SimplifyCFG.cpp revision 263508
150397Sobrien//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 2169689Skan// 3132718Skan// The LLVM Compiler Infrastructure 450397Sobrien// 5132718Skan// This file is distributed under the University of Illinois Open Source 650397Sobrien// License. See LICENSE.TXT for details. 7132718Skan// 850397Sobrien//===----------------------------------------------------------------------===// 950397Sobrien// 1050397Sobrien// Peephole optimize the CFG. 1150397Sobrien// 12132718Skan//===----------------------------------------------------------------------===// 1350397Sobrien 1450397Sobrien#define DEBUG_TYPE "simplifycfg" 1550397Sobrien#include "llvm/Transforms/Utils/Local.h" 1650397Sobrien#include "llvm/ADT/DenseMap.h" 1750397Sobrien#include "llvm/ADT/STLExtras.h" 18132718Skan#include "llvm/ADT/SetVector.h" 19169689Skan#include "llvm/ADT/SmallPtrSet.h" 20169689Skan#include "llvm/ADT/SmallVector.h" 2150397Sobrien#include "llvm/ADT/Statistic.h" 2250397Sobrien#include "llvm/Analysis/ConstantFolding.h" 2350397Sobrien#include "llvm/Analysis/InstructionSimplify.h" 24132718Skan#include "llvm/Analysis/TargetTransformInfo.h" 25132718Skan#include "llvm/Analysis/ValueTracking.h" 2650397Sobrien#include "llvm/IR/Constants.h" 2750397Sobrien#include "llvm/IR/DataLayout.h" 2890075Sobrien#include "llvm/IR/DerivedTypes.h" 2950397Sobrien#include "llvm/IR/GlobalVariable.h" 3050397Sobrien#include "llvm/IR/IRBuilder.h" 3150397Sobrien#include "llvm/IR/Instructions.h" 3250397Sobrien#include "llvm/IR/IntrinsicInst.h" 3350397Sobrien#include "llvm/IR/LLVMContext.h" 3450397Sobrien#include "llvm/IR/MDBuilder.h" 3550397Sobrien#include "llvm/IR/Metadata.h" 36117395Skan#include "llvm/IR/Module.h" 3752284Sobrien#include "llvm/IR/Operator.h" 3850397Sobrien#include "llvm/IR/Type.h" 39132718Skan#include "llvm/Support/CFG.h" 4050397Sobrien#include "llvm/Support/CommandLine.h" 4150397Sobrien#include "llvm/Support/ConstantRange.h" 42132718Skan#include "llvm/Support/Debug.h" 4350397Sobrien#include "llvm/Support/NoFolder.h" 4450397Sobrien#include "llvm/Support/PatternMatch.h" 4550397Sobrien#include "llvm/Support/raw_ostream.h" 4650397Sobrien#include "llvm/Transforms/Utils/BasicBlockUtils.h" 4750397Sobrien#include <algorithm> 4890075Sobrien#include <map> 4950397Sobrien#include <set> 5050397Sobrienusing namespace llvm; 5150397Sobrienusing namespace PatternMatch; 5250397Sobrien 5350397Sobrienstatic cl::opt<unsigned> 5450397SobrienPHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 5550397Sobrien cl::desc("Control the amount of phi node folding to perform (default = 1)")); 5650397Sobrien 5750397Sobrienstatic cl::opt<bool> 5852284SobrienDupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 5950397Sobrien cl::desc("Duplicate return instructions into unconditional branches")); 6052284Sobrien 6150397Sobrienstatic cl::opt<bool> 6250397SobrienSinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 63132718Skan cl::desc("Sink common instructions down to the end block")); 64132718Skan 65132718Skanstatic cl::opt<bool> 6650397SobrienHoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true), 6750397Sobrien cl::desc("Hoist conditional stores if an unconditional store preceeds")); 68132718Skan 6952284SobrienSTATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 7050397SobrienSTATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 71132718SkanSTATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 72117395SkanSTATISTIC(NumSpeculations, "Number of speculative executed instructions"); 7350397Sobrien 7450397Sobriennamespace { 7550397Sobrien /// ValueEqualityComparisonCase - Represents a case of a switch. 7650397Sobrien struct ValueEqualityComparisonCase { 7750397Sobrien ConstantInt *Value; 7850397Sobrien BasicBlock *Dest; 7950397Sobrien 8050397Sobrien ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 81169689Skan : Value(Value), Dest(Dest) {} 82169689Skan 83169689Skan bool operator<(ValueEqualityComparisonCase RHS) const { 8450397Sobrien // Comparing pointers is ok as we only rely on the order for uniquing. 8550397Sobrien return Value < RHS.Value; 8650397Sobrien } 8750397Sobrien 8850397Sobrien bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 89169689Skan }; 90169689Skan 9152284Sobrienclass SimplifyCFGOpt { 9250397Sobrien const TargetTransformInfo &TTI; 9350397Sobrien const DataLayout *const TD; 94169689Skan Value *isValueEqualityComparison(TerminatorInst *TI); 9550397Sobrien BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 96169689Skan std::vector<ValueEqualityComparisonCase> &Cases); 97169689Skan bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 98169689Skan BasicBlock *Pred, 99169689Skan IRBuilder<> &Builder); 100169689Skan bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 101169689Skan IRBuilder<> &Builder); 102169689Skan 10350397Sobrien bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 104169689Skan bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 105169689Skan bool SimplifyUnreachable(UnreachableInst *UI); 106169689Skan bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 107169689Skan bool SimplifyIndirectBr(IndirectBrInst *IBI); 108169689Skan bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 109169689Skan bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 110169689Skan 111169689Skanpublic: 112169689Skan SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD) 11350397Sobrien : TTI(TTI), TD(TD) {} 114117395Skan bool run(BasicBlock *BB); 11550397Sobrien}; 11650397Sobrien} 11750397Sobrien 11850397Sobrien/// SafeToMergeTerminators - Return true if it is safe to merge these two 11950397Sobrien/// terminator instructions together. 12050397Sobrien/// 12150397Sobrienstatic bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 12250397Sobrien if (SI1 == SI2) return false; // Can't merge with self! 12350397Sobrien 12450397Sobrien // It is not safe to merge these two switch instructions if they have a common 125132718Skan // successor, and if that successor has a PHI node, and if *that* PHI node has 12650397Sobrien // conflicting incoming values from the two switch blocks. 127132718Skan BasicBlock *SI1BB = SI1->getParent(); 128132718Skan BasicBlock *SI2BB = SI2->getParent(); 129132718Skan SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 130132718Skan 13152284Sobrien for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 132132718Skan if (SI1Succs.count(*I)) 13350397Sobrien for (BasicBlock::iterator BBI = (*I)->begin(); 13452284Sobrien isa<PHINode>(BBI); ++BBI) { 13552284Sobrien PHINode *PN = cast<PHINode>(BBI); 13652284Sobrien if (PN->getIncomingValueForBlock(SI1BB) != 137169689Skan PN->getIncomingValueForBlock(SI2BB)) 13850397Sobrien return false; 13952284Sobrien } 14052284Sobrien 14152284Sobrien return true; 14252284Sobrien} 14352284Sobrien 14452284Sobrien/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 14552284Sobrien/// to merge these two terminator instructions together, where SI1 is an 14652284Sobrien/// unconditional branch. PhiNodes will store all PHI nodes in common 14750397Sobrien/// successors. 14850397Sobrien/// 14952284Sobrienstatic bool isProfitableToFoldUnconditional(BranchInst *SI1, 15050397Sobrien BranchInst *SI2, 15152284Sobrien Instruction *Cond, 15250397Sobrien SmallVectorImpl<PHINode*> &PhiNodes) { 15350397Sobrien if (SI1 == SI2) return false; // Can't merge with self! 15450397Sobrien assert(SI1->isUnconditional() && SI2->isConditional()); 15550397Sobrien 156132718Skan // We fold the unconditional branch if we can easily update all PHI nodes in 157169689Skan // common successors: 158132718Skan // 1> We have a constant incoming value for the conditional branch; 15950397Sobrien // 2> We have "Cond" as the incoming value for the unconditional branch; 16050397Sobrien // 3> SI2->getCondition() and Cond have same operands. 16150397Sobrien CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 162117395Skan if (!Ci2) return false; 163117395Skan if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 164117395Skan Cond->getOperand(1) == Ci2->getOperand(1)) && 165132718Skan !(Cond->getOperand(0) == Ci2->getOperand(1) && 16650397Sobrien Cond->getOperand(1) == Ci2->getOperand(0))) 16750397Sobrien return false; 16850397Sobrien 16950397Sobrien BasicBlock *SI1BB = SI1->getParent(); 17050397Sobrien BasicBlock *SI2BB = SI2->getParent(); 17152284Sobrien SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 172169689Skan for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 173169689Skan if (SI1Succs.count(*I)) 174169689Skan for (BasicBlock::iterator BBI = (*I)->begin(); 175132718Skan isa<PHINode>(BBI); ++BBI) { 176117395Skan PHINode *PN = cast<PHINode>(BBI); 177117395Skan if (PN->getIncomingValueForBlock(SI1BB) != Cond || 17850397Sobrien !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 179132718Skan return false; 18050397Sobrien PhiNodes.push_back(PN); 18152284Sobrien } 182169689Skan return true; 18352284Sobrien} 18452284Sobrien 18550397Sobrien/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 18650397Sobrien/// now be entries in it from the 'NewPred' block. The values that will be 18750397Sobrien/// flowing into the PHI nodes will be the same as those coming in from 18850397Sobrien/// ExistPred, an existing predecessor of Succ. 18950397Sobrienstatic void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 19050397Sobrien BasicBlock *ExistPred) { 19150397Sobrien if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 19250397Sobrien 19350397Sobrien PHINode *PN; 19450397Sobrien for (BasicBlock::iterator I = Succ->begin(); 195132718Skan (PN = dyn_cast<PHINode>(I)); ++I) 19650397Sobrien PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 197132718Skan} 198132718Skan 199132718Skan/// ComputeSpeculationCost - Compute an abstract "cost" of speculating the 200132718Skan/// given instruction, which is assumed to be safe to speculate. 1 means 20150397Sobrien/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 202132718Skanstatic unsigned ComputeSpeculationCost(const User *I) { 20350397Sobrien assert(isSafeToSpeculativelyExecute(I) && 20450397Sobrien "Instruction is not safe to speculatively execute!"); 20550397Sobrien switch (Operator::getOpcode(I)) { 206169689Skan default: 207169689Skan // In doubt, be conservative. 208169689Skan return UINT_MAX; 209169689Skan case Instruction::GetElementPtr: 210169689Skan // GEPs are cheap if all indices are constant. 211169689Skan if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 212169689Skan return UINT_MAX; 213169689Skan return 1; 214169689Skan case Instruction::Load: 215169689Skan case Instruction::Add: 216169689Skan case Instruction::Sub: 217169689Skan case Instruction::And: 218169689Skan case Instruction::Or: 219169689Skan case Instruction::Xor: 220169689Skan case Instruction::Shl: 221169689Skan case Instruction::LShr: 222169689Skan case Instruction::AShr: 223169689Skan case Instruction::ICmp: 224169689Skan case Instruction::Trunc: 225169689Skan case Instruction::ZExt: 226169689Skan case Instruction::SExt: 227169689Skan return 1; // These are all cheap. 228169689Skan 229169689Skan case Instruction::Call: 230169689Skan case Instruction::Select: 23190075Sobrien return 2; 23250397Sobrien } 233169689Skan} 23450397Sobrien 23550397Sobrien/// DominatesMergePoint - If we have a merge point of an "if condition" as 23652284Sobrien/// accepted above, return true if the specified value dominates the block. We 237169689Skan/// don't handle the true generality of domination here, just a special case 23850397Sobrien/// which works well enough for us. 23950397Sobrien/// 24050397Sobrien/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 241117395Skan/// see if V (which must be an instruction) and its recursive operands 242117395Skan/// that do not dominate BB have a combined cost lower than CostRemaining and 243117395Skan/// are non-trapping. If both are true, the instruction is inserted into the 244117395Skan/// set and true is returned. 245117395Skan/// 246117395Skan/// The cost for most non-trapping instructions is defined as 1 except for 247117395Skan/// Select whose cost is 2. 248117395Skan/// 249169689Skan/// After this function returns, CostRemaining is decreased by the cost of 250117395Skan/// V plus its non-dominating operands. If that cost is greater than 251117395Skan/// CostRemaining, false is returned and CostRemaining is undefined. 252117395Skanstatic bool DominatesMergePoint(Value *V, BasicBlock *BB, 253117395Skan SmallPtrSet<Instruction*, 4> *AggressiveInsts, 25452284Sobrien unsigned &CostRemaining) { 25550397Sobrien Instruction *I = dyn_cast<Instruction>(V); 256132718Skan if (!I) { 257169689Skan // Non-instructions all dominate instructions, but not all constantexprs 258132718Skan // can be executed unconditionally. 25950397Sobrien if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 26050397Sobrien if (C->canTrap()) 26150397Sobrien return false; 26290075Sobrien return true; 26390075Sobrien } 26490075Sobrien BasicBlock *PBB = I->getParent(); 26590075Sobrien 26690075Sobrien // We don't want to allow weird loops that might have the "if condition" in 267117395Skan // the bottom of this block. 268169689Skan if (PBB == BB) return false; 26990075Sobrien 27090075Sobrien // If this instruction is defined in a block that contains an unconditional 27190075Sobrien // branch to BB, then it must be in the 'conditional' part of the "if 272169689Skan // statement". If not, it definitely dominates the region. 273169689Skan BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 274169689Skan if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 275169689Skan return true; 276169689Skan 277169689Skan // If we aren't allowing aggressive promotion anymore, then don't consider 278169689Skan // instructions in the 'if region'. 279169689Skan if (AggressiveInsts == 0) return false; 280169689Skan 281169689Skan // If we have seen this instruction before, don't count it again. 282169689Skan if (AggressiveInsts->count(I)) return true; 283169689Skan 284169689Skan // Okay, it looks like the instruction IS in the "condition". Check to 285169689Skan // see if it's a cheap instruction to unconditionally compute, and if it 286169689Skan // only uses stuff defined outside of the condition. If so, hoist it out. 287169689Skan if (!isSafeToSpeculativelyExecute(I)) 288169689Skan return false; 289169689Skan 290169689Skan unsigned Cost = ComputeSpeculationCost(I); 291169689Skan 292169689Skan if (Cost > CostRemaining) 293169689Skan return false; 294169689Skan 295169689Skan CostRemaining -= Cost; 296169689Skan 297169689Skan // Okay, we can only really hoist these out if their operands do 298169689Skan // not take us over the cost threshold. 299169689Skan for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 300169689Skan if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 301169689Skan return false; 302169689Skan // Okay, it's safe to do this! Remember this instruction. 303169689Skan AggressiveInsts->insert(I); 304169689Skan return true; 305169689Skan} 306169689Skan 307169689Skan/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 308169689Skan/// and PointerNullValue. Return NULL if value is not a constant int. 309169689Skanstatic ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 310169689Skan // Normal constant int. 311169689Skan ConstantInt *CI = dyn_cast<ConstantInt>(V); 312169689Skan if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 313169689Skan return CI; 314169689Skan 315169689Skan // This is some kind of pointer constant. Turn it into a pointer-sized 316169689Skan // ConstantInt if possible. 317169689Skan IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); 318169689Skan 319169689Skan // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 320169689Skan if (isa<ConstantPointerNull>(V)) 321169689Skan return ConstantInt::get(PtrTy, 0); 322169689Skan 323169689Skan // IntToPtr const int. 324169689Skan if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 325169689Skan if (CE->getOpcode() == Instruction::IntToPtr) 326169689Skan if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 327169689Skan // The constant is very likely to have the right type already. 328169689Skan if (CI->getType() == PtrTy) 329169689Skan return CI; 330169689Skan else 331169689Skan return cast<ConstantInt> 332169689Skan (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 333169689Skan } 334169689Skan return 0; 335169689Skan} 336169689Skan 33790075Sobrien/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 33890075Sobrien/// collection of icmp eq/ne instructions that compare a value against a 33990075Sobrien/// constant, return the value being compared, and stick the constant into the 34090075Sobrien/// Values vector. 341169689Skanstatic Value * 34290075SobrienGatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 34390075Sobrien const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 34490075Sobrien Instruction *I = dyn_cast<Instruction>(V); 34590075Sobrien if (I == 0) return 0; 34690075Sobrien 347169689Skan // If this is an icmp against a constant, handle this as one of the cases. 34890075Sobrien if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 34990075Sobrien if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 350169689Skan Value *RHSVal; 351169689Skan ConstantInt *RHSC; 35250397Sobrien 35350397Sobrien if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 354169689Skan // (x & ~2^x) == y --> x == y || x == y|2^x 355169689Skan // This undoes a transformation done by instcombine to fuse 2 compares. 356169689Skan if (match(ICI->getOperand(0), 357132718Skan m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) { 358132718Skan APInt Not = ~RHSC->getValue(); 359132718Skan if (Not.isPowerOf2()) { 360132718Skan Vals.push_back(C); 36150397Sobrien Vals.push_back( 362132718Skan ConstantInt::get(C->getContext(), C->getValue() | Not)); 363132718Skan UsedICmps++; 364132718Skan return RHSVal; 365132718Skan } 366132718Skan } 367132718Skan 368132718Skan UsedICmps++; 369169689Skan Vals.push_back(C); 370132718Skan return I->getOperand(0); 371132718Skan } 372132718Skan 373132718Skan // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 374132718Skan // the set. 375132718Skan ConstantRange Span = 376132718Skan ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 377132718Skan 378169689Skan // Shift the range if the compare is fed by an add. This is the range 379132718Skan // compare idiom as emitted by instcombine. 380132718Skan bool hasAdd = 381132718Skan match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC))); 382132718Skan if (hasAdd) 383169689Skan Span = Span.subtract(RHSC->getValue()); 384169689Skan 385132718Skan // If this is an and/!= check then we want to optimize "x ugt 2" into 386117395Skan // x != 0 && x != 1. 387117395Skan if (!isEQ) 38850397Sobrien Span = Span.inverse(); 38950397Sobrien 390169689Skan // If there are a ton of values, we don't want to make a ginormous switch. 39152284Sobrien if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 39252284Sobrien return 0; 39352284Sobrien 394169689Skan for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 395169689Skan Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 39650397Sobrien UsedICmps++; 39750397Sobrien return hasAdd ? RHSVal : I->getOperand(0); 39850397Sobrien } 399169689Skan return 0; 400169689Skan } 401169689Skan 402169689Skan // Otherwise, we can only handle an | or &, depending on isEQ. 403169689Skan if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 404169689Skan return 0; 40550397Sobrien 40650397Sobrien unsigned NumValsBeforeLHS = Vals.size(); 407132718Skan unsigned UsedICmpsBeforeLHS = UsedICmps; 408169689Skan if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 409169689Skan isEQ, UsedICmps)) { 41050397Sobrien unsigned NumVals = Vals.size(); 411169689Skan unsigned UsedICmpsBeforeRHS = UsedICmps; 412169689Skan if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 413169689Skan isEQ, UsedICmps)) { 41450397Sobrien if (LHS == RHS) 41550397Sobrien return LHS; 41650397Sobrien Vals.resize(NumVals); 41750397Sobrien UsedICmps = UsedICmpsBeforeRHS; 41850397Sobrien } 41950397Sobrien 42050397Sobrien // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 42150397Sobrien // set it and return success. 