SimplifyCFG.cpp revision 59324297650c12a8dccf1a7ad650a9e895fdc17e
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/Constants.h" 17#include "llvm/DerivedTypes.h" 18#include "llvm/GlobalVariable.h" 19#include "llvm/IRBuilder.h" 20#include "llvm/Instructions.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/LLVMContext.h" 23#include "llvm/MDBuilder.h" 24#include "llvm/Metadata.h" 25#include "llvm/Operator.h" 26#include "llvm/Type.h" 27#include "llvm/ADT/DenseMap.h" 28#include "llvm/ADT/STLExtras.h" 29#include "llvm/ADT/SetVector.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/Analysis/InstructionSimplify.h" 34#include "llvm/Analysis/ValueTracking.h" 35#include "llvm/Support/CFG.h" 36#include "llvm/Support/CommandLine.h" 37#include "llvm/Support/ConstantRange.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/NoFolder.h" 40#include "llvm/Support/raw_ostream.h" 41#include "llvm/Target/TargetData.h" 42#include "llvm/Transforms/Utils/BasicBlockUtils.h" 43#include <algorithm> 44#include <set> 45#include <map> 46using namespace llvm; 47 48static cl::opt<unsigned> 49PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 50 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 51 52static cl::opt<bool> 53DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 54 cl::desc("Duplicate return instructions into unconditional branches")); 55 56STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 57 58namespace { 59 /// ValueEqualityComparisonCase - Represents a case of a switch. 60 struct ValueEqualityComparisonCase { 61 ConstantInt *Value; 62 BasicBlock *Dest; 63 64 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 65 : Value(Value), Dest(Dest) {} 66 67 bool operator<(ValueEqualityComparisonCase RHS) const { 68 // Comparing pointers is ok as we only rely on the order for uniquing. 69 return Value < RHS.Value; 70 } 71 }; 72 73class SimplifyCFGOpt { 74 const TargetData *const TD; 75 76 Value *isValueEqualityComparison(TerminatorInst *TI); 77 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 78 std::vector<ValueEqualityComparisonCase> &Cases); 79 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 80 BasicBlock *Pred, 81 IRBuilder<> &Builder); 82 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 83 IRBuilder<> &Builder); 84 85 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 86 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 87 bool SimplifyUnreachable(UnreachableInst *UI); 88 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 89 bool SimplifyIndirectBr(IndirectBrInst *IBI); 90 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 91 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 92 93public: 94 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} 95 bool run(BasicBlock *BB); 96}; 97} 98 99/// SafeToMergeTerminators - Return true if it is safe to merge these two 100/// terminator instructions together. 101/// 102static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 103 if (SI1 == SI2) return false; // Can't merge with self! 104 105 // It is not safe to merge these two switch instructions if they have a common 106 // successor, and if that successor has a PHI node, and if *that* PHI node has 107 // conflicting incoming values from the two switch blocks. 108 BasicBlock *SI1BB = SI1->getParent(); 109 BasicBlock *SI2BB = SI2->getParent(); 110 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 111 112 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 113 if (SI1Succs.count(*I)) 114 for (BasicBlock::iterator BBI = (*I)->begin(); 115 isa<PHINode>(BBI); ++BBI) { 116 PHINode *PN = cast<PHINode>(BBI); 117 if (PN->getIncomingValueForBlock(SI1BB) != 118 PN->getIncomingValueForBlock(SI2BB)) 119 return false; 120 } 121 122 return true; 123} 124 125/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 126/// to merge these two terminator instructions together, where SI1 is an 127/// unconditional branch. PhiNodes will store all PHI nodes in common 128/// successors. 129/// 130static bool isProfitableToFoldUnconditional(BranchInst *SI1, 131 BranchInst *SI2, 132 Instruction *Cond, 133 SmallVectorImpl<PHINode*> &PhiNodes) { 134 if (SI1 == SI2) return false; // Can't merge with self! 135 assert(SI1->isUnconditional() && SI2->isConditional()); 136 137 // We fold the unconditional branch if we can easily update all PHI nodes in 138 // common successors: 139 // 1> We have a constant incoming value for the conditional branch; 140 // 2> We have "Cond" as the incoming value for the unconditional branch; 141 // 3> SI2->getCondition() and Cond have same operands. 142 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 143 if (!Ci2) return false; 144 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 145 Cond->getOperand(1) == Ci2->getOperand(1)) && 146 !(Cond->getOperand(0) == Ci2->getOperand(1) && 147 Cond->getOperand(1) == Ci2->getOperand(0))) 148 return false; 149 150 BasicBlock *SI1BB = SI1->getParent(); 151 BasicBlock *SI2BB = SI2->getParent(); 152 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 153 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 154 if (SI1Succs.count(*I)) 155 for (BasicBlock::iterator BBI = (*I)->begin(); 156 isa<PHINode>(BBI); ++BBI) { 157 PHINode *PN = cast<PHINode>(BBI); 158 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 159 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 160 return false; 161 PhiNodes.push_back(PN); 162 } 163 return true; 164} 165 166/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 167/// now be entries in it from the 'NewPred' block. The values that will be 168/// flowing into the PHI nodes will be the same as those coming in from 169/// ExistPred, an existing predecessor of Succ. 170static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 171 BasicBlock *ExistPred) { 172 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 173 174 PHINode *PN; 175 for (BasicBlock::iterator I = Succ->begin(); 176 (PN = dyn_cast<PHINode>(I)); ++I) 177 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 178} 179 180 181/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 182/// least one PHI node in it), check to see if the merge at this block is due 183/// to an "if condition". If so, return the boolean condition that determines 184/// which entry into BB will be taken. Also, return by references the block 185/// that will be entered from if the condition is true, and the block that will 186/// be entered if the condition is false. 187/// 188/// This does no checking to see if the true/false blocks have large or unsavory 189/// instructions in them. 190static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 191 BasicBlock *&IfFalse) { 192 PHINode *SomePHI = cast<PHINode>(BB->begin()); 193 assert(SomePHI->getNumIncomingValues() == 2 && 194 "Function can only handle blocks with 2 predecessors!"); 195 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 196 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 197 198 // We can only handle branches. Other control flow will be lowered to 199 // branches if possible anyway. 200 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 201 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 202 if (Pred1Br == 0 || Pred2Br == 0) 203 return 0; 204 205 // Eliminate code duplication by ensuring that Pred1Br is conditional if 206 // either are. 207 if (Pred2Br->isConditional()) { 208 // If both branches are conditional, we don't have an "if statement". In 209 // reality, we could transform this case, but since the condition will be 210 // required anyway, we stand no chance of eliminating it, so the xform is 211 // probably not profitable. 212 if (Pred1Br->isConditional()) 213 return 0; 214 215 std::swap(Pred1, Pred2); 216 std::swap(Pred1Br, Pred2Br); 217 } 218 219 if (Pred1Br->isConditional()) { 220 // The only thing we have to watch out for here is to make sure that Pred2 221 // doesn't have incoming edges from other blocks. If it does, the condition 222 // doesn't dominate BB. 223 if (Pred2->getSinglePredecessor() == 0) 224 return 0; 225 226 // If we found a conditional branch predecessor, make sure that it branches 227 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 228 if (Pred1Br->getSuccessor(0) == BB && 229 Pred1Br->getSuccessor(1) == Pred2) { 230 IfTrue = Pred1; 231 IfFalse = Pred2; 232 } else if (Pred1Br->getSuccessor(0) == Pred2 && 233 Pred1Br->getSuccessor(1) == BB) { 234 IfTrue = Pred2; 235 IfFalse = Pred1; 236 } else { 237 // We know that one arm of the conditional goes to BB, so the other must 238 // go somewhere unrelated, and this must not be an "if statement". 239 return 0; 240 } 241 242 return Pred1Br->getCondition(); 243 } 244 245 // Ok, if we got here, both predecessors end with an unconditional branch to 246 // BB. Don't panic! If both blocks only have a single (identical) 247 // predecessor, and THAT is a conditional branch, then we're all ok! 248 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 249 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 250 return 0; 251 252 // Otherwise, if this is a conditional branch, then we can use it! 253 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 254 if (BI == 0) return 0; 255 256 assert(BI->isConditional() && "Two successors but not conditional?"); 257 if (BI->getSuccessor(0) == Pred1) { 258 IfTrue = Pred1; 259 IfFalse = Pred2; 260 } else { 261 IfTrue = Pred2; 262 IfFalse = Pred1; 263 } 264 return BI->getCondition(); 265} 266 267/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 268/// given instruction, which is assumed to be safe to speculate. 1 means 269/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 270static unsigned ComputeSpeculationCost(const User *I) { 271 assert(isSafeToSpeculativelyExecute(I) && 272 "Instruction is not safe to speculatively execute!"); 273 switch (Operator::getOpcode(I)) { 274 default: 275 // In doubt, be conservative. 276 return UINT_MAX; 277 case Instruction::GetElementPtr: 278 // GEPs are cheap if all indices are constant. 279 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 280 return UINT_MAX; 281 return 1; 282 case Instruction::Load: 283 case Instruction::Add: 284 case Instruction::Sub: 285 case Instruction::And: 286 case Instruction::Or: 287 case Instruction::Xor: 288 case Instruction::Shl: 289 case Instruction::LShr: 290 case Instruction::AShr: 291 case Instruction::ICmp: 292 case Instruction::Trunc: 293 case Instruction::ZExt: 294 case Instruction::SExt: 295 return 1; // These are all cheap. 296 297 case Instruction::Call: 298 case Instruction::Select: 299 return 2; 300 } 301} 302 303/// DominatesMergePoint - If we have a merge point of an "if condition" as 304/// accepted above, return true if the specified value dominates the block. We 305/// don't handle the true generality of domination here, just a special case 306/// which works well enough for us. 307/// 308/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 309/// see if V (which must be an instruction) and its recursive operands 310/// that do not dominate BB have a combined cost lower than CostRemaining and 311/// are non-trapping. If both are true, the instruction is inserted into the 312/// set and true is returned. 313/// 314/// The cost for most non-trapping instructions is defined as 1 except for 315/// Select whose cost is 2. 316/// 317/// After this function returns, CostRemaining is decreased by the cost of 318/// V plus its non-dominating operands. If that cost is greater than 319/// CostRemaining, false is returned and CostRemaining is undefined. 320static bool DominatesMergePoint(Value *V, BasicBlock *BB, 321 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 322 unsigned &CostRemaining) { 323 Instruction *I = dyn_cast<Instruction>(V); 324 if (!I) { 325 // Non-instructions all dominate instructions, but not all constantexprs 326 // can be executed unconditionally. 327 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 328 if (C->canTrap()) 329 return false; 330 return true; 331 } 332 BasicBlock *PBB = I->getParent(); 333 334 // We don't want to allow weird loops that might have the "if condition" in 335 // the bottom of this block. 336 if (PBB == BB) return false; 337 338 // If this instruction is defined in a block that contains an unconditional 339 // branch to BB, then it must be in the 'conditional' part of the "if 340 // statement". If not, it definitely dominates the region. 341 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 342 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 343 return true; 344 345 // If we aren't allowing aggressive promotion anymore, then don't consider 346 // instructions in the 'if region'. 347 if (AggressiveInsts == 0) return false; 348 349 // If we have seen this instruction before, don't count it again. 350 if (AggressiveInsts->count(I)) return true; 351 352 // Okay, it looks like the instruction IS in the "condition". Check to 353 // see if it's a cheap instruction to unconditionally compute, and if it 354 // only uses stuff defined outside of the condition. If so, hoist it out. 355 if (!isSafeToSpeculativelyExecute(I)) 356 return false; 357 358 unsigned Cost = ComputeSpeculationCost(I); 359 360 if (Cost > CostRemaining) 361 return false; 362 363 CostRemaining -= Cost; 364 365 // Okay, we can only really hoist these out if their operands do 366 // not take us over the cost threshold. 367 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 368 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 369 return false; 370 // Okay, it's safe to do this! Remember this instruction. 371 AggressiveInsts->insert(I); 372 return true; 373} 374 375/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 376/// and PointerNullValue. Return NULL if value is not a constant int. 377static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { 378 // Normal constant int. 379 ConstantInt *CI = dyn_cast<ConstantInt>(V); 380 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 381 return CI; 382 383 // This is some kind of pointer constant. Turn it into a pointer-sized 384 // ConstantInt if possible. 385 IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); 386 387 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 388 if (isa<ConstantPointerNull>(V)) 389 return ConstantInt::get(PtrTy, 0); 390 391 // IntToPtr const int. 392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 393 if (CE->getOpcode() == Instruction::IntToPtr) 394 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 395 // The constant is very likely to have the right type already. 396 if (CI->getType() == PtrTy) 397 return CI; 398 else 399 return cast<ConstantInt> 400 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 401 } 402 return 0; 403} 404 405/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 406/// collection of icmp eq/ne instructions that compare a value against a 407/// constant, return the value being compared, and stick the constant into the 408/// Values vector. 409static Value * 410GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 411 const TargetData *TD, bool isEQ, unsigned &UsedICmps) { 412 Instruction *I = dyn_cast<Instruction>(V); 413 if (I == 0) return 0; 414 415 // If this is an icmp against a constant, handle this as one of the cases. 