1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Peephole optimize the CFG. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "simplifycfg" 15#include "llvm/Transforms/Utils/Local.h" 16#include "llvm/ADT/DenseMap.h" 17#include "llvm/ADT/STLExtras.h" 18#include "llvm/ADT/SetVector.h" 19#include "llvm/ADT/SmallPtrSet.h" 20#include "llvm/ADT/SmallVector.h" 21#include "llvm/ADT/Statistic.h" 22#include "llvm/Analysis/InstructionSimplify.h" 23#include "llvm/Analysis/TargetTransformInfo.h" 24#include "llvm/Analysis/ValueTracking.h" 25#include "llvm/IR/Constants.h" 26#include "llvm/IR/DataLayout.h" 27#include "llvm/IR/DerivedTypes.h" 28#include "llvm/IR/GlobalVariable.h" 29#include "llvm/IR/IRBuilder.h" 30#include "llvm/IR/Instructions.h" 31#include "llvm/IR/IntrinsicInst.h" 32#include "llvm/IR/LLVMContext.h" 33#include "llvm/IR/MDBuilder.h" 34#include "llvm/IR/Metadata.h" 35#include "llvm/IR/Module.h" 36#include "llvm/IR/Operator.h" 37#include "llvm/IR/Type.h" 38#include "llvm/Support/CFG.h" 39#include "llvm/Support/CommandLine.h" 40#include "llvm/Support/ConstantRange.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/NoFolder.h" 43#include "llvm/Support/PatternMatch.h" 44#include "llvm/Support/raw_ostream.h" 45#include "llvm/Transforms/Utils/BasicBlockUtils.h" 46#include <algorithm> 47#include <map> 48#include <set> 49using namespace llvm; 50using namespace PatternMatch; 51 52static cl::opt<unsigned> 53PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 54 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 55 56static cl::opt<bool> 57DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 58 cl::desc("Duplicate return instructions into unconditional branches")); 59 60static cl::opt<bool> 61SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 62 cl::desc("Sink common instructions down to the end block")); 63 64static cl::opt<bool> 65HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true), 66 cl::desc("Hoist conditional stores if an unconditional store preceeds")); 67 68STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 69STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 70STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 71STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 72 73namespace { 74 /// ValueEqualityComparisonCase - Represents a case of a switch. 75 struct ValueEqualityComparisonCase { 76 ConstantInt *Value; 77 BasicBlock *Dest; 78 79 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 80 : Value(Value), Dest(Dest) {} 81 82 bool operator<(ValueEqualityComparisonCase RHS) const { 83 // Comparing pointers is ok as we only rely on the order for uniquing. 84 return Value < RHS.Value; 85 } 86 87 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 88 }; 89 90class SimplifyCFGOpt { 91 const TargetTransformInfo &TTI; 92 const DataLayout *const TD; 93 Value *isValueEqualityComparison(TerminatorInst *TI); 94 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 95 std::vector<ValueEqualityComparisonCase> &Cases); 96 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 97 BasicBlock *Pred, 98 IRBuilder<> &Builder); 99 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 100 IRBuilder<> &Builder); 101 102 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 103 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 104 bool SimplifyUnreachable(UnreachableInst *UI); 105 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 106 bool SimplifyIndirectBr(IndirectBrInst *IBI); 107 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 108 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 109 110public: 111 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD) 112 : TTI(TTI), TD(TD) {} 113 bool run(BasicBlock *BB); 114}; 115} 116 117/// SafeToMergeTerminators - Return true if it is safe to merge these two 118/// terminator instructions together. 119/// 120static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 121 if (SI1 == SI2) return false; // Can't merge with self! 122 123 // It is not safe to merge these two switch instructions if they have a common 124 // successor, and if that successor has a PHI node, and if *that* PHI node has 125 // conflicting incoming values from the two switch blocks. 126 BasicBlock *SI1BB = SI1->getParent(); 127 BasicBlock *SI2BB = SI2->getParent(); 128 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 129 130 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 131 if (SI1Succs.count(*I)) 132 for (BasicBlock::iterator BBI = (*I)->begin(); 133 isa<PHINode>(BBI); ++BBI) { 134 PHINode *PN = cast<PHINode>(BBI); 135 if (PN->getIncomingValueForBlock(SI1BB) != 136 PN->getIncomingValueForBlock(SI2BB)) 137 return false; 138 } 139 140 return true; 141} 142 143/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 144/// to merge these two terminator instructions together, where SI1 is an 145/// unconditional branch. PhiNodes will store all PHI nodes in common 146/// successors. 147/// 148static bool isProfitableToFoldUnconditional(BranchInst *SI1, 149 BranchInst *SI2, 150 Instruction *Cond, 151 SmallVectorImpl<PHINode*> &PhiNodes) { 152 if (SI1 == SI2) return false; // Can't merge with self! 153 assert(SI1->isUnconditional() && SI2->isConditional()); 154 155 // We fold the unconditional branch if we can easily update all PHI nodes in 156 // common successors: 157 // 1> We have a constant incoming value for the conditional branch; 158 // 2> We have "Cond" as the incoming value for the unconditional branch; 159 // 3> SI2->getCondition() and Cond have same operands. 160 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 161 if (!Ci2) return false; 162 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 163 Cond->getOperand(1) == Ci2->getOperand(1)) && 164 !(Cond->getOperand(0) == Ci2->getOperand(1) && 165 Cond->getOperand(1) == Ci2->getOperand(0))) 166 return false; 167 168 BasicBlock *SI1BB = SI1->getParent(); 169 BasicBlock *SI2BB = SI2->getParent(); 170 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 171 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 172 if (SI1Succs.count(*I)) 173 for (BasicBlock::iterator BBI = (*I)->begin(); 174 isa<PHINode>(BBI); ++BBI) { 175 PHINode *PN = cast<PHINode>(BBI); 176 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 177 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 178 return false; 179 PhiNodes.push_back(PN); 180 } 181 return true; 182} 183 184/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 185/// now be entries in it from the 'NewPred' block. The values that will be 186/// flowing into the PHI nodes will be the same as those coming in from 187/// ExistPred, an existing predecessor of Succ. 188static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 189 BasicBlock *ExistPred) { 190 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 191 192 PHINode *PN; 193 for (BasicBlock::iterator I = Succ->begin(); 194 (PN = dyn_cast<PHINode>(I)); ++I) 195 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 196} 197 198/// ComputeSpeculationCost - Compute an abstract "cost" of speculating the 199/// given instruction, which is assumed to be safe to speculate. 1 means 200/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 201static unsigned ComputeSpeculationCost(const User *I) { 202 assert(isSafeToSpeculativelyExecute(I) && 203 "Instruction is not safe to speculatively execute!"); 204 switch (Operator::getOpcode(I)) { 205 default: 206 // In doubt, be conservative. 207 return UINT_MAX; 208 case Instruction::GetElementPtr: 209 // GEPs are cheap if all indices are constant. 210 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 211 return UINT_MAX; 212 return 1; 213 case Instruction::Load: 214 case Instruction::Add: 215 case Instruction::Sub: 216 case Instruction::And: 217 case Instruction::Or: 218 case Instruction::Xor: 219 case Instruction::Shl: 220 case Instruction::LShr: 221 case Instruction::AShr: 222 case Instruction::ICmp: 223 case Instruction::Trunc: 224 case Instruction::ZExt: 225 case Instruction::SExt: 226 return 1; // These are all cheap. 227 228 case Instruction::Call: 229 case Instruction::Select: 230 return 2; 231 } 232} 233 234/// DominatesMergePoint - If we have a merge point of an "if condition" as 235/// accepted above, return true if the specified value dominates the block. We 236/// don't handle the true generality of domination here, just a special case 237/// which works well enough for us. 238/// 239/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 240/// see if V (which must be an instruction) and its recursive operands 241/// that do not dominate BB have a combined cost lower than CostRemaining and 242/// are non-trapping. If both are true, the instruction is inserted into the 243/// set and true is returned. 244/// 245/// The cost for most non-trapping instructions is defined as 1 except for 246/// Select whose cost is 2. 247/// 248/// After this function returns, CostRemaining is decreased by the cost of 249/// V plus its non-dominating operands. If that cost is greater than 250/// CostRemaining, false is returned and CostRemaining is undefined. 251static bool DominatesMergePoint(Value *V, BasicBlock *BB, 252 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 253 unsigned &CostRemaining) { 254 Instruction *I = dyn_cast<Instruction>(V); 255 if (!I) { 256 // Non-instructions all dominate instructions, but not all constantexprs 257 // can be executed unconditionally. 258 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 259 if (C->canTrap()) 260 return false; 261 return true; 262 } 263 BasicBlock *PBB = I->getParent(); 264 265 // We don't want to allow weird loops that might have the "if condition" in 266 // the bottom of this block. 267 if (PBB == BB) return false; 268 269 // If this instruction is defined in a block that contains an unconditional 270 // branch to BB, then it must be in the 'conditional' part of the "if 271 // statement". If not, it definitely dominates the region. 272 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 273 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 274 return true; 275 276 // If we aren't allowing aggressive promotion anymore, then don't consider 277 // instructions in the 'if region'. 278 if (AggressiveInsts == 0) return false; 279 280 // If we have seen this instruction before, don't count it again. 281 if (AggressiveInsts->count(I)) return true; 282 283 // Okay, it looks like the instruction IS in the "condition". Check to 284 // see if it's a cheap instruction to unconditionally compute, and if it 285 // only uses stuff defined outside of the condition. If so, hoist it out. 286 if (!isSafeToSpeculativelyExecute(I)) 287 return false; 288 289 unsigned Cost = ComputeSpeculationCost(I); 290 291 if (Cost > CostRemaining) 292 return false; 293 294 CostRemaining -= Cost; 295 296 // Okay, we can only really hoist these out if their operands do 297 // not take us over the cost threshold. 298 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 299 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 300 return false; 301 // Okay, it's safe to do this! Remember this instruction. 302 AggressiveInsts->insert(I); 303 return true; 304} 305 306/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 307/// and PointerNullValue. Return NULL if value is not a constant int. 308static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 309 // Normal constant int. 310 ConstantInt *CI = dyn_cast<ConstantInt>(V); 311 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 312 return CI; 313 314 // This is some kind of pointer constant. Turn it into a pointer-sized 315 // ConstantInt if possible. 316 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); 317 318 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 319 if (isa<ConstantPointerNull>(V)) 320 return ConstantInt::get(PtrTy, 0); 321 322 // IntToPtr const int. 323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 324 if (CE->getOpcode() == Instruction::IntToPtr) 325 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 326 // The constant is very likely to have the right type already. 327 if (CI->getType() == PtrTy) 328 return CI; 329 else 330 return cast<ConstantInt> 331 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 332 } 333 return 0; 334} 335 336/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 337/// collection of icmp eq/ne instructions that compare a value against a 338/// constant, return the value being compared, and stick the constant into the 339/// Values vector. 340static Value * 341GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 342 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 343 Instruction *I = dyn_cast<Instruction>(V); 344 if (I == 0) return 0; 345 346 // If this is an icmp against a constant, handle this as one of the cases. 347 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 348 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 349 Value *RHSVal; 350 ConstantInt *RHSC; 351 352 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 353 // (x & ~2^x) == y --> x == y || x == y|2^x 354 // This undoes a transformation done by instcombine to fuse 2 compares. 355 if (match(ICI->getOperand(0), 356 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) { 357 APInt Not = ~RHSC->getValue(); 358 if (Not.isPowerOf2()) { 359 Vals.push_back(C); 360 Vals.push_back( 361 ConstantInt::get(C->getContext(), C->getValue() | Not)); 362 UsedICmps++; 363 return RHSVal; 364 } 365 } 366 367 UsedICmps++; 368 Vals.push_back(C); 369 return I->getOperand(0); 370 } 371 372 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 373 // the set. 374 ConstantRange Span = 375 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 376 377 // Shift the range if the compare is fed by an add. This is the range 378 // compare idiom as emitted by instcombine. 379 bool hasAdd = 380 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC))); 381 if (hasAdd) 382 Span = Span.subtract(RHSC->getValue()); 383 384 // If this is an and/!= check then we want to optimize "x ugt 2" into 385 // x != 0 && x != 1. 386 if (!isEQ) 387 Span = Span.inverse(); 388 389 // If there are a ton of values, we don't want to make a ginormous switch. 390 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 391 return 0; 392 393 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 394 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 395 UsedICmps++; 396 return hasAdd ? RHSVal : I->getOperand(0); 397 } 398 return 0; 399 } 400 401 // Otherwise, we can only handle an | or &, depending on isEQ. 402 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 403 return 0; 404 405 unsigned NumValsBeforeLHS = Vals.size(); 406 unsigned UsedICmpsBeforeLHS = UsedICmps; 407 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 408 isEQ, UsedICmps)) { 409 unsigned NumVals = Vals.size(); 410 unsigned UsedICmpsBeforeRHS = UsedICmps; 411 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 412 isEQ, UsedICmps)) { 413 if (LHS == RHS) 414 return LHS; 415 Vals.resize(NumVals); 416 UsedICmps = UsedICmpsBeforeRHS; 417 } 418 419 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 420 // set it and return success. 421 if (Extra == 0 || Extra == I->getOperand(1)) { 422 Extra = I->getOperand(1); 423 return LHS; 424 } 425 426 Vals.resize(NumValsBeforeLHS); 427 UsedICmps = UsedICmpsBeforeLHS; 428 return 0; 429 } 430 431 // If the LHS can't be folded in, but Extra is available and RHS can, try to 432 // use LHS as Extra. 433 if (Extra == 0 || Extra == I->getOperand(0)) { 434 Value *OldExtra = Extra; 435 Extra = I->getOperand(0); 436 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 437 isEQ, UsedICmps)) 438 return RHS; 439 assert(Vals.size() == NumValsBeforeLHS); 440 Extra = OldExtra; 441 } 442 443 return 0; 444} 445 446static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 447 Instruction *Cond = 0; 448 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 449 Cond = dyn_cast<Instruction>(SI->getCondition()); 450 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 451 if (BI->isConditional()) 452 Cond = dyn_cast<Instruction>(BI->getCondition()); 453 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 454 Cond = dyn_cast<Instruction>(IBI->getAddress()); 455 } 456 457 TI->eraseFromParent(); 458 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 459} 460 461/// isValueEqualityComparison - Return true if the specified terminator checks 462/// to see if a value is equal to constant integer value. 463Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 464 Value *CV = 0; 465 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 466 // Do not permit merging of large switch instructions into their 467 // predecessors unless there is only one predecessor. 468 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 469 pred_end(SI->getParent())) <= 128) 470 CV = SI->getCondition(); 471 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 472 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 473 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 474 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD)) 475 CV = ICI->getOperand(0); 476 477 // Unwrap any lossless ptrtoint cast. 478 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 479 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 480 CV = PTII->getOperand(0); 481 return CV; 482} 483 484/// GetValueEqualityComparisonCases - Given a value comparison instruction, 485/// decode all of the 'cases' that it represents and return the 'default' block. 486BasicBlock *SimplifyCFGOpt:: 487GetValueEqualityComparisonCases(TerminatorInst *TI, 488 std::vector<ValueEqualityComparisonCase> 489 &Cases) { 490 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 491 Cases.reserve(SI->getNumCases()); 492 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 493 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 494 i.getCaseSuccessor())); 495 return SI->getDefaultDest(); 496 } 497 498 BranchInst *BI = cast<BranchInst>(TI); 499 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 500 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 501 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 502 TD), 503 Succ)); 504 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 505} 506 507 508/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 509/// in the list that match the specified block. 510static void EliminateBlockCases(BasicBlock *BB, 511 std::vector<ValueEqualityComparisonCase> &Cases) { 512 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 513} 514 515/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 516/// well. 517static bool 518ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 519 std::vector<ValueEqualityComparisonCase > &C2) { 520 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 521 522 // Make V1 be smaller than V2. 523 if (V1->size() > V2->size()) 524 std::swap(V1, V2); 525 526 if (V1->size() == 0) return false; 527 if (V1->size() == 1) { 528 // Just scan V2. 529 ConstantInt *TheVal = (*V1)[0].Value; 530 for (unsigned i = 0, e = V2->size(); i != e; ++i) 531 if (TheVal == (*V2)[i].Value) 532 return true; 533 } 534 535 // Otherwise, just sort both lists and compare element by element. 