42250397Sobrien if (Extra == 0 || Extra == I->getOperand(1)) { 42350397Sobrien Extra = I->getOperand(1); 42450397Sobrien return LHS; 42550397Sobrien } 426132718Skan 427132718Skan Vals.resize(NumValsBeforeLHS); 428132718Skan UsedICmps = UsedICmpsBeforeLHS; 429132718Skan return 0; 430132718Skan } 431132718Skan 432132718Skan // If the LHS can't be folded in, but Extra is available and RHS can, try to 433132718Skan // use LHS as Extra. 434132718Skan if (Extra == 0 || Extra == I->getOperand(0)) { 435132718Skan Value *OldExtra = Extra; 436132718Skan Extra = I->getOperand(0); 437132718Skan if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 438132718Skan isEQ, UsedICmps)) 439132718Skan return RHS; 440132718Skan assert(Vals.size() == NumValsBeforeLHS); 441132718Skan Extra = OldExtra; 442132718Skan } 443132718Skan 444132718Skan return 0; 445132718Skan} 446132718Skan 447132718Skanstatic void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 448132718Skan Instruction *Cond = 0; 449132718Skan if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 45050397Sobrien Cond = dyn_cast<Instruction>(SI->getCondition()); 45150397Sobrien } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 45250397Sobrien if (BI->isConditional()) 45350397Sobrien Cond = dyn_cast<Instruction>(BI->getCondition()); 45450397Sobrien } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 455132718Skan Cond = dyn_cast<Instruction>(IBI->getAddress()); 45652284Sobrien } 45750397Sobrien 458169689Skan TI->eraseFromParent(); 45950397Sobrien if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 460132718Skan} 461132718Skan 462132718Skan/// isValueEqualityComparison - Return true if the specified terminator checks 463132718Skan/// to see if a value is equal to constant integer value. 464132718SkanValue *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 465132718Skan Value *CV = 0; 46650397Sobrien if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 46752284Sobrien // Do not permit merging of large switch instructions into their 46852284Sobrien // predecessors unless there is only one predecessor. 46952284Sobrien if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 470132718Skan pred_end(SI->getParent())) <= 128) 471132718Skan CV = SI->getCondition(); 472132718Skan } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 47352284Sobrien if (BI->isConditional() && BI->getCondition()->hasOneUse()) 474132718Skan if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 475132718Skan if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD)) 476132718Skan CV = ICI->getOperand(0); 477132718Skan 478132718Skan // Unwrap any lossless ptrtoint cast. 479132718Skan if (TD && CV) { 480132718Skan if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) { 481169689Skan Value *Ptr = PTII->getPointerOperand(); 482132718Skan if (PTII->getType() == TD->getIntPtrType(Ptr->getType())) 483169689Skan CV = Ptr; 484169689Skan } 485169689Skan } 486169689Skan return CV; 48750397Sobrien} 488132718Skan 489132718Skan/// GetValueEqualityComparisonCases - Given a value comparison instruction, 490132718Skan/// decode all of the 'cases' that it represents and return the 'default' block. 49152284SobrienBasicBlock *SimplifyCFGOpt:: 492132718SkanGetValueEqualityComparisonCases(TerminatorInst *TI, 49350397Sobrien std::vector<ValueEqualityComparisonCase> 49450397Sobrien &Cases) { 495169689Skan if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 49652284Sobrien Cases.reserve(SI->getNumCases()); 49750397Sobrien for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 49850397Sobrien Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 49950397Sobrien i.getCaseSuccessor())); 50050397Sobrien return SI->getDefaultDest(); 50150397Sobrien } 502132718Skan 503132718Skan BranchInst *BI = cast<BranchInst>(TI); 50450397Sobrien ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 50550397Sobrien BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 50650397Sobrien Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 50750397Sobrien TD), 50850397Sobrien Succ)); 50950397Sobrien return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 51090075Sobrien} 51150397Sobrien 51250397Sobrien 51350397Sobrien/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 51490075Sobrien/// in the list that match the specified block. 51590075Sobrienstatic void EliminateBlockCases(BasicBlock *BB, 51690075Sobrien std::vector<ValueEqualityComparisonCase> &Cases) { 51790075Sobrien Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 51890075Sobrien} 51990075Sobrien 52090075Sobrien/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 52190075Sobrien/// well. 52290075Sobrienstatic bool 52390075SobrienValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 52490075Sobrien std::vector<ValueEqualityComparisonCase > &C2) { 52590075Sobrien std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 52690075Sobrien 52790075Sobrien // Make V1 be smaller than V2. 528169689Skan if (V1->size() > V2->size()) 52990075Sobrien std::swap(V1, V2); 53090075Sobrien 53190075Sobrien if (V1->size() == 0) return false; 53290075Sobrien if (V1->size() == 1) { 533117395Skan // Just scan V2. 534169689Skan ConstantInt *TheVal = (*V1)[0].Value; 53590075Sobrien for (unsigned i = 0, e = V2->size(); i != e; ++i) 53690075Sobrien if (TheVal == (*V2)[i].Value) 53790075Sobrien return true; 538169689Skan } 539169689Skan 540169689Skan // Otherwise, just sort both lists and compare element by element. 54190075Sobrien array_pod_sort(V1->begin(), V1->end()); 542169689Skan array_pod_sort(V2->begin(), V2->end()); 54390075Sobrien unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 544169689Skan while (i1 != e1 && i2 != e2) { 54590075Sobrien if ((*V1)[i1].Value == (*V2)[i2].Value) 54690075Sobrien return true; 54750397Sobrien if ((*V1)[i1].Value < (*V2)[i2].Value) 54890075Sobrien ++i1; 54950397Sobrien else 55050397Sobrien ++i2; 551169689Skan } 552169689Skan return false; 55350397Sobrien} 55450397Sobrien 555169689Skan/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 556169689Skan/// terminator instruction and its block is known to only have a single 557169689Skan/// predecessor block, check to see if that predecessor is also a value 558169689Skan/// comparison with the same value, and if that comparison determines the 55950397Sobrien/// outcome of this comparison. If so, simplify TI. This does a very limited 56050397Sobrien/// form of jump threading. 56150397Sobrienbool SimplifyCFGOpt:: 56250397SobrienSimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 56350397Sobrien BasicBlock *Pred, 564132718Skan IRBuilder<> &Builder) { 565132718Skan Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 566169689Skan if (!PredVal) return false; // Not a value comparison in predecessor. 567169689Skan 568132718Skan Value *ThisVal = isValueEqualityComparison(TI); 569132718Skan assert(ThisVal && "This isn't a value comparison!!"); 57050397Sobrien if (ThisVal != PredVal) return false; // Different predicates. 57150397Sobrien 57252284Sobrien // TODO: Preserve branch weight metadata, similarly to how 57352284Sobrien // FoldValueComparisonIntoPredecessors preserves it. 57452284Sobrien 57552284Sobrien // Find out information about when control will move from Pred to TI's block. 57652284Sobrien std::vector<ValueEqualityComparisonCase> PredCases; 57752284Sobrien BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 57890075Sobrien PredCases); 57950397Sobrien EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 58050397Sobrien 58150397Sobrien // Find information about how control leaves this block. 582 std::vector<ValueEqualityComparisonCase> ThisCases; 583 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 584 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 585 586 // If TI's block is the default block from Pred's comparison, potentially 587 // simplify TI based on this knowledge. 588 if (PredDef == TI->getParent()) { 589 // If we are here, we know that the value is none of those cases listed in 590 // PredCases. If there are any cases in ThisCases that are in PredCases, we 591 // can simplify TI. 592 if (!ValuesOverlap(PredCases, ThisCases)) 593 return false; 594 595 if (isa<BranchInst>(TI)) { 596 // Okay, one of the successors of this condbr is dead. Convert it to a 597 // uncond br. 598 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 599 // Insert the new branch. 600 Instruction *NI = Builder.CreateBr(ThisDef); 601 (void) NI; 602 603 // Remove PHI node entries for the dead edge. 604 ThisCases[0].Dest->removePredecessor(TI->getParent()); 605 606 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 607 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 608 609 EraseTerminatorInstAndDCECond(TI); 610 return true; 611 } 612 613 SwitchInst *SI = cast<SwitchInst>(TI); 614 // Okay, TI has cases that are statically dead, prune them away. 615 SmallPtrSet<Constant*, 16> DeadCases; 616 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 617 DeadCases.insert(PredCases[i].Value); 618 619 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 620 << "Through successor TI: " << *TI); 621 622 // Collect branch weights into a vector. 623 SmallVector<uint32_t, 8> Weights; 624 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 625 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 626 if (HasWeight) 627 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 628 ++MD_i) { 629 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 630 assert(CI); 631 Weights.push_back(CI->getValue().getZExtValue()); 632 } 633 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 634 --i; 635 if (DeadCases.count(i.getCaseValue())) { 636 if (HasWeight) { 637 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 638 Weights.pop_back(); 639 } 640 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 641 SI->removeCase(i); 642 } 643 } 644 if (HasWeight && Weights.size() >= 2) 645 SI->setMetadata(LLVMContext::MD_prof, 646 MDBuilder(SI->getParent()->getContext()). 647 createBranchWeights(Weights)); 648 649 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 650 return true; 651 } 652 653 // Otherwise, TI's block must correspond to some matched value. Find out 654 // which value (or set of values) this is. 655 ConstantInt *TIV = 0; 656 BasicBlock *TIBB = TI->getParent(); 657 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 658 if (PredCases[i].Dest == TIBB) { 659 if (TIV != 0) 660 return false; // Cannot handle multiple values coming to this block. 661 TIV = PredCases[i].Value; 662 } 663 assert(TIV && "No edge from pred to succ?"); 664 665 // Okay, we found the one constant that our value can be if we get into TI's 666 // BB. Find out which successor will unconditionally be branched to. 667 BasicBlock *TheRealDest = 0; 668 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 669 if (ThisCases[i].Value == TIV) { 670 TheRealDest = ThisCases[i].Dest; 671 break; 672 } 673 674 // If not handled by any explicit cases, it is handled by the default case. 675 if (TheRealDest == 0) TheRealDest = ThisDef; 676 677 // Remove PHI node entries for dead edges. 678 BasicBlock *CheckEdge = TheRealDest; 679 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 680 if (*SI != CheckEdge) 681 (*SI)->removePredecessor(TIBB); 682 else 683 CheckEdge = 0; 684 685 // Insert the new branch. 686 Instruction *NI = Builder.CreateBr(TheRealDest); 687 (void) NI; 688 689 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 690 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 691 692 EraseTerminatorInstAndDCECond(TI); 693 return true; 694} 695 696namespace { 697 /// ConstantIntOrdering - This class implements a stable ordering of constant 698 /// integers that does not depend on their address. This is important for 699 /// applications that sort ConstantInt's to ensure uniqueness. 700 struct ConstantIntOrdering { 701 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 702 return LHS->getValue().ult(RHS->getValue()); 703 } 704 }; 705} 706 707static int ConstantIntSortPredicate(ConstantInt *const *P1, 708 ConstantInt *const *P2) { 709 const ConstantInt *LHS = *P1; 710 const ConstantInt *RHS = *P2; 711 if (LHS->getValue().ult(RHS->getValue())) 712 return 1; 713 if (LHS->getValue() == RHS->getValue()) 714 return 0; 715 return -1; 716} 717 718static inline bool HasBranchWeights(const Instruction* I) { 719 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 720 if (ProfMD && ProfMD->getOperand(0)) 721 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 722 return MDS->getString().equals("branch_weights"); 723 724 return false; 725} 726 727/// Get Weights of a given TerminatorInst, the default weight is at the front 728/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 729/// metadata. 730static void GetBranchWeights(TerminatorInst *TI, 731 SmallVectorImpl<uint64_t> &Weights) { 732 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 733 assert(MD); 734 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 735 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 736 assert(CI); 737 Weights.push_back(CI->getValue().getZExtValue()); 738 } 739 740 // If TI is a conditional eq, the default case is the false case, 741 // and the corresponding branch-weight data is at index 2. We swap the 742 // default weight to be the first entry. 743 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 744 assert(Weights.size() == 2); 745 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 746 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 747 std::swap(Weights.front(), Weights.back()); 748 } 749} 750 751/// Sees if any of the weights are too big for a uint32_t, and halves all the 752/// weights if any are. 753static void FitWeights(MutableArrayRef<uint64_t> Weights) { 754 bool Halve = false; 755 for (unsigned i = 0; i < Weights.size(); ++i) 756 if (Weights[i] > UINT_MAX) { 757 Halve = true; 758 break; 759 } 760 761 if (! Halve) 762 return; 763 764 for (unsigned i = 0; i < Weights.size(); ++i) 765 Weights[i] /= 2; 766} 767 768/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 769/// equality comparison instruction (either a switch or a branch on "X == c"). 770/// See if any of the predecessors of the terminator block are value comparisons 771/// on the same value. If so, and if safe to do so, fold them together. 772bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 773 IRBuilder<> &Builder) { 774 BasicBlock *BB = TI->getParent(); 775 Value *CV = isValueEqualityComparison(TI); // CondVal 776 assert(CV && "Not a comparison?"); 777 bool Changed = false; 778 779 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 780 while (!Preds.empty()) { 781 BasicBlock *Pred = Preds.pop_back_val(); 782 783 // See if the predecessor is a comparison with the same value. 784 TerminatorInst *PTI = Pred->getTerminator(); 785 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 786 787 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 788 // Figure out which 'cases' to copy from SI to PSI. 789 std::vector<ValueEqualityComparisonCase> BBCases; 790 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 791 792 std::vector<ValueEqualityComparisonCase> PredCases; 793 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 794 795 // Based on whether the default edge from PTI goes to BB or not, fill in 796 // PredCases and PredDefault with the new switch cases we would like to 797 // build. 798 SmallVector<BasicBlock*, 8> NewSuccessors; 799 800 // Update the branch weight metadata along the way 801 SmallVector<uint64_t, 8> Weights; 802 bool PredHasWeights = HasBranchWeights(PTI); 803 bool SuccHasWeights = HasBranchWeights(TI); 804 805 if (PredHasWeights) { 806 GetBranchWeights(PTI, Weights); 807 // branch-weight metadata is inconsistent here. 808 if (Weights.size() != 1 + PredCases.size()) 809 PredHasWeights = SuccHasWeights = false; 810 } else if (SuccHasWeights) 811 // If there are no predecessor weights but there are successor weights, 812 // populate Weights with 1, which will later be scaled to the sum of 813 // successor's weights 814 Weights.assign(1 + PredCases.size(), 1); 815 816 SmallVector<uint64_t, 8> SuccWeights; 817 if (SuccHasWeights) { 818 GetBranchWeights(TI, SuccWeights); 819 // branch-weight metadata is inconsistent here. 820 if (SuccWeights.size() != 1 + BBCases.size()) 821 PredHasWeights = SuccHasWeights = false; 822 } else if (PredHasWeights) 823 SuccWeights.assign(1 + BBCases.size(), 1); 824 825 if (PredDefault == BB) { 826 // If this is the default destination from PTI, only the edges in TI 827 // that don't occur in PTI, or that branch to BB will be activated. 828 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 829 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 830 if (PredCases[i].Dest != BB) 831 PTIHandled.insert(PredCases[i].Value); 832 else { 833 // The default destination is BB, we don't need explicit targets. 834 std::swap(PredCases[i], PredCases.back()); 835 836 if (PredHasWeights || SuccHasWeights) { 837 // Increase weight for the default case. 838 Weights[0] += Weights[i+1]; 839 std::swap(Weights[i+1], Weights.back()); 840 Weights.pop_back(); 841 } 842 843 PredCases.pop_back(); 844 --i; --e; 845 } 846 847 // Reconstruct the new switch statement we will be building. 848 if (PredDefault != BBDefault) { 849 PredDefault->removePredecessor(Pred); 850 PredDefault = BBDefault; 851 NewSuccessors.push_back(BBDefault); 852 } 853 854 unsigned CasesFromPred = Weights.size(); 855 uint64_t ValidTotalSuccWeight = 0; 856 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 857 if (!PTIHandled.count(BBCases[i].Value) && 858 BBCases[i].Dest != BBDefault) { 859 PredCases.push_back(BBCases[i]); 860 NewSuccessors.push_back(BBCases[i].Dest); 861 if (SuccHasWeights || PredHasWeights) { 862 // The default weight is at index 0, so weight for the ith case 863 // should be at index i+1. Scale the cases from successor by 864 // PredDefaultWeight (Weights[0]). 865 Weights.push_back(Weights[0] * SuccWeights[i+1]); 866 ValidTotalSuccWeight += SuccWeights[i+1]; 867 } 868 } 869 870 if (SuccHasWeights || PredHasWeights) { 871 ValidTotalSuccWeight += SuccWeights[0]; 872 // Scale the cases from predecessor by ValidTotalSuccWeight. 873 for (unsigned i = 1; i < CasesFromPred; ++i) 874 Weights[i] *= ValidTotalSuccWeight; 875 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 876 Weights[0] *= SuccWeights[0]; 877 } 878 } else { 879 // If this is not the default destination from PSI, only the edges 880 // in SI that occur in PSI with a destination of BB will be 881 // activated. 882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 883 std::map<ConstantInt*, uint64_t> WeightsForHandled; 884 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 885 if (PredCases[i].Dest == BB) { 886 PTIHandled.insert(PredCases[i].Value); 887 888 if (PredHasWeights || SuccHasWeights) { 889 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 890 std::swap(Weights[i+1], Weights.back()); 891 Weights.pop_back(); 892 } 893 894 std::swap(PredCases[i], PredCases.back()); 895 PredCases.pop_back(); 896 --i; --e; 897 } 898 899 // Okay, now we know which constants were sent to BB from the 900 // predecessor. Figure out where they will all go now. 901 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 902 if (PTIHandled.count(BBCases[i].Value)) { 903 // If this is one we are capable of getting... 904 if (PredHasWeights || SuccHasWeights) 905 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 906 PredCases.push_back(BBCases[i]); 907 NewSuccessors.push_back(BBCases[i].Dest); 908 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 909 } 910 911 // If there are any constants vectored to BB that TI doesn't handle, 912 // they must go to the default destination of TI. 913 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 914 PTIHandled.begin(), 915 E = PTIHandled.end(); I != E; ++I) { 916 if (PredHasWeights || SuccHasWeights) 917 Weights.push_back(WeightsForHandled[*I]); 918 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 919 NewSuccessors.push_back(BBDefault); 920 } 921 } 922 923 // Okay, at this point, we know which new successor Pred will get. Make 924 // sure we update the number of entries in the PHI nodes for these 925 // successors. 