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/Instructions.h" 18 #include "llvm/IntrinsicInst.h" 19 #include "llvm/Type.h" 20 #include "llvm/DerivedTypes.h" 21 #include "llvm/GlobalVariable.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Target/TargetData.h" 24 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 25 #include "llvm/ADT/DenseMap.h" 26 #include "llvm/ADT/SmallVector.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/ADT/STLExtras.h" 30 #include "llvm/Support/CFG.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Support/ConstantRange.h" 33 #include "llvm/Support/Debug.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include <algorithm> 36 #include <set> 37 #include <map> 38 using namespace llvm; 39 40 static cl::opt<bool> 41 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 42 cl::desc("Duplicate return instructions into unconditional branches")); 43 44 STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 45 46 namespace { 47 class SimplifyCFGOpt { 48 const TargetData *const TD; 49 50 Value *isValueEqualityComparison(TerminatorInst *TI); 51 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 52 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases); 53 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 54 BasicBlock *Pred); 55 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI); 56 57 bool SimplifyReturn(ReturnInst *RI); 58 bool SimplifyUnwind(UnwindInst *UI); 59 bool SimplifyUnreachable(UnreachableInst *UI); 60 bool SimplifySwitch(SwitchInst *SI); 61 bool SimplifyIndirectBr(IndirectBrInst *IBI); 62 bool SimplifyUncondBranch(BranchInst *BI); 63 bool SimplifyCondBranch(BranchInst *BI); 64 65 public: 66 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} 67 bool run(BasicBlock *BB); 68 }; 69 } 70 71 /// SafeToMergeTerminators - Return true if it is safe to merge these two 72 /// terminator instructions together. 73 /// 74 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 75 if (SI1 == SI2) return false; // Can't merge with self! 76 77 // It is not safe to merge these two switch instructions if they have a common 78 // successor, and if that successor has a PHI node, and if *that* PHI node has 79 // conflicting incoming values from the two switch blocks. 80 BasicBlock *SI1BB = SI1->getParent(); 81 BasicBlock *SI2BB = SI2->getParent(); 82 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 83 84 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 85 if (SI1Succs.count(*I)) 86 for (BasicBlock::iterator BBI = (*I)->begin(); 87 isa<PHINode>(BBI); ++BBI) { 88 PHINode *PN = cast<PHINode>(BBI); 89 if (PN->getIncomingValueForBlock(SI1BB) != 90 PN->getIncomingValueForBlock(SI2BB)) 91 return false; 92 } 93 94 return true; 95 } 96 97 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 98 /// now be entries in it from the 'NewPred' block. The values that will be 99 /// flowing into the PHI nodes will be the same as those coming in from 100 /// ExistPred, an existing predecessor of Succ. 101 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 102 BasicBlock *ExistPred) { 103 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 104 105 PHINode *PN; 106 for (BasicBlock::iterator I = Succ->begin(); 107 (PN = dyn_cast<PHINode>(I)); ++I) 108 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 109 } 110 111 112 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at 113 /// least one PHI node in it), check to see if the merge at this block is due 114 /// to an "if condition". If so, return the boolean condition that determines 115 /// which entry into BB will be taken. Also, return by references the block 116 /// that will be entered from if the condition is true, and the block that will 117 /// be entered if the condition is false. 118 /// 119 /// This does no checking to see if the true/false blocks have large or unsavory 120 /// instructions in them. 121 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 122 BasicBlock *&IfFalse) { 123 PHINode *SomePHI = cast<PHINode>(BB->begin()); 124 assert(SomePHI->getNumIncomingValues() == 2 && 125 "Function can only handle blocks with 2 predecessors!"); 126 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 127 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 128 129 // We can only handle branches. Other control flow will be lowered to 130 // branches if possible anyway. 131 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 132 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 133 if (Pred1Br == 0 || Pred2Br == 0) 134 return 0; 135 136 // Eliminate code duplication by ensuring that Pred1Br is conditional if 137 // either are. 138 if (Pred2Br->isConditional()) { 139 // If both branches are conditional, we don't have an "if statement". In 140 // reality, we could transform this case, but since the condition will be 141 // required anyway, we stand no chance of eliminating it, so the xform is 142 // probably not profitable. 143 if (Pred1Br->isConditional()) 144 return 0; 145 146 std::swap(Pred1, Pred2); 147 std::swap(Pred1Br, Pred2Br); 148 } 149 150 if (Pred1Br->isConditional()) { 151 // The only thing we have to watch out for here is to make sure that Pred2 152 // doesn't have incoming edges from other blocks. If it does, the condition 153 // doesn't dominate BB. 154 if (Pred2->getSinglePredecessor() == 0) 155 return 0; 156 157 // If we found a conditional branch predecessor, make sure that it branches 158 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 159 if (Pred1Br->getSuccessor(0) == BB && 160 Pred1Br->getSuccessor(1) == Pred2) { 161 IfTrue = Pred1; 162 IfFalse = Pred2; 163 } else if (Pred1Br->getSuccessor(0) == Pred2 && 164 Pred1Br->getSuccessor(1) == BB) { 165 IfTrue = Pred2; 166 IfFalse = Pred1; 167 } else { 168 // We know that one arm of the conditional goes to BB, so the other must 169 // go somewhere unrelated, and this must not be an "if statement". 170 return 0; 171 } 172 173 return Pred1Br->getCondition(); 174 } 175 176 // Ok, if we got here, both predecessors end with an unconditional branch to 177 // BB. Don't panic! If both blocks only have a single (identical) 178 // predecessor, and THAT is a conditional branch, then we're all ok! 179 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 180 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 181 return 0; 182 183 // Otherwise, if this is a conditional branch, then we can use it! 184 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 185 if (BI == 0) return 0; 186 187 assert(BI->isConditional() && "Two successors but not conditional?"); 188 if (BI->getSuccessor(0) == Pred1) { 189 IfTrue = Pred1; 190 IfFalse = Pred2; 191 } else { 192 IfTrue = Pred2; 193 IfFalse = Pred1; 194 } 195 return BI->getCondition(); 196 } 197 198 /// DominatesMergePoint - If we have a merge point of an "if condition" as 199 /// accepted above, return true if the specified value dominates the block. We 200 /// don't handle the true generality of domination here, just a special case 201 /// which works well enough for us. 202 /// 203 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 204 /// see if V (which must be an instruction) is cheap to compute and is 205 /// non-trapping. If both are true, the instruction is inserted into the set 206 /// and true is returned. 207 static bool DominatesMergePoint(Value *V, BasicBlock *BB, 208 SmallPtrSet<Instruction*, 4> *AggressiveInsts) { 209 Instruction *I = dyn_cast<Instruction>(V); 210 if (!I) { 211 // Non-instructions all dominate instructions, but not all constantexprs 212 // can be executed unconditionally. 213 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 214 if (C->canTrap()) 215 return false; 216 return true; 217 } 218 BasicBlock *PBB = I->getParent(); 219 220 // We don't want to allow weird loops that might have the "if condition" in 221 // the bottom of this block. 222 if (PBB == BB) return false; 223 224 // If this instruction is defined in a block that contains an unconditional 225 // branch to BB, then it must be in the 'conditional' part of the "if 226 // statement". If not, it definitely dominates the region. 227 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 228 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 229 return true; 230 231 // If we aren't allowing aggressive promotion anymore, then don't consider 232 // instructions in the 'if region'. 233 if (AggressiveInsts == 0) return false; 234 235 // Okay, it looks like the instruction IS in the "condition". Check to 236 // see if it's a cheap instruction to unconditionally compute, and if it 237 // only uses stuff defined outside of the condition. If so, hoist it out. 238 if (!I->isSafeToSpeculativelyExecute()) 239 return false; 240 241 switch (I->getOpcode()) { 242 default: return false; // Cannot hoist this out safely. 243 case Instruction::Load: 244 // We have to check to make sure there are no instructions before the 245 // load in its basic block, as we are going to hoist the load out to its 246 // predecessor. 247 if (PBB->getFirstNonPHIOrDbg() != I) 248 return false; 249 break; 250 case Instruction::GetElementPtr: 251 // GEPs are cheap if all indices are constant. 252 if (!cast<GetElementPtrInst>(I)->hasAllConstantIndices()) 253 return false; 254 break; 255 case Instruction::Add: 256 case Instruction::Sub: 257 case Instruction::And: 258 case Instruction::Or: 259 case Instruction::Xor: 260 case Instruction::Shl: 261 case Instruction::LShr: 262 case Instruction::AShr: 263 case Instruction::ICmp: 264 break; // These are all cheap and non-trapping instructions. 265 } 266 267 // Okay, we can only really hoist these out if their operands are not 268 // defined in the conditional region. 269 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 270 if (!DominatesMergePoint(*i, BB, 0)) 271 return false; 272 // Okay, it's safe to do this! Remember this instruction. 273 AggressiveInsts->insert(I); 274 return true; 275 } 276 277 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 278 /// and PointerNullValue. Return NULL if value is not a constant int. 279 static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { 280 // Normal constant int. 281 ConstantInt *CI = dyn_cast<ConstantInt>(V); 282 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 283 return CI; 284 285 // This is some kind of pointer constant. Turn it into a pointer-sized 286 // ConstantInt if possible. 287 const IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); 288 289 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 290 if (isa<ConstantPointerNull>(V)) 291 return ConstantInt::get(PtrTy, 0); 292 293 // IntToPtr const int. 294 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 295 if (CE->getOpcode() == Instruction::IntToPtr) 296 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 297 // The constant is very likely to have the right type already. 298 if (CI->getType() == PtrTy) 299 return CI; 300 else 301 return cast<ConstantInt> 302 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 303 } 304 return 0; 305 } 306 307 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 308 /// collection of icmp eq/ne instructions that compare a value against a 309 /// constant, return the value being compared, and stick the constant into the 310 /// Values vector. 311 static Value * 312 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 313 const TargetData *TD, bool isEQ, unsigned &UsedICmps) { 314 Instruction *I = dyn_cast<Instruction>(V); 315 if (I == 0) return 0; 316 317 // If this is an icmp against a constant, handle this as one of the cases. 318 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 319 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 320 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 321 UsedICmps++; 322 Vals.push_back(C); 323 return I->getOperand(0); 324 } 325 326 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 327 // the set. 328 ConstantRange Span = 329 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 330 331 // If this is an and/!