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 // Skip debug info if it is not identical. 811 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 812 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 813 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 814 while (isa<DbgInfoIntrinsic>(I1)) 815 I1 = BB1_Itr++; 816 while (isa<DbgInfoIntrinsic>(I2)) 817 I2 = BB2_Itr++; 818 } 819 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 820 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 821 return false; 822 823 // If we get here, we can hoist at least one instruction. 824 BasicBlock *BIParent = BI->getParent(); 825 826 do { 827 // If we are hoisting the terminator instruction, don't move one (making a 828 // broken BB), instead clone it, and remove BI. 829 if (isa<TerminatorInst>(I1)) 830 goto HoistTerminator; 831 832 // For a normal instruction, we just move one to right before the branch, 833 // then replace all uses of the other with the first. Finally, we remove 834 // the now redundant second instruction. 835 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 836 if (!I2->use_empty()) 837 I2->replaceAllUsesWith(I1); 838 I1->intersectOptionalDataWith(I2); 839 I2->eraseFromParent(); 840 841 I1 = BB1_Itr++; 842 I2 = BB2_Itr++; 843 // Skip debug info if it is not identical. 844 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 845 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 846 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 847 while (isa<DbgInfoIntrinsic>(I1)) 848 I1 = BB1_Itr++; 849 while (isa<DbgInfoIntrinsic>(I2)) 850 I2 = BB2_Itr++; 851 } 852 } while (I1->isIdenticalToWhenDefined(I2)); 853 854 return true; 855 856 HoistTerminator: 857 // It may not be possible to hoist an invoke. 858 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 859 return true; 860 861 // Okay, it is safe to hoist the terminator. 862 Instruction *NT = I1->clone(); 863 BIParent->getInstList().insert(BI, NT); 864 if (!NT->getType()->isVoidTy()) { 865 I1->replaceAllUsesWith(NT); 866 I2->replaceAllUsesWith(NT); 867 NT->takeName(I1); 868 } 869 870 // Hoisting one of the terminators from our successor is a great thing. 871 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 872 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 873 // nodes, so we insert select instruction to compute the final result. 874 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 875 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 876 PHINode *PN; 877 for (BasicBlock::iterator BBI = SI->begin(); 878 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 879 Value *BB1V = PN->getIncomingValueForBlock(BB1); 880 Value *BB2V = PN->getIncomingValueForBlock(BB2); 881 if (BB1V == BB2V) continue; 882 883 // These values do not agree. Insert a select instruction before NT 884 // that determines the right value. 885 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 886 if (SI == 0) 887 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 888 BB1V->getName()+"."+BB2V->getName(), NT); 889 // Make the PHI node use the select for all incoming values for BB1/BB2 890 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 891 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 892 PN->setIncomingValue(i, SI); 893 } 894 } 895 896 // Update any PHI nodes in our new successors. 897 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 898 AddPredecessorToBlock(*SI, BIParent, BB1); 899 900 EraseTerminatorInstAndDCECond(BI); 901 return true; 902 } 903 904 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 905 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code 906 /// (for now, restricted to a single instruction that's side effect free) from 907 /// the BB1 into the branch block to speculatively execute it. 908 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 909 // Only speculatively execution a single instruction (not counting the 910 // terminator) for now. 911 Instruction *HInst = NULL; 912 Instruction *Term = BB1->getTerminator(); 913 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 914 BBI != BBE; ++BBI) { 915 Instruction *I = BBI; 916 // Skip debug info. 917 if (isa<DbgInfoIntrinsic>(I)) continue; 918 if (I == Term) break; 919 920 if (HInst) 921 return false; 922 HInst = I; 923 } 924 if (!HInst) 925 return false; 926 927 // Be conservative for now. FP select instruction can often be expensive. 928 Value *BrCond = BI->getCondition(); 929 if (isa<FCmpInst>(BrCond)) 930 return false; 931 932 // If BB1 is actually on the false edge of the conditional branch, remember 933 // to swap the select operands later. 934 bool Invert = false; 935 if (BB1 != BI->getSuccessor(0)) { 936 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 937 Invert = true; 938 } 939 940 // Turn 941 // BB: 942 // %t1 = icmp 943 // br i1 %t1, label %BB1, label %BB2 944 // BB1: 945 // %t3 = add %t2, c 946 // br label BB2 947 // BB2: 948 // => 949 // BB: 950 // %t1 = icmp 951 // %t4 = add %t2, c 952 // %t3 = select i1 %t1, %t2, %t3 953 switch (HInst->getOpcode()) { 954 default: return false; // Not safe / profitable to hoist. 955 case Instruction::Add: 956 case Instruction::Sub: 957 // Not worth doing for vector ops. 958 if (HInst->getType()->isVectorTy()) 959 return false; 960 break; 961 case Instruction::And: 962 case Instruction::Or: 963 case Instruction::Xor: 964 case Instruction::Shl: 965 case Instruction::LShr: 966 case Instruction::AShr: 967 // Don't mess with vector operations. 968 if (HInst->getType()->isVectorTy()) 969 return false; 970 break; // These are all cheap and non-trapping instructions. 971 } 972 973 // If the instruction is obviously dead, don't try to predicate it. 974 if (HInst->use_empty()) { 975 HInst->eraseFromParent(); 976 return true; 977 } 978 979 // Can we speculatively execute the instruction? And what is the value 980 // if the condition is false? Consider the phi uses, if the incoming value 981 // from the "if" block are all the same V, then V is the value of the 982 // select if the condition is false. 983 BasicBlock *BIParent = BI->getParent(); 984 SmallVector<PHINode*, 4> PHIUses; 985 Value *FalseV = NULL; 986 987 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 988 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end(); 989 UI != E; ++UI) { 990 // Ignore any user that is not a PHI node in BB2. These can only occur in 991 // unreachable blocks, because they would not be dominated by the instr. 992 PHINode *PN = dyn_cast<PHINode>(*UI); 993 if (!PN || PN->getParent() != BB2) 994 return false; 995 PHIUses.push_back(PN); 996 997 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 998 if (!FalseV) 999 FalseV = PHIV; 1000 else if (FalseV != PHIV) 1001 return false; // Inconsistent value when condition is false. 1002 } 1003 1004 assert(FalseV && "Must have at least one user, and it must be a PHI"); 1005 1006 // Do not hoist the instruction if any of its operands are defined but not 1007 // used in this BB. The transformation will prevent the operand from 1008 // being sunk into the use block. 1009 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1010 i != e; ++i) { 1011 Instruction *OpI = dyn_cast<Instruction>(*i); 1012 if (OpI && OpI->getParent() == BIParent && 1013 !OpI->isUsedInBasicBlock(BIParent)) 1014 return false; 1015 } 1016 1017 // If we get here, we can hoist the instruction. Try to place it 1018 // before the icmp instruction preceding the conditional branch. 1019 BasicBlock::iterator InsertPos = BI; 1020 if (InsertPos != BIParent->begin()) 1021 --InsertPos; 1022 // Skip debug info between condition and branch. 1023 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos)) 1024 --InsertPos; 1025 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) { 1026 SmallPtrSet<Instruction *, 4> BB1Insns; 1027 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 1028 BB1I != BB1E; ++BB1I) 1029 BB1Insns.insert(BB1I); 1030 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 1031 UI != UE; ++UI) { 1032 Instruction *Use = cast<Instruction>(*UI); 1033 if (!BB1Insns.count(Use)) continue; 1034 1035 // If BrCond uses the instruction that place it just before 1036 // branch instruction. 1037 InsertPos = BI; 1038 break; 1039 } 1040 } else 1041 InsertPos = BI; 1042 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst); 1043 1044 // Create a select whose true value is the speculatively executed value and 1045 // false value is the previously determined FalseV. 