536 array_pod_sort(V1->begin(), V1->end()); 537 array_pod_sort(V2->begin(), V2->end()); 538 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 539 while (i1 != e1 && i2 != e2) { 540 if ((*V1)[i1].Value == (*V2)[i2].Value) 541 return true; 542 if ((*V1)[i1].Value < (*V2)[i2].Value) 543 ++i1; 544 else 545 ++i2; 546 } 547 return false; 548} 549 550/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 551/// terminator instruction and its block is known to only have a single 552/// predecessor block, check to see if that predecessor is also a value 553/// comparison with the same value, and if that comparison determines the 554/// outcome of this comparison. If so, simplify TI. This does a very limited 555/// form of jump threading. 556bool SimplifyCFGOpt:: 557SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 558 BasicBlock *Pred, 559 IRBuilder<> &Builder) { 560 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 561 if (!PredVal) return false; // Not a value comparison in predecessor. 562 563 Value *ThisVal = isValueEqualityComparison(TI); 564 assert(ThisVal && "This isn't a value comparison!!"); 565 if (ThisVal != PredVal) return false; // Different predicates. 566 567 // TODO: Preserve branch weight metadata, similarly to how 568 // FoldValueComparisonIntoPredecessors preserves it. 569 570 // Find out information about when control will move from Pred to TI's block. 571 std::vector<ValueEqualityComparisonCase> PredCases; 572 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 573 PredCases); 574 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 575 576 // Find information about how control leaves this block. 577 std::vector<ValueEqualityComparisonCase> ThisCases; 578 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 579 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 580 581 // If TI's block is the default block from Pred's comparison, potentially 582 // simplify TI based on this knowledge. 583 if (PredDef == TI->getParent()) { 584 // If we are here, we know that the value is none of those cases listed in 585 // PredCases. If there are any cases in ThisCases that are in PredCases, we 586 // can simplify TI. 587 if (!ValuesOverlap(PredCases, ThisCases)) 588 return false; 589 590 if (isa<BranchInst>(TI)) { 591 // Okay, one of the successors of this condbr is dead. Convert it to a 592 // uncond br. 593 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 594 // Insert the new branch. 595 Instruction *NI = Builder.CreateBr(ThisDef); 596 (void) NI; 597 598 // Remove PHI node entries for the dead edge. 599 ThisCases[0].Dest->removePredecessor(TI->getParent()); 600 601 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 602 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 603 604 EraseTerminatorInstAndDCECond(TI); 605 return true; 606 } 607 608 SwitchInst *SI = cast<SwitchInst>(TI); 609 // Okay, TI has cases that are statically dead, prune them away. 610 SmallPtrSet<Constant*, 16> DeadCases; 611 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 612 DeadCases.insert(PredCases[i].Value); 613 614 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 615 << "Through successor TI: " << *TI); 616 617 // Collect branch weights into a vector. 618 SmallVector<uint32_t, 8> Weights; 619 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 620 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 621 if (HasWeight) 622 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 623 ++MD_i) { 624 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 625 assert(CI); 626 Weights.push_back(CI->getValue().getZExtValue()); 627 } 628 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 629 --i; 630 if (DeadCases.count(i.getCaseValue())) { 631 if (HasWeight) { 632 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 633 Weights.pop_back(); 634 } 635 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 636 SI->removeCase(i); 637 } 638 } 639 if (HasWeight && Weights.size() >= 2) 640 SI->setMetadata(LLVMContext::MD_prof, 641 MDBuilder(SI->getParent()->getContext()). 642 createBranchWeights(Weights)); 643 644 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 645 return true; 646 } 647 648 // Otherwise, TI's block must correspond to some matched value. Find out 649 // which value (or set of values) this is. 650 ConstantInt *TIV = 0; 651 BasicBlock *TIBB = TI->getParent(); 652 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 653 if (PredCases[i].Dest == TIBB) { 654 if (TIV != 0) 655 return false; // Cannot handle multiple values coming to this block. 656 TIV = PredCases[i].Value; 657 } 658 assert(TIV && "No edge from pred to succ?"); 659 660 // Okay, we found the one constant that our value can be if we get into TI's 661 // BB. Find out which successor will unconditionally be branched to. 662 BasicBlock *TheRealDest = 0; 663 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 664 if (ThisCases[i].Value == TIV) { 665 TheRealDest = ThisCases[i].Dest; 666 break; 667 } 668 669 // If not handled by any explicit cases, it is handled by the default case. 670 if (TheRealDest == 0) TheRealDest = ThisDef; 671 672 // Remove PHI node entries for dead edges. 673 BasicBlock *CheckEdge = TheRealDest; 674 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 675 if (*SI != CheckEdge) 676 (*SI)->removePredecessor(TIBB); 677 else 678 CheckEdge = 0; 679 680 // Insert the new branch. 681 Instruction *NI = Builder.CreateBr(TheRealDest); 682 (void) NI; 683 684 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 685 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 686 687 EraseTerminatorInstAndDCECond(TI); 688 return true; 689} 690 691namespace { 692 /// ConstantIntOrdering - This class implements a stable ordering of constant 693 /// integers that does not depend on their address. This is important for 694 /// applications that sort ConstantInt's to ensure uniqueness. 695 struct ConstantIntOrdering { 696 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 697 return LHS->getValue().ult(RHS->getValue()); 698 } 699 }; 700} 701 702static int ConstantIntSortPredicate(const void *P1, const void *P2) { 703 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 704 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 705 if (LHS->getValue().ult(RHS->getValue())) 706 return 1; 707 if (LHS->getValue() == RHS->getValue()) 708 return 0; 709 return -1; 710} 711 712static inline bool HasBranchWeights(const Instruction* I) { 713 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 714 if (ProfMD && ProfMD->getOperand(0)) 715 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 716 return MDS->getString().equals("branch_weights"); 717 718 return false; 719} 720 721/// Get Weights of a given TerminatorInst, the default weight is at the front 722/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 723/// metadata. 724static void GetBranchWeights(TerminatorInst *TI, 725 SmallVectorImpl<uint64_t> &Weights) { 726 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 727 assert(MD); 728 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 729 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 730 assert(CI); 731 Weights.push_back(CI->getValue().getZExtValue()); 732 } 733 734 // If TI is a conditional eq, the default case is the false case, 735 // and the corresponding branch-weight data is at index 2. We swap the 736 // default weight to be the first entry. 737 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 738 assert(Weights.size() == 2); 739 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 740 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 741 std::swap(Weights.front(), Weights.back()); 742 } 743} 744 745/// Sees if any of the weights are too big for a uint32_t, and halves all the 746/// weights if any are. 747static void FitWeights(MutableArrayRef<uint64_t> Weights) { 748 bool Halve = false; 749 for (unsigned i = 0; i < Weights.size(); ++i) 750 if (Weights[i] > UINT_MAX) { 751 Halve = true; 752 break; 753 } 754 755 if (! Halve) 756 return; 757 758 for (unsigned i = 0; i < Weights.size(); ++i) 759 Weights[i] /= 2; 760} 761 762/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 763/// equality comparison instruction (either a switch or a branch on "X == c"). 764/// See if any of the predecessors of the terminator block are value comparisons 765/// on the same value. If so, and if safe to do so, fold them together. 766bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 767 IRBuilder<> &Builder) { 768 BasicBlock *BB = TI->getParent(); 769 Value *CV = isValueEqualityComparison(TI); // CondVal 770 assert(CV && "Not a comparison?"); 771 bool Changed = false; 772 773 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 774 while (!Preds.empty()) { 775 BasicBlock *Pred = Preds.pop_back_val(); 776 777 // See if the predecessor is a comparison with the same value. 778 TerminatorInst *PTI = Pred->getTerminator(); 779 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 780 781 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 782 // Figure out which 'cases' to copy from SI to PSI. 783 std::vector<ValueEqualityComparisonCase> BBCases; 784 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 785 786 std::vector<ValueEqualityComparisonCase> PredCases; 787 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 788 789 // Based on whether the default edge from PTI goes to BB or not, fill in 790 // PredCases and PredDefault with the new switch cases we would like to 791 // build. 792 SmallVector<BasicBlock*, 8> NewSuccessors; 793 794 // Update the branch weight metadata along the way 795 SmallVector<uint64_t, 8> Weights; 796 bool PredHasWeights = HasBranchWeights(PTI); 797 bool SuccHasWeights = HasBranchWeights(TI); 798 799 if (PredHasWeights) { 800 GetBranchWeights(PTI, Weights); 801 // branch-weight metadata is inconsistent here. 802 if (Weights.size() != 1 + PredCases.size()) 803 PredHasWeights = SuccHasWeights = false; 804 } else if (SuccHasWeights) 805 // If there are no predecessor weights but there are successor weights, 806 // populate Weights with 1, which will later be scaled to the sum of 807 // successor's weights 808 Weights.assign(1 + PredCases.size(), 1); 809 810 SmallVector<uint64_t, 8> SuccWeights; 811 if (SuccHasWeights) { 812 GetBranchWeights(TI, SuccWeights); 813 // branch-weight metadata is inconsistent here. 814 if (SuccWeights.size() != 1 + BBCases.size()) 815 PredHasWeights = SuccHasWeights = false; 816 } else if (PredHasWeights) 817 SuccWeights.assign(1 + BBCases.size(), 1); 818 819 if (PredDefault == BB) { 820 // If this is the default destination from PTI, only the edges in TI 821 // that don't occur in PTI, or that branch to BB will be activated. 822 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 823 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 824 if (PredCases[i].Dest != BB) 825 PTIHandled.insert(PredCases[i].Value); 826 else { 827 // The default destination is BB, we don't need explicit targets. 828 std::swap(PredCases[i], PredCases.back()); 829 830 if (PredHasWeights || SuccHasWeights) { 831 // Increase weight for the default case. 832 Weights[0] += Weights[i+1]; 833 std::swap(Weights[i+1], Weights.back()); 834 Weights.pop_back(); 835 } 836 837 PredCases.pop_back(); 838 --i; --e; 839 } 840 841 // Reconstruct the new switch statement we will be building. 842 if (PredDefault != BBDefault) { 843 PredDefault->removePredecessor(Pred); 844 PredDefault = BBDefault; 845 NewSuccessors.push_back(BBDefault); 846 } 847 848 unsigned CasesFromPred = Weights.size(); 849 uint64_t ValidTotalSuccWeight = 0; 850 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 851 if (!PTIHandled.count(BBCases[i].Value) && 852 BBCases[i].Dest != BBDefault) { 853 PredCases.push_back(BBCases[i]); 854 NewSuccessors.push_back(BBCases[i].Dest); 855 if (SuccHasWeights || PredHasWeights) { 856 // The default weight is at index 0, so weight for the ith case 857 // should be at index i+1. Scale the cases from successor by 858 // PredDefaultWeight (Weights[0]). 859 Weights.push_back(Weights[0] * SuccWeights[i+1]); 860 ValidTotalSuccWeight += SuccWeights[i+1]; 861 } 862 } 863 864 if (SuccHasWeights || PredHasWeights) { 865 ValidTotalSuccWeight += SuccWeights[0]; 866 // Scale the cases from predecessor by ValidTotalSuccWeight. 867 for (unsigned i = 1; i < CasesFromPred; ++i) 868 Weights[i] *= ValidTotalSuccWeight; 869 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 870 Weights[0] *= SuccWeights[0]; 871 } 872 } else { 873 // If this is not the default destination from PSI, only the edges 874 // in SI that occur in PSI with a destination of BB will be 875 // activated. 876 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 877 std::map<ConstantInt*, uint64_t> WeightsForHandled; 878 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 879 if (PredCases[i].Dest == BB) { 880 PTIHandled.insert(PredCases[i].Value); 881 882 if (PredHasWeights || SuccHasWeights) { 883 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 884 std::swap(Weights[i+1], Weights.back()); 885 Weights.pop_back(); 886 } 887 888 std::swap(PredCases[i], PredCases.back()); 889 PredCases.pop_back(); 890 --i; --e; 891 } 892 893 // Okay, now we know which constants were sent to BB from the 894 // predecessor. Figure out where they will all go now. 895 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 896 if (PTIHandled.count(BBCases[i].Value)) { 897 // If this is one we are capable of getting... 898 if (PredHasWeights || SuccHasWeights) 899 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 900 PredCases.push_back(BBCases[i]); 901 NewSuccessors.push_back(BBCases[i].Dest); 902 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 903 } 904 905 // If there are any constants vectored to BB that TI doesn't handle, 906 // they must go to the default destination of TI. 907 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 908 PTIHandled.begin(), 909 E = PTIHandled.end(); I != E; ++I) { 910 if (PredHasWeights || SuccHasWeights) 911 Weights.push_back(WeightsForHandled[*I]); 912 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 913 NewSuccessors.push_back(BBDefault); 914 } 915 } 916 917 // Okay, at this point, we know which new successor Pred will get. Make 918 // sure we update the number of entries in the PHI nodes for these 919 // successors. 920 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 921 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 922 923 Builder.SetInsertPoint(PTI); 924 // Convert pointer to int before we switch. 925 if (CV->getType()->isPointerTy()) { 926 assert(TD && "Cannot switch on pointer without DataLayout"); 927 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 928 "magicptr"); 929 } 930 931 // Now that the successors are updated, create the new Switch instruction. 932 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 933 PredCases.size()); 934 NewSI->setDebugLoc(PTI->getDebugLoc()); 935 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 936 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 937 938 if (PredHasWeights || SuccHasWeights) { 939 // Halve the weights if any of them cannot fit in an uint32_t 940 FitWeights(Weights); 941 942 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 943 944 NewSI->setMetadata(LLVMContext::MD_prof, 945 MDBuilder(BB->getContext()). 946 createBranchWeights(MDWeights)); 947 } 948 949 EraseTerminatorInstAndDCECond(PTI); 950 951 // Okay, last check. If BB is still a successor of PSI, then we must 952 // have an infinite loop case. If so, add an infinitely looping block 953 // to handle the case to preserve the behavior of the code. 954 BasicBlock *InfLoopBlock = 0; 955 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 956 if (NewSI->getSuccessor(i) == BB) { 957 if (InfLoopBlock == 0) { 958 // Insert it at the end of the function, because it's either code, 959 // or it won't matter if it's hot. :) 960 InfLoopBlock = BasicBlock::Create(BB->getContext(), 961 "infloop", BB->getParent()); 962 BranchInst::Create(InfLoopBlock, InfLoopBlock); 963 } 964 NewSI->setSuccessor(i, InfLoopBlock); 965 } 966 967 Changed = true; 968 } 969 } 970 return Changed; 971} 972 973// isSafeToHoistInvoke - If we would need to insert a select that uses the 974// value of this invoke (comments in HoistThenElseCodeToIf explain why we 975// would need to do this), we can't hoist the invoke, as there is nowhere 976// to put the select in this case. 977static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 978 Instruction *I1, Instruction *I2) { 979 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 980 PHINode *PN; 981 for (BasicBlock::iterator BBI = SI->begin(); 982 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 983 Value *BB1V = PN->getIncomingValueForBlock(BB1); 984 Value *BB2V = PN->getIncomingValueForBlock(BB2); 985 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 986 return false; 987 } 988 } 989 } 990 return true; 991} 992 993/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 994/// BB2, hoist any common code in the two blocks up into the branch block. The 995/// caller of this function guarantees that BI's block dominates BB1 and BB2. 996static bool HoistThenElseCodeToIf(BranchInst *BI) { 997 // This does very trivial matching, with limited scanning, to find identical 998 // instructions in the two blocks. In particular, we don't want to get into 999 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1000 // such, we currently just scan for obviously identical instructions in an 1001 // identical order. 1002 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1003 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1004 1005 BasicBlock::iterator BB1_Itr = BB1->begin(); 1006 BasicBlock::iterator BB2_Itr = BB2->begin(); 1007 1008 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1009 // Skip debug info if it is not identical. 1010 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1011 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1012 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1013 while (isa<DbgInfoIntrinsic>(I1)) 1014 I1 = BB1_Itr++; 1015 while (isa<DbgInfoIntrinsic>(I2)) 1016 I2 = BB2_Itr++; 1017 } 1018 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1019 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1020 return false; 1021 1022 BasicBlock *BIParent = BI->getParent(); 1023 1024 bool Changed = false; 1025 do { 1026 // If we are hoisting the terminator instruction, don't move one (making a 1027 // broken BB), instead clone it, and remove BI. 1028 if (isa<TerminatorInst>(I1)) 1029 goto HoistTerminator; 1030 1031 // For a normal instruction, we just move one to right before the branch, 1032 // then replace all uses of the other with the first. Finally, we remove 1033 // the now redundant second instruction. 1034 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1035 if (!I2->use_empty()) 1036 I2->replaceAllUsesWith(I1); 1037 I1->intersectOptionalDataWith(I2); 1038 I2->eraseFromParent(); 1039 Changed = true; 1040 1041 I1 = BB1_Itr++; 1042 I2 = BB2_Itr++; 1043 // Skip debug info if it is not identical. 1044 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1045 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1046 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1047 while (isa<DbgInfoIntrinsic>(I1)) 1048 I1 = BB1_Itr++; 1049 while (isa<DbgInfoIntrinsic>(I2)) 1050 I2 = BB2_Itr++; 1051 } 1052 } while (I1->isIdenticalToWhenDefined(I2)); 1053 1054 return true; 1055 1056HoistTerminator: 1057 // It may not be possible to hoist an invoke. 