SimplifyCFG.cpp revision ddb97a2bf1f06d7d1b0a1af8e74c67917b43b606
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/Support/CFG.h" 23#include "llvm/Support/Debug.h" 24#include "llvm/Support/raw_ostream.h" 25#include "llvm/Analysis/ConstantFolding.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Transforms/Utils/BasicBlockUtils.h" 28#include "llvm/ADT/DenseMap.h" 29#include "llvm/ADT/SmallVector.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33#include <algorithm> 34#include <set> 35#include <map> 36using namespace llvm; 37 38STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 39 40namespace { 41class SimplifyCFGOpt { 42 const TargetData *const TD; 43 44 Value *isValueEqualityComparison(TerminatorInst *TI); 45 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 46 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases); 47 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 48 BasicBlock *Pred); 49 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI); 50 51public: 52 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} 53 bool run(BasicBlock *BB); 54}; 55} 56 57/// SafeToMergeTerminators - Return true if it is safe to merge these two 58/// terminator instructions together. 59/// 60static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 61 if (SI1 == SI2) return false; // Can't merge with self! 62 63 // It is not safe to merge these two switch instructions if they have a common 64 // successor, and if that successor has a PHI node, and if *that* PHI node has 65 // conflicting incoming values from the two switch blocks. 66 BasicBlock *SI1BB = SI1->getParent(); 67 BasicBlock *SI2BB = SI2->getParent(); 68 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 69 70 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 71 if (SI1Succs.count(*I)) 72 for (BasicBlock::iterator BBI = (*I)->begin(); 73 isa<PHINode>(BBI); ++BBI) { 74 PHINode *PN = cast<PHINode>(BBI); 75 if (PN->getIncomingValueForBlock(SI1BB) != 76 PN->getIncomingValueForBlock(SI2BB)) 77 return false; 78 } 79 80 return true; 81} 82 83/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 84/// now be entries in it from the 'NewPred' block. The values that will be 85/// flowing into the PHI nodes will be the same as those coming in from 86/// ExistPred, an existing predecessor of Succ. 87static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 88 BasicBlock *ExistPred) { 89 assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != 90 succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); 91 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 92 93 PHINode *PN; 94 for (BasicBlock::iterator I = Succ->begin(); 95 (PN = dyn_cast<PHINode>(I)); ++I) 96 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 97} 98 99 100/// GetIfCondition - Given a basic block (BB) with two predecessors (and 101/// presumably PHI nodes in it), check to see if the merge at this block is due 102/// to an "if condition". If so, return the boolean condition that determines 103/// which entry into BB will be taken. Also, return by references the block 104/// that will be entered from if the condition is true, and the block that will 105/// be entered if the condition is false. 106/// 107/// 108static Value *GetIfCondition(BasicBlock *BB, 109 BasicBlock *&IfTrue, BasicBlock *&IfFalse) { 110 assert(std::distance(pred_begin(BB), pred_end(BB)) == 2 && 111 "Function can only handle blocks with 2 predecessors!"); 112 BasicBlock *Pred1 = *pred_begin(BB); 113 BasicBlock *Pred2 = *++pred_begin(BB); 114 115 // We can only handle branches. Other control flow will be lowered to 116 // branches if possible anyway. 117 if (!isa<BranchInst>(Pred1->getTerminator()) || 118 !isa<BranchInst>(Pred2->getTerminator())) 119 return 0; 120 BranchInst *Pred1Br = cast<BranchInst>(Pred1->getTerminator()); 121 BranchInst *Pred2Br = cast<BranchInst>(Pred2->getTerminator()); 122 123 // Eliminate code duplication by ensuring that Pred1Br is conditional if 124 // either are. 125 if (Pred2Br->isConditional()) { 126 // If both branches are conditional, we don't have an "if statement". In 127 // reality, we could transform this case, but since the condition will be 128 // required anyway, we stand no chance of eliminating it, so the xform is 129 // probably not profitable. 130 if (Pred1Br->isConditional()) 131 return 0; 132 133 std::swap(Pred1, Pred2); 134 std::swap(Pred1Br, Pred2Br); 135 } 136 137 if (Pred1Br->isConditional()) { 138 // If we found a conditional branch predecessor, make sure that it branches 139 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 140 if (Pred1Br->getSuccessor(0) == BB && 141 Pred1Br->getSuccessor(1) == Pred2) { 142 IfTrue = Pred1; 143 IfFalse = Pred2; 144 } else if (Pred1Br->getSuccessor(0) == Pred2 && 145 Pred1Br->getSuccessor(1) == BB) { 146 IfTrue = Pred2; 147 IfFalse = Pred1; 148 } else { 149 // We know that one arm of the conditional goes to BB, so the other must 150 // go somewhere unrelated, and this must not be an "if statement". 151 return 0; 152 } 153 154 // The only thing we have to watch out for here is to make sure that Pred2 155 // doesn't have incoming edges from other blocks. If it does, the condition 156 // doesn't dominate BB. 157 if (++pred_begin(Pred2) != pred_end(Pred2)) 158 return 0; 159 160 return Pred1Br->getCondition(); 161 } 162 163 // Ok, if we got here, both predecessors end with an unconditional branch to 164 // BB. Don't panic! If both blocks only have a single (identical) 165 // predecessor, and THAT is a conditional branch, then we're all ok! 166 if (pred_begin(Pred1) == pred_end(Pred1) || 167 ++pred_begin(Pred1) != pred_end(Pred1) || 168 pred_begin(Pred2) == pred_end(Pred2) || 169 ++pred_begin(Pred2) != pred_end(Pred2) || 170 *pred_begin(Pred1) != *pred_begin(Pred2)) 171 return 0; 172 173 // Otherwise, if this is a conditional branch, then we can use it! 174 BasicBlock *CommonPred = *pred_begin(Pred1); 175 if (BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator())) { 176 assert(BI->isConditional() && "Two successors but not conditional?"); 177 if (BI->getSuccessor(0) == Pred1) { 178 IfTrue = Pred1; 179 IfFalse = Pred2; 180 } else { 181 IfTrue = Pred2; 182 IfFalse = Pred1; 183 } 184 return BI->getCondition(); 185 } 186 return 0; 187} 188 189/// DominatesMergePoint - If we have a merge point of an "if condition" as 190/// accepted above, return true if the specified value dominates the block. We 191/// don't handle the true generality of domination here, just a special case 192/// which works well enough for us. 193/// 194/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 195/// see if V (which must be an instruction) is cheap to compute and is 196/// non-trapping. If both are true, the instruction is inserted into the set 197/// and true is returned. 198static bool DominatesMergePoint(Value *V, BasicBlock *BB, 199 std::set<Instruction*> *AggressiveInsts) { 200 Instruction *I = dyn_cast<Instruction>(V); 201 if (!I) { 202 // Non-instructions all dominate instructions, but not all constantexprs 203 // can be executed unconditionally. 204 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 205 if (C->canTrap()) 206 return false; 207 return true; 208 } 209 BasicBlock *PBB = I->getParent(); 210 211 // We don't want to allow weird loops that might have the "if condition" in 212 // the bottom of this block. 213 if (PBB == BB) return false; 214 215 // If this instruction is defined in a block that contains an unconditional 216 // branch to BB, then it must be in the 'conditional' part of the "if 217 // statement". 218 if (BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator())) 219 if (BI->isUnconditional() && BI->getSuccessor(0) == BB) { 220 if (!AggressiveInsts) return false; 221 // Okay, it looks like the instruction IS in the "condition". Check to 222 // see if it's a cheap instruction to unconditionally compute, and if it 223 // only uses stuff defined outside of the condition. If so, hoist it out. 224 if (!I->isSafeToSpeculativelyExecute()) 225 return false; 226 227 switch (I->getOpcode()) { 228 default: return false; // Cannot hoist this out safely. 229 case Instruction::Load: { 230 // We have to check to make sure there are no instructions before the 231 // load in its basic block, as we are going to hoist the loop out to 232 // its predecessor. 233 BasicBlock::iterator IP = PBB->begin(); 234 while (isa<DbgInfoIntrinsic>(IP)) 235 IP++; 236 if (IP != BasicBlock::iterator(I)) 237 return false; 238 break; 239 } 240 case Instruction::Add: 241 case Instruction::Sub: 242 case Instruction::And: 243 case Instruction::Or: 244 case Instruction::Xor: 245 case Instruction::Shl: 246 case Instruction::LShr: 247 case Instruction::AShr: 248 case Instruction::ICmp: 249 break; // These are all cheap and non-trapping instructions. 250 } 251 252 // Okay, we can only really hoist these out if their operands are not 253 // defined in the conditional region. 254 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 255 if (!DominatesMergePoint(*i, BB, 0)) 256 return false; 257 // Okay, it's safe to do this! Remember this instruction. 258 AggressiveInsts->insert(I); 259 } 260 261 return true; 262} 263 264/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 265/// and PointerNullValue. Return NULL if value is not a constant int. 266static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { 267 // Normal constant int. 268 ConstantInt *CI = dyn_cast<ConstantInt>(V); 269 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 270 return CI; 271 272 // This is some kind of pointer constant. Turn it into a pointer-sized 273 // ConstantInt if possible. 274 const IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); 275 276 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 277 if (isa<ConstantPointerNull>(V)) 278 return ConstantInt::get(PtrTy, 0); 279 280 // IntToPtr const int. 281 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 282 if (CE->getOpcode() == Instruction::IntToPtr) 283 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 284 // The constant is very likely to have the right type already. 285 if (CI->getType() == PtrTy) 286 return CI; 287 else 288 return cast<ConstantInt> 289 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 290 } 291 return 0; 292} 293 294/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 295/// collection of icmp eq/ne instructions that compare a value against a 296/// constant, return the value being compared, and stick the constant into the 297/// Values vector. 298static Value * 299GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 300 const TargetData *TD, bool isEQ) { 301 Instruction *I = dyn_cast<Instruction>(V); 302 if (I == 0) return 0; 303 304 // If this is an icmp against a constant, handle this as one of the cases. 