926 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 927 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 928 929 Builder.SetInsertPoint(PTI); 930 // Convert pointer to int before we switch. 931 if (CV->getType()->isPointerTy()) { 932 assert(TD && "Cannot switch on pointer without DataLayout"); 933 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getType()), 934 "magicptr"); 935 } 936 937 // Now that the successors are updated, create the new Switch instruction. 938 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 939 PredCases.size()); 940 NewSI->setDebugLoc(PTI->getDebugLoc()); 941 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 942 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 943 944 if (PredHasWeights || SuccHasWeights) { 945 // Halve the weights if any of them cannot fit in an uint32_t 946 FitWeights(Weights); 947 948 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 949 950 NewSI->setMetadata(LLVMContext::MD_prof, 951 MDBuilder(BB->getContext()). 952 createBranchWeights(MDWeights)); 953 } 954 955 EraseTerminatorInstAndDCECond(PTI); 956 957 // Okay, last check. If BB is still a successor of PSI, then we must 958 // have an infinite loop case. If so, add an infinitely looping block 959 // to handle the case to preserve the behavior of the code. 960 BasicBlock *InfLoopBlock = 0; 961 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 962 if (NewSI->getSuccessor(i) == BB) { 963 if (InfLoopBlock == 0) { 964 // Insert it at the end of the function, because it's either code, 965 // or it won't matter if it's hot. :) 966 InfLoopBlock = BasicBlock::Create(BB->getContext(), 967 "infloop", BB->getParent()); 968 BranchInst::Create(InfLoopBlock, InfLoopBlock); 969 } 970 NewSI->setSuccessor(i, InfLoopBlock); 971 } 972 973 Changed = true; 974 } 975 } 976 return Changed; 977} 978 979// isSafeToHoistInvoke - If we would need to insert a select that uses the 980// value of this invoke (comments in HoistThenElseCodeToIf explain why we 981// would need to do this), we can't hoist the invoke, as there is nowhere 982// to put the select in this case. 983static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 984 Instruction *I1, Instruction *I2) { 985 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 986 PHINode *PN; 987 for (BasicBlock::iterator BBI = SI->begin(); 988 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 989 Value *BB1V = PN->getIncomingValueForBlock(BB1); 990 Value *BB2V = PN->getIncomingValueForBlock(BB2); 991 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 992 return false; 993 } 994 } 995 } 996 return true; 997} 998 999/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1000/// BB2, hoist any common code in the two blocks up into the branch block. The 1001/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1002static bool HoistThenElseCodeToIf(BranchInst *BI) { 1003 // This does very trivial matching, with limited scanning, to find identical 1004 // instructions in the two blocks. In particular, we don't want to get into 1005 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1006 // such, we currently just scan for obviously identical instructions in an 1007 // identical order. 1008 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1009 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1010 1011 BasicBlock::iterator BB1_Itr = BB1->begin(); 1012 BasicBlock::iterator BB2_Itr = BB2->begin(); 1013 1014 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1015 // Skip debug info if it is not identical. 1016 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1017 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1018 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1019 while (isa<DbgInfoIntrinsic>(I1)) 1020 I1 = BB1_Itr++; 1021 while (isa<DbgInfoIntrinsic>(I2)) 1022 I2 = BB2_Itr++; 1023 } 1024 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1025 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1026 return false; 1027 1028 BasicBlock *BIParent = BI->getParent(); 1029 1030 bool Changed = false; 1031 do { 1032 // If we are hoisting the terminator instruction, don't move one (making a 1033 // broken BB), instead clone it, and remove BI. 1034 if (isa<TerminatorInst>(I1)) 1035 goto HoistTerminator; 1036 1037 // For a normal instruction, we just move one to right before the branch, 1038 // then replace all uses of the other with the first. Finally, we remove 1039 // the now redundant second instruction. 1040 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1041 if (!I2->use_empty()) 1042 I2->replaceAllUsesWith(I1); 1043 I1->intersectOptionalDataWith(I2); 1044 I2->eraseFromParent(); 1045 Changed = true; 1046 1047 I1 = BB1_Itr++; 1048 I2 = BB2_Itr++; 1049 // Skip debug info if it is not identical. 1050 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1051 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1052 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1053 while (isa<DbgInfoIntrinsic>(I1)) 1054 I1 = BB1_Itr++; 1055 while (isa<DbgInfoIntrinsic>(I2)) 1056 I2 = BB2_Itr++; 1057 } 1058 } while (I1->isIdenticalToWhenDefined(I2)); 1059 1060 return true; 1061 1062HoistTerminator: 1063 // It may not be possible to hoist an invoke. 1064 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1065 return Changed; 1066 1067 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1068 PHINode *PN; 1069 for (BasicBlock::iterator BBI = SI->begin(); 1070 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1071 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1072 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1073 if (BB1V == BB2V) 1074 continue; 1075 1076 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V)) 1077 return Changed; 1078 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V)) 1079 return Changed; 1080 } 1081 } 1082 1083 // Okay, it is safe to hoist the terminator. 1084 Instruction *NT = I1->clone(); 1085 BIParent->getInstList().insert(BI, NT); 1086 if (!NT->getType()->isVoidTy()) { 1087 I1->replaceAllUsesWith(NT); 1088 I2->replaceAllUsesWith(NT); 1089 NT->takeName(I1); 1090 } 1091 1092 IRBuilder<true, NoFolder> Builder(NT); 1093 // Hoisting one of the terminators from our successor is a great thing. 1094 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1095 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1096 // nodes, so we insert select instruction to compute the final result. 1097 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1098 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1099 PHINode *PN; 1100 for (BasicBlock::iterator BBI = SI->begin(); 1101 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1102 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1103 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1104 if (BB1V == BB2V) continue; 1105 1106 // These values do not agree. Insert a select instruction before NT 1107 // that determines the right value. 1108 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1109 if (SI == 0) 1110 SI = cast<SelectInst> 1111 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1112 BB1V->getName()+"."+BB2V->getName())); 1113 1114 // Make the PHI node use the select for all incoming values for BB1/BB2 1115 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1116 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1117 PN->setIncomingValue(i, SI); 1118 } 1119 } 1120 1121 // Update any PHI nodes in our new successors. 1122 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1123 AddPredecessorToBlock(*SI, BIParent, BB1); 1124 1125 EraseTerminatorInstAndDCECond(BI); 1126 return true; 1127} 1128 1129/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1130/// check whether BBEnd has only two predecessors and the other predecessor 1131/// ends with an unconditional branch. If it is true, sink any common code 1132/// in the two predecessors to BBEnd. 1133static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1134 assert(BI1->isUnconditional()); 1135 BasicBlock *BB1 = BI1->getParent(); 1136 BasicBlock *BBEnd = BI1->getSuccessor(0); 1137 1138 // Check that BBEnd has two predecessors and the other predecessor ends with 1139 // an unconditional branch. 1140 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1141 BasicBlock *Pred0 = *PI++; 1142 if (PI == PE) // Only one predecessor. 1143 return false; 1144 BasicBlock *Pred1 = *PI++; 1145 if (PI != PE) // More than two predecessors. 1146 return false; 1147 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1148 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1149 if (!BI2 || !BI2->isUnconditional()) 1150 return false; 1151 1152 // Gather the PHI nodes in BBEnd. 1153 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1154 Instruction *FirstNonPhiInBBEnd = 0; 1155 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1156 I != E; ++I) { 1157 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1158 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1159 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1160 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1161 } else { 1162 FirstNonPhiInBBEnd = &*I; 1163 break; 1164 } 1165 } 1166 if (!FirstNonPhiInBBEnd) 1167 return false; 1168 1169 1170 // This does very trivial matching, with limited scanning, to find identical 1171 // instructions in the two blocks. We scan backward for obviously identical 1172 // instructions in an identical order. 1173 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1174 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1175 RE2 = BB2->getInstList().rend(); 1176 // Skip debug info. 1177 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1178 if (RI1 == RE1) 1179 return false; 1180 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1181 if (RI2 == RE2) 1182 return false; 1183 // Skip the unconditional branches. 1184 ++RI1; 1185 ++RI2; 1186 1187 bool Changed = false; 1188 while (RI1 != RE1 && RI2 != RE2) { 1189 // Skip debug info. 1190 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1191 if (RI1 == RE1) 1192 return Changed; 1193 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1194 if (RI2 == RE2) 1195 return Changed; 1196 1197 Instruction *I1 = &*RI1, *I2 = &*RI2; 1198 // I1 and I2 should have a single use in the same PHI node, and they 1199 // perform the same operation. 1200 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1201 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1202 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1203 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1204 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1205 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1206 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1207 !I1->hasOneUse() || !I2->hasOneUse() || 1208 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1209 MapValueFromBB1ToBB2[I1].first != I2) 1210 return Changed; 1211 1212 // Check whether we should swap the operands of ICmpInst. 1213 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1214 bool SwapOpnds = false; 1215 if (ICmp1 && ICmp2 && 1216 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1217 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1218 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1219 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1220 ICmp2->swapOperands(); 1221 SwapOpnds = true; 1222 } 1223 if (!I1->isSameOperationAs(I2)) { 1224 if (SwapOpnds) 1225 ICmp2->swapOperands(); 1226 return Changed; 1227 } 1228 1229 // The operands should be either the same or they need to be generated 1230 // with a PHI node after sinking. We only handle the case where there is 1231 // a single pair of different operands. 1232 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1233 unsigned Op1Idx = 0; 1234 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1235 if (I1->getOperand(I) == I2->getOperand(I)) 1236 continue; 1237 // Early exit if we have more-than one pair of different operands or 1238 // the different operand is already in MapValueFromBB1ToBB2. 1239 // Early exit if we need a PHI node to replace a constant. 1240 if (DifferentOp1 || 1241 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1242 MapValueFromBB1ToBB2.end() || 1243 isa<Constant>(I1->getOperand(I)) || 1244 isa<Constant>(I2->getOperand(I))) { 1245 // If we can't sink the instructions, undo the swapping. 1246 if (SwapOpnds) 1247 ICmp2->swapOperands(); 1248 return Changed; 1249 } 1250 DifferentOp1 = I1->getOperand(I); 1251 Op1Idx = I; 1252 DifferentOp2 = I2->getOperand(I); 1253 } 1254 1255 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1256 // remove (I1, I2) from MapValueFromBB1ToBB2. 1257 if (DifferentOp1) { 1258 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1259 DifferentOp1->getName() + ".sink", 1260 BBEnd->begin()); 1261 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1262 // I1 should use NewPN instead of DifferentOp1. 1263 I1->setOperand(Op1Idx, NewPN); 1264 NewPN->addIncoming(DifferentOp1, BB1); 1265 NewPN->addIncoming(DifferentOp2, BB2); 1266 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1267 } 1268 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1269 MapValueFromBB1ToBB2.erase(I1); 1270 1271 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1272 DEBUG(dbgs() << " " << *I2 << "\n";); 1273 // We need to update RE1 and RE2 if we are going to sink the first 1274 // instruction in the basic block down. 1275 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1276 // Sink the instruction. 1277 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1278 if (!OldPN->use_empty()) 1279 OldPN->replaceAllUsesWith(I1); 1280 OldPN->eraseFromParent(); 1281 1282 if (!I2->use_empty()) 1283 I2->replaceAllUsesWith(I1); 1284 I1->intersectOptionalDataWith(I2); 1285 I2->eraseFromParent(); 1286 1287 if (UpdateRE1) 1288 RE1 = BB1->getInstList().rend(); 1289 if (UpdateRE2) 1290 RE2 = BB2->getInstList().rend(); 1291 FirstNonPhiInBBEnd = I1; 1292 NumSinkCommons++; 1293 Changed = true; 1294 } 1295 return Changed; 1296} 1297 1298/// \brief Determine if we can hoist sink a sole store instruction out of a 1299/// conditional block. 1300/// 1301/// We are looking for code like the following: 1302/// BrBB: 1303/// store i32 %add, i32* %arrayidx2 1304/// ... // No other stores or function calls (we could be calling a memory 1305/// ... // function). 1306/// %cmp = icmp ult %x, %y 1307/// br i1 %cmp, label %EndBB, label %ThenBB 1308/// ThenBB: 1309/// store i32 %add5, i32* %arrayidx2 1310/// br label EndBB 1311/// EndBB: 1312/// ... 1313/// We are going to transform this into: 1314/// BrBB: 1315/// store i32 %add, i32* %arrayidx2 1316/// ... // 1317/// %cmp = icmp ult %x, %y 1318/// %add.add5 = select i1 %cmp, i32 %add, %add5 1319/// store i32 %add.add5, i32* %arrayidx2 1320/// ... 1321/// 1322/// \return The pointer to the value of the previous store if the store can be 1323/// hoisted into the predecessor block. 0 otherwise. 1324static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, 1325 BasicBlock *StoreBB, BasicBlock *EndBB) { 1326 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); 1327 if (!StoreToHoist) 1328 return 0; 1329 1330 // Volatile or atomic. 1331 if (!StoreToHoist->isSimple()) 1332 return 0; 1333 1334 Value *StorePtr = StoreToHoist->getPointerOperand(); 1335 1336 // Look for a store to the same pointer in BrBB. 1337 unsigned MaxNumInstToLookAt = 10; 1338 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), 1339 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { 1340 Instruction *CurI = &*RI; 1341 1342 // Could be calling an instruction that effects memory like free(). 1343 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) 1344 return 0; 1345 1346 StoreInst *SI = dyn_cast<StoreInst>(CurI); 1347 // Found the previous store make sure it stores to the same location. 1348 if (SI && SI->getPointerOperand() == StorePtr) 1349 // Found the previous store, return its value operand. 1350 return SI->getValueOperand(); 1351 else if (SI) 1352 return 0; // Unknown store. 1353 } 1354 1355 return 0; 1356} 1357 1358/// \brief Speculate a conditional basic block flattening the CFG. 1359/// 1360/// Note that this is a very risky transform currently. Speculating 1361/// instructions like this is most often not desirable. Instead, there is an MI 1362/// pass which can do it with full awareness of the resource constraints. 1363/// However, some cases are "obvious" and we should do directly. An example of 1364/// this is speculating a single, reasonably cheap instruction. 1365/// 1366/// There is only one distinct advantage to flattening the CFG at the IR level: 1367/// it makes very common but simplistic optimizations such as are common in 1368/// instcombine and the DAG combiner more powerful by removing CFG edges and 1369/// modeling their effects with easier to reason about SSA value graphs. 1370/// 1371/// 1372/// An illustration of this transform is turning this IR: 1373/// \code 1374/// BB: 1375/// %cmp = icmp ult %x, %y 1376/// br i1 %cmp, label %EndBB, label %ThenBB 1377/// ThenBB: 1378/// %sub = sub %x, %y 1379/// br label BB2 1380/// EndBB: 1381/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1382/// ... 1383/// \endcode 1384/// 1385/// Into this IR: 1386/// \code 1387/// BB: 1388/// %cmp = icmp ult %x, %y 1389/// %sub = sub %x, %y 1390/// %cond = select i1 %cmp, 0, %sub 1391/// ... 1392/// \endcode 1393/// 1394/// \returns true if the conditional block is removed. 1395static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1396 // Be conservative for now. FP select instruction can often be expensive. 1397 Value *BrCond = BI->getCondition(); 1398 if (isa<FCmpInst>(BrCond)) 1399 return false; 1400 1401 BasicBlock *BB = BI->getParent(); 1402 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1403 1404 // If ThenBB is actually on the false edge of the conditional branch, remember 1405 // to swap the select operands later. 1406 bool Invert = false; 1407 if (ThenBB != BI->getSuccessor(0)) { 1408 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1409 Invert = true; 1410 } 1411 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1412 1413 // Keep a count of how many times instructions are used within CondBB when 1414 // they are candidates for sinking into CondBB. Specifically: 1415 // - They are defined in BB, and 1416 // - They have no side effects, and 1417 // - All of their uses are in CondBB. 1418 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1419 1420 unsigned SpeculationCost = 0; 1421 Value *SpeculatedStoreValue = 0; 1422 StoreInst *SpeculatedStore = 0; 1423 for (BasicBlock::iterator BBI = ThenBB->begin(), 1424 BBE = llvm::prior(ThenBB->end()); 1425 BBI != BBE; ++BBI) { 1426 Instruction *I = BBI; 1427 // Skip debug info. 1428 if (isa<DbgInfoIntrinsic>(I)) 1429 continue; 1430 1431 // Only speculatively execution a single instruction (not counting the 1432 // terminator) for now. 1433 ++SpeculationCost; 1434 if (SpeculationCost > 1) 1435 return false; 1436 1437 // Don't hoist the instruction if it's unsafe or expensive. 1438 if (!isSafeToSpeculativelyExecute(I) && 1439 !(HoistCondStores && 1440 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, 1441 EndBB)))) 1442 return false; 1443 if (!SpeculatedStoreValue && 1444 ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1445 return false; 1446 1447 // Store the store speculation candidate. 1448 if (SpeculatedStoreValue) 1449 SpeculatedStore = cast<StoreInst>(I); 1450 1451 // Do not hoist the instruction if any of its operands are defined but not 1452 // used in BB. The transformation will prevent the operand from 1453 // being sunk into the use block. 1454 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1455 i != e; ++i) { 1456 Instruction *OpI = dyn_cast<Instruction>(*i); 1457 if (!OpI || OpI->getParent() != BB || 1458 OpI->mayHaveSideEffects()) 1459 continue; // Not a candidate for sinking. 1460 1461 ++SinkCandidateUseCounts[OpI]; 1462 } 1463 } 1464 1465 // Consider any sink candidates which are only used in CondBB as costs for 1466 // speculation. Note, while we iterate over a DenseMap here, we are summing 1467 // and so iteration order isn't significant. 1468 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1469 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1470 I != E; ++I) 1471 if (I->first->getNumUses() == I->second) { 1472 ++SpeculationCost; 1473 if (SpeculationCost > 1) 1474 return false; 1475 } 1476 1477 // Check that the PHI nodes can be converted to selects. 