1058 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1059 return Changed; 1060 1061 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1062 PHINode *PN; 1063 for (BasicBlock::iterator BBI = SI->begin(); 1064 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1065 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1066 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1067 if (BB1V == BB2V) 1068 continue; 1069 1070 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V)) 1071 return Changed; 1072 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V)) 1073 return Changed; 1074 } 1075 } 1076 1077 // Okay, it is safe to hoist the terminator. 1078 Instruction *NT = I1->clone(); 1079 BIParent->getInstList().insert(BI, NT); 1080 if (!NT->getType()->isVoidTy()) { 1081 I1->replaceAllUsesWith(NT); 1082 I2->replaceAllUsesWith(NT); 1083 NT->takeName(I1); 1084 } 1085 1086 IRBuilder<true, NoFolder> Builder(NT); 1087 // Hoisting one of the terminators from our successor is a great thing. 1088 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1089 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1090 // nodes, so we insert select instruction to compute the final result. 1091 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1092 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1093 PHINode *PN; 1094 for (BasicBlock::iterator BBI = SI->begin(); 1095 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1096 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1097 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1098 if (BB1V == BB2V) continue; 1099 1100 // These values do not agree. Insert a select instruction before NT 1101 // that determines the right value. 1102 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1103 if (SI == 0) 1104 SI = cast<SelectInst> 1105 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1106 BB1V->getName()+"."+BB2V->getName())); 1107 1108 // Make the PHI node use the select for all incoming values for BB1/BB2 1109 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1110 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1111 PN->setIncomingValue(i, SI); 1112 } 1113 } 1114 1115 // Update any PHI nodes in our new successors. 1116 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1117 AddPredecessorToBlock(*SI, BIParent, BB1); 1118 1119 EraseTerminatorInstAndDCECond(BI); 1120 return true; 1121} 1122 1123/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1124/// check whether BBEnd has only two predecessors and the other predecessor 1125/// ends with an unconditional branch. If it is true, sink any common code 1126/// in the two predecessors to BBEnd. 1127static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1128 assert(BI1->isUnconditional()); 1129 BasicBlock *BB1 = BI1->getParent(); 1130 BasicBlock *BBEnd = BI1->getSuccessor(0); 1131 1132 // Check that BBEnd has two predecessors and the other predecessor ends with 1133 // an unconditional branch. 1134 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1135 BasicBlock *Pred0 = *PI++; 1136 if (PI == PE) // Only one predecessor. 1137 return false; 1138 BasicBlock *Pred1 = *PI++; 1139 if (PI != PE) // More than two predecessors. 1140 return false; 1141 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1142 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1143 if (!BI2 || !BI2->isUnconditional()) 1144 return false; 1145 1146 // Gather the PHI nodes in BBEnd. 1147 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1148 Instruction *FirstNonPhiInBBEnd = 0; 1149 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1150 I != E; ++I) { 1151 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1152 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1153 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1154 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1155 } else { 1156 FirstNonPhiInBBEnd = &*I; 1157 break; 1158 } 1159 } 1160 if (!FirstNonPhiInBBEnd) 1161 return false; 1162 1163 1164 // This does very trivial matching, with limited scanning, to find identical 1165 // instructions in the two blocks. We scan backward for obviously identical 1166 // instructions in an identical order. 1167 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1168 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1169 RE2 = BB2->getInstList().rend(); 1170 // Skip debug info. 1171 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1172 if (RI1 == RE1) 1173 return false; 1174 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1175 if (RI2 == RE2) 1176 return false; 1177 // Skip the unconditional branches. 1178 ++RI1; 1179 ++RI2; 1180 1181 bool Changed = false; 1182 while (RI1 != RE1 && RI2 != RE2) { 1183 // Skip debug info. 1184 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1185 if (RI1 == RE1) 1186 return Changed; 1187 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1188 if (RI2 == RE2) 1189 return Changed; 1190 1191 Instruction *I1 = &*RI1, *I2 = &*RI2; 1192 // I1 and I2 should have a single use in the same PHI node, and they 1193 // perform the same operation. 1194 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1195 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1196 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1197 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1198 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1199 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1200 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1201 !I1->hasOneUse() || !I2->hasOneUse() || 1202 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1203 MapValueFromBB1ToBB2[I1].first != I2) 1204 return Changed; 1205 1206 // Check whether we should swap the operands of ICmpInst. 1207 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1208 bool SwapOpnds = false; 1209 if (ICmp1 && ICmp2 && 1210 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1211 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1212 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1213 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1214 ICmp2->swapOperands(); 1215 SwapOpnds = true; 1216 } 1217 if (!I1->isSameOperationAs(I2)) { 1218 if (SwapOpnds) 1219 ICmp2->swapOperands(); 1220 return Changed; 1221 } 1222 1223 // The operands should be either the same or they need to be generated 1224 // with a PHI node after sinking. We only handle the case where there is 1225 // a single pair of different operands. 1226 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1227 unsigned Op1Idx = 0; 1228 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1229 if (I1->getOperand(I) == I2->getOperand(I)) 1230 continue; 1231 // Early exit if we have more-than one pair of different operands or 1232 // the different operand is already in MapValueFromBB1ToBB2. 1233 // Early exit if we need a PHI node to replace a constant. 1234 if (DifferentOp1 || 1235 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1236 MapValueFromBB1ToBB2.end() || 1237 isa<Constant>(I1->getOperand(I)) || 1238 isa<Constant>(I2->getOperand(I))) { 1239 // If we can't sink the instructions, undo the swapping. 1240 if (SwapOpnds) 1241 ICmp2->swapOperands(); 1242 return Changed; 1243 } 1244 DifferentOp1 = I1->getOperand(I); 1245 Op1Idx = I; 1246 DifferentOp2 = I2->getOperand(I); 1247 } 1248 1249 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1250 // remove (I1, I2) from MapValueFromBB1ToBB2. 1251 if (DifferentOp1) { 1252 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1253 DifferentOp1->getName() + ".sink", 1254 BBEnd->begin()); 1255 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1256 // I1 should use NewPN instead of DifferentOp1. 1257 I1->setOperand(Op1Idx, NewPN); 1258 NewPN->addIncoming(DifferentOp1, BB1); 1259 NewPN->addIncoming(DifferentOp2, BB2); 1260 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1261 } 1262 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1263 MapValueFromBB1ToBB2.erase(I1); 1264 1265 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1266 DEBUG(dbgs() << " " << *I2 << "\n";); 1267 // We need to update RE1 and RE2 if we are going to sink the first 1268 // instruction in the basic block down. 1269 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1270 // Sink the instruction. 1271 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1272 if (!OldPN->use_empty()) 1273 OldPN->replaceAllUsesWith(I1); 1274 OldPN->eraseFromParent(); 1275 1276 if (!I2->use_empty()) 1277 I2->replaceAllUsesWith(I1); 1278 I1->intersectOptionalDataWith(I2); 1279 I2->eraseFromParent(); 1280 1281 if (UpdateRE1) 1282 RE1 = BB1->getInstList().rend(); 1283 if (UpdateRE2) 1284 RE2 = BB2->getInstList().rend(); 1285 FirstNonPhiInBBEnd = I1; 1286 NumSinkCommons++; 1287 Changed = true; 1288 } 1289 return Changed; 1290} 1291 1292/// \brief Determine if we can hoist sink a sole store instruction out of a 1293/// conditional block. 1294/// 1295/// We are looking for code like the following: 1296/// BrBB: 1297/// store i32 %add, i32* %arrayidx2 1298/// ... // No other stores or function calls (we could be calling a memory 1299/// ... // function). 1300/// %cmp = icmp ult %x, %y 1301/// br i1 %cmp, label %EndBB, label %ThenBB 1302/// ThenBB: 1303/// store i32 %add5, i32* %arrayidx2 1304/// br label EndBB 1305/// EndBB: 1306/// ... 1307/// We are going to transform this into: 1308/// BrBB: 1309/// store i32 %add, i32* %arrayidx2 1310/// ... // 1311/// %cmp = icmp ult %x, %y 1312/// %add.add5 = select i1 %cmp, i32 %add, %add5 1313/// store i32 %add.add5, i32* %arrayidx2 1314/// ... 1315/// 1316/// \return The pointer to the value of the previous store if the store can be 1317/// hoisted into the predecessor block. 0 otherwise. 1318static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, 1319 BasicBlock *StoreBB, BasicBlock *EndBB) { 1320 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); 1321 if (!StoreToHoist) 1322 return 0; 1323 1324 // Volatile or atomic. 1325 if (!StoreToHoist->isSimple()) 1326 return 0; 1327 1328 Value *StorePtr = StoreToHoist->getPointerOperand(); 1329 1330 // Look for a store to the same pointer in BrBB. 1331 unsigned MaxNumInstToLookAt = 10; 1332 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), 1333 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { 1334 Instruction *CurI = &*RI; 1335 1336 // Could be calling an instruction that effects memory like free(). 1337 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) 1338 return 0; 1339 1340 StoreInst *SI = dyn_cast<StoreInst>(CurI); 1341 // Found the previous store make sure it stores to the same location. 1342 if (SI && SI->getPointerOperand() == StorePtr) 1343 // Found the previous store, return its value operand. 1344 return SI->getValueOperand(); 1345 else if (SI) 1346 return 0; // Unknown store. 1347 } 1348 1349 return 0; 1350} 1351 1352/// \brief Speculate a conditional basic block flattening the CFG. 1353/// 1354/// Note that this is a very risky transform currently. Speculating 1355/// instructions like this is most often not desirable. Instead, there is an MI 1356/// pass which can do it with full awareness of the resource constraints. 1357/// However, some cases are "obvious" and we should do directly. An example of 1358/// this is speculating a single, reasonably cheap instruction. 1359/// 1360/// There is only one distinct advantage to flattening the CFG at the IR level: 1361/// it makes very common but simplistic optimizations such as are common in 1362/// instcombine and the DAG combiner more powerful by removing CFG edges and 1363/// modeling their effects with easier to reason about SSA value graphs. 1364/// 1365/// 1366/// An illustration of this transform is turning this IR: 1367/// \code 1368/// BB: 1369/// %cmp = icmp ult %x, %y 1370/// br i1 %cmp, label %EndBB, label %ThenBB 1371/// ThenBB: 1372/// %sub = sub %x, %y 1373/// br label BB2 1374/// EndBB: 1375/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1376/// ... 1377/// \endcode 1378/// 1379/// Into this IR: 1380/// \code 1381/// BB: 1382/// %cmp = icmp ult %x, %y 1383/// %sub = sub %x, %y 1384/// %cond = select i1 %cmp, 0, %sub 1385/// ... 1386/// \endcode 1387/// 1388/// \returns true if the conditional block is removed. 1389static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1390 // Be conservative for now. FP select instruction can often be expensive. 1391 Value *BrCond = BI->getCondition(); 1392 if (isa<FCmpInst>(BrCond)) 1393 return false; 1394 1395 BasicBlock *BB = BI->getParent(); 1396 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1397 1398 // If ThenBB is actually on the false edge of the conditional branch, remember 1399 // to swap the select operands later. 1400 bool Invert = false; 1401 if (ThenBB != BI->getSuccessor(0)) { 1402 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1403 Invert = true; 1404 } 1405 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1406 1407 // Keep a count of how many times instructions are used within CondBB when 1408 // they are candidates for sinking into CondBB. Specifically: 1409 // - They are defined in BB, and 1410 // - They have no side effects, and 1411 // - All of their uses are in CondBB. 1412 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1413 1414 unsigned SpeculationCost = 0; 1415 Value *SpeculatedStoreValue = 0; 1416 StoreInst *SpeculatedStore = 0; 1417 for (BasicBlock::iterator BBI = ThenBB->begin(), 1418 BBE = llvm::prior(ThenBB->end()); 1419 BBI != BBE; ++BBI) { 1420 Instruction *I = BBI; 1421 // Skip debug info. 1422 if (isa<DbgInfoIntrinsic>(I)) 1423 continue; 1424 1425 // Only speculatively execution a single instruction (not counting the 1426 // terminator) for now. 1427 ++SpeculationCost; 1428 if (SpeculationCost > 1) 1429 return false; 1430 1431 // Don't hoist the instruction if it's unsafe or expensive. 1432 if (!isSafeToSpeculativelyExecute(I) && 1433 !(HoistCondStores && 1434 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, 1435 EndBB)))) 1436 return false; 1437 if (!SpeculatedStoreValue && 1438 ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1439 return false; 1440 1441 // Store the store speculation candidate. 1442 if (SpeculatedStoreValue) 1443 SpeculatedStore = cast<StoreInst>(I); 1444 1445 // Do not hoist the instruction if any of its operands are defined but not 1446 // used in BB. The transformation will prevent the operand from 1447 // being sunk into the use block. 1448 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1449 i != e; ++i) { 1450 Instruction *OpI = dyn_cast<Instruction>(*i); 1451 if (!OpI || OpI->getParent() != BB || 1452 OpI->mayHaveSideEffects()) 1453 continue; // Not a candidate for sinking. 1454 1455 ++SinkCandidateUseCounts[OpI]; 1456 } 1457 } 1458 1459 // Consider any sink candidates which are only used in CondBB as costs for 1460 // speculation. Note, while we iterate over a DenseMap here, we are summing 1461 // and so iteration order isn't significant. 1462 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1463 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1464 I != E; ++I) 1465 if (I->first->getNumUses() == I->second) { 1466 ++SpeculationCost; 1467 if (SpeculationCost > 1) 1468 return false; 1469 } 1470 1471 // Check that the PHI nodes can be converted to selects. 1472 bool HaveRewritablePHIs = false; 1473 for (BasicBlock::iterator I = EndBB->begin(); 1474 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1475 Value *OrigV = PN->getIncomingValueForBlock(BB); 1476 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1477 1478 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf. 1479 // Skip PHIs which are trivial. 1480 if (ThenV == OrigV) 1481 continue; 1482 1483 HaveRewritablePHIs = true; 1484 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV); 1485 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV); 1486 if (!OrigCE && !ThenCE) 1487 continue; // Known safe and cheap. 1488 1489 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) || 1490 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE))) 1491 return false; 1492 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0; 1493 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0; 1494 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold) 1495 return false; 1496 1497 // Account for the cost of an unfolded ConstantExpr which could end up 1498 // getting expanded into Instructions. 1499 // FIXME: This doesn't account for how many operations are combined in the 1500 // constant expression. 1501 ++SpeculationCost; 1502 if (SpeculationCost > 1) 1503 return false; 1504 } 1505 1506 // If there are no PHIs to process, bail early. This helps ensure idempotence 1507 // as well. 1508 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) 1509 return false; 1510 1511 // If we get here, we can hoist the instruction and if-convert. 1512 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1513 1514 // Insert a select of the value of the speculated store. 1515 if (SpeculatedStoreValue) { 1516 IRBuilder<true, NoFolder> Builder(BI); 1517 Value *TrueV = SpeculatedStore->getValueOperand(); 1518 Value *FalseV = SpeculatedStoreValue; 1519 if (Invert) 1520 std::swap(TrueV, FalseV); 1521 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + 1522 "." + FalseV->getName()); 1523 SpeculatedStore->setOperand(0, S); 1524 } 1525 1526 // Hoist the instructions. 1527 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1528 llvm::prior(ThenBB->end())); 1529 1530 // Insert selects and rewrite the PHI operands. 1531 IRBuilder<true, NoFolder> Builder(BI); 1532 for (BasicBlock::iterator I = EndBB->begin(); 1533 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1534 unsigned OrigI = PN->getBasicBlockIndex(BB); 1535 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1536 Value *OrigV = PN->getIncomingValue(OrigI); 1537 Value *ThenV = PN->getIncomingValue(ThenI); 1538 1539 // Skip PHIs which are trivial. 1540 if (OrigV == ThenV) 1541 continue; 1542 1543 // Create a select whose true value is the speculatively executed value and 1544 // false value is the preexisting value. Swap them if the branch 1545 // destinations were inverted. 1546 Value *TrueV = ThenV, *FalseV = OrigV; 1547 if (Invert) 1548 std::swap(TrueV, FalseV); 1549 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1550 TrueV->getName() + "." + FalseV->getName()); 1551 PN->setIncomingValue(OrigI, V); 1552 PN->setIncomingValue(ThenI, V); 1553 } 1554 1555 ++NumSpeculations; 1556 return true; 1557} 1558 1559/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1560/// across this block. 1561static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1562 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1563 unsigned Size = 0; 1564 1565 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1566 if (isa<DbgInfoIntrinsic>(BBI)) 1567 continue; 1568 if (Size > 10) return false; // Don't clone large BB's. 1569 ++Size; 1570 1571 // We can only support instructions that do not define values that are 1572 // live outside of the current basic block. 