }8250  }8251 8252  if (DTU) {8253    std::vector<DominatorTree::UpdateType> Updates;8254    Updates.reserve(RemovedSuccs.size());8255    for (auto *RemovedSucc : RemovedSuccs)8256      Updates.push_back({DominatorTree::Delete, BB, RemovedSucc});8257    DTU->applyUpdates(Updates);8258  }8259 8260  if (IBI->getNumDestinations() == 0) {8261    // If the indirectbr has no successors, change it to unreachable.8262    new UnreachableInst(IBI->getContext(), IBI->getIterator());8263    eraseTerminatorAndDCECond(IBI);8264    return true;8265  }8266 8267  if (IBI->getNumDestinations() == 1) {8268    // If the indirectbr has one successor, change it to a direct branch.8269    BranchInst::Create(IBI->getDestination(0), IBI->getIterator());8270    eraseTerminatorAndDCECond(IBI);8271    return true;8272  }8273  if (HasBranchWeights) {8274    SmallVector<uint64_t> NewBranchWeights(IBI->getNumDestinations());8275    for (size_t I = 0, E = IBI->getNumDestinations(); I < E; ++I)8276      NewBranchWeights[I] += TargetWeight.find(IBI->getDestination(I))->second;8277    setFittedBranchWeights(*IBI, NewBranchWeights, /*IsExpected=*/false);8278  }8279  if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {8280    if (simplifyIndirectBrOnSelect(IBI, SI))8281      return requestResimplify();8282  }8283  return Changed;8284}8285 8286/// Given an block with only a single landing pad and a unconditional branch8287/// try to find another basic block which this one can be merged with.  This8288/// handles cases where we have multiple invokes with unique landing pads, but8289/// a shared handler.8290///8291/// We specifically choose to not worry about merging non-empty blocks8292/// here.  That is a PRE/scheduling problem and is best solved elsewhere.  In8293/// practice, the optimizer produces empty landing pad blocks quite frequently8294/// when dealing with exception dense code.  (see: instcombine, gvn, if-else8295/// sinking in this file)8296///8297/// This is primarily a code size optimization.  We need to avoid performing8298/// any transform which might inhibit optimization (such as our ability to8299/// specialize a particular handler via tail commoning).  We do this by not8300/// merging any blocks which require us to introduce a phi.  Since the same8301/// values are flowing through both blocks, we don't lose any ability to8302/// specialize.  If anything, we make such specialization more likely.8303///8304/// TODO - This transformation could remove entries from a phi in the target8305/// block when the inputs in the phi are the same for the two blocks being8306/// merged.  In some cases, this could result in removal of the PHI entirely.8307static bool tryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,8308                                 BasicBlock *BB, DomTreeUpdater *DTU) {8309  auto Succ = BB->getUniqueSuccessor();8310  assert(Succ);8311  // If there's a phi in the successor block, we'd likely have to introduce8312  // a phi into the merged landing pad block.8313  if (isa<PHINode>(*Succ->begin()))8314    return false;8315 8316  for (BasicBlock *OtherPred : predecessors(Succ)) {8317    if (BB == OtherPred)8318      continue;8319    BasicBlock::iterator I = OtherPred->begin();8320    LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);8321    if (!LPad2 || !LPad2->isIdenticalTo(LPad))8322      continue;8323    ++I;8324    BranchInst *BI2 = dyn_cast<BranchInst>(I);8325    if (!BI2 || !BI2->isIdenticalTo(BI))8326      continue;8327 8328    std::vector<DominatorTree::UpdateType> Updates;8329 8330    // We've found an identical block.  Update our predecessors to take that8331    // path instead and make ourselves dead.8332    SmallSetVector<BasicBlock *, 16> UniquePreds(pred_begin(BB), pred_end(BB));8333    for (BasicBlock *Pred : UniquePreds) {8334      InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());8335      assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&8336             "unexpected successor");8337      II->setUnwindDest(OtherPred);8338      if (DTU) {8339        Updates.push_back({DominatorTree::Insert, Pred, OtherPred});8340        Updates.push_back({DominatorTree::Delete, Pred, BB});8341      }8342    }8343 8344    SmallSetVector<BasicBlock *, 16> UniqueSuccs(succ_begin(BB), succ_end(BB));8345    for (BasicBlock *Succ : UniqueSuccs) {8346      