305 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 306 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) 307 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 308 Vals.push_back(C); 309 return I->getOperand(0); 310 } 311 return 0; 312 } 313 314 // Otherwise, we can only handle an | or &, depending on isEQ. 315 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 316 return 0; 317 318 unsigned NumValsBeforeLHS = Vals.size(); 319 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 320 isEQ)) { 321 unsigned NumVals = Vals.size(); 322 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 323 isEQ)) { 324 if (LHS == RHS) 325 return LHS; 326 } 327 Vals.resize(NumVals); 328 329 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 330 // set it and return success. 331 if (Extra == 0 || Extra == I->getOperand(1)) { 332 Extra = I->getOperand(1); 333 return LHS; 334 } 335 336 Vals.resize(NumValsBeforeLHS); 337 return 0; 338 } 339 340 // If the LHS can't be folded in, but Extra is available and RHS can, try to 341 // use LHS as Extra. 342 if (Extra == 0 || Extra == I->getOperand(0)) { 343 Extra = I->getOperand(0); 344 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 345 isEQ)) 346 return RHS; 347 Vals.resize(NumValsBeforeLHS); 348 Extra = 0; 349 } 350 351 return 0; 352} 353 354static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 355 Instruction* Cond = 0; 356 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 357 Cond = dyn_cast<Instruction>(SI->getCondition()); 358 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 359 if (BI->isConditional()) 360 Cond = dyn_cast<Instruction>(BI->getCondition()); 361 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 362 Cond = dyn_cast<Instruction>(IBI->getAddress()); 363 } 364 365 TI->eraseFromParent(); 366 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 367} 368 369/// isValueEqualityComparison - Return true if the specified terminator checks 370/// to see if a value is equal to constant integer value. 371Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 372 Value *CV = 0; 373 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 374 // Do not permit merging of large switch instructions into their 375 // predecessors unless there is only one predecessor. 376 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 377 pred_end(SI->getParent())) <= 128) 378 CV = SI->getCondition(); 379 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 380 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 381 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 382 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 383 ICI->getPredicate() == ICmpInst::ICMP_NE) && 384 GetConstantInt(ICI->getOperand(1), TD)) 385 CV = ICI->getOperand(0); 386 387 // Unwrap any lossless ptrtoint cast. 388 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 389 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 390 CV = PTII->getOperand(0); 391 return CV; 392} 393 394/// GetValueEqualityComparisonCases - Given a value comparison instruction, 395/// decode all of the 'cases' that it represents and return the 'default' block. 396BasicBlock *SimplifyCFGOpt:: 397GetValueEqualityComparisonCases(TerminatorInst *TI, 398 std::vector<std::pair<ConstantInt*, 399 BasicBlock*> > &Cases) { 400 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 401 Cases.reserve(SI->getNumCases()); 402 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 403 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i))); 404 return SI->getDefaultDest(); 405 } 406 407 BranchInst *BI = cast<BranchInst>(TI); 408 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 409 Cases.push_back(std::make_pair(GetConstantInt(ICI->getOperand(1), TD), 410 BI->getSuccessor(ICI->getPredicate() == 411 ICmpInst::ICMP_NE))); 412 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 413} 414 415 416/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 417/// in the list that match the specified block. 418static void EliminateBlockCases(BasicBlock *BB, 419 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 420 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 421 if (Cases[i].second == BB) { 422 Cases.erase(Cases.begin()+i); 423 --i; --e; 424 } 425} 426 427/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 428/// well. 429static bool 430ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 431 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 432 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 433 434 // Make V1 be smaller than V2. 435 if (V1->size() > V2->size()) 436 std::swap(V1, V2); 437 438 if (V1->size() == 0) return false; 439 if (V1->size() == 1) { 440 // Just scan V2. 441 ConstantInt *TheVal = (*V1)[0].first; 442 for (unsigned i = 0, e = V2->size(); i != e; ++i) 443 if (TheVal == (*V2)[i].first) 444 return true; 445 } 446 447 // Otherwise, just sort both lists and compare element by element. 448 array_pod_sort(V1->begin(), V1->end()); 449 array_pod_sort(V2->begin(), V2->end()); 450 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 451 while (i1 != e1 && i2 != e2) { 452 if ((*V1)[i1].first == (*V2)[i2].first) 453 return true; 454 if ((*V1)[i1].first < (*V2)[i2].first) 455 ++i1; 456 else 457 ++i2; 458 } 459 return false; 460} 461 462/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 463/// terminator instruction and its block is known to only have a single 464/// predecessor block, check to see if that predecessor is also a value 465/// comparison with the same value, and if that comparison determines the 466/// outcome of this comparison. If so, simplify TI. This does a very limited 467/// form of jump threading. 468bool SimplifyCFGOpt:: 469SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 470 BasicBlock *Pred) { 471 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 472 if (!PredVal) return false; // Not a value comparison in predecessor. 473 474 Value *ThisVal = isValueEqualityComparison(TI); 475 assert(ThisVal && "This isn't a value comparison!!"); 476 if (ThisVal != PredVal) return false; // Different predicates. 477 478 // Find out information about when control will move from Pred to TI's block. 479 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 480 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 481 PredCases); 482 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 483 484 // Find information about how control leaves this block. 485 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 486 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 487 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 488 489 // If TI's block is the default block from Pred's comparison, potentially 490 // simplify TI based on this knowledge. 491 if (PredDef == TI->getParent()) { 492 // If we are here, we know that the value is none of those cases listed in 493 // PredCases. If there are any cases in ThisCases that are in PredCases, we 494 // can simplify TI. 495 if (!ValuesOverlap(PredCases, ThisCases)) 496 return false; 497 498 if (isa<BranchInst>(TI)) { 499 // Okay, one of the successors of this condbr is dead. Convert it to a 500 // uncond br. 501 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 502 // Insert the new branch. 503 Instruction *NI = BranchInst::Create(ThisDef, TI); 504 (void) NI; 505 506 // Remove PHI node entries for the dead edge. 507 ThisCases[0].second->removePredecessor(TI->getParent()); 508 509 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 510 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 511 512 EraseTerminatorInstAndDCECond(TI); 513 return true; 514 } 515 516 SwitchInst *SI = cast<SwitchInst>(TI); 517 // Okay, TI has cases that are statically dead, prune them away. 518 SmallPtrSet<Constant*, 16> DeadCases; 519 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 520 DeadCases.insert(PredCases[i].first); 521 522 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 523 << "Through successor TI: " << *TI); 524 525 for (unsigned i = SI->getNumCases()-1; i != 0; --i) 526 if (DeadCases.count(SI->getCaseValue(i))) { 527 SI->getSuccessor(i)->removePredecessor(TI->getParent()); 528 SI->removeCase(i); 529 } 530 531 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 532 return true; 533 } 534 535 // Otherwise, TI's block must correspond to some matched value. Find out 536 // which value (or set of values) this is. 537 ConstantInt *TIV = 0; 538 BasicBlock *TIBB = TI->getParent(); 539 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 540 if (PredCases[i].second == TIBB) { 541 if (TIV != 0) 542 return false; // Cannot handle multiple values coming to this block. 543 TIV = PredCases[i].first; 544 } 545 assert(TIV && "No edge from pred to succ?"); 546 547 // Okay, we found the one constant that our value can be if we get into TI's 548 // BB. Find out which successor will unconditionally be branched to. 549 BasicBlock *TheRealDest = 0; 550 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 551 if (ThisCases[i].first == TIV) { 552 TheRealDest = ThisCases[i].second; 553 break; 554 } 555 556 // If not handled by any explicit cases, it is handled by the default case. 557 if (TheRealDest == 0) TheRealDest = ThisDef; 558 559 // Remove PHI node entries for dead edges. 560 BasicBlock *CheckEdge = TheRealDest; 561 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 562 if (*SI != CheckEdge) 563 (*SI)->removePredecessor(TIBB); 564 else 565 CheckEdge = 0; 566 567 // Insert the new branch. 568 Instruction *NI = BranchInst::Create(TheRealDest, TI); 569 (void) NI; 570 571 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 572 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 573 574 EraseTerminatorInstAndDCECond(TI); 575 return true; 576} 577 578namespace { 579 /// ConstantIntOrdering - This class implements a stable ordering of constant 580 /// integers that does not depend on their address. This is important for 581 /// applications that sort ConstantInt's to ensure uniqueness. 582 struct ConstantIntOrdering { 583 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 584 return LHS->getValue().ult(RHS->getValue()); 585 } 586 }; 587} 588 589static int ConstantIntSortPredicate(const void *P1, const void *P2) { 590 const ConstantInt *LHS = *(const ConstantInt**)P1; 591 const ConstantInt *RHS = *(const ConstantInt**)P2; 592 return LHS->getValue().ult(RHS->getValue()); 593} 594 595/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 596/// equality comparison instruction (either a switch or a branch on "X == c"). 597/// See if any of the predecessors of the terminator block are value comparisons 598/// on the same value. If so, and if safe to do so, fold them together. 599bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { 600 BasicBlock *BB = TI->getParent(); 601 Value *CV = isValueEqualityComparison(TI); // CondVal 602 assert(CV && "Not a comparison?"); 603 bool Changed = false; 604 605 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 606 while (!Preds.empty()) { 607 BasicBlock *Pred = Preds.pop_back_val(); 608 609 // See if the predecessor is a comparison with the same value. 610 TerminatorInst *PTI = Pred->getTerminator(); 611 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 612 613 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 614 // Figure out which 'cases' to copy from SI to PSI. 615 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 616 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 617 618 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 619 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 620 621 // Based on whether the default edge from PTI goes to BB or not, fill in 622 // PredCases and PredDefault with the new switch cases we would like to 623 // build. 