1046 SelectInst *SI; 1047 if (Invert) 1048 SI = SelectInst::Create(BrCond, FalseV, HInst, 1049 FalseV->getName() + "." + HInst->getName(), BI); 1050 else 1051 SI = SelectInst::Create(BrCond, HInst, FalseV, 1052 HInst->getName() + "." + FalseV->getName(), BI); 1053 1054 // Make the PHI node use the select for all incoming values for "then" and 1055 // "if" blocks. 1056 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1057 PHINode *PN = PHIUses[i]; 1058 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1059 if (PN->getIncomingBlock(j) == BB1 || PN->getIncomingBlock(j) == BIParent) 1060 PN->setIncomingValue(j, SI); 1061 } 1062 1063 ++NumSpeculations; 1064 return true; 1065 } 1066 1067 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1068 /// across this block. 1069 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1070 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1071 unsigned Size = 0; 1072 1073 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1074 if (isa<DbgInfoIntrinsic>(BBI)) 1075 continue; 1076 if (Size > 10) return false; // Don't clone large BB's. 1077 ++Size; 1078 1079 // We can only support instructions that do not define values that are 1080 // live outside of the current basic block. 1081 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1082 UI != E; ++UI) { 1083 Instruction *U = cast<Instruction>(*UI); 1084 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1085 } 1086 1087 // Looks ok, continue checking. 1088 } 1089 1090 return true; 1091 } 1092 1093 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1094 /// that is defined in the same block as the branch and if any PHI entries are 1095 /// constants, thread edges corresponding to that entry to be branches to their 1096 /// ultimate destination. 1097 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { 1098 BasicBlock *BB = BI->getParent(); 1099 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1100 // NOTE: we currently cannot transform this case if the PHI node is used 1101 // outside of the block. 1102 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1103 return false; 1104 1105 // Degenerate case of a single entry PHI. 1106 if (PN->getNumIncomingValues() == 1) { 1107 FoldSingleEntryPHINodes(PN->getParent()); 1108 return true; 1109 } 1110 1111 // Now we know that this block has multiple preds and two succs. 1112 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1113 1114 // Okay, this is a simple enough basic block. See if any phi values are 1115 // constants. 1116 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1117 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1118 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1119 1120 // Okay, we now know that all edges from PredBB should be revectored to 1121 // branch to RealDest. 1122 BasicBlock *PredBB = PN->getIncomingBlock(i); 1123 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1124 1125 if (RealDest == BB) continue; // Skip self loops. 1126 1127 // The dest block might have PHI nodes, other predecessors and other 1128 // difficult cases. Instead of being smart about this, just insert a new 1129 // block that jumps to the destination block, effectively splitting 1130 // the edge we are about to create. 1131 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1132 RealDest->getName()+".critedge", 1133 RealDest->getParent(), RealDest); 1134 BranchInst::Create(RealDest, EdgeBB); 1135 1136 // Update PHI nodes. 1137 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1138 1139 // BB may have instructions that are being threaded over. Clone these 1140 // instructions into EdgeBB. We know that there will be no uses of the 1141 // cloned instructions outside of EdgeBB. 1142 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1143 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1144 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1145 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1146 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1147 continue; 1148 } 1149 // Clone the instruction. 1150 Instruction *N = BBI->clone(); 1151 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1152 1153 // Update operands due to translation. 1154 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1155 i != e; ++i) { 1156 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1157 if (PI != TranslateMap.end()) 1158 *i = PI->second; 1159 } 1160 1161 // Check for trivial simplification. 1162 if (Value *V = SimplifyInstruction(N, TD)) { 1163 TranslateMap[BBI] = V; 1164 delete N; // Instruction folded away, don't need actual inst 1165 } else { 1166 // Insert the new instruction into its new home. 1167 EdgeBB->getInstList().insert(InsertPt, N); 1168 if (!BBI->use_empty()) 1169 TranslateMap[BBI] = N; 1170 } 1171 } 1172 1173 // Loop over all of the edges from PredBB to BB, changing them to branch 1174 // to EdgeBB instead. 1175 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1176 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1177 if (PredBBTI->getSuccessor(i) == BB) { 1178 BB->removePredecessor(PredBB); 1179 PredBBTI->setSuccessor(i, EdgeBB); 1180 } 1181 1182 // Recurse, simplifying any other constants. 1183 return FoldCondBranchOnPHI(BI, TD) | true; 1184 } 1185 1186 return false; 1187 } 1188 1189 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1190 /// PHI node, see if we can eliminate it. 1191 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { 1192 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1193 // statement", which has a very simple dominance structure. Basically, we 1194 // are trying to find the condition that is being branched on, which 1195 // subsequently causes this merge to happen. We really want control 1196 // dependence information for this check, but simplifycfg can't keep it up 1197 // to date, and this catches most of the cases we care about anyway. 1198 BasicBlock *BB = PN->getParent(); 1199 BasicBlock *IfTrue, *IfFalse; 1200 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1201 if (!IfCond || 1202 // Don't bother if the branch will be constant folded trivially. 1203 isa<ConstantInt>(IfCond)) 1204 return false; 1205 1206 // Okay, we found that we can merge this two-entry phi node into a select. 1207 // Doing so would require us to fold *all* two entry phi nodes in this block. 1208 // At some point this becomes non-profitable (particularly if the target 1209 // doesn't support cmov's). Only do this transformation if there are two or 1210 // fewer PHI nodes in this block. 1211 unsigned NumPhis = 0; 1212 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1213 if (NumPhis > 2) 1214 return false; 1215 1216 // Loop over the PHI's seeing if we can promote them all to select 1217 // instructions. While we are at it, keep track of the instructions 1218 // that need to be moved to the dominating block. 1219 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1220 1221 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1222 PHINode *PN = cast<PHINode>(II++); 1223 if (Value *V = SimplifyInstruction(PN, TD)) { 1224 PN->replaceAllUsesWith(V); 1225 PN->eraseFromParent(); 1226 continue; 1227 } 1228 1229 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts) || 1230 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts)) 1231 return false; 1232 } 1233 1234 // If we folded the the first phi, PN dangles at this point. Refresh it. If 1235 // we ran out of PHIs then we simplified them all. 1236 PN = dyn_cast<PHINode>(BB->begin()); 1237 if (PN == 0) return true; 1238 1239 // Don't fold i1 branches on PHIs which contain binary operators. These can 1240 // often be turned into switches and other things. 1241 if (PN->getType()->isIntegerTy(1) && 1242 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1243 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1244 isa<BinaryOperator>(IfCond))) 1245 return false; 1246 1247 // If we all PHI nodes are promotable, check to make sure that all 1248 // instructions in the predecessor blocks can be promoted as well. If 1249 // not, we won't be able to get rid of the control flow, so it's not 1250 // worth promoting to select instructions. 1251 BasicBlock *DomBlock = 0; 1252 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1253 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1254 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1255 IfBlock1 = 0; 1256 } else { 1257 DomBlock = *pred_begin(IfBlock1); 1258 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1259 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1260 // This is not an aggressive instruction that we can promote. 