= check then we want to optimize "x ugt 2" into 332 // x != 0 && x != 1. 333 if (!isEQ) 334 Span = Span.inverse(); 335 336 // If there are a ton of values, we don't want to make a ginormous switch. 337 if (Span.getSetSize().ugt(8) || Span.isEmptySet() || 338 // We don't handle wrapped sets yet. 339 Span.isWrappedSet()) 340 return 0; 341 342 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 343 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 344 UsedICmps++; 345 return I->getOperand(0); 346 } 347 return 0; 348 } 349 350 // Otherwise, we can only handle an | or &, depending on isEQ. 351 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 352 return 0; 353 354 unsigned NumValsBeforeLHS = Vals.size(); 355 unsigned UsedICmpsBeforeLHS = UsedICmps; 356 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 357 isEQ, UsedICmps)) { 358 unsigned NumVals = Vals.size(); 359 unsigned UsedICmpsBeforeRHS = UsedICmps; 360 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 361 isEQ, UsedICmps)) { 362 if (LHS == RHS) 363 return LHS; 364 Vals.resize(NumVals); 365 UsedICmps = UsedICmpsBeforeRHS; 366 } 367 368 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 369 // set it and return success. 370 if (Extra == 0 || Extra == I->getOperand(1)) { 371 Extra = I->getOperand(1); 372 return LHS; 373 } 374 375 Vals.resize(NumValsBeforeLHS); 376 UsedICmps = UsedICmpsBeforeLHS; 377 return 0; 378 } 379 380 // If the LHS can't be folded in, but Extra is available and RHS can, try to 381 // use LHS as Extra. 382 if (Extra == 0 || Extra == I->getOperand(0)) { 383 Value *OldExtra = Extra; 384 Extra = I->getOperand(0); 385 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 386 isEQ, UsedICmps)) 387 return RHS; 388 assert(Vals.size() == NumValsBeforeLHS); 389 Extra = OldExtra; 390 } 391 392 return 0; 393 } 394 395 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 396 Instruction* Cond = 0; 397 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 398 Cond = dyn_cast<Instruction>(SI->getCondition()); 399 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 400 if (BI->isConditional()) 401 Cond = dyn_cast<Instruction>(BI->getCondition()); 402 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 403 Cond = dyn_cast<Instruction>(IBI->getAddress()); 404 } 405 406 TI->eraseFromParent(); 407 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 408 } 409 410 /// isValueEqualityComparison - Return true if the specified terminator checks 411 /// to see if a value is equal to constant integer value. 412 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 413 Value *CV = 0; 414 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 415 // Do not permit merging of large switch instructions into their 416 // predecessors unless there is only one predecessor. 417 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 418 pred_end(SI->getParent())) <= 128) 419 CV = SI->getCondition(); 420 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 421 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 422 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 423 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 424 ICI->getPredicate() == ICmpInst::ICMP_NE) && 425 GetConstantInt(ICI->getOperand(1), TD)) 426 CV = ICI->getOperand(0); 427 428 // Unwrap any lossless ptrtoint cast. 429 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 430 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 431 CV = PTII->getOperand(0); 432 return CV; 433 } 434 435 /// GetValueEqualityComparisonCases - Given a value comparison instruction, 436 /// decode all of the 'cases' that it represents and return the 'default' block. 437 BasicBlock *SimplifyCFGOpt:: 438 GetValueEqualityComparisonCases(TerminatorInst *TI, 439 std::vector<std::pair<ConstantInt*, 440 BasicBlock*> > &Cases) { 441 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 442 Cases.reserve(SI->getNumCases()); 443 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 444 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i))); 445 return SI->getDefaultDest(); 446 } 447 448 BranchInst *BI = cast<BranchInst>(TI); 449 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 450 Cases.push_back(std::make_pair(GetConstantInt(ICI->getOperand(1), TD), 451 BI->getSuccessor(ICI->getPredicate() == 452 ICmpInst::ICMP_NE))); 453 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 454 } 455 456 457 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 458 /// in the list that match the specified block. 459 static void EliminateBlockCases(BasicBlock *BB, 460 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 461 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 462 if (Cases[i].second == BB) { 463 Cases.erase(Cases.begin()+i); 464 --i; --e; 465 } 466 } 467 468 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 469 /// well. 470 static bool 471 ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 472 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 473 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 474 475 // Make V1 be smaller than V2. 476 if (V1->size() > V2->size()) 477 std::swap(V1, V2); 478 479 if (V1->size() == 0) return false; 480 if (V1->size() == 1) { 481 // Just scan V2. 482 ConstantInt *TheVal = (*V1)[0].first; 483 for (unsigned i = 0, e = V2->size(); i != e; ++i) 484 if (TheVal == (*V2)[i].first) 485 return true; 486 } 487 488 // Otherwise, just sort both lists and compare element by element. 489 array_pod_sort(V1->begin(), V1->end()); 490 array_pod_sort(V2->begin(), V2->end()); 491 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 492 while (i1 != e1 && i2 != e2) { 493 if ((*V1)[i1].first == (*V2)[i2].first) 494 return true; 495 if ((*V1)[i1].first < (*V2)[i2].first) 496 ++i1; 497 else 498 ++i2; 499 } 500 return false; 501 } 502 503 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 504 /// terminator instruction and its block is known to only have a single 505 /// predecessor block, check to see if that predecessor is also a value 506 /// comparison with the same value, and if that comparison determines the 507 /// outcome of this comparison. If so, simplify TI. This does a very limited 508 /// form of jump threading. 509 bool SimplifyCFGOpt:: 510 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 511 BasicBlock *Pred) { 512 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 513 if (!PredVal) return false; // Not a value comparison in predecessor. 514 515 Value *ThisVal = isValueEqualityComparison(TI); 516 assert(ThisVal && "This isn't a value comparison!!"); 517 if (ThisVal != PredVal) return false; // Different predicates. 518 519 // Find out information about when control will move from Pred to TI's block. 520 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 521 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 522 PredCases); 523 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 524 525 // Find information about how control leaves this block. 526 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 527 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 528 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 529 530 // If TI's block is the default block from Pred's comparison, potentially 531 // simplify TI based on this knowledge. 532 if (PredDef == TI->getParent()) { 533 // If we are here, we know that the value is none of those cases listed in 534 // PredCases. If there are any cases in ThisCases that are in PredCases, we 535 // can simplify TI. 536 if (!ValuesOverlap(PredCases, ThisCases)) 537 return false; 538 539 if (isa<BranchInst>(TI)) { 540 // Okay, one of the successors of this condbr is dead. Convert it to a 541 // uncond br. 542 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 543 // Insert the new branch. 544 Instruction *NI = BranchInst::Create(ThisDef, TI); 545 (void) NI; 546 547 // Remove PHI node entries for the dead edge. 548 ThisCases[0].second->removePredecessor(TI->getParent()); 549 550 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 551 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 552 553 EraseTerminatorInstAndDCECond(TI); 554 return true; 555 } 556 557 SwitchInst *SI = cast<SwitchInst>(TI); 558 // Okay, TI has cases that are statically dead, prune them away. 559 SmallPtrSet<Constant*, 16> DeadCases; 560 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 561 DeadCases.insert(PredCases[i].first); 562 563 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 564 << "Through successor TI: " << *TI); 565 566 for (unsigned i = SI->getNumCases()-1; i != 0; --i) 567 if (DeadCases.count(SI->getCaseValue(i))) { 568 SI->getSuccessor(i)->removePredecessor(TI->getParent()); 569 SI->removeCase(i); 570 } 571 572 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 573 return true; 574 } 575 576 // Otherwise, TI's block must correspond to some matched value. Find out 577 // which value (or set of values) this is. 578 ConstantInt *TIV = 0; 579 BasicBlock *TIBB = TI->getParent(); 580 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 581 if (PredCases[i].second == TIBB) { 582 if (TIV != 0) 583 return false; // Cannot handle multiple values coming to this block. 584 TIV = PredCases[i].first; 585 } 586 assert(TIV && "No edge from pred to succ?"); 587 588 // Okay, we found the one constant that our value can be if we get into TI's 589 // BB. Find out which successor will unconditionally be branched to. 590 BasicBlock *TheRealDest = 0; 591 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 592 if (ThisCases[i].first == TIV) { 593 TheRealDest = ThisCases[i].second; 594 break; 595 } 596 597 // If not handled by any explicit cases, it is handled by the default case. 598 if (TheRealDest == 0) TheRealDest = ThisDef; 599 600 // Remove PHI node entries for dead edges. 601 BasicBlock *CheckEdge = TheRealDest; 602 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 603 if (*SI != CheckEdge) 604 (*SI)->removePredecessor(TIBB); 605 else 606 CheckEdge = 0; 607 608 // Insert the new branch. 609 Instruction *NI = BranchInst::Create(TheRealDest, TI); 610 (void) NI; 611 612 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 613 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 614 615 EraseTerminatorInstAndDCECond(TI); 616 return true; 617 } 618 619 namespace { 620 /// ConstantIntOrdering - This class implements a stable ordering of constant 621 /// integers that does not depend on their address. This is important for 622 /// applications that sort ConstantInt's to ensure uniqueness. 623 struct ConstantIntOrdering { 624 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 625 return LHS->getValue().ult(RHS->getValue()); 626 } 627 }; 628 } 629 630 static int ConstantIntSortPredicate(const void *P1, const void *P2) { 631 const ConstantInt *LHS = *(const ConstantInt**)P1; 632 const ConstantInt *RHS = *(const ConstantInt**)P2; 633 if (LHS->getValue().ult(RHS->getValue())) 634 return 1; 635 if (LHS->getValue() == RHS->getValue()) 636 return 0; 637 return -1; 638 } 639 640 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value 641 /// equality comparison instruction (either a switch or a branch on "X == c"). 642 /// See if any of the predecessors of the terminator block are value comparisons 643 /// on the same value. If so, and if safe to do so, fold them together. 644 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { 645 BasicBlock *BB = TI->getParent(); 646 Value *CV = isValueEqualityComparison(TI); // CondVal 647 assert(CV && "Not a comparison?"); 648 bool Changed = false; 649 650 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 651 while (!Preds.empty()) { 652 BasicBlock *Pred = Preds.pop_back_val(); 653 654 // See if the predecessor is a comparison with the same value. 