624 SmallVector<BasicBlock*, 8> NewSuccessors; 625 626 if (PredDefault == BB) { 627 // If this is the default destination from PTI, only the edges in TI 628 // that don't occur in PTI, or that branch to BB will be activated. 629 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 630 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 631 if (PredCases[i].second != BB) 632 PTIHandled.insert(PredCases[i].first); 633 else { 634 // The default destination is BB, we don't need explicit targets. 635 std::swap(PredCases[i], PredCases.back()); 636 PredCases.pop_back(); 637 --i; --e; 638 } 639 640 // Reconstruct the new switch statement we will be building. 641 if (PredDefault != BBDefault) { 642 PredDefault->removePredecessor(Pred); 643 PredDefault = BBDefault; 644 NewSuccessors.push_back(BBDefault); 645 } 646 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 647 if (!PTIHandled.count(BBCases[i].first) && 648 BBCases[i].second != BBDefault) { 649 PredCases.push_back(BBCases[i]); 650 NewSuccessors.push_back(BBCases[i].second); 651 } 652 653 } else { 654 // If this is not the default destination from PSI, only the edges 655 // in SI that occur in PSI with a destination of BB will be 656 // activated. 657 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 658 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 659 if (PredCases[i].second == BB) { 660 PTIHandled.insert(PredCases[i].first); 661 std::swap(PredCases[i], PredCases.back()); 662 PredCases.pop_back(); 663 --i; --e; 664 } 665 666 // Okay, now we know which constants were sent to BB from the 667 // predecessor. Figure out where they will all go now. 668 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 669 if (PTIHandled.count(BBCases[i].first)) { 670 // If this is one we are capable of getting... 671 PredCases.push_back(BBCases[i]); 672 NewSuccessors.push_back(BBCases[i].second); 673 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 674 } 675 676 // If there are any constants vectored to BB that TI doesn't handle, 677 // they must go to the default destination of TI. 678 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 679 PTIHandled.begin(), 680 E = PTIHandled.end(); I != E; ++I) { 681 PredCases.push_back(std::make_pair(*I, BBDefault)); 682 NewSuccessors.push_back(BBDefault); 683 } 684 } 685 686 // Okay, at this point, we know which new successor Pred will get. Make 687 // sure we update the number of entries in the PHI nodes for these 688 // successors. 689 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 690 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 691 692 // Convert pointer to int before we switch. 693 if (CV->getType()->isPointerTy()) { 694 assert(TD && "Cannot switch on pointer without TargetData"); 695 CV = new PtrToIntInst(CV, TD->getIntPtrType(CV->getContext()), 696 "magicptr", PTI); 697 } 698 699 // Now that the successors are updated, create the new Switch instruction. 700 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, 701 PredCases.size(), PTI); 702 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 703 NewSI->addCase(PredCases[i].first, PredCases[i].second); 704 705 EraseTerminatorInstAndDCECond(PTI); 706 707 // Okay, last check. If BB is still a successor of PSI, then we must 708 // have an infinite loop case. If so, add an infinitely looping block 709 // to handle the case to preserve the behavior of the code. 710 BasicBlock *InfLoopBlock = 0; 711 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 712 if (NewSI->getSuccessor(i) == BB) { 713 if (InfLoopBlock == 0) { 714 // Insert it at the end of the function, because it's either code, 715 // or it won't matter if it's hot. :) 716 InfLoopBlock = BasicBlock::Create(BB->getContext(), 717 "infloop", BB->getParent()); 718 BranchInst::Create(InfLoopBlock, InfLoopBlock); 719 } 720 NewSI->setSuccessor(i, InfLoopBlock); 721 } 722 723 Changed = true; 724 } 725 } 726 return Changed; 727} 728 729// isSafeToHoistInvoke - If we would need to insert a select that uses the 730// value of this invoke (comments in HoistThenElseCodeToIf explain why we 731// would need to do this), we can't hoist the invoke, as there is nowhere 732// to put the select in this case. 733static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 734 Instruction *I1, Instruction *I2) { 735 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 736 PHINode *PN; 737 for (BasicBlock::iterator BBI = SI->begin(); 738 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 739 Value *BB1V = PN->getIncomingValueForBlock(BB1); 740 Value *BB2V = PN->getIncomingValueForBlock(BB2); 741 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 742 return false; 743 } 744 } 745 } 746 return true; 747} 748 749/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 750/// BB2, hoist any common code in the two blocks up into the branch block. The 751/// caller of this function guarantees that BI's block dominates BB1 and BB2. 752static bool HoistThenElseCodeToIf(BranchInst *BI) { 753 // This does very trivial matching, with limited scanning, to find identical 754 // instructions in the two blocks. In particular, we don't want to get into 755 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 756 // such, we currently just scan for obviously identical instructions in an 757 // identical order. 758 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 759 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 760 761 BasicBlock::iterator BB1_Itr = BB1->begin(); 762 BasicBlock::iterator BB2_Itr = BB2->begin(); 763 764 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 765 while (isa<DbgInfoIntrinsic>(I1)) 766 I1 = BB1_Itr++; 767 while (isa<DbgInfoIntrinsic>(I2)) 768 I2 = BB2_Itr++; 769 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) || 770 !I1->isIdenticalToWhenDefined(I2) || 771 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 772 return false; 773 774 // If we get here, we can hoist at least one instruction. 775 BasicBlock *BIParent = BI->getParent(); 776 777 do { 778 // If we are hoisting the terminator instruction, don't move one (making a 779 // broken BB), instead clone it, and remove BI. 780 if (isa<TerminatorInst>(I1)) 781 goto HoistTerminator; 782 783 // For a normal instruction, we just move one to right before the branch, 784 // then replace all uses of the other with the first. Finally, we remove 785 // the now redundant second instruction. 786 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 787 if (!I2->use_empty()) 788 I2->replaceAllUsesWith(I1); 789 I1->intersectOptionalDataWith(I2); 790 BB2->getInstList().erase(I2); 791 792 I1 = BB1_Itr++; 793 while (isa<DbgInfoIntrinsic>(I1)) 794 I1 = BB1_Itr++; 795 I2 = BB2_Itr++; 796 while (isa<DbgInfoIntrinsic>(I2)) 797 I2 = BB2_Itr++; 798 } while (I1->getOpcode() == I2->getOpcode() && 799 I1->isIdenticalToWhenDefined(I2)); 800 801 return true; 802 803HoistTerminator: 804 // It may not be possible to hoist an invoke. 805 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 806 return true; 807 808 // Okay, it is safe to hoist the terminator. 809 Instruction *NT = I1->clone(); 810 BIParent->getInstList().insert(BI, NT); 811 if (!NT->getType()->isVoidTy()) { 812 I1->replaceAllUsesWith(NT); 813 I2->replaceAllUsesWith(NT); 814 NT->takeName(I1); 815 } 816 817 // Hoisting one of the terminators from our successor is a great thing. 818 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 819 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 820 // nodes, so we insert select instruction to compute the final result. 821 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 822 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 823 PHINode *PN; 824 for (BasicBlock::iterator BBI = SI->begin(); 825 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 826 Value *BB1V = PN->getIncomingValueForBlock(BB1); 827 Value *BB2V = PN->getIncomingValueForBlock(BB2); 828 if (BB1V == BB2V) continue; 829 830 // These values do not agree. Insert a select instruction before NT 831 // that determines the right value. 832 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 833 if (SI == 0) 834 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 835 BB1V->getName()+"."+BB2V->getName(), NT); 836 // Make the PHI node use the select for all incoming values for BB1/BB2 837 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 838 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 839 PN->setIncomingValue(i, SI); 840 } 841 } 842 843 // Update any PHI nodes in our new successors. 844 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 845 AddPredecessorToBlock(*SI, BIParent, BB1); 846 847 EraseTerminatorInstAndDCECond(BI); 848 return true; 849} 850 851/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 852/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 853/// (for now, restricted to a single instruction that's side effect free) from 854/// the BB1 into the branch block to speculatively execute it. 855static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 856 // Only speculatively execution a single instruction (not counting the 857 // terminator) for now. 858 Instruction *HInst = NULL; 859 Instruction *Term = BB1->getTerminator(); 860 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 861 BBI != BBE; ++BBI) { 862 Instruction *I = BBI; 863 // Skip debug info. 864 if (isa<DbgInfoIntrinsic>(I)) continue; 865 if (I == Term) break; 866 867 if (HInst) 868 return false; 869 HInst = I; 870 } 871 if (!HInst) 872 return false; 873 874 // Be conservative for now. FP select instruction can often be expensive. 875 Value *BrCond = BI->getCondition(); 876 if (isa<FCmpInst>(BrCond)) 877 return false; 878 879 // If BB1 is actually on the false edge of the conditional branch, remember 880 // to swap the select operands later. 881 bool Invert = false; 882 if (BB1 != BI->getSuccessor(0)) { 883 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 884 Invert = true; 885 } 886 887 // Turn 888 // BB: 889 // %t1 = icmp 890 // br i1 %t1, label %BB1, label %BB2 891 // BB1: 892 // %t3 = add %t2, c 893 // br label BB2 894 // BB2: 895 // => 896 // BB: 897 // %t1 = icmp 898 // %t4 = add %t2, c 899 // %t3 = select i1 %t1, %t2, %t3 900 switch (HInst->getOpcode()) { 901 default: return false; // Not safe / profitable to hoist. 902 case Instruction::Add: 903 case Instruction::Sub: 904 // Not worth doing for vector ops. 905 if (HInst->getType()->isVectorTy()) 906 return false; 907 break; 908 case Instruction::And: 909 case Instruction::Or: 910 case Instruction::Xor: 911 case Instruction::Shl: 912 case Instruction::LShr: 913 case Instruction::AShr: 914 // Don't mess with vector operations. 915 if (HInst->getType()->isVectorTy()) 916 return false; 917 break; // These are all cheap and non-trapping instructions. 918 } 919 920 // If the instruction is obviously dead, don't try to predicate it. 921 if (HInst->use_empty()) { 922 HInst->eraseFromParent(); 923 return true; 924 } 925 926 // Can we speculatively execute the instruction? And what is the value 927 // if the condition is false? Consider the phi uses, if the incoming value 928 // from the "if" block are all the same V, then V is the value of the 929 // select if the condition is false. 