416 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 417 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 418 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 419 UsedICmps++; 420 Vals.push_back(C); 421 return I->getOperand(0); 422 } 423 424 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 425 // the set. 426 ConstantRange Span = 427 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 428 429 // If this is an and/!= check then we want to optimize "x ugt 2" into 430 // x != 0 && x != 1. 431 if (!isEQ) 432 Span = Span.inverse(); 433 434 // If there are a ton of values, we don't want to make a ginormous switch. 435 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 436 return 0; 437 438 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 439 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 440 UsedICmps++; 441 return I->getOperand(0); 442 } 443 return 0; 444 } 445 446 // Otherwise, we can only handle an | or &, depending on isEQ. 447 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 448 return 0; 449 450 unsigned NumValsBeforeLHS = Vals.size(); 451 unsigned UsedICmpsBeforeLHS = UsedICmps; 452 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 453 isEQ, UsedICmps)) { 454 unsigned NumVals = Vals.size(); 455 unsigned UsedICmpsBeforeRHS = UsedICmps; 456 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 457 isEQ, UsedICmps)) { 458 if (LHS == RHS) 459 return LHS; 460 Vals.resize(NumVals); 461 UsedICmps = UsedICmpsBeforeRHS; 462 } 463 464 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 465 // set it and return success. 466 if (Extra == 0 || Extra == I->getOperand(1)) { 467 Extra = I->getOperand(1); 468 return LHS; 469 } 470 471 Vals.resize(NumValsBeforeLHS); 472 UsedICmps = UsedICmpsBeforeLHS; 473 return 0; 474 } 475 476 // If the LHS can't be folded in, but Extra is available and RHS can, try to 477 // use LHS as Extra. 478 if (Extra == 0 || Extra == I->getOperand(0)) { 479 Value *OldExtra = Extra; 480 Extra = I->getOperand(0); 481 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 482 isEQ, UsedICmps)) 483 return RHS; 484 assert(Vals.size() == NumValsBeforeLHS); 485 Extra = OldExtra; 486 } 487 488 return 0; 489} 490 491static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 492 Instruction *Cond = 0; 493 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 494 Cond = dyn_cast<Instruction>(SI->getCondition()); 495 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 496 if (BI->isConditional()) 497 Cond = dyn_cast<Instruction>(BI->getCondition()); 498 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 499 Cond = dyn_cast<Instruction>(IBI->getAddress()); 500 } 501 502 TI->eraseFromParent(); 503 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 504} 505 506/// isValueEqualityComparison - Return true if the specified terminator checks 507/// to see if a value is equal to constant integer value. 508Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 509 Value *CV = 0; 510 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 511 // Do not permit merging of large switch instructions into their 512 // predecessors unless there is only one predecessor. 513 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 514 pred_end(SI->getParent())) <= 128) 515 CV = SI->getCondition(); 516 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 517 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 518 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 519 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 520 ICI->getPredicate() == ICmpInst::ICMP_NE) && 521 GetConstantInt(ICI->getOperand(1), TD)) 522 CV = ICI->getOperand(0); 523 524 // Unwrap any lossless ptrtoint cast. 525 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 526 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 527 CV = PTII->getOperand(0); 528 return CV; 529} 530 531/// GetValueEqualityComparisonCases - Given a value comparison instruction, 532/// decode all of the 'cases' that it represents and return the 'default' block. 533BasicBlock *SimplifyCFGOpt:: 534GetValueEqualityComparisonCases(TerminatorInst *TI, 535 std::vector<ValueEqualityComparisonCase> 536 &Cases) { 537 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 538 Cases.reserve(SI->getNumCases()); 539 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 540 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 541 i.getCaseSuccessor())); 542 return SI->getDefaultDest(); 543 } 544 545 BranchInst *BI = cast<BranchInst>(TI); 546 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 547 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 548 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 549 TD), 550 Succ)); 551 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 552} 553 554 555/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 556/// in the list that match the specified block. 557static void EliminateBlockCases(BasicBlock *BB, 558 std::vector<ValueEqualityComparisonCase> &Cases) { 559 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 560 if (Cases[i].Dest == BB) { 561 Cases.erase(Cases.begin()+i); 562 --i; --e; 563 } 564} 565 566/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 567/// well. 568static bool 569ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 570 std::vector<ValueEqualityComparisonCase > &C2) { 571 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 572 573 // Make V1 be smaller than V2. 574 if (V1->size() > V2->size()) 575 std::swap(V1, V2); 576 577 if (V1->size() == 0) return false; 578 if (V1->size() == 1) { 579 // Just scan V2. 580 ConstantInt *TheVal = (*V1)[0].Value; 581 for (unsigned i = 0, e = V2->size(); i != e; ++i) 582 if (TheVal == (*V2)[i].Value) 583 return true; 584 } 585 586 // Otherwise, just sort both lists and compare element by element. 587 array_pod_sort(V1->begin(), V1->end()); 588 array_pod_sort(V2->begin(), V2->end()); 589 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 590 while (i1 != e1 && i2 != e2) { 591 if ((*V1)[i1].Value == (*V2)[i2].Value) 592 return true; 593 if ((*V1)[i1].Value < (*V2)[i2].Value) 594 ++i1; 595 else 596 ++i2; 597 } 598 return false; 599} 600 601/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 602/// terminator instruction and its block is known to only have a single 603/// predecessor block, check to see if that predecessor is also a value 604/// comparison with the same value, and if that comparison determines the 605/// outcome of this comparison. If so, simplify TI. This does a very limited 606/// form of jump threading. 607bool SimplifyCFGOpt:: 608SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 609 BasicBlock *Pred, 610 IRBuilder<> &Builder) { 611 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 612 if (!PredVal) return false; // Not a value comparison in predecessor. 613 614 Value *ThisVal = isValueEqualityComparison(TI); 615 assert(ThisVal && "This isn't a value comparison!!"); 616 if (ThisVal != PredVal) return false; // Different predicates. 617 618 // TODO: Preserve branch weight metadata, similarly to how 619 // FoldValueComparisonIntoPredecessors preserves it. 620 621 // Find out information about when control will move from Pred to TI's block. 622 std::vector<ValueEqualityComparisonCase> PredCases; 623 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 624 PredCases); 625 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 626 627 // Find information about how control leaves this block. 628 std::vector<ValueEqualityComparisonCase> ThisCases; 629 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 630 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 631 632 // If TI's block is the default block from Pred's comparison, potentially 633 // simplify TI based on this knowledge. 634 if (PredDef == TI->getParent()) { 635 // If we are here, we know that the value is none of those cases listed in 636 // PredCases. If there are any cases in ThisCases that are in PredCases, we 637 // can simplify TI. 638 if (!ValuesOverlap(PredCases, ThisCases)) 639 return false; 640 641 if (isa<BranchInst>(TI)) { 642 // Okay, one of the successors of this condbr is dead. Convert it to a 643 // uncond br. 644 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 645 // Insert the new branch. 646 Instruction *NI = Builder.CreateBr(ThisDef); 647 (void) NI; 648 649 // Remove PHI node entries for the dead edge. 650 ThisCases[0].Dest->removePredecessor(TI->getParent()); 651 652 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 653 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 654 655 EraseTerminatorInstAndDCECond(TI); 656 return true; 657 } 658 659 SwitchInst *SI = cast<SwitchInst>(TI); 660 // Okay, TI has cases that are statically dead, prune them away. 661 SmallPtrSet<Constant*, 16> DeadCases; 662 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 663 DeadCases.insert(PredCases[i].Value); 664 665 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 666 << "Through successor TI: " << *TI); 667 668 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 669 --i; 670 if (DeadCases.count(i.getCaseValue())) { 671 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 672 SI->removeCase(i); 673 } 674 } 675 676 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 677 return true; 678 } 679 680 // Otherwise, TI's block must correspond to some matched value. Find out 681 // which value (or set of values) this is. 682 ConstantInt *TIV = 0; 683 BasicBlock *TIBB = TI->getParent(); 684 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 685 if (PredCases[i].Dest == TIBB) { 686 if (TIV != 0) 687 return false; // Cannot handle multiple values coming to this block. 688 TIV = PredCases[i].Value; 689 } 690 assert(TIV && "No edge from pred to succ?"); 691 692 // Okay, we found the one constant that our value can be if we get into TI's 693 // BB. Find out which successor will unconditionally be branched to. 694 BasicBlock *TheRealDest = 0; 695 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 696 if (ThisCases[i].Value == TIV) { 697 TheRealDest = ThisCases[i].Dest; 698 break; 699 } 700 701 // If not handled by any explicit cases, it is handled by the default case. 702 if (TheRealDest == 0) TheRealDest = ThisDef; 703 704 // Remove PHI node entries for dead edges. 705 BasicBlock *CheckEdge = TheRealDest; 706 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 707 if (*SI != CheckEdge) 708 (*SI)->removePredecessor(TIBB); 709 else 710 CheckEdge = 0; 711 712 // Insert the new branch. 713 Instruction *NI = Builder.CreateBr(TheRealDest); 714 (void) NI; 715 716 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 717 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 718 719 EraseTerminatorInstAndDCECond(TI); 720 return true; 721} 722 723namespace { 724 /// ConstantIntOrdering - This class implements a stable ordering of constant 725 /// integers that does not depend on their address. This is important for 726 /// applications that sort ConstantInt's to ensure uniqueness. 727 struct ConstantIntOrdering { 728 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 729 return LHS->getValue().ult(RHS->getValue()); 730 } 731 }; 732} 733 734static int ConstantIntSortPredicate(const void *P1, const void *P2) { 735 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 736 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 737 if (LHS->getValue().ult(RHS->getValue())) 738 return 1; 739 if (LHS->getValue() == RHS->getValue()) 740 return 0; 741 return -1; 742} 743 744static inline bool HasBranchWeights(const Instruction* I) { 745 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 746 if (ProfMD && ProfMD->getOperand(0)) 747 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 748 return MDS->getString().equals("branch_weights"); 749 750 return false; 751} 752 753/// Tries to get a branch weight for the given instruction, returns NULL if it 754/// can't. Pos starts at 0. 755static ConstantInt* GetWeight(Instruction* I, int Pos) { 756 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 757 if (ProfMD && ProfMD->getOperand(0)) { 758 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) { 759 if (MDS->getString().equals("branch_weights")) { 760 assert(ProfMD->getNumOperands() >= 3); 761 return dyn_cast<ConstantInt>(ProfMD->getOperand(1 + Pos)); 762 } 763 } 764 } 765 766 return 0; 767} 768 769/// Scale the given weights based on the successor TI's metadata. Scaling is 770/// done by multiplying every weight by the sum of the successor's weights. 771static void ScaleWeights(Instruction* STI, MutableArrayRef<uint64_t> Weights) { 772 // Sum the successor's weights 773 assert(HasBranchWeights(STI)); 774 unsigned Scale = 0; 775 MDNode* ProfMD = STI->getMetadata(LLVMContext::MD_prof); 776 for (unsigned i = 1; i < ProfMD->getNumOperands(); ++i) { 777 ConstantInt* CI = dyn_cast<ConstantInt>(ProfMD->getOperand(i)); 778 assert(CI); 779 Scale += CI->getValue().getZExtValue(); 780 } 781 782 // Skip default, as it's replaced during the folding 783 for (unsigned i = 1; i < Weights.size(); ++i) { 784 Weights[i] *= Scale; 785 } 786} 787 788/// Sees if any of the weights are too big for a uint32_t, and halves all the 789/// weights if any are. 790static void FitWeights(MutableArrayRef<uint64_t> Weights) { 791 bool Halve = false; 792 for (unsigned i = 0; i < Weights.size(); ++i) 793 if (Weights[i] > UINT_MAX) { 794 Halve = true; 795 break; 796 } 797 798 if (! Halve) 799 return; 800 801 for (unsigned i = 0; i < Weights.size(); ++i) 802 Weights[i] /= 2; 803} 804 805/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 806/// equality comparison instruction (either a switch or a branch on "X == c"). 807/// See if any of the predecessors of the terminator block are value comparisons 808/// on the same value. If so, and if safe to do so, fold them together. 809bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 810 IRBuilder<> &Builder) { 811 BasicBlock *BB = TI->getParent(); 812 Value *CV = isValueEqualityComparison(TI); // CondVal 813 assert(CV && "Not a comparison?"); 814 bool Changed = false; 815 816 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 817 while (!Preds.empty()) { 818 BasicBlock *Pred = Preds.pop_back_val(); 819 820 // See if the predecessor is a comparison with the same value. 821 TerminatorInst *PTI = Pred->getTerminator(); 822 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 823 824 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 825 // Figure out which 'cases' to copy from SI to PSI. 826 std::vector<ValueEqualityComparisonCase> BBCases; 827 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 828 829 std::vector<ValueEqualityComparisonCase> PredCases; 830 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 831 832 // Based on whether the default edge from PTI goes to BB or not, fill in 833 // PredCases and PredDefault with the new switch cases we would like to 834 // build. 835 SmallVector<BasicBlock*, 8> NewSuccessors; 836 837 // Update the branch weight metadata along the way 838 SmallVector<uint64_t, 8> Weights; 839 uint64_t PredDefaultWeight = 0; 840 bool PredHasWeights = HasBranchWeights(PTI); 841 bool SuccHasWeights = HasBranchWeights(TI); 842 843 if (PredHasWeights) { 844 MDNode* MD = PTI->getMetadata(LLVMContext::MD_prof); 845 assert(MD); 846 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 847 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 848 assert(CI); 849 Weights.push_back(CI->getValue().getZExtValue()); 850 } 851 852 // If the predecessor is a conditional eq, then swap the default weight 853 // to be the first entry. 