1261 // Because of this, we won't be able to get rid of the control 1262 // flow, so the xform is not worth it. 1263 return false; 1264 } 1265 } 1266 1267 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1268 IfBlock2 = 0; 1269 } else { 1270 DomBlock = *pred_begin(IfBlock2); 1271 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1272 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1273 // This is not an aggressive instruction that we can promote. 1274 // Because of this, we won't be able to get rid of the control 1275 // flow, so the xform is not worth it. 1276 return false; 1277 } 1278 } 1279 1280 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1281 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1282 1283 // If we can still promote the PHI nodes after this gauntlet of tests, 1284 // do all of the PHI's now. 1285 Instruction *InsertPt = DomBlock->getTerminator(); 1286 1287 // Move all 'aggressive' instructions, which are defined in the 1288 // conditional parts of the if's up to the dominating block. 1289 if (IfBlock1) 1290 DomBlock->getInstList().splice(InsertPt, 1291 IfBlock1->getInstList(), IfBlock1->begin(), 1292 IfBlock1->getTerminator()); 1293 if (IfBlock2) 1294 DomBlock->getInstList().splice(InsertPt, 1295 IfBlock2->getInstList(), IfBlock2->begin(), 1296 IfBlock2->getTerminator()); 1297 1298 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1299 // Change the PHI node into a select instruction. 1300 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1301 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1302 1303 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", InsertPt); 1304 PN->replaceAllUsesWith(NV); 1305 NV->takeName(PN); 1306 PN->eraseFromParent(); 1307 } 1308 1309 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1310 // has been flattened. Change DomBlock to jump directly to our new block to 1311 // avoid other simplifycfg's kicking in on the diamond. 1312 TerminatorInst *OldTI = DomBlock->getTerminator(); 1313 BranchInst::Create(BB, OldTI); 1314 OldTI->eraseFromParent(); 1315 return true; 1316 } 1317 1318 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1319 /// to two returning blocks, try to merge them together into one return, 1320 /// introducing a select if the return values disagree. 1321 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1322 assert(BI->isConditional() && "Must be a conditional branch"); 1323 BasicBlock *TrueSucc = BI->getSuccessor(0); 1324 BasicBlock *FalseSucc = BI->getSuccessor(1); 1325 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1326 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1327 1328 // Check to ensure both blocks are empty (just a return) or optionally empty 1329 // with PHI nodes. If there are other instructions, merging would cause extra 1330 // computation on one path or the other. 1331 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1332 return false; 1333 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1334 return false; 1335 1336 // Okay, we found a branch that is going to two return nodes. If 1337 // there is no return value for this function, just change the 1338 // branch into a return. 1339 if (FalseRet->getNumOperands() == 0) { 1340 TrueSucc->removePredecessor(BI->getParent()); 1341 FalseSucc->removePredecessor(BI->getParent()); 1342 ReturnInst::Create(BI->getContext(), 0, BI); 1343 EraseTerminatorInstAndDCECond(BI); 1344 return true; 1345 } 1346 1347 // Otherwise, figure out what the true and false return values are 1348 // so we can insert a new select instruction. 1349 Value *TrueValue = TrueRet->getReturnValue(); 1350 Value *FalseValue = FalseRet->getReturnValue(); 1351 1352 // Unwrap any PHI nodes in the return blocks. 1353 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1354 if (TVPN->getParent() == TrueSucc) 1355 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1356 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1357 if (FVPN->getParent() == FalseSucc) 1358 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1359 1360 // In order for this transformation to be safe, we must be able to 1361 // unconditionally execute both operands to the return. This is 1362 // normally the case, but we could have a potentially-trapping 1363 // constant expression that prevents this transformation from being 1364 // safe. 1365 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1366 if (TCV->canTrap()) 1367 return false; 1368 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1369 if (FCV->canTrap()) 1370 return false; 1371 1372 // Okay, we collected all the mapped values and checked them for sanity, and 1373 // defined to really do this transformation. First, update the CFG. 1374 TrueSucc->removePredecessor(BI->getParent()); 1375 FalseSucc->removePredecessor(BI->getParent()); 1376 1377 // Insert select instructions where needed. 1378 Value *BrCond = BI->getCondition(); 1379 if (TrueValue) { 1380 // Insert a select if the results differ. 1381 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1382 } else if (isa<UndefValue>(TrueValue)) { 1383 TrueValue = FalseValue; 1384 } else { 1385 TrueValue = SelectInst::Create(BrCond, TrueValue, 1386 FalseValue, "retval", BI); 1387 } 1388 } 1389 1390 Value *RI = !TrueValue ? 1391 ReturnInst::Create(BI->getContext(), BI) : 1392 ReturnInst::Create(BI->getContext(), TrueValue, BI); 1393 (void) RI; 1394 1395 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1396 << "\n " << *BI << "NewRet = " << *RI 1397 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1398 1399 EraseTerminatorInstAndDCECond(BI); 1400 1401 return true; 1402 } 1403 1404 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1405 /// predecessor branches to us and one of our successors, fold the block into 1406 /// the predecessor and use logical operations to pick the right destination. 1407 bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1408 BasicBlock *BB = BI->getParent(); 1409 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1410 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1411 Cond->getParent() != BB || !Cond->hasOneUse()) 1412 return false; 1413 1414 // Only allow this if the condition is a simple instruction that can be 1415 // executed unconditionally. It must be in the same block as the branch, and 1416 // must be at the front of the block. 1417 BasicBlock::iterator FrontIt = BB->front(); 1418 1419 // Ignore dbg intrinsics. 1420 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1421 1422 // Allow a single instruction to be hoisted in addition to the compare 1423 // that feeds the branch. We later ensure that any values that _it_ uses 1424 // were also live in the predecessor, so that we don't unnecessarily create 1425 // register pressure or inhibit out-of-order execution. 1426 Instruction *BonusInst = 0; 1427 if (&*FrontIt != Cond && 1428 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1429 FrontIt->isSafeToSpeculativelyExecute()) { 1430 BonusInst = &*FrontIt; 1431 ++FrontIt; 1432 1433 // Ignore dbg intrinsics. 1434 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1435 } 1436 1437 // Only a single bonus inst is allowed. 1438 if (&*FrontIt != Cond) 1439 return false; 1440 1441 // Make sure the instruction after the condition is the cond branch. 1442 BasicBlock::iterator CondIt = Cond; ++CondIt; 1443 1444 // Ingore dbg intrinsics. 1445 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1446 1447 if (&*CondIt != BI) 1448 return false; 1449 1450 // Cond is known to be a compare or binary operator. Check to make sure that 1451 // neither operand is a potentially-trapping constant expression. 1452 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1453 if (CE->canTrap()) 1454 return false; 1455 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1456 if (CE->canTrap()) 1457 return false; 1458 1459 // Finally, don't infinitely unroll conditional loops. 1460 BasicBlock *TrueDest = BI->getSuccessor(0); 1461 BasicBlock *FalseDest = BI->getSuccessor(1); 1462 if (TrueDest == BB || FalseDest == BB) 1463 return false; 1464 1465 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1466 BasicBlock *PredBlock = *PI; 1467 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1468 1469 // Check that we have two conditional branches. If there is a PHI node in 1470 // the common successor, verify that the same value flows in from both 1471 // blocks. 1472 if (PBI == 0 || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI)) 1473 continue; 1474 1475 // Determine if the two branches share a common destination. 