930 BasicBlock *BIParent = BI->getParent(); 931 SmallVector<PHINode*, 4> PHIUses; 932 Value *FalseV = NULL; 933 934 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 935 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end(); 936 UI != E; ++UI) { 937 // Ignore any user that is not a PHI node in BB2. These can only occur in 938 // unreachable blocks, because they would not be dominated by the instr. 939 PHINode *PN = dyn_cast<PHINode>(*UI); 940 if (!PN || PN->getParent() != BB2) 941 return false; 942 PHIUses.push_back(PN); 943 944 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 945 if (!FalseV) 946 FalseV = PHIV; 947 else if (FalseV != PHIV) 948 return false; // Inconsistent value when condition is false. 949 } 950 951 assert(FalseV && "Must have at least one user, and it must be a PHI"); 952 953 // Do not hoist the instruction if any of its operands are defined but not 954 // used in this BB. The transformation will prevent the operand from 955 // being sunk into the use block. 956 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 957 i != e; ++i) { 958 Instruction *OpI = dyn_cast<Instruction>(*i); 959 if (OpI && OpI->getParent() == BIParent && 960 !OpI->isUsedInBasicBlock(BIParent)) 961 return false; 962 } 963 964 // If we get here, we can hoist the instruction. Try to place it 965 // before the icmp instruction preceding the conditional branch. 966 BasicBlock::iterator InsertPos = BI; 967 if (InsertPos != BIParent->begin()) 968 --InsertPos; 969 // Skip debug info between condition and branch. 970 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos)) 971 --InsertPos; 972 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) { 973 SmallPtrSet<Instruction *, 4> BB1Insns; 974 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 975 BB1I != BB1E; ++BB1I) 976 BB1Insns.insert(BB1I); 977 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 978 UI != UE; ++UI) { 979 Instruction *Use = cast<Instruction>(*UI); 980 if (!BB1Insns.count(Use)) continue; 981 982 // If BrCond uses the instruction that place it just before 983 // branch instruction. 984 InsertPos = BI; 985 break; 986 } 987 } else 988 InsertPos = BI; 989 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst); 990 991 // Create a select whose true value is the speculatively executed value and 992 // false value is the previously determined FalseV. 993 SelectInst *SI; 994 if (Invert) 995 SI = SelectInst::Create(BrCond, FalseV, HInst, 996 FalseV->getName() + "." + HInst->getName(), BI); 997 else 998 SI = SelectInst::Create(BrCond, HInst, FalseV, 999 HInst->getName() + "." + FalseV->getName(), BI); 1000 1001 // Make the PHI node use the select for all incoming values for "then" and 1002 // "if" blocks. 1003 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1004 PHINode *PN = PHIUses[i]; 1005 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1006 if (PN->getIncomingBlock(j) == BB1 || PN->getIncomingBlock(j) == BIParent) 1007 PN->setIncomingValue(j, SI); 1008 } 1009 1010 ++NumSpeculations; 1011 return true; 1012} 1013 1014/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1015/// across this block. 1016static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1017 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1018 unsigned Size = 0; 1019 1020 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1021 if (isa<DbgInfoIntrinsic>(BBI)) 1022 continue; 1023 if (Size > 10) return false; // Don't clone large BB's. 1024 ++Size; 1025 1026 // We can only support instructions that do not define values that are 1027 // live outside of the current basic block. 1028 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1029 UI != E; ++UI) { 1030 Instruction *U = cast<Instruction>(*UI); 1031 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1032 } 1033 1034 // Looks ok, continue checking. 1035 } 1036 1037 return true; 1038} 1039 1040/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1041/// that is defined in the same block as the branch and if any PHI entries are 1042/// constants, thread edges corresponding to that entry to be branches to their 1043/// ultimate destination. 1044static bool FoldCondBranchOnPHI(BranchInst *BI) { 1045 BasicBlock *BB = BI->getParent(); 1046 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1047 // NOTE: we currently cannot transform this case if the PHI node is used 1048 // outside of the block. 1049 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1050 return false; 1051 1052 // Degenerate case of a single entry PHI. 1053 if (PN->getNumIncomingValues() == 1) { 1054 FoldSingleEntryPHINodes(PN->getParent()); 1055 return true; 1056 } 1057 1058 // Now we know that this block has multiple preds and two succs. 1059 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1060 1061 // Okay, this is a simple enough basic block. See if any phi values are 1062 // constants. 1063 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1064 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1065 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1066 1067 // Okay, we now know that all edges from PredBB should be revectored to 1068 // branch to RealDest. 1069 BasicBlock *PredBB = PN->getIncomingBlock(i); 1070 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1071 1072 if (RealDest == BB) continue; // Skip self loops. 1073 1074 // The dest block might have PHI nodes, other predecessors and other 1075 // difficult cases. Instead of being smart about this, just insert a new 1076 // block that jumps to the destination block, effectively splitting 1077 // the edge we are about to create. 1078 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1079 RealDest->getName()+".critedge", 1080 RealDest->getParent(), RealDest); 1081 BranchInst::Create(RealDest, EdgeBB); 1082 PHINode *PN; 1083 for (BasicBlock::iterator BBI = RealDest->begin(); 1084 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1085 Value *V = PN->getIncomingValueForBlock(BB); 1086 PN->addIncoming(V, EdgeBB); 1087 } 1088 1089 // BB may have instructions that are being threaded over. Clone these 1090 // instructions into EdgeBB. We know that there will be no uses of the 1091 // cloned instructions outside of EdgeBB. 1092 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1093 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1094 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1095 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1096 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1097 continue; 1098 } 1099 // Clone the instruction. 1100 Instruction *N = BBI->clone(); 1101 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1102 1103 // Update operands due to translation. 1104 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1105 i != e; ++i) { 1106 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1107 if (PI != TranslateMap.end()) 1108 *i = PI->second; 1109 } 1110 1111 // Check for trivial simplification. 1112 if (Constant *C = ConstantFoldInstruction(N)) { 1113 TranslateMap[BBI] = C; 1114 delete N; // Constant folded away, don't need actual inst 1115 } else { 1116 // Insert the new instruction into its new home. 1117 EdgeBB->getInstList().insert(InsertPt, N); 1118 if (!BBI->use_empty()) 1119 TranslateMap[BBI] = N; 1120 } 1121 } 1122 1123 // Loop over all of the edges from PredBB to BB, changing them to branch 1124 // to EdgeBB instead. 1125 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1126 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1127 if (PredBBTI->getSuccessor(i) == BB) { 1128 BB->removePredecessor(PredBB); 1129 PredBBTI->setSuccessor(i, EdgeBB); 1130 } 1131 1132 // Recurse, simplifying any other constants. 1133 return FoldCondBranchOnPHI(BI) | true; 1134 } 1135 1136 return false; 1137} 1138 1139/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1140/// PHI node, see if we can eliminate it. 1141static bool FoldTwoEntryPHINode(PHINode *PN) { 1142 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1143 // statement", which has a very simple dominance structure. Basically, we 1144 // are trying to find the condition that is being branched on, which 1145 // subsequently causes this merge to happen. We really want control 1146 // dependence information for this check, but simplifycfg can't keep it up 1147 // to date, and this catches most of the cases we care about anyway. 1148 // 1149 BasicBlock *BB = PN->getParent(); 1150 BasicBlock *IfTrue, *IfFalse; 1151 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1152 if (!IfCond) return false; 1153 1154 // Okay, we found that we can merge this two-entry phi node into a select. 1155 // Doing so would require us to fold *all* two entry phi nodes in this block. 1156 // At some point this becomes non-profitable (particularly if the target 1157 // doesn't support cmov's). Only do this transformation if there are two or 1158 // fewer PHI nodes in this block. 1159 unsigned NumPhis = 0; 1160 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1161 if (NumPhis > 2) 1162 return false; 1163 1164 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1165 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1166 1167 // Loop over the PHI's seeing if we can promote them all to select 1168 // instructions. While we are at it, keep track of the instructions 1169 // that need to be moved to the dominating block. 1170 std::set<Instruction*> AggressiveInsts; 1171 1172 BasicBlock::iterator AfterPHIIt = BB->begin(); 1173 while (isa<PHINode>(AfterPHIIt)) { 1174 PHINode *PN = cast<PHINode>(AfterPHIIt++); 1175 if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) { 1176 if (PN->getIncomingValue(0) != PN) 1177 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 1178 else 1179 PN->replaceAllUsesWith(UndefValue::get(PN->getType())); 1180 } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, 1181 &AggressiveInsts) || 1182 !DominatesMergePoint(PN->getIncomingValue(1), BB, 1183 &AggressiveInsts)) { 1184 return false; 1185 } 1186 } 1187 1188 // If we all PHI nodes are promotable, check to make sure that all 1189 // instructions in the predecessor blocks can be promoted as well. If 1190 // not, we won't be able to get rid of the control flow, so it's not 1191 // worth promoting to select instructions. 1192 BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; 1193 PN = cast<PHINode>(BB->begin()); 1194 BasicBlock *Pred = PN->getIncomingBlock(0); 1195 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1196 IfBlock1 = Pred; 1197 DomBlock = *pred_begin(Pred); 1198 for (BasicBlock::iterator I = Pred->begin(); 1199 !isa<TerminatorInst>(I); ++I) 1200 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1201 // This is not an aggressive instruction that we can promote. 