854 if (BranchInst* BI = dyn_cast<BranchInst>(PTI)) { 855 assert(Weights.size() == 2); 856 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 857 858 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) { 859 std::swap(Weights.front(), Weights.back()); 860 } 861 } 862 863 PredDefaultWeight = Weights.front(); 864 } else if (SuccHasWeights) { 865 // If there are no predecessor weights but there are successor weights, 866 // populate Weights with 1, which will later be scaled to the sum of 867 // successor's weights 868 Weights.assign(1 + PredCases.size(), 1); 869 PredDefaultWeight = 1; 870 } 871 872 uint64_t SuccDefaultWeight = 0; 873 if (SuccHasWeights) { 874 int Index = 0; 875 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 876 ICmpInst* ICI = dyn_cast<ICmpInst>(BI->getCondition()); 877 assert(ICI); 878 879 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 880 Index = 1; 881 } 882 883 SuccDefaultWeight = GetWeight(TI, Index)->getValue().getZExtValue(); 884 } 885 886 if (PredDefault == BB) { 887 // If this is the default destination from PTI, only the edges in TI 888 // that don't occur in PTI, or that branch to BB will be activated. 889 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 890 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 891 if (PredCases[i].Dest != BB) 892 PTIHandled.insert(PredCases[i].Value); 893 else { 894 // The default destination is BB, we don't need explicit targets. 895 std::swap(PredCases[i], PredCases.back()); 896 897 if (PredHasWeights) { 898 std::swap(Weights[i+1], Weights.back()); 899 Weights.pop_back(); 900 } 901 902 PredCases.pop_back(); 903 --i; --e; 904 } 905 906 // Reconstruct the new switch statement we will be building. 907 if (PredDefault != BBDefault) { 908 PredDefault->removePredecessor(Pred); 909 PredDefault = BBDefault; 910 NewSuccessors.push_back(BBDefault); 911 } 912 913 if (SuccHasWeights) { 914 ScaleWeights(TI, Weights); 915 Weights.front() *= SuccDefaultWeight; 916 } else if (PredHasWeights) { 917 Weights.front() /= (1 + BBCases.size()); 918 } 919 920 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 921 if (!PTIHandled.count(BBCases[i].Value) && 922 BBCases[i].Dest != BBDefault) { 923 PredCases.push_back(BBCases[i]); 924 NewSuccessors.push_back(BBCases[i].Dest); 925 if (SuccHasWeights) { 926 Weights.push_back(PredDefaultWeight * 927 GetWeight(TI, i)->getValue().getZExtValue()); 928 } else if (PredHasWeights) { 929 // Split the old default's weight amongst the children 930 assert(PredDefaultWeight != 0); 931 Weights.push_back(PredDefaultWeight / (1 + BBCases.size())); 932 } 933 } 934 935 } else { 936 // FIXME: preserve branch weight metadata, similarly to the 'then' 937 // above. For now, drop it. 938 PredHasWeights = false; 939 SuccHasWeights = false; 940 941 // If this is not the default destination from PSI, only the edges 942 // in SI that occur in PSI with a destination of BB will be 943 // activated. 944 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 945 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 946 if (PredCases[i].Dest == BB) { 947 PTIHandled.insert(PredCases[i].Value); 948 std::swap(PredCases[i], PredCases.back()); 949 PredCases.pop_back(); 950 --i; --e; 951 } 952 953 // Okay, now we know which constants were sent to BB from the 954 // predecessor. Figure out where they will all go now. 955 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 956 if (PTIHandled.count(BBCases[i].Value)) { 957 // If this is one we are capable of getting... 958 PredCases.push_back(BBCases[i]); 959 NewSuccessors.push_back(BBCases[i].Dest); 960 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 961 } 962 963 // If there are any constants vectored to BB that TI doesn't handle, 964 // they must go to the default destination of TI. 965 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 966 PTIHandled.begin(), 967 E = PTIHandled.end(); I != E; ++I) { 968 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 969 NewSuccessors.push_back(BBDefault); 970 } 971 } 972 973 // Okay, at this point, we know which new successor Pred will get. Make 974 // sure we update the number of entries in the PHI nodes for these 975 // successors. 976 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 977 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 978 979 Builder.SetInsertPoint(PTI); 980 // Convert pointer to int before we switch. 981 if (CV->getType()->isPointerTy()) { 982 assert(TD && "Cannot switch on pointer without TargetData"); 983 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 984 "magicptr"); 985 } 986 987 // Now that the successors are updated, create the new Switch instruction. 988 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 989 PredCases.size()); 990 NewSI->setDebugLoc(PTI->getDebugLoc()); 991 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 992 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 993 994 if (PredHasWeights || SuccHasWeights) { 995 // Halve the weights if any of them cannot fit in an uint32_t 996 FitWeights(Weights); 997 998 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 999 1000 NewSI->setMetadata(LLVMContext::MD_prof, 1001 MDBuilder(BB->getContext()). 1002 createBranchWeights(MDWeights)); 1003 } 1004 1005 EraseTerminatorInstAndDCECond(PTI); 1006 1007 // Okay, last check. If BB is still a successor of PSI, then we must 1008 // have an infinite loop case. If so, add an infinitely looping block 1009 // to handle the case to preserve the behavior of the code. 1010 BasicBlock *InfLoopBlock = 0; 1011 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1012 if (NewSI->getSuccessor(i) == BB) { 1013 if (InfLoopBlock == 0) { 1014 // Insert it at the end of the function, because it's either code, 1015 // or it won't matter if it's hot. :) 1016 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1017 "infloop", BB->getParent()); 1018 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1019 } 1020 NewSI->setSuccessor(i, InfLoopBlock); 1021 } 1022 1023 Changed = true; 1024 } 1025 } 1026 return Changed; 1027} 1028 1029// isSafeToHoistInvoke - If we would need to insert a select that uses the 1030// value of this invoke (comments in HoistThenElseCodeToIf explain why we 1031// would need to do this), we can't hoist the invoke, as there is nowhere 1032// to put the select in this case. 1033static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1034 Instruction *I1, Instruction *I2) { 1035 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1036 PHINode *PN; 1037 for (BasicBlock::iterator BBI = SI->begin(); 1038 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1039 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1040 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1041 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1042 return false; 1043 } 1044 } 1045 } 1046 return true; 1047} 1048 1049/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1050/// BB2, hoist any common code in the two blocks up into the branch block. The 1051/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1052static bool HoistThenElseCodeToIf(BranchInst *BI) { 1053 // This does very trivial matching, with limited scanning, to find identical 1054 // instructions in the two blocks. In particular, we don't want to get into 1055 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1056 // such, we currently just scan for obviously identical instructions in an 1057 // identical order. 1058 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1059 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1060 1061 BasicBlock::iterator BB1_Itr = BB1->begin(); 1062 BasicBlock::iterator BB2_Itr = BB2->begin(); 1063 1064 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1065 // Skip debug info if it is not identical. 1066 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1067 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1068 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1069 while (isa<DbgInfoIntrinsic>(I1)) 1070 I1 = BB1_Itr++; 1071 while (isa<DbgInfoIntrinsic>(I2)) 1072 I2 = BB2_Itr++; 1073 } 1074 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1075 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1076 return false; 1077 1078 // If we get here, we can hoist at least one instruction. 1079 BasicBlock *BIParent = BI->getParent(); 1080 1081 do { 1082 // If we are hoisting the terminator instruction, don't move one (making a 1083 // broken BB), instead clone it, and remove BI. 1084 if (isa<TerminatorInst>(I1)) 1085 goto HoistTerminator; 1086 1087 // For a normal instruction, we just move one to right before the branch, 1088 // then replace all uses of the other with the first. Finally, we remove 1089 // the now redundant second instruction. 1090 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1091 if (!I2->use_empty()) 1092 I2->replaceAllUsesWith(I1); 1093 I1->intersectOptionalDataWith(I2); 1094 I2->eraseFromParent(); 1095 1096 I1 = BB1_Itr++; 1097 I2 = BB2_Itr++; 1098 // Skip debug info if it is not identical. 1099 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1100 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1101 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1102 while (isa<DbgInfoIntrinsic>(I1)) 1103 I1 = BB1_Itr++; 1104 while (isa<DbgInfoIntrinsic>(I2)) 1105 I2 = BB2_Itr++; 1106 } 1107 } while (I1->isIdenticalToWhenDefined(I2)); 1108 1109 return true; 1110 1111HoistTerminator: 1112 // It may not be possible to hoist an invoke. 1113 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1114 return true; 1115 1116 // Okay, it is safe to hoist the terminator. 1117 Instruction *NT = I1->clone(); 1118 BIParent->getInstList().insert(BI, NT); 1119 if (!NT->getType()->isVoidTy()) { 1120 I1->replaceAllUsesWith(NT); 1121 I2->replaceAllUsesWith(NT); 1122 NT->takeName(I1); 1123 } 1124 1125 IRBuilder<true, NoFolder> Builder(NT); 1126 // Hoisting one of the terminators from our successor is a great thing. 1127 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1128 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1129 // nodes, so we insert select instruction to compute the final result. 1130 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1131 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1132 PHINode *PN; 1133 for (BasicBlock::iterator BBI = SI->begin(); 1134 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1135 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1136 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1137 if (BB1V == BB2V) continue; 1138 1139 // These values do not agree. Insert a select instruction before NT 1140 // that determines the right value. 1141 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1142 if (SI == 0) 1143 SI = cast<SelectInst> 1144 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1145 BB1V->getName()+"."+BB2V->getName())); 1146 1147 // Make the PHI node use the select for all incoming values for BB1/BB2 1148 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1149 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1150 PN->setIncomingValue(i, SI); 1151 } 1152 } 1153 1154 // Update any PHI nodes in our new successors. 1155 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1156 AddPredecessorToBlock(*SI, BIParent, BB1); 1157 1158 EraseTerminatorInstAndDCECond(BI); 1159 return true; 1160} 1161 1162/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 1163/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 1164/// (for now, restricted to a single instruction that's side effect free) from 1165/// the BB1 into the branch block to speculatively execute it. 1166/// 1167/// Turn 1168/// BB: 1169/// %t1 = icmp 1170/// br i1 %t1, label %BB1, label %BB2 1171/// BB1: 1172/// %t3 = add %t2, c 1173/// br label BB2 1174/// BB2: 1175/// => 1176/// BB: 1177/// %t1 = icmp 1178/// %t4 = add %t2, c 1179/// %t3 = select i1 %t1, %t2, %t3 1180static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 1181 // Only speculatively execution a single instruction (not counting the 1182 // terminator) for now. 1183 Instruction *HInst = NULL; 1184 Instruction *Term = BB1->getTerminator(); 1185 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 1186 BBI != BBE; ++BBI) { 1187 Instruction *I = BBI; 1188 // Skip debug info. 1189 if (isa<DbgInfoIntrinsic>(I)) continue; 1190 if (I == Term) break; 1191 1192 if (HInst) 1193 return false; 1194 HInst = I; 1195 } 1196 1197 BasicBlock *BIParent = BI->getParent(); 1198 1199 // Check the instruction to be hoisted, if there is one. 1200 if (HInst) { 1201 // Don't hoist the instruction if it's unsafe or expensive. 1202 if (!isSafeToSpeculativelyExecute(HInst)) 1203 return false; 1204 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold) 1205 return false; 1206 1207 // Do not hoist the instruction if any of its operands are defined but not 1208 // used in this BB. The transformation will prevent the operand from 1209 // being sunk into the use block. 1210 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1211 i != e; ++i) { 1212 Instruction *OpI = dyn_cast<Instruction>(*i); 1213 if (OpI && OpI->getParent() == BIParent && 1214 !OpI->mayHaveSideEffects() && 1215 !OpI->isUsedInBasicBlock(BIParent)) 1216 return false; 1217 } 1218 } 1219 1220 // Be conservative for now. FP select instruction can often be expensive. 1221 Value *BrCond = BI->getCondition(); 1222 if (isa<FCmpInst>(BrCond)) 1223 return false; 1224 1225 // If BB1 is actually on the false edge of the conditional branch, remember 1226 // to swap the select operands later. 1227 bool Invert = false; 1228 if (BB1 != BI->getSuccessor(0)) { 1229 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1230 Invert = true; 1231 } 1232 1233 // Collect interesting PHIs, and scan for hazards. 1234 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs; 1235 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1236 for (BasicBlock::iterator I = BB2->begin(); 1237 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1238 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1239 Value *BIParentV = PN->getIncomingValueForBlock(BIParent); 1240 1241 // Skip PHIs which are trivial. 1242 if (BB1V == BIParentV) 1243 continue; 1244 1245 // Check for saftey. 1246 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) { 1247 // An unfolded ConstantExpr could end up getting expanded into 1248 // Instructions. Don't speculate this and another instruction at 1249 // the same time. 1250 if (HInst) 1251 return false; 1252 if (!isSafeToSpeculativelyExecute(CE)) 1253 return false; 1254 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1255 return false; 1256 } 1257 1258 // Ok, we may insert a select for this PHI. 1259 PHIs.insert(std::make_pair(BB1V, BIParentV)); 1260 } 1261 1262 // If there are no PHIs to process, bail early. This helps ensure idempotence 1263 // as well. 1264 if (PHIs.empty()) 1265 return false; 1266 1267 // If we get here, we can hoist the instruction and if-convert. 1268 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); 1269 1270 // Hoist the instruction. 1271 if (HInst) 1272 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst); 1273 1274 // Insert selects and rewrite the PHI operands. 