1478 bool HaveRewritablePHIs = false; 1479 for (BasicBlock::iterator I = EndBB->begin(); 1480 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1481 Value *OrigV = PN->getIncomingValueForBlock(BB); 1482 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1483 1484 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf. 1485 // Skip PHIs which are trivial. 1486 if (ThenV == OrigV) 1487 continue; 1488 1489 HaveRewritablePHIs = true; 1490 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV); 1491 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV); 1492 if (!OrigCE && !ThenCE) 1493 continue; // Known safe and cheap. 1494 1495 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) || 1496 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE))) 1497 return false; 1498 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0; 1499 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0; 1500 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold) 1501 return false; 1502 1503 // Account for the cost of an unfolded ConstantExpr which could end up 1504 // getting expanded into Instructions. 1505 // FIXME: This doesn't account for how many operations are combined in the 1506 // constant expression. 1507 ++SpeculationCost; 1508 if (SpeculationCost > 1) 1509 return false; 1510 } 1511 1512 // If there are no PHIs to process, bail early. This helps ensure idempotence 1513 // as well. 1514 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) 1515 return false; 1516 1517 // If we get here, we can hoist the instruction and if-convert. 1518 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1519 1520 // Insert a select of the value of the speculated store. 1521 if (SpeculatedStoreValue) { 1522 IRBuilder<true, NoFolder> Builder(BI); 1523 Value *TrueV = SpeculatedStore->getValueOperand(); 1524 Value *FalseV = SpeculatedStoreValue; 1525 if (Invert) 1526 std::swap(TrueV, FalseV); 1527 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + 1528 "." + FalseV->getName()); 1529 SpeculatedStore->setOperand(0, S); 1530 } 1531 1532 // Hoist the instructions. 1533 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1534 llvm::prior(ThenBB->end())); 1535 1536 // Insert selects and rewrite the PHI operands. 1537 IRBuilder<true, NoFolder> Builder(BI); 1538 for (BasicBlock::iterator I = EndBB->begin(); 1539 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1540 unsigned OrigI = PN->getBasicBlockIndex(BB); 1541 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1542 Value *OrigV = PN->getIncomingValue(OrigI); 1543 Value *ThenV = PN->getIncomingValue(ThenI); 1544 1545 // Skip PHIs which are trivial. 1546 if (OrigV == ThenV) 1547 continue; 1548 1549 // Create a select whose true value is the speculatively executed value and 1550 // false value is the preexisting value. Swap them if the branch 1551 // destinations were inverted. 1552 Value *TrueV = ThenV, *FalseV = OrigV; 1553 if (Invert) 1554 std::swap(TrueV, FalseV); 1555 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1556 TrueV->getName() + "." + FalseV->getName()); 1557 PN->setIncomingValue(OrigI, V); 1558 PN->setIncomingValue(ThenI, V); 1559 } 1560 1561 ++NumSpeculations; 1562 return true; 1563} 1564 1565/// \returns True if this block contains a CallInst with the NoDuplicate 1566/// attribute. 1567static bool HasNoDuplicateCall(const BasicBlock *BB) { 1568 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 1569 const CallInst *CI = dyn_cast<CallInst>(I); 1570 if (!CI) 1571 continue; 1572 if (CI->cannotDuplicate()) 1573 return true; 1574 } 1575 return false; 1576} 1577 1578/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1579/// across this block. 1580static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1581 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1582 unsigned Size = 0; 1583 1584 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1585 if (isa<DbgInfoIntrinsic>(BBI)) 1586 continue; 1587 if (Size > 10) return false; // Don't clone large BB's. 1588 ++Size; 1589 1590 // We can only support instructions that do not define values that are 1591 // live outside of the current basic block. 1592 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1593 UI != E; ++UI) { 1594 Instruction *U = cast<Instruction>(*UI); 1595 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1596 } 1597 1598 // Looks ok, continue checking. 1599 } 1600 1601 return true; 1602} 1603 1604/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1605/// that is defined in the same block as the branch and if any PHI entries are 1606/// constants, thread edges corresponding to that entry to be branches to their 1607/// ultimate destination. 1608static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1609 BasicBlock *BB = BI->getParent(); 1610 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1611 // NOTE: we currently cannot transform this case if the PHI node is used 1612 // outside of the block. 1613 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1614 return false; 1615 1616 // Degenerate case of a single entry PHI. 1617 if (PN->getNumIncomingValues() == 1) { 1618 FoldSingleEntryPHINodes(PN->getParent()); 1619 return true; 1620 } 1621 1622 // Now we know that this block has multiple preds and two succs. 1623 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1624 1625 if (HasNoDuplicateCall(BB)) return false; 1626 1627 // Okay, this is a simple enough basic block. See if any phi values are 1628 // constants. 1629 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1630 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1631 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1632 1633 // Okay, we now know that all edges from PredBB should be revectored to 1634 // branch to RealDest. 1635 BasicBlock *PredBB = PN->getIncomingBlock(i); 1636 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1637 1638 if (RealDest == BB) continue; // Skip self loops. 1639 // Skip if the predecessor's terminator is an indirect branch. 1640 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1641 1642 // The dest block might have PHI nodes, other predecessors and other 1643 // difficult cases. Instead of being smart about this, just insert a new 1644 // block that jumps to the destination block, effectively splitting 1645 // the edge we are about to create. 1646 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1647 RealDest->getName()+".critedge", 1648 RealDest->getParent(), RealDest); 1649 BranchInst::Create(RealDest, EdgeBB); 1650 1651 // Update PHI nodes. 1652 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1653 1654 // BB may have instructions that are being threaded over. Clone these 1655 // instructions into EdgeBB. We know that there will be no uses of the 1656 // cloned instructions outside of EdgeBB. 1657 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1658 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1659 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1660 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1661 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1662 continue; 1663 } 1664 // Clone the instruction. 1665 Instruction *N = BBI->clone(); 1666 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1667 1668 // Update operands due to translation. 1669 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1670 i != e; ++i) { 1671 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1672 if (PI != TranslateMap.end()) 1673 *i = PI->second; 1674 } 1675 1676 // Check for trivial simplification. 1677 if (Value *V = SimplifyInstruction(N, TD)) { 1678 TranslateMap[BBI] = V; 1679 delete N; // Instruction folded away, don't need actual inst 1680 } else { 1681 // Insert the new instruction into its new home. 1682 EdgeBB->getInstList().insert(InsertPt, N); 1683 if (!BBI->use_empty()) 1684 TranslateMap[BBI] = N; 1685 } 1686 } 1687 1688 // Loop over all of the edges from PredBB to BB, changing them to branch 1689 // to EdgeBB instead. 1690 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1691 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1692 if (PredBBTI->getSuccessor(i) == BB) { 1693 BB->removePredecessor(PredBB); 1694 PredBBTI->setSuccessor(i, EdgeBB); 1695 } 1696 1697 // Recurse, simplifying any other constants. 1698 return FoldCondBranchOnPHI(BI, TD) | true; 1699 } 1700 1701 return false; 1702} 1703 1704/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1705/// PHI node, see if we can eliminate it. 1706static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1707 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1708 // statement", which has a very simple dominance structure. Basically, we 1709 // are trying to find the condition that is being branched on, which 1710 // subsequently causes this merge to happen. We really want control 1711 // dependence information for this check, but simplifycfg can't keep it up 1712 // to date, and this catches most of the cases we care about anyway. 1713 BasicBlock *BB = PN->getParent(); 1714 BasicBlock *IfTrue, *IfFalse; 1715 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1716 if (!IfCond || 1717 // Don't bother if the branch will be constant folded trivially. 1718 isa<ConstantInt>(IfCond)) 1719 return false; 1720 1721 // Okay, we found that we can merge this two-entry phi node into a select. 1722 // Doing so would require us to fold *all* two entry phi nodes in this block. 1723 // At some point this becomes non-profitable (particularly if the target 1724 // doesn't support cmov's). Only do this transformation if there are two or 1725 // fewer PHI nodes in this block. 1726 unsigned NumPhis = 0; 1727 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1728 if (NumPhis > 2) 1729 return false; 1730 1731 // Loop over the PHI's seeing if we can promote them all to select 1732 // instructions. While we are at it, keep track of the instructions 1733 // that need to be moved to the dominating block. 1734 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1735 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1736 MaxCostVal1 = PHINodeFoldingThreshold; 1737 1738 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1739 PHINode *PN = cast<PHINode>(II++); 1740 if (Value *V = SimplifyInstruction(PN, TD)) { 1741 PN->replaceAllUsesWith(V); 1742 PN->eraseFromParent(); 1743 continue; 1744 } 1745 1746 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1747 MaxCostVal0) || 1748 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1749 MaxCostVal1)) 1750 return false; 1751 } 1752 1753 // If we folded the first phi, PN dangles at this point. Refresh it. If 1754 // we ran out of PHIs then we simplified them all. 1755 PN = dyn_cast<PHINode>(BB->begin()); 1756 if (PN == 0) return true; 1757 1758 // Don't fold i1 branches on PHIs which contain binary operators. These can 1759 // often be turned into switches and other things. 1760 if (PN->getType()->isIntegerTy(1) && 1761 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1762 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1763 isa<BinaryOperator>(IfCond))) 1764 return false; 1765 1766 // If we all PHI nodes are promotable, check to make sure that all 1767 // instructions in the predecessor blocks can be promoted as well. If 1768 // not, we won't be able to get rid of the control flow, so it's not 1769 // worth promoting to select instructions. 1770 BasicBlock *DomBlock = 0; 1771 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1772 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1773 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1774 IfBlock1 = 0; 1775 } else { 1776 DomBlock = *pred_begin(IfBlock1); 1777 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1778 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1779 // This is not an aggressive instruction that we can promote. 1780 // Because of this, we won't be able to get rid of the control 1781 // flow, so the xform is not worth it. 1782 return false; 1783 } 1784 } 1785 1786 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1787 IfBlock2 = 0; 1788 } else { 1789 DomBlock = *pred_begin(IfBlock2); 1790 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1791 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1792 // This is not an aggressive instruction that we can promote. 1793 // Because of this, we won't be able to get rid of the control 1794 // flow, so the xform is not worth it. 1795 return false; 1796 } 1797 } 1798 1799 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1800 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1801 1802 // If we can still promote the PHI nodes after this gauntlet of tests, 1803 // do all of the PHI's now. 1804 Instruction *InsertPt = DomBlock->getTerminator(); 1805 IRBuilder<true, NoFolder> Builder(InsertPt); 1806 1807 // Move all 'aggressive' instructions, which are defined in the 1808 // conditional parts of the if's up to the dominating block. 1809 if (IfBlock1) 1810 DomBlock->getInstList().splice(InsertPt, 1811 IfBlock1->getInstList(), IfBlock1->begin(), 1812 IfBlock1->getTerminator()); 1813 if (IfBlock2) 1814 DomBlock->getInstList().splice(InsertPt, 1815 IfBlock2->getInstList(), IfBlock2->begin(), 1816 IfBlock2->getTerminator()); 1817 1818 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1819 // Change the PHI node into a select instruction. 1820 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1821 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1822 1823 SelectInst *NV = 1824 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1825 PN->replaceAllUsesWith(NV); 1826 NV->takeName(PN); 1827 PN->eraseFromParent(); 1828 } 1829 1830 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1831 // has been flattened. Change DomBlock to jump directly to our new block to 1832 // avoid other simplifycfg's kicking in on the diamond. 1833 TerminatorInst *OldTI = DomBlock->getTerminator(); 1834 Builder.SetInsertPoint(OldTI); 1835 Builder.CreateBr(BB); 1836 OldTI->eraseFromParent(); 1837 return true; 1838} 1839 1840/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1841/// to two returning blocks, try to merge them together into one return, 1842/// introducing a select if the return values disagree. 1843static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1844 IRBuilder<> &Builder) { 1845 assert(BI->isConditional() && "Must be a conditional branch"); 1846 BasicBlock *TrueSucc = BI->getSuccessor(0); 1847 BasicBlock *FalseSucc = BI->getSuccessor(1); 1848 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1849 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1850 1851 // Check to ensure both blocks are empty (just a return) or optionally empty 1852 // with PHI nodes. If there are other instructions, merging would cause extra 1853 // computation on one path or the other. 1854 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1855 return false; 1856 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1857 return false; 1858 1859 Builder.SetInsertPoint(BI); 1860 // Okay, we found a branch that is going to two return nodes. If 1861 // there is no return value for this function, just change the 1862 // branch into a return. 1863 if (FalseRet->getNumOperands() == 0) { 1864 TrueSucc->removePredecessor(BI->getParent()); 1865 FalseSucc->removePredecessor(BI->getParent()); 1866 Builder.CreateRetVoid(); 1867 EraseTerminatorInstAndDCECond(BI); 1868 return true; 1869 } 1870 1871 // Otherwise, figure out what the true and false return values are 1872 // so we can insert a new select instruction. 1873 Value *TrueValue = TrueRet->getReturnValue(); 1874 Value *FalseValue = FalseRet->getReturnValue(); 1875 1876 // Unwrap any PHI nodes in the return blocks. 1877 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1878 if (TVPN->getParent() == TrueSucc) 1879 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1880 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1881 if (FVPN->getParent() == FalseSucc) 1882 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1883 1884 // In order for this transformation to be safe, we must be able to 1885 // unconditionally execute both operands to the return. This is 1886 // normally the case, but we could have a potentially-trapping 1887 // constant expression that prevents this transformation from being 1888 // safe. 1889 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1890 if (TCV->canTrap()) 1891 return false; 1892 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1893 if (FCV->canTrap()) 1894 return false; 1895 1896 // Okay, we collected all the mapped values and checked them for sanity, and 1897 // defined to really do this transformation. First, update the CFG. 1898 TrueSucc->removePredecessor(BI->getParent()); 1899 FalseSucc->removePredecessor(BI->getParent()); 1900 1901 // Insert select instructions where needed. 1902 Value *BrCond = BI->getCondition(); 1903 if (TrueValue) { 1904 // Insert a select if the results differ. 1905 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1906 } else if (isa<UndefValue>(TrueValue)) { 1907 TrueValue = FalseValue; 1908 } else { 1909 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1910 FalseValue, "retval"); 1911 } 1912 } 1913 1914 Value *RI = !TrueValue ? 1915 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1916 1917 (void) RI; 1918 1919 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1920 << "\n " << *BI << "NewRet = " << *RI 1921 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1922 1923 EraseTerminatorInstAndDCECond(BI); 1924 1925 return true; 1926} 1927 1928/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1929/// probabilities of the branch taking each edge. Fills in the two APInt 1930/// parameters and return true, or returns false if no or invalid metadata was 1931/// found. 1932static bool ExtractBranchMetadata(BranchInst *BI, 1933 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1934 assert(BI->isConditional() && 1935 "Looking for probabilities on unconditional branch?"); 1936 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1937 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1938 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1939 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1940 if (!CITrue || !CIFalse) return false; 1941 ProbTrue = CITrue->getValue().getZExtValue(); 1942 ProbFalse = CIFalse->getValue().getZExtValue(); 1943 return true; 1944} 1945 1946/// checkCSEInPredecessor - Return true if the given instruction is available 1947/// in its predecessor block. If yes, the instruction will be removed. 1948/// 1949static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1950 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1951 return false; 1952 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1953 Instruction *PBI = &*I; 1954 // Check whether Inst and PBI generate the same value. 1955 if (Inst->isIdenticalTo(PBI)) { 1956 Inst->replaceAllUsesWith(PBI); 1957 Inst->eraseFromParent(); 1958 return true; 1959 } 1960 } 1961 return false; 1962} 1963 1964/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1965/// predecessor branches to us and one of our successors, fold the block into 1966/// the predecessor and use logical operations to pick the right destination. 1967bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1968 BasicBlock *BB = BI->getParent(); 1969 1970 Instruction *Cond = 0; 1971 if (BI->isConditional()) 1972 Cond = dyn_cast<Instruction>(BI->getCondition()); 1973 else { 1974 // For unconditional branch, check for a simple CFG pattern, where 1975 // BB has a single predecessor and BB's successor is also its predecessor's 1976 // successor. If such pattern exisits, check for CSE between BB and its 1977 // predecessor. 1978 if (BasicBlock *PB = BB->getSinglePredecessor()) 1979 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1980 if (PBI->isConditional() && 1981 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1982 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1983 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1984 I != E; ) { 1985 Instruction *Curr = I++; 1986 if (isa<CmpInst>(Curr)) { 1987 Cond = Curr; 1988 break; 1989 } 1990 // Quit if we can't remove this instruction. 