655 TerminatorInst *PTI = Pred->getTerminator(); 656 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 657 658 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 659 // Figure out which 'cases' to copy from SI to PSI. 660 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 661 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 662 663 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 664 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 665 666 // Based on whether the default edge from PTI goes to BB or not, fill in 667 // PredCases and PredDefault with the new switch cases we would like to 668 // build. 669 SmallVector<BasicBlock*, 8> NewSuccessors; 670 671 if (PredDefault == BB) { 672 // If this is the default destination from PTI, only the edges in TI 673 // that don't occur in PTI, or that branch to BB will be activated. 674 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 675 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 676 if (PredCases[i].second != BB) 677 PTIHandled.insert(PredCases[i].first); 678 else { 679 // The default destination is BB, we don't need explicit targets. 680 std::swap(PredCases[i], PredCases.back()); 681 PredCases.pop_back(); 682 --i; --e; 683 } 684 685 // Reconstruct the new switch statement we will be building. 686 if (PredDefault != BBDefault) { 687 PredDefault->removePredecessor(Pred); 688 PredDefault = BBDefault; 689 NewSuccessors.push_back(BBDefault); 690 } 691 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 692 if (!PTIHandled.count(BBCases[i].first) && 693 BBCases[i].second != BBDefault) { 694 PredCases.push_back(BBCases[i]); 695 NewSuccessors.push_back(BBCases[i].second); 696 } 697 698 } else { 699 // If this is not the default destination from PSI, only the edges 700 // in SI that occur in PSI with a destination of BB will be 701 // activated. 702 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 703 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 704 if (PredCases[i].second == BB) { 705 PTIHandled.insert(PredCases[i].first); 706 std::swap(PredCases[i], PredCases.back()); 707 PredCases.pop_back(); 708 --i; --e; 709 } 710 711 // Okay, now we know which constants were sent to BB from the 712 // predecessor. Figure out where they will all go now. 713 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 714 if (PTIHandled.count(BBCases[i].first)) { 715 // If this is one we are capable of getting... 716 PredCases.push_back(BBCases[i]); 717 NewSuccessors.push_back(BBCases[i].second); 718 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 719 } 720 721 // If there are any constants vectored to BB that TI doesn't handle, 722 // they must go to the default destination of TI. 723 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 724 PTIHandled.begin(), 725 E = PTIHandled.end(); I != E; ++I) { 726 PredCases.push_back(std::make_pair(*I, BBDefault)); 727 NewSuccessors.push_back(BBDefault); 728 } 729 } 730 731 // Okay, at this point, we know which new successor Pred will get. Make 732 // sure we update the number of entries in the PHI nodes for these 733 // successors. 734 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 735 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 736 737 // Convert pointer to int before we switch. 738 if (CV->getType()->isPointerTy()) { 739 assert(TD && "Cannot switch on pointer without TargetData"); 740 CV = new PtrToIntInst(CV, TD->getIntPtrType(CV->getContext()), 741 "magicptr", PTI); 742 } 743 744 // Now that the successors are updated, create the new Switch instruction. 745 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, 746 PredCases.size(), PTI); 747 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 748 NewSI->addCase(PredCases[i].first, PredCases[i].second); 749 750 EraseTerminatorInstAndDCECond(PTI); 751 752 // Okay, last check. If BB is still a successor of PSI, then we must 753 // have an infinite loop case. If so, add an infinitely looping block 754 // to handle the case to preserve the behavior of the code. 755 BasicBlock *InfLoopBlock = 0; 756 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 757 if (NewSI->getSuccessor(i) == BB) { 758 if (InfLoopBlock == 0) { 759 // Insert it at the end of the function, because it's either code, 760 // or it won't matter if it's hot. :) 761 InfLoopBlock = BasicBlock::Create(BB->getContext(), 762 "infloop", BB->getParent()); 763 BranchInst::Create(InfLoopBlock, InfLoopBlock); 764 } 765 NewSI->setSuccessor(i, InfLoopBlock); 766 } 767 768 Changed = true; 769 } 770 } 771 return Changed; 772 } 773 774 // isSafeToHoistInvoke - If we would need to insert a select that uses the 775 // value of this invoke (comments in HoistThenElseCodeToIf explain why we 776 // would need to do this), we can't hoist the invoke, as there is nowhere 777 // to put the select in this case. 778 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 779 Instruction *I1, Instruction *I2) { 780 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 781 PHINode *PN; 782 for (BasicBlock::iterator BBI = SI->begin(); 783 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 784 Value *BB1V = PN->getIncomingValueForBlock(BB1); 785 Value *BB2V = PN->getIncomingValueForBlock(BB2); 786 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 787 return false; 788 } 789 } 790 } 791 return true; 792 } 793 794 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 795 /// BB2, hoist any common code in the two blocks up into the branch block. The 796 /// caller of this function guarantees that BI's block dominates BB1 and BB2. 797 static bool HoistThenElseCodeToIf(BranchInst *BI) { 798 // This does very trivial matching, with limited scanning, to find identical 799 // instructions in the two blocks. In particular, we don't want to get into 800 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 801 // such, we currently just scan for obviously identical instructions in an 802 // identical order. 803 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 804 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 805 806 BasicBlock::iterator BB1_Itr = BB1->begin(); 807 BasicBlock::iterator BB2_Itr = BB2->begin(); 808 809 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 810 while (isa<DbgInfoIntrinsic>(I1)) 811 I1 = BB1_Itr++; 812 while (isa<DbgInfoIntrinsic>(I2)) 813 I2 = BB2_Itr++; 814 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) || 815 !I1->isIdenticalToWhenDefined(I2) || 816 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 817 return false; 818 819 // If we get here, we can hoist at least one instruction. 820 BasicBlock *BIParent = BI->getParent(); 821 822 do { 823 // If we are hoisting the terminator instruction, don't move one (making a 824 // broken BB), instead clone it, and remove BI. 825 if (isa<TerminatorInst>(I1)) 826 goto HoistTerminator; 827 828 // For a normal instruction, we just move one to right before the branch, 829 // then replace all uses of the other with the first. Finally, we remove 830 // the now redundant second instruction. 831 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 832 if (!I2->use_empty()) 833 I2->replaceAllUsesWith(I1); 834 I1->intersectOptionalDataWith(I2); 835 I2->eraseFromParent(); 836 837 I1 = BB1_Itr++; 838 while (isa<DbgInfoIntrinsic>(I1)) 839 I1 = BB1_Itr++; 840 I2 = BB2_Itr++; 841 while (isa<DbgInfoIntrinsic>(I2)) 842 I2 = BB2_Itr++; 843 } while (I1->getOpcode() == I2->getOpcode() && 844 I1->isIdenticalToWhenDefined(I2)); 845 846 return true; 847 848 HoistTerminator: 849 // It may not be possible to hoist an invoke. 850 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 851 return true; 852 853 // Okay, it is safe to hoist the terminator. 854 Instruction *NT = I1->clone(); 855 BIParent->getInstList().insert(BI, NT); 856 if (!NT->getType()->isVoidTy()) { 857 I1->replaceAllUsesWith(NT); 858 I2->replaceAllUsesWith(NT); 859 NT->takeName(I1); 860 } 861 862 // Hoisting one of the terminators from our successor is a great thing. 863 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 864 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 865 // nodes, so we insert select instruction to compute the final result. 866 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 867 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 868 PHINode *PN; 869 for (BasicBlock::iterator BBI = SI->begin(); 870 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 871 Value *BB1V = PN->getIncomingValueForBlock(BB1); 872 Value *BB2V = PN->getIncomingValueForBlock(BB2); 873 if (BB1V == BB2V) continue; 874 875 // These values do not agree. Insert a select instruction before NT 876 // that determines the right value. 877 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 878 if (SI == 0) 879 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 880 BB1V->getName()+"."+BB2V->getName(), NT); 881 // Make the PHI node use the select for all incoming values for BB1/BB2 882 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 883 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 884 PN->setIncomingValue(i, SI); 885 } 886 } 887 888 // Update any PHI nodes in our new successors. 889 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 890 AddPredecessorToBlock(*SI, BIParent, BB1); 891 892 EraseTerminatorInstAndDCECond(BI); 893 return true; 894 } 895 896 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 897 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code 898 /// (for now, restricted to a single instruction that's side effect free) from 899 /// the BB1 into the branch block to speculatively execute it. 900 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 901 // Only speculatively execution a single instruction (not counting the 902 // terminator) for now. 903 Instruction *HInst = NULL; 904 Instruction *Term = BB1->getTerminator(); 905 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 906 BBI != BBE; ++BBI) { 907 Instruction *I = BBI; 908 // Skip debug info. 909 if (isa<DbgInfoIntrinsic>(I)) continue; 910 if (I == Term) break; 911 912 if (HInst) 913 return false; 914 HInst = I; 915 } 916 if (!HInst) 917 return false; 918 919 // Be conservative for now. FP select instruction can often be expensive. 920 Value *BrCond = BI->getCondition(); 921 if (isa<FCmpInst>(BrCond)) 922 return false; 923 924 // If BB1 is actually on the false edge of the conditional branch, remember 925 // to swap the select operands later. 926 bool Invert = false; 927 if (BB1 != BI->getSuccessor(0)) { 928 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 929 Invert = true; 930 } 931 932 // Turn 933 // BB: 934 // %t1 = icmp 935 // br i1 %t1, label %BB1, label %BB2 936 // BB1: 937 // %t3 = add %t2, c 938 // br label BB2 939 // BB2: 940 // => 941 // BB: 942 // %t1 = icmp 943 // %t4 = add %t2, c 944 // %t3 = select i1 %t1, %t2, %t3 945 switch (HInst->getOpcode()) { 946 default: return false; // Not safe / profitable to hoist. 947 case Instruction::Add: 948 case Instruction::Sub: 949 // Not worth doing for vector ops. 950 if (HInst->getType()->isVectorTy()) 951 return false; 952 break; 953 case Instruction::And: 954 case Instruction::Or: 955 case Instruction::Xor: 956 case Instruction::Shl: 957 case Instruction::LShr: 958 case Instruction::AShr: 959 // Don't mess with vector operations. 960 if (HInst->getType()->isVectorTy()) 961 return false; 962 break; // These are all cheap and non-trapping instructions. 963 } 964 965 // If the instruction is obviously dead, don't try to predicate it. 966 if (HInst->use_empty()) { 967 HInst->eraseFromParent(); 968 return true; 969 } 970 971 // Can we speculatively execute the instruction? And what is the value 972 // if the condition is false? Consider the phi uses, if the incoming value 973 // from the "if" block are all the same V, then V is the value of the 974 // select if the condition is false. 975 BasicBlock *BIParent = BI->getParent(); 976 SmallVector<PHINode*, 4> PHIUses; 977 Value *FalseV = NULL; 978 979 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 980 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end(); 981 UI != E; ++UI) { 982 // Ignore any user that is not a PHI node in BB2. These can only occur in 983 // unreachable blocks, because they would not be dominated by the instr. 