1275 IRBuilder<true, NoFolder> Builder(BI); 1276 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1277 Value *TrueV = PHIs[i].first; 1278 Value *FalseV = PHIs[i].second; 1279 1280 // Create a select whose true value is the speculatively executed value and 1281 // false value is the previously determined FalseV. 1282 SelectInst *SI; 1283 if (Invert) 1284 SI = cast<SelectInst> 1285 (Builder.CreateSelect(BrCond, FalseV, TrueV, 1286 FalseV->getName() + "." + TrueV->getName())); 1287 else 1288 SI = cast<SelectInst> 1289 (Builder.CreateSelect(BrCond, TrueV, FalseV, 1290 TrueV->getName() + "." + FalseV->getName())); 1291 1292 // Make the PHI node use the select for all incoming values for "then" and 1293 // "if" blocks. 1294 for (BasicBlock::iterator I = BB2->begin(); 1295 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1296 unsigned BB1I = PN->getBasicBlockIndex(BB1); 1297 unsigned BIParentI = PN->getBasicBlockIndex(BIParent); 1298 Value *BB1V = PN->getIncomingValue(BB1I); 1299 Value *BIParentV = PN->getIncomingValue(BIParentI); 1300 if (TrueV == BB1V && FalseV == BIParentV) { 1301 PN->setIncomingValue(BB1I, SI); 1302 PN->setIncomingValue(BIParentI, SI); 1303 } 1304 } 1305 } 1306 1307 ++NumSpeculations; 1308 return true; 1309} 1310 1311/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1312/// across this block. 1313static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1314 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1315 unsigned Size = 0; 1316 1317 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1318 if (isa<DbgInfoIntrinsic>(BBI)) 1319 continue; 1320 if (Size > 10) return false; // Don't clone large BB's. 1321 ++Size; 1322 1323 // We can only support instructions that do not define values that are 1324 // live outside of the current basic block. 1325 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1326 UI != E; ++UI) { 1327 Instruction *U = cast<Instruction>(*UI); 1328 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1329 } 1330 1331 // Looks ok, continue checking. 1332 } 1333 1334 return true; 1335} 1336 1337/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1338/// that is defined in the same block as the branch and if any PHI entries are 1339/// constants, thread edges corresponding to that entry to be branches to their 1340/// ultimate destination. 1341static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { 1342 BasicBlock *BB = BI->getParent(); 1343 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1344 // NOTE: we currently cannot transform this case if the PHI node is used 1345 // outside of the block. 1346 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1347 return false; 1348 1349 // Degenerate case of a single entry PHI. 1350 if (PN->getNumIncomingValues() == 1) { 1351 FoldSingleEntryPHINodes(PN->getParent()); 1352 return true; 1353 } 1354 1355 // Now we know that this block has multiple preds and two succs. 1356 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1357 1358 // Okay, this is a simple enough basic block. See if any phi values are 1359 // constants. 1360 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1361 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1362 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1363 1364 // Okay, we now know that all edges from PredBB should be revectored to 1365 // branch to RealDest. 1366 BasicBlock *PredBB = PN->getIncomingBlock(i); 1367 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1368 1369 if (RealDest == BB) continue; // Skip self loops. 1370 // Skip if the predecessor's terminator is an indirect branch. 1371 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1372 1373 // The dest block might have PHI nodes, other predecessors and other 1374 // difficult cases. Instead of being smart about this, just insert a new 1375 // block that jumps to the destination block, effectively splitting 1376 // the edge we are about to create. 1377 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1378 RealDest->getName()+".critedge", 1379 RealDest->getParent(), RealDest); 1380 BranchInst::Create(RealDest, EdgeBB); 1381 1382 // Update PHI nodes. 1383 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1384 1385 // BB may have instructions that are being threaded over. Clone these 1386 // instructions into EdgeBB. We know that there will be no uses of the 1387 // cloned instructions outside of EdgeBB. 1388 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1389 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1390 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1391 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1392 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1393 continue; 1394 } 1395 // Clone the instruction. 1396 Instruction *N = BBI->clone(); 1397 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1398 1399 // Update operands due to translation. 1400 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1401 i != e; ++i) { 1402 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1403 if (PI != TranslateMap.end()) 1404 *i = PI->second; 1405 } 1406 1407 // Check for trivial simplification. 1408 if (Value *V = SimplifyInstruction(N, TD)) { 1409 TranslateMap[BBI] = V; 1410 delete N; // Instruction folded away, don't need actual inst 1411 } else { 1412 // Insert the new instruction into its new home. 1413 EdgeBB->getInstList().insert(InsertPt, N); 1414 if (!BBI->use_empty()) 1415 TranslateMap[BBI] = N; 1416 } 1417 } 1418 1419 // Loop over all of the edges from PredBB to BB, changing them to branch 1420 // to EdgeBB instead. 1421 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1422 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1423 if (PredBBTI->getSuccessor(i) == BB) { 1424 BB->removePredecessor(PredBB); 1425 PredBBTI->setSuccessor(i, EdgeBB); 1426 } 1427 1428 // Recurse, simplifying any other constants. 1429 return FoldCondBranchOnPHI(BI, TD) | true; 1430 } 1431 1432 return false; 1433} 1434 1435/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1436/// PHI node, see if we can eliminate it. 1437static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { 1438 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1439 // statement", which has a very simple dominance structure. Basically, we 1440 // are trying to find the condition that is being branched on, which 1441 // subsequently causes this merge to happen. We really want control 1442 // dependence information for this check, but simplifycfg can't keep it up 1443 // to date, and this catches most of the cases we care about anyway. 1444 BasicBlock *BB = PN->getParent(); 1445 BasicBlock *IfTrue, *IfFalse; 1446 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1447 if (!IfCond || 1448 // Don't bother if the branch will be constant folded trivially. 1449 isa<ConstantInt>(IfCond)) 1450 return false; 1451 1452 // Okay, we found that we can merge this two-entry phi node into a select. 1453 // Doing so would require us to fold *all* two entry phi nodes in this block. 1454 // At some point this becomes non-profitable (particularly if the target 1455 // doesn't support cmov's). Only do this transformation if there are two or 1456 // fewer PHI nodes in this block. 1457 unsigned NumPhis = 0; 1458 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1459 if (NumPhis > 2) 1460 return false; 1461 1462 // Loop over the PHI's seeing if we can promote them all to select 1463 // instructions. While we are at it, keep track of the instructions 1464 // that need to be moved to the dominating block. 1465 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1466 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1467 MaxCostVal1 = PHINodeFoldingThreshold; 1468 1469 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1470 PHINode *PN = cast<PHINode>(II++); 1471 if (Value *V = SimplifyInstruction(PN, TD)) { 1472 PN->replaceAllUsesWith(V); 1473 PN->eraseFromParent(); 1474 continue; 1475 } 1476 1477 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1478 MaxCostVal0) || 1479 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1480 MaxCostVal1)) 1481 return false; 1482 } 1483 1484 // If we folded the first phi, PN dangles at this point. Refresh it. If 1485 // we ran out of PHIs then we simplified them all. 1486 PN = dyn_cast<PHINode>(BB->begin()); 1487 if (PN == 0) return true; 1488 1489 // Don't fold i1 branches on PHIs which contain binary operators. These can 1490 // often be turned into switches and other things. 1491 if (PN->getType()->isIntegerTy(1) && 1492 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1493 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1494 isa<BinaryOperator>(IfCond))) 1495 return false; 1496 1497 // If we all PHI nodes are promotable, check to make sure that all 1498 // instructions in the predecessor blocks can be promoted as well. If 1499 // not, we won't be able to get rid of the control flow, so it's not 1500 // worth promoting to select instructions. 1501 BasicBlock *DomBlock = 0; 1502 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1503 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1504 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1505 IfBlock1 = 0; 1506 } else { 1507 DomBlock = *pred_begin(IfBlock1); 1508 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1509 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1510 // This is not an aggressive instruction that we can promote. 1511 // Because of this, we won't be able to get rid of the control 1512 // flow, so the xform is not worth it. 1513 return false; 1514 } 1515 } 1516 1517 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1518 IfBlock2 = 0; 1519 } else { 1520 DomBlock = *pred_begin(IfBlock2); 1521 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1522 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1523 // This is not an aggressive instruction that we can promote. 1524 // Because of this, we won't be able to get rid of the control 1525 // flow, so the xform is not worth it. 1526 return false; 1527 } 1528 } 1529 1530 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1531 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1532 1533 // If we can still promote the PHI nodes after this gauntlet of tests, 1534 // do all of the PHI's now. 1535 Instruction *InsertPt = DomBlock->getTerminator(); 1536 IRBuilder<true, NoFolder> Builder(InsertPt); 1537 1538 // Move all 'aggressive' instructions, which are defined in the 1539 // conditional parts of the if's up to the dominating block. 1540 if (IfBlock1) 1541 DomBlock->getInstList().splice(InsertPt, 1542 IfBlock1->getInstList(), IfBlock1->begin(), 1543 IfBlock1->getTerminator()); 1544 if (IfBlock2) 1545 DomBlock->getInstList().splice(InsertPt, 1546 IfBlock2->getInstList(), IfBlock2->begin(), 1547 IfBlock2->getTerminator()); 1548 1549 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1550 // Change the PHI node into a select instruction. 1551 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1552 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1553 1554 SelectInst *NV = 1555 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1556 PN->replaceAllUsesWith(NV); 1557 NV->takeName(PN); 1558 PN->eraseFromParent(); 1559 } 1560 1561 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1562 // has been flattened. Change DomBlock to jump directly to our new block to 1563 // avoid other simplifycfg's kicking in on the diamond. 1564 TerminatorInst *OldTI = DomBlock->getTerminator(); 1565 Builder.SetInsertPoint(OldTI); 1566 Builder.CreateBr(BB); 1567 OldTI->eraseFromParent(); 1568 return true; 1569} 1570 1571/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1572/// to two returning blocks, try to merge them together into one return, 1573/// introducing a select if the return values disagree. 1574static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1575 IRBuilder<> &Builder) { 1576 assert(BI->isConditional() && "Must be a conditional branch"); 1577 BasicBlock *TrueSucc = BI->getSuccessor(0); 1578 BasicBlock *FalseSucc = BI->getSuccessor(1); 1579 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1580 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1581 1582 // Check to ensure both blocks are empty (just a return) or optionally empty 1583 // with PHI nodes. If there are other instructions, merging would cause extra 1584 // computation on one path or the other. 1585 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1586 return false; 1587 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1588 return false; 1589 1590 Builder.SetInsertPoint(BI); 1591 // Okay, we found a branch that is going to two return nodes. If 1592 // there is no return value for this function, just change the 1593 // branch into a return. 1594 if (FalseRet->getNumOperands() == 0) { 1595 TrueSucc->removePredecessor(BI->getParent()); 1596 FalseSucc->removePredecessor(BI->getParent()); 1597 Builder.CreateRetVoid(); 1598 EraseTerminatorInstAndDCECond(BI); 1599 return true; 1600 } 1601 1602 // Otherwise, figure out what the true and false return values are 1603 // so we can insert a new select instruction. 1604 Value *TrueValue = TrueRet->getReturnValue(); 1605 Value *FalseValue = FalseRet->getReturnValue(); 1606 1607 // Unwrap any PHI nodes in the return blocks. 1608 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1609 if (TVPN->getParent() == TrueSucc) 1610 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1611 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1612 if (FVPN->getParent() == FalseSucc) 1613 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1614 1615 // In order for this transformation to be safe, we must be able to 1616 // unconditionally execute both operands to the return. This is 1617 // normally the case, but we could have a potentially-trapping 1618 // constant expression that prevents this transformation from being 1619 // safe. 1620 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1621 if (TCV->canTrap()) 1622 return false; 1623 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1624 if (FCV->canTrap()) 1625 return false; 1626 1627 // Okay, we collected all the mapped values and checked them for sanity, and 1628 // defined to really do this transformation. First, update the CFG. 1629 TrueSucc->removePredecessor(BI->getParent()); 1630 FalseSucc->removePredecessor(BI->getParent()); 1631 1632 // Insert select instructions where needed. 1633 Value *BrCond = BI->getCondition(); 1634 if (TrueValue) { 1635 // Insert a select if the results differ. 1636 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1637 } else if (isa<UndefValue>(TrueValue)) { 1638 TrueValue = FalseValue; 1639 } else { 1640 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1641 FalseValue, "retval"); 1642 } 1643 } 1644 1645 Value *RI = !TrueValue ? 1646 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1647 1648 (void) RI; 1649 1650 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1651 << "\n " << *BI << "NewRet = " << *RI 1652 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1653 1654 EraseTerminatorInstAndDCECond(BI); 1655 1656 return true; 1657} 1658 1659/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1660/// probabilities of the branch taking each edge. Fills in the two APInt 1661/// parameters and return true, or returns false if no or invalid metadata was 1662/// found. 