1991 if (!checkCSEInPredecessor(Curr, PB)) 1992 return false; 1993 } 1994 } 1995 1996 if (Cond == 0) 1997 return false; 1998 } 1999 2000 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 2001 Cond->getParent() != BB || !Cond->hasOneUse()) 2002 return false; 2003 2004 // Only allow this if the condition is a simple instruction that can be 2005 // executed unconditionally. It must be in the same block as the branch, and 2006 // must be at the front of the block. 2007 BasicBlock::iterator FrontIt = BB->front(); 2008 2009 // Ignore dbg intrinsics. 2010 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2011 2012 // Allow a single instruction to be hoisted in addition to the compare 2013 // that feeds the branch. We later ensure that any values that _it_ uses 2014 // were also live in the predecessor, so that we don't unnecessarily create 2015 // register pressure or inhibit out-of-order execution. 2016 Instruction *BonusInst = 0; 2017 if (&*FrontIt != Cond && 2018 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 2019 isSafeToSpeculativelyExecute(FrontIt)) { 2020 BonusInst = &*FrontIt; 2021 ++FrontIt; 2022 2023 // Ignore dbg intrinsics. 2024 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2025 } 2026 2027 // Only a single bonus inst is allowed. 2028 if (&*FrontIt != Cond) 2029 return false; 2030 2031 // Make sure the instruction after the condition is the cond branch. 2032 BasicBlock::iterator CondIt = Cond; ++CondIt; 2033 2034 // Ingore dbg intrinsics. 2035 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 2036 2037 if (&*CondIt != BI) 2038 return false; 2039 2040 // Cond is known to be a compare or binary operator. Check to make sure that 2041 // neither operand is a potentially-trapping constant expression. 2042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 2043 if (CE->canTrap()) 2044 return false; 2045 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 2046 if (CE->canTrap()) 2047 return false; 2048 2049 // Finally, don't infinitely unroll conditional loops. 2050 BasicBlock *TrueDest = BI->getSuccessor(0); 2051 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 2052 if (TrueDest == BB || FalseDest == BB) 2053 return false; 2054 2055 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2056 BasicBlock *PredBlock = *PI; 2057 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 2058 2059 // Check that we have two conditional branches. If there is a PHI node in 2060 // the common successor, verify that the same value flows in from both 2061 // blocks. 2062 SmallVector<PHINode*, 4> PHIs; 2063 if (PBI == 0 || PBI->isUnconditional() || 2064 (BI->isConditional() && 2065 !SafeToMergeTerminators(BI, PBI)) || 2066 (!BI->isConditional() && 2067 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2068 continue; 2069 2070 // Determine if the two branches share a common destination. 2071 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2072 bool InvertPredCond = false; 2073 2074 if (BI->isConditional()) { 2075 if (PBI->getSuccessor(0) == TrueDest) 2076 Opc = Instruction::Or; 2077 else if (PBI->getSuccessor(1) == FalseDest) 2078 Opc = Instruction::And; 2079 else if (PBI->getSuccessor(0) == FalseDest) 2080 Opc = Instruction::And, InvertPredCond = true; 2081 else if (PBI->getSuccessor(1) == TrueDest) 2082 Opc = Instruction::Or, InvertPredCond = true; 2083 else 2084 continue; 2085 } else { 2086 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2087 continue; 2088 } 2089 2090 // Ensure that any values used in the bonus instruction are also used 2091 // by the terminator of the predecessor. This means that those values 2092 // must already have been resolved, so we won't be inhibiting the 2093 // out-of-order core by speculating them earlier. We also allow 2094 // instructions that are used by the terminator's condition because it 2095 // exposes more merging opportunities. 2096 bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() && 2097 *BonusInst->use_begin() == Cond); 2098 2099 if (BonusInst && !UsedByBranch) { 2100 // Collect the values used by the bonus inst 2101 SmallPtrSet<Value*, 4> UsedValues; 2102 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2103 OE = BonusInst->op_end(); OI != OE; ++OI) { 2104 Value *V = *OI; 2105 if (!isa<Constant>(V) && !isa<Argument>(V)) 2106 UsedValues.insert(V); 2107 } 2108 2109 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2110 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2111 2112 // Walk up to four levels back up the use-def chain of the predecessor's 2113 // terminator to see if all those values were used. The choice of four 2114 // levels is arbitrary, to provide a compile-time-cost bound. 2115 while (!Worklist.empty()) { 2116 std::pair<Value*, unsigned> Pair = Worklist.back(); 2117 Worklist.pop_back(); 2118 2119 if (Pair.second >= 4) continue; 2120 UsedValues.erase(Pair.first); 2121 if (UsedValues.empty()) break; 2122 2123 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2124 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2125 OI != OE; ++OI) 2126 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2127 } 2128 } 2129 2130 if (!UsedValues.empty()) return false; 2131 } 2132 2133 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2134 IRBuilder<> Builder(PBI); 2135 2136 // If we need to invert the condition in the pred block to match, do so now. 2137 if (InvertPredCond) { 2138 Value *NewCond = PBI->getCondition(); 2139 2140 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2141 CmpInst *CI = cast<CmpInst>(NewCond); 2142 CI->setPredicate(CI->getInversePredicate()); 2143 } else { 2144 NewCond = Builder.CreateNot(NewCond, 2145 PBI->getCondition()->getName()+".not"); 2146 } 2147 2148 PBI->setCondition(NewCond); 2149 PBI->swapSuccessors(); 2150 } 2151 2152 // If we have a bonus inst, clone it into the predecessor block. 2153 Instruction *NewBonus = 0; 2154 if (BonusInst) { 2155 NewBonus = BonusInst->clone(); 2156 PredBlock->getInstList().insert(PBI, NewBonus); 2157 NewBonus->takeName(BonusInst); 2158 BonusInst->setName(BonusInst->getName()+".old"); 2159 } 2160 2161 // Clone Cond into the predecessor basic block, and or/and the 2162 // two conditions together. 2163 Instruction *New = Cond->clone(); 2164 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2165 PredBlock->getInstList().insert(PBI, New); 2166 New->takeName(Cond); 2167 Cond->setName(New->getName()+".old"); 2168 2169 if (BI->isConditional()) { 2170 Instruction *NewCond = 2171 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2172 New, "or.cond")); 2173 PBI->setCondition(NewCond); 2174 2175 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2176 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2177 PredFalseWeight); 2178 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2179 SuccFalseWeight); 2180 SmallVector<uint64_t, 8> NewWeights; 2181 2182 if (PBI->getSuccessor(0) == BB) { 2183 if (PredHasWeights && SuccHasWeights) { 2184 // PBI: br i1 %x, BB, FalseDest 2185 // BI: br i1 %y, TrueDest, FalseDest 2186 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2187 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2188 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2189 // TrueWeight for PBI * FalseWeight for BI. 2190 // We assume that total weights of a BranchInst can fit into 32 bits. 2191 // Therefore, we will not have overflow using 64-bit arithmetic. 2192 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2193 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2194 } 2195 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2196 PBI->setSuccessor(0, TrueDest); 2197 } 2198 if (PBI->getSuccessor(1) == BB) { 2199 if (PredHasWeights && SuccHasWeights) { 2200 // PBI: br i1 %x, TrueDest, BB 2201 // BI: br i1 %y, TrueDest, FalseDest 2202 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2203 // FalseWeight for PBI * TrueWeight for BI. 2204 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2205 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2206 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2207 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2208 } 2209 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2210 PBI->setSuccessor(1, FalseDest); 2211 } 2212 if (NewWeights.size() == 2) { 2213 // Halve the weights if any of them cannot fit in an uint32_t 2214 FitWeights(NewWeights); 2215 2216 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2217 PBI->setMetadata(LLVMContext::MD_prof, 2218 MDBuilder(BI->getContext()). 2219 createBranchWeights(MDWeights)); 2220 } else 2221 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2222 } else { 2223 // Update PHI nodes in the common successors. 2224 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2225 ConstantInt *PBI_C = cast<ConstantInt>( 2226 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2227 assert(PBI_C->getType()->isIntegerTy(1)); 2228 Instruction *MergedCond = 0; 2229 if (PBI->getSuccessor(0) == TrueDest) { 2230 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2231 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2232 // is false: !PBI_Cond and BI_Value 2233 Instruction *NotCond = 2234 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2235 "not.cond")); 2236 MergedCond = 2237 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2238 NotCond, New, 2239 "and.cond")); 2240 if (PBI_C->isOne()) 2241 MergedCond = 2242 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2243 PBI->getCondition(), MergedCond, 2244 "or.cond")); 2245 } else { 2246 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2247 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2248 // is false: PBI_Cond and BI_Value 2249 MergedCond = 2250 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2251 PBI->getCondition(), New, 2252 "and.cond")); 2253 if (PBI_C->isOne()) { 2254 Instruction *NotCond = 2255 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2256 "not.cond")); 2257 MergedCond = 2258 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2259 NotCond, MergedCond, 2260 "or.cond")); 2261 } 2262 } 2263 // Update PHI Node. 2264 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2265 MergedCond); 2266 } 2267 // Change PBI from Conditional to Unconditional. 2268 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2269 EraseTerminatorInstAndDCECond(PBI); 2270 PBI = New_PBI; 2271 } 2272 2273 // TODO: If BB is reachable from all paths through PredBlock, then we 2274 // could replace PBI's branch probabilities with BI's. 2275 2276 // Copy any debug value intrinsics into the end of PredBlock. 2277 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2278 if (isa<DbgInfoIntrinsic>(*I)) 2279 I->clone()->insertBefore(PBI); 2280 2281 return true; 2282 } 2283 return false; 2284} 2285 2286/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2287/// predecessor of another block, this function tries to simplify it. We know 2288/// that PBI and BI are both conditional branches, and BI is in one of the 2289/// successor blocks of PBI - PBI branches to BI. 2290static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2291 assert(PBI->isConditional() && BI->isConditional()); 2292 BasicBlock *BB = BI->getParent(); 2293 2294 // If this block ends with a branch instruction, and if there is a 2295 // predecessor that ends on a branch of the same condition, make 2296 // this conditional branch redundant. 2297 if (PBI->getCondition() == BI->getCondition() && 2298 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2299 // Okay, the outcome of this conditional branch is statically 2300 // knowable. If this block had a single pred, handle specially. 2301 if (BB->getSinglePredecessor()) { 2302 // Turn this into a branch on constant. 2303 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2304 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2305 CondIsTrue)); 2306 return true; // Nuke the branch on constant. 2307 } 2308 2309 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2310 // in the constant and simplify the block result. Subsequent passes of 2311 // simplifycfg will thread the block. 2312 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2313 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2314 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2315 std::distance(PB, PE), 2316 BI->getCondition()->getName() + ".pr", 2317 BB->begin()); 2318 // Okay, we're going to insert the PHI node. Since PBI is not the only 2319 // predecessor, compute the PHI'd conditional value for all of the preds. 2320 // Any predecessor where the condition is not computable we keep symbolic. 2321 for (pred_iterator PI = PB; PI != PE; ++PI) { 2322 BasicBlock *P = *PI; 2323 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2324 PBI != BI && PBI->isConditional() && 2325 PBI->getCondition() == BI->getCondition() && 2326 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2327 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2328 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2329 CondIsTrue), P); 2330 } else { 2331 NewPN->addIncoming(BI->getCondition(), P); 2332 } 2333 } 2334 2335 BI->setCondition(NewPN); 2336 return true; 2337 } 2338 } 2339 2340 // If this is a conditional branch in an empty block, and if any 2341 // predecessors is a conditional branch to one of our destinations, 2342 // fold the conditions into logical ops and one cond br. 2343 BasicBlock::iterator BBI = BB->begin(); 2344 // Ignore dbg intrinsics. 2345 while (isa<DbgInfoIntrinsic>(BBI)) 2346 ++BBI; 2347 if (&*BBI != BI) 2348 return false; 2349 2350 2351 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2352 if (CE->canTrap()) 2353 return false; 2354 2355 int PBIOp, BIOp; 2356 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2357 PBIOp = BIOp = 0; 2358 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2359 PBIOp = 0, BIOp = 1; 2360 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2361 PBIOp = 1, BIOp = 0; 2362 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2363 PBIOp = BIOp = 1; 2364 else 2365 return false; 2366 2367 // Check to make sure that the other destination of this branch 2368 // isn't BB itself. If so, this is an infinite loop that will 2369 // keep getting unwound. 2370 if (PBI->getSuccessor(PBIOp) == BB) 2371 return false; 2372 2373 // Do not perform this transformation if it would require 2374 // insertion of a large number of select instructions. For targets 2375 // without predication/cmovs, this is a big pessimization. 2376 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2377 2378 unsigned NumPhis = 0; 2379 for (BasicBlock::iterator II = CommonDest->begin(); 2380 isa<PHINode>(II); ++II, ++NumPhis) 2381 if (NumPhis > 2) // Disable this xform. 2382 return false; 2383 2384 // Finally, if everything is ok, fold the branches to logical ops. 2385 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2386 2387 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2388 << "AND: " << *BI->getParent()); 2389 2390 2391 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2392 // branch in it, where one edge (OtherDest) goes back to itself but the other 2393 // exits. We don't *know* that the program avoids the infinite loop 2394 // (even though that seems likely). If we do this xform naively, we'll end up 2395 // recursively unpeeling the loop. Since we know that (after the xform is 2396 // done) that the block *is* infinite if reached, we just make it an obviously 2397 // infinite loop with no cond branch. 2398 if (OtherDest == BB) { 2399 // Insert it at the end of the function, because it's either code, 2400 // or it won't matter if it's hot. :) 2401 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2402 "infloop", BB->getParent()); 2403 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2404 OtherDest = InfLoopBlock; 2405 } 2406 2407 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2408 2409 // BI may have other predecessors. Because of this, we leave 2410 // it alone, but modify PBI. 2411 2412 // Make sure we get to CommonDest on True&True directions. 2413 Value *PBICond = PBI->getCondition(); 2414 IRBuilder<true, NoFolder> Builder(PBI); 2415 if (PBIOp) 2416 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2417 2418 Value *BICond = BI->getCondition(); 2419 if (BIOp) 2420 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2421 2422 // Merge the conditions. 2423 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2424 2425 // Modify PBI to branch on the new condition to the new dests. 2426 PBI->setCondition(Cond); 2427 PBI->setSuccessor(0, CommonDest); 2428 PBI->setSuccessor(1, OtherDest); 2429 2430 // Update branch weight for PBI. 2431 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2432 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2433 PredFalseWeight); 2434 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2435 SuccFalseWeight); 2436 if (PredHasWeights && SuccHasWeights) { 2437 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2438 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2439 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2440 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2441 // The weight to CommonDest should be PredCommon * SuccTotal + 2442 // PredOther * SuccCommon. 2443 // The weight to OtherDest should be PredOther * SuccOther. 2444 SmallVector<uint64_t, 2> NewWeights; 2445 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2446 PredOther * SuccCommon); 2447 NewWeights.push_back(PredOther * SuccOther); 2448 // Halve the weights if any of them cannot fit in an uint32_t 2449 FitWeights(NewWeights); 2450 2451 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2452 PBI->setMetadata(LLVMContext::MD_prof, 2453 MDBuilder(BI->getContext()). 2454 createBranchWeights(MDWeights)); 2455 } 2456 2457 // OtherDest may have phi nodes. If so, add an entry from PBI's 2458 // block that are identical to the entries for BI's block. 2459 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2460 2461 // We know that the CommonDest already had an edge from PBI to 2462 // it. If it has PHIs though, the PHIs may have different 2463 // entries for BB and PBI's BB. If so, insert a select to make 2464 // them agree. 2465 PHINode *PN; 2466 for (BasicBlock::iterator II = CommonDest->begin(); 2467 (PN = dyn_cast<PHINode>(II)); ++II) { 2468 Value *BIV = PN->getIncomingValueForBlock(BB); 2469 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2470 Value *PBIV = PN->getIncomingValue(PBBIdx); 2471 if (BIV != PBIV) { 2472 // Insert a select in PBI to pick the right value. 2473 Value *NV = cast<SelectInst> 2474 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2475 PN->setIncomingValue(PBBIdx, NV); 2476 } 2477 } 2478 2479 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2480 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2481 2482 // This basic block is probably dead. We know it has at least 2483 // one fewer predecessor. 2484 return true; 2485} 2486 2487// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2488// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2489// Takes care of updating the successors and removing the old terminator. 2490// Also makes sure not to introduce new successors by assuming that edges to 2491// non-successor TrueBBs and FalseBBs aren't reachable. 2492static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2493 BasicBlock *TrueBB, BasicBlock *FalseBB, 2494 uint32_t TrueWeight, 2495 uint32_t FalseWeight){ 2496 // Remove any superfluous successor edges from the CFG. 2497 // First, figure out which successors to preserve. 2498 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2499 // successor. 2500 BasicBlock *KeepEdge1 = TrueBB; 2501 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2502 2503 // Then remove the rest. 2504 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2505 BasicBlock *Succ = OldTerm->getSuccessor(I); 2506 // Make sure only to keep exactly one copy of each edge. 2507 if (Succ == KeepEdge1) 2508 KeepEdge1 = 0; 2509 else if (Succ == KeepEdge2) 2510 KeepEdge2 = 0; 2511 else 2512 Succ->removePredecessor(OldTerm->getParent()); 2513 } 2514 2515 IRBuilder<> Builder(OldTerm); 2516 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2517 2518 // Insert an appropriate new terminator. 