1573 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1574 UI != E; ++UI) { 1575 Instruction *U = cast<Instruction>(*UI); 1576 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1577 } 1578 1579 // Looks ok, continue checking. 1580 } 1581 1582 return true; 1583} 1584 1585/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1586/// that is defined in the same block as the branch and if any PHI entries are 1587/// constants, thread edges corresponding to that entry to be branches to their 1588/// ultimate destination. 1589static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1590 BasicBlock *BB = BI->getParent(); 1591 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1592 // NOTE: we currently cannot transform this case if the PHI node is used 1593 // outside of the block. 1594 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1595 return false; 1596 1597 // Degenerate case of a single entry PHI. 1598 if (PN->getNumIncomingValues() == 1) { 1599 FoldSingleEntryPHINodes(PN->getParent()); 1600 return true; 1601 } 1602 1603 // Now we know that this block has multiple preds and two succs. 1604 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1605 1606 // Okay, this is a simple enough basic block. See if any phi values are 1607 // constants. 1608 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1609 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1610 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1611 1612 // Okay, we now know that all edges from PredBB should be revectored to 1613 // branch to RealDest. 1614 BasicBlock *PredBB = PN->getIncomingBlock(i); 1615 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1616 1617 if (RealDest == BB) continue; // Skip self loops. 1618 // Skip if the predecessor's terminator is an indirect branch. 1619 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1620 1621 // The dest block might have PHI nodes, other predecessors and other 1622 // difficult cases. Instead of being smart about this, just insert a new 1623 // block that jumps to the destination block, effectively splitting 1624 // the edge we are about to create. 1625 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1626 RealDest->getName()+".critedge", 1627 RealDest->getParent(), RealDest); 1628 BranchInst::Create(RealDest, EdgeBB); 1629 1630 // Update PHI nodes. 1631 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1632 1633 // BB may have instructions that are being threaded over. Clone these 1634 // instructions into EdgeBB. We know that there will be no uses of the 1635 // cloned instructions outside of EdgeBB. 1636 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1637 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1638 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1639 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1640 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1641 continue; 1642 } 1643 // Clone the instruction. 1644 Instruction *N = BBI->clone(); 1645 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1646 1647 // Update operands due to translation. 1648 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1649 i != e; ++i) { 1650 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1651 if (PI != TranslateMap.end()) 1652 *i = PI->second; 1653 } 1654 1655 // Check for trivial simplification. 1656 if (Value *V = SimplifyInstruction(N, TD)) { 1657 TranslateMap[BBI] = V; 1658 delete N; // Instruction folded away, don't need actual inst 1659 } else { 1660 // Insert the new instruction into its new home. 1661 EdgeBB->getInstList().insert(InsertPt, N); 1662 if (!BBI->use_empty()) 1663 TranslateMap[BBI] = N; 1664 } 1665 } 1666 1667 // Loop over all of the edges from PredBB to BB, changing them to branch 1668 // to EdgeBB instead. 1669 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1670 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1671 if (PredBBTI->getSuccessor(i) == BB) { 1672 BB->removePredecessor(PredBB); 1673 PredBBTI->setSuccessor(i, EdgeBB); 1674 } 1675 1676 // Recurse, simplifying any other constants. 1677 return FoldCondBranchOnPHI(BI, TD) | true; 1678 } 1679 1680 return false; 1681} 1682 1683/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1684/// PHI node, see if we can eliminate it. 1685static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1686 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1687 // statement", which has a very simple dominance structure. Basically, we 1688 // are trying to find the condition that is being branched on, which 1689 // subsequently causes this merge to happen. We really want control 1690 // dependence information for this check, but simplifycfg can't keep it up 1691 // to date, and this catches most of the cases we care about anyway. 1692 BasicBlock *BB = PN->getParent(); 1693 BasicBlock *IfTrue, *IfFalse; 1694 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1695 if (!IfCond || 1696 // Don't bother if the branch will be constant folded trivially. 1697 isa<ConstantInt>(IfCond)) 1698 return false; 1699 1700 // Okay, we found that we can merge this two-entry phi node into a select. 1701 // Doing so would require us to fold *all* two entry phi nodes in this block. 1702 // At some point this becomes non-profitable (particularly if the target 1703 // doesn't support cmov's). Only do this transformation if there are two or 1704 // fewer PHI nodes in this block. 1705 unsigned NumPhis = 0; 1706 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1707 if (NumPhis > 2) 1708 return false; 1709 1710 // Loop over the PHI's seeing if we can promote them all to select 1711 // instructions. While we are at it, keep track of the instructions 1712 // that need to be moved to the dominating block. 1713 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1714 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1715 MaxCostVal1 = PHINodeFoldingThreshold; 1716 1717 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1718 PHINode *PN = cast<PHINode>(II++); 1719 if (Value *V = SimplifyInstruction(PN, TD)) { 1720 PN->replaceAllUsesWith(V); 1721 PN->eraseFromParent(); 1722 continue; 1723 } 1724 1725 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1726 MaxCostVal0) || 1727 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1728 MaxCostVal1)) 1729 return false; 1730 } 1731 1732 // If we folded the first phi, PN dangles at this point. Refresh it. If 1733 // we ran out of PHIs then we simplified them all. 1734 PN = dyn_cast<PHINode>(BB->begin()); 1735 if (PN == 0) return true; 1736 1737 // Don't fold i1 branches on PHIs which contain binary operators. These can 1738 // often be turned into switches and other things. 1739 if (PN->getType()->isIntegerTy(1) && 1740 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1741 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1742 isa<BinaryOperator>(IfCond))) 1743 return false; 1744 1745 // If we all PHI nodes are promotable, check to make sure that all 1746 // instructions in the predecessor blocks can be promoted as well. If 1747 // not, we won't be able to get rid of the control flow, so it's not 1748 // worth promoting to select instructions. 1749 BasicBlock *DomBlock = 0; 1750 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1751 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1752 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1753 IfBlock1 = 0; 1754 } else { 1755 DomBlock = *pred_begin(IfBlock1); 1756 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1757 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1758 // This is not an aggressive instruction that we can promote. 1759 // Because of this, we won't be able to get rid of the control 1760 // flow, so the xform is not worth it. 1761 return false; 1762 } 1763 } 1764 1765 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1766 IfBlock2 = 0; 1767 } else { 1768 DomBlock = *pred_begin(IfBlock2); 1769 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1770 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1771 // This is not an aggressive instruction that we can promote. 1772 // Because of this, we won't be able to get rid of the control 1773 // flow, so the xform is not worth it. 1774 return false; 1775 } 1776 } 1777 1778 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1779 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1780 1781 // If we can still promote the PHI nodes after this gauntlet of tests, 1782 // do all of the PHI's now. 1783 Instruction *InsertPt = DomBlock->getTerminator(); 1784 IRBuilder<true, NoFolder> Builder(InsertPt); 1785 1786 // Move all 'aggressive' instructions, which are defined in the 1787 // conditional parts of the if's up to the dominating block. 1788 if (IfBlock1) 1789 DomBlock->getInstList().splice(InsertPt, 1790 IfBlock1->getInstList(), IfBlock1->begin(), 1791 IfBlock1->getTerminator()); 1792 if (IfBlock2) 1793 DomBlock->getInstList().splice(InsertPt, 1794 IfBlock2->getInstList(), IfBlock2->begin(), 1795 IfBlock2->getTerminator()); 1796 1797 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1798 // Change the PHI node into a select instruction. 1799 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1800 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1801 1802 SelectInst *NV = 1803 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1804 PN->replaceAllUsesWith(NV); 1805 NV->takeName(PN); 1806 PN->eraseFromParent(); 1807 } 1808 1809 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1810 // has been flattened. Change DomBlock to jump directly to our new block to 1811 // avoid other simplifycfg's kicking in on the diamond. 1812 TerminatorInst *OldTI = DomBlock->getTerminator(); 1813 Builder.SetInsertPoint(OldTI); 1814 Builder.CreateBr(BB); 1815 OldTI->eraseFromParent(); 1816 return true; 1817} 1818 1819/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1820/// to two returning blocks, try to merge them together into one return, 1821/// introducing a select if the return values disagree. 1822static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1823 IRBuilder<> &Builder) { 1824 assert(BI->isConditional() && "Must be a conditional branch"); 1825 BasicBlock *TrueSucc = BI->getSuccessor(0); 1826 BasicBlock *FalseSucc = BI->getSuccessor(1); 1827 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1828 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1829 1830 // Check to ensure both blocks are empty (just a return) or optionally empty 1831 // with PHI nodes. If there are other instructions, merging would cause extra 1832 // computation on one path or the other. 1833 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1834 return false; 1835 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1836 return false; 1837 1838 Builder.SetInsertPoint(BI); 1839 // Okay, we found a branch that is going to two return nodes. If 1840 // there is no return value for this function, just change the 1841 // branch into a return. 1842 if (FalseRet->getNumOperands() == 0) { 1843 TrueSucc->removePredecessor(BI->getParent()); 1844 FalseSucc->removePredecessor(BI->getParent()); 1845 Builder.CreateRetVoid(); 1846 EraseTerminatorInstAndDCECond(BI); 1847 return true; 1848 } 1849 1850 // Otherwise, figure out what the true and false return values are 1851 // so we can insert a new select instruction. 1852 Value *TrueValue = TrueRet->getReturnValue(); 1853 Value *FalseValue = FalseRet->getReturnValue(); 1854 1855 // Unwrap any PHI nodes in the return blocks. 1856 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1857 if (TVPN->getParent() == TrueSucc) 1858 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1859 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1860 if (FVPN->getParent() == FalseSucc) 1861 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1862 1863 // In order for this transformation to be safe, we must be able to 1864 // unconditionally execute both operands to the return. This is 1865 // normally the case, but we could have a potentially-trapping 1866 // constant expression that prevents this transformation from being 1867 // safe. 1868 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1869 if (TCV->canTrap()) 1870 return false; 1871 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1872 if (FCV->canTrap()) 1873 return false; 1874 1875 // Okay, we collected all the mapped values and checked them for sanity, and 1876 // defined to really do this transformation. First, update the CFG. 1877 TrueSucc->removePredecessor(BI->getParent()); 1878 FalseSucc->removePredecessor(BI->getParent()); 1879 1880 // Insert select instructions where needed. 1881 Value *BrCond = BI->getCondition(); 1882 if (TrueValue) { 1883 // Insert a select if the results differ. 1884 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1885 } else if (isa<UndefValue>(TrueValue)) { 1886 TrueValue = FalseValue; 1887 } else { 1888 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1889 FalseValue, "retval"); 1890 } 1891 } 1892 1893 Value *RI = !TrueValue ? 1894 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1895 1896 (void) RI; 1897 1898 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1899 << "\n " << *BI << "NewRet = " << *RI 1900 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1901 1902 EraseTerminatorInstAndDCECond(BI); 1903 1904 return true; 1905} 1906 1907/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1908/// probabilities of the branch taking each edge. Fills in the two APInt 1909/// parameters and return true, or returns false if no or invalid metadata was 1910/// found. 1911static bool ExtractBranchMetadata(BranchInst *BI, 1912 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1913 assert(BI->isConditional() && 1914 "Looking for probabilities on unconditional branch?"); 1915 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1916 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1917 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1918 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1919 if (!CITrue || !CIFalse) return false; 1920 ProbTrue = CITrue->getValue().getZExtValue(); 1921 ProbFalse = CIFalse->getValue().getZExtValue(); 1922 return true; 1923} 1924 1925/// checkCSEInPredecessor - Return true if the given instruction is available 1926/// in its predecessor block. If yes, the instruction will be removed. 1927/// 1928static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1929 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1930 return false; 1931 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1932 Instruction *PBI = &*I; 1933 // Check whether Inst and PBI generate the same value. 1934 if (Inst->isIdenticalTo(PBI)) { 1935 Inst->replaceAllUsesWith(PBI); 1936 Inst->eraseFromParent(); 1937 return true; 1938 } 1939 } 1940 return false; 1941} 1942 1943/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1944/// predecessor branches to us and one of our successors, fold the block into 1945/// the predecessor and use logical operations to pick the right destination. 1946bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1947 BasicBlock *BB = BI->getParent(); 1948 1949 Instruction *Cond = 0; 1950 if (BI->isConditional()) 1951 Cond = dyn_cast<Instruction>(BI->getCondition()); 1952 else { 1953 // For unconditional branch, check for a simple CFG pattern, where 1954 // BB has a single predecessor and BB's successor is also its predecessor's 1955 // successor. If such pattern exisits, check for CSE between BB and its 1956 // predecessor. 1957 if (BasicBlock *PB = BB->getSinglePredecessor()) 1958 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1959 if (PBI->isConditional() && 1960 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1961 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1962 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1963 I != E; ) { 1964 Instruction *Curr = I++; 1965 if (isa<CmpInst>(Curr)) { 1966 Cond = Curr; 1967 break; 1968 } 1969 // Quit if we can't remove this instruction. 1970 if (!checkCSEInPredecessor(Curr, PB)) 1971 return false; 1972 } 1973 } 1974 1975 if (Cond == 0) 1976 return false; 1977 } 1978 1979 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1980 Cond->getParent() != BB || !Cond->hasOneUse()) 1981 return false; 1982 1983 // Only allow this if the condition is a simple instruction that can be 1984 // executed unconditionally. It must be in the same block as the branch, and 1985 // must be at the front of the block. 1986 BasicBlock::iterator FrontIt = BB->front(); 1987 1988 // Ignore dbg intrinsics. 1989 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1990 1991 // Allow a single instruction to be hoisted in addition to the compare 1992 // that feeds the branch. We later ensure that any values that _it_ uses 1993 // were also live in the predecessor, so that we don't unnecessarily create 1994 // register pressure or inhibit out-of-order execution. 1995 Instruction *BonusInst = 0; 1996 if (&*FrontIt != Cond && 1997 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1998 isSafeToSpeculativelyExecute(FrontIt)) { 1999 BonusInst = &*FrontIt; 2000 ++FrontIt; 2001 2002 // Ignore dbg intrinsics. 2003 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2004 } 2005 2006 // Only a single bonus inst is allowed. 2007 if (&*FrontIt != Cond) 2008 return false; 2009 2010 // Make sure the instruction after the condition is the cond branch. 2011 BasicBlock::iterator CondIt = Cond; ++CondIt; 2012 2013 // Ingore dbg intrinsics. 2014 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 2015 2016 if (&*CondIt != BI) 2017 return false; 2018 2019 // Cond is known to be a compare or binary operator. Check to make sure that 2020 // neither operand is a potentially-trapping constant expression. 2021 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 2022 if (CE->canTrap()) 2023 return false; 2024 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 2025 if (CE->canTrap()) 2026 return false; 2027 2028 // Finally, don't infinitely unroll conditional loops. 2029 BasicBlock *TrueDest = BI->getSuccessor(0); 2030 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 2031 if (TrueDest == BB || FalseDest == BB) 2032 return false; 2033 2034 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2035 BasicBlock *PredBlock = *PI; 2036 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 2037 2038 // Check that we have two conditional branches. If there is a PHI node in 2039 // the common successor, verify that the same value flows in from both 2040 // blocks. 2041 SmallVector<PHINode*, 4> PHIs; 2042 if (PBI == 0 || PBI->isUnconditional() || 2043 (BI->isConditional() && 2044 !SafeToMergeTerminators(BI, PBI)) || 2045 (!BI->isConditional() && 2046 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2047 continue; 2048 2049 // Determine if the two branches share a common destination. 