984 PHINode *PN = dyn_cast<PHINode>(*UI); 985 if (!PN || PN->getParent() != BB2) 986 return false; 987 PHIUses.push_back(PN); 988 989 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 990 if (!FalseV) 991 FalseV = PHIV; 992 else if (FalseV != PHIV) 993 return false; // Inconsistent value when condition is false. 994 } 995 996 assert(FalseV && "Must have at least one user, and it must be a PHI"); 997 998 // Do not hoist the instruction if any of its operands are defined but not 999 // used in this BB. The transformation will prevent the operand from 1000 // being sunk into the use block. 1001 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1002 i != e; ++i) { 1003 Instruction *OpI = dyn_cast<Instruction>(*i); 1004 if (OpI && OpI->getParent() == BIParent && 1005 !OpI->isUsedInBasicBlock(BIParent)) 1006 return false; 1007 } 1008 1009 // If we get here, we can hoist the instruction. Try to place it 1010 // before the icmp instruction preceding the conditional branch. 1011 BasicBlock::iterator InsertPos = BI; 1012 if (InsertPos != BIParent->begin()) 1013 --InsertPos; 1014 // Skip debug info between condition and branch. 1015 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos)) 1016 --InsertPos; 1017 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) { 1018 SmallPtrSet<Instruction *, 4> BB1Insns; 1019 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 1020 BB1I != BB1E; ++BB1I) 1021 BB1Insns.insert(BB1I); 1022 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 1023 UI != UE; ++UI) { 1024 Instruction *Use = cast<Instruction>(*UI); 1025 if (!BB1Insns.count(Use)) continue; 1026 1027 // If BrCond uses the instruction that place it just before 1028 // branch instruction. 1029 InsertPos = BI; 1030 break; 1031 } 1032 } else 1033 InsertPos = BI; 1034 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst); 1035 1036 // Create a select whose true value is the speculatively executed value and 1037 // false value is the previously determined FalseV. 1038 SelectInst *SI; 1039 if (Invert) 1040 SI = SelectInst::Create(BrCond, FalseV, HInst, 1041 FalseV->getName() + "." + HInst->getName(), BI); 1042 else 1043 SI = SelectInst::Create(BrCond, HInst, FalseV, 1044 HInst->getName() + "." + FalseV->getName(), BI); 1045 1046 // Make the PHI node use the select for all incoming values for "then" and 1047 // "if" blocks. 1048 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1049 PHINode *PN = PHIUses[i]; 1050 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1051 if (PN->getIncomingBlock(j) == BB1 || PN->getIncomingBlock(j) == BIParent) 1052 PN->setIncomingValue(j, SI); 1053 } 1054 1055 ++NumSpeculations; 1056 return true; 1057 } 1058 1059 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1060 /// across this block. 1061 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1062 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1063 unsigned Size = 0; 1064 1065 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1066 if (isa<DbgInfoIntrinsic>(BBI)) 1067 continue; 1068 if (Size > 10) return false; // Don't clone large BB's. 1069 ++Size; 1070 1071 // We can only support instructions that do not define values that are 1072 // live outside of the current basic block. 1073 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1074 UI != E; ++UI) { 1075 Instruction *U = cast<Instruction>(*UI); 1076 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1077 } 1078 1079 // Looks ok, continue checking. 1080 } 1081 1082 return true; 1083 } 1084 1085 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1086 /// that is defined in the same block as the branch and if any PHI entries are 1087 /// constants, thread edges corresponding to that entry to be branches to their 1088 /// ultimate destination. 1089 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { 1090 BasicBlock *BB = BI->getParent(); 1091 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1092 // NOTE: we currently cannot transform this case if the PHI node is used 1093 // outside of the block. 1094 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1095 return false; 1096 1097 // Degenerate case of a single entry PHI. 1098 if (PN->getNumIncomingValues() == 1) { 1099 FoldSingleEntryPHINodes(PN->getParent()); 1100 return true; 1101 } 1102 1103 // Now we know that this block has multiple preds and two succs. 1104 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1105 1106 // Okay, this is a simple enough basic block. See if any phi values are 1107 // constants. 1108 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1109 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1110 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1111 1112 // Okay, we now know that all edges from PredBB should be revectored to 1113 // branch to RealDest. 1114 BasicBlock *PredBB = PN->getIncomingBlock(i); 1115 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1116 1117 if (RealDest == BB) continue; // Skip self loops. 1118 1119 // The dest block might have PHI nodes, other predecessors and other 1120 // difficult cases. Instead of being smart about this, just insert a new 1121 // block that jumps to the destination block, effectively splitting 1122 // the edge we are about to create. 1123 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1124 RealDest->getName()+".critedge", 1125 RealDest->getParent(), RealDest); 1126 BranchInst::Create(RealDest, EdgeBB); 1127 1128 // Update PHI nodes. 1129 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1130 1131 // BB may have instructions that are being threaded over. Clone these 1132 // instructions into EdgeBB. We know that there will be no uses of the 1133 // cloned instructions outside of EdgeBB. 1134 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1135 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1136 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1137 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1138 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1139 continue; 1140 } 1141 // Clone the instruction. 1142 Instruction *N = BBI->clone(); 1143 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1144 1145 // Update operands due to translation. 1146 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1147 i != e; ++i) { 1148 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1149 if (PI != TranslateMap.end()) 1150 *i = PI->second; 1151 } 1152 1153 // Check for trivial simplification. 1154 if (Value *V = SimplifyInstruction(N, TD)) { 1155 TranslateMap[BBI] = V; 1156 delete N; // Instruction folded away, don't need actual inst 1157 } else { 1158 // Insert the new instruction into its new home. 1159 EdgeBB->getInstList().insert(InsertPt, N); 1160 if (!BBI->use_empty()) 1161 TranslateMap[BBI] = N; 1162 } 1163 } 1164 1165 // Loop over all of the edges from PredBB to BB, changing them to branch 1166 // to EdgeBB instead. 1167 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1168 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1169 if (PredBBTI->getSuccessor(i) == BB) { 1170 BB->removePredecessor(PredBB); 1171 PredBBTI->setSuccessor(i, EdgeBB); 1172 } 1173 1174 // Recurse, simplifying any other constants. 1175 return FoldCondBranchOnPHI(BI, TD) | true; 1176 } 1177 1178 return false; 1179 } 1180 1181 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1182 /// PHI node, see if we can eliminate it. 1183 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { 1184 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1185 // statement", which has a very simple dominance structure. Basically, we 1186 // are trying to find the condition that is being branched on, which 1187 // subsequently causes this merge to happen. We really want control 1188 // dependence information for this check, but simplifycfg can't keep it up 1189 // to date, and this catches most of the cases we care about anyway. 1190 BasicBlock *BB = PN->getParent(); 1191 BasicBlock *IfTrue, *IfFalse; 1192 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1193 if (!IfCond || 1194 // Don't bother if the branch will be constant folded trivially. 1195 isa<ConstantInt>(IfCond)) 1196 return false; 1197 1198 // Okay, we found that we can merge this two-entry phi node into a select. 1199 // Doing so would require us to fold *all* two entry phi nodes in this block. 1200 // At some point this becomes non-profitable (particularly if the target 1201 // doesn't support cmov's). Only do this transformation if there are two or 1202 // fewer PHI nodes in this block. 1203 unsigned NumPhis = 0; 1204 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1205 if (NumPhis > 2) 1206 return false; 1207 1208 // Loop over the PHI's seeing if we can promote them all to select 1209 // instructions. While we are at it, keep track of the instructions 1210 // that need to be moved to the dominating block. 1211 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1212 1213 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1214 PHINode *PN = cast<PHINode>(II++); 1215 if (Value *V = SimplifyInstruction(PN, TD)) { 1216 PN->replaceAllUsesWith(V); 1217 PN->eraseFromParent(); 1218 continue; 1219 } 1220 1221 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts) || 1222 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts)) 1223 return false; 1224 } 1225 1226 // If we folded the the first phi, PN dangles at this point. Refresh it. If 1227 // we ran out of PHIs then we simplified them all. 1228 PN = dyn_cast<PHINode>(BB->begin()); 1229 if (PN == 0) return true; 1230 1231 // Don't fold i1 branches on PHIs which contain binary operators. These can 1232 // often be turned into switches and other things. 1233 if (PN->getType()->isIntegerTy(1) && 1234 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1235 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1236 isa<BinaryOperator>(IfCond))) 1237 return false; 1238 1239 // If we all PHI nodes are promotable, check to make sure that all 1240 // instructions in the predecessor blocks can be promoted as well. If 1241 // not, we won't be able to get rid of the control flow, so it's not 1242 // worth promoting to select instructions. 1243 BasicBlock *DomBlock = 0; 1244 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1245 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1246 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1247 IfBlock1 = 0; 1248 } else { 1249 DomBlock = *pred_begin(IfBlock1); 1250 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1251 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1252 // This is not an aggressive instruction that we can promote. 1253 // Because of this, we won't be able to get rid of the control 1254 // flow, so the xform is not worth it. 1255 return false; 1256 } 1257 } 1258 1259 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1260 IfBlock2 = 0; 1261 } else { 1262 DomBlock = *pred_begin(IfBlock2); 1263 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1264 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1265 // This is not an aggressive instruction that we can promote. 1266 // Because of this, we won't be able to get rid of the control 1267 // flow, so the xform is not worth it. 1268 return false; 1269 } 1270 } 1271 1272 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1273 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1274 1275 // If we can still promote the PHI nodes after this gauntlet of tests, 1276 // do all of the PHI's now. 1277 Instruction *InsertPt = DomBlock->getTerminator(); 1278 1279 // Move all 'aggressive' instructions, which are defined in the 1280 // conditional parts of the if's up to the dominating block. 1281 if (IfBlock1) 1282 DomBlock->getInstList().splice(InsertPt, 1283 IfBlock1->getInstList(), IfBlock1->begin(), 1284 IfBlock1->getTerminator()); 1285 if (IfBlock2) 1286 DomBlock->getInstList().splice(InsertPt, 1287 IfBlock2->getInstList(), IfBlock2->begin(), 1288 IfBlock2->getTerminator()); 1289 1290 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1291 // Change the PHI node into a select instruction. 