1476 Instruction::BinaryOps Opc; 1477 bool InvertPredCond = false; 1478 1479 if (PBI->getSuccessor(0) == TrueDest) 1480 Opc = Instruction::Or; 1481 else if (PBI->getSuccessor(1) == FalseDest) 1482 Opc = Instruction::And; 1483 else if (PBI->getSuccessor(0) == FalseDest) 1484 Opc = Instruction::And, InvertPredCond = true; 1485 else if (PBI->getSuccessor(1) == TrueDest) 1486 Opc = Instruction::Or, InvertPredCond = true; 1487 else 1488 continue; 1489 1490 // Ensure that any values used in the bonus instruction are also used 1491 // by the terminator of the predecessor. This means that those values 1492 // must already have been resolved, so we won't be inhibiting the 1493 // out-of-order core by speculating them earlier. 1494 if (BonusInst) { 1495 // Collect the values used by the bonus inst 1496 SmallPtrSet<Value*, 4> UsedValues; 1497 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1498 OE = BonusInst->op_end(); OI != OE; ++OI) { 1499 Value* V = *OI; 1500 if (!isa<Constant>(V)) 1501 UsedValues.insert(V); 1502 } 1503 1504 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1505 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1506 1507 // Walk up to four levels back up the use-def chain of the predecessor's 1508 // terminator to see if all those values were used. The choice of four 1509 // levels is arbitrary, to provide a compile-time-cost bound. 1510 while (!Worklist.empty()) { 1511 std::pair<Value*, unsigned> Pair = Worklist.back(); 1512 Worklist.pop_back(); 1513 1514 if (Pair.second >= 4) continue; 1515 UsedValues.erase(Pair.first); 1516 if (UsedValues.empty()) break; 1517 1518 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 1519 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1520 OI != OE; ++OI) 1521 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1522 } 1523 } 1524 1525 if (!UsedValues.empty()) return false; 1526 } 1527 1528 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1529 1530 // If we need to invert the condition in the pred block to match, do so now. 1531 if (InvertPredCond) { 1532 Value *NewCond = PBI->getCondition(); 1533 1534 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1535 CmpInst *CI = cast<CmpInst>(NewCond); 1536 CI->setPredicate(CI->getInversePredicate()); 1537 } else { 1538 NewCond = BinaryOperator::CreateNot(NewCond, 1539 PBI->getCondition()->getName()+".not", PBI); 1540 } 1541 1542 PBI->setCondition(NewCond); 1543 BasicBlock *OldTrue = PBI->getSuccessor(0); 1544 BasicBlock *OldFalse = PBI->getSuccessor(1); 1545 PBI->setSuccessor(0, OldFalse); 1546 PBI->setSuccessor(1, OldTrue); 1547 } 1548 1549 // If we have a bonus inst, clone it into the predecessor block. 1550 Instruction *NewBonus = 0; 1551 if (BonusInst) { 1552 NewBonus = BonusInst->clone(); 1553 PredBlock->getInstList().insert(PBI, NewBonus); 1554 NewBonus->takeName(BonusInst); 1555 BonusInst->setName(BonusInst->getName()+".old"); 1556 } 1557 1558 // Clone Cond into the predecessor basic block, and or/and the 1559 // two conditions together. 1560 Instruction *New = Cond->clone(); 1561 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1562 PredBlock->getInstList().insert(PBI, New); 1563 New->takeName(Cond); 1564 Cond->setName(New->getName()+".old"); 1565 1566 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1567 New, "or.cond", PBI); 1568 PBI->setCondition(NewCond); 1569 if (PBI->getSuccessor(0) == BB) { 1570 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1571 PBI->setSuccessor(0, TrueDest); 1572 } 1573 if (PBI->getSuccessor(1) == BB) { 1574 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1575 PBI->setSuccessor(1, FalseDest); 1576 } 1577 1578 // Copy any debug value intrinsics into the end of PredBlock. 1579 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 1580 if (isa<DbgInfoIntrinsic>(*I)) 1581 I->clone()->insertBefore(PBI); 1582 1583 return true; 1584 } 1585 return false; 1586 } 1587 1588 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1589 /// predecessor of another block, this function tries to simplify it. We know 1590 /// that PBI and BI are both conditional branches, and BI is in one of the 1591 /// successor blocks of PBI - PBI branches to BI. 1592 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1593 assert(PBI->isConditional() && BI->isConditional()); 1594 BasicBlock *BB = BI->getParent(); 1595 1596 // If this block ends with a branch instruction, and if there is a 1597 // predecessor that ends on a branch of the same condition, make 1598 // this conditional branch redundant. 1599 if (PBI->getCondition() == BI->getCondition() && 1600 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1601 // Okay, the outcome of this conditional branch is statically 1602 // knowable. If this block had a single pred, handle specially. 1603 if (BB->getSinglePredecessor()) { 1604 // Turn this into a branch on constant. 1605 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1606 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1607 CondIsTrue)); 1608 return true; // Nuke the branch on constant. 1609 } 1610 1611 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1612 // in the constant and simplify the block result. Subsequent passes of 1613 // simplifycfg will thread the block. 1614 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1615 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 1616 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 1617 std::distance(PB, PE), 1618 BI->getCondition()->getName() + ".pr", 1619 BB->begin()); 1620 // Okay, we're going to insert the PHI node. Since PBI is not the only 1621 // predecessor, compute the PHI'd conditional value for all of the preds. 1622 // Any predecessor where the condition is not computable we keep symbolic. 1623 for (pred_iterator PI = PB; PI != PE; ++PI) { 1624 BasicBlock *P = *PI; 1625 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 1626 PBI != BI && PBI->isConditional() && 1627 PBI->getCondition() == BI->getCondition() && 1628 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1629 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1630 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1631 CondIsTrue), P); 1632 } else { 1633 NewPN->addIncoming(BI->getCondition(), P); 1634 } 1635 } 1636 1637 BI->setCondition(NewPN); 1638 return true; 1639 } 1640 } 1641 1642 // If this is a conditional branch in an empty block, and if any 1643 // predecessors is a conditional branch to one of our destinations, 1644 // fold the conditions into logical ops and one cond br. 1645 BasicBlock::iterator BBI = BB->begin(); 1646 // Ignore dbg intrinsics. 1647 while (isa<DbgInfoIntrinsic>(BBI)) 1648 ++BBI; 1649 if (&*BBI != BI) 1650 return false; 1651 1652 1653 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1654 if (CE->canTrap()) 1655 return false; 1656 1657 int PBIOp, BIOp; 1658 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1659 PBIOp = BIOp = 0; 1660 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1661 PBIOp = 0, BIOp = 1; 1662 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1663 PBIOp = 1, BIOp = 0; 1664 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1665 PBIOp = BIOp = 1; 1666 else 1667 return false; 1668 1669 // Check to make sure that the other destination of this branch 1670 // isn't BB itself. If so, this is an infinite loop that will 1671 // keep getting unwound. 1672 if (PBI->getSuccessor(PBIOp) == BB) 1673 return false; 1674 1675 // Do not perform this transformation if it would require 1676 // insertion of a large number of select instructions. For targets 1677 // without predication/cmovs, this is a big pessimization. 1678 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1679 1680 unsigned NumPhis = 0; 1681 for (BasicBlock::iterator II = CommonDest->begin(); 1682 isa<PHINode>(II); ++II, ++NumPhis) 1683 if (NumPhis > 2) // Disable this xform. 1684 return false; 1685 1686 // Finally, if everything is ok, fold the branches to logical ops. 1687 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1688 1689 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 1690 << "AND: " << *BI->getParent()); 1691 1692 1693 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1694 // branch in it, where one edge (OtherDest) goes back to itself but the other 1695 // exits. We don't *know* that the program avoids the infinite loop 1696 // (even though that seems likely). If we do this xform naively, we'll end up 1697 // recursively unpeeling the loop. Since we know that (after the xform is 1698 // done) that the block *is* infinite if reached, we just make it an obviously 1699 // infinite loop with no cond branch. 