2050 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2051 bool InvertPredCond = false; 2052 2053 if (BI->isConditional()) { 2054 if (PBI->getSuccessor(0) == TrueDest) 2055 Opc = Instruction::Or; 2056 else if (PBI->getSuccessor(1) == FalseDest) 2057 Opc = Instruction::And; 2058 else if (PBI->getSuccessor(0) == FalseDest) 2059 Opc = Instruction::And, InvertPredCond = true; 2060 else if (PBI->getSuccessor(1) == TrueDest) 2061 Opc = Instruction::Or, InvertPredCond = true; 2062 else 2063 continue; 2064 } else { 2065 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2066 continue; 2067 } 2068 2069 // Ensure that any values used in the bonus instruction are also used 2070 // by the terminator of the predecessor. This means that those values 2071 // must already have been resolved, so we won't be inhibiting the 2072 // out-of-order core by speculating them earlier. 2073 if (BonusInst) { 2074 // Collect the values used by the bonus inst 2075 SmallPtrSet<Value*, 4> UsedValues; 2076 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2077 OE = BonusInst->op_end(); OI != OE; ++OI) { 2078 Value *V = *OI; 2079 if (!isa<Constant>(V)) 2080 UsedValues.insert(V); 2081 } 2082 2083 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2084 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2085 2086 // Walk up to four levels back up the use-def chain of the predecessor's 2087 // terminator to see if all those values were used. The choice of four 2088 // levels is arbitrary, to provide a compile-time-cost bound. 2089 while (!Worklist.empty()) { 2090 std::pair<Value*, unsigned> Pair = Worklist.back(); 2091 Worklist.pop_back(); 2092 2093 if (Pair.second >= 4) continue; 2094 UsedValues.erase(Pair.first); 2095 if (UsedValues.empty()) break; 2096 2097 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2098 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2099 OI != OE; ++OI) 2100 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2101 } 2102 } 2103 2104 if (!UsedValues.empty()) return false; 2105 } 2106 2107 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2108 IRBuilder<> Builder(PBI); 2109 2110 // If we need to invert the condition in the pred block to match, do so now. 2111 if (InvertPredCond) { 2112 Value *NewCond = PBI->getCondition(); 2113 2114 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2115 CmpInst *CI = cast<CmpInst>(NewCond); 2116 CI->setPredicate(CI->getInversePredicate()); 2117 } else { 2118 NewCond = Builder.CreateNot(NewCond, 2119 PBI->getCondition()->getName()+".not"); 2120 } 2121 2122 PBI->setCondition(NewCond); 2123 PBI->swapSuccessors(); 2124 } 2125 2126 // If we have a bonus inst, clone it into the predecessor block. 2127 Instruction *NewBonus = 0; 2128 if (BonusInst) { 2129 NewBonus = BonusInst->clone(); 2130 PredBlock->getInstList().insert(PBI, NewBonus); 2131 NewBonus->takeName(BonusInst); 2132 BonusInst->setName(BonusInst->getName()+".old"); 2133 } 2134 2135 // Clone Cond into the predecessor basic block, and or/and the 2136 // two conditions together. 2137 Instruction *New = Cond->clone(); 2138 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2139 PredBlock->getInstList().insert(PBI, New); 2140 New->takeName(Cond); 2141 Cond->setName(New->getName()+".old"); 2142 2143 if (BI->isConditional()) { 2144 Instruction *NewCond = 2145 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2146 New, "or.cond")); 2147 PBI->setCondition(NewCond); 2148 2149 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2150 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2151 PredFalseWeight); 2152 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2153 SuccFalseWeight); 2154 SmallVector<uint64_t, 8> NewWeights; 2155 2156 if (PBI->getSuccessor(0) == BB) { 2157 if (PredHasWeights && SuccHasWeights) { 2158 // PBI: br i1 %x, BB, FalseDest 2159 // BI: br i1 %y, TrueDest, FalseDest 2160 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2161 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2162 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2163 // TrueWeight for PBI * FalseWeight for BI. 2164 // We assume that total weights of a BranchInst can fit into 32 bits. 2165 // Therefore, we will not have overflow using 64-bit arithmetic. 2166 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2167 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2168 } 2169 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2170 PBI->setSuccessor(0, TrueDest); 2171 } 2172 if (PBI->getSuccessor(1) == BB) { 2173 if (PredHasWeights && SuccHasWeights) { 2174 // PBI: br i1 %x, TrueDest, BB 2175 // BI: br i1 %y, TrueDest, FalseDest 2176 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2177 // FalseWeight for PBI * TrueWeight for BI. 2178 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2179 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2180 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2181 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2182 } 2183 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2184 PBI->setSuccessor(1, FalseDest); 2185 } 2186 if (NewWeights.size() == 2) { 2187 // Halve the weights if any of them cannot fit in an uint32_t 2188 FitWeights(NewWeights); 2189 2190 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2191 PBI->setMetadata(LLVMContext::MD_prof, 2192 MDBuilder(BI->getContext()). 2193 createBranchWeights(MDWeights)); 2194 } else 2195 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2196 } else { 2197 // Update PHI nodes in the common successors. 2198 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2199 ConstantInt *PBI_C = cast<ConstantInt>( 2200 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2201 assert(PBI_C->getType()->isIntegerTy(1)); 2202 Instruction *MergedCond = 0; 2203 if (PBI->getSuccessor(0) == TrueDest) { 2204 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2205 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2206 // is false: !PBI_Cond and BI_Value 2207 Instruction *NotCond = 2208 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2209 "not.cond")); 2210 MergedCond = 2211 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2212 NotCond, New, 2213 "and.cond")); 2214 if (PBI_C->isOne()) 2215 MergedCond = 2216 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2217 PBI->getCondition(), MergedCond, 2218 "or.cond")); 2219 } else { 2220 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2221 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2222 // is false: PBI_Cond and BI_Value 2223 MergedCond = 2224 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2225 PBI->getCondition(), New, 2226 "and.cond")); 2227 if (PBI_C->isOne()) { 2228 Instruction *NotCond = 2229 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2230 "not.cond")); 2231 MergedCond = 2232 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2233 NotCond, MergedCond, 2234 "or.cond")); 2235 } 2236 } 2237 // Update PHI Node. 2238 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2239 MergedCond); 2240 } 2241 // Change PBI from Conditional to Unconditional. 2242 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2243 EraseTerminatorInstAndDCECond(PBI); 2244 PBI = New_PBI; 2245 } 2246 2247 // TODO: If BB is reachable from all paths through PredBlock, then we 2248 // could replace PBI's branch probabilities with BI's. 2249 2250 // Copy any debug value intrinsics into the end of PredBlock. 2251 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2252 if (isa<DbgInfoIntrinsic>(*I)) 2253 I->clone()->insertBefore(PBI); 2254 2255 return true; 2256 } 2257 return false; 2258} 2259 2260/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2261/// predecessor of another block, this function tries to simplify it. We know 2262/// that PBI and BI are both conditional branches, and BI is in one of the 2263/// successor blocks of PBI - PBI branches to BI. 2264static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2265 assert(PBI->isConditional() && BI->isConditional()); 2266 BasicBlock *BB = BI->getParent(); 2267 2268 // If this block ends with a branch instruction, and if there is a 2269 // predecessor that ends on a branch of the same condition, make 2270 // this conditional branch redundant. 2271 if (PBI->getCondition() == BI->getCondition() && 2272 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2273 // Okay, the outcome of this conditional branch is statically 2274 // knowable. If this block had a single pred, handle specially. 2275 if (BB->getSinglePredecessor()) { 2276 // Turn this into a branch on constant. 2277 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2278 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2279 CondIsTrue)); 2280 return true; // Nuke the branch on constant. 2281 } 2282 2283 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2284 // in the constant and simplify the block result. Subsequent passes of 2285 // simplifycfg will thread the block. 2286 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2287 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2288 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2289 std::distance(PB, PE), 2290 BI->getCondition()->getName() + ".pr", 2291 BB->begin()); 2292 // Okay, we're going to insert the PHI node. Since PBI is not the only 2293 // predecessor, compute the PHI'd conditional value for all of the preds. 2294 // Any predecessor where the condition is not computable we keep symbolic. 2295 for (pred_iterator PI = PB; PI != PE; ++PI) { 2296 BasicBlock *P = *PI; 2297 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2298 PBI != BI && PBI->isConditional() && 2299 PBI->getCondition() == BI->getCondition() && 2300 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2301 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2302 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2303 CondIsTrue), P); 2304 } else { 2305 NewPN->addIncoming(BI->getCondition(), P); 2306 } 2307 } 2308 2309 BI->setCondition(NewPN); 2310 return true; 2311 } 2312 } 2313 2314 // If this is a conditional branch in an empty block, and if any 2315 // predecessors is a conditional branch to one of our destinations, 2316 // fold the conditions into logical ops and one cond br. 2317 BasicBlock::iterator BBI = BB->begin(); 2318 // Ignore dbg intrinsics. 2319 while (isa<DbgInfoIntrinsic>(BBI)) 2320 ++BBI; 2321 if (&*BBI != BI) 2322 return false; 2323 2324 2325 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2326 if (CE->canTrap()) 2327 return false; 2328 2329 int PBIOp, BIOp; 2330 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2331 PBIOp = BIOp = 0; 2332 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2333 PBIOp = 0, BIOp = 1; 2334 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2335 PBIOp = 1, BIOp = 0; 2336 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2337 PBIOp = BIOp = 1; 2338 else 2339 return false; 2340 2341 // Check to make sure that the other destination of this branch 2342 // isn't BB itself. If so, this is an infinite loop that will 2343 // keep getting unwound. 2344 if (PBI->getSuccessor(PBIOp) == BB) 2345 return false; 2346 2347 // Do not perform this transformation if it would require 2348 // insertion of a large number of select instructions. For targets 2349 // without predication/cmovs, this is a big pessimization. 2350 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2351 2352 unsigned NumPhis = 0; 2353 for (BasicBlock::iterator II = CommonDest->begin(); 2354 isa<PHINode>(II); ++II, ++NumPhis) 2355 if (NumPhis > 2) // Disable this xform. 2356 return false; 2357 2358 // Finally, if everything is ok, fold the branches to logical ops. 2359 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2360 2361 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2362 << "AND: " << *BI->getParent()); 2363 2364 2365 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2366 // branch in it, where one edge (OtherDest) goes back to itself but the other 2367 // exits. We don't *know* that the program avoids the infinite loop 2368 // (even though that seems likely). If we do this xform naively, we'll end up 2369 // recursively unpeeling the loop. Since we know that (after the xform is 2370 // done) that the block *is* infinite if reached, we just make it an obviously 2371 // infinite loop with no cond branch. 2372 if (OtherDest == BB) { 2373 // Insert it at the end of the function, because it's either code, 2374 // or it won't matter if it's hot. :) 2375 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2376 "infloop", BB->getParent()); 2377 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2378 OtherDest = InfLoopBlock; 2379 } 2380 2381 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2382 2383 // BI may have other predecessors. Because of this, we leave 2384 // it alone, but modify PBI. 2385 2386 // Make sure we get to CommonDest on True&True directions. 2387 Value *PBICond = PBI->getCondition(); 2388 IRBuilder<true, NoFolder> Builder(PBI); 2389 if (PBIOp) 2390 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2391 2392 Value *BICond = BI->getCondition(); 2393 if (BIOp) 2394 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2395 2396 // Merge the conditions. 2397 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2398 2399 // Modify PBI to branch on the new condition to the new dests. 2400 PBI->setCondition(Cond); 2401 PBI->setSuccessor(0, CommonDest); 2402 PBI->setSuccessor(1, OtherDest); 2403 2404 // Update branch weight for PBI. 2405 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2406 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2407 PredFalseWeight); 2408 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2409 SuccFalseWeight); 2410 if (PredHasWeights && SuccHasWeights) { 2411 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2412 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2413 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2414 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2415 // The weight to CommonDest should be PredCommon * SuccTotal + 2416 // PredOther * SuccCommon. 2417 // The weight to OtherDest should be PredOther * SuccOther. 2418 SmallVector<uint64_t, 2> NewWeights; 2419 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2420 PredOther * SuccCommon); 2421 NewWeights.push_back(PredOther * SuccOther); 2422 // Halve the weights if any of them cannot fit in an uint32_t 2423 FitWeights(NewWeights); 2424 2425 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2426 PBI->setMetadata(LLVMContext::MD_prof, 2427 MDBuilder(BI->getContext()). 2428 createBranchWeights(MDWeights)); 2429 } 2430 2431 // OtherDest may have phi nodes. If so, add an entry from PBI's 2432 // block that are identical to the entries for BI's block. 2433 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2434 2435 // We know that the CommonDest already had an edge from PBI to 2436 // it. If it has PHIs though, the PHIs may have different 2437 // entries for BB and PBI's BB. If so, insert a select to make 2438 // them agree. 2439 PHINode *PN; 2440 for (BasicBlock::iterator II = CommonDest->begin(); 2441 (PN = dyn_cast<PHINode>(II)); ++II) { 2442 Value *BIV = PN->getIncomingValueForBlock(BB); 2443 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2444 Value *PBIV = PN->getIncomingValue(PBBIdx); 2445 if (BIV != PBIV) { 2446 // Insert a select in PBI to pick the right value. 2447 Value *NV = cast<SelectInst> 2448 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2449 PN->setIncomingValue(PBBIdx, NV); 2450 } 2451 } 2452 2453 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2454 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2455 2456 // This basic block is probably dead. We know it has at least 2457 // one fewer predecessor. 2458 return true; 2459} 2460 2461// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2462// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2463// Takes care of updating the successors and removing the old terminator. 2464// Also makes sure not to introduce new successors by assuming that edges to 2465// non-successor TrueBBs and FalseBBs aren't reachable. 2466static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2467 BasicBlock *TrueBB, BasicBlock *FalseBB, 2468 uint32_t TrueWeight, 2469 uint32_t FalseWeight){ 2470 // Remove any superfluous successor edges from the CFG. 2471 // First, figure out which successors to preserve. 2472 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2473 // successor. 2474 BasicBlock *KeepEdge1 = TrueBB; 2475 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2476 2477 // Then remove the rest. 2478 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2479 BasicBlock *Succ = OldTerm->getSuccessor(I); 2480 // Make sure only to keep exactly one copy of each edge. 2481 if (Succ == KeepEdge1) 2482 KeepEdge1 = 0; 2483 else if (Succ == KeepEdge2) 2484 KeepEdge2 = 0; 2485 else 2486 Succ->removePredecessor(OldTerm->getParent()); 2487 } 2488 2489 IRBuilder<> Builder(OldTerm); 2490 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2491 2492 // Insert an appropriate new terminator. 2493 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2494 if (TrueBB == FalseBB) 2495 // We were only looking for one successor, and it was present. 2496 // Create an unconditional branch to it. 2497 Builder.CreateBr(TrueBB); 2498 else { 2499 // We found both of the successors we were looking for. 2500 // Create a conditional branch sharing the condition of the select. 2501 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2502 if (TrueWeight != FalseWeight) 2503 NewBI->setMetadata(LLVMContext::MD_prof, 2504 MDBuilder(OldTerm->getContext()). 2505 createBranchWeights(TrueWeight, FalseWeight)); 2506 } 2507 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2508 // Neither of the selected blocks were successors, so this 2509 // terminator must be unreachable. 2510 new UnreachableInst(OldTerm->getContext(), OldTerm); 2511 } else { 2512 // One of the selected values was a successor, but the other wasn't. 2513 // Insert an unconditional branch to the one that was found; 2514 // the edge to the one that wasn't must be unreachable. 2515 if (KeepEdge1 == 0) 2516 // Only TrueBB was found. 2517 Builder.CreateBr(TrueBB); 2518 else 2519 // Only FalseBB was found. 2520 Builder.CreateBr(FalseBB); 2521 } 2522 2523 EraseTerminatorInstAndDCECond(OldTerm); 2524 return true; 2525} 2526 2527// SimplifySwitchOnSelect - Replaces 2528// (switch (select cond, X, Y)) on constant X, Y 2529// with a branch - conditional if X and Y lead to distinct BBs, 2530// unconditional otherwise. 2531static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2532 // Check for constant integer values in the select. 2533 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2534 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2535 if (!TrueVal || !FalseVal) 2536 return false; 2537 2538 // Find the relevant condition and destinations. 2539 Value *Condition = Select->getCondition(); 2540 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2541 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2542 2543 // Get weight for TrueBB and FalseBB. 2544 uint32_t TrueWeight = 0, FalseWeight = 0; 2545 SmallVector<uint64_t, 8> Weights; 2546 bool HasWeights = HasBranchWeights(SI); 2547 if (HasWeights) { 2548 GetBranchWeights(SI, Weights); 2549 if (Weights.size() == 1 + SI->getNumCases()) { 2550 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2551 getSuccessorIndex()]; 2552 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2553 getSuccessorIndex()]; 2554 } 2555 } 2556 2557 // Perform the actual simplification. 