1663static bool ExtractBranchMetadata(BranchInst *BI, 1664 APInt &ProbTrue, APInt &ProbFalse) { 1665 assert(BI->isConditional() && 1666 "Looking for probabilities on unconditional branch?"); 1667 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1668 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1669 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1670 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1671 if (!CITrue || !CIFalse) return false; 1672 ProbTrue = CITrue->getValue(); 1673 ProbFalse = CIFalse->getValue(); 1674 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 && 1675 "Branch probability metadata must be 32-bit integers"); 1676 return true; 1677} 1678 1679/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In 1680/// the event of overflow, logically-shifts all four inputs right until the 1681/// multiply fits. 1682static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D, 1683 unsigned &BitsLost) { 1684 BitsLost = 0; 1685 bool Overflow = false; 1686 APInt Result = A.umul_ov(B, Overflow); 1687 if (Overflow) { 1688 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A); 1689 do { 1690 B = B.lshr(1); 1691 ++BitsLost; 1692 } while (B.ugt(MaxB)); 1693 A = A.lshr(BitsLost); 1694 C = C.lshr(BitsLost); 1695 D = D.lshr(BitsLost); 1696 Result = A * B; 1697 } 1698 return Result; 1699} 1700 1701/// checkCSEInPredecessor - Return true if the given instruction is available 1702/// in its predecessor block. If yes, the instruction will be removed. 1703/// 1704static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1705 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1706 return false; 1707 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1708 Instruction *PBI = &*I; 1709 // Check whether Inst and PBI generate the same value. 1710 if (Inst->isIdenticalTo(PBI)) { 1711 Inst->replaceAllUsesWith(PBI); 1712 Inst->eraseFromParent(); 1713 return true; 1714 } 1715 } 1716 return false; 1717} 1718 1719/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1720/// predecessor branches to us and one of our successors, fold the block into 1721/// the predecessor and use logical operations to pick the right destination. 1722bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1723 BasicBlock *BB = BI->getParent(); 1724 1725 Instruction *Cond = 0; 1726 if (BI->isConditional()) 1727 Cond = dyn_cast<Instruction>(BI->getCondition()); 1728 else { 1729 // For unconditional branch, check for a simple CFG pattern, where 1730 // BB has a single predecessor and BB's successor is also its predecessor's 1731 // successor. If such pattern exisits, check for CSE between BB and its 1732 // predecessor. 1733 if (BasicBlock *PB = BB->getSinglePredecessor()) 1734 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1735 if (PBI->isConditional() && 1736 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1737 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1738 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1739 I != E; ) { 1740 Instruction *Curr = I++; 1741 if (isa<CmpInst>(Curr)) { 1742 Cond = Curr; 1743 break; 1744 } 1745 // Quit if we can't remove this instruction. 1746 if (!checkCSEInPredecessor(Curr, PB)) 1747 return false; 1748 } 1749 } 1750 1751 if (Cond == 0) 1752 return false; 1753 } 1754 1755 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1756 Cond->getParent() != BB || !Cond->hasOneUse()) 1757 return false; 1758 1759 // Only allow this if the condition is a simple instruction that can be 1760 // executed unconditionally. It must be in the same block as the branch, and 1761 // must be at the front of the block. 1762 BasicBlock::iterator FrontIt = BB->front(); 1763 1764 // Ignore dbg intrinsics. 1765 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1766 1767 // Allow a single instruction to be hoisted in addition to the compare 1768 // that feeds the branch. We later ensure that any values that _it_ uses 1769 // were also live in the predecessor, so that we don't unnecessarily create 1770 // register pressure or inhibit out-of-order execution. 1771 Instruction *BonusInst = 0; 1772 if (&*FrontIt != Cond && 1773 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1774 isSafeToSpeculativelyExecute(FrontIt)) { 1775 BonusInst = &*FrontIt; 1776 ++FrontIt; 1777 1778 // Ignore dbg intrinsics. 1779 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1780 } 1781 1782 // Only a single bonus inst is allowed. 1783 if (&*FrontIt != Cond) 1784 return false; 1785 1786 // Make sure the instruction after the condition is the cond branch. 1787 BasicBlock::iterator CondIt = Cond; ++CondIt; 1788 1789 // Ingore dbg intrinsics. 1790 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1791 1792 if (&*CondIt != BI) 1793 return false; 1794 1795 // Cond is known to be a compare or binary operator. Check to make sure that 1796 // neither operand is a potentially-trapping constant expression. 1797 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1798 if (CE->canTrap()) 1799 return false; 1800 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1801 if (CE->canTrap()) 1802 return false; 1803 1804 // Finally, don't infinitely unroll conditional loops. 1805 BasicBlock *TrueDest = BI->getSuccessor(0); 1806 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 1807 if (TrueDest == BB || FalseDest == BB) 1808 return false; 1809 1810 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1811 BasicBlock *PredBlock = *PI; 1812 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1813 1814 // Check that we have two conditional branches. If there is a PHI node in 1815 // the common successor, verify that the same value flows in from both 1816 // blocks. 1817 SmallVector<PHINode*, 4> PHIs; 1818 if (PBI == 0 || PBI->isUnconditional() || 1819 (BI->isConditional() && 1820 !SafeToMergeTerminators(BI, PBI)) || 1821 (!BI->isConditional() && 1822 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 1823 continue; 1824 1825 // Determine if the two branches share a common destination. 1826 Instruction::BinaryOps Opc; 1827 bool InvertPredCond = false; 1828 1829 if (BI->isConditional()) { 1830 if (PBI->getSuccessor(0) == TrueDest) 1831 Opc = Instruction::Or; 1832 else if (PBI->getSuccessor(1) == FalseDest) 1833 Opc = Instruction::And; 1834 else if (PBI->getSuccessor(0) == FalseDest) 1835 Opc = Instruction::And, InvertPredCond = true; 1836 else if (PBI->getSuccessor(1) == TrueDest) 1837 Opc = Instruction::Or, InvertPredCond = true; 1838 else 1839 continue; 1840 } else { 1841 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 1842 continue; 1843 } 1844 1845 // Ensure that any values used in the bonus instruction are also used 1846 // by the terminator of the predecessor. This means that those values 1847 // must already have been resolved, so we won't be inhibiting the 1848 // out-of-order core by speculating them earlier. 1849 if (BonusInst) { 1850 // Collect the values used by the bonus inst 1851 SmallPtrSet<Value*, 4> UsedValues; 1852 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1853 OE = BonusInst->op_end(); OI != OE; ++OI) { 1854 Value *V = *OI; 1855 if (!isa<Constant>(V)) 1856 UsedValues.insert(V); 1857 } 1858 1859 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1860 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1861 1862 // Walk up to four levels back up the use-def chain of the predecessor's 1863 // terminator to see if all those values were used. The choice of four 1864 // levels is arbitrary, to provide a compile-time-cost bound. 1865 while (!Worklist.empty()) { 1866 std::pair<Value*, unsigned> Pair = Worklist.back(); 1867 Worklist.pop_back(); 1868 1869 if (Pair.second >= 4) continue; 1870 UsedValues.erase(Pair.first); 1871 if (UsedValues.empty()) break; 1872 1873 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 1874 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1875 OI != OE; ++OI) 1876 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1877 } 1878 } 1879 1880 if (!UsedValues.empty()) return false; 1881 } 1882 1883 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1884 IRBuilder<> Builder(PBI); 1885 1886 // If we need to invert the condition in the pred block to match, do so now. 1887 if (InvertPredCond) { 1888 Value *NewCond = PBI->getCondition(); 1889 1890 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1891 CmpInst *CI = cast<CmpInst>(NewCond); 1892 CI->setPredicate(CI->getInversePredicate()); 1893 } else { 1894 NewCond = Builder.CreateNot(NewCond, 1895 PBI->getCondition()->getName()+".not"); 1896 } 1897 1898 PBI->setCondition(NewCond); 1899 PBI->swapSuccessors(); 1900 } 1901 1902 // If we have a bonus inst, clone it into the predecessor block. 1903 Instruction *NewBonus = 0; 1904 if (BonusInst) { 1905 NewBonus = BonusInst->clone(); 1906 PredBlock->getInstList().insert(PBI, NewBonus); 1907 NewBonus->takeName(BonusInst); 1908 BonusInst->setName(BonusInst->getName()+".old"); 1909 } 1910 1911 // Clone Cond into the predecessor basic block, and or/and the 1912 // two conditions together. 1913 Instruction *New = Cond->clone(); 1914 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1915 PredBlock->getInstList().insert(PBI, New); 1916 New->takeName(Cond); 1917 Cond->setName(New->getName()+".old"); 1918 1919 if (BI->isConditional()) { 1920 Instruction *NewCond = 1921 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 1922 New, "or.cond")); 1923 PBI->setCondition(NewCond); 1924 1925 if (PBI->getSuccessor(0) == BB) { 1926 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1927 PBI->setSuccessor(0, TrueDest); 1928 } 1929 if (PBI->getSuccessor(1) == BB) { 1930 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1931 PBI->setSuccessor(1, FalseDest); 1932 } 1933 } else { 1934 // Update PHI nodes in the common successors. 1935 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1936 ConstantInt *PBI_C = cast<ConstantInt>( 1937 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 1938 assert(PBI_C->getType()->isIntegerTy(1)); 1939 Instruction *MergedCond = 0; 1940 if (PBI->getSuccessor(0) == TrueDest) { 1941 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 1942 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 1943 // is false: !PBI_Cond and BI_Value 1944 Instruction *NotCond = 1945 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 1946 "not.cond")); 1947 MergedCond = 1948 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 1949 NotCond, New, 1950 "and.cond")); 1951 if (PBI_C->isOne()) 1952 MergedCond = 1953 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 1954 PBI->getCondition(), MergedCond, 1955 "or.cond")); 1956 } else { 1957 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 1958 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 1959 // is false: PBI_Cond and BI_Value 1960 MergedCond = 1961 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 1962 PBI->getCondition(), New, 1963 "and.cond")); 1964 if (PBI_C->isOne()) { 1965 Instruction *NotCond = 1966 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 1967 "not.cond")); 1968 MergedCond = 1969 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 1970 NotCond, MergedCond, 1971 "or.cond")); 1972 } 1973 } 1974 // Update PHI Node. 1975 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 1976 MergedCond); 1977 } 1978 // Change PBI from Conditional to Unconditional. 1979 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 1980 EraseTerminatorInstAndDCECond(PBI); 1981 PBI = New_PBI; 1982 } 1983 1984 // TODO: If BB is reachable from all paths through PredBlock, then we 1985 // could replace PBI's branch probabilities with BI's. 1986 1987 // Merge probability data into PredBlock's branch. 1988 APInt A, B, C, D; 1989 if (PBI->isConditional() && BI->isConditional() && 1990 ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) { 1991 // Given IR which does: 1992 // bbA: 1993 // br i1 %x, label %bbB, label %bbC 1994 // bbB: 1995 // br i1 %y, label %bbD, label %bbC 1996 // Let's call the probability that we take the edge from %bbA to %bbB 1997 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to 1998 // %bbC probability 'd'. 1999 // 2000 // We transform the IR into: 2001 // bbA: 2002 // br i1 %z, label %bbD, label %bbC 2003 // where the probability of going to %bbD is (a*c) and going to bbC is 2004 // (b+a*d). 2005 // 2006 // Probabilities aren't stored as ratios directly. Using branch weights, 2007 // we get: 2008 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D. 2009 2010 // In the event of overflow, we want to drop the LSB of the input 2011 // probabilities. 2012 unsigned BitsLost; 2013 2014 // Ignore overflow result on ProbTrue. 2015 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost); 2016 2017 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost); 2018 if (BitsLost) { 2019 ProbTrue = ProbTrue.lshr(BitsLost*2); 2020 } 2021 2022 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost); 2023 if (BitsLost) { 2024 ProbTrue = ProbTrue.lshr(BitsLost*2); 2025 Tmp1 = Tmp1.lshr(BitsLost*2); 2026 } 2027 2028 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost); 2029 if (BitsLost) { 2030 ProbTrue = ProbTrue.lshr(BitsLost*2); 2031 Tmp1 = Tmp1.lshr(BitsLost*2); 2032 Tmp2 = Tmp2.lshr(BitsLost*2); 2033 } 2034 2035 bool Overflow1 = false, Overflow2 = false; 2036 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1); 2037 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2); 2038 2039 if (Overflow1 || Overflow2) { 2040 ProbTrue = ProbTrue.lshr(1); 2041 Tmp1 = Tmp1.lshr(1); 2042 Tmp2 = Tmp2.lshr(1); 2043 Tmp3 = Tmp3.lshr(1); 2044 Tmp4 = Tmp2 + Tmp3; 2045 ProbFalse = Tmp4 + Tmp1; 2046 } 2047 2048 // The sum of branch weights must fit in 32-bits. 2049 if (ProbTrue.isNegative() && ProbFalse.isNegative()) { 2050 ProbTrue = ProbTrue.lshr(1); 2051 ProbFalse = ProbFalse.lshr(1); 2052 } 2053 2054 if (ProbTrue != ProbFalse) { 2055 // Normalize the result. 2056 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse); 2057 ProbTrue = ProbTrue.udiv(GCD); 2058 ProbFalse = ProbFalse.udiv(GCD); 2059 2060 MDBuilder MDB(BI->getContext()); 2061 MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(), 2062 ProbFalse.getZExtValue()); 2063 PBI->setMetadata(LLVMContext::MD_prof, N); 2064 } else { 2065 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2066 } 2067 } else { 2068 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2069 } 2070 2071 // Copy any debug value intrinsics into the end of PredBlock. 2072 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2073 if (isa<DbgInfoIntrinsic>(*I)) 2074 I->clone()->insertBefore(PBI); 2075 2076 return true; 2077 } 2078 return false; 2079} 2080 2081/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2082/// predecessor of another block, this function tries to simplify it. We know 2083/// that PBI and BI are both conditional branches, and BI is in one of the 2084/// successor blocks of PBI - PBI branches to BI. 2085static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2086 assert(PBI->isConditional() && BI->isConditional()); 2087 BasicBlock *BB = BI->getParent(); 2088 2089 // If this block ends with a branch instruction, and if there is a 2090 // predecessor that ends on a branch of the same condition, make 2091 // this conditional branch redundant. 2092 if (PBI->getCondition() == BI->getCondition() && 2093 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2094 // Okay, the outcome of this conditional branch is statically 2095 // knowable. If this block had a single pred, handle specially. 