1700 if (OtherDest == BB) { 1701 // Insert it at the end of the function, because it's either code, 1702 // or it won't matter if it's hot. :) 1703 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 1704 "infloop", BB->getParent()); 1705 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1706 OtherDest = InfLoopBlock; 1707 } 1708 1709 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1710 1711 // BI may have other predecessors. Because of this, we leave 1712 // it alone, but modify PBI. 1713 1714 // Make sure we get to CommonDest on True&True directions. 1715 Value *PBICond = PBI->getCondition(); 1716 if (PBIOp) 1717 PBICond = BinaryOperator::CreateNot(PBICond, 1718 PBICond->getName()+".not", 1719 PBI); 1720 Value *BICond = BI->getCondition(); 1721 if (BIOp) 1722 BICond = BinaryOperator::CreateNot(BICond, 1723 BICond->getName()+".not", 1724 PBI); 1725 // Merge the conditions. 1726 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1727 1728 // Modify PBI to branch on the new condition to the new dests. 1729 PBI->setCondition(Cond); 1730 PBI->setSuccessor(0, CommonDest); 1731 PBI->setSuccessor(1, OtherDest); 1732 1733 // OtherDest may have phi nodes. If so, add an entry from PBI's 1734 // block that are identical to the entries for BI's block. 1735 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 1736 1737 // We know that the CommonDest already had an edge from PBI to 1738 // it. If it has PHIs though, the PHIs may have different 1739 // entries for BB and PBI's BB. If so, insert a select to make 1740 // them agree. 1741 PHINode *PN; 1742 for (BasicBlock::iterator II = CommonDest->begin(); 1743 (PN = dyn_cast<PHINode>(II)); ++II) { 1744 Value *BIV = PN->getIncomingValueForBlock(BB); 1745 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1746 Value *PBIV = PN->getIncomingValue(PBBIdx); 1747 if (BIV != PBIV) { 1748 // Insert a select in PBI to pick the right value. 1749 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1750 PBIV->getName()+".mux", PBI); 1751 PN->setIncomingValue(PBBIdx, NV); 1752 } 1753 } 1754 1755 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 1756 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1757 1758 // This basic block is probably dead. We know it has at least 1759 // one fewer predecessor. 1760 return true; 1761 } 1762 1763 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 1764 // branch to TrueBB if Cond is true or to FalseBB if Cond is false. 1765 // Takes care of updating the successors and removing the old terminator. 1766 // Also makes sure not to introduce new successors by assuming that edges to 1767 // non-successor TrueBBs and FalseBBs aren't reachable. 1768 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 1769 BasicBlock *TrueBB, BasicBlock *FalseBB){ 1770 // Remove any superfluous successor edges from the CFG. 1771 // First, figure out which successors to preserve. 1772 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 1773 // successor. 1774 BasicBlock *KeepEdge1 = TrueBB; 1775 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 1776 1777 // Then remove the rest. 1778 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 1779 BasicBlock *Succ = OldTerm->getSuccessor(I); 1780 // Make sure only to keep exactly one copy of each edge. 1781 if (Succ == KeepEdge1) 1782 KeepEdge1 = 0; 1783 else if (Succ == KeepEdge2) 1784 KeepEdge2 = 0; 1785 else 1786 Succ->removePredecessor(OldTerm->getParent()); 1787 } 1788 1789 // Insert an appropriate new terminator. 1790 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 1791 if (TrueBB == FalseBB) 1792 // We were only looking for one successor, and it was present. 1793 // Create an unconditional branch to it. 1794 BranchInst::Create(TrueBB, OldTerm); 1795 else 1796 // We found both of the successors we were looking for. 1797 // Create a conditional branch sharing the condition of the select. 1798 BranchInst::Create(TrueBB, FalseBB, Cond, OldTerm); 1799 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 1800 // Neither of the selected blocks were successors, so this 1801 // terminator must be unreachable. 1802 new UnreachableInst(OldTerm->getContext(), OldTerm); 1803 } else { 1804 // One of the selected values was a successor, but the other wasn't. 1805 // Insert an unconditional branch to the one that was found; 1806 // the edge to the one that wasn't must be unreachable. 1807 if (KeepEdge1 == 0) 1808 // Only TrueBB was found. 1809 BranchInst::Create(TrueBB, OldTerm); 1810 else 1811 // Only FalseBB was found. 1812 BranchInst::Create(FalseBB, OldTerm); 1813 } 1814 1815 EraseTerminatorInstAndDCECond(OldTerm); 1816 return true; 1817 } 1818 1819 // SimplifySwitchOnSelect - Replaces 1820 // (switch (select cond, X, Y)) on constant X, Y 1821 // with a branch - conditional if X and Y lead to distinct BBs, 1822 // unconditional otherwise. 1823 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 1824 // Check for constant integer values in the select. 1825 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 1826 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 1827 if (!TrueVal || !FalseVal) 1828 return false; 1829 1830 // Find the relevant condition and destinations. 1831 Value *Condition = Select->getCondition(); 1832 BasicBlock *TrueBB = SI->getSuccessor(SI->findCaseValue(TrueVal)); 1833 BasicBlock *FalseBB = SI->getSuccessor(SI->findCaseValue(FalseVal)); 1834 1835 // Perform the actual simplification. 1836 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB); 1837 } 1838 1839 // SimplifyIndirectBrOnSelect - Replaces 1840 // (indirectbr (select cond, blockaddress(@fn, BlockA), 1841 // blockaddress(@fn, BlockB))) 1842 // with 1843 // (br cond, BlockA, BlockB). 1844 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 1845 // Check that both operands of the select are block addresses. 1846 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 1847 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 1848 if (!TBA || !FBA) 1849 return false; 1850 1851 // Extract the actual blocks. 1852 BasicBlock *TrueBB = TBA->getBasicBlock(); 1853 BasicBlock *FalseBB = FBA->getBasicBlock(); 1854 1855 // Perform the actual simplification. 1856 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); 1857 } 1858 1859 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 1860 /// instruction (a seteq/setne with a constant) as the only instruction in a 1861 /// block that ends with an uncond branch. We are looking for a very specific 1862 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 1863 /// this case, we merge the first two "or's of icmp" into a switch, but then the 1864 /// default value goes to an uncond block with a seteq in it, we get something 1865 /// like: 1866 /// 1867 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 1868 /// DEFAULT: 1869 /// %tmp = icmp eq i8 %A, 92 1870 /// br label %end 1871 /// end: 1872 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 1873 /// 1874 /// We prefer to split the edge to 'end' so that there is a true/false entry to 1875 /// the PHI, merging the third icmp into the switch. 1876 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 1877 const TargetData *TD) { 1878 BasicBlock *BB = ICI->getParent(); 1879 // If the block has any PHIs in it or the icmp has multiple uses, it is too 1880 // complex. 1881 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 1882 1883 Value *V = ICI->getOperand(0); 1884 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 1885 1886 // The pattern we're looking for is where our only predecessor is a switch on 1887 // 'V' and this block is the default case for the switch. In this case we can 1888 // fold the compared value into the switch to simplify things. 