1202 // Because of this, we won't be able to get rid of the control 1203 // flow, so the xform is not worth it. 1204 return false; 1205 } 1206 } 1207 1208 Pred = PN->getIncomingBlock(1); 1209 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1210 IfBlock2 = Pred; 1211 DomBlock = *pred_begin(Pred); 1212 for (BasicBlock::iterator I = Pred->begin(); 1213 !isa<TerminatorInst>(I); ++I) 1214 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1215 // This is not an aggressive instruction that we can promote. 1216 // Because of this, we won't be able to get rid of the control 1217 // flow, so the xform is not worth it. 1218 return false; 1219 } 1220 } 1221 1222 // If we can still promote the PHI nodes after this gauntlet of tests, 1223 // do all of the PHI's now. 1224 1225 // Move all 'aggressive' instructions, which are defined in the 1226 // conditional parts of the if's up to the dominating block. 1227 if (IfBlock1) 1228 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1229 IfBlock1->getInstList(), IfBlock1->begin(), 1230 IfBlock1->getTerminator()); 1231 if (IfBlock2) 1232 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1233 IfBlock2->getInstList(), IfBlock2->begin(), 1234 IfBlock2->getTerminator()); 1235 1236 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1237 // Change the PHI node into a select instruction. 1238 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1239 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1240 1241 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); 1242 PN->replaceAllUsesWith(NV); 1243 NV->takeName(PN); 1244 1245 BB->getInstList().erase(PN); 1246 } 1247 return true; 1248} 1249 1250/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1251/// to two returning blocks, try to merge them together into one return, 1252/// introducing a select if the return values disagree. 1253static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1254 assert(BI->isConditional() && "Must be a conditional branch"); 1255 BasicBlock *TrueSucc = BI->getSuccessor(0); 1256 BasicBlock *FalseSucc = BI->getSuccessor(1); 1257 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1258 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1259 1260 // Check to ensure both blocks are empty (just a return) or optionally empty 1261 // with PHI nodes. If there are other instructions, merging would cause extra 1262 // computation on one path or the other. 1263 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1264 return false; 1265 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1266 return false; 1267 1268 // Okay, we found a branch that is going to two return nodes. If 1269 // there is no return value for this function, just change the 1270 // branch into a return. 1271 if (FalseRet->getNumOperands() == 0) { 1272 TrueSucc->removePredecessor(BI->getParent()); 1273 FalseSucc->removePredecessor(BI->getParent()); 1274 ReturnInst::Create(BI->getContext(), 0, BI); 1275 EraseTerminatorInstAndDCECond(BI); 1276 return true; 1277 } 1278 1279 // Otherwise, figure out what the true and false return values are 1280 // so we can insert a new select instruction. 1281 Value *TrueValue = TrueRet->getReturnValue(); 1282 Value *FalseValue = FalseRet->getReturnValue(); 1283 1284 // Unwrap any PHI nodes in the return blocks. 1285 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1286 if (TVPN->getParent() == TrueSucc) 1287 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1288 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1289 if (FVPN->getParent() == FalseSucc) 1290 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1291 1292 // In order for this transformation to be safe, we must be able to 1293 // unconditionally execute both operands to the return. This is 1294 // normally the case, but we could have a potentially-trapping 1295 // constant expression that prevents this transformation from being 1296 // safe. 1297 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1298 if (TCV->canTrap()) 1299 return false; 1300 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1301 if (FCV->canTrap()) 1302 return false; 1303 1304 // Okay, we collected all the mapped values and checked them for sanity, and 1305 // defined to really do this transformation. First, update the CFG. 1306 TrueSucc->removePredecessor(BI->getParent()); 1307 FalseSucc->removePredecessor(BI->getParent()); 1308 1309 // Insert select instructions where needed. 1310 Value *BrCond = BI->getCondition(); 1311 if (TrueValue) { 1312 // Insert a select if the results differ. 1313 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1314 } else if (isa<UndefValue>(TrueValue)) { 1315 TrueValue = FalseValue; 1316 } else { 1317 TrueValue = SelectInst::Create(BrCond, TrueValue, 1318 FalseValue, "retval", BI); 1319 } 1320 } 1321 1322 Value *RI = !TrueValue ? 1323 ReturnInst::Create(BI->getContext(), BI) : 1324 ReturnInst::Create(BI->getContext(), TrueValue, BI); 1325 (void) RI; 1326 1327 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1328 << "\n " << *BI << "NewRet = " << *RI 1329 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1330 1331 EraseTerminatorInstAndDCECond(BI); 1332 1333 return true; 1334} 1335 1336/// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, 1337/// and if a predecessor branches to us and one of our successors, fold the 1338/// setcc into the predecessor and use logical operations to pick the right 1339/// destination. 1340bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1341 BasicBlock *BB = BI->getParent(); 1342 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1343 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1344 Cond->getParent() != BB || !Cond->hasOneUse()) 1345 return false; 1346 1347 // Only allow this if the condition is a simple instruction that can be 1348 // executed unconditionally. It must be in the same block as the branch, and 1349 // must be at the front of the block. 1350 BasicBlock::iterator FrontIt = BB->front(); 1351 // Ignore dbg intrinsics. 1352 while(isa<DbgInfoIntrinsic>(FrontIt)) 1353 ++FrontIt; 1354 1355 // Allow a single instruction to be hoisted in addition to the compare 1356 // that feeds the branch. We later ensure that any values that _it_ uses 1357 // were also live in the predecessor, so that we don't unnecessarily create 1358 // register pressure or inhibit out-of-order execution. 1359 Instruction *BonusInst = 0; 1360 if (&*FrontIt != Cond && 1361 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1362 FrontIt->isSafeToSpeculativelyExecute()) { 1363 BonusInst = &*FrontIt; 1364 ++FrontIt; 1365 } 1366 1367 // Only a single bonus inst is allowed. 1368 if (&*FrontIt != Cond) 1369 return false; 1370 1371 // Make sure the instruction after the condition is the cond branch. 1372 BasicBlock::iterator CondIt = Cond; ++CondIt; 1373 // Ingore dbg intrinsics. 1374 while(isa<DbgInfoIntrinsic>(CondIt)) 1375 ++CondIt; 1376 if (&*CondIt != BI) { 1377 assert (!isa<DbgInfoIntrinsic>(CondIt) && "Hey do not forget debug info!"); 1378 return false; 1379 } 1380 1381 // Cond is known to be a compare or binary operator. Check to make sure that 1382 // neither operand is a potentially-trapping constant expression. 1383 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1384 if (CE->canTrap()) 1385 return false; 1386 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1387 if (CE->canTrap()) 1388 return false; 1389 1390 1391 // Finally, don't infinitely unroll conditional loops. 1392 BasicBlock *TrueDest = BI->getSuccessor(0); 1393 BasicBlock *FalseDest = BI->getSuccessor(1); 1394 if (TrueDest == BB || FalseDest == BB) 1395 return false; 1396 1397 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1398 BasicBlock *PredBlock = *PI; 1399 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1400 1401 // Check that we have two conditional branches. If there is a PHI node in 1402 // the common successor, verify that the same value flows in from both 1403 // blocks. 1404 if (PBI == 0 || PBI->isUnconditional() || 1405 !SafeToMergeTerminators(BI, PBI)) 1406 continue; 1407 1408 // Ensure that any values used in the bonus instruction are also used 1409 // by the terminator of the predecessor. This means that those values 1410 // must already have been resolved, so we won't be inhibiting the 1411 // out-of-order core by speculating them earlier. 1412 if (BonusInst) { 1413 // Collect the values used by the bonus inst 1414 SmallPtrSet<Value*, 4> UsedValues; 1415 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1416 OE = BonusInst->op_end(); OI != OE; ++OI) { 1417 Value* V = *OI; 1418 if (!isa<Constant>(V)) 1419 UsedValues.insert(V); 1420 } 1421 1422 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1423 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1424 1425 // Walk up to four levels back up the use-def chain of the predecessor's 1426 // terminator to see if all those values were used. The choice of four 1427 // levels is arbitrary, to provide a compile-time-cost bound. 1428 while (!Worklist.empty()) { 1429 std::pair<Value*, unsigned> Pair = Worklist.back(); 1430 Worklist.pop_back(); 1431 1432 if (Pair.second >= 4) continue; 1433 UsedValues.erase(Pair.first); 1434 if (UsedValues.empty()) break; 1435 1436 if (Instruction* I = dyn_cast<Instruction>(Pair.first)) { 1437 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1438 OI != OE; ++OI) 1439 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1440 } 1441 } 1442 1443 if (!UsedValues.empty()) return false; 1444 } 1445 1446 Instruction::BinaryOps Opc; 1447 bool InvertPredCond = false; 1448 1449 if (PBI->getSuccessor(0) == TrueDest) 1450 Opc = Instruction::Or; 1451 else if (PBI->getSuccessor(1) == FalseDest) 1452 Opc = Instruction::And; 1453 else if (PBI->getSuccessor(0) == FalseDest) 1454 Opc = Instruction::And, InvertPredCond = true; 1455 else if (PBI->getSuccessor(1) == TrueDest) 1456 Opc = Instruction::Or, InvertPredCond = true; 1457 else 1458 continue; 1459 1460 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1461 1462 // If we need to invert the condition in the pred block to match, do so now. 1463 if (InvertPredCond) { 1464 Value *NewCond = 1465 BinaryOperator::CreateNot(PBI->getCondition(), 1466 PBI->getCondition()->getName()+".not", PBI); 1467 PBI->setCondition(NewCond); 1468 BasicBlock *OldTrue = PBI->getSuccessor(0); 1469 BasicBlock *OldFalse = PBI->getSuccessor(1); 1470 PBI->setSuccessor(0, OldFalse); 1471 PBI->setSuccessor(1, OldTrue); 1472 } 1473 1474 // If we have a bonus inst, clone it into the predecessor block. 1475 Instruction *NewBonus = 0; 1476 if (BonusInst) { 1477 NewBonus = BonusInst->clone(); 1478 PredBlock->getInstList().insert(PBI, NewBonus); 1479 NewBonus->takeName(BonusInst); 1480 BonusInst->setName(BonusInst->getName()+".old"); 1481 } 1482 1483 // Clone Cond into the predecessor basic block, and or/and the 1484 // two conditions together. 