2096 if (BB->getSinglePredecessor()) { 2097 // Turn this into a branch on constant. 2098 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2099 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2100 CondIsTrue)); 2101 return true; // Nuke the branch on constant. 2102 } 2103 2104 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2105 // in the constant and simplify the block result. Subsequent passes of 2106 // simplifycfg will thread the block. 2107 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2108 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2109 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2110 std::distance(PB, PE), 2111 BI->getCondition()->getName() + ".pr", 2112 BB->begin()); 2113 // Okay, we're going to insert the PHI node. Since PBI is not the only 2114 // predecessor, compute the PHI'd conditional value for all of the preds. 2115 // Any predecessor where the condition is not computable we keep symbolic. 2116 for (pred_iterator PI = PB; PI != PE; ++PI) { 2117 BasicBlock *P = *PI; 2118 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2119 PBI != BI && PBI->isConditional() && 2120 PBI->getCondition() == BI->getCondition() && 2121 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2122 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2123 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2124 CondIsTrue), P); 2125 } else { 2126 NewPN->addIncoming(BI->getCondition(), P); 2127 } 2128 } 2129 2130 BI->setCondition(NewPN); 2131 return true; 2132 } 2133 } 2134 2135 // If this is a conditional branch in an empty block, and if any 2136 // predecessors is a conditional branch to one of our destinations, 2137 // fold the conditions into logical ops and one cond br. 2138 BasicBlock::iterator BBI = BB->begin(); 2139 // Ignore dbg intrinsics. 2140 while (isa<DbgInfoIntrinsic>(BBI)) 2141 ++BBI; 2142 if (&*BBI != BI) 2143 return false; 2144 2145 2146 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2147 if (CE->canTrap()) 2148 return false; 2149 2150 int PBIOp, BIOp; 2151 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2152 PBIOp = BIOp = 0; 2153 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2154 PBIOp = 0, BIOp = 1; 2155 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2156 PBIOp = 1, BIOp = 0; 2157 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2158 PBIOp = BIOp = 1; 2159 else 2160 return false; 2161 2162 // Check to make sure that the other destination of this branch 2163 // isn't BB itself. If so, this is an infinite loop that will 2164 // keep getting unwound. 2165 if (PBI->getSuccessor(PBIOp) == BB) 2166 return false; 2167 2168 // Do not perform this transformation if it would require 2169 // insertion of a large number of select instructions. For targets 2170 // without predication/cmovs, this is a big pessimization. 2171 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2172 2173 unsigned NumPhis = 0; 2174 for (BasicBlock::iterator II = CommonDest->begin(); 2175 isa<PHINode>(II); ++II, ++NumPhis) 2176 if (NumPhis > 2) // Disable this xform. 2177 return false; 2178 2179 // Finally, if everything is ok, fold the branches to logical ops. 2180 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2181 2182 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2183 << "AND: " << *BI->getParent()); 2184 2185 2186 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2187 // branch in it, where one edge (OtherDest) goes back to itself but the other 2188 // exits. We don't *know* that the program avoids the infinite loop 2189 // (even though that seems likely). If we do this xform naively, we'll end up 2190 // recursively unpeeling the loop. Since we know that (after the xform is 2191 // done) that the block *is* infinite if reached, we just make it an obviously 2192 // infinite loop with no cond branch. 2193 if (OtherDest == BB) { 2194 // Insert it at the end of the function, because it's either code, 2195 // or it won't matter if it's hot. :) 2196 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2197 "infloop", BB->getParent()); 2198 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2199 OtherDest = InfLoopBlock; 2200 } 2201 2202 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2203 2204 // BI may have other predecessors. Because of this, we leave 2205 // it alone, but modify PBI. 2206 2207 // Make sure we get to CommonDest on True&True directions. 2208 Value *PBICond = PBI->getCondition(); 2209 IRBuilder<true, NoFolder> Builder(PBI); 2210 if (PBIOp) 2211 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2212 2213 Value *BICond = BI->getCondition(); 2214 if (BIOp) 2215 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2216 2217 // Merge the conditions. 2218 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2219 2220 // Modify PBI to branch on the new condition to the new dests. 2221 PBI->setCondition(Cond); 2222 PBI->setSuccessor(0, CommonDest); 2223 PBI->setSuccessor(1, OtherDest); 2224 2225 // OtherDest may have phi nodes. If so, add an entry from PBI's 2226 // block that are identical to the entries for BI's block. 2227 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2228 2229 // We know that the CommonDest already had an edge from PBI to 2230 // it. If it has PHIs though, the PHIs may have different 2231 // entries for BB and PBI's BB. If so, insert a select to make 2232 // them agree. 2233 PHINode *PN; 2234 for (BasicBlock::iterator II = CommonDest->begin(); 2235 (PN = dyn_cast<PHINode>(II)); ++II) { 2236 Value *BIV = PN->getIncomingValueForBlock(BB); 2237 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2238 Value *PBIV = PN->getIncomingValue(PBBIdx); 2239 if (BIV != PBIV) { 2240 // Insert a select in PBI to pick the right value. 2241 Value *NV = cast<SelectInst> 2242 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2243 PN->setIncomingValue(PBBIdx, NV); 2244 } 2245 } 2246 2247 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2248 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2249 2250 // This basic block is probably dead. We know it has at least 2251 // one fewer predecessor. 2252 return true; 2253} 2254 2255// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2256// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2257// Takes care of updating the successors and removing the old terminator. 2258// Also makes sure not to introduce new successors by assuming that edges to 2259// non-successor TrueBBs and FalseBBs aren't reachable. 2260static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2261 BasicBlock *TrueBB, BasicBlock *FalseBB){ 2262 // Remove any superfluous successor edges from the CFG. 2263 // First, figure out which successors to preserve. 2264 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2265 // successor. 2266 BasicBlock *KeepEdge1 = TrueBB; 2267 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2268 2269 // Then remove the rest. 2270 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2271 BasicBlock *Succ = OldTerm->getSuccessor(I); 2272 // Make sure only to keep exactly one copy of each edge. 2273 if (Succ == KeepEdge1) 2274 KeepEdge1 = 0; 2275 else if (Succ == KeepEdge2) 2276 KeepEdge2 = 0; 2277 else 2278 Succ->removePredecessor(OldTerm->getParent()); 2279 } 2280 2281 IRBuilder<> Builder(OldTerm); 2282 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2283 2284 // Insert an appropriate new terminator. 2285 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2286 if (TrueBB == FalseBB) 2287 // We were only looking for one successor, and it was present. 2288 // Create an unconditional branch to it. 2289 Builder.CreateBr(TrueBB); 2290 else 2291 // We found both of the successors we were looking for. 2292 // Create a conditional branch sharing the condition of the select. 2293 Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2294 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2295 // Neither of the selected blocks were successors, so this 2296 // terminator must be unreachable. 2297 new UnreachableInst(OldTerm->getContext(), OldTerm); 2298 } else { 2299 // One of the selected values was a successor, but the other wasn't. 2300 // Insert an unconditional branch to the one that was found; 2301 // the edge to the one that wasn't must be unreachable. 2302 if (KeepEdge1 == 0) 2303 // Only TrueBB was found. 2304 Builder.CreateBr(TrueBB); 2305 else 2306 // Only FalseBB was found. 2307 Builder.CreateBr(FalseBB); 2308 } 2309 2310 EraseTerminatorInstAndDCECond(OldTerm); 2311 return true; 2312} 2313 2314// SimplifySwitchOnSelect - Replaces 2315// (switch (select cond, X, Y)) on constant X, Y 2316// with a branch - conditional if X and Y lead to distinct BBs, 2317// unconditional otherwise. 2318static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2319 // Check for constant integer values in the select. 2320 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2321 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2322 if (!TrueVal || !FalseVal) 2323 return false; 2324 2325 // Find the relevant condition and destinations. 2326 Value *Condition = Select->getCondition(); 2327 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2328 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2329 2330 // Perform the actual simplification. 2331 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB); 2332} 2333 2334// SimplifyIndirectBrOnSelect - Replaces 2335// (indirectbr (select cond, blockaddress(@fn, BlockA), 2336// blockaddress(@fn, BlockB))) 2337// with 2338// (br cond, BlockA, BlockB). 2339static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2340 // Check that both operands of the select are block addresses. 2341 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2342 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2343 if (!TBA || !FBA) 2344 return false; 2345 2346 // Extract the actual blocks. 2347 BasicBlock *TrueBB = TBA->getBasicBlock(); 2348 BasicBlock *FalseBB = FBA->getBasicBlock(); 2349 2350 // Perform the actual simplification. 2351 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); 2352} 2353 2354/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2355/// instruction (a seteq/setne with a constant) as the only instruction in a 2356/// block that ends with an uncond branch. We are looking for a very specific 2357/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2358/// this case, we merge the first two "or's of icmp" into a switch, but then the 2359/// default value goes to an uncond block with a seteq in it, we get something 2360/// like: 2361/// 2362/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2363/// DEFAULT: 2364/// %tmp = icmp eq i8 %A, 92 2365/// br label %end 2366/// end: 2367/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2368/// 2369/// We prefer to split the edge to 'end' so that there is a true/false entry to 2370/// the PHI, merging the third icmp into the switch. 2371static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 2372 const TargetData *TD, 2373 IRBuilder<> &Builder) { 2374 BasicBlock *BB = ICI->getParent(); 2375 2376 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2377 // complex. 2378 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2379 2380 Value *V = ICI->getOperand(0); 2381 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2382 2383 // The pattern we're looking for is where our only predecessor is a switch on 2384 // 'V' and this block is the default case for the switch. In this case we can 2385 // fold the compared value into the switch to simplify things. 2386 BasicBlock *Pred = BB->getSinglePredecessor(); 2387 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2388 2389 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2390 if (SI->getCondition() != V) 2391 return false; 2392 2393 // If BB is reachable on a non-default case, then we simply know the value of 2394 // V in this block. Substitute it and constant fold the icmp instruction 2395 // away. 2396 if (SI->getDefaultDest() != BB) { 2397 ConstantInt *VVal = SI->findCaseDest(BB); 2398 assert(VVal && "Should have a unique destination value"); 2399 ICI->setOperand(0, VVal); 2400 2401 if (Value *V = SimplifyInstruction(ICI, TD)) { 2402 ICI->replaceAllUsesWith(V); 2403 ICI->eraseFromParent(); 2404 } 2405 // BB is now empty, so it is likely to simplify away. 2406 return SimplifyCFG(BB) | true; 2407 } 2408 2409 // Ok, the block is reachable from the default dest. If the constant we're 2410 // comparing exists in one of the other edges, then we can constant fold ICI 2411 // and zap it. 2412 if (SI->findCaseValue(Cst) != SI->case_default()) { 2413 Value *V; 2414 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2415 V = ConstantInt::getFalse(BB->getContext()); 2416 else 2417 V = ConstantInt::getTrue(BB->getContext()); 2418 2419 ICI->replaceAllUsesWith(V); 2420 ICI->eraseFromParent(); 2421 // BB is now empty, so it is likely to simplify away. 2422 return SimplifyCFG(BB) | true; 2423 } 2424 2425 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2426 // the block. 2427 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2428 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2429 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2430 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2431 return false; 2432 2433 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2434 // true in the PHI. 2435 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2436 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2437 2438 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2439 std::swap(DefaultCst, NewCst); 2440 2441 // Replace ICI (which is used by the PHI for the default value) with true or 2442 // false depending on if it is EQ or NE. 2443 ICI->replaceAllUsesWith(DefaultCst); 2444 ICI->eraseFromParent(); 2445 2446 // Okay, the switch goes to this block on a default value. Add an edge from 2447 // the switch to the merge point on the compared value. 2448 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2449 BB->getParent(), BB); 2450 SI->addCase(Cst, NewBB); 2451 2452 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2453 Builder.SetInsertPoint(NewBB); 2454 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2455 Builder.CreateBr(SuccBlock); 2456 PHIUse->addIncoming(NewCst, NewBB); 2457 return true; 2458} 2459 2460/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2461/// Check to see if it is branching on an or/and chain of icmp instructions, and 2462/// fold it into a switch instruction if so. 2463static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, 2464 IRBuilder<> &Builder) { 2465 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2466 if (Cond == 0) return false; 2467 2468 2469 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2470 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2471 // 'setne's and'ed together, collect them. 2472 Value *CompVal = 0; 2473 std::vector<ConstantInt*> Values; 2474 bool TrueWhenEqual = true; 2475 Value *ExtraCase = 0; 2476 unsigned UsedICmps = 0; 2477 2478 if (Cond->getOpcode() == Instruction::Or) { 2479 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2480 UsedICmps); 2481 } else if (Cond->getOpcode() == Instruction::And) { 2482 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2483 UsedICmps); 2484 TrueWhenEqual = false; 2485 } 2486 2487 // If we didn't have a multiply compared value, fail. 2488 if (CompVal == 0) return false; 2489 2490 // Avoid turning single icmps into a switch. 