Succ->removePredecessor(BB);8347      if (DTU)8348        Updates.push_back({DominatorTree::Delete, BB, Succ});8349    }8350 8351    IRBuilder<> Builder(BI);8352    Builder.CreateUnreachable();8353    BI->eraseFromParent();8354    if (DTU)8355      DTU->applyUpdates(Updates);8356    return true;8357  }8358  return false;8359}8360 8361bool SimplifyCFGOpt::simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder) {8362  return Branch->isUnconditional() ? simplifyUncondBranch(Branch, Builder)8363                                   : simplifyCondBranch(Branch, Builder);8364}8365 8366bool SimplifyCFGOpt::simplifyUncondBranch(BranchInst *BI,8367                                          IRBuilder<> &Builder) {8368  BasicBlock *BB = BI->getParent();8369  BasicBlock *Succ = BI->getSuccessor(0);8370 8371  // If the Terminator is the only non-phi instruction, simplify the block.8372  // If LoopHeader is provided, check if the block or its successor is a loop8373  // header. (This is for early invocations before loop simplify and8374  // vectorization to keep canonical loop forms for nested loops. These blocks8375  // can be eliminated when the pass is invoked later in the back-end.)8376  // Note that if BB has only one predecessor then we do not introduce new8377  // backedge, so we can eliminate BB.8378  bool NeedCanonicalLoop =8379      Options.NeedCanonicalLoop &&8380      (!LoopHeaders.empty() && BB->hasNPredecessorsOrMore(2) &&8381       (is_contained(LoopHeaders, BB) || is_contained(LoopHeaders, Succ)));8382  BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();8383  if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&8384      !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB, DTU))8385    return true;8386 8387  // If the only instruction in the block is a seteq/setne comparison against a8388  // constant, try to simplify the block.8389  if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {8390    if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {8391      ++I;8392      if (I->isTerminator() &&8393          tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))8394        return true;8395      if (isa<SelectInst>(I) && I->getNextNode()->isTerminator() &&8396          tryToSimplifyUncondBranchWithICmpSelectInIt(ICI, cast<SelectInst>(I),8397                                                      Builder))8398        return true;8399    }8400  }8401 8402  // See if we can merge an empty landing pad block with another which is8403  // equivalent.8404  if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {8405    ++I;8406    if (I->isTerminator() && tryToMergeLandingPad(LPad, BI, BB, DTU))8407      return true;8408  }8409 8410  // If this basic block is ONLY a compare and a branch, and if a predecessor8411  // branches to us and our successor, fold the comparison into the8412  // predecessor and use logical operations to update the incoming value8413  // for PHI nodes in common successor.8414  if (Options.SpeculateBlocks &&8415      foldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI,8416                             Options.BonusInstThreshold))8417    return requestResimplify();8418  return false;8419}8420 8421static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {8422  BasicBlock *PredPred = nullptr;8423  for (auto *P : predecessors(BB)) {8424    BasicBlock *PPred = P->getSinglePredecessor();8425    if (!PPred || (PredPred && PredPred != PPred))8426      return nullptr;8427    PredPred = PPred;8428  }8429  return PredPred;8430}8431 8432/// Fold the following pattern:8433/// bb0:8434///   br i1 %cond1, label %bb1, label %bb28435/// bb1:8436///   br i1 %cond2, label %bb3, label %bb48437/// bb2:8438///   br i1 %cond2, label %bb4, label %bb38439/// bb3:8440///   ...8441/// bb4:8442///   ...8443/// into8444/// bb0:8445///   %cond = xor i1 %cond1, %cond28446///   br i1 %cond, label %bb4, label %bb38447/// bb3:8448///   ...8449/// bb4:8450///   ...8451/// NOTE: %cond2 always dominates the terminator of bb0.8452static bool mergeNestedCondBranch(BranchInst *BI, DomTreeUpdater *DTU) {8453  BasicBlock *BB = BI->getParent();8454  BasicBlock *BB1 = BI->getSuccessor(0);8455  BasicBlock *BB2 = BI->getSuccessor(1);8456  auto IsSimpleSuccessor = [BB](BasicBlock *Succ, BranchInst *&SuccBI) {8457    if (Succ == BB)8458      return false;8459    if (&Succ->front() != Succ->getTerminator())8460      return false;8461    SuccBI = dyn_cast<BranchInst>(Succ->getTerminator());8462    if (!SuccBI || !SuccBI->isConditional())8463      return false;8464    BasicBlock *Succ1 = SuccBI->getSuccessor(0);8465    BasicBlock *Succ2 = SuccBI->getSuccessor(1);8466    return Succ1 != Succ && Succ2 != Succ && Succ1 != BB && Succ2 != BB &&8467           !isa<PHINode>(Succ1->front()) && !isa<PHINode>(Succ2->front());8468  };8469  BranchInst *BB1BI, *BB2BI;8470  if (!IsSimpleSuccessor(BB1, BB1BI) || !IsSimpleSuccessor(BB2, BB2BI))8471    return false;8472 8473  if (BB1BI->getCondition() != BB2BI->getCondition() ||8474      BB1BI->getSuccessor(0) != BB2BI->getSuccessor(1) ||8475      BB1BI->getSuccessor(1) != BB2BI->getSuccessor(0))8476    return false;8477 8478  BasicBlock *BB3 = BB1BI->getSuccessor(0);8479  BasicBlock *BB4 = BB1BI->getSuccessor(1);8480  IRBuilder<> Builder(BI);8481  BI->setCondition(8482      Builder.CreateXor(BI->getCondition(), BB1BI->getCondition()));8483  BB1->removePredecessor(BB);8484  BI->setSuccessor(0, BB4);8485  BB2->removePredecessor(BB);8486  BI->setSuccessor(1, BB3);8487  if (DTU) {8488    SmallVector<DominatorTree::UpdateType, 4> Updates;8489    Updates.push_back({DominatorTree::Delete, BB, BB1});8490    Updates.push_back({DominatorTree::Insert, BB, BB4});8491    Updates.push_back({DominatorTree::Delete, BB, BB2});8492    Updates.push_back({DominatorTree::Insert, BB, BB3});8493 8494    DTU->applyUpdates(Updates);8495  }8496  bool HasWeight = false;8497  uint64_t BBTWeight, BBFWeight;8498  if (extractBranchWeights(*BI, BBTWeight, BBFWeight))8499    HasWeight = true;8500  else8501    BBTWeight = BBFWeight = 1;8502  uint64_t BB1TWeight, BB1FWeight;8503  if (extractBranchWeights(*BB1BI, BB1TWeight, BB1FWeight))8504    HasWeight = true;8505  else8506    BB1TWeight = BB1FWeight = 1;8507  uint64_t BB2TWeight, BB2FWeight;8508  if (extractBranchWeights(*BB2BI, BB2TWeight, BB2FWeight))8509    HasWeight = true;8510  else8511    BB2TWeight = BB2FWeight = 1;8512  if (HasWeight) {8513    uint64_t Weights[2] = {BBTWeight * BB1FWeight + BBFWeight * BB2TWeight,8514                           BBTWeight * BB1TWeight + BBFWeight * BB2FWeight};8515    setFittedBranchWeights(*BI, Weights, /*IsExpected=*/false,8516                           /*ElideAllZero=*/true);8517  }8518  return true;8519}8520 8521bool SimplifyCFGOpt::simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {8522  assert(8523      !isa<ConstantInt>(BI->getCondition()) &&8524      BI->getSuccessor(0) != BI->getSuccessor(1) &&8525      "Tautological conditional branch should have been eliminated already.");8526 8527  BasicBlock *BB = BI->getParent();8528  if (!Options.SimplifyCondBranch ||8529      BI->getFunction()->hasFnAttribute(Attribute::OptForFuzzing))8530    return false;8531 8532  // Conditional branch8533  if (isValueEqualityComparison(BI)) {8534    // If we only have one predecessor, and if it is a branch on this value,8535    // see if that predecessor totally determines the outcome of this8536    // switch.8537    if (BasicBlock *OnlyPred = BB->getSinglePredecessor())8538      if (simplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))8539        return requestResimplify();8540 8541    // This block must be empty, except for the setcond inst, if it exists.8542    // Ignore dbg and pseudo intrinsics.8543    auto I = BB->instructionsWithoutDebug(true).begin();8544    if (&*I == BI) {8545      if (foldValueComparisonIntoPredecessors(BI, Builder))8546        return requestResimplify();8547    } else if (&*I == cast<Instruction>(BI->getCondition())) {8548      ++I;8549      if (&*I == BI && foldValueComparisonIntoPredecessors(BI, Builder))8550        return requestResimplify();8551    }8552  }8553 8554  // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.8555  if (simplifyBranchOnICmpChain(BI, Builder, DL))8556    return true;8557 8558  // If this basic block has dominating predecessor blocks and the dominating8559  // blocks' conditions imply BI's condition, we know the direction of BI.8560  std::optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);8561  if (Imp) {8562    // Turn this into a branch on constant.8563    auto *OldCond = BI->getCondition();8564    ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())8565                             : ConstantInt::getFalse(BB->getContext());8566    BI->setCondition(TorF);8567    RecursivelyDeleteTriviallyDeadInstructions(OldCond);8568    return requestResimplify();8569  }8570 8571  // If this basic block is ONLY a compare and a branch, and if a predecessor8572  // branches to us and one of our successors, fold the comparison into the8573  // predecessor and use logical operations to pick the right destination.8574  if (Options.SpeculateBlocks &&8575      foldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI,8576                             Options.BonusInstThreshold))8577    return requestResimplify();8578 8579  // We have a conditional branch to two blocks that are only reachable8580  // from BI.  We know that the condbr dominates the two blocks, so see if8581  // there is any identical code in the "then" and "else" blocks.  If so, we8582  // can hoist it up to the branching block.8583  if (BI->getSuccessor(0)->getSinglePredecessor()) {8584    if (BI->getSuccessor(1)->getSinglePredecessor()) {8585      if (HoistCommon &&8586          hoistCommonCodeFromSuccessors(BI, !Options.HoistCommonInsts))8587        return requestResimplify();8588 8589      if (BI && Options.HoistLoadsStoresWithCondFaulting &&8590          isProfitableToSpeculate(BI, std::nullopt, TTI)) {8591        SmallVector<Instruction *, 2> SpeculatedConditionalLoadsStores;8592        auto CanSpeculateConditionalLoadsStores = [&]() {8593          for (auto *Succ : successors(BB)) {8594            for (Instruction &I : *Succ) {8595              if (I.isTerminator()) {8596                if (I.getNumSuccessors() > 1)8597                  return false;8598                continue;8599              } else if (!isSafeCheapLoadStore(&I, TTI) ||8600                         SpeculatedConditionalLoadsStores.size() ==8601                             HoistLoadsStoresWithCondFaultingThreshold) {8602                return false;8603              }8604              SpeculatedConditionalLoadsStores.push_back(&I);8605            }8606          }8607          return !SpeculatedConditionalLoadsStores.empty();8608        };8609 8610        if (CanSpeculateConditionalLoadsStores()) {8611          hoistConditionalLoadsStores(BI, SpeculatedConditionalLoadsStores,8612                                      std::nullopt, nullptr);8613          return requestResimplify();8614        }8615      }8616    } else {8617      // If Successor #1 has multiple preds, we may be able to conditionally8618      // execute Successor #0 if it branches to Successor #1.8619      Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();8620      if (Succ0TI->getNumSuccessors() == 1 &&8621          Succ0TI->getSuccessor(0) == BI->getSuccessor(1))8622        if (speculativelyExecuteBB(BI, BI->getSuccessor(0)))8623          return requestResimplify();8624    }8625  } else if (BI->getSuccessor(1)->getSinglePredecessor()) {8626    // If Successor #0 has multiple preds, we may be able to conditionally8627    // execute Successor #1 if it branches to Successor #0.8628    Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();8629    if (Succ1TI->getNumSuccessors() == 1 &&8630        Succ1TI->getSuccessor(0) == BI->getSuccessor(0))8631      if (speculativelyExecuteBB(BI, BI->getSuccessor(1)))8632        return requestResimplify();8633  }8634 8635  // If this is a branch on something for which we know the constant value in8636  // predecessors (e.g. a phi node in the current block), thread control8637  // through this block.8638  if (foldCondBranchOnValueKnownInPredecessor(BI))8639    return requestResimplify();8640 8641  // Scan predecessor blocks for conditional branches.8642  for (BasicBlock *Pred : predecessors(BB))8643    if (BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator()))8644      if (PBI != BI && PBI->isConditional())8645        if (SimplifyCondBranchToCondBranch(PBI, BI, DTU, DL, TTI))8646          return requestResimplify();8647 8648  // Look for diamond patterns.8649  if (MergeCondStores)8650    if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))8651      if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))8652        if (PBI != BI && PBI->isConditional())8653          if (mergeConditionalStores(PBI, BI, DTU, DL, TTI))8654            return requestResimplify();8655 8656  // Look for nested conditional branches.8657  if (mergeNestedCondBranch(BI, DTU))8658    return requestResimplify();8659 8660  return false;8661}8662 8663/// Check if passing a value to an instruction will cause undefined behavior.8664static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified) {8665  assert(V->getType() == I->getType() && "Mismatched types");8666  Constant *C = dyn_cast<Constant>(V);8667  if (!C)8668    return false;8669 8670  if (I->use_empty())8671    return false;8672 8673  if (C->isNullValue() || isa<UndefValue>(C)) {8674    // Only look at the first use we can handle, avoid hurting compile time with8675    // long uselists8676    auto FindUse = llvm::find_if(I->uses(), [](auto &U) {8677      auto *Use = cast<Instruction>(U.getUser());8678      // Change this list when we want to add new instructions.8679      switch (Use->getOpcode()) {8680      default:8681        return false;8682      case Instruction::GetElementPtr:8683      case Instruction::Ret:8684      case Instruction::BitCast:8685      case Instruction::Load:8686      case Instruction::Store:8687      case Instruction::Call:8688      case Instruction::CallBr:8689      case Instruction::Invoke:8690      case Instruction::UDiv:8691      case Instruction::URem:8692        // Note: signed div/rem of INT_MIN / -1 is also immediate UB, not8693        // implemented to avoid code complexity as it is unclear how useful such8694        // logic is.8695      case Instruction::SDiv:8696      case Instruction::SRem:8697        return true;8698      }8699    });8700    if (FindUse == I->use_end())8701      return false;8702    auto &Use = *FindUse;8703    auto *User = cast<Instruction>(Use.getUser());8704    // Bail out if User is not in the same BB as I or User == I or User comes8705    // before I in the block. The latter two can be the case if User is a8706    // PHI node.8707    if (User->getParent() != I->getParent() || User == I ||8708        User->comesBefore(I))8709      return false;8710 8711    // Now make sure that there are no instructions in between that can alter8712    // control flow (eg. calls)8713    auto InstrRange =8714        make_range(std::next(I->getIterator()), User->getIterator());8715    if (any_of(InstrRange, [](Instruction &I) {8716          return !isGuaranteedToTransferExecutionToSuccessor(&I);8717        }))8718      return false;8719 8720    // Look through GEPs. A load from a GEP derived from NULL is still undefined8721    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User))8722      if (GEP->getPointerOperand() == I) {8723        // The type of GEP may differ from the type of base pointer.8724        // Bail out on vector GEPs, as they are not handled by other checks.8725        if (GEP->getType()->isVectorTy())8726          return false;8727        // The current base address is null, there are four cases to consider:8728        // getelementptr (TY, null, 0)                 -> null8729        // getelementptr (TY, null, not zero)          -> may be modified8730        // getelementptr inbounds (TY, null, 0)        -> null8731        // getelementptr inbounds (TY, null, not zero) -> poison iff null is8732        // undefined?8733        if (!GEP->hasAllZeroIndices() &&8734            (!GEP->isInBounds() ||8735             NullPointerIsDefined(GEP->getFunction(),8736                                  GEP->getPointerAddressSpace())))8737          PtrValueMayBeModified = true;8738        return passingValueIsAlwaysUndefined(V, GEP, PtrValueMayBeModified);8739      }8740 8741    // Look through return.8742    if (ReturnInst *Ret = dyn_cast<ReturnInst>(User)) {8743      bool HasNoUndefAttr =8744          Ret->getFunction()->hasRetAttribute(Attribute::NoUndef);8745      // Return undefined to a noundef return value is undefined.8746      if (isa<UndefValue>(C) && HasNoUndefAttr)8747        return true;8748      // Return null to a nonnull+noundef return value is undefined.8749      if (C->isNullValue() && HasNoUndefAttr &&8750          Ret->getFunction()->hasRetAttribute(Attribute::NonNull)) {8751        return !PtrValueMayBeModified;8752      }8753    }8754 8755    // Load from null is undefined.8756    if (LoadInst *LI = dyn_cast<LoadInst>(User))8757      if (!LI->isVolatile())8758        return !NullPointerIsDefined(LI->getFunction(),8759                                     LI->getPointerAddressSpace());8760 8761    // Store to null is undefined.8762    if (StoreInst *SI = dyn_cast<StoreInst>(User))8763      if (!SI->isVolatile())8764        return (!NullPointerIsDefined(SI->getFunction(),8765                                      SI->getPointerAddressSpace())) &&8766               SI->getPointerOperand() == I;8767 8768    // llvm.assume(false/undef) always triggers immediate UB.8769    if (auto *Assume = dyn_cast<AssumeInst>(User)) {8770      // Ignore assume operand bundles.8771      if (I == Assume->getArgOperand(0))8772        return true;8773    }8774 8775    if (auto *CB = dyn_cast<CallBase>(User)) {8776      if (C->isNullValue() && NullPointerIsDefined(CB->getFunction()))8777        return false;8778      // A call to null is undefined.8779      if (CB->getCalledOperand() == I)8780        return true;8781 8782      if (CB->isArgOperand(&Use)) {8783        unsigned ArgIdx = CB->getArgOperandNo(&Use);8784        // Passing null to a nonnnull+noundef argument is undefined.8785        if (isa<ConstantPointerNull>(C) &&8786            CB->paramHasNonNullAttr(ArgIdx, /*AllowUndefOrPoison=*/false))8787          return !PtrValueMayBeModified;8788        // Passing undef to a noundef argument is undefined.8789        if (isa<UndefValue>(C) && CB->isPassingUndefUB(ArgIdx))8790          return true;8791      }8792    }8793    // Div/Rem by zero is immediate UB8794    if (match(User, m_BinOp(m_Value(), m_Specific(I))) && User->isIntDivRem())8795      return true;8796  }8797  return false;8798}8799 8800/// If BB has an incoming value that will always trigger undefined behavior8801/// (eg. null pointer dereference), remove the branch leading here.8802static bool removeUndefIntroducingPredecessor(BasicBlock *BB,8803                                              DomTreeUpdater *DTU,8804                                              AssumptionCache *AC) {8805  for (PHINode &PHI : BB->phis())8806    for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)8807      if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {8808        BasicBlock *Predecessor = PHI.getIncomingBlock(i);8809        Instruction *T = Predecessor->getTerminator();8810        IRBuilder<> Builder(T);8811        if (BranchInst *BI = dyn_cast<BranchInst>(T)) {8812          BB->removePredecessor(Predecessor);8813          // Turn unconditional branches into unreachables and remove the dead8814          // destination from conditional branches.8815          if (BI->isUnconditional())8816            Builder.CreateUnreachable();8817          else {8818            // Preserve guarding condition in assume, because it might not be8819            // inferrable from any dominating condition.8820            Value *Cond = BI->getCondition();8821            CallInst *Assumption;8822            if (BI->getSuccessor(0) == BB)8823              Assumption = Builder.CreateAssumption(Builder.CreateNot(Cond));8824            else8825              Assumption = Builder.CreateAssumption(Cond);8826            if (AC)8827              AC->registerAssumption(cast<AssumeInst>(Assumption));8828            Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)8829                                                       : BI->getSuccessor(0));8830          }8831          BI->eraseFromParent();8832          if (DTU)8833            DTU->applyUpdates({{DominatorTree::Delete, Predecessor, BB}});8834          return true;8835        } else if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) {8836          // Redirect all branches leading to UB into8837          // a newly created unreachable block.8838          BasicBlock *Unreachable = BasicBlock::Create(8839              Predecessor->getContext(), "unreachable", BB->getParent(), BB);8840          Builder.SetInsertPoint(Unreachable);8841          // The new block contains only one instruction: Unreachable8842          Builder.CreateUnreachable();8843          for (const auto &Case : SI->cases())8844            if (Case.getCaseSuccessor() == BB) {8845              BB->removePredecessor(Predecessor);8846              Case.setSuccessor(Unreachable);8847            }8848          if (SI->getDefaultDest() == BB) {8849            BB->removePredecessor(Predecessor);8850            SI->setDefaultDest(Unreachable);8851          }8852 8853          if (DTU)8854            DTU->applyUpdates(8855                { { DominatorTree::Insert, Predecessor, Unreachable },8856                  { DominatorTree::Delete, Predecessor, BB } });8857          return true;8858        }8859      }8860 8861  return false;8862}8863 8864bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {8865  bool Changed = false;8866 8867  assert(BB && BB->getParent() && "Block not embedded in function!");8868  assert(BB->getTerminator() && "Degenerate basic block encountered!");8869 8870  // Remove basic blocks that have no predecessors (except the entry block)...8871  // or that just have themself as a predecessor.  