1292 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1293 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1294 1295 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", InsertPt); 1296 PN->replaceAllUsesWith(NV); 1297 NV->takeName(PN); 1298 PN->eraseFromParent(); 1299 } 1300 1301 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1302 // has been flattened. Change DomBlock to jump directly to our new block to 1303 // avoid other simplifycfg's kicking in on the diamond. 1304 TerminatorInst *OldTI = DomBlock->getTerminator(); 1305 BranchInst::Create(BB, OldTI); 1306 OldTI->eraseFromParent(); 1307 return true; 1308 } 1309 1310 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1311 /// to two returning blocks, try to merge them together into one return, 1312 /// introducing a select if the return values disagree. 1313 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1314 assert(BI->isConditional() && "Must be a conditional branch"); 1315 BasicBlock *TrueSucc = BI->getSuccessor(0); 1316 BasicBlock *FalseSucc = BI->getSuccessor(1); 1317 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1318 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1319 1320 // Check to ensure both blocks are empty (just a return) or optionally empty 1321 // with PHI nodes. If there are other instructions, merging would cause extra 1322 // computation on one path or the other. 1323 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1324 return false; 1325 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1326 return false; 1327 1328 // Okay, we found a branch that is going to two return nodes. If 1329 // there is no return value for this function, just change the 1330 // branch into a return. 1331 if (FalseRet->getNumOperands() == 0) { 1332 TrueSucc->removePredecessor(BI->getParent()); 1333 FalseSucc->removePredecessor(BI->getParent()); 1334 ReturnInst::Create(BI->getContext(), 0, BI); 1335 EraseTerminatorInstAndDCECond(BI); 1336 return true; 1337 } 1338 1339 // Otherwise, figure out what the true and false return values are 1340 // so we can insert a new select instruction. 1341 Value *TrueValue = TrueRet->getReturnValue(); 1342 Value *FalseValue = FalseRet->getReturnValue(); 1343 1344 // Unwrap any PHI nodes in the return blocks. 1345 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1346 if (TVPN->getParent() == TrueSucc) 1347 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1348 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1349 if (FVPN->getParent() == FalseSucc) 1350 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1351 1352 // In order for this transformation to be safe, we must be able to 1353 // unconditionally execute both operands to the return. This is 1354 // normally the case, but we could have a potentially-trapping 1355 // constant expression that prevents this transformation from being 1356 // safe. 1357 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1358 if (TCV->canTrap()) 1359 return false; 1360 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1361 if (FCV->canTrap()) 1362 return false; 1363 1364 // Okay, we collected all the mapped values and checked them for sanity, and 1365 // defined to really do this transformation. First, update the CFG. 1366 TrueSucc->removePredecessor(BI->getParent()); 1367 FalseSucc->removePredecessor(BI->getParent()); 1368 1369 // Insert select instructions where needed. 1370 Value *BrCond = BI->getCondition(); 1371 if (TrueValue) { 1372 // Insert a select if the results differ. 1373 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1374 } else if (isa<UndefValue>(TrueValue)) { 1375 TrueValue = FalseValue; 1376 } else { 1377 TrueValue = SelectInst::Create(BrCond, TrueValue, 1378 FalseValue, "retval", BI); 1379 } 1380 } 1381 1382 Value *RI = !TrueValue ? 1383 ReturnInst::Create(BI->getContext(), BI) : 1384 ReturnInst::Create(BI->getContext(), TrueValue, BI); 1385 (void) RI; 1386 1387 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1388 << "\n " << *BI << "NewRet = " << *RI 1389 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1390 1391 EraseTerminatorInstAndDCECond(BI); 1392 1393 return true; 1394 } 1395 1396 /// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, 1397 /// and if a predecessor branches to us and one of our successors, fold the 1398 /// setcc into the predecessor and use logical operations to pick the right 1399 /// destination. 1400 bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1401 BasicBlock *BB = BI->getParent(); 1402 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1403 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1404 Cond->getParent() != BB || !Cond->hasOneUse()) 1405 return false; 1406 1407 // Only allow this if the condition is a simple instruction that can be 1408 // executed unconditionally. It must be in the same block as the branch, and 1409 // must be at the front of the block. 1410 BasicBlock::iterator FrontIt = BB->front(); 1411 // Ignore dbg intrinsics. 1412 while (isa<DbgInfoIntrinsic>(FrontIt)) 1413 ++FrontIt; 1414 1415 // Allow a single instruction to be hoisted in addition to the compare 1416 // that feeds the branch. We later ensure that any values that _it_ uses 1417 // were also live in the predecessor, so that we don't unnecessarily create 1418 // register pressure or inhibit out-of-order execution. 1419 Instruction *BonusInst = 0; 1420 if (&*FrontIt != Cond && 1421 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1422 FrontIt->isSafeToSpeculativelyExecute()) { 1423 BonusInst = &*FrontIt; 1424 ++FrontIt; 1425 } 1426 1427 // Only a single bonus inst is allowed. 1428 if (&*FrontIt != Cond) 1429 return false; 1430 1431 // Make sure the instruction after the condition is the cond branch. 1432 BasicBlock::iterator CondIt = Cond; ++CondIt; 1433 // Ingore dbg intrinsics. 1434 while(isa<DbgInfoIntrinsic>(CondIt)) 1435 ++CondIt; 1436 if (&*CondIt != BI) { 1437 assert (!isa<DbgInfoIntrinsic>(CondIt) && "Hey do not forget debug info!"); 1438 return false; 1439 } 1440 1441 // Cond is known to be a compare or binary operator. Check to make sure that 1442 // neither operand is a potentially-trapping constant expression. 1443 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1444 if (CE->canTrap()) 1445 return false; 1446 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1447 if (CE->canTrap()) 1448 return false; 1449 1450 1451 // Finally, don't infinitely unroll conditional loops. 1452 BasicBlock *TrueDest = BI->getSuccessor(0); 1453 BasicBlock *FalseDest = BI->getSuccessor(1); 1454 if (TrueDest == BB || FalseDest == BB) 1455 return false; 1456 1457 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1458 BasicBlock *PredBlock = *PI; 1459 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1460 1461 // Check that we have two conditional branches. If there is a PHI node in 1462 // the common successor, verify that the same value flows in from both 1463 // blocks. 1464 if (PBI == 0 || PBI->isUnconditional() || 1465 !SafeToMergeTerminators(BI, PBI)) 1466 continue; 1467 1468 // Ensure that any values used in the bonus instruction are also used 1469 // by the terminator of the predecessor. This means that those values 1470 // must already have been resolved, so we won't be inhibiting the 1471 // out-of-order core by speculating them earlier. 1472 if (BonusInst) { 1473 // Collect the values used by the bonus inst 1474 SmallPtrSet<Value*, 4> UsedValues; 1475 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1476 OE = BonusInst->op_end(); OI != OE; ++OI) { 1477 Value* V = *OI; 1478 if (!isa<Constant>(V)) 1479 UsedValues.insert(V); 1480 } 1481 1482 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1483 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1484 1485 // Walk up to four levels back up the use-def chain of the predecessor's 1486 // terminator to see if all those values were used. The choice of four 1487 // levels is arbitrary, to provide a compile-time-cost bound. 1488 while (!Worklist.empty()) { 1489 std::pair<Value*, unsigned> Pair = Worklist.back(); 1490 Worklist.pop_back(); 1491 1492 if (Pair.second >= 4) continue; 1493 UsedValues.erase(Pair.first); 1494 if (UsedValues.empty()) break; 1495 1496 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 1497 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1498 OI != OE; ++OI) 1499 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1500 } 1501 } 1502 1503 if (!UsedValues.empty()) return false; 1504 } 1505 1506 Instruction::BinaryOps Opc; 1507 bool InvertPredCond = false; 1508 1509 if (PBI->getSuccessor(0) == TrueDest) 1510 Opc = Instruction::Or; 1511 else if (PBI->getSuccessor(1) == FalseDest) 1512 Opc = Instruction::And; 1513 else if (PBI->getSuccessor(0) == FalseDest) 1514 Opc = Instruction::And, InvertPredCond = true; 1515 else if (PBI->getSuccessor(1) == TrueDest) 1516 Opc = Instruction::Or, InvertPredCond = true; 1517 else 1518 continue; 1519 1520 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1521 1522 // If we need to invert the condition in the pred block to match, do so now. 1523 if (InvertPredCond) { 1524 Value *NewCond = PBI->getCondition(); 1525 1526 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1527 CmpInst *CI = cast<CmpInst>(NewCond); 1528 CI->setPredicate(CI->getInversePredicate()); 1529 } else { 1530 NewCond = BinaryOperator::CreateNot(NewCond, 1531 PBI->getCondition()->getName()+".not", PBI); 1532 } 1533 1534 PBI->setCondition(NewCond); 1535 BasicBlock *OldTrue = PBI->getSuccessor(0); 1536 BasicBlock *OldFalse = PBI->getSuccessor(1); 1537 PBI->setSuccessor(0, OldFalse); 1538 PBI->setSuccessor(1, OldTrue); 1539 } 1540 1541 // If we have a bonus inst, clone it into the predecessor block. 1542 Instruction *NewBonus = 0; 1543 if (BonusInst) { 1544 NewBonus = BonusInst->clone(); 1545 PredBlock->getInstList().insert(PBI, NewBonus); 1546 NewBonus->takeName(BonusInst); 1547 BonusInst->setName(BonusInst->getName()+".old"); 1548 } 1549 1550 // Clone Cond into the predecessor basic block, and or/and the 1551 // two conditions together. 1552 Instruction *New = Cond->clone(); 1553 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1554 PredBlock->getInstList().insert(PBI, New); 1555 New->takeName(Cond); 1556 Cond->setName(New->getName()+".old"); 1557 1558 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1559 New, "or.cond", PBI); 1560 PBI->setCondition(NewCond); 1561 if (PBI->getSuccessor(0) == BB) { 1562 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1563 PBI->setSuccessor(0, TrueDest); 1564 } 1565 if (PBI->getSuccessor(1) == BB) { 1566 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1567 PBI->setSuccessor(1, FalseDest); 1568 } 1569 return true; 1570 } 1571 return false; 1572 } 1573 1574 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1575 /// predecessor of another block, this function tries to simplify it. We know 1576 /// that PBI and BI are both conditional branches, and BI is in one of the 1577 /// successor blocks of PBI - PBI branches to BI. 1578 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1579 assert(PBI->isConditional() && BI->isConditional()); 1580 BasicBlock *BB = BI->getParent(); 1581 1582 // If this block ends with a branch instruction, and if there is a 1583 // predecessor that ends on a branch of the same condition, make 1584 // this conditional branch redundant. 1585 if (PBI->getCondition() == BI->getCondition() && 1586 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1587 // Okay, the outcome of this conditional branch is statically 1588 // knowable. If this block had a single pred, handle specially. 1589 if (BB->getSinglePredecessor()) { 1590 // Turn this into a branch on constant. 1591 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1592 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1593 CondIsTrue)); 1594 return true; // Nuke the branch on constant. 