2558 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2559 TrueWeight, FalseWeight); 2560} 2561 2562// SimplifyIndirectBrOnSelect - Replaces 2563// (indirectbr (select cond, blockaddress(@fn, BlockA), 2564// blockaddress(@fn, BlockB))) 2565// with 2566// (br cond, BlockA, BlockB). 2567static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2568 // Check that both operands of the select are block addresses. 2569 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2570 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2571 if (!TBA || !FBA) 2572 return false; 2573 2574 // Extract the actual blocks. 2575 BasicBlock *TrueBB = TBA->getBasicBlock(); 2576 BasicBlock *FalseBB = FBA->getBasicBlock(); 2577 2578 // Perform the actual simplification. 2579 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2580 0, 0); 2581} 2582 2583/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2584/// instruction (a seteq/setne with a constant) as the only instruction in a 2585/// block that ends with an uncond branch. We are looking for a very specific 2586/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2587/// this case, we merge the first two "or's of icmp" into a switch, but then the 2588/// default value goes to an uncond block with a seteq in it, we get something 2589/// like: 2590/// 2591/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2592/// DEFAULT: 2593/// %tmp = icmp eq i8 %A, 92 2594/// br label %end 2595/// end: 2596/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2597/// 2598/// We prefer to split the edge to 'end' so that there is a true/false entry to 2599/// the PHI, merging the third icmp into the switch. 2600static bool TryToSimplifyUncondBranchWithICmpInIt( 2601 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2602 const DataLayout *TD) { 2603 BasicBlock *BB = ICI->getParent(); 2604 2605 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2606 // complex. 2607 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2608 2609 Value *V = ICI->getOperand(0); 2610 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2611 2612 // The pattern we're looking for is where our only predecessor is a switch on 2613 // 'V' and this block is the default case for the switch. In this case we can 2614 // fold the compared value into the switch to simplify things. 2615 BasicBlock *Pred = BB->getSinglePredecessor(); 2616 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2617 2618 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2619 if (SI->getCondition() != V) 2620 return false; 2621 2622 // If BB is reachable on a non-default case, then we simply know the value of 2623 // V in this block. Substitute it and constant fold the icmp instruction 2624 // away. 2625 if (SI->getDefaultDest() != BB) { 2626 ConstantInt *VVal = SI->findCaseDest(BB); 2627 assert(VVal && "Should have a unique destination value"); 2628 ICI->setOperand(0, VVal); 2629 2630 if (Value *V = SimplifyInstruction(ICI, TD)) { 2631 ICI->replaceAllUsesWith(V); 2632 ICI->eraseFromParent(); 2633 } 2634 // BB is now empty, so it is likely to simplify away. 2635 return SimplifyCFG(BB, TTI, TD) | true; 2636 } 2637 2638 // Ok, the block is reachable from the default dest. If the constant we're 2639 // comparing exists in one of the other edges, then we can constant fold ICI 2640 // and zap it. 2641 if (SI->findCaseValue(Cst) != SI->case_default()) { 2642 Value *V; 2643 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2644 V = ConstantInt::getFalse(BB->getContext()); 2645 else 2646 V = ConstantInt::getTrue(BB->getContext()); 2647 2648 ICI->replaceAllUsesWith(V); 2649 ICI->eraseFromParent(); 2650 // BB is now empty, so it is likely to simplify away. 2651 return SimplifyCFG(BB, TTI, TD) | true; 2652 } 2653 2654 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2655 // the block. 2656 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2657 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2658 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2659 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2660 return false; 2661 2662 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2663 // true in the PHI. 2664 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2665 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2666 2667 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2668 std::swap(DefaultCst, NewCst); 2669 2670 // Replace ICI (which is used by the PHI for the default value) with true or 2671 // false depending on if it is EQ or NE. 2672 ICI->replaceAllUsesWith(DefaultCst); 2673 ICI->eraseFromParent(); 2674 2675 // Okay, the switch goes to this block on a default value. Add an edge from 2676 // the switch to the merge point on the compared value. 2677 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2678 BB->getParent(), BB); 2679 SmallVector<uint64_t, 8> Weights; 2680 bool HasWeights = HasBranchWeights(SI); 2681 if (HasWeights) { 2682 GetBranchWeights(SI, Weights); 2683 if (Weights.size() == 1 + SI->getNumCases()) { 2684 // Split weight for default case to case for "Cst". 2685 Weights[0] = (Weights[0]+1) >> 1; 2686 Weights.push_back(Weights[0]); 2687 2688 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2689 SI->setMetadata(LLVMContext::MD_prof, 2690 MDBuilder(SI->getContext()). 2691 createBranchWeights(MDWeights)); 2692 } 2693 } 2694 SI->addCase(Cst, NewBB); 2695 2696 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2697 Builder.SetInsertPoint(NewBB); 2698 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2699 Builder.CreateBr(SuccBlock); 2700 PHIUse->addIncoming(NewCst, NewBB); 2701 return true; 2702} 2703 2704/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2705/// Check to see if it is branching on an or/and chain of icmp instructions, and 2706/// fold it into a switch instruction if so. 2707static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2708 IRBuilder<> &Builder) { 2709 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2710 if (Cond == 0) return false; 2711 2712 2713 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2714 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2715 // 'setne's and'ed together, collect them. 2716 Value *CompVal = 0; 2717 std::vector<ConstantInt*> Values; 2718 bool TrueWhenEqual = true; 2719 Value *ExtraCase = 0; 2720 unsigned UsedICmps = 0; 2721 2722 if (Cond->getOpcode() == Instruction::Or) { 2723 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2724 UsedICmps); 2725 } else if (Cond->getOpcode() == Instruction::And) { 2726 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2727 UsedICmps); 2728 TrueWhenEqual = false; 2729 } 2730 2731 // If we didn't have a multiply compared value, fail. 2732 if (CompVal == 0) return false; 2733 2734 // Avoid turning single icmps into a switch. 2735 if (UsedICmps <= 1) 2736 return false; 2737 2738 // There might be duplicate constants in the list, which the switch 2739 // instruction can't handle, remove them now. 2740 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2741 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2742 2743 // If Extra was used, we require at least two switch values to do the 2744 // transformation. A switch with one value is just an cond branch. 2745 if (ExtraCase && Values.size() < 2) return false; 2746 2747 // TODO: Preserve branch weight metadata, similarly to how 2748 // FoldValueComparisonIntoPredecessors preserves it. 2749 2750 // Figure out which block is which destination. 2751 BasicBlock *DefaultBB = BI->getSuccessor(1); 2752 BasicBlock *EdgeBB = BI->getSuccessor(0); 2753 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2754 2755 BasicBlock *BB = BI->getParent(); 2756 2757 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2758 << " cases into SWITCH. BB is:\n" << *BB); 2759 2760 // If there are any extra values that couldn't be folded into the switch 2761 // then we evaluate them with an explicit branch first. Split the block 2762 // right before the condbr to handle it. 2763 if (ExtraCase) { 2764 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2765 // Remove the uncond branch added to the old block. 2766 TerminatorInst *OldTI = BB->getTerminator(); 2767 Builder.SetInsertPoint(OldTI); 2768 2769 if (TrueWhenEqual) 2770 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2771 else 2772 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2773 2774 OldTI->eraseFromParent(); 2775 2776 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2777 // for the edge we just added. 2778 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2779 2780 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2781 << "\nEXTRABB = " << *BB); 2782 BB = NewBB; 2783 } 2784 2785 Builder.SetInsertPoint(BI); 2786 // Convert pointer to int before we switch. 2787 if (CompVal->getType()->isPointerTy()) { 2788 assert(TD && "Cannot switch on pointer without DataLayout"); 2789 CompVal = Builder.CreatePtrToInt(CompVal, 2790 TD->getIntPtrType(CompVal->getContext()), 2791 "magicptr"); 2792 } 2793 2794 // Create the new switch instruction now. 2795 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2796 2797 // Add all of the 'cases' to the switch instruction. 2798 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2799 New->addCase(Values[i], EdgeBB); 2800 2801 // We added edges from PI to the EdgeBB. As such, if there were any 2802 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2803 // the number of edges added. 2804 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2805 isa<PHINode>(BBI); ++BBI) { 2806 PHINode *PN = cast<PHINode>(BBI); 2807 Value *InVal = PN->getIncomingValueForBlock(BB); 2808 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2809 PN->addIncoming(InVal, BB); 2810 } 2811 2812 // Erase the old branch instruction. 2813 EraseTerminatorInstAndDCECond(BI); 2814 2815 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2816 return true; 2817} 2818 2819bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2820 // If this is a trivial landing pad that just continues unwinding the caught 2821 // exception then zap the landing pad, turning its invokes into calls. 2822 BasicBlock *BB = RI->getParent(); 2823 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2824 if (RI->getValue() != LPInst) 2825 // Not a landing pad, or the resume is not unwinding the exception that 2826 // caused control to branch here. 2827 return false; 2828 2829 // Check that there are no other instructions except for debug intrinsics. 2830 BasicBlock::iterator I = LPInst, E = RI; 2831 while (++I != E) 2832 if (!isa<DbgInfoIntrinsic>(I)) 2833 return false; 2834 2835 // Turn all invokes that unwind here into calls and delete the basic block. 2836 bool InvokeRequiresTableEntry = false; 2837 bool Changed = false; 2838 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2839 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2840 2841 if (II->hasFnAttr(Attribute::UWTable)) { 2842 // Don't remove an `invoke' instruction if the ABI requires an entry into 2843 // the table. 2844 InvokeRequiresTableEntry = true; 2845 continue; 2846 } 2847 2848 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2849 2850 // Insert a call instruction before the invoke. 2851 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2852 Call->takeName(II); 2853 Call->setCallingConv(II->getCallingConv()); 2854 Call->setAttributes(II->getAttributes()); 2855 Call->setDebugLoc(II->getDebugLoc()); 2856 2857 // Anything that used the value produced by the invoke instruction now uses 2858 // the value produced by the call instruction. Note that we do this even 2859 // for void functions and calls with no uses so that the callgraph edge is 2860 // updated. 2861 II->replaceAllUsesWith(Call); 2862 BB->removePredecessor(II->getParent()); 2863 2864 // Insert a branch to the normal destination right before the invoke. 2865 BranchInst::Create(II->getNormalDest(), II); 2866 2867 // Finally, delete the invoke instruction! 2868 II->eraseFromParent(); 2869 Changed = true; 2870 } 2871 2872 if (!InvokeRequiresTableEntry) 2873 // The landingpad is now unreachable. Zap it. 2874 BB->eraseFromParent(); 2875 2876 return Changed; 2877} 2878 2879bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2880 BasicBlock *BB = RI->getParent(); 2881 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2882 2883 // Find predecessors that end with branches. 2884 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2885 SmallVector<BranchInst*, 8> CondBranchPreds; 2886 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2887 BasicBlock *P = *PI; 2888 TerminatorInst *PTI = P->getTerminator(); 2889 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2890 if (BI->isUnconditional()) 2891 UncondBranchPreds.push_back(P); 2892 else 2893 CondBranchPreds.push_back(BI); 2894 } 2895 } 2896 2897 // If we found some, do the transformation! 2898 if (!UncondBranchPreds.empty() && DupRet) { 2899 while (!UncondBranchPreds.empty()) { 2900 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2901 DEBUG(dbgs() << "FOLDING: " << *BB 2902 << "INTO UNCOND BRANCH PRED: " << *Pred); 2903 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2904 } 2905 2906 // If we eliminated all predecessors of the block, delete the block now. 2907 if (pred_begin(BB) == pred_end(BB)) 2908 // We know there are no successors, so just nuke the block. 2909 BB->eraseFromParent(); 2910 2911 return true; 2912 } 2913 2914 // Check out all of the conditional branches going to this return 2915 // instruction. If any of them just select between returns, change the 2916 // branch itself into a select/return pair. 2917 while (!CondBranchPreds.empty()) { 2918 BranchInst *BI = CondBranchPreds.pop_back_val(); 2919 2920 // Check to see if the non-BB successor is also a return block. 2921 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2922 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2923 SimplifyCondBranchToTwoReturns(BI, Builder)) 2924 return true; 2925 } 2926 return false; 2927} 2928 2929bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2930 BasicBlock *BB = UI->getParent(); 2931 2932 bool Changed = false; 2933 2934 // If there are any instructions immediately before the unreachable that can 2935 // be removed, do so. 2936 while (UI != BB->begin()) { 2937 BasicBlock::iterator BBI = UI; 2938 --BBI; 2939 // Do not delete instructions that can have side effects which might cause 2940 // the unreachable to not be reachable; specifically, calls and volatile 2941 // operations may have this effect. 2942 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2943 2944 if (BBI->mayHaveSideEffects()) { 2945 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2946 if (SI->isVolatile()) 2947 break; 2948 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2949 if (LI->isVolatile()) 2950 break; 2951 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2952 if (RMWI->isVolatile()) 2953 break; 2954 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2955 if (CXI->isVolatile()) 2956 break; 2957 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2958 !isa<LandingPadInst>(BBI)) { 2959 break; 2960 } 2961 // Note that deleting LandingPad's here is in fact okay, although it 2962 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2963 // all the predecessors of this block will be the unwind edges of Invokes, 2964 // and we can therefore guarantee this block will be erased. 2965 } 2966 2967 // Delete this instruction (any uses are guaranteed to be dead) 2968 if (!BBI->use_empty()) 2969 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2970 BBI->eraseFromParent(); 2971 Changed = true; 2972 } 2973 2974 // If the unreachable instruction is the first in the block, take a gander 2975 // at all of the predecessors of this instruction, and simplify them. 2976 if (&BB->front() != UI) return Changed; 2977 2978 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2979 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2980 TerminatorInst *TI = Preds[i]->getTerminator(); 2981 IRBuilder<> Builder(TI); 2982 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2983 if (BI->isUnconditional()) { 2984 if (BI->getSuccessor(0) == BB) { 2985 new UnreachableInst(TI->getContext(), TI); 2986 TI->eraseFromParent(); 2987 Changed = true; 2988 } 2989 } else { 2990 if (BI->getSuccessor(0) == BB) { 2991 Builder.CreateBr(BI->getSuccessor(1)); 2992 EraseTerminatorInstAndDCECond(BI); 2993 } else if (BI->getSuccessor(1) == BB) { 2994 Builder.CreateBr(BI->getSuccessor(0)); 2995 EraseTerminatorInstAndDCECond(BI); 2996 Changed = true; 2997 } 2998 } 2999 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 3000 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3001 i != e; ++i) 3002 if (i.getCaseSuccessor() == BB) { 3003 BB->removePredecessor(SI->getParent()); 3004 SI->removeCase(i); 3005 --i; --e; 3006 Changed = true; 3007 } 3008 // If the default value is unreachable, figure out the most popular 3009 // destination and make it the default. 3010 if (SI->getDefaultDest() == BB) { 3011 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 3012 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3013 i != e; ++i) { 3014 std::pair<unsigned, unsigned> &entry = 3015 Popularity[i.getCaseSuccessor()]; 3016 if (entry.first == 0) { 3017 entry.first = 1; 3018 entry.second = i.getCaseIndex(); 3019 } else { 3020 entry.first++; 3021 } 3022 } 3023 3024 // Find the most popular block. 3025 unsigned MaxPop = 0; 3026 unsigned MaxIndex = 0; 3027 BasicBlock *MaxBlock = 0; 3028 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 3029 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 3030 if (I->second.first > MaxPop || 3031 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 3032 MaxPop = I->second.first; 3033 MaxIndex = I->second.second; 3034 MaxBlock = I->first; 3035 } 3036 } 3037 if (MaxBlock) { 3038 // Make this the new default, allowing us to delete any explicit 3039 // edges to it. 3040 SI->setDefaultDest(MaxBlock); 3041 Changed = true; 3042 3043 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3044 // it. 3045 if (isa<PHINode>(MaxBlock->begin())) 3046 for (unsigned i = 0; i != MaxPop-1; ++i) 3047 MaxBlock->removePredecessor(SI->getParent()); 3048 3049 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3050 i != e; ++i) 3051 if (i.getCaseSuccessor() == MaxBlock) { 3052 SI->removeCase(i); 3053 --i; --e; 3054 } 3055 } 3056 } 3057 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3058 if (II->getUnwindDest() == BB) { 3059 // Convert the invoke to a call instruction. This would be a good 3060 // place to note that the call does not throw though. 3061 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3062 II->removeFromParent(); // Take out of symbol table 3063 3064 // Insert the call now... 3065 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3066 Builder.SetInsertPoint(BI); 3067 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3068 Args, II->getName()); 3069 CI->setCallingConv(II->getCallingConv()); 3070 CI->setAttributes(II->getAttributes()); 3071 // If the invoke produced a value, the call does now instead. 3072 II->replaceAllUsesWith(CI); 3073 delete II; 3074 Changed = true; 3075 } 3076 } 3077 } 3078 3079 // If this block is now dead, remove it. 3080 if (pred_begin(BB) == pred_end(BB) && 3081 BB != &BB->getParent()->getEntryBlock()) { 3082 // We know there are no successors, so just nuke the block. 