2519 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2520 if (TrueBB == FalseBB) 2521 // We were only looking for one successor, and it was present. 2522 // Create an unconditional branch to it. 2523 Builder.CreateBr(TrueBB); 2524 else { 2525 // We found both of the successors we were looking for. 2526 // Create a conditional branch sharing the condition of the select. 2527 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2528 if (TrueWeight != FalseWeight) 2529 NewBI->setMetadata(LLVMContext::MD_prof, 2530 MDBuilder(OldTerm->getContext()). 2531 createBranchWeights(TrueWeight, FalseWeight)); 2532 } 2533 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2534 // Neither of the selected blocks were successors, so this 2535 // terminator must be unreachable. 2536 new UnreachableInst(OldTerm->getContext(), OldTerm); 2537 } else { 2538 // One of the selected values was a successor, but the other wasn't. 2539 // Insert an unconditional branch to the one that was found; 2540 // the edge to the one that wasn't must be unreachable. 2541 if (KeepEdge1 == 0) 2542 // Only TrueBB was found. 2543 Builder.CreateBr(TrueBB); 2544 else 2545 // Only FalseBB was found. 2546 Builder.CreateBr(FalseBB); 2547 } 2548 2549 EraseTerminatorInstAndDCECond(OldTerm); 2550 return true; 2551} 2552 2553// SimplifySwitchOnSelect - Replaces 2554// (switch (select cond, X, Y)) on constant X, Y 2555// with a branch - conditional if X and Y lead to distinct BBs, 2556// unconditional otherwise. 2557static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2558 // Check for constant integer values in the select. 2559 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2560 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2561 if (!TrueVal || !FalseVal) 2562 return false; 2563 2564 // Find the relevant condition and destinations. 2565 Value *Condition = Select->getCondition(); 2566 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2567 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2568 2569 // Get weight for TrueBB and FalseBB. 2570 uint32_t TrueWeight = 0, FalseWeight = 0; 2571 SmallVector<uint64_t, 8> Weights; 2572 bool HasWeights = HasBranchWeights(SI); 2573 if (HasWeights) { 2574 GetBranchWeights(SI, Weights); 2575 if (Weights.size() == 1 + SI->getNumCases()) { 2576 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2577 getSuccessorIndex()]; 2578 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2579 getSuccessorIndex()]; 2580 } 2581 } 2582 2583 // Perform the actual simplification. 2584 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2585 TrueWeight, FalseWeight); 2586} 2587 2588// SimplifyIndirectBrOnSelect - Replaces 2589// (indirectbr (select cond, blockaddress(@fn, BlockA), 2590// blockaddress(@fn, BlockB))) 2591// with 2592// (br cond, BlockA, BlockB). 2593static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2594 // Check that both operands of the select are block addresses. 2595 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2596 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2597 if (!TBA || !FBA) 2598 return false; 2599 2600 // Extract the actual blocks. 2601 BasicBlock *TrueBB = TBA->getBasicBlock(); 2602 BasicBlock *FalseBB = FBA->getBasicBlock(); 2603 2604 // Perform the actual simplification. 2605 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2606 0, 0); 2607} 2608 2609/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2610/// instruction (a seteq/setne with a constant) as the only instruction in a 2611/// block that ends with an uncond branch. We are looking for a very specific 2612/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2613/// this case, we merge the first two "or's of icmp" into a switch, but then the 2614/// default value goes to an uncond block with a seteq in it, we get something 2615/// like: 2616/// 2617/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2618/// DEFAULT: 2619/// %tmp = icmp eq i8 %A, 92 2620/// br label %end 2621/// end: 2622/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2623/// 2624/// We prefer to split the edge to 'end' so that there is a true/false entry to 2625/// the PHI, merging the third icmp into the switch. 2626static bool TryToSimplifyUncondBranchWithICmpInIt( 2627 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2628 const DataLayout *TD) { 2629 BasicBlock *BB = ICI->getParent(); 2630 2631 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2632 // complex. 2633 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2634 2635 Value *V = ICI->getOperand(0); 2636 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2637 2638 // The pattern we're looking for is where our only predecessor is a switch on 2639 // 'V' and this block is the default case for the switch. In this case we can 2640 // fold the compared value into the switch to simplify things. 2641 BasicBlock *Pred = BB->getSinglePredecessor(); 2642 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2643 2644 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2645 if (SI->getCondition() != V) 2646 return false; 2647 2648 // If BB is reachable on a non-default case, then we simply know the value of 2649 // V in this block. Substitute it and constant fold the icmp instruction 2650 // away. 2651 if (SI->getDefaultDest() != BB) { 2652 ConstantInt *VVal = SI->findCaseDest(BB); 2653 assert(VVal && "Should have a unique destination value"); 2654 ICI->setOperand(0, VVal); 2655 2656 if (Value *V = SimplifyInstruction(ICI, TD)) { 2657 ICI->replaceAllUsesWith(V); 2658 ICI->eraseFromParent(); 2659 } 2660 // BB is now empty, so it is likely to simplify away. 2661 return SimplifyCFG(BB, TTI, TD) | true; 2662 } 2663 2664 // Ok, the block is reachable from the default dest. If the constant we're 2665 // comparing exists in one of the other edges, then we can constant fold ICI 2666 // and zap it. 2667 if (SI->findCaseValue(Cst) != SI->case_default()) { 2668 Value *V; 2669 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2670 V = ConstantInt::getFalse(BB->getContext()); 2671 else 2672 V = ConstantInt::getTrue(BB->getContext()); 2673 2674 ICI->replaceAllUsesWith(V); 2675 ICI->eraseFromParent(); 2676 // BB is now empty, so it is likely to simplify away. 2677 return SimplifyCFG(BB, TTI, TD) | true; 2678 } 2679 2680 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2681 // the block. 2682 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2683 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2684 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2685 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2686 return false; 2687 2688 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2689 // true in the PHI. 2690 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2691 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2692 2693 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2694 std::swap(DefaultCst, NewCst); 2695 2696 // Replace ICI (which is used by the PHI for the default value) with true or 2697 // false depending on if it is EQ or NE. 2698 ICI->replaceAllUsesWith(DefaultCst); 2699 ICI->eraseFromParent(); 2700 2701 // Okay, the switch goes to this block on a default value. Add an edge from 2702 // the switch to the merge point on the compared value. 2703 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2704 BB->getParent(), BB); 2705 SmallVector<uint64_t, 8> Weights; 2706 bool HasWeights = HasBranchWeights(SI); 2707 if (HasWeights) { 2708 GetBranchWeights(SI, Weights); 2709 if (Weights.size() == 1 + SI->getNumCases()) { 2710 // Split weight for default case to case for "Cst". 2711 Weights[0] = (Weights[0]+1) >> 1; 2712 Weights.push_back(Weights[0]); 2713 2714 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2715 SI->setMetadata(LLVMContext::MD_prof, 2716 MDBuilder(SI->getContext()). 2717 createBranchWeights(MDWeights)); 2718 } 2719 } 2720 SI->addCase(Cst, NewBB); 2721 2722 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2723 Builder.SetInsertPoint(NewBB); 2724 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2725 Builder.CreateBr(SuccBlock); 2726 PHIUse->addIncoming(NewCst, NewBB); 2727 return true; 2728} 2729 2730/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2731/// Check to see if it is branching on an or/and chain of icmp instructions, and 2732/// fold it into a switch instruction if so. 2733static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2734 IRBuilder<> &Builder) { 2735 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2736 if (Cond == 0) return false; 2737 2738 2739 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2740 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2741 // 'setne's and'ed together, collect them. 2742 Value *CompVal = 0; 2743 std::vector<ConstantInt*> Values; 2744 bool TrueWhenEqual = true; 2745 Value *ExtraCase = 0; 2746 unsigned UsedICmps = 0; 2747 2748 if (Cond->getOpcode() == Instruction::Or) { 2749 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2750 UsedICmps); 2751 } else if (Cond->getOpcode() == Instruction::And) { 2752 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2753 UsedICmps); 2754 TrueWhenEqual = false; 2755 } 2756 2757 // If we didn't have a multiply compared value, fail. 2758 if (CompVal == 0) return false; 2759 2760 // Avoid turning single icmps into a switch. 2761 if (UsedICmps <= 1) 2762 return false; 2763 2764 // There might be duplicate constants in the list, which the switch 2765 // instruction can't handle, remove them now. 2766 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2767 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2768 2769 // If Extra was used, we require at least two switch values to do the 2770 // transformation. A switch with one value is just an cond branch. 2771 if (ExtraCase && Values.size() < 2) return false; 2772 2773 // TODO: Preserve branch weight metadata, similarly to how 2774 // FoldValueComparisonIntoPredecessors preserves it. 2775 2776 // Figure out which block is which destination. 2777 BasicBlock *DefaultBB = BI->getSuccessor(1); 2778 BasicBlock *EdgeBB = BI->getSuccessor(0); 2779 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2780 2781 BasicBlock *BB = BI->getParent(); 2782 2783 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2784 << " cases into SWITCH. BB is:\n" << *BB); 2785 2786 // If there are any extra values that couldn't be folded into the switch 2787 // then we evaluate them with an explicit branch first. Split the block 2788 // right before the condbr to handle it. 2789 if (ExtraCase) { 2790 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2791 // Remove the uncond branch added to the old block. 2792 TerminatorInst *OldTI = BB->getTerminator(); 2793 Builder.SetInsertPoint(OldTI); 2794 2795 if (TrueWhenEqual) 2796 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2797 else 2798 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2799 2800 OldTI->eraseFromParent(); 2801 2802 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2803 // for the edge we just added. 2804 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2805 2806 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2807 << "\nEXTRABB = " << *BB); 2808 BB = NewBB; 2809 } 2810 2811 Builder.SetInsertPoint(BI); 2812 // Convert pointer to int before we switch. 2813 if (CompVal->getType()->isPointerTy()) { 2814 assert(TD && "Cannot switch on pointer without DataLayout"); 2815 CompVal = Builder.CreatePtrToInt(CompVal, 2816 TD->getIntPtrType(CompVal->getType()), 2817 "magicptr"); 2818 } 2819 2820 // Create the new switch instruction now. 2821 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2822 2823 // Add all of the 'cases' to the switch instruction. 2824 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2825 New->addCase(Values[i], EdgeBB); 2826 2827 // We added edges from PI to the EdgeBB. As such, if there were any 2828 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2829 // the number of edges added. 2830 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2831 isa<PHINode>(BBI); ++BBI) { 2832 PHINode *PN = cast<PHINode>(BBI); 2833 Value *InVal = PN->getIncomingValueForBlock(BB); 2834 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2835 PN->addIncoming(InVal, BB); 2836 } 2837 2838 // Erase the old branch instruction. 2839 EraseTerminatorInstAndDCECond(BI); 2840 2841 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2842 return true; 2843} 2844 2845bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2846 // If this is a trivial landing pad that just continues unwinding the caught 2847 // exception then zap the landing pad, turning its invokes into calls. 2848 BasicBlock *BB = RI->getParent(); 2849 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2850 if (RI->getValue() != LPInst) 2851 // Not a landing pad, or the resume is not unwinding the exception that 2852 // caused control to branch here. 2853 return false; 2854 2855 // Check that there are no other instructions except for debug intrinsics. 2856 BasicBlock::iterator I = LPInst, E = RI; 2857 while (++I != E) 2858 if (!isa<DbgInfoIntrinsic>(I)) 2859 return false; 2860 2861 // Turn all invokes that unwind here into calls and delete the basic block. 2862 bool InvokeRequiresTableEntry = false; 2863 bool Changed = false; 2864 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2865 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2866 2867 if (II->hasFnAttr(Attribute::UWTable)) { 2868 // Don't remove an `invoke' instruction if the ABI requires an entry into 2869 // the table. 2870 InvokeRequiresTableEntry = true; 2871 continue; 2872 } 2873 2874 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2875 2876 // Insert a call instruction before the invoke. 2877 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2878 Call->takeName(II); 2879 Call->setCallingConv(II->getCallingConv()); 2880 Call->setAttributes(II->getAttributes()); 2881 Call->setDebugLoc(II->getDebugLoc()); 2882 2883 // Anything that used the value produced by the invoke instruction now uses 2884 // the value produced by the call instruction. Note that we do this even 2885 // for void functions and calls with no uses so that the callgraph edge is 2886 // updated. 2887 II->replaceAllUsesWith(Call); 2888 BB->removePredecessor(II->getParent()); 2889 2890 // Insert a branch to the normal destination right before the invoke. 2891 BranchInst::Create(II->getNormalDest(), II); 2892 2893 // Finally, delete the invoke instruction! 2894 II->eraseFromParent(); 2895 Changed = true; 2896 } 2897 2898 if (!InvokeRequiresTableEntry) 2899 // The landingpad is now unreachable. Zap it. 2900 BB->eraseFromParent(); 2901 2902 return Changed; 2903} 2904 2905bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2906 BasicBlock *BB = RI->getParent(); 2907 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2908 2909 // Find predecessors that end with branches. 2910 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2911 SmallVector<BranchInst*, 8> CondBranchPreds; 2912 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2913 BasicBlock *P = *PI; 2914 TerminatorInst *PTI = P->getTerminator(); 2915 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2916 if (BI->isUnconditional()) 2917 UncondBranchPreds.push_back(P); 2918 else 2919 CondBranchPreds.push_back(BI); 2920 } 2921 } 2922 2923 // If we found some, do the transformation! 2924 if (!UncondBranchPreds.empty() && DupRet) { 2925 while (!UncondBranchPreds.empty()) { 2926 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2927 DEBUG(dbgs() << "FOLDING: " << *BB 2928 << "INTO UNCOND BRANCH PRED: " << *Pred); 2929 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2930 } 2931 2932 // If we eliminated all predecessors of the block, delete the block now. 2933 if (pred_begin(BB) == pred_end(BB)) 2934 // We know there are no successors, so just nuke the block. 2935 BB->eraseFromParent(); 2936 2937 return true; 2938 } 2939 2940 // Check out all of the conditional branches going to this return 2941 // instruction. If any of them just select between returns, change the 2942 // branch itself into a select/return pair. 2943 while (!CondBranchPreds.empty()) { 2944 BranchInst *BI = CondBranchPreds.pop_back_val(); 2945 2946 // Check to see if the non-BB successor is also a return block. 2947 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2948 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2949 SimplifyCondBranchToTwoReturns(BI, Builder)) 2950 return true; 2951 } 2952 return false; 2953} 2954 2955bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2956 BasicBlock *BB = UI->getParent(); 2957 2958 bool Changed = false; 2959 2960 // If there are any instructions immediately before the unreachable that can 2961 // be removed, do so. 2962 while (UI != BB->begin()) { 2963 BasicBlock::iterator BBI = UI; 2964 --BBI; 2965 // Do not delete instructions that can have side effects which might cause 2966 // the unreachable to not be reachable; specifically, calls and volatile 2967 // operations may have this effect. 2968 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2969 2970 if (BBI->mayHaveSideEffects()) { 2971 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2972 if (SI->isVolatile()) 2973 break; 2974 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2975 if (LI->isVolatile()) 2976 break; 2977 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2978 if (RMWI->isVolatile()) 2979 break; 2980 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2981 if (CXI->isVolatile()) 2982 break; 2983 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2984 !isa<LandingPadInst>(BBI)) { 2985 break; 2986 } 2987 // Note that deleting LandingPad's here is in fact okay, although it 2988 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2989 // all the predecessors of this block will be the unwind edges of Invokes, 2990 // and we can therefore guarantee this block will be erased. 2991 } 2992 2993 // Delete this instruction (any uses are guaranteed to be dead) 2994 if (!BBI->use_empty()) 2995 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2996 BBI->eraseFromParent(); 2997 Changed = true; 2998 } 2999 3000 // If the unreachable instruction is the first in the block, take a gander 3001 // at all of the predecessors of this instruction, and simplify them. 3002 if (&BB->front() != UI) return Changed; 3003 3004 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 3005 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 3006 TerminatorInst *TI = Preds[i]->getTerminator(); 3007 IRBuilder<> Builder(TI); 3008 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 3009 if (BI->isUnconditional()) { 3010 if (BI->getSuccessor(0) == BB) { 3011 new UnreachableInst(TI->getContext(), TI); 3012 TI->eraseFromParent(); 3013 Changed = true; 3014 } 3015 } else { 3016 if (BI->getSuccessor(0) == BB) { 3017 Builder.CreateBr(BI->getSuccessor(1)); 3018 EraseTerminatorInstAndDCECond(BI); 3019 } else if (BI->getSuccessor(1) == BB) { 3020 Builder.CreateBr(BI->getSuccessor(0)); 3021 EraseTerminatorInstAndDCECond(BI); 3022 Changed = true; 3023 } 3024 } 3025 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 3026 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3027 i != e; ++i) 3028 if (i.getCaseSuccessor() == BB) { 3029 BB->removePredecessor(SI->getParent()); 3030 SI->removeCase(i); 3031 --i; --e; 3032 Changed = true; 3033 } 3034 // If the default value is unreachable, figure out the most popular 3035 // destination and make it the default. 3036 if (SI->getDefaultDest() == BB) { 3037 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 3038 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3039 i != e; ++i) { 3040 std::pair<unsigned, unsigned> &entry = 3041 Popularity[i.getCaseSuccessor()]; 3042 if (entry.first == 0) { 3043 entry.first = 1; 3044 entry.second = i.getCaseIndex(); 3045 } else { 3046 entry.first++; 3047 } 3048 } 3049 3050 // Find the most popular block. 