1485 Instruction *New = Cond->clone(); 1486 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1487 PredBlock->getInstList().insert(PBI, New); 1488 New->takeName(Cond); 1489 Cond->setName(New->getName()+".old"); 1490 1491 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1492 New, "or.cond", PBI); 1493 PBI->setCondition(NewCond); 1494 if (PBI->getSuccessor(0) == BB) { 1495 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1496 PBI->setSuccessor(0, TrueDest); 1497 } 1498 if (PBI->getSuccessor(1) == BB) { 1499 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1500 PBI->setSuccessor(1, FalseDest); 1501 } 1502 return true; 1503 } 1504 return false; 1505} 1506 1507/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1508/// predecessor of another block, this function tries to simplify it. We know 1509/// that PBI and BI are both conditional branches, and BI is in one of the 1510/// successor blocks of PBI - PBI branches to BI. 1511static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1512 assert(PBI->isConditional() && BI->isConditional()); 1513 BasicBlock *BB = BI->getParent(); 1514 1515 // If this block ends with a branch instruction, and if there is a 1516 // predecessor that ends on a branch of the same condition, make 1517 // this conditional branch redundant. 1518 if (PBI->getCondition() == BI->getCondition() && 1519 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1520 // Okay, the outcome of this conditional branch is statically 1521 // knowable. If this block had a single pred, handle specially. 1522 if (BB->getSinglePredecessor()) { 1523 // Turn this into a branch on constant. 1524 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1525 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1526 CondIsTrue)); 1527 return true; // Nuke the branch on constant. 1528 } 1529 1530 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1531 // in the constant and simplify the block result. Subsequent passes of 1532 // simplifycfg will thread the block. 1533 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1534 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 1535 BI->getCondition()->getName() + ".pr", 1536 BB->begin()); 1537 // Okay, we're going to insert the PHI node. Since PBI is not the only 1538 // predecessor, compute the PHI'd conditional value for all of the preds. 1539 // Any predecessor where the condition is not computable we keep symbolic. 1540 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1541 BasicBlock *P = *PI; 1542 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 1543 PBI != BI && PBI->isConditional() && 1544 PBI->getCondition() == BI->getCondition() && 1545 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1546 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1547 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1548 CondIsTrue), P); 1549 } else { 1550 NewPN->addIncoming(BI->getCondition(), P); 1551 } 1552 } 1553 1554 BI->setCondition(NewPN); 1555 return true; 1556 } 1557 } 1558 1559 // If this is a conditional branch in an empty block, and if any 1560 // predecessors is a conditional branch to one of our destinations, 1561 // fold the conditions into logical ops and one cond br. 1562 BasicBlock::iterator BBI = BB->begin(); 1563 // Ignore dbg intrinsics. 1564 while (isa<DbgInfoIntrinsic>(BBI)) 1565 ++BBI; 1566 if (&*BBI != BI) 1567 return false; 1568 1569 1570 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1571 if (CE->canTrap()) 1572 return false; 1573 1574 int PBIOp, BIOp; 1575 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1576 PBIOp = BIOp = 0; 1577 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1578 PBIOp = 0, BIOp = 1; 1579 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1580 PBIOp = 1, BIOp = 0; 1581 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1582 PBIOp = BIOp = 1; 1583 else 1584 return false; 1585 1586 // Check to make sure that the other destination of this branch 1587 // isn't BB itself. If so, this is an infinite loop that will 1588 // keep getting unwound. 1589 if (PBI->getSuccessor(PBIOp) == BB) 1590 return false; 1591 1592 // Do not perform this transformation if it would require 1593 // insertion of a large number of select instructions. For targets 1594 // without predication/cmovs, this is a big pessimization. 1595 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1596 1597 unsigned NumPhis = 0; 1598 for (BasicBlock::iterator II = CommonDest->begin(); 1599 isa<PHINode>(II); ++II, ++NumPhis) 1600 if (NumPhis > 2) // Disable this xform. 1601 return false; 1602 1603 // Finally, if everything is ok, fold the branches to logical ops. 1604 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1605 1606 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 1607 << "AND: " << *BI->getParent()); 1608 1609 1610 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1611 // branch in it, where one edge (OtherDest) goes back to itself but the other 1612 // exits. We don't *know* that the program avoids the infinite loop 1613 // (even though that seems likely). If we do this xform naively, we'll end up 1614 // recursively unpeeling the loop. Since we know that (after the xform is 1615 // done) that the block *is* infinite if reached, we just make it an obviously 1616 // infinite loop with no cond branch. 1617 if (OtherDest == BB) { 1618 // Insert it at the end of the function, because it's either code, 1619 // or it won't matter if it's hot. :) 1620 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 1621 "infloop", BB->getParent()); 1622 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1623 OtherDest = InfLoopBlock; 1624 } 1625 1626 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1627 1628 // BI may have other predecessors. Because of this, we leave 1629 // it alone, but modify PBI. 1630 1631 // Make sure we get to CommonDest on True&True directions. 1632 Value *PBICond = PBI->getCondition(); 1633 if (PBIOp) 1634 PBICond = BinaryOperator::CreateNot(PBICond, 1635 PBICond->getName()+".not", 1636 PBI); 1637 Value *BICond = BI->getCondition(); 1638 if (BIOp) 1639 BICond = BinaryOperator::CreateNot(BICond, 1640 BICond->getName()+".not", 1641 PBI); 1642 // Merge the conditions. 1643 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1644 1645 // Modify PBI to branch on the new condition to the new dests. 1646 PBI->setCondition(Cond); 1647 PBI->setSuccessor(0, CommonDest); 1648 PBI->setSuccessor(1, OtherDest); 1649 1650 // OtherDest may have phi nodes. If so, add an entry from PBI's 1651 // block that are identical to the entries for BI's block. 1652 PHINode *PN; 1653 for (BasicBlock::iterator II = OtherDest->begin(); 1654 (PN = dyn_cast<PHINode>(II)); ++II) { 1655 Value *V = PN->getIncomingValueForBlock(BB); 1656 PN->addIncoming(V, PBI->getParent()); 1657 } 1658 1659 // We know that the CommonDest already had an edge from PBI to 1660 // it. If it has PHIs though, the PHIs may have different 1661 // entries for BB and PBI's BB. If so, insert a select to make 1662 // them agree. 1663 for (BasicBlock::iterator II = CommonDest->begin(); 1664 (PN = dyn_cast<PHINode>(II)); ++II) { 1665 Value *BIV = PN->getIncomingValueForBlock(BB); 1666 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1667 Value *PBIV = PN->getIncomingValue(PBBIdx); 1668 if (BIV != PBIV) { 1669 // Insert a select in PBI to pick the right value. 1670 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1671 PBIV->getName()+".mux", PBI); 1672 PN->setIncomingValue(PBBIdx, NV); 1673 } 1674 } 1675 1676 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 1677 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1678 1679 // This basic block is probably dead. We know it has at least 1680 // one fewer predecessor. 1681 return true; 1682} 1683 1684// SimplifyIndirectBrOnSelect - Replaces 1685// (indirectbr (select cond, blockaddress(@fn, BlockA), 1686// blockaddress(@fn, BlockB))) 1687// with 1688// (br cond, BlockA, BlockB). 1689static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 1690 // Check that both operands of the select are block addresses. 1691 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 1692 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 1693 if (!TBA || !FBA) 1694 return false; 1695 1696 // Extract the actual blocks. 1697 BasicBlock *TrueBB = TBA->getBasicBlock(); 1698 BasicBlock *FalseBB = FBA->getBasicBlock(); 1699 1700 // Remove any superfluous successor edges from the CFG. 1701 // First, figure out which successors to preserve. 1702 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 1703 // successor. 1704 BasicBlock *KeepEdge1 = TrueBB; 1705 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 1706 1707 // Then remove the rest. 1708 for (unsigned I = 0, E = IBI->getNumSuccessors(); I != E; ++I) { 1709 BasicBlock *Succ = IBI->getSuccessor(I); 1710 // Make sure only to keep exactly one copy of each edge. 1711 if (Succ == KeepEdge1) 1712 KeepEdge1 = 0; 1713 else if (Succ == KeepEdge2) 1714 KeepEdge2 = 0; 1715 else 1716 Succ->removePredecessor(IBI->getParent()); 1717 } 1718 1719 // Insert an appropriate new terminator. 1720 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 1721 if (TrueBB == FalseBB) 1722 // We were only looking for one successor, and it was present. 1723 // Create an unconditional branch to it. 1724 BranchInst::Create(TrueBB, IBI); 1725 else 1726 // We found both of the successors we were looking for. 1727 // Create a conditional branch sharing the condition of the select. 1728 BranchInst::Create(TrueBB, FalseBB, SI->getCondition(), IBI); 1729 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 1730 // Neither of the selected blocks were successors, so this 1731 // indirectbr must be unreachable. 1732 new UnreachableInst(IBI->getContext(), IBI); 1733 } else { 1734 // One of the selected values was a successor, but the other wasn't. 1735 // Insert an unconditional branch to the one that was found; 1736 // the edge to the one that wasn't must be unreachable. 1737 if (KeepEdge1 == 0) 1738 // Only TrueBB was found. 1739 BranchInst::Create(TrueBB, IBI); 1740 else 1741 // Only FalseBB was found. 1742 BranchInst::Create(FalseBB, IBI); 1743 } 1744 1745 EraseTerminatorInstAndDCECond(IBI); 1746 return true; 1747} 1748 1749/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 1750/// instruction (a seteq/setne with a constant) as the only instruction in a 1751/// block that ends with an uncond branch. We are looking for a very specific 1752/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 1753/// this case, we merge the first two "or's of icmp" into a switch, but then the 1754/// default value goes to an uncond block with a seteq in it, we get something 1755/// like: 1756/// 1757/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 1758/// DEFAULT: 1759/// %tmp = icmp eq i8 %A, 92 1760/// br label %end 1761/// end: 1762/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 1763/// 1764/// We prefer to split the edge to 'end' so that there is a true/false entry to 1765/// the PHI, merging the third icmp into the switch. 1766static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI) { 1767 BasicBlock *BB = ICI->getParent(); 1768 // If the block has any PHIs in it or the icmp has multiple uses, it is too 1769 // complex. 1770 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 1771 1772 Value *V = ICI->getOperand(0); 1773 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 1774 1775 // The pattern we're looking for is where our only predecessor is a switch on 1776 // 'V' and this block is the default case for the switch. In this case we can 1777 // fold the compared value into the switch to simplify things. 