1595 } 1596 1597 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1598 // in the constant and simplify the block result. Subsequent passes of 1599 // simplifycfg will thread the block. 1600 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1601 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 1602 BI->getCondition()->getName() + ".pr", 1603 BB->begin()); 1604 // Okay, we're going to insert the PHI node. Since PBI is not the only 1605 // predecessor, compute the PHI'd conditional value for all of the preds. 1606 // Any predecessor where the condition is not computable we keep symbolic. 1607 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1608 BasicBlock *P = *PI; 1609 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 1610 PBI != BI && PBI->isConditional() && 1611 PBI->getCondition() == BI->getCondition() && 1612 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1613 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1614 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1615 CondIsTrue), P); 1616 } else { 1617 NewPN->addIncoming(BI->getCondition(), P); 1618 } 1619 } 1620 1621 BI->setCondition(NewPN); 1622 return true; 1623 } 1624 } 1625 1626 // If this is a conditional branch in an empty block, and if any 1627 // predecessors is a conditional branch to one of our destinations, 1628 // fold the conditions into logical ops and one cond br. 1629 BasicBlock::iterator BBI = BB->begin(); 1630 // Ignore dbg intrinsics. 1631 while (isa<DbgInfoIntrinsic>(BBI)) 1632 ++BBI; 1633 if (&*BBI != BI) 1634 return false; 1635 1636 1637 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1638 if (CE->canTrap()) 1639 return false; 1640 1641 int PBIOp, BIOp; 1642 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1643 PBIOp = BIOp = 0; 1644 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1645 PBIOp = 0, BIOp = 1; 1646 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1647 PBIOp = 1, BIOp = 0; 1648 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1649 PBIOp = BIOp = 1; 1650 else 1651 return false; 1652 1653 // Check to make sure that the other destination of this branch 1654 // isn't BB itself. If so, this is an infinite loop that will 1655 // keep getting unwound. 1656 if (PBI->getSuccessor(PBIOp) == BB) 1657 return false; 1658 1659 // Do not perform this transformation if it would require 1660 // insertion of a large number of select instructions. For targets 1661 // without predication/cmovs, this is a big pessimization. 1662 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1663 1664 unsigned NumPhis = 0; 1665 for (BasicBlock::iterator II = CommonDest->begin(); 1666 isa<PHINode>(II); ++II, ++NumPhis) 1667 if (NumPhis > 2) // Disable this xform. 1668 return false; 1669 1670 // Finally, if everything is ok, fold the branches to logical ops. 1671 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1672 1673 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 1674 << "AND: " << *BI->getParent()); 1675 1676 1677 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1678 // branch in it, where one edge (OtherDest) goes back to itself but the other 1679 // exits. We don't *know* that the program avoids the infinite loop 1680 // (even though that seems likely). If we do this xform naively, we'll end up 1681 // recursively unpeeling the loop. Since we know that (after the xform is 1682 // done) that the block *is* infinite if reached, we just make it an obviously 1683 // infinite loop with no cond branch. 1684 if (OtherDest == BB) { 1685 // Insert it at the end of the function, because it's either code, 1686 // or it won't matter if it's hot. :) 1687 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 1688 "infloop", BB->getParent()); 1689 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1690 OtherDest = InfLoopBlock; 1691 } 1692 1693 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1694 1695 // BI may have other predecessors. Because of this, we leave 1696 // it alone, but modify PBI. 1697 1698 // Make sure we get to CommonDest on True&True directions. 1699 Value *PBICond = PBI->getCondition(); 1700 if (PBIOp) 1701 PBICond = BinaryOperator::CreateNot(PBICond, 1702 PBICond->getName()+".not", 1703 PBI); 1704 Value *BICond = BI->getCondition(); 1705 if (BIOp) 1706 BICond = BinaryOperator::CreateNot(BICond, 1707 BICond->getName()+".not", 1708 PBI); 1709 // Merge the conditions. 1710 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1711 1712 // Modify PBI to branch on the new condition to the new dests. 1713 PBI->setCondition(Cond); 1714 PBI->setSuccessor(0, CommonDest); 1715 PBI->setSuccessor(1, OtherDest); 1716 1717 // OtherDest may have phi nodes. If so, add an entry from PBI's 1718 // block that are identical to the entries for BI's block. 1719 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 1720 1721 // We know that the CommonDest already had an edge from PBI to 1722 // it. If it has PHIs though, the PHIs may have different 1723 // entries for BB and PBI's BB. If so, insert a select to make 1724 // them agree. 1725 PHINode *PN; 1726 for (BasicBlock::iterator II = CommonDest->begin(); 1727 (PN = dyn_cast<PHINode>(II)); ++II) { 1728 Value *BIV = PN->getIncomingValueForBlock(BB); 1729 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1730 Value *PBIV = PN->getIncomingValue(PBBIdx); 1731 if (BIV != PBIV) { 1732 // Insert a select in PBI to pick the right value. 1733 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1734 PBIV->getName()+".mux", PBI); 1735 PN->setIncomingValue(PBBIdx, NV); 1736 } 1737 } 1738 1739 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 1740 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1741 1742 // This basic block is probably dead. We know it has at least 1743 // one fewer predecessor. 1744 return true; 1745 } 1746 1747 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 1748 // branch to TrueBB if Cond is true or to FalseBB if Cond is false. 1749 // Takes care of updating the successors and removing the old terminator. 1750 // Also makes sure not to introduce new successors by assuming that edges to 1751 // non-successor TrueBBs and FalseBBs aren't reachable. 1752 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 1753 BasicBlock *TrueBB, BasicBlock *FalseBB){ 1754 // Remove any superfluous successor edges from the CFG. 1755 // First, figure out which successors to preserve. 1756 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 1757 // successor. 1758 BasicBlock *KeepEdge1 = TrueBB; 1759 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 1760 1761 // Then remove the rest. 1762 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 1763 BasicBlock *Succ = OldTerm->getSuccessor(I); 1764 // Make sure only to keep exactly one copy of each edge. 1765 if (Succ == KeepEdge1) 1766 KeepEdge1 = 0; 1767 else if (Succ == KeepEdge2) 1768 KeepEdge2 = 0; 1769 else 1770 Succ->removePredecessor(OldTerm->getParent()); 1771 } 1772 1773 // Insert an appropriate new terminator. 1774 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 1775 if (TrueBB == FalseBB) 1776 // We were only looking for one successor, and it was present. 1777 // Create an unconditional branch to it. 1778 BranchInst::Create(TrueBB, OldTerm); 1779 else 1780 // We found both of the successors we were looking for. 1781 // Create a conditional branch sharing the condition of the select. 1782 BranchInst::Create(TrueBB, FalseBB, Cond, OldTerm); 1783 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 1784 // Neither of the selected blocks were successors, so this 1785 // terminator must be unreachable. 1786 new UnreachableInst(OldTerm->getContext(), OldTerm); 1787 } else { 1788 // One of the selected values was a successor, but the other wasn't. 1789 // Insert an unconditional branch to the one that was found; 1790 // the edge to the one that wasn't must be unreachable. 1791 if (KeepEdge1 == 0) 1792 // Only TrueBB was found. 1793 BranchInst::Create(TrueBB, OldTerm); 1794 else 1795 // Only FalseBB was found. 1796 BranchInst::Create(FalseBB, OldTerm); 1797 } 1798 1799 EraseTerminatorInstAndDCECond(OldTerm); 1800 return true; 1801 } 1802 1803 // SimplifyIndirectBrOnSelect - Replaces 1804 // (indirectbr (select cond, blockaddress(@fn, BlockA), 1805 // blockaddress(@fn, BlockB))) 1806 // with 1807 // (br cond, BlockA, BlockB). 1808 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 1809 // Check that both operands of the select are block addresses. 1810 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 1811 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 1812 if (!TBA || !FBA) 1813 return false; 1814 1815 // Extract the actual blocks. 1816 BasicBlock *TrueBB = TBA->getBasicBlock(); 1817 BasicBlock *FalseBB = FBA->getBasicBlock(); 1818 1819 // Perform the actual simplification. 1820 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); 1821 } 1822 1823 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 1824 /// instruction (a seteq/setne with a constant) as the only instruction in a 1825 /// block that ends with an uncond branch. We are looking for a very specific 1826 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 1827 /// this case, we merge the first two "or's of icmp" into a switch, but then the 1828 /// default value goes to an uncond block with a seteq in it, we get something 1829 /// like: 1830 /// 1831 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 1832 /// DEFAULT: 1833 /// %tmp = icmp eq i8 %A, 92 1834 /// br label %end 1835 /// end: 1836 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 1837 /// 1838 /// We prefer to split the edge to 'end' so that there is a true/false entry to 1839 /// the PHI, merging the third icmp into the switch. 1840 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 1841 const TargetData *TD) { 1842 BasicBlock *BB = ICI->getParent(); 1843 // If the block has any PHIs in it or the icmp has multiple uses, it is too 1844 // complex. 1845 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 1846 1847 Value *V = ICI->getOperand(0); 1848 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 1849 1850 // The pattern we're looking for is where our only predecessor is a switch on 1851 // 'V' and this block is the default case for the switch. In this case we can 1852 // fold the compared value into the switch to simplify things. 1853 BasicBlock *Pred = BB->getSinglePredecessor(); 1854 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 1855 1856 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 1857 if (SI->getCondition() != V) 1858 return false; 1859 1860 // If BB is reachable on a non-default case, then we simply know the value of 1861 // V in this block. Substitute it and constant fold the icmp instruction 1862 // away. 1863 if (SI->getDefaultDest() != BB) { 1864 ConstantInt *VVal = SI->findCaseDest(BB); 1865 assert(VVal && "Should have a unique destination value"); 1866 ICI->setOperand(0, VVal); 1867 1868 if (Value *V = SimplifyInstruction(ICI, TD)) { 1869 ICI->replaceAllUsesWith(V); 1870 ICI->eraseFromParent(); 1871 } 1872 // BB is now empty, so it is likely to simplify away. 1873 return SimplifyCFG(BB) | true; 1874 } 1875 1876 // Ok, the block is reachable from the default dest. If the constant we're 1877 // comparing exists in one of the other edges, then we can constant fold ICI 1878 // and zap it. 1879 if (SI->findCaseValue(Cst) != 0) { 1880 Value *V; 1881 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1882 V = ConstantInt::getFalse(BB->getContext()); 1883 else 1884 V = ConstantInt::getTrue(BB->getContext()); 1885 1886 ICI->replaceAllUsesWith(V); 1887 ICI->eraseFromParent(); 1888 // BB is now empty, so it is likely to simplify away. 1889 return SimplifyCFG(BB) | true; 1890 } 1891 1892 // The use of the icmp has to be in the 'end' block, by the only PHI node in 1893 // the block. 