3083 BB->eraseFromParent(); 3084 return true; 3085 } 3086 3087 return Changed; 3088} 3089 3090/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3091/// integer range comparison into a sub, an icmp and a branch. 3092static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3093 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3094 3095 // Make sure all cases point to the same destination and gather the values. 3096 SmallVector<ConstantInt *, 16> Cases; 3097 SwitchInst::CaseIt I = SI->case_begin(); 3098 Cases.push_back(I.getCaseValue()); 3099 SwitchInst::CaseIt PrevI = I++; 3100 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3101 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3102 return false; 3103 Cases.push_back(I.getCaseValue()); 3104 } 3105 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3106 3107 // Sort the case values, then check if they form a range we can transform. 3108 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3109 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3110 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3111 return false; 3112 } 3113 3114 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3115 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3116 3117 Value *Sub = SI->getCondition(); 3118 if (!Offset->isNullValue()) 3119 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3120 Value *Cmp; 3121 // If NumCases overflowed, then all possible values jump to the successor. 3122 if (NumCases->isNullValue() && SI->getNumCases() != 0) 3123 Cmp = ConstantInt::getTrue(SI->getContext()); 3124 else 3125 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3126 BranchInst *NewBI = Builder.CreateCondBr( 3127 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3128 3129 // Update weight for the newly-created conditional branch. 3130 SmallVector<uint64_t, 8> Weights; 3131 bool HasWeights = HasBranchWeights(SI); 3132 if (HasWeights) { 3133 GetBranchWeights(SI, Weights); 3134 if (Weights.size() == 1 + SI->getNumCases()) { 3135 // Combine all weights for the cases to be the true weight of NewBI. 3136 // We assume that the sum of all weights for a Terminator can fit into 32 3137 // bits. 3138 uint32_t NewTrueWeight = 0; 3139 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3140 NewTrueWeight += (uint32_t)Weights[I]; 3141 NewBI->setMetadata(LLVMContext::MD_prof, 3142 MDBuilder(SI->getContext()). 3143 createBranchWeights(NewTrueWeight, 3144 (uint32_t)Weights[0])); 3145 } 3146 } 3147 3148 // Prune obsolete incoming values off the successor's PHI nodes. 3149 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3150 isa<PHINode>(BBI); ++BBI) { 3151 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3152 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3153 } 3154 SI->eraseFromParent(); 3155 3156 return true; 3157} 3158 3159/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3160/// and use it to remove dead cases. 3161static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3162 Value *Cond = SI->getCondition(); 3163 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3164 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3165 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3166 3167 // Gather dead cases. 3168 SmallVector<ConstantInt*, 8> DeadCases; 3169 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3170 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3171 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3172 DeadCases.push_back(I.getCaseValue()); 3173 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3174 << I.getCaseValue() << "' is dead.\n"); 3175 } 3176 } 3177 3178 SmallVector<uint64_t, 8> Weights; 3179 bool HasWeight = HasBranchWeights(SI); 3180 if (HasWeight) { 3181 GetBranchWeights(SI, Weights); 3182 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3183 } 3184 3185 // Remove dead cases from the switch. 3186 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3187 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3188 assert(Case != SI->case_default() && 3189 "Case was not found. Probably mistake in DeadCases forming."); 3190 if (HasWeight) { 3191 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3192 Weights.pop_back(); 3193 } 3194 3195 // Prune unused values from PHI nodes. 3196 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3197 SI->removeCase(Case); 3198 } 3199 if (HasWeight) { 3200 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3201 SI->setMetadata(LLVMContext::MD_prof, 3202 MDBuilder(SI->getParent()->getContext()). 3203 createBranchWeights(MDWeights)); 3204 } 3205 3206 return !DeadCases.empty(); 3207} 3208 3209/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3210/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3211/// by an unconditional branch), look at the phi node for BB in the successor 3212/// block and see if the incoming value is equal to CaseValue. If so, return 3213/// the phi node, and set PhiIndex to BB's index in the phi node. 3214static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3215 BasicBlock *BB, 3216 int *PhiIndex) { 3217 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3218 return NULL; // BB must be empty to be a candidate for simplification. 3219 if (!BB->getSinglePredecessor()) 3220 return NULL; // BB must be dominated by the switch. 3221 3222 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3223 if (!Branch || !Branch->isUnconditional()) 3224 return NULL; // Terminator must be unconditional branch. 3225 3226 BasicBlock *Succ = Branch->getSuccessor(0); 3227 3228 BasicBlock::iterator I = Succ->begin(); 3229 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3230 int Idx = PHI->getBasicBlockIndex(BB); 3231 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3232 3233 Value *InValue = PHI->getIncomingValue(Idx); 3234 if (InValue != CaseValue) continue; 3235 3236 *PhiIndex = Idx; 3237 return PHI; 3238 } 3239 3240 return NULL; 3241} 3242 3243/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3244/// instruction to a phi node dominated by the switch, if that would mean that 3245/// some of the destination blocks of the switch can be folded away. 3246/// Returns true if a change is made. 3247static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3248 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3249 ForwardingNodesMap ForwardingNodes; 3250 3251 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3252 ConstantInt *CaseValue = I.getCaseValue(); 3253 BasicBlock *CaseDest = I.getCaseSuccessor(); 3254 3255 int PhiIndex; 3256 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3257 &PhiIndex); 3258 if (!PHI) continue; 3259 3260 ForwardingNodes[PHI].push_back(PhiIndex); 3261 } 3262 3263 bool Changed = false; 3264 3265 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3266 E = ForwardingNodes.end(); I != E; ++I) { 3267 PHINode *Phi = I->first; 3268 SmallVectorImpl<int> &Indexes = I->second; 3269 3270 if (Indexes.size() < 2) continue; 3271 3272 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3273 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3274 Changed = true; 3275 } 3276 3277 return Changed; 3278} 3279 3280/// ValidLookupTableConstant - Return true if the backend will be able to handle 3281/// initializing an array of constants like C. 3282static bool ValidLookupTableConstant(Constant *C) { 3283 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3284 return CE->isGEPWithNoNotionalOverIndexing(); 3285 3286 return isa<ConstantFP>(C) || 3287 isa<ConstantInt>(C) || 3288 isa<ConstantPointerNull>(C) || 3289 isa<GlobalValue>(C) || 3290 isa<UndefValue>(C); 3291} 3292 3293/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3294/// its constant value in ConstantPool, returning 0 if it's not there. 3295static Constant *LookupConstant(Value *V, 3296 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3297 if (Constant *C = dyn_cast<Constant>(V)) 3298 return C; 3299 return ConstantPool.lookup(V); 3300} 3301 3302/// ConstantFold - Try to fold instruction I into a constant. This works for 3303/// simple instructions such as binary operations where both operands are 3304/// constant or can be replaced by constants from the ConstantPool. Returns the 3305/// resulting constant on success, 0 otherwise. 3306static Constant *ConstantFold(Instruction *I, 3307 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3308 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3309 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3310 if (!A) 3311 return 0; 3312 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3313 if (!B) 3314 return 0; 3315 return ConstantExpr::get(BO->getOpcode(), A, B); 3316 } 3317 3318 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3319 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3320 if (!A) 3321 return 0; 3322 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3323 if (!B) 3324 return 0; 3325 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3326 } 3327 3328 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3329 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3330 if (!A) 3331 return 0; 3332 if (A->isAllOnesValue()) 3333 return LookupConstant(Select->getTrueValue(), ConstantPool); 3334 if (A->isNullValue()) 3335 return LookupConstant(Select->getFalseValue(), ConstantPool); 3336 return 0; 3337 } 3338 3339 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3340 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3341 if (!A) 3342 return 0; 3343 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3344 } 3345 3346 return 0; 3347} 3348 3349/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3350/// at the common destination basic block, *CommonDest, for one of the case 3351/// destionations CaseDest corresponding to value CaseVal (0 for the default 3352/// case), of a switch instruction SI. 3353static bool 3354GetCaseResults(SwitchInst *SI, 3355 ConstantInt *CaseVal, 3356 BasicBlock *CaseDest, 3357 BasicBlock **CommonDest, 3358 SmallVectorImpl<std::pair<PHINode*,Constant*> > &Res) { 3359 // The block from which we enter the common destination. 3360 BasicBlock *Pred = SI->getParent(); 3361 3362 // If CaseDest is empty except for some side-effect free instructions through 3363 // which we can constant-propagate the CaseVal, continue to its successor. 3364 SmallDenseMap<Value*, Constant*> ConstantPool; 3365 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3366 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3367 ++I) { 3368 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3369 // If the terminator is a simple branch, continue to the next block. 3370 if (T->getNumSuccessors() != 1) 3371 return false; 3372 Pred = CaseDest; 3373 CaseDest = T->getSuccessor(0); 3374 } else if (isa<DbgInfoIntrinsic>(I)) { 3375 // Skip debug intrinsic. 3376 continue; 3377 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3378 // Instruction is side-effect free and constant. 3379 ConstantPool.insert(std::make_pair(I, C)); 3380 } else { 3381 break; 3382 } 3383 } 3384 3385 // If we did not have a CommonDest before, use the current one. 3386 if (!*CommonDest) 3387 *CommonDest = CaseDest; 3388 // If the destination isn't the common one, abort. 3389 if (CaseDest != *CommonDest) 3390 return false; 3391 3392 // Get the values for this case from phi nodes in the destination block. 3393 BasicBlock::iterator I = (*CommonDest)->begin(); 3394 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3395 int Idx = PHI->getBasicBlockIndex(Pred); 3396 if (Idx == -1) 3397 continue; 3398 3399 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3400 ConstantPool); 3401 if (!ConstVal) 3402 return false; 3403 3404 // Note: If the constant comes from constant-propagating the case value 3405 // through the CaseDest basic block, it will be safe to remove the 3406 // instructions in that block. They cannot be used (except in the phi nodes 3407 // we visit) outside CaseDest, because that block does not dominate its 3408 // successor. If it did, we would not be in this phi node. 3409 3410 // Be conservative about which kinds of constants we support. 3411 if (!ValidLookupTableConstant(ConstVal)) 3412 return false; 3413 3414 Res.push_back(std::make_pair(PHI, ConstVal)); 3415 } 3416 3417 return true; 3418} 3419 3420namespace { 3421 /// SwitchLookupTable - This class represents a lookup table that can be used 3422 /// to replace a switch. 3423 class SwitchLookupTable { 3424 public: 3425 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3426 /// with the contents of Values, using DefaultValue to fill any holes in the 3427 /// table. 3428 SwitchLookupTable(Module &M, 3429 uint64_t TableSize, 3430 ConstantInt *Offset, 3431 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, 3432 Constant *DefaultValue, 3433 const DataLayout *TD); 3434 3435 /// BuildLookup - Build instructions with Builder to retrieve the value at 3436 /// the position given by Index in the lookup table. 3437 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3438 3439 /// WouldFitInRegister - Return true if a table with TableSize elements of 3440 /// type ElementType would fit in a target-legal register. 3441 static bool WouldFitInRegister(const DataLayout *TD, 3442 uint64_t TableSize, 3443 const Type *ElementType); 3444 3445 private: 3446 // Depending on the contents of the table, it can be represented in 3447 // different ways. 3448 enum { 3449 // For tables where each element contains the same value, we just have to 3450 // store that single value and return it for each lookup. 3451 SingleValueKind, 3452 3453 // For small tables with integer elements, we can pack them into a bitmap 3454 // that fits into a target-legal register. Values are retrieved by 3455 // shift and mask operations. 3456 BitMapKind, 3457 3458 // The table is stored as an array of values. Values are retrieved by load 3459 // instructions from the table. 3460 ArrayKind 3461 } Kind; 3462 3463 // For SingleValueKind, this is the single value. 3464 Constant *SingleValue; 3465 3466 // For BitMapKind, this is the bitmap. 3467 ConstantInt *BitMap; 3468 IntegerType *BitMapElementTy; 3469 3470 // For ArrayKind, this is the array. 3471 GlobalVariable *Array; 3472 }; 3473} 3474 3475SwitchLookupTable::SwitchLookupTable(Module &M, 3476 uint64_t TableSize, 3477 ConstantInt *Offset, 3478 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, 3479 Constant *DefaultValue, 3480 const DataLayout *TD) 3481 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3482 assert(Values.size() && "Can't build lookup table without values!"); 3483 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3484 3485 // If all values in the table are equal, this is that value. 3486 SingleValue = Values.begin()->second; 3487 3488 // Build up the table contents. 3489 SmallVector<Constant*, 64> TableContents(TableSize); 3490 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3491 ConstantInt *CaseVal = Values[I].first; 3492 Constant *CaseRes = Values[I].second; 3493 assert(CaseRes->getType() == DefaultValue->getType()); 3494 3495 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3496 .getLimitedValue(); 3497 TableContents[Idx] = CaseRes; 3498 3499 if (CaseRes != SingleValue) 3500 SingleValue = 0; 3501 } 3502 3503 // Fill in any holes in the table with the default result. 3504 if (Values.size() < TableSize) { 3505 for (uint64_t I = 0; I < TableSize; ++I) { 3506 if (!TableContents[I]) 3507 TableContents[I] = DefaultValue; 3508 } 3509 3510 if (DefaultValue != SingleValue) 3511 SingleValue = 0; 3512 } 3513 3514 // If each element in the table contains the same value, we only need to store 3515 // that single value. 3516 if (SingleValue) { 3517 Kind = SingleValueKind; 3518 return; 3519 } 3520 3521 // If the type is integer and the table fits in a register, build a bitmap. 3522 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3523 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3524 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3525 for (uint64_t I = TableSize; I > 0; --I) { 3526 TableInt <<= IT->getBitWidth(); 3527 // Insert values into the bitmap. Undef values are set to zero. 3528 if (!isa<UndefValue>(TableContents[I - 1])) { 3529 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3530 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3531 } 3532 } 3533 BitMap = ConstantInt::get(M.getContext(), TableInt); 3534 BitMapElementTy = IT; 3535 Kind = BitMapKind; 3536 ++NumBitMaps; 3537 return; 3538 } 3539 3540 // Store the table in an array. 3541 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3542 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3543 3544 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3545 GlobalVariable::PrivateLinkage, 3546 Initializer, 3547 "switch.table"); 3548 Array->setUnnamedAddr(true); 3549 Kind = ArrayKind; 3550} 3551 3552Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3553 switch (Kind) { 3554 case SingleValueKind: 3555 return SingleValue; 3556 case BitMapKind: { 3557 // Type of the bitmap (e.g. i59). 3558 IntegerType *MapTy = BitMap->getType(); 3559 3560 // Cast Index to the same type as the bitmap. 3561 // Note: The Index is <= the number of elements in the table, so 3562 // truncating it to the width of the bitmask is safe. 3563 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3564 3565 // Multiply the shift amount by the element width. 3566 ShiftAmt = Builder.CreateMul(ShiftAmt, 3567 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3568 "switch.shiftamt"); 3569 3570 // Shift down. 3571 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3572 "switch.downshift"); 3573 // Mask off. 3574 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3575 "switch.masked"); 3576 } 3577 case ArrayKind: { 3578 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3579 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3580 "switch.gep"); 3581 return Builder.CreateLoad(GEP, "switch.load"); 3582 } 3583 } 3584 llvm_unreachable("Unknown lookup table kind!"); 3585} 3586 3587bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3588 uint64_t TableSize, 3589 const Type *ElementType) { 3590 if (!TD) 3591 return false; 3592 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3593 if (!IT) 3594 return false; 3595 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3596 // are <= 15, we could try to narrow the type. 3597 3598 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3599 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3600 return false; 3601 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3602} 3603 3604/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3605/// for this switch, based on the number of cases, size of the table and the 3606/// types of the results. 3607static bool ShouldBuildLookupTable(SwitchInst *SI, 3608 uint64_t TableSize, 3609 const TargetTransformInfo &TTI, 3610 const DataLayout *TD, 3611 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3612 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3613 return false; // TableSize overflowed, or mul below might overflow. 3614 3615 bool AllTablesFitInRegister = true; 3616 bool HasIllegalType = false; 3617 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3618 E = ResultTypes.end(); I != E; ++I) { 3619 Type *Ty = I->second; 3620 3621 // Saturate this flag to true. 3622 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3623 3624 // Saturate this flag to false. 