1778 BasicBlock *Pred = BB->getSinglePredecessor(); 1779 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 1780 1781 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 1782 if (SI->getCondition() != V) 1783 return false; 1784 1785 // If BB is reachable on a non-default case, then we simply know the value of 1786 // V in this block. Substitute it and constant fold the icmp instruction 1787 // away. 1788 if (SI->getDefaultDest() != BB) { 1789 ConstantInt *VVal = SI->findCaseDest(BB); 1790 assert(VVal && "Should have a unique destination value"); 1791 ICI->setOperand(0, VVal); 1792 1793 if (Constant *C = ConstantFoldInstruction(ICI)) { 1794 ICI->replaceAllUsesWith(C); 1795 ICI->eraseFromParent(); 1796 } 1797 // BB is now empty, so it is likely to simplify away. 1798 return SimplifyCFG(BB) | true; 1799 } 1800 1801 // Ok, the block is reachable from the default dest. If the constant we're 1802 // comparing exists in one of the other edges, then we can constant fold ICI 1803 // and zap it. 1804 if (SI->findCaseValue(Cst) != 0) { 1805 Value *V; 1806 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1807 V = ConstantInt::getFalse(BB->getContext()); 1808 else 1809 V = ConstantInt::getTrue(BB->getContext()); 1810 1811 ICI->replaceAllUsesWith(V); 1812 ICI->eraseFromParent(); 1813 // BB is now empty, so it is likely to simplify away. 1814 return SimplifyCFG(BB) | true; 1815 } 1816 1817 // The use of the icmp has to be in the 'end' block, by the only PHI node in 1818 // the block. 1819 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 1820 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 1821 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 1822 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 1823 return false; 1824 1825 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 1826 // true in the PHI. 1827 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 1828 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 1829 1830 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 1831 std::swap(DefaultCst, NewCst); 1832 1833 // Replace ICI (which is used by the PHI for the default value) with true or 1834 // false depending on if it is EQ or NE. 1835 ICI->replaceAllUsesWith(DefaultCst); 1836 ICI->eraseFromParent(); 1837 1838 // Okay, the switch goes to this block on a default value. Add an edge from 1839 // the switch to the merge point on the compared value. 1840 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 1841 BB->getParent(), BB); 1842 SI->addCase(Cst, NewBB); 1843 1844 // NewBB branches to the phi block, add the uncond branch and the phi entry. 1845 BranchInst::Create(SuccBlock, NewBB); 1846 PHIUse->addIncoming(NewCst, NewBB); 1847 return true; 1848} 1849 1850/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 1851/// Check to see if it is branching on an or/and chain of icmp instructions, and 1852/// fold it into a switch instruction if so. 1853static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD) { 1854 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1855 if (Cond == 0) return false; 1856 1857 1858 // Change br (X == 0 | X == 1), T, F into a switch instruction. 1859 // If this is a bunch of seteq's or'd together, or if it's a bunch of 1860 // 'setne's and'ed together, collect them. 1861 Value *CompVal = 0; 1862 std::vector<ConstantInt*> Values; 1863 bool TrueWhenEqual = true; 1864 Value *ExtraCase = 0; 1865 1866 if (Cond->getOpcode() == Instruction::Or) { 1867 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true); 1868 } else if (Cond->getOpcode() == Instruction::And) { 1869 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false); 1870 TrueWhenEqual = false; 1871 } 1872 1873 // If we didn't have a multiply compared value, fail. 1874 if (CompVal == 0) return false; 1875 1876 // There might be duplicate constants in the list, which the switch 1877 // instruction can't handle, remove them now. 1878 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 1879 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 1880 1881 // If Extra was used, we require at least two switch values to do the 1882 // transformation. A switch with one value is just an cond branch. 1883 if (ExtraCase && Values.size() < 2) return false; 1884 1885 // Figure out which block is which destination. 1886 BasicBlock *DefaultBB = BI->getSuccessor(1); 1887 BasicBlock *EdgeBB = BI->getSuccessor(0); 1888 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 1889 1890 BasicBlock *BB = BI->getParent(); 1891 1892 // If there are any extra values that couldn't be folded into the switch 1893 // then we evaluate them with an explicit branch first. Split the block 1894 // right before the condbr to handle it. 1895 if (ExtraCase) { 1896 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 1897 // Remove the uncond branch added to the old block. 1898 TerminatorInst *OldTI = BB->getTerminator(); 1899 1900 BranchInst::Create(EdgeBB, NewBB, ExtraCase, OldTI); 1901 OldTI->eraseFromParent(); 1902 BB = NewBB; 1903 } 1904 1905 // Convert pointer to int before we switch. 1906 if (CompVal->getType()->isPointerTy()) { 1907 assert(TD && "Cannot switch on pointer without TargetData"); 1908 CompVal = new PtrToIntInst(CompVal, 1909 TD->getIntPtrType(CompVal->getContext()), 1910 "magicptr", BI); 1911 } 1912 1913 // Create the new switch instruction now. 1914 SwitchInst *New = 1915 SwitchInst::Create(CompVal, DefaultBB, Values.size(), BI); 1916 1917 // Add all of the 'cases' to the switch instruction. 1918 for (unsigned i = 0, e = Values.size(); i != e; ++i) 1919 New->addCase(Values[i], EdgeBB); 1920 1921 // We added edges from PI to the EdgeBB. As such, if there were any 1922 // PHI nodes in EdgeBB, they need entries to be added corresponding to 1923 // the number of edges added. 1924 for (BasicBlock::iterator BBI = EdgeBB->begin(); 1925 isa<PHINode>(BBI); ++BBI) { 1926 PHINode *PN = cast<PHINode>(BBI); 1927 Value *InVal = PN->getIncomingValueForBlock(BB); 1928 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 1929 PN->addIncoming(InVal, BB); 1930 } 1931 1932 // Erase the old branch instruction. 1933 EraseTerminatorInstAndDCECond(BI); 1934 return true; 1935} 1936 1937bool SimplifyCFGOpt::run(BasicBlock *BB) { 1938 bool Changed = false; 1939 Function *Fn = BB->getParent(); 1940 1941 assert(BB && Fn && "Block not embedded in function!"); 1942 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 1943 1944 // Remove basic blocks that have no predecessors (except the entry block)... 1945 // or that just have themself as a predecessor. These are unreachable. 1946 if ((pred_begin(BB) == pred_end(BB) && BB != &Fn->getEntryBlock()) || 1947 BB->getSinglePredecessor() == BB) { 1948 DEBUG(dbgs() << "Removing BB: \n" << *BB); 1949 DeleteDeadBlock(BB); 1950 return true; 1951 } 1952 1953 // Check to see if we can constant propagate this terminator instruction 1954 // away... 1955 Changed |= ConstantFoldTerminator(BB); 1956 1957 // Check for and eliminate duplicate PHI nodes in this block. 1958 Changed |= EliminateDuplicatePHINodes(BB); 1959 1960 // Merge basic blocks into their predecessor if there is only one distinct 1961 // pred, and if there is only one distinct successor of the predecessor, and 1962 // if there are no PHI nodes. 1963 // 1964 if (MergeBlockIntoPredecessor(BB)) 1965 return true; 1966 1967 // If there is a trivial two-entry PHI node in this basic block, and we can 1968 // eliminate it, do so now. 1969 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 1970 if (PN->getNumIncomingValues() == 2) 1971 Changed |= FoldTwoEntryPHINode(PN); 1972 1973 // If this is a returning block with only PHI nodes in it, fold the return 1974 // instruction into any unconditional branch predecessors. 1975 // 1976 // If any predecessor is a conditional branch that just selects among 1977 // different return values, fold the replace the branch/return with a select 1978 // and return. 1979 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1980 if (BB->getFirstNonPHIOrDbg()->isTerminator()) { 1981 // Find predecessors that end with branches. 1982 SmallVector<BasicBlock*, 8> UncondBranchPreds; 1983 SmallVector<BranchInst*, 8> CondBranchPreds; 1984 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1985 BasicBlock *P = *PI; 1986 TerminatorInst *PTI = P->getTerminator(); 1987 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 1988 if (BI->isUnconditional()) 1989 UncondBranchPreds.push_back(P); 1990 else 1991 CondBranchPreds.push_back(BI); 1992 } 1993 } 1994 1995 // If we found some, do the transformation! 1996 if (!UncondBranchPreds.empty()) { 1997 while (!UncondBranchPreds.empty()) { 1998 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 1999 DEBUG(dbgs() << "FOLDING: " << *BB 2000 << "INTO UNCOND BRANCH PRED: " << *Pred); 2001 Instruction *UncondBranch = Pred->getTerminator(); 2002 // Clone the return and add it to the end of the predecessor. 2003 Instruction *NewRet = RI->clone(); 2004 Pred->getInstList().push_back(NewRet); 2005 2006 // If the return instruction returns a value, and if the value was a 2007 // PHI node in "BB", propagate the right value into the return. 2008 for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end(); 2009 i != e; ++i) 2010 if (PHINode *PN = dyn_cast<PHINode>(*i)) 2011 if (PN->getParent() == BB) 2012 *i = PN->getIncomingValueForBlock(Pred); 2013 2014 // Update any PHI nodes in the returning block to realize that we no 2015 // longer branch to them. 2016 BB->removePredecessor(Pred); 2017 Pred->getInstList().erase(UncondBranch); 2018 } 2019 2020 // If we eliminated all predecessors of the block, delete the block now. 2021 if (pred_begin(BB) == pred_end(BB)) 2022 // We know there are no successors, so just nuke the block. 2023 Fn->getBasicBlockList().erase(BB); 2024 2025 return true; 2026 } 2027 2028 // Check out all of the conditional branches going to this return 2029 // instruction. If any of them just select between returns, change the 2030 // branch itself into a select/return pair. 2031 while (!CondBranchPreds.empty()) { 2032 BranchInst *BI = CondBranchPreds.pop_back_val(); 2033 2034 // Check to see if the non-BB successor is also a return block. 2035 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2036 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2037 SimplifyCondBranchToTwoReturns(BI)) 2038 return true; 2039 } 2040 } 2041 } else if (isa<UnwindInst>(BB->begin())) { 2042 // Check to see if the first instruction in this block is just an unwind. 2043 // If so, replace any invoke instructions which use this as an exception 2044 // destination with call instructions. 2045 // 2046 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2047 while (!Preds.empty()) { 2048 BasicBlock *Pred = Preds.back(); 2049 InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator()); 2050 if (II && II->getUnwindDest() == BB) { 2051 // Insert a new branch instruction before the invoke, because this 2052 // is now a fall through. 