3051 unsigned MaxPop = 0; 3052 unsigned MaxIndex = 0; 3053 BasicBlock *MaxBlock = 0; 3054 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 3055 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 3056 if (I->second.first > MaxPop || 3057 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 3058 MaxPop = I->second.first; 3059 MaxIndex = I->second.second; 3060 MaxBlock = I->first; 3061 } 3062 } 3063 if (MaxBlock) { 3064 // Make this the new default, allowing us to delete any explicit 3065 // edges to it. 3066 SI->setDefaultDest(MaxBlock); 3067 Changed = true; 3068 3069 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3070 // it. 3071 if (isa<PHINode>(MaxBlock->begin())) 3072 for (unsigned i = 0; i != MaxPop-1; ++i) 3073 MaxBlock->removePredecessor(SI->getParent()); 3074 3075 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3076 i != e; ++i) 3077 if (i.getCaseSuccessor() == MaxBlock) { 3078 SI->removeCase(i); 3079 --i; --e; 3080 } 3081 } 3082 } 3083 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3084 if (II->getUnwindDest() == BB) { 3085 // Convert the invoke to a call instruction. This would be a good 3086 // place to note that the call does not throw though. 3087 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3088 II->removeFromParent(); // Take out of symbol table 3089 3090 // Insert the call now... 3091 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3092 Builder.SetInsertPoint(BI); 3093 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3094 Args, II->getName()); 3095 CI->setCallingConv(II->getCallingConv()); 3096 CI->setAttributes(II->getAttributes()); 3097 // If the invoke produced a value, the call does now instead. 3098 II->replaceAllUsesWith(CI); 3099 delete II; 3100 Changed = true; 3101 } 3102 } 3103 } 3104 3105 // If this block is now dead, remove it. 3106 if (pred_begin(BB) == pred_end(BB) && 3107 BB != &BB->getParent()->getEntryBlock()) { 3108 // We know there are no successors, so just nuke the block. 3109 BB->eraseFromParent(); 3110 return true; 3111 } 3112 3113 return Changed; 3114} 3115 3116/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3117/// integer range comparison into a sub, an icmp and a branch. 3118static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3119 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3120 3121 // Make sure all cases point to the same destination and gather the values. 3122 SmallVector<ConstantInt *, 16> Cases; 3123 SwitchInst::CaseIt I = SI->case_begin(); 3124 Cases.push_back(I.getCaseValue()); 3125 SwitchInst::CaseIt PrevI = I++; 3126 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3127 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3128 return false; 3129 Cases.push_back(I.getCaseValue()); 3130 } 3131 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3132 3133 // Sort the case values, then check if they form a range we can transform. 3134 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3135 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3136 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3137 return false; 3138 } 3139 3140 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3141 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3142 3143 Value *Sub = SI->getCondition(); 3144 if (!Offset->isNullValue()) 3145 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3146 Value *Cmp; 3147 // If NumCases overflowed, then all possible values jump to the successor. 3148 if (NumCases->isNullValue() && SI->getNumCases() != 0) 3149 Cmp = ConstantInt::getTrue(SI->getContext()); 3150 else 3151 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3152 BranchInst *NewBI = Builder.CreateCondBr( 3153 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3154 3155 // Update weight for the newly-created conditional branch. 3156 SmallVector<uint64_t, 8> Weights; 3157 bool HasWeights = HasBranchWeights(SI); 3158 if (HasWeights) { 3159 GetBranchWeights(SI, Weights); 3160 if (Weights.size() == 1 + SI->getNumCases()) { 3161 // Combine all weights for the cases to be the true weight of NewBI. 3162 // We assume that the sum of all weights for a Terminator can fit into 32 3163 // bits. 3164 uint32_t NewTrueWeight = 0; 3165 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3166 NewTrueWeight += (uint32_t)Weights[I]; 3167 NewBI->setMetadata(LLVMContext::MD_prof, 3168 MDBuilder(SI->getContext()). 3169 createBranchWeights(NewTrueWeight, 3170 (uint32_t)Weights[0])); 3171 } 3172 } 3173 3174 // Prune obsolete incoming values off the successor's PHI nodes. 3175 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3176 isa<PHINode>(BBI); ++BBI) { 3177 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3178 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3179 } 3180 SI->eraseFromParent(); 3181 3182 return true; 3183} 3184 3185/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3186/// and use it to remove dead cases. 3187static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3188 Value *Cond = SI->getCondition(); 3189 unsigned Bits = Cond->getType()->getIntegerBitWidth(); 3190 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3191 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3192 3193 // Gather dead cases. 3194 SmallVector<ConstantInt*, 8> DeadCases; 3195 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3196 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3197 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3198 DeadCases.push_back(I.getCaseValue()); 3199 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3200 << I.getCaseValue() << "' is dead.\n"); 3201 } 3202 } 3203 3204 SmallVector<uint64_t, 8> Weights; 3205 bool HasWeight = HasBranchWeights(SI); 3206 if (HasWeight) { 3207 GetBranchWeights(SI, Weights); 3208 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3209 } 3210 3211 // Remove dead cases from the switch. 3212 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3213 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3214 assert(Case != SI->case_default() && 3215 "Case was not found. Probably mistake in DeadCases forming."); 3216 if (HasWeight) { 3217 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3218 Weights.pop_back(); 3219 } 3220 3221 // Prune unused values from PHI nodes. 3222 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3223 SI->removeCase(Case); 3224 } 3225 if (HasWeight) { 3226 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3227 SI->setMetadata(LLVMContext::MD_prof, 3228 MDBuilder(SI->getParent()->getContext()). 3229 createBranchWeights(MDWeights)); 3230 } 3231 3232 return !DeadCases.empty(); 3233} 3234 3235/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3236/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3237/// by an unconditional branch), look at the phi node for BB in the successor 3238/// block and see if the incoming value is equal to CaseValue. If so, return 3239/// the phi node, and set PhiIndex to BB's index in the phi node. 3240static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3241 BasicBlock *BB, 3242 int *PhiIndex) { 3243 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3244 return NULL; // BB must be empty to be a candidate for simplification. 3245 if (!BB->getSinglePredecessor()) 3246 return NULL; // BB must be dominated by the switch. 3247 3248 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3249 if (!Branch || !Branch->isUnconditional()) 3250 return NULL; // Terminator must be unconditional branch. 3251 3252 BasicBlock *Succ = Branch->getSuccessor(0); 3253 3254 BasicBlock::iterator I = Succ->begin(); 3255 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3256 int Idx = PHI->getBasicBlockIndex(BB); 3257 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3258 3259 Value *InValue = PHI->getIncomingValue(Idx); 3260 if (InValue != CaseValue) continue; 3261 3262 *PhiIndex = Idx; 3263 return PHI; 3264 } 3265 3266 return NULL; 3267} 3268 3269/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3270/// instruction to a phi node dominated by the switch, if that would mean that 3271/// some of the destination blocks of the switch can be folded away. 3272/// Returns true if a change is made. 3273static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3274 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3275 ForwardingNodesMap ForwardingNodes; 3276 3277 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3278 ConstantInt *CaseValue = I.getCaseValue(); 3279 BasicBlock *CaseDest = I.getCaseSuccessor(); 3280 3281 int PhiIndex; 3282 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3283 &PhiIndex); 3284 if (!PHI) continue; 3285 3286 ForwardingNodes[PHI].push_back(PhiIndex); 3287 } 3288 3289 bool Changed = false; 3290 3291 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3292 E = ForwardingNodes.end(); I != E; ++I) { 3293 PHINode *Phi = I->first; 3294 SmallVectorImpl<int> &Indexes = I->second; 3295 3296 if (Indexes.size() < 2) continue; 3297 3298 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3299 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3300 Changed = true; 3301 } 3302 3303 return Changed; 3304} 3305 3306/// ValidLookupTableConstant - Return true if the backend will be able to handle 3307/// initializing an array of constants like C. 3308static bool ValidLookupTableConstant(Constant *C) { 3309 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3310 return CE->isGEPWithNoNotionalOverIndexing(); 3311 3312 return isa<ConstantFP>(C) || 3313 isa<ConstantInt>(C) || 3314 isa<ConstantPointerNull>(C) || 3315 isa<GlobalValue>(C) || 3316 isa<UndefValue>(C); 3317} 3318 3319/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3320/// its constant value in ConstantPool, returning 0 if it's not there. 3321static Constant *LookupConstant(Value *V, 3322 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3323 if (Constant *C = dyn_cast<Constant>(V)) 3324 return C; 3325 return ConstantPool.lookup(V); 3326} 3327 3328/// ConstantFold - Try to fold instruction I into a constant. This works for 3329/// simple instructions such as binary operations where both operands are 3330/// constant or can be replaced by constants from the ConstantPool. Returns the 3331/// resulting constant on success, 0 otherwise. 3332static Constant * 3333ConstantFold(Instruction *I, 3334 const SmallDenseMap<Value *, Constant *> &ConstantPool, 3335 const DataLayout *DL) { 3336 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3337 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3338 if (!A) 3339 return 0; 3340 if (A->isAllOnesValue()) 3341 return LookupConstant(Select->getTrueValue(), ConstantPool); 3342 if (A->isNullValue()) 3343 return LookupConstant(Select->getFalseValue(), ConstantPool); 3344 return 0; 3345 } 3346 3347 SmallVector<Constant *, 4> COps; 3348 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) { 3349 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool)) 3350 COps.push_back(A); 3351 else 3352 return 0; 3353 } 3354 3355 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) 3356 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0], 3357 COps[1], DL); 3358 3359 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL); 3360} 3361 3362/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3363/// at the common destination basic block, *CommonDest, for one of the case 3364/// destionations CaseDest corresponding to value CaseVal (0 for the default 3365/// case), of a switch instruction SI. 3366static bool 3367GetCaseResults(SwitchInst *SI, 3368 ConstantInt *CaseVal, 3369 BasicBlock *CaseDest, 3370 BasicBlock **CommonDest, 3371 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res, 3372 const DataLayout *DL) { 3373 // The block from which we enter the common destination. 3374 BasicBlock *Pred = SI->getParent(); 3375 3376 // If CaseDest is empty except for some side-effect free instructions through 3377 // which we can constant-propagate the CaseVal, continue to its successor. 3378 SmallDenseMap<Value*, Constant*> ConstantPool; 3379 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3380 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3381 ++I) { 3382 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3383 // If the terminator is a simple branch, continue to the next block. 3384 if (T->getNumSuccessors() != 1) 3385 return false; 3386 Pred = CaseDest; 3387 CaseDest = T->getSuccessor(0); 3388 } else if (isa<DbgInfoIntrinsic>(I)) { 3389 // Skip debug intrinsic. 3390 continue; 3391 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) { 3392 // Instruction is side-effect free and constant. 3393 ConstantPool.insert(std::make_pair(I, C)); 3394 } else { 3395 break; 3396 } 3397 } 3398 3399 // If we did not have a CommonDest before, use the current one. 3400 if (!*CommonDest) 3401 *CommonDest = CaseDest; 3402 // If the destination isn't the common one, abort. 3403 if (CaseDest != *CommonDest) 3404 return false; 3405 3406 // Get the values for this case from phi nodes in the destination block. 3407 BasicBlock::iterator I = (*CommonDest)->begin(); 3408 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3409 int Idx = PHI->getBasicBlockIndex(Pred); 3410 if (Idx == -1) 3411 continue; 3412 3413 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3414 ConstantPool); 3415 if (!ConstVal) 3416 return false; 3417 3418 // Note: If the constant comes from constant-propagating the case value 3419 // through the CaseDest basic block, it will be safe to remove the 3420 // instructions in that block. They cannot be used (except in the phi nodes 3421 // we visit) outside CaseDest, because that block does not dominate its 3422 // successor. If it did, we would not be in this phi node. 3423 3424 // Be conservative about which kinds of constants we support. 3425 if (!ValidLookupTableConstant(ConstVal)) 3426 return false; 3427 3428 Res.push_back(std::make_pair(PHI, ConstVal)); 3429 } 3430 3431 return true; 3432} 3433 3434namespace { 3435 /// SwitchLookupTable - This class represents a lookup table that can be used 3436 /// to replace a switch. 3437 class SwitchLookupTable { 3438 public: 3439 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3440 /// with the contents of Values, using DefaultValue to fill any holes in the 3441 /// table. 3442 SwitchLookupTable(Module &M, 3443 uint64_t TableSize, 3444 ConstantInt *Offset, 3445 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, 3446 Constant *DefaultValue, 3447 const DataLayout *TD); 3448 3449 /// BuildLookup - Build instructions with Builder to retrieve the value at 3450 /// the position given by Index in the lookup table. 3451 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3452 3453 /// WouldFitInRegister - Return true if a table with TableSize elements of 3454 /// type ElementType would fit in a target-legal register. 3455 static bool WouldFitInRegister(const DataLayout *TD, 3456 uint64_t TableSize, 3457 const Type *ElementType); 3458 3459 private: 3460 // Depending on the contents of the table, it can be represented in 3461 // different ways. 3462 enum { 3463 // For tables where each element contains the same value, we just have to 3464 // store that single value and return it for each lookup. 3465 SingleValueKind, 3466 3467 // For small tables with integer elements, we can pack them into a bitmap 3468 // that fits into a target-legal register. Values are retrieved by 3469 // shift and mask operations. 3470 BitMapKind, 3471 3472 // The table is stored as an array of values. Values are retrieved by load 3473 // instructions from the table. 3474 ArrayKind 3475 } Kind; 3476 3477 // For SingleValueKind, this is the single value. 3478 Constant *SingleValue; 3479 3480 // For BitMapKind, this is the bitmap. 3481 ConstantInt *BitMap; 3482 IntegerType *BitMapElementTy; 3483 3484 // For ArrayKind, this is the array. 3485 GlobalVariable *Array; 3486 }; 3487} 3488 3489SwitchLookupTable::SwitchLookupTable(Module &M, 3490 uint64_t TableSize, 3491 ConstantInt *Offset, 3492 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, 3493 Constant *DefaultValue, 3494 const DataLayout *TD) 3495 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3496 assert(Values.size() && "Can't build lookup table without values!"); 3497 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3498 3499 // If all values in the table are equal, this is that value. 3500 SingleValue = Values.begin()->second; 3501 3502 // Build up the table contents. 3503 SmallVector<Constant*, 64> TableContents(TableSize); 3504 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3505 ConstantInt *CaseVal = Values[I].first; 3506 Constant *CaseRes = Values[I].second; 3507 assert(CaseRes->getType() == DefaultValue->getType()); 3508 3509 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3510 .getLimitedValue(); 3511 TableContents[Idx] = CaseRes; 3512 3513 if (CaseRes != SingleValue) 3514 SingleValue = 0; 3515 } 3516 3517 // Fill in any holes in the table with the default result. 3518 if (Values.size() < TableSize) { 3519 for (uint64_t I = 0; I < TableSize; ++I) { 3520 if (!TableContents[I]) 3521 TableContents[I] = DefaultValue; 3522 } 3523 3524 if (DefaultValue != SingleValue) 3525 SingleValue = 0; 3526 } 3527 3528 // If each element in the table contains the same value, we only need to store 3529 // that single value. 3530 if (SingleValue) { 3531 Kind = SingleValueKind; 3532 return; 3533 } 3534 3535 // If the type is integer and the table fits in a register, build a bitmap. 3536 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3537 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3538 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3539 for (uint64_t I = TableSize; I > 0; --I) { 3540 TableInt <<= IT->getBitWidth(); 3541 // Insert values into the bitmap. Undef values are set to zero. 3542 if (!isa<UndefValue>(TableContents[I - 1])) { 3543 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3544 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3545 } 3546 } 3547 BitMap = ConstantInt::get(M.getContext(), TableInt); 3548 BitMapElementTy = IT; 3549 Kind = BitMapKind; 3550 ++NumBitMaps; 3551 return; 3552 } 3553 3554 // Store the table in an array. 3555 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3556 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3557 3558 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3559 GlobalVariable::PrivateLinkage, 3560 Initializer, 3561 "switch.table"); 3562 Array->setUnnamedAddr(true); 3563 Kind = ArrayKind; 3564} 3565 3566Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3567 switch (Kind) { 3568 case SingleValueKind: 3569 return SingleValue; 3570 case BitMapKind: { 3571 // Type of the bitmap (e.g. i59). 3572 IntegerType *MapTy = BitMap->getType(); 3573 3574 // Cast Index to the same type as the bitmap. 3575 // Note: The Index is <= the number of elements in the table, so 3576 // truncating it to the width of the bitmask is safe. 3577 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3578 3579 // Multiply the shift amount by the element width. 3580 ShiftAmt = Builder.CreateMul(ShiftAmt, 3581 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3582 "switch.shiftamt"); 3583 3584 // Shift down. 3585 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3586 "switch.downshift"); 3587 // Mask off. 3588 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3589 "switch.masked"); 3590 } 3591 case ArrayKind: { 3592 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3593 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3594 "switch.gep"); 3595 return Builder.CreateLoad(GEP, "switch.load"); 3596 } 3597 } 3598 llvm_unreachable("Unknown lookup table kind!"); 3599} 3600 3601bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3602 uint64_t TableSize, 3603 const Type *ElementType) { 3604 if (!TD) 3605 return false; 3606 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3607 if (!IT) 3608 return false; 3609 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3610 // are <= 15, we could try to narrow the type. 3611 3612 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3613 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3614 return false; 3615 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3616} 3617 3618/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3619/// for this switch, based on the number of cases, size of the table and the 3620/// types of the results. 3621static bool ShouldBuildLookupTable(SwitchInst *SI, 3622 uint64_t TableSize, 3623 const TargetTransformInfo &TTI, 3624 const DataLayout *TD, 3625 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3626 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3627 return false; // TableSize overflowed, or mul below might overflow. 