1894 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 1895 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 1896 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 1897 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 1898 return false; 1899 1900 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 1901 // true in the PHI. 1902 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 1903 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 1904 1905 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1906 std::swap(DefaultCst, NewCst); 1907 1908 // Replace ICI (which is used by the PHI for the default value) with true or 1909 // false depending on if it is EQ or NE. 1910 ICI->replaceAllUsesWith(DefaultCst); 1911 ICI->eraseFromParent(); 1912 1913 // Okay, the switch goes to this block on a default value. Add an edge from 1914 // the switch to the merge point on the compared value. 1915 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 1916 BB->getParent(), BB); 1917 SI->addCase(Cst, NewBB); 1918 1919 // NewBB branches to the phi block, add the uncond branch and the phi entry. 1920 BranchInst::Create(SuccBlock, NewBB); 1921 PHIUse->addIncoming(NewCst, NewBB); 1922 return true; 1923 } 1924 1925 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 1926 /// Check to see if it is branching on an or/and chain of icmp instructions, and 1927 /// fold it into a switch instruction if so. 1928 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD) { 1929 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1930 if (Cond == 0) return false; 1931 1932 1933 // Change br (X == 0 | X == 1), T, F into a switch instruction. 1934 // If this is a bunch of seteq's or'd together, or if it's a bunch of 1935 // 'setne's and'ed together, collect them. 1936 Value *CompVal = 0; 1937 std::vector<ConstantInt*> Values; 1938 bool TrueWhenEqual = true; 1939 Value *ExtraCase = 0; 1940 unsigned UsedICmps = 0; 1941 1942 if (Cond->getOpcode() == Instruction::Or) { 1943 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 1944 UsedICmps); 1945 } else if (Cond->getOpcode() == Instruction::And) { 1946 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 1947 UsedICmps); 1948 TrueWhenEqual = false; 1949 } 1950 1951 // If we didn't have a multiply compared value, fail. 1952 if (CompVal == 0) return false; 1953 1954 // Avoid turning single icmps into a switch. 1955 if (UsedICmps <= 1) 1956 return false; 1957 1958 // There might be duplicate constants in the list, which the switch 1959 // instruction can't handle, remove them now. 1960 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 1961 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 1962 1963 // If Extra was used, we require at least two switch values to do the 1964 // transformation. A switch with one value is just an cond branch. 1965 if (ExtraCase && Values.size() < 2) return false; 1966 1967 // Figure out which block is which destination. 1968 BasicBlock *DefaultBB = BI->getSuccessor(1); 1969 BasicBlock *EdgeBB = BI->getSuccessor(0); 1970 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 1971 1972 BasicBlock *BB = BI->getParent(); 1973 1974 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 1975 << " cases into SWITCH. BB is:\n" << *BB); 1976 1977 // If there are any extra values that couldn't be folded into the switch 1978 // then we evaluate them with an explicit branch first. Split the block 1979 // right before the condbr to handle it. 1980 if (ExtraCase) { 1981 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 1982 // Remove the uncond branch added to the old block. 1983 TerminatorInst *OldTI = BB->getTerminator(); 1984 1985 if (TrueWhenEqual) 1986 BranchInst::Create(EdgeBB, NewBB, ExtraCase, OldTI); 1987 else 1988 BranchInst::Create(NewBB, EdgeBB, ExtraCase, OldTI); 1989 1990 OldTI->eraseFromParent(); 1991 1992 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 1993 // for the edge we just added. 1994 AddPredecessorToBlock(EdgeBB, BB, NewBB); 1995 1996 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 1997 << "\nEXTRABB = " << *BB); 1998 BB = NewBB; 1999 } 2000 2001 // Convert pointer to int before we switch. 2002 if (CompVal->getType()->isPointerTy()) { 2003 assert(TD && "Cannot switch on pointer without TargetData"); 2004 CompVal = new PtrToIntInst(CompVal, 2005 TD->getIntPtrType(CompVal->getContext()), 2006 "magicptr", BI); 2007 } 2008 2009 // Create the new switch instruction now. 2010 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, Values.size(), BI); 2011 2012 // Add all of the 'cases' to the switch instruction. 2013 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2014 New->addCase(Values[i], EdgeBB); 2015 2016 // We added edges from PI to the EdgeBB. As such, if there were any 2017 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2018 // the number of edges added. 2019 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2020 isa<PHINode>(BBI); ++BBI) { 2021 PHINode *PN = cast<PHINode>(BBI); 2022 Value *InVal = PN->getIncomingValueForBlock(BB); 2023 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2024 PN->addIncoming(InVal, BB); 2025 } 2026 2027 // Erase the old branch instruction. 2028 EraseTerminatorInstAndDCECond(BI); 2029 2030 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2031 return true; 2032 } 2033 2034 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI) { 2035 BasicBlock *BB = RI->getParent(); 2036 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2037 2038 // Find predecessors that end with branches. 2039 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2040 SmallVector<BranchInst*, 8> CondBranchPreds; 2041 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2042 BasicBlock *P = *PI; 2043 TerminatorInst *PTI = P->getTerminator(); 2044 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2045 if (BI->isUnconditional()) 2046 UncondBranchPreds.push_back(P); 2047 else 2048 CondBranchPreds.push_back(BI); 2049 } 2050 } 2051 2052 // If we found some, do the transformation! 2053 if (!UncondBranchPreds.empty() && DupRet) { 2054 while (!UncondBranchPreds.empty()) { 2055 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2056 DEBUG(dbgs() << "FOLDING: " << *BB 2057 << "INTO UNCOND BRANCH PRED: " << *Pred); 2058 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2059 } 2060 2061 // If we eliminated all predecessors of the block, delete the block now. 2062 if (pred_begin(BB) == pred_end(BB)) 2063 // We know there are no successors, so just nuke the block. 2064 BB->eraseFromParent(); 2065 2066 return true; 2067 } 2068 2069 // Check out all of the conditional branches going to this return 2070 // instruction. If any of them just select between returns, change the 2071 // branch itself into a select/return pair. 2072 while (!CondBranchPreds.empty()) { 2073 BranchInst *BI = CondBranchPreds.pop_back_val(); 2074 2075 // Check to see if the non-BB successor is also a return block. 2076 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2077 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2078 SimplifyCondBranchToTwoReturns(BI)) 2079 return true; 2080 } 2081 return false; 2082 } 2083 2084 bool SimplifyCFGOpt::SimplifyUnwind(UnwindInst *UI) { 2085 // Check to see if the first instruction in this block is just an unwind. 2086 // If so, replace any invoke instructions which use this as an exception 2087 // destination with call instructions. 2088 BasicBlock *BB = UI->getParent(); 2089 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2090 2091 bool Changed = false; 2092 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2093 while (!Preds.empty()) { 2094 BasicBlock *Pred = Preds.back(); 2095 InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator()); 2096 if (II && II->getUnwindDest() == BB) { 2097 // Insert a new branch instruction before the invoke, because this 2098 // is now a fall through. 2099 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2100 Pred->getInstList().remove(II); // Take out of symbol table 2101 2102 // Insert the call now. 2103 SmallVector<Value*,8> Args(II->op_begin(), II->op_end()-3); 2104 CallInst *CI = CallInst::Create(II->getCalledValue(), 2105 Args.begin(), Args.end(), 2106 II->getName(), BI); 2107 CI->setCallingConv(II->getCallingConv()); 2108 CI->setAttributes(II->getAttributes()); 2109 // If the invoke produced a value, the Call now does instead. 2110 II->replaceAllUsesWith(CI); 2111 delete II; 2112 Changed = true; 2113 } 2114 2115 Preds.pop_back(); 2116 } 2117 2118 // If this block is now dead (and isn't the entry block), remove it. 2119 if (pred_begin(BB) == pred_end(BB) && 2120 BB != &BB->getParent()->getEntryBlock()) { 2121 // We know there are no successors, so just nuke the block. 2122 BB->eraseFromParent(); 2123 return true; 2124 } 2125 2126 return Changed; 2127 } 2128 2129 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2130 BasicBlock *BB = UI->getParent(); 2131 2132 bool Changed = false; 2133 2134 // If there are any instructions immediately before the unreachable that can 2135 // be removed, do so. 2136 while (UI != BB->begin()) { 2137 BasicBlock::iterator BBI = UI; 2138 --BBI; 2139 // Do not delete instructions that can have side effects, like calls 2140 // (which may never return) and volatile loads and stores. 2141 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2142 2143 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) 2144 if (SI->isVolatile()) 2145 break; 2146 2147 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) 2148 if (LI->isVolatile()) 2149 break; 2150 2151 // Delete this instruction 2152 BBI->eraseFromParent(); 2153 Changed = true; 2154 } 2155 2156 // If the unreachable instruction is the first in the block, take a gander 2157 // at all of the predecessors of this instruction, and simplify them. 2158 if (&BB->front() != UI) return Changed; 2159 2160 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2161 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2162 TerminatorInst *TI = Preds[i]->getTerminator(); 2163 2164 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2165 if (BI->isUnconditional()) { 2166 if (BI->getSuccessor(0) == BB) { 2167 new UnreachableInst(TI->getContext(), TI); 2168 TI->eraseFromParent(); 2169 Changed = true; 2170 } 2171 } else { 2172 if (BI->getSuccessor(0) == BB) { 2173 BranchInst::Create(BI->getSuccessor(1), BI); 2174 EraseTerminatorInstAndDCECond(BI); 2175 } else if (BI->getSuccessor(1) == BB) { 2176 BranchInst::Create(BI->getSuccessor(0), BI); 2177 EraseTerminatorInstAndDCECond(BI); 2178 Changed = true; 2179 } 2180 } 2181 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2182 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2183 if (SI->getSuccessor(i) == BB) { 2184 BB->removePredecessor(SI->getParent()); 2185 SI->removeCase(i); 2186 --i; --e; 2187 Changed = true; 2188 } 2189 // If the default value is unreachable, figure out the most popular 2190 // destination and make it the default. 2191 if (SI->getSuccessor(0) == BB) { 2192 std::map<BasicBlock*, unsigned> Popularity; 2193 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2194 Popularity[SI->getSuccessor(i)]++; 2195 2196 // Find the most popular block. 2197 unsigned MaxPop = 0; 2198 BasicBlock *MaxBlock = 0; 2199 for (std::map<BasicBlock*, unsigned>::iterator 2200 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2201 if (I->second > MaxPop) { 2202 MaxPop = I->second; 2203 MaxBlock = I->first; 2204 } 2205 } 2206 if (MaxBlock) { 2207 // Make this the new default, allowing us to delete any explicit 2208 // edges to it. 2209 SI->setSuccessor(0, MaxBlock); 2210 Changed = true; 2211 2212 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2213 // it. 2214 if (isa<PHINode>(MaxBlock->begin())) 2215 for (unsigned i = 0; i != MaxPop-1; ++i) 2216 MaxBlock->removePredecessor(SI->getParent()); 2217 2218 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2219 if (SI->getSuccessor(i) == MaxBlock) { 2220 SI->removeCase(i); 2221 --i; --e; 2222 } 2223 } 2224 } 2225 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2226 if (II->getUnwindDest() == BB) { 2227 // Convert the invoke to a call instruction. This would be a good 2228 // place to note that the call does not throw though. 