1889 BasicBlock *Pred = BB->getSinglePredecessor(); 1890 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 1891 1892 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 1893 if (SI->getCondition() != V) 1894 return false; 1895 1896 // If BB is reachable on a non-default case, then we simply know the value of 1897 // V in this block. Substitute it and constant fold the icmp instruction 1898 // away. 1899 if (SI->getDefaultDest() != BB) { 1900 ConstantInt *VVal = SI->findCaseDest(BB); 1901 assert(VVal && "Should have a unique destination value"); 1902 ICI->setOperand(0, VVal); 1903 1904 if (Value *V = SimplifyInstruction(ICI, TD)) { 1905 ICI->replaceAllUsesWith(V); 1906 ICI->eraseFromParent(); 1907 } 1908 // BB is now empty, so it is likely to simplify away. 1909 return SimplifyCFG(BB) | true; 1910 } 1911 1912 // Ok, the block is reachable from the default dest. If the constant we're 1913 // comparing exists in one of the other edges, then we can constant fold ICI 1914 // and zap it. 1915 if (SI->findCaseValue(Cst) != 0) { 1916 Value *V; 1917 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1918 V = ConstantInt::getFalse(BB->getContext()); 1919 else 1920 V = ConstantInt::getTrue(BB->getContext()); 1921 1922 ICI->replaceAllUsesWith(V); 1923 ICI->eraseFromParent(); 1924 // BB is now empty, so it is likely to simplify away. 1925 return SimplifyCFG(BB) | true; 1926 } 1927 1928 // The use of the icmp has to be in the 'end' block, by the only PHI node in 1929 // the block. 1930 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 1931 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 1932 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 1933 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 1934 return false; 1935 1936 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 1937 // true in the PHI. 1938 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 1939 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 1940 1941 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1942 std::swap(DefaultCst, NewCst); 1943 1944 // Replace ICI (which is used by the PHI for the default value) with true or 1945 // false depending on if it is EQ or NE. 1946 ICI->replaceAllUsesWith(DefaultCst); 1947 ICI->eraseFromParent(); 1948 1949 // Okay, the switch goes to this block on a default value. Add an edge from 1950 // the switch to the merge point on the compared value. 1951 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 1952 BB->getParent(), BB); 1953 SI->addCase(Cst, NewBB); 1954 1955 // NewBB branches to the phi block, add the uncond branch and the phi entry. 1956 BranchInst::Create(SuccBlock, NewBB); 1957 PHIUse->addIncoming(NewCst, NewBB); 1958 return true; 1959 } 1960 1961 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 1962 /// Check to see if it is branching on an or/and chain of icmp instructions, and 1963 /// fold it into a switch instruction if so. 1964 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD) { 1965 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1966 if (Cond == 0) return false; 1967 1968 1969 // Change br (X == 0 | X == 1), T, F into a switch instruction. 1970 // If this is a bunch of seteq's or'd together, or if it's a bunch of 1971 // 'setne's and'ed together, collect them. 1972 Value *CompVal = 0; 1973 std::vector<ConstantInt*> Values; 1974 bool TrueWhenEqual = true; 1975 Value *ExtraCase = 0; 1976 unsigned UsedICmps = 0; 1977 1978 if (Cond->getOpcode() == Instruction::Or) { 1979 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 1980 UsedICmps); 1981 } else if (Cond->getOpcode() == Instruction::And) { 1982 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 1983 UsedICmps); 1984 TrueWhenEqual = false; 1985 } 1986 1987 // If we didn't have a multiply compared value, fail. 1988 if (CompVal == 0) return false; 1989 1990 // Avoid turning single icmps into a switch. 1991 if (UsedICmps <= 1) 1992 return false; 1993 1994 // There might be duplicate constants in the list, which the switch 1995 // instruction can't handle, remove them now. 1996 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 1997 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 1998 1999 // If Extra was used, we require at least two switch values to do the 2000 // transformation. A switch with one value is just an cond branch. 2001 if (ExtraCase && Values.size() < 2) return false; 2002 2003 // Figure out which block is which destination. 2004 BasicBlock *DefaultBB = BI->getSuccessor(1); 2005 BasicBlock *EdgeBB = BI->getSuccessor(0); 2006 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2007 2008 BasicBlock *BB = BI->getParent(); 2009 2010 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2011 << " cases into SWITCH. BB is:\n" << *BB); 2012 2013 // If there are any extra values that couldn't be folded into the switch 2014 // then we evaluate them with an explicit branch first. Split the block 2015 // right before the condbr to handle it. 2016 if (ExtraCase) { 2017 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2018 // Remove the uncond branch added to the old block. 2019 TerminatorInst *OldTI = BB->getTerminator(); 2020 2021 if (TrueWhenEqual) 2022 BranchInst::Create(EdgeBB, NewBB, ExtraCase, OldTI); 2023 else 2024 BranchInst::Create(NewBB, EdgeBB, ExtraCase, OldTI); 2025 2026 OldTI->eraseFromParent(); 2027 2028 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2029 // for the edge we just added. 2030 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2031 2032 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2033 << "\nEXTRABB = " << *BB); 2034 BB = NewBB; 2035 } 2036 2037 // Convert pointer to int before we switch. 2038 if (CompVal->getType()->isPointerTy()) { 2039 assert(TD && "Cannot switch on pointer without TargetData"); 2040 CompVal = new PtrToIntInst(CompVal, 2041 TD->getIntPtrType(CompVal->getContext()), 2042 "magicptr", BI); 2043 } 2044 2045 // Create the new switch instruction now. 2046 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, Values.size(), BI); 2047 2048 // Add all of the 'cases' to the switch instruction. 2049 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2050 New->addCase(Values[i], EdgeBB); 2051 2052 // We added edges from PI to the EdgeBB. As such, if there were any 2053 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2054 // the number of edges added. 2055 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2056 isa<PHINode>(BBI); ++BBI) { 2057 PHINode *PN = cast<PHINode>(BBI); 2058 Value *InVal = PN->getIncomingValueForBlock(BB); 2059 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2060 PN->addIncoming(InVal, BB); 2061 } 2062 2063 // Erase the old branch instruction. 2064 EraseTerminatorInstAndDCECond(BI); 2065 2066 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2067 return true; 2068 } 2069 2070 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI) { 2071 BasicBlock *BB = RI->getParent(); 2072 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2073 2074 // Find predecessors that end with branches. 2075 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2076 SmallVector<BranchInst*, 8> CondBranchPreds; 2077 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2078 BasicBlock *P = *PI; 2079 TerminatorInst *PTI = P->getTerminator(); 2080 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2081 if (BI->isUnconditional()) 2082 UncondBranchPreds.push_back(P); 2083 else 2084 CondBranchPreds.push_back(BI); 2085 } 2086 } 2087 2088 // If we found some, do the transformation! 2089 if (!UncondBranchPreds.empty() && DupRet) { 2090 while (!UncondBranchPreds.empty()) { 2091 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2092 DEBUG(dbgs() << "FOLDING: " << *BB 2093 << "INTO UNCOND BRANCH PRED: " << *Pred); 2094 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2095 } 2096 2097 // If we eliminated all predecessors of the block, delete the block now. 2098 if (pred_begin(BB) == pred_end(BB)) 2099 // We know there are no successors, so just nuke the block. 2100 BB->eraseFromParent(); 2101 2102 return true; 2103 } 2104 2105 // Check out all of the conditional branches going to this return 2106 // instruction. If any of them just select between returns, change the 2107 // branch itself into a select/return pair. 2108 while (!CondBranchPreds.empty()) { 2109 BranchInst *BI = CondBranchPreds.pop_back_val(); 2110 2111 // Check to see if the non-BB successor is also a return block. 2112 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2113 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2114 SimplifyCondBranchToTwoReturns(BI)) 2115 return true; 2116 } 2117 return false; 2118 } 2119 2120 bool SimplifyCFGOpt::SimplifyUnwind(UnwindInst *UI) { 2121 // Check to see if the first instruction in this block is just an unwind. 2122 // If so, replace any invoke instructions which use this as an exception 2123 // destination with call instructions. 2124 BasicBlock *BB = UI->getParent(); 2125 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2126 2127 bool Changed = false; 2128 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2129 while (!Preds.empty()) { 2130 BasicBlock *Pred = Preds.back(); 2131 InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator()); 2132 if (II && II->getUnwindDest() == BB) { 2133 // Insert a new branch instruction before the invoke, because this 2134 // is now a fall through. 