2491 if (UsedICmps <= 1) 2492 return false; 2493 2494 // There might be duplicate constants in the list, which the switch 2495 // instruction can't handle, remove them now. 2496 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2497 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2498 2499 // If Extra was used, we require at least two switch values to do the 2500 // transformation. A switch with one value is just an cond branch. 2501 if (ExtraCase && Values.size() < 2) return false; 2502 2503 // TODO: Preserve branch weight metadata, similarly to how 2504 // FoldValueComparisonIntoPredecessors preserves it. 2505 2506 // Figure out which block is which destination. 2507 BasicBlock *DefaultBB = BI->getSuccessor(1); 2508 BasicBlock *EdgeBB = BI->getSuccessor(0); 2509 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2510 2511 BasicBlock *BB = BI->getParent(); 2512 2513 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2514 << " cases into SWITCH. BB is:\n" << *BB); 2515 2516 // If there are any extra values that couldn't be folded into the switch 2517 // then we evaluate them with an explicit branch first. Split the block 2518 // right before the condbr to handle it. 2519 if (ExtraCase) { 2520 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2521 // Remove the uncond branch added to the old block. 2522 TerminatorInst *OldTI = BB->getTerminator(); 2523 Builder.SetInsertPoint(OldTI); 2524 2525 if (TrueWhenEqual) 2526 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2527 else 2528 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2529 2530 OldTI->eraseFromParent(); 2531 2532 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2533 // for the edge we just added. 2534 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2535 2536 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2537 << "\nEXTRABB = " << *BB); 2538 BB = NewBB; 2539 } 2540 2541 Builder.SetInsertPoint(BI); 2542 // Convert pointer to int before we switch. 2543 if (CompVal->getType()->isPointerTy()) { 2544 assert(TD && "Cannot switch on pointer without TargetData"); 2545 CompVal = Builder.CreatePtrToInt(CompVal, 2546 TD->getIntPtrType(CompVal->getContext()), 2547 "magicptr"); 2548 } 2549 2550 // Create the new switch instruction now. 2551 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2552 2553 // Add all of the 'cases' to the switch instruction. 2554 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2555 New->addCase(Values[i], EdgeBB); 2556 2557 // We added edges from PI to the EdgeBB. As such, if there were any 2558 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2559 // the number of edges added. 2560 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2561 isa<PHINode>(BBI); ++BBI) { 2562 PHINode *PN = cast<PHINode>(BBI); 2563 Value *InVal = PN->getIncomingValueForBlock(BB); 2564 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2565 PN->addIncoming(InVal, BB); 2566 } 2567 2568 // Erase the old branch instruction. 2569 EraseTerminatorInstAndDCECond(BI); 2570 2571 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2572 return true; 2573} 2574 2575bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2576 // If this is a trivial landing pad that just continues unwinding the caught 2577 // exception then zap the landing pad, turning its invokes into calls. 2578 BasicBlock *BB = RI->getParent(); 2579 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2580 if (RI->getValue() != LPInst) 2581 // Not a landing pad, or the resume is not unwinding the exception that 2582 // caused control to branch here. 2583 return false; 2584 2585 // Check that there are no other instructions except for debug intrinsics. 2586 BasicBlock::iterator I = LPInst, E = RI; 2587 while (++I != E) 2588 if (!isa<DbgInfoIntrinsic>(I)) 2589 return false; 2590 2591 // Turn all invokes that unwind here into calls and delete the basic block. 2592 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2593 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2594 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2595 // Insert a call instruction before the invoke. 2596 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2597 Call->takeName(II); 2598 Call->setCallingConv(II->getCallingConv()); 2599 Call->setAttributes(II->getAttributes()); 2600 Call->setDebugLoc(II->getDebugLoc()); 2601 2602 // Anything that used the value produced by the invoke instruction now uses 2603 // the value produced by the call instruction. Note that we do this even 2604 // for void functions and calls with no uses so that the callgraph edge is 2605 // updated. 2606 II->replaceAllUsesWith(Call); 2607 BB->removePredecessor(II->getParent()); 2608 2609 // Insert a branch to the normal destination right before the invoke. 2610 BranchInst::Create(II->getNormalDest(), II); 2611 2612 // Finally, delete the invoke instruction! 2613 II->eraseFromParent(); 2614 } 2615 2616 // The landingpad is now unreachable. Zap it. 2617 BB->eraseFromParent(); 2618 return true; 2619} 2620 2621bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2622 BasicBlock *BB = RI->getParent(); 2623 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2624 2625 // Find predecessors that end with branches. 2626 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2627 SmallVector<BranchInst*, 8> CondBranchPreds; 2628 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2629 BasicBlock *P = *PI; 2630 TerminatorInst *PTI = P->getTerminator(); 2631 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2632 if (BI->isUnconditional()) 2633 UncondBranchPreds.push_back(P); 2634 else 2635 CondBranchPreds.push_back(BI); 2636 } 2637 } 2638 2639 // If we found some, do the transformation! 2640 if (!UncondBranchPreds.empty() && DupRet) { 2641 while (!UncondBranchPreds.empty()) { 2642 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2643 DEBUG(dbgs() << "FOLDING: " << *BB 2644 << "INTO UNCOND BRANCH PRED: " << *Pred); 2645 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2646 } 2647 2648 // If we eliminated all predecessors of the block, delete the block now. 2649 if (pred_begin(BB) == pred_end(BB)) 2650 // We know there are no successors, so just nuke the block. 2651 BB->eraseFromParent(); 2652 2653 return true; 2654 } 2655 2656 // Check out all of the conditional branches going to this return 2657 // instruction. If any of them just select between returns, change the 2658 // branch itself into a select/return pair. 2659 while (!CondBranchPreds.empty()) { 2660 BranchInst *BI = CondBranchPreds.pop_back_val(); 2661 2662 // Check to see if the non-BB successor is also a return block. 2663 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2664 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2665 SimplifyCondBranchToTwoReturns(BI, Builder)) 2666 return true; 2667 } 2668 return false; 2669} 2670 2671bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2672 BasicBlock *BB = UI->getParent(); 2673 2674 bool Changed = false; 2675 2676 // If there are any instructions immediately before the unreachable that can 2677 // be removed, do so. 2678 while (UI != BB->begin()) { 2679 BasicBlock::iterator BBI = UI; 2680 --BBI; 2681 // Do not delete instructions that can have side effects which might cause 2682 // the unreachable to not be reachable; specifically, calls and volatile 2683 // operations may have this effect. 2684 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2685 2686 if (BBI->mayHaveSideEffects()) { 2687 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2688 if (SI->isVolatile()) 2689 break; 2690 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2691 if (LI->isVolatile()) 2692 break; 2693 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2694 if (RMWI->isVolatile()) 2695 break; 2696 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2697 if (CXI->isVolatile()) 2698 break; 2699 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2700 !isa<LandingPadInst>(BBI)) { 2701 break; 2702 } 2703 // Note that deleting LandingPad's here is in fact okay, although it 2704 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2705 // all the predecessors of this block will be the unwind edges of Invokes, 2706 // and we can therefore guarantee this block will be erased. 2707 } 2708 2709 // Delete this instruction (any uses are guaranteed to be dead) 2710 if (!BBI->use_empty()) 2711 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2712 BBI->eraseFromParent(); 2713 Changed = true; 2714 } 2715 2716 // If the unreachable instruction is the first in the block, take a gander 2717 // at all of the predecessors of this instruction, and simplify them. 2718 if (&BB->front() != UI) return Changed; 2719 2720 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2721 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2722 TerminatorInst *TI = Preds[i]->getTerminator(); 2723 IRBuilder<> Builder(TI); 2724 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2725 if (BI->isUnconditional()) { 2726 if (BI->getSuccessor(0) == BB) { 2727 new UnreachableInst(TI->getContext(), TI); 2728 TI->eraseFromParent(); 2729 Changed = true; 2730 } 2731 } else { 2732 if (BI->getSuccessor(0) == BB) { 2733 Builder.CreateBr(BI->getSuccessor(1)); 2734 EraseTerminatorInstAndDCECond(BI); 2735 } else if (BI->getSuccessor(1) == BB) { 2736 Builder.CreateBr(BI->getSuccessor(0)); 2737 EraseTerminatorInstAndDCECond(BI); 2738 Changed = true; 2739 } 2740 } 2741 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2742 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2743 i != e; ++i) 2744 if (i.getCaseSuccessor() == BB) { 2745 BB->removePredecessor(SI->getParent()); 2746 SI->removeCase(i); 2747 --i; --e; 2748 Changed = true; 2749 } 2750 // If the default value is unreachable, figure out the most popular 2751 // destination and make it the default. 2752 if (SI->getDefaultDest() == BB) { 2753 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2754 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2755 i != e; ++i) { 2756 std::pair<unsigned, unsigned> &entry = 2757 Popularity[i.getCaseSuccessor()]; 2758 if (entry.first == 0) { 2759 entry.first = 1; 2760 entry.second = i.getCaseIndex(); 2761 } else { 2762 entry.first++; 2763 } 2764 } 2765 2766 // Find the most popular block. 2767 unsigned MaxPop = 0; 2768 unsigned MaxIndex = 0; 2769 BasicBlock *MaxBlock = 0; 2770 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2771 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2772 if (I->second.first > MaxPop || 2773 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2774 MaxPop = I->second.first; 2775 MaxIndex = I->second.second; 2776 MaxBlock = I->first; 2777 } 2778 } 2779 if (MaxBlock) { 2780 // Make this the new default, allowing us to delete any explicit 2781 // edges to it. 2782 SI->setDefaultDest(MaxBlock); 2783 Changed = true; 2784 2785 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2786 // it. 2787 if (isa<PHINode>(MaxBlock->begin())) 2788 for (unsigned i = 0; i != MaxPop-1; ++i) 2789 MaxBlock->removePredecessor(SI->getParent()); 2790 2791 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2792 i != e; ++i) 2793 if (i.getCaseSuccessor() == MaxBlock) { 2794 SI->removeCase(i); 2795 --i; --e; 2796 } 2797 } 2798 } 2799 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2800 if (II->getUnwindDest() == BB) { 2801 // Convert the invoke to a call instruction. This would be a good 2802 // place to note that the call does not throw though. 2803 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2804 II->removeFromParent(); // Take out of symbol table 2805 2806 // Insert the call now... 2807 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2808 Builder.SetInsertPoint(BI); 2809 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2810 Args, II->getName()); 2811 CI->setCallingConv(II->getCallingConv()); 2812 CI->setAttributes(II->getAttributes()); 2813 // If the invoke produced a value, the call does now instead. 2814 II->replaceAllUsesWith(CI); 2815 delete II; 2816 Changed = true; 2817 } 2818 } 2819 } 2820 2821 // If this block is now dead, remove it. 2822 if (pred_begin(BB) == pred_end(BB) && 2823 BB != &BB->getParent()->getEntryBlock()) { 2824 // We know there are no successors, so just nuke the block. 2825 BB->eraseFromParent(); 2826 return true; 2827 } 2828 2829 return Changed; 2830} 2831 2832/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 2833/// integer range comparison into a sub, an icmp and a branch. 2834static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 2835 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 2836 2837 // Make sure all cases point to the same destination and gather the values. 2838 SmallVector<ConstantInt *, 16> Cases; 2839 SwitchInst::CaseIt I = SI->case_begin(); 2840 Cases.push_back(I.getCaseValue()); 2841 SwitchInst::CaseIt PrevI = I++; 2842 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 2843 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 2844 return false; 2845 Cases.push_back(I.getCaseValue()); 2846 } 2847 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 2848 2849 // Sort the case values, then check if they form a range we can transform. 2850 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 2851 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 2852 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 2853 return false; 2854 } 2855 2856 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 2857 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 2858 2859 Value *Sub = SI->getCondition(); 2860 if (!Offset->isNullValue()) 2861 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 2862 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 2863 Builder.CreateCondBr( 2864 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 2865 2866 // Prune obsolete incoming values off the successor's PHI nodes. 2867 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 2868 isa<PHINode>(BBI); ++BBI) { 2869 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 2870 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 2871 } 2872 SI->eraseFromParent(); 2873 2874 return true; 2875} 2876 2877/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 2878/// and use it to remove dead cases. 2879static bool EliminateDeadSwitchCases(SwitchInst *SI) { 2880 Value *Cond = SI->getCondition(); 2881 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 2882 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 2883 ComputeMaskedBits(Cond, KnownZero, KnownOne); 2884 2885 // Gather dead cases. 2886 SmallVector<ConstantInt*, 8> DeadCases; 2887 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 2888 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 2889 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 2890 DeadCases.push_back(I.getCaseValue()); 2891 DEBUG(dbgs() << "SimplifyCFG: switch case '" 2892 << I.getCaseValue() << "' is dead.\n"); 2893 } 2894 } 2895 2896 // Remove dead cases from the switch. 2897 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 2898 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 2899 assert(Case != SI->case_default() && 2900 "Case was not found. Probably mistake in DeadCases forming."); 2901 // Prune unused values from PHI nodes. 