These are unreachable.8872  if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||8873      BB->getSinglePredecessor() == BB) {8874    LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB);8875    DeleteDeadBlock(BB, DTU);8876    return true;8877  }8878 8879  // Check to see if we can constant propagate this terminator instruction8880  // away...8881  Changed |= ConstantFoldTerminator(BB, /*DeleteDeadConditions=*/true,8882                                    /*TLI=*/nullptr, DTU);8883 8884  // Check for and eliminate duplicate PHI nodes in this block.8885  Changed |= EliminateDuplicatePHINodes(BB);8886 8887  // Check for and remove branches that will always cause undefined behavior.8888  if (removeUndefIntroducingPredecessor(BB, DTU, Options.AC))8889    return requestResimplify();8890 8891  // Merge basic blocks into their predecessor if there is only one distinct8892  // pred, and if there is only one distinct successor of the predecessor, and8893  // if there are no PHI nodes.8894  if (MergeBlockIntoPredecessor(BB, DTU))8895    return true;8896 8897  if (SinkCommon && Options.SinkCommonInsts)8898    if (sinkCommonCodeFromPredecessors(BB, DTU) ||8899        mergeCompatibleInvokes(BB, DTU)) {8900      // sinkCommonCodeFromPredecessors() does not automatically CSE PHI's,8901      // so we may now how duplicate PHI's.8902      // Let's rerun EliminateDuplicatePHINodes() first,8903      // before foldTwoEntryPHINode() potentially converts them into select's,8904      // after which we'd need a whole EarlyCSE pass run to cleanup them.8905      return true;8906    }8907 8908  IRBuilder<> Builder(BB);8909 8910  if (Options.SpeculateBlocks &&8911      !BB->getParent()->hasFnAttribute(Attribute::OptForFuzzing)) {8912    // If there is a trivial two-entry PHI node in this basic block, and we can8913    // eliminate it, do so now.8914    if (auto *PN = dyn_cast<PHINode>(BB->begin()))8915      if (PN->getNumIncomingValues() == 2)8916        if (foldTwoEntryPHINode(PN, TTI, DTU, Options.AC, DL,8917                                Options.SpeculateUnpredictables))8918          return true;8919  }8920 8921  Instruction *Terminator = BB->getTerminator();8922  Builder.SetInsertPoint(Terminator);8923  switch (Terminator->getOpcode()) {8924  case Instruction::Br:8925    Changed |= simplifyBranch(cast<BranchInst>(Terminator), Builder);8926    break;8927  case Instruction::Resume:8928    Changed |= simplifyResume(cast<ResumeInst>(Terminator), Builder);8929    break;8930  case Instruction::CleanupRet:8931    Changed |= simplifyCleanupReturn(cast<CleanupReturnInst>(Terminator));8932    break;8933  case Instruction::Switch:8934    Changed |= simplifySwitch(cast<SwitchInst>(Terminator), Builder);8935    break;8936  case Instruction::Unreachable:8937    Changed |= simplifyUnreachable(cast<UnreachableInst>(Terminator));8938    break;8939  case Instruction::IndirectBr:8940    Changed |= simplifyIndirectBr(cast<IndirectBrInst>(Terminator));8941    break;8942  }8943 8944  return Changed;8945}8946 8947bool SimplifyCFGOpt::run(BasicBlock *BB) {8948  bool Changed = false;8949 8950  // Repeated simplify BB as long as resimplification is requested.8951  do {8952    Resimplify = false;8953 8954    // Perform one round of simplifcation. Resimplify flag will be set if8955    // another iteration is requested.8956    Changed |= simplifyOnce(BB);8957  } while (Resimplify);8958 8959  return Changed;8960}8961 8962bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,8963                       DomTreeUpdater *DTU, const SimplifyCFGOptions &Options,8964                       ArrayRef<WeakVH> LoopHeaders) {8965  return SimplifyCFGOpt(TTI, DTU, BB->getDataLayout(), LoopHeaders,8966                        Options)8967      .run(BB);8968}8969