2135 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2136 Pred->getInstList().remove(II); // Take out of symbol table 2137 2138 // Insert the call now. 2139 SmallVector<Value*,8> Args(II->op_begin(), II->op_end()-3); 2140 CallInst *CI = CallInst::Create(II->getCalledValue(), 2141 Args.begin(), Args.end(), 2142 II->getName(), BI); 2143 CI->setCallingConv(II->getCallingConv()); 2144 CI->setAttributes(II->getAttributes()); 2145 // If the invoke produced a value, the Call now does instead. 2146 II->replaceAllUsesWith(CI); 2147 delete II; 2148 Changed = true; 2149 } 2150 2151 Preds.pop_back(); 2152 } 2153 2154 // If this block is now dead (and isn't the entry block), remove it. 2155 if (pred_begin(BB) == pred_end(BB) && 2156 BB != &BB->getParent()->getEntryBlock()) { 2157 // We know there are no successors, so just nuke the block. 2158 BB->eraseFromParent(); 2159 return true; 2160 } 2161 2162 return Changed; 2163 } 2164 2165 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2166 BasicBlock *BB = UI->getParent(); 2167 2168 bool Changed = false; 2169 2170 // If there are any instructions immediately before the unreachable that can 2171 // be removed, do so. 2172 while (UI != BB->begin()) { 2173 BasicBlock::iterator BBI = UI; 2174 --BBI; 2175 // Do not delete instructions that can have side effects, like calls 2176 // (which may never return) and volatile loads and stores. 2177 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2178 2179 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) 2180 if (SI->isVolatile()) 2181 break; 2182 2183 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) 2184 if (LI->isVolatile()) 2185 break; 2186 2187 // Delete this instruction (any uses are guaranteed to be dead) 2188 if (!BBI->use_empty()) 2189 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2190 BBI->eraseFromParent(); 2191 Changed = true; 2192 } 2193 2194 // If the unreachable instruction is the first in the block, take a gander 2195 // at all of the predecessors of this instruction, and simplify them. 2196 if (&BB->front() != UI) return Changed; 2197 2198 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2199 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2200 TerminatorInst *TI = Preds[i]->getTerminator(); 2201 2202 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2203 if (BI->isUnconditional()) { 2204 if (BI->getSuccessor(0) == BB) { 2205 new UnreachableInst(TI->getContext(), TI); 2206 TI->eraseFromParent(); 2207 Changed = true; 2208 } 2209 } else { 2210 if (BI->getSuccessor(0) == BB) { 2211 BranchInst::Create(BI->getSuccessor(1), BI); 2212 EraseTerminatorInstAndDCECond(BI); 2213 } else if (BI->getSuccessor(1) == BB) { 2214 BranchInst::Create(BI->getSuccessor(0), BI); 2215 EraseTerminatorInstAndDCECond(BI); 2216 Changed = true; 2217 } 2218 } 2219 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2220 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2221 if (SI->getSuccessor(i) == BB) { 2222 BB->removePredecessor(SI->getParent()); 2223 SI->removeCase(i); 2224 --i; --e; 2225 Changed = true; 2226 } 2227 // If the default value is unreachable, figure out the most popular 2228 // destination and make it the default. 2229 if (SI->getSuccessor(0) == BB) { 2230 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2231 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 2232 std::pair<unsigned, unsigned>& entry = 2233 Popularity[SI->getSuccessor(i)]; 2234 if (entry.first == 0) { 2235 entry.first = 1; 2236 entry.second = i; 2237 } else { 2238 entry.first++; 2239 } 2240 } 2241 2242 // Find the most popular block. 2243 unsigned MaxPop = 0; 2244 unsigned MaxIndex = 0; 2245 BasicBlock *MaxBlock = 0; 2246 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2247 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2248 if (I->second.first > MaxPop || 2249 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2250 MaxPop = I->second.first; 2251 MaxIndex = I->second.second; 2252 MaxBlock = I->first; 2253 } 2254 } 2255 if (MaxBlock) { 2256 // Make this the new default, allowing us to delete any explicit 2257 // edges to it. 2258 SI->setSuccessor(0, MaxBlock); 2259 Changed = true; 2260 2261 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2262 // it. 2263 if (isa<PHINode>(MaxBlock->begin())) 2264 for (unsigned i = 0; i != MaxPop-1; ++i) 2265 MaxBlock->removePredecessor(SI->getParent()); 2266 2267 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2268 if (SI->getSuccessor(i) == MaxBlock) { 2269 SI->removeCase(i); 2270 --i; --e; 2271 } 2272 } 2273 } 2274 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2275 if (II->getUnwindDest() == BB) { 2276 // Convert the invoke to a call instruction. This would be a good 2277 // place to note that the call does not throw though. 2278 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2279 II->removeFromParent(); // Take out of symbol table 2280 2281 // Insert the call now... 2282 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2283 CallInst *CI = CallInst::Create(II->getCalledValue(), 2284 Args.begin(), Args.end(), 2285 II->getName(), BI); 2286 CI->setCallingConv(II->getCallingConv()); 2287 CI->setAttributes(II->getAttributes()); 2288 // If the invoke produced a value, the call does now instead. 2289 II->replaceAllUsesWith(CI); 2290 delete II; 2291 Changed = true; 2292 } 2293 } 2294 } 2295 2296 // If this block is now dead, remove it. 2297 if (pred_begin(BB) == pred_end(BB) && 2298 BB != &BB->getParent()->getEntryBlock()) { 2299 // We know there are no successors, so just nuke the block. 2300 BB->eraseFromParent(); 2301 return true; 2302 } 2303 2304 return Changed; 2305 } 2306 2307 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 2308 /// integer range comparison into a sub, an icmp and a branch. 2309 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI) { 2310 assert(SI->getNumCases() > 2 && "Degenerate switch?"); 2311 2312 // Make sure all cases point to the same destination and gather the values. 2313 SmallVector<ConstantInt *, 16> Cases; 2314 Cases.push_back(SI->getCaseValue(1)); 2315 for (unsigned I = 2, E = SI->getNumCases(); I != E; ++I) { 2316 if (SI->getSuccessor(I-1) != SI->getSuccessor(I)) 2317 return false; 2318 Cases.push_back(SI->getCaseValue(I)); 2319 } 2320 assert(Cases.size() == SI->getNumCases()-1 && "Not all cases gathered"); 2321 2322 // Sort the case values, then check if they form a range we can transform. 2323 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 2324 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 2325 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 2326 return false; 2327 } 2328 2329 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 2330 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()-1); 2331 2332 Value *Sub = SI->getCondition(); 2333 if (!Offset->isNullValue()) 2334 Sub = BinaryOperator::CreateAdd(Sub, Offset, Sub->getName()+".off", SI); 2335 Value *Cmp = new ICmpInst(SI, ICmpInst::ICMP_ULT, Sub, NumCases, "switch"); 2336 BranchInst::Create(SI->getSuccessor(1), SI->getDefaultDest(), Cmp, SI); 2337 2338 // Prune obsolete incoming values off the successor's PHI nodes. 2339 for (BasicBlock::iterator BBI = SI->getSuccessor(1)->begin(); 2340 isa<PHINode>(BBI); ++BBI) { 2341 for (unsigned I = 0, E = SI->getNumCases()-2; I != E; ++I) 2342 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 2343 } 2344 SI->eraseFromParent(); 2345 2346 return true; 2347 } 2348 2349 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI) { 2350 // If this switch is too complex to want to look at, ignore it. 2351 if (!isValueEqualityComparison(SI)) 2352 return false; 2353 2354 BasicBlock *BB = SI->getParent(); 2355 2356 // If we only have one predecessor, and if it is a branch on this value, 2357 // see if that predecessor totally determines the outcome of this switch. 2358 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2359 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 2360 return SimplifyCFG(BB) | true; 2361 2362 Value *Cond = SI->getCondition(); 2363 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 2364 if (SimplifySwitchOnSelect(SI, Select)) 2365 return SimplifyCFG(BB) | true; 2366 2367 // If the block only contains the switch, see if we can fold the block 2368 // away into any preds. 2369 BasicBlock::iterator BBI = BB->begin(); 2370 // Ignore dbg intrinsics. 