2902 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 2903 SI->removeCase(Case); 2904 } 2905 2906 return !DeadCases.empty(); 2907} 2908 2909/// FindPHIForConditionForwarding - If BB would be eligible for simplification 2910/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 2911/// by an unconditional branch), look at the phi node for BB in the successor 2912/// block and see if the incoming value is equal to CaseValue. If so, return 2913/// the phi node, and set PhiIndex to BB's index in the phi node. 2914static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 2915 BasicBlock *BB, 2916 int *PhiIndex) { 2917 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 2918 return NULL; // BB must be empty to be a candidate for simplification. 2919 if (!BB->getSinglePredecessor()) 2920 return NULL; // BB must be dominated by the switch. 2921 2922 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 2923 if (!Branch || !Branch->isUnconditional()) 2924 return NULL; // Terminator must be unconditional branch. 2925 2926 BasicBlock *Succ = Branch->getSuccessor(0); 2927 2928 BasicBlock::iterator I = Succ->begin(); 2929 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 2930 int Idx = PHI->getBasicBlockIndex(BB); 2931 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 2932 2933 Value *InValue = PHI->getIncomingValue(Idx); 2934 if (InValue != CaseValue) continue; 2935 2936 *PhiIndex = Idx; 2937 return PHI; 2938 } 2939 2940 return NULL; 2941} 2942 2943/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 2944/// instruction to a phi node dominated by the switch, if that would mean that 2945/// some of the destination blocks of the switch can be folded away. 2946/// Returns true if a change is made. 2947static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 2948 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 2949 ForwardingNodesMap ForwardingNodes; 2950 2951 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 2952 ConstantInt *CaseValue = I.getCaseValue(); 2953 BasicBlock *CaseDest = I.getCaseSuccessor(); 2954 2955 int PhiIndex; 2956 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 2957 &PhiIndex); 2958 if (!PHI) continue; 2959 2960 ForwardingNodes[PHI].push_back(PhiIndex); 2961 } 2962 2963 bool Changed = false; 2964 2965 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 2966 E = ForwardingNodes.end(); I != E; ++I) { 2967 PHINode *Phi = I->first; 2968 SmallVector<int,4> &Indexes = I->second; 2969 2970 if (Indexes.size() < 2) continue; 2971 2972 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 2973 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 2974 Changed = true; 2975 } 2976 2977 return Changed; 2978} 2979 2980bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 2981 // If this switch is too complex to want to look at, ignore it. 2982 if (!isValueEqualityComparison(SI)) 2983 return false; 2984 2985 BasicBlock *BB = SI->getParent(); 2986 2987 // If we only have one predecessor, and if it is a branch on this value, 2988 // see if that predecessor totally determines the outcome of this switch. 2989 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2990 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 2991 return SimplifyCFG(BB) | true; 2992 2993 Value *Cond = SI->getCondition(); 2994 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 2995 if (SimplifySwitchOnSelect(SI, Select)) 2996 return SimplifyCFG(BB) | true; 2997 2998 // If the block only contains the switch, see if we can fold the block 2999 // away into any preds. 3000 BasicBlock::iterator BBI = BB->begin(); 3001 // Ignore dbg intrinsics. 3002 while (isa<DbgInfoIntrinsic>(BBI)) 3003 ++BBI; 3004 if (SI == &*BBI) 3005 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3006 return SimplifyCFG(BB) | true; 3007 3008 // Try to transform the switch into an icmp and a branch. 3009 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3010 return SimplifyCFG(BB) | true; 3011 3012 // Remove unreachable cases. 3013 if (EliminateDeadSwitchCases(SI)) 3014 return SimplifyCFG(BB) | true; 3015 3016 if (ForwardSwitchConditionToPHI(SI)) 3017 return SimplifyCFG(BB) | true; 3018 3019 return false; 3020} 3021 3022bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3023 BasicBlock *BB = IBI->getParent(); 3024 bool Changed = false; 3025 3026 // Eliminate redundant destinations. 3027 SmallPtrSet<Value *, 8> Succs; 3028 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3029 BasicBlock *Dest = IBI->getDestination(i); 3030 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3031 Dest->removePredecessor(BB); 3032 IBI->removeDestination(i); 3033 --i; --e; 3034 Changed = true; 3035 } 3036 } 3037 3038 if (IBI->getNumDestinations() == 0) { 3039 // If the indirectbr has no successors, change it to unreachable. 3040 new UnreachableInst(IBI->getContext(), IBI); 3041 EraseTerminatorInstAndDCECond(IBI); 3042 return true; 3043 } 3044 3045 if (IBI->getNumDestinations() == 1) { 3046 // If the indirectbr has one successor, change it to a direct branch. 3047 BranchInst::Create(IBI->getDestination(0), IBI); 3048 EraseTerminatorInstAndDCECond(IBI); 3049 return true; 3050 } 3051 3052 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3053 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3054 return SimplifyCFG(BB) | true; 3055 } 3056 return Changed; 3057} 3058 3059bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3060 BasicBlock *BB = BI->getParent(); 3061 3062 // If the Terminator is the only non-phi instruction, simplify the block. 3063 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3064 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3065 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3066 return true; 3067 3068 // If the only instruction in the block is a seteq/setne comparison 3069 // against a constant, try to simplify the block. 3070 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3071 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3072 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3073 ; 3074 if (I->isTerminator() && 3075 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) 3076 return true; 3077 } 3078 3079 // If this basic block is ONLY a compare and a branch, and if a predecessor 3080 // branches to us and our successor, fold the comparison into the 3081 // predecessor and use logical operations to update the incoming value 3082 // for PHI nodes in common successor. 3083 if (FoldBranchToCommonDest(BI)) 3084 return SimplifyCFG(BB) | true; 3085 return false; 3086} 3087 3088 3089bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3090 BasicBlock *BB = BI->getParent(); 3091 3092 // Conditional branch 3093 if (isValueEqualityComparison(BI)) { 3094 // If we only have one predecessor, and if it is a branch on this value, 3095 // see if that predecessor totally determines the outcome of this 3096 // switch. 3097 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3098 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3099 return SimplifyCFG(BB) | true; 3100 3101 // This block must be empty, except for the setcond inst, if it exists. 3102 // Ignore dbg intrinsics. 3103 BasicBlock::iterator I = BB->begin(); 3104 // Ignore dbg intrinsics. 3105 while (isa<DbgInfoIntrinsic>(I)) 3106 ++I; 3107 if (&*I == BI) { 3108 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3109 return SimplifyCFG(BB) | true; 3110 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3111 ++I; 3112 // Ignore dbg intrinsics. 3113 while (isa<DbgInfoIntrinsic>(I)) 3114 ++I; 3115 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3116 return SimplifyCFG(BB) | true; 3117 } 3118 } 3119 3120 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3121 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3122 return true; 3123 3124 // If this basic block is ONLY a compare and a branch, and if a predecessor 3125 // branches to us and one of our successors, fold the comparison into the 3126 // predecessor and use logical operations to pick the right destination. 3127 if (FoldBranchToCommonDest(BI)) 3128 return SimplifyCFG(BB) | true; 3129 3130 // We have a conditional branch to two blocks that are only reachable 3131 // from BI. We know that the condbr dominates the two blocks, so see if 3132 // there is any identical code in the "then" and "else" blocks. If so, we 3133 // can hoist it up to the branching block. 3134 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3135 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3136 if (HoistThenElseCodeToIf(BI)) 3137 return SimplifyCFG(BB) | true; 3138 } else { 3139 // If Successor #1 has multiple preds, we may be able to conditionally 3140 // execute Successor #0 if it branches to successor #1. 3141 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3142 if (Succ0TI->getNumSuccessors() == 1 && 3143 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3144 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3145 return SimplifyCFG(BB) | true; 3146 } 3147 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3148 // If Successor #0 has multiple preds, we may be able to conditionally 3149 // execute Successor #1 if it branches to successor #0. 3150 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3151 if (Succ1TI->getNumSuccessors() == 1 && 3152 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3153 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3154 return SimplifyCFG(BB) | true; 3155 } 3156 3157 // If this is a branch on a phi node in the current block, thread control 3158 // through this block if any PHI node entries are constants. 3159 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3160 if (PN->getParent() == BI->getParent()) 3161 if (FoldCondBranchOnPHI(BI, TD)) 3162 return SimplifyCFG(BB) | true; 3163 3164 // Scan predecessor blocks for conditional branches. 3165 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3166 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3167 if (PBI != BI && PBI->isConditional()) 3168 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3169 return SimplifyCFG(BB) | true; 3170 3171 return false; 3172} 3173 3174/// Check if passing a value to an instruction will cause undefined behavior. 3175static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3176 Constant *C = dyn_cast<Constant>(V); 3177 if (!C) 3178 return false; 3179 3180 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time 3181 return false; 3182 3183 if (C->isNullValue()) { 3184 Instruction *Use = I->use_back(); 3185 3186 // Now make sure that there are no instructions in between that can alter 3187 // control flow (eg. calls) 3188 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3189 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3190 return false; 3191 3192 // Look through GEPs. A load from a GEP derived from NULL is still undefined 3193 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 3194 if (GEP->getPointerOperand() == I) 3195 return passingValueIsAlwaysUndefined(V, GEP); 3196 3197 // Look through bitcasts. 3198 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 3199 return passingValueIsAlwaysUndefined(V, BC); 3200 3201 // Load from null is undefined. 3202 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 3203 return LI->getPointerAddressSpace() == 0; 3204 3205 // Store to null is undefined. 3206 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 3207 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 3208 } 3209 return false; 3210} 3211 3212/// If BB has an incoming value that will always trigger undefined behavior 3213/// (eg. null pointer dereference), remove the branch leading here. 3214static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 3215 for (BasicBlock::iterator i = BB->begin(); 3216 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 3217 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 3218 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 3219 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 3220 IRBuilder<> Builder(T); 3221 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 3222 BB->removePredecessor(PHI->getIncomingBlock(i)); 3223 // Turn uncoditional branches into unreachables and remove the dead 3224 // destination from conditional branches. 3225 if (BI->isUnconditional()) 3226 Builder.CreateUnreachable(); 3227 else 3228 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 3229 BI->getSuccessor(0)); 3230 BI->eraseFromParent(); 3231 return true; 3232 } 3233 // TODO: SwitchInst. 3234 } 3235 3236 return false; 3237} 3238 3239bool SimplifyCFGOpt::run(BasicBlock *BB) { 3240 bool Changed = false; 3241 3242 assert(BB && BB->getParent() && "Block not embedded in function!"); 3243 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 3244 3245 // Remove basic blocks that have no predecessors (except the entry block)... 3246 // or that just have themself as a predecessor. These are unreachable. 3247 if ((pred_begin(BB) == pred_end(BB) && 3248 BB != &BB->getParent()->getEntryBlock()) || 3249 BB->getSinglePredecessor() == BB) { 3250 DEBUG(dbgs() << "Removing BB: \n" << *BB); 3251 DeleteDeadBlock(BB); 3252 return true; 3253 } 3254 3255 // Check to see if we can constant propagate this terminator instruction 3256 // away... 3257 Changed |= ConstantFoldTerminator(BB, true); 3258 3259 // Check for and eliminate duplicate PHI nodes in this block. 3260 Changed |= EliminateDuplicatePHINodes(BB); 3261 3262 // Check for and remove branches that will always cause undefined behavior. 3263 Changed |= removeUndefIntroducingPredecessor(BB); 3264 3265 // Merge basic blocks into their predecessor if there is only one distinct 3266 // pred, and if there is only one distinct successor of the predecessor, and 3267 // if there are no PHI nodes. 3268 // 3269 if (MergeBlockIntoPredecessor(BB)) 3270 return true; 3271 3272 IRBuilder<> Builder(BB); 3273 3274 // If there is a trivial two-entry PHI node in this basic block, and we can 3275 // eliminate it, do so now. 3276 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 3277 if (PN->getNumIncomingValues() == 2) 3278 Changed |= FoldTwoEntryPHINode(PN, TD); 3279 3280 Builder.SetInsertPoint(BB->getTerminator()); 3281 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 3282 if (BI->isUnconditional()) { 3283 if (SimplifyUncondBranch(BI, Builder)) return true; 3284 } else { 3285 if (SimplifyCondBranch(BI, Builder)) return true; 3286 } 3287 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 3288 if (SimplifyReturn(RI, Builder)) return true; 3289 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 3290 if (SimplifyResume(RI, Builder)) return true; 3291 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 3292 if (SimplifySwitch(SI, Builder)) return true; 3293 } else if (UnreachableInst *UI = 3294 dyn_cast<UnreachableInst>(BB->getTerminator())) { 3295 if (SimplifyUnreachable(UI)) return true; 3296 } else if (IndirectBrInst *IBI = 3297 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 3298 if (SimplifyIndirectBr(IBI)) return true; 3299 } 3300 3301 return Changed; 3302} 3303 3304/// SimplifyCFG - This function is used to do simplification of a CFG. For 3305/// example, it adjusts branches to branches to eliminate the extra hop, it 3306/// eliminates unreachable basic blocks, and does other "peephole" optimization 3307/// of the CFG. It returns true if a modification was made. 3308/// 3309bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 3310 return SimplifyCFGOpt(TD).run(BB); 3311} 3312