2229 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2230 II->removeFromParent(); // Take out of symbol table 2231 2232 // Insert the call now... 2233 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2234 CallInst *CI = CallInst::Create(II->getCalledValue(), 2235 Args.begin(), Args.end(), 2236 II->getName(), BI); 2237 CI->setCallingConv(II->getCallingConv()); 2238 CI->setAttributes(II->getAttributes()); 2239 // If the invoke produced a value, the call does now instead. 2240 II->replaceAllUsesWith(CI); 2241 delete II; 2242 Changed = true; 2243 } 2244 } 2245 } 2246 2247 // If this block is now dead, remove it. 2248 if (pred_begin(BB) == pred_end(BB) && 2249 BB != &BB->getParent()->getEntryBlock()) { 2250 // We know there are no successors, so just nuke the block. 2251 BB->eraseFromParent(); 2252 return true; 2253 } 2254 2255 return Changed; 2256 } 2257 2258 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 2259 /// integer range comparison into a sub, an icmp and a branch. 2260 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI) { 2261 assert(SI->getNumCases() > 2 && "Degenerate switch?"); 2262 2263 // Make sure all cases point to the same destination and gather the values. 2264 SmallVector<ConstantInt *, 16> Cases; 2265 Cases.push_back(SI->getCaseValue(1)); 2266 for (unsigned I = 2, E = SI->getNumCases(); I != E; ++I) { 2267 if (SI->getSuccessor(I-1) != SI->getSuccessor(I)) 2268 return false; 2269 Cases.push_back(SI->getCaseValue(I)); 2270 } 2271 assert(Cases.size() == SI->getNumCases()-1 && "Not all cases gathered"); 2272 2273 // Sort the case values, then check if they form a range we can transform. 2274 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 2275 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 2276 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 2277 return false; 2278 } 2279 2280 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 2281 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()-1); 2282 2283 Value *Sub = SI->getCondition(); 2284 if (!Offset->isNullValue()) 2285 Sub = BinaryOperator::CreateAdd(Sub, Offset, Sub->getName()+".off", SI); 2286 Value *Cmp = new ICmpInst(SI, ICmpInst::ICMP_ULT, Sub, NumCases, "switch"); 2287 BranchInst::Create(SI->getSuccessor(1), SI->getDefaultDest(), Cmp, SI); 2288 2289 // Prune obsolete incoming values off the successor's PHI nodes. 2290 for (BasicBlock::iterator BBI = SI->getSuccessor(1)->begin(); 2291 isa<PHINode>(BBI); ++BBI) { 2292 for (unsigned I = 0, E = SI->getNumCases()-2; I != E; ++I) 2293 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 2294 } 2295 SI->eraseFromParent(); 2296 2297 return true; 2298 } 2299 2300 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI) { 2301 // If this switch is too complex to want to look at, ignore it. 2302 if (!isValueEqualityComparison(SI)) 2303 return false; 2304 2305 BasicBlock *BB = SI->getParent(); 2306 2307 // If we only have one predecessor, and if it is a branch on this value, 2308 // see if that predecessor totally determines the outcome of this switch. 2309 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2310 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 2311 return SimplifyCFG(BB) | true; 2312 2313 // If the block only contains the switch, see if we can fold the block 2314 // away into any preds. 2315 BasicBlock::iterator BBI = BB->begin(); 2316 // Ignore dbg intrinsics. 2317 while (isa<DbgInfoIntrinsic>(BBI)) 2318 ++BBI; 2319 if (SI == &*BBI) 2320 if (FoldValueComparisonIntoPredecessors(SI)) 2321 return SimplifyCFG(BB) | true; 2322 2323 // Try to transform the switch into an icmp and a branch. 2324 if (TurnSwitchRangeIntoICmp(SI)) 2325 return SimplifyCFG(BB) | true; 2326 2327 return false; 2328 } 2329 2330 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 2331 BasicBlock *BB = IBI->getParent(); 2332 bool Changed = false; 2333 2334 // Eliminate redundant destinations. 2335 SmallPtrSet<Value *, 8> Succs; 2336 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 2337 BasicBlock *Dest = IBI->getDestination(i); 2338 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 2339 Dest->removePredecessor(BB); 2340 IBI->removeDestination(i); 2341 --i; --e; 2342 Changed = true; 2343 } 2344 } 2345 2346 if (IBI->getNumDestinations() == 0) { 2347 // If the indirectbr has no successors, change it to unreachable. 2348 new UnreachableInst(IBI->getContext(), IBI); 2349 EraseTerminatorInstAndDCECond(IBI); 2350 return true; 2351 } 2352 2353 if (IBI->getNumDestinations() == 1) { 2354 // If the indirectbr has one successor, change it to a direct branch. 2355 BranchInst::Create(IBI->getDestination(0), IBI); 2356 EraseTerminatorInstAndDCECond(IBI); 2357 return true; 2358 } 2359 2360 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 2361 if (SimplifyIndirectBrOnSelect(IBI, SI)) 2362 return SimplifyCFG(BB) | true; 2363 } 2364 return Changed; 2365 } 2366 2367 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI) { 2368 BasicBlock *BB = BI->getParent(); 2369 2370 // If the Terminator is the only non-phi instruction, simplify the block. 2371 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg(); 2372 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 2373 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 2374 return true; 2375 2376 // If the only instruction in the block is a seteq/setne comparison 2377 // against a constant, try to simplify the block. 2378 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 2379 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 2380 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 2381 ; 2382 if (I->isTerminator() && TryToSimplifyUncondBranchWithICmpInIt(ICI, TD)) 2383 return true; 2384 } 2385 2386 return false; 2387 } 2388 2389 2390 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI) { 2391 BasicBlock *BB = BI->getParent(); 2392 2393 // Conditional branch 2394 if (isValueEqualityComparison(BI)) { 2395 // If we only have one predecessor, and if it is a branch on this value, 2396 // see if that predecessor totally determines the outcome of this 2397 // switch. 2398 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2399 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 2400 return SimplifyCFG(BB) | true; 2401 2402 // This block must be empty, except for the setcond inst, if it exists. 2403 // Ignore dbg intrinsics. 2404 BasicBlock::iterator I = BB->begin(); 2405 // Ignore dbg intrinsics. 2406 while (isa<DbgInfoIntrinsic>(I)) 2407 ++I; 2408 if (&*I == BI) { 2409 if (FoldValueComparisonIntoPredecessors(BI)) 2410 return SimplifyCFG(BB) | true; 2411 } else if (&*I == cast<Instruction>(BI->getCondition())){ 2412 ++I; 2413 // Ignore dbg intrinsics. 2414 while (isa<DbgInfoIntrinsic>(I)) 2415 ++I; 2416 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI)) 2417 return SimplifyCFG(BB) | true; 2418 } 2419 } 2420 2421 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 2422 if (SimplifyBranchOnICmpChain(BI, TD)) 2423 return true; 2424 2425 // We have a conditional branch to two blocks that are only reachable 2426 // from BI. We know that the condbr dominates the two blocks, so see if 2427 // there is any identical code in the "then" and "else" blocks. If so, we 2428 // can hoist it up to the branching block. 2429 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 2430 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2431 if (HoistThenElseCodeToIf(BI)) 2432 return SimplifyCFG(BB) | true; 2433 } else { 2434 // If Successor #1 has multiple preds, we may be able to conditionally 2435 // execute Successor #0 if it branches to successor #1. 2436 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 2437 if (Succ0TI->getNumSuccessors() == 1 && 2438 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 2439 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 2440 return SimplifyCFG(BB) | true; 2441 } 2442 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2443 // If Successor #0 has multiple preds, we may be able to conditionally 2444 // execute Successor #1 if it branches to successor #0. 2445 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 2446 if (Succ1TI->getNumSuccessors() == 1 && 2447 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 2448 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 2449 return SimplifyCFG(BB) | true; 2450 } 2451 2452 // If this is a branch on a phi node in the current block, thread control 2453 // through this block if any PHI node entries are constants. 2454 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 2455 if (PN->getParent() == BI->getParent()) 2456 if (FoldCondBranchOnPHI(BI, TD)) 2457 return SimplifyCFG(BB) | true; 2458 2459 // If this basic block is ONLY a setcc and a branch, and if a predecessor 2460 // branches to us and one of our successors, fold the setcc into the 2461 // predecessor and use logical operations to pick the right destination. 2462 if (FoldBranchToCommonDest(BI)) 2463 return SimplifyCFG(BB) | true; 2464 2465 // Scan predecessor blocks for conditional branches. 2466 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2467 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2468 if (PBI != BI && PBI->isConditional()) 2469 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2470 return SimplifyCFG(BB) | true; 2471 2472 return false; 2473 } 2474 2475 bool SimplifyCFGOpt::run(BasicBlock *BB) { 2476 bool Changed = false; 2477 2478 assert(BB && BB->getParent() && "Block not embedded in function!"); 2479 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 2480 2481 // Remove basic blocks that have no predecessors (except the entry block)... 2482 // or that just have themself as a predecessor. These are unreachable. 2483 if ((pred_begin(BB) == pred_end(BB) && 2484 BB != &BB->getParent()->getEntryBlock()) || 2485 BB->getSinglePredecessor() == BB) { 2486 DEBUG(dbgs() << "Removing BB: \n" << *BB); 2487 DeleteDeadBlock(BB); 2488 return true; 2489 } 2490 2491 // Check to see if we can constant propagate this terminator instruction 2492 // away... 2493 Changed |= ConstantFoldTerminator(BB); 2494 2495 // Check for and eliminate duplicate PHI nodes in this block. 2496 Changed |= EliminateDuplicatePHINodes(BB); 2497 2498 // Merge basic blocks into their predecessor if there is only one distinct 2499 // pred, and if there is only one distinct successor of the predecessor, and 2500 // if there are no PHI nodes. 2501 // 2502 if (MergeBlockIntoPredecessor(BB)) 2503 return true; 2504 2505 // If there is a trivial two-entry PHI node in this basic block, and we can 2506 // eliminate it, do so now. 2507 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 2508 if (PN->getNumIncomingValues() == 2) 2509 Changed |= FoldTwoEntryPHINode(PN, TD); 2510 2511 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 2512 if (BI->isUnconditional()) { 2513 if (SimplifyUncondBranch(BI)) return true; 2514 } else { 2515 if (SimplifyCondBranch(BI)) return true; 2516 } 2517 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 2518 if (SimplifyReturn(RI)) return true; 2519 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2520 if (SimplifySwitch(SI)) return true; 2521 } else if (UnreachableInst *UI = 2522 dyn_cast<UnreachableInst>(BB->getTerminator())) { 2523 if (SimplifyUnreachable(UI)) return true; 2524 } else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 2525 if (SimplifyUnwind(UI)) return true; 2526 } else if (IndirectBrInst *IBI = 2527 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 2528 if (SimplifyIndirectBr(IBI)) return true; 2529 } 2530 2531 return Changed; 2532 } 2533 2534 /// SimplifyCFG - This function is used to do simplification of a CFG. For 2535 /// example, it adjusts branches to branches to eliminate the extra hop, it 2536 /// eliminates unreachable basic blocks, and does other "peephole" optimization 2537 /// of the CFG. It returns true if a modification was made. 2538 /// 2539 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 2540 return SimplifyCFGOpt(TD).run(BB); 2541 } 2542