2371 while (isa<DbgInfoIntrinsic>(BBI)) 2372 ++BBI; 2373 if (SI == &*BBI) 2374 if (FoldValueComparisonIntoPredecessors(SI)) 2375 return SimplifyCFG(BB) | true; 2376 2377 // Try to transform the switch into an icmp and a branch. 2378 if (TurnSwitchRangeIntoICmp(SI)) 2379 return SimplifyCFG(BB) | true; 2380 2381 return false; 2382 } 2383 2384 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 2385 BasicBlock *BB = IBI->getParent(); 2386 bool Changed = false; 2387 2388 // Eliminate redundant destinations. 2389 SmallPtrSet<Value *, 8> Succs; 2390 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 2391 BasicBlock *Dest = IBI->getDestination(i); 2392 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 2393 Dest->removePredecessor(BB); 2394 IBI->removeDestination(i); 2395 --i; --e; 2396 Changed = true; 2397 } 2398 } 2399 2400 if (IBI->getNumDestinations() == 0) { 2401 // If the indirectbr has no successors, change it to unreachable. 2402 new UnreachableInst(IBI->getContext(), IBI); 2403 EraseTerminatorInstAndDCECond(IBI); 2404 return true; 2405 } 2406 2407 if (IBI->getNumDestinations() == 1) { 2408 // If the indirectbr has one successor, change it to a direct branch. 2409 BranchInst::Create(IBI->getDestination(0), IBI); 2410 EraseTerminatorInstAndDCECond(IBI); 2411 return true; 2412 } 2413 2414 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 2415 if (SimplifyIndirectBrOnSelect(IBI, SI)) 2416 return SimplifyCFG(BB) | true; 2417 } 2418 return Changed; 2419 } 2420 2421 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI) { 2422 BasicBlock *BB = BI->getParent(); 2423 2424 // If the Terminator is the only non-phi instruction, simplify the block. 2425 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg(); 2426 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 2427 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 2428 return true; 2429 2430 // If the only instruction in the block is a seteq/setne comparison 2431 // against a constant, try to simplify the block. 2432 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 2433 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 2434 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 2435 ; 2436 if (I->isTerminator() && TryToSimplifyUncondBranchWithICmpInIt(ICI, TD)) 2437 return true; 2438 } 2439 2440 return false; 2441 } 2442 2443 2444 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI) { 2445 BasicBlock *BB = BI->getParent(); 2446 2447 // Conditional branch 2448 if (isValueEqualityComparison(BI)) { 2449 // If we only have one predecessor, and if it is a branch on this value, 2450 // see if that predecessor totally determines the outcome of this 2451 // switch. 2452 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2453 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 2454 return SimplifyCFG(BB) | true; 2455 2456 // This block must be empty, except for the setcond inst, if it exists. 2457 // Ignore dbg intrinsics. 2458 BasicBlock::iterator I = BB->begin(); 2459 // Ignore dbg intrinsics. 2460 while (isa<DbgInfoIntrinsic>(I)) 2461 ++I; 2462 if (&*I == BI) { 2463 if (FoldValueComparisonIntoPredecessors(BI)) 2464 return SimplifyCFG(BB) | true; 2465 } else if (&*I == cast<Instruction>(BI->getCondition())){ 2466 ++I; 2467 // Ignore dbg intrinsics. 2468 while (isa<DbgInfoIntrinsic>(I)) 2469 ++I; 2470 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI)) 2471 return SimplifyCFG(BB) | true; 2472 } 2473 } 2474 2475 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 2476 if (SimplifyBranchOnICmpChain(BI, TD)) 2477 return true; 2478 2479 // We have a conditional branch to two blocks that are only reachable 2480 // from BI. We know that the condbr dominates the two blocks, so see if 2481 // there is any identical code in the "then" and "else" blocks. If so, we 2482 // can hoist it up to the branching block. 2483 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 2484 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2485 if (HoistThenElseCodeToIf(BI)) 2486 return SimplifyCFG(BB) | true; 2487 } else { 2488 // If Successor #1 has multiple preds, we may be able to conditionally 2489 // execute Successor #0 if it branches to successor #1. 2490 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 2491 if (Succ0TI->getNumSuccessors() == 1 && 2492 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 2493 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 2494 return SimplifyCFG(BB) | true; 2495 } 2496 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2497 // If Successor #0 has multiple preds, we may be able to conditionally 2498 // execute Successor #1 if it branches to successor #0. 2499 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 2500 if (Succ1TI->getNumSuccessors() == 1 && 2501 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 2502 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 2503 return SimplifyCFG(BB) | true; 2504 } 2505 2506 // If this is a branch on a phi node in the current block, thread control 2507 // through this block if any PHI node entries are constants. 2508 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 2509 if (PN->getParent() == BI->getParent()) 2510 if (FoldCondBranchOnPHI(BI, TD)) 2511 return SimplifyCFG(BB) | true; 2512 2513 // If this basic block is ONLY a setcc and a branch, and if a predecessor 2514 // branches to us and one of our successors, fold the setcc into the 2515 // predecessor and use logical operations to pick the right destination. 2516 if (FoldBranchToCommonDest(BI)) 2517 return SimplifyCFG(BB) | true; 2518 2519 // Scan predecessor blocks for conditional branches. 2520 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2521 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2522 if (PBI != BI && PBI->isConditional()) 2523 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2524 return SimplifyCFG(BB) | true; 2525 2526 return false; 2527 } 2528 2529 bool SimplifyCFGOpt::run(BasicBlock *BB) { 2530 bool Changed = false; 2531 2532 assert(BB && BB->getParent() && "Block not embedded in function!"); 2533 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 2534 2535 // Remove basic blocks that have no predecessors (except the entry block)... 2536 // or that just have themself as a predecessor. These are unreachable. 2537 if ((pred_begin(BB) == pred_end(BB) && 2538 BB != &BB->getParent()->getEntryBlock()) || 2539 BB->getSinglePredecessor() == BB) { 2540 DEBUG(dbgs() << "Removing BB: \n" << *BB); 2541 DeleteDeadBlock(BB); 2542 return true; 2543 } 2544 2545 // Check to see if we can constant propagate this terminator instruction 2546 // away... 2547 Changed |= ConstantFoldTerminator(BB); 2548 2549 // Check for and eliminate duplicate PHI nodes in this block. 2550 Changed |= EliminateDuplicatePHINodes(BB); 2551 2552 // Merge basic blocks into their predecessor if there is only one distinct 2553 // pred, and if there is only one distinct successor of the predecessor, and 2554 // if there are no PHI nodes. 2555 // 2556 if (MergeBlockIntoPredecessor(BB)) 2557 return true; 2558 2559 // If there is a trivial two-entry PHI node in this basic block, and we can 2560 // eliminate it, do so now. 2561 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 2562 if (PN->getNumIncomingValues() == 2) 2563 Changed |= FoldTwoEntryPHINode(PN, TD); 2564 2565 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 2566 if (BI->isUnconditional()) { 2567 if (SimplifyUncondBranch(BI)) return true; 2568 } else { 2569 if (SimplifyCondBranch(BI)) return true; 2570 } 2571 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 2572 if (SimplifyReturn(RI)) return true; 2573 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2574 if (SimplifySwitch(SI)) return true; 2575 } else if (UnreachableInst *UI = 2576 dyn_cast<UnreachableInst>(BB->getTerminator())) { 2577 if (SimplifyUnreachable(UI)) return true; 2578 } else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 2579 if (SimplifyUnwind(UI)) return true; 2580 } else if (IndirectBrInst *IBI = 2581 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 2582 if (SimplifyIndirectBr(IBI)) return true; 2583 } 2584 2585 return Changed; 2586 } 2587 2588 /// SimplifyCFG - This function is used to do simplification of a CFG. For 2589 /// example, it adjusts branches to branches to eliminate the extra hop, it 2590 /// eliminates unreachable basic blocks, and does other "peephole" optimization 2591 /// of the CFG. It returns true if a modification was made. 2592 /// 2593 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 2594 return SimplifyCFGOpt(TD).run(BB); 2595 } 2596