2053 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2054 Pred->getInstList().remove(II); // Take out of symbol table 2055 2056 // Insert the call now. 2057 SmallVector<Value*,8> Args(II->op_begin(), II->op_end()-3); 2058 CallInst *CI = CallInst::Create(II->getCalledValue(), 2059 Args.begin(), Args.end(), 2060 II->getName(), BI); 2061 CI->setCallingConv(II->getCallingConv()); 2062 CI->setAttributes(II->getAttributes()); 2063 // If the invoke produced a value, the Call now does instead. 2064 II->replaceAllUsesWith(CI); 2065 delete II; 2066 Changed = true; 2067 } 2068 2069 Preds.pop_back(); 2070 } 2071 2072 // If this block is now dead (and isn't the entry block), remove it. 2073 if (pred_begin(BB) == pred_end(BB) && BB != &Fn->getEntryBlock()) { 2074 // We know there are no successors, so just nuke the block. 2075 Fn->getBasicBlockList().erase(BB); 2076 return true; 2077 } 2078 2079 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2080 if (isValueEqualityComparison(SI)) { 2081 // If we only have one predecessor, and if it is a branch on this value, 2082 // see if that predecessor totally determines the outcome of this switch. 2083 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2084 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 2085 return SimplifyCFG(BB) || 1; 2086 2087 // If the block only contains the switch, see if we can fold the block 2088 // away into any preds. 2089 BasicBlock::iterator BBI = BB->begin(); 2090 // Ignore dbg intrinsics. 2091 while (isa<DbgInfoIntrinsic>(BBI)) 2092 ++BBI; 2093 if (SI == &*BBI) 2094 if (FoldValueComparisonIntoPredecessors(SI)) 2095 return SimplifyCFG(BB) || 1; 2096 } 2097 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 2098 if (BI->isUnconditional()) { 2099 // If the Terminator is the only non-phi instruction, simplify the block. 2100 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg(); 2101 if (I->isTerminator() && BB != &Fn->getEntryBlock() && 2102 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 2103 return true; 2104 2105 // If the only instruction in the block is a seteq/setne comparison 2106 // against a constant, try to simplify the block. 2107 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 2108 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 2109 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 2110 ; 2111 if (I->isTerminator() && 2112 TryToSimplifyUncondBranchWithICmpInIt(ICI)) 2113 return true; 2114 } 2115 2116 } else { // Conditional branch 2117 if (isValueEqualityComparison(BI)) { 2118 // If we only have one predecessor, and if it is a branch on this value, 2119 // see if that predecessor totally determines the outcome of this 2120 // switch. 2121 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2122 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 2123 return SimplifyCFG(BB) | true; 2124 2125 // This block must be empty, except for the setcond inst, if it exists. 2126 // Ignore dbg intrinsics. 2127 BasicBlock::iterator I = BB->begin(); 2128 // Ignore dbg intrinsics. 2129 while (isa<DbgInfoIntrinsic>(I)) 2130 ++I; 2131 if (&*I == BI) { 2132 if (FoldValueComparisonIntoPredecessors(BI)) 2133 return SimplifyCFG(BB) | true; 2134 } else if (&*I == cast<Instruction>(BI->getCondition())){ 2135 ++I; 2136 // Ignore dbg intrinsics. 2137 while (isa<DbgInfoIntrinsic>(I)) 2138 ++I; 2139 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI)) 2140 return SimplifyCFG(BB) | true; 2141 } 2142 } 2143 2144 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 2145 if (SimplifyBranchOnICmpChain(BI, TD)) 2146 return true; 2147 2148 // If this is a branch on a phi node in the current block, thread control 2149 // through this block if any PHI node entries are constants. 2150 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 2151 if (PN->getParent() == BI->getParent()) 2152 if (FoldCondBranchOnPHI(BI)) 2153 return SimplifyCFG(BB) | true; 2154 2155 // If this basic block is ONLY a setcc and a branch, and if a predecessor 2156 // branches to us and one of our successors, fold the setcc into the 2157 // predecessor and use logical operations to pick the right destination. 2158 if (FoldBranchToCommonDest(BI)) 2159 return SimplifyCFG(BB) | true; 2160 2161 // Scan predecessor blocks for conditional branches. 2162 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2163 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2164 if (PBI != BI && PBI->isConditional()) 2165 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2166 return SimplifyCFG(BB) | true; 2167 } 2168 } else if (isa<UnreachableInst>(BB->getTerminator())) { 2169 // If there are any instructions immediately before the unreachable that can 2170 // be removed, do so. 2171 Instruction *Unreachable = BB->getTerminator(); 2172 while (Unreachable != BB->begin()) { 2173 BasicBlock::iterator BBI = Unreachable; 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 2188 BB->getInstList().erase(BBI); 2189 Changed = true; 2190 } 2191 2192 // If the unreachable instruction is the first in the block, take a gander 2193 // at all of the predecessors of this instruction, and simplify them. 2194 if (&BB->front() == Unreachable) { 2195 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2196 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2197 TerminatorInst *TI = Preds[i]->getTerminator(); 2198 2199 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2200 if (BI->isUnconditional()) { 2201 if (BI->getSuccessor(0) == BB) { 2202 new UnreachableInst(TI->getContext(), TI); 2203 TI->eraseFromParent(); 2204 Changed = true; 2205 } 2206 } else { 2207 if (BI->getSuccessor(0) == BB) { 2208 BranchInst::Create(BI->getSuccessor(1), BI); 2209 EraseTerminatorInstAndDCECond(BI); 2210 } else if (BI->getSuccessor(1) == BB) { 2211 BranchInst::Create(BI->getSuccessor(0), BI); 2212 EraseTerminatorInstAndDCECond(BI); 2213 Changed = true; 2214 } 2215 } 2216 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2217 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2218 if (SI->getSuccessor(i) == BB) { 2219 BB->removePredecessor(SI->getParent()); 2220 SI->removeCase(i); 2221 --i; --e; 2222 Changed = true; 2223 } 2224 // If the default value is unreachable, figure out the most popular 2225 // destination and make it the default. 2226 if (SI->getSuccessor(0) == BB) { 2227 std::map<BasicBlock*, unsigned> Popularity; 2228 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2229 Popularity[SI->getSuccessor(i)]++; 2230 2231 // Find the most popular block. 2232 unsigned MaxPop = 0; 2233 BasicBlock *MaxBlock = 0; 2234 for (std::map<BasicBlock*, unsigned>::iterator 2235 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2236 if (I->second > MaxPop) { 2237 MaxPop = I->second; 2238 MaxBlock = I->first; 2239 } 2240 } 2241 if (MaxBlock) { 2242 // Make this the new default, allowing us to delete any explicit 2243 // edges to it. 2244 SI->setSuccessor(0, MaxBlock); 2245 Changed = true; 2246 2247 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2248 // it. 2249 if (isa<PHINode>(MaxBlock->begin())) 2250 for (unsigned i = 0; i != MaxPop-1; ++i) 2251 MaxBlock->removePredecessor(SI->getParent()); 2252 2253 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2254 if (SI->getSuccessor(i) == MaxBlock) { 2255 SI->removeCase(i); 2256 --i; --e; 2257 } 2258 } 2259 } 2260 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2261 if (II->getUnwindDest() == BB) { 2262 // Convert the invoke to a call instruction. This would be a good 2263 // place to note that the call does not throw though. 2264 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2265 II->removeFromParent(); // Take out of symbol table 2266 2267 // Insert the call now... 2268 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2269 CallInst *CI = CallInst::Create(II->getCalledValue(), 2270 Args.begin(), Args.end(), 2271 II->getName(), BI); 2272 CI->setCallingConv(II->getCallingConv()); 2273 CI->setAttributes(II->getAttributes()); 2274 // If the invoke produced a value, the call does now instead. 2275 II->replaceAllUsesWith(CI); 2276 delete II; 2277 Changed = true; 2278 } 2279 } 2280 } 2281 2282 // If this block is now dead, remove it. 2283 if (pred_begin(BB) == pred_end(BB) && BB != &Fn->getEntryBlock()) { 2284 // We know there are no successors, so just nuke the block. 2285 Fn->getBasicBlockList().erase(BB); 2286 return true; 2287 } 2288 } 2289 } else if (IndirectBrInst *IBI = 2290 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 2291 // Eliminate redundant destinations. 2292 SmallPtrSet<Value *, 8> Succs; 2293 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 2294 BasicBlock *Dest = IBI->getDestination(i); 2295 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 2296 Dest->removePredecessor(BB); 2297 IBI->removeDestination(i); 2298 --i; --e; 2299 Changed = true; 2300 } 2301 } 2302 2303 if (IBI->getNumDestinations() == 0) { 2304 // If the indirectbr has no successors, change it to unreachable. 2305 new UnreachableInst(IBI->getContext(), IBI); 2306 EraseTerminatorInstAndDCECond(IBI); 2307 Changed = true; 2308 } else if (IBI->getNumDestinations() == 1) { 2309 // If the indirectbr has one successor, change it to a direct branch. 2310 BranchInst::Create(IBI->getDestination(0), IBI); 2311 EraseTerminatorInstAndDCECond(IBI); 2312 Changed = true; 2313 } else if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 2314 if (SimplifyIndirectBrOnSelect(IBI, SI)) 2315 return SimplifyCFG(BB) | true; 2316 } 2317 } 2318 2319 // Otherwise, if this block only has a single predecessor, and if that block 2320 // is a conditional branch, see if we can hoist any code from this block up 2321 // into our predecessor. 2322 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) { 2323 BranchInst *BI = dyn_cast<BranchInst>(OnlyPred->getTerminator()); 2324 if (BI && BI->isConditional()) { 2325 // Get the other block. 2326 BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB); 2327 pred_iterator PI = pred_begin(OtherBB); 2328 ++PI; 2329 2330 if (PI == pred_end(OtherBB)) { 2331 // We have a conditional branch to two blocks that are only reachable 2332 // from the condbr. We know that the condbr dominates the two blocks, 2333 // so see if there is any identical code in the "then" and "else" 2334 // blocks. If so, we can hoist it up to the branching block. 2335 Changed |= HoistThenElseCodeToIf(BI); 2336 } else { 2337 BasicBlock* OnlySucc = NULL; 2338 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 2339 SI != SE; ++SI) { 2340 if (!OnlySucc) 2341 OnlySucc = *SI; 2342 else if (*SI != OnlySucc) { 2343 OnlySucc = 0; // There are multiple distinct successors! 2344 break; 2345 } 2346 } 2347 2348 if (OnlySucc == OtherBB) { 2349 // If BB's only successor is the other successor of the predecessor, 2350 // i.e. a triangle, see if we can hoist any code from this block up 2351 // to the "if" block. 2352 Changed |= SpeculativelyExecuteBB(BI, BB); 2353 } 2354 } 2355 } 2356 } 2357 2358 return Changed; 2359} 2360 2361/// SimplifyCFG - This function is used to do simplification of a CFG. For 2362/// example, it adjusts branches to branches to eliminate the extra hop, it 2363/// eliminates unreachable basic blocks, and does other "peephole" optimization 2364/// of the CFG. It returns true if a modification was made. 2365/// 2366bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 2367 return SimplifyCFGOpt(TD).run(BB); 2368} 2369