3628 3629 bool AllTablesFitInRegister = true; 3630 bool HasIllegalType = false; 3631 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3632 E = ResultTypes.end(); I != E; ++I) { 3633 Type *Ty = I->second; 3634 3635 // Saturate this flag to true. 3636 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3637 3638 // Saturate this flag to false. 3639 AllTablesFitInRegister = AllTablesFitInRegister && 3640 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3641 3642 // If both flags saturate, we're done. NOTE: This *only* works with 3643 // saturating flags, and all flags have to saturate first due to the 3644 // non-deterministic behavior of iterating over a dense map. 3645 if (HasIllegalType && !AllTablesFitInRegister) 3646 break; 3647 } 3648 3649 // If each table would fit in a register, we should build it anyway. 3650 if (AllTablesFitInRegister) 3651 return true; 3652 3653 // Don't build a table that doesn't fit in-register if it has illegal types. 3654 if (HasIllegalType) 3655 return false; 3656 3657 // The table density should be at least 40%. This is the same criterion as for 3658 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3659 // FIXME: Find the best cut-off. 3660 return SI->getNumCases() * 10 >= TableSize * 4; 3661} 3662 3663/// SwitchToLookupTable - If the switch is only used to initialize one or more 3664/// phi nodes in a common successor block with different constant values, 3665/// replace the switch with lookup tables. 3666static bool SwitchToLookupTable(SwitchInst *SI, 3667 IRBuilder<> &Builder, 3668 const TargetTransformInfo &TTI, 3669 const DataLayout* TD) { 3670 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3671 3672 // Only build lookup table when we have a target that supports it. 3673 if (!TTI.shouldBuildLookupTables()) 3674 return false; 3675 3676 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3677 // split off a dense part and build a lookup table for that. 3678 3679 // FIXME: This creates arrays of GEPs to constant strings, which means each 3680 // GEP needs a runtime relocation in PIC code. We should just build one big 3681 // string and lookup indices into that. 3682 3683 // Ignore the switch if the number of cases is too small. 3684 // This is similar to the check when building jump tables in 3685 // SelectionDAGBuilder::handleJTSwitchCase. 3686 // FIXME: Determine the best cut-off. 3687 if (SI->getNumCases() < 4) 3688 return false; 3689 3690 // Figure out the corresponding result for each case value and phi node in the 3691 // common destination, as well as the the min and max case values. 3692 assert(SI->case_begin() != SI->case_end()); 3693 SwitchInst::CaseIt CI = SI->case_begin(); 3694 ConstantInt *MinCaseVal = CI.getCaseValue(); 3695 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3696 3697 BasicBlock *CommonDest = 0; 3698 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3699 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3700 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3701 SmallDenseMap<PHINode*, Type*> ResultTypes; 3702 SmallVector<PHINode*, 4> PHIs; 3703 3704 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3705 ConstantInt *CaseVal = CI.getCaseValue(); 3706 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3707 MinCaseVal = CaseVal; 3708 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3709 MaxCaseVal = CaseVal; 3710 3711 // Resulting value at phi nodes for this case value. 3712 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3713 ResultsTy Results; 3714 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3715 Results, TD)) 3716 return false; 3717 3718 // Append the result from this case to the list for each phi. 3719 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3720 if (!ResultLists.count(I->first)) 3721 PHIs.push_back(I->first); 3722 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3723 } 3724 } 3725 3726 // Get the resulting values for the default case. 3727 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3728 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3729 DefaultResultsList, TD)) 3730 return false; 3731 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3732 PHINode *PHI = DefaultResultsList[I].first; 3733 Constant *Result = DefaultResultsList[I].second; 3734 DefaultResults[PHI] = Result; 3735 ResultTypes[PHI] = Result->getType(); 3736 } 3737 3738 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3739 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3740 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3741 return false; 3742 3743 // Create the BB that does the lookups. 3744 Module &Mod = *CommonDest->getParent()->getParent(); 3745 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3746 "switch.lookup", 3747 CommonDest->getParent(), 3748 CommonDest); 3749 3750 // Compute the table index value. 3751 Builder.SetInsertPoint(SI); 3752 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3753 "switch.tableidx"); 3754 3755 // Compute the maximum table size representable by the integer type we are 3756 // switching upon. 3757 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits(); 3758 uint64_t MaxTableSize = CaseSize > 63? UINT64_MAX : 1ULL << CaseSize; 3759 assert(MaxTableSize >= TableSize && 3760 "It is impossible for a switch to have more entries than the max " 3761 "representable value of its input integer type's size."); 3762 3763 // If we have a fully covered lookup table, unconditionally branch to the 3764 // lookup table BB. Otherwise, check if the condition value is within the case 3765 // range. If it is so, branch to the new BB. Otherwise branch to SI's default 3766 // destination. 3767 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize; 3768 if (GeneratingCoveredLookupTable) { 3769 Builder.CreateBr(LookupBB); 3770 SI->getDefaultDest()->removePredecessor(SI->getParent()); 3771 } else { 3772 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3773 MinCaseVal->getType(), TableSize)); 3774 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3775 } 3776 3777 // Populate the BB that does the lookups. 3778 Builder.SetInsertPoint(LookupBB); 3779 bool ReturnedEarly = false; 3780 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3781 PHINode *PHI = PHIs[I]; 3782 3783 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3784 DefaultResults[PHI], TD); 3785 3786 Value *Result = Table.BuildLookup(TableIndex, Builder); 3787 3788 // If the result is used to return immediately from the function, we want to 3789 // do that right here. 3790 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3791 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3792 Builder.CreateRet(Result); 3793 ReturnedEarly = true; 3794 break; 3795 } 3796 3797 PHI->addIncoming(Result, LookupBB); 3798 } 3799 3800 if (!ReturnedEarly) 3801 Builder.CreateBr(CommonDest); 3802 3803 // Remove the switch. 3804 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) { 3805 BasicBlock *Succ = SI->getSuccessor(i); 3806 3807 if (Succ == SI->getDefaultDest()) 3808 continue; 3809 Succ->removePredecessor(SI->getParent()); 3810 } 3811 SI->eraseFromParent(); 3812 3813 ++NumLookupTables; 3814 return true; 3815} 3816 3817bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3818 BasicBlock *BB = SI->getParent(); 3819 3820 if (isValueEqualityComparison(SI)) { 3821 // If we only have one predecessor, and if it is a branch on this value, 3822 // see if that predecessor totally determines the outcome of this switch. 3823 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3824 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3825 return SimplifyCFG(BB, TTI, TD) | true; 3826 3827 Value *Cond = SI->getCondition(); 3828 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3829 if (SimplifySwitchOnSelect(SI, Select)) 3830 return SimplifyCFG(BB, TTI, TD) | true; 3831 3832 // If the block only contains the switch, see if we can fold the block 3833 // away into any preds. 3834 BasicBlock::iterator BBI = BB->begin(); 3835 // Ignore dbg intrinsics. 3836 while (isa<DbgInfoIntrinsic>(BBI)) 3837 ++BBI; 3838 if (SI == &*BBI) 3839 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3840 return SimplifyCFG(BB, TTI, TD) | true; 3841 } 3842 3843 // Try to transform the switch into an icmp and a branch. 3844 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3845 return SimplifyCFG(BB, TTI, TD) | true; 3846 3847 // Remove unreachable cases. 3848 if (EliminateDeadSwitchCases(SI)) 3849 return SimplifyCFG(BB, TTI, TD) | true; 3850 3851 if (ForwardSwitchConditionToPHI(SI)) 3852 return SimplifyCFG(BB, TTI, TD) | true; 3853 3854 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3855 return SimplifyCFG(BB, TTI, TD) | true; 3856 3857 return false; 3858} 3859 3860bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3861 BasicBlock *BB = IBI->getParent(); 3862 bool Changed = false; 3863 3864 // Eliminate redundant destinations. 3865 SmallPtrSet<Value *, 8> Succs; 3866 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3867 BasicBlock *Dest = IBI->getDestination(i); 3868 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3869 Dest->removePredecessor(BB); 3870 IBI->removeDestination(i); 3871 --i; --e; 3872 Changed = true; 3873 } 3874 } 3875 3876 if (IBI->getNumDestinations() == 0) { 3877 // If the indirectbr has no successors, change it to unreachable. 3878 new UnreachableInst(IBI->getContext(), IBI); 3879 EraseTerminatorInstAndDCECond(IBI); 3880 return true; 3881 } 3882 3883 if (IBI->getNumDestinations() == 1) { 3884 // If the indirectbr has one successor, change it to a direct branch. 3885 BranchInst::Create(IBI->getDestination(0), IBI); 3886 EraseTerminatorInstAndDCECond(IBI); 3887 return true; 3888 } 3889 3890 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3891 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3892 return SimplifyCFG(BB, TTI, TD) | true; 3893 } 3894 return Changed; 3895} 3896 3897bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3898 BasicBlock *BB = BI->getParent(); 3899 3900 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3901 return true; 3902 3903 // If the Terminator is the only non-phi instruction, simplify the block. 3904 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3905 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3906 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3907 return true; 3908 3909 // If the only instruction in the block is a seteq/setne comparison 3910 // against a constant, try to simplify the block. 3911 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3912 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3913 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3914 ; 3915 if (I->isTerminator() && 3916 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3917 return true; 3918 } 3919 3920 // If this basic block is ONLY a compare and a branch, and if a predecessor 3921 // branches to us and our successor, fold the comparison into the 3922 // predecessor and use logical operations to update the incoming value 3923 // for PHI nodes in common successor. 3924 if (FoldBranchToCommonDest(BI)) 3925 return SimplifyCFG(BB, TTI, TD) | true; 3926 return false; 3927} 3928 3929 3930bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3931 BasicBlock *BB = BI->getParent(); 3932 3933 // Conditional branch 3934 if (isValueEqualityComparison(BI)) { 3935 // If we only have one predecessor, and if it is a branch on this value, 3936 // see if that predecessor totally determines the outcome of this 3937 // switch. 3938 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3939 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3940 return SimplifyCFG(BB, TTI, TD) | true; 3941 3942 // This block must be empty, except for the setcond inst, if it exists. 3943 // Ignore dbg intrinsics. 3944 BasicBlock::iterator I = BB->begin(); 3945 // Ignore dbg intrinsics. 3946 while (isa<DbgInfoIntrinsic>(I)) 3947 ++I; 3948 if (&*I == BI) { 3949 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3950 return SimplifyCFG(BB, TTI, TD) | true; 3951 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3952 ++I; 3953 // Ignore dbg intrinsics. 3954 while (isa<DbgInfoIntrinsic>(I)) 3955 ++I; 3956 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3957 return SimplifyCFG(BB, TTI, TD) | true; 3958 } 3959 } 3960 3961 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3962 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3963 return true; 3964 3965 // If this basic block is ONLY a compare and a branch, and if a predecessor 3966 // branches to us and one of our successors, fold the comparison into the 3967 // predecessor and use logical operations to pick the right destination. 3968 if (FoldBranchToCommonDest(BI)) 3969 return SimplifyCFG(BB, TTI, TD) | true; 3970 3971 // We have a conditional branch to two blocks that are only reachable 3972 // from BI. We know that the condbr dominates the two blocks, so see if 3973 // there is any identical code in the "then" and "else" blocks. If so, we 3974 // can hoist it up to the branching block. 3975 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3976 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3977 if (HoistThenElseCodeToIf(BI)) 3978 return SimplifyCFG(BB, TTI, TD) | true; 3979 } else { 3980 // If Successor #1 has multiple preds, we may be able to conditionally 3981 // execute Successor #0 if it branches to successor #1. 3982 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3983 if (Succ0TI->getNumSuccessors() == 1 && 3984 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3985 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3986 return SimplifyCFG(BB, TTI, TD) | true; 3987 } 3988 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3989 // If Successor #0 has multiple preds, we may be able to conditionally 3990 // execute Successor #1 if it branches to successor #0. 3991 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3992 if (Succ1TI->getNumSuccessors() == 1 && 3993 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3994 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3995 return SimplifyCFG(BB, TTI, TD) | true; 3996 } 3997 3998 // If this is a branch on a phi node in the current block, thread control 3999 // through this block if any PHI node entries are constants. 4000 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 4001 if (PN->getParent() == BI->getParent()) 4002 if (FoldCondBranchOnPHI(BI, TD)) 4003 return SimplifyCFG(BB, TTI, TD) | true; 4004 4005 // Scan predecessor blocks for conditional branches. 4006 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 4007 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 4008 if (PBI != BI && PBI->isConditional()) 4009 if (SimplifyCondBranchToCondBranch(PBI, BI)) 4010 return SimplifyCFG(BB, TTI, TD) | true; 4011 4012 return false; 4013} 4014 4015/// Check if passing a value to an instruction will cause undefined behavior. 4016static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 4017 Constant *C = dyn_cast<Constant>(V); 4018 if (!C) 4019 return false; 4020 4021 if (I->use_empty()) 4022 return false; 4023 4024 if (C->isNullValue()) { 4025 // Only look at the first use, avoid hurting compile time with long uselists 4026 User *Use = *I->use_begin(); 4027 4028 // Now make sure that there are no instructions in between that can alter 4029 // control flow (eg. calls) 4030 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 4031 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 4032 return false; 4033 4034 // Look through GEPs. A load from a GEP derived from NULL is still undefined 4035 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 4036 if (GEP->getPointerOperand() == I) 4037 return passingValueIsAlwaysUndefined(V, GEP); 4038 4039 // Look through bitcasts. 4040 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 4041 return passingValueIsAlwaysUndefined(V, BC); 4042 4043 // Load from null is undefined. 4044 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 4045 if (!LI->isVolatile()) 4046 return LI->getPointerAddressSpace() == 0; 4047 4048 // Store to null is undefined. 4049 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 4050 if (!SI->isVolatile()) 4051 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 4052 } 4053 return false; 4054} 4055 4056/// If BB has an incoming value that will always trigger undefined behavior 4057/// (eg. null pointer dereference), remove the branch leading here. 4058static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 4059 for (BasicBlock::iterator i = BB->begin(); 4060 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 4061 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 4062 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 4063 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 4064 IRBuilder<> Builder(T); 4065 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 4066 BB->removePredecessor(PHI->getIncomingBlock(i)); 4067 // Turn uncoditional branches into unreachables and remove the dead 4068 // destination from conditional branches. 4069 if (BI->isUnconditional()) 4070 Builder.CreateUnreachable(); 4071 else 4072 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 4073 BI->getSuccessor(0)); 4074 BI->eraseFromParent(); 4075 return true; 4076 } 4077 // TODO: SwitchInst. 4078 } 4079 4080 return false; 4081} 4082 4083bool SimplifyCFGOpt::run(BasicBlock *BB) { 4084 bool Changed = false; 4085 4086 assert(BB && BB->getParent() && "Block not embedded in function!"); 4087 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4088 4089 // Remove basic blocks that have no predecessors (except the entry block)... 4090 // or that just have themself as a predecessor. These are unreachable. 4091 if ((pred_begin(BB) == pred_end(BB) && 4092 BB != &BB->getParent()->getEntryBlock()) || 4093 BB->getSinglePredecessor() == BB) { 4094 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4095 DeleteDeadBlock(BB); 4096 return true; 4097 } 4098 4099 // Check to see if we can constant propagate this terminator instruction 4100 // away... 4101 Changed |= ConstantFoldTerminator(BB, true); 4102 4103 // Check for and eliminate duplicate PHI nodes in this block. 4104 Changed |= EliminateDuplicatePHINodes(BB); 4105 4106 // Check for and remove branches that will always cause undefined behavior. 4107 Changed |= removeUndefIntroducingPredecessor(BB); 4108 4109 // Merge basic blocks into their predecessor if there is only one distinct 4110 // pred, and if there is only one distinct successor of the predecessor, and 4111 // if there are no PHI nodes. 4112 // 4113 if (MergeBlockIntoPredecessor(BB)) 4114 return true; 4115 4116 IRBuilder<> Builder(BB); 4117 4118 // If there is a trivial two-entry PHI node in this basic block, and we can 4119 // eliminate it, do so now. 4120 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4121 if (PN->getNumIncomingValues() == 2) 4122 Changed |= FoldTwoEntryPHINode(PN, TD); 4123 4124 Builder.SetInsertPoint(BB->getTerminator()); 4125 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4126 if (BI->isUnconditional()) { 4127 if (SimplifyUncondBranch(BI, Builder)) return true; 4128 } else { 4129 if (SimplifyCondBranch(BI, Builder)) return true; 4130 } 4131 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4132 if (SimplifyReturn(RI, Builder)) return true; 4133 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4134 if (SimplifyResume(RI, Builder)) return true; 4135 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4136 if (SimplifySwitch(SI, Builder)) return true; 4137 } else if (UnreachableInst *UI = 4138 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4139 if (SimplifyUnreachable(UI)) return true; 4140 } else if (IndirectBrInst *IBI = 4141 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4142 if (SimplifyIndirectBr(IBI)) return true; 4143 } 4144 4145 return Changed; 4146} 4147 4148/// SimplifyCFG - This function is used to do simplification of a CFG. For 4149/// example, it adjusts branches to branches to eliminate the extra hop, it 4150/// eliminates unreachable basic blocks, and does other "peephole" optimization 4151/// of the CFG. It returns true if a modification was made. 4152/// 4153bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4154 const DataLayout *TD) { 4155 return SimplifyCFGOpt(TTI, TD).run(BB); 4156} 4157