3625 AllTablesFitInRegister = AllTablesFitInRegister && 3626 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3627 3628 // If both flags saturate, we're done. NOTE: This *only* works with 3629 // saturating flags, and all flags have to saturate first due to the 3630 // non-deterministic behavior of iterating over a dense map. 3631 if (HasIllegalType && !AllTablesFitInRegister) 3632 break; 3633 } 3634 3635 // If each table would fit in a register, we should build it anyway. 3636 if (AllTablesFitInRegister) 3637 return true; 3638 3639 // Don't build a table that doesn't fit in-register if it has illegal types. 3640 if (HasIllegalType) 3641 return false; 3642 3643 // The table density should be at least 40%. This is the same criterion as for 3644 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3645 // FIXME: Find the best cut-off. 3646 return SI->getNumCases() * 10 >= TableSize * 4; 3647} 3648 3649/// SwitchToLookupTable - If the switch is only used to initialize one or more 3650/// phi nodes in a common successor block with different constant values, 3651/// replace the switch with lookup tables. 3652static bool SwitchToLookupTable(SwitchInst *SI, 3653 IRBuilder<> &Builder, 3654 const TargetTransformInfo &TTI, 3655 const DataLayout* TD) { 3656 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3657 3658 // Only build lookup table when we have a target that supports it. 3659 if (!TTI.shouldBuildLookupTables()) 3660 return false; 3661 3662 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3663 // split off a dense part and build a lookup table for that. 3664 3665 // FIXME: This creates arrays of GEPs to constant strings, which means each 3666 // GEP needs a runtime relocation in PIC code. We should just build one big 3667 // string and lookup indices into that. 3668 3669 // Ignore the switch if the number of cases is too small. 3670 // This is similar to the check when building jump tables in 3671 // SelectionDAGBuilder::handleJTSwitchCase. 3672 // FIXME: Determine the best cut-off. 3673 if (SI->getNumCases() < 4) 3674 return false; 3675 3676 // Figure out the corresponding result for each case value and phi node in the 3677 // common destination, as well as the the min and max case values. 3678 assert(SI->case_begin() != SI->case_end()); 3679 SwitchInst::CaseIt CI = SI->case_begin(); 3680 ConstantInt *MinCaseVal = CI.getCaseValue(); 3681 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3682 3683 BasicBlock *CommonDest = 0; 3684 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3685 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3686 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3687 SmallDenseMap<PHINode*, Type*> ResultTypes; 3688 SmallVector<PHINode*, 4> PHIs; 3689 3690 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3691 ConstantInt *CaseVal = CI.getCaseValue(); 3692 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3693 MinCaseVal = CaseVal; 3694 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3695 MaxCaseVal = CaseVal; 3696 3697 // Resulting value at phi nodes for this case value. 3698 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3699 ResultsTy Results; 3700 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3701 Results)) 3702 return false; 3703 3704 // Append the result from this case to the list for each phi. 3705 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3706 if (!ResultLists.count(I->first)) 3707 PHIs.push_back(I->first); 3708 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3709 } 3710 } 3711 3712 // Get the resulting values for the default case. 3713 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3714 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3715 DefaultResultsList)) 3716 return false; 3717 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3718 PHINode *PHI = DefaultResultsList[I].first; 3719 Constant *Result = DefaultResultsList[I].second; 3720 DefaultResults[PHI] = Result; 3721 ResultTypes[PHI] = Result->getType(); 3722 } 3723 3724 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3725 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3726 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3727 return false; 3728 3729 // Create the BB that does the lookups. 3730 Module &Mod = *CommonDest->getParent()->getParent(); 3731 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3732 "switch.lookup", 3733 CommonDest->getParent(), 3734 CommonDest); 3735 3736 // Check whether the condition value is within the case range, and branch to 3737 // the new BB. 3738 Builder.SetInsertPoint(SI); 3739 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3740 "switch.tableidx"); 3741 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3742 MinCaseVal->getType(), TableSize)); 3743 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3744 3745 // Populate the BB that does the lookups. 3746 Builder.SetInsertPoint(LookupBB); 3747 bool ReturnedEarly = false; 3748 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3749 PHINode *PHI = PHIs[I]; 3750 3751 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3752 DefaultResults[PHI], TD); 3753 3754 Value *Result = Table.BuildLookup(TableIndex, Builder); 3755 3756 // If the result is used to return immediately from the function, we want to 3757 // do that right here. 3758 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3759 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3760 Builder.CreateRet(Result); 3761 ReturnedEarly = true; 3762 break; 3763 } 3764 3765 PHI->addIncoming(Result, LookupBB); 3766 } 3767 3768 if (!ReturnedEarly) 3769 Builder.CreateBr(CommonDest); 3770 3771 // Remove the switch. 3772 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3773 BasicBlock *Succ = SI->getSuccessor(i); 3774 if (Succ == SI->getDefaultDest()) continue; 3775 Succ->removePredecessor(SI->getParent()); 3776 } 3777 SI->eraseFromParent(); 3778 3779 ++NumLookupTables; 3780 return true; 3781} 3782 3783bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3784 BasicBlock *BB = SI->getParent(); 3785 3786 if (isValueEqualityComparison(SI)) { 3787 // If we only have one predecessor, and if it is a branch on this value, 3788 // see if that predecessor totally determines the outcome of this switch. 3789 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3790 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3791 return SimplifyCFG(BB, TTI, TD) | true; 3792 3793 Value *Cond = SI->getCondition(); 3794 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3795 if (SimplifySwitchOnSelect(SI, Select)) 3796 return SimplifyCFG(BB, TTI, TD) | true; 3797 3798 // If the block only contains the switch, see if we can fold the block 3799 // away into any preds. 3800 BasicBlock::iterator BBI = BB->begin(); 3801 // Ignore dbg intrinsics. 3802 while (isa<DbgInfoIntrinsic>(BBI)) 3803 ++BBI; 3804 if (SI == &*BBI) 3805 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3806 return SimplifyCFG(BB, TTI, TD) | true; 3807 } 3808 3809 // Try to transform the switch into an icmp and a branch. 3810 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3811 return SimplifyCFG(BB, TTI, TD) | true; 3812 3813 // Remove unreachable cases. 3814 if (EliminateDeadSwitchCases(SI)) 3815 return SimplifyCFG(BB, TTI, TD) | true; 3816 3817 if (ForwardSwitchConditionToPHI(SI)) 3818 return SimplifyCFG(BB, TTI, TD) | true; 3819 3820 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3821 return SimplifyCFG(BB, TTI, TD) | true; 3822 3823 return false; 3824} 3825 3826bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3827 BasicBlock *BB = IBI->getParent(); 3828 bool Changed = false; 3829 3830 // Eliminate redundant destinations. 3831 SmallPtrSet<Value *, 8> Succs; 3832 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3833 BasicBlock *Dest = IBI->getDestination(i); 3834 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3835 Dest->removePredecessor(BB); 3836 IBI->removeDestination(i); 3837 --i; --e; 3838 Changed = true; 3839 } 3840 } 3841 3842 if (IBI->getNumDestinations() == 0) { 3843 // If the indirectbr has no successors, change it to unreachable. 3844 new UnreachableInst(IBI->getContext(), IBI); 3845 EraseTerminatorInstAndDCECond(IBI); 3846 return true; 3847 } 3848 3849 if (IBI->getNumDestinations() == 1) { 3850 // If the indirectbr has one successor, change it to a direct branch. 3851 BranchInst::Create(IBI->getDestination(0), IBI); 3852 EraseTerminatorInstAndDCECond(IBI); 3853 return true; 3854 } 3855 3856 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3857 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3858 return SimplifyCFG(BB, TTI, TD) | true; 3859 } 3860 return Changed; 3861} 3862 3863bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3864 BasicBlock *BB = BI->getParent(); 3865 3866 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3867 return true; 3868 3869 // If the Terminator is the only non-phi instruction, simplify the block. 3870 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3871 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3872 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3873 return true; 3874 3875 // If the only instruction in the block is a seteq/setne comparison 3876 // against a constant, try to simplify the block. 3877 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3878 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3879 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3880 ; 3881 if (I->isTerminator() && 3882 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3883 return true; 3884 } 3885 3886 // If this basic block is ONLY a compare and a branch, and if a predecessor 3887 // branches to us and our successor, fold the comparison into the 3888 // predecessor and use logical operations to update the incoming value 3889 // for PHI nodes in common successor. 3890 if (FoldBranchToCommonDest(BI)) 3891 return SimplifyCFG(BB, TTI, TD) | true; 3892 return false; 3893} 3894 3895 3896bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3897 BasicBlock *BB = BI->getParent(); 3898 3899 // Conditional branch 3900 if (isValueEqualityComparison(BI)) { 3901 // If we only have one predecessor, and if it is a branch on this value, 3902 // see if that predecessor totally determines the outcome of this 3903 // switch. 3904 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3905 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3906 return SimplifyCFG(BB, TTI, TD) | true; 3907 3908 // This block must be empty, except for the setcond inst, if it exists. 3909 // Ignore dbg intrinsics. 3910 BasicBlock::iterator I = BB->begin(); 3911 // Ignore dbg intrinsics. 3912 while (isa<DbgInfoIntrinsic>(I)) 3913 ++I; 3914 if (&*I == BI) { 3915 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3916 return SimplifyCFG(BB, TTI, TD) | true; 3917 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3918 ++I; 3919 // Ignore dbg intrinsics. 3920 while (isa<DbgInfoIntrinsic>(I)) 3921 ++I; 3922 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3923 return SimplifyCFG(BB, TTI, TD) | true; 3924 } 3925 } 3926 3927 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3928 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3929 return true; 3930 3931 // If this basic block is ONLY a compare and a branch, and if a predecessor 3932 // branches to us and one of our successors, fold the comparison into the 3933 // predecessor and use logical operations to pick the right destination. 3934 if (FoldBranchToCommonDest(BI)) 3935 return SimplifyCFG(BB, TTI, TD) | true; 3936 3937 // We have a conditional branch to two blocks that are only reachable 3938 // from BI. We know that the condbr dominates the two blocks, so see if 3939 // there is any identical code in the "then" and "else" blocks. If so, we 3940 // can hoist it up to the branching block. 3941 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3942 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3943 if (HoistThenElseCodeToIf(BI)) 3944 return SimplifyCFG(BB, TTI, TD) | true; 3945 } else { 3946 // If Successor #1 has multiple preds, we may be able to conditionally 3947 // execute Successor #0 if it branches to successor #1. 3948 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3949 if (Succ0TI->getNumSuccessors() == 1 && 3950 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3951 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3952 return SimplifyCFG(BB, TTI, TD) | true; 3953 } 3954 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3955 // If Successor #0 has multiple preds, we may be able to conditionally 3956 // execute Successor #1 if it branches to successor #0. 3957 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3958 if (Succ1TI->getNumSuccessors() == 1 && 3959 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3960 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3961 return SimplifyCFG(BB, TTI, TD) | true; 3962 } 3963 3964 // If this is a branch on a phi node in the current block, thread control 3965 // through this block if any PHI node entries are constants. 3966 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3967 if (PN->getParent() == BI->getParent()) 3968 if (FoldCondBranchOnPHI(BI, TD)) 3969 return SimplifyCFG(BB, TTI, TD) | true; 3970 3971 // Scan predecessor blocks for conditional branches. 3972 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3973 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3974 if (PBI != BI && PBI->isConditional()) 3975 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3976 return SimplifyCFG(BB, TTI, TD) | true; 3977 3978 return false; 3979} 3980 3981/// Check if passing a value to an instruction will cause undefined behavior. 3982static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3983 Constant *C = dyn_cast<Constant>(V); 3984 if (!C) 3985 return false; 3986 3987 if (I->use_empty()) 3988 return false; 3989 3990 if (C->isNullValue()) { 3991 // Only look at the first use, avoid hurting compile time with long uselists 3992 User *Use = *I->use_begin(); 3993 3994 // Now make sure that there are no instructions in between that can alter 3995 // control flow (eg. calls) 3996 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3997 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3998 return false; 3999 4000 // Look through GEPs. A load from a GEP derived from NULL is still undefined 4001 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 4002 if (GEP->getPointerOperand() == I) 4003 return passingValueIsAlwaysUndefined(V, GEP); 4004 4005 // Look through bitcasts. 4006 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 4007 return passingValueIsAlwaysUndefined(V, BC); 4008 4009 // Load from null is undefined. 4010 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 4011 if (!LI->isVolatile()) 4012 return LI->getPointerAddressSpace() == 0; 4013 4014 // Store to null is undefined. 4015 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 4016 if (!SI->isVolatile()) 4017 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 4018 } 4019 return false; 4020} 4021 4022/// If BB has an incoming value that will always trigger undefined behavior 4023/// (eg. null pointer dereference), remove the branch leading here. 4024static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 4025 for (BasicBlock::iterator i = BB->begin(); 4026 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 4027 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 4028 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 4029 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 4030 IRBuilder<> Builder(T); 4031 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 4032 BB->removePredecessor(PHI->getIncomingBlock(i)); 4033 // Turn uncoditional branches into unreachables and remove the dead 4034 // destination from conditional branches. 4035 if (BI->isUnconditional()) 4036 Builder.CreateUnreachable(); 4037 else 4038 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 4039 BI->getSuccessor(0)); 4040 BI->eraseFromParent(); 4041 return true; 4042 } 4043 // TODO: SwitchInst. 4044 } 4045 4046 return false; 4047} 4048 4049bool SimplifyCFGOpt::run(BasicBlock *BB) { 4050 bool Changed = false; 4051 4052 assert(BB && BB->getParent() && "Block not embedded in function!"); 4053 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4054 4055 // Remove basic blocks that have no predecessors (except the entry block)... 4056 // or that just have themself as a predecessor. These are unreachable. 4057 if ((pred_begin(BB) == pred_end(BB) && 4058 BB != &BB->getParent()->getEntryBlock()) || 4059 BB->getSinglePredecessor() == BB) { 4060 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4061 DeleteDeadBlock(BB); 4062 return true; 4063 } 4064 4065 // Check to see if we can constant propagate this terminator instruction 4066 // away... 4067 Changed |= ConstantFoldTerminator(BB, true); 4068 4069 // Check for and eliminate duplicate PHI nodes in this block. 4070 Changed |= EliminateDuplicatePHINodes(BB); 4071 4072 // Check for and remove branches that will always cause undefined behavior. 4073 Changed |= removeUndefIntroducingPredecessor(BB); 4074 4075 // Merge basic blocks into their predecessor if there is only one distinct 4076 // pred, and if there is only one distinct successor of the predecessor, and 4077 // if there are no PHI nodes. 4078 // 4079 if (MergeBlockIntoPredecessor(BB)) 4080 return true; 4081 4082 IRBuilder<> Builder(BB); 4083 4084 // If there is a trivial two-entry PHI node in this basic block, and we can 4085 // eliminate it, do so now. 4086 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4087 if (PN->getNumIncomingValues() == 2) 4088 Changed |= FoldTwoEntryPHINode(PN, TD); 4089 4090 Builder.SetInsertPoint(BB->getTerminator()); 4091 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4092 if (BI->isUnconditional()) { 4093 if (SimplifyUncondBranch(BI, Builder)) return true; 4094 } else { 4095 if (SimplifyCondBranch(BI, Builder)) return true; 4096 } 4097 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4098 if (SimplifyReturn(RI, Builder)) return true; 4099 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4100 if (SimplifyResume(RI, Builder)) return true; 4101 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4102 if (SimplifySwitch(SI, Builder)) return true; 4103 } else if (UnreachableInst *UI = 4104 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4105 if (SimplifyUnreachable(UI)) return true; 4106 } else if (IndirectBrInst *IBI = 4107 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4108 if (SimplifyIndirectBr(IBI)) return true; 4109 } 4110 4111 return Changed; 4112} 4113 4114/// SimplifyCFG - This function is used to do simplification of a CFG. For 4115/// example, it adjusts branches to branches to eliminate the extra hop, it 4116/// eliminates unreachable basic blocks, and does other "peephole" optimization 4117/// of the CFG. It returns true if a modification was made. 4118/// 4119bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4120 const DataLayout *TD) { 4121 return SimplifyCFGOpt(TTI, TD).run(BB); 4122} 4123