SimplifyCFG.cpp revision ee28e0fdd1955d9fe7acc05b3250ceee759ac3c6
1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Peephole optimize the CFG. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "simplifycfg" 15#include "llvm/Transforms/Utils/Local.h" 16#include "llvm/Constants.h" 17#include "llvm/DerivedTypes.h" 18#include "llvm/GlobalVariable.h" 19#include "llvm/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/LLVMContext.h" 22#include "llvm/Metadata.h" 23#include "llvm/Operator.h" 24#include "llvm/Type.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Analysis/ValueTracking.h" 27#include "llvm/Target/TargetData.h" 28#include "llvm/Transforms/Utils/BasicBlockUtils.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/SetVector.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/ADT/STLExtras.h" 35#include "llvm/Support/CFG.h" 36#include "llvm/Support/CommandLine.h" 37#include "llvm/Support/ConstantRange.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MDBuilder.h" 41#include "llvm/Support/NoFolder.h" 42#include "llvm/Support/raw_ostream.h" 43#include <algorithm> 44#include <set> 45#include <map> 46using namespace llvm; 47 48static cl::opt<unsigned> 49PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 50 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 51 52static cl::opt<bool> 53DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 54 cl::desc("Duplicate return instructions into unconditional branches")); 55 56STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 57 58namespace { 59 /// ValueEqualityComparisonCase - Represents a case of a switch. 60 struct ValueEqualityComparisonCase { 61 ConstantInt *Value; 62 BasicBlock *Dest; 63 64 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 65 : Value(Value), Dest(Dest) {} 66 67 bool operator<(ValueEqualityComparisonCase RHS) const { 68 // Comparing pointers is ok as we only rely on the order for uniquing. 69 return Value < RHS.Value; 70 } 71 }; 72 73class SimplifyCFGOpt { 74 const TargetData *const TD; 75 76 Value *isValueEqualityComparison(TerminatorInst *TI); 77 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 78 std::vector<ValueEqualityComparisonCase> &Cases); 79 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 80 BasicBlock *Pred, 81 IRBuilder<> &Builder); 82 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 83 IRBuilder<> &Builder); 84 85 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 86 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 87 bool SimplifyUnreachable(UnreachableInst *UI); 88 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 89 bool SimplifyIndirectBr(IndirectBrInst *IBI); 90 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 91 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 92 93public: 94 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} 95 bool run(BasicBlock *BB); 96}; 97} 98 99/// SafeToMergeTerminators - Return true if it is safe to merge these two 100/// terminator instructions together. 101/// 102static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 103 if (SI1 == SI2) return false; // Can't merge with self! 104 105 // It is not safe to merge these two switch instructions if they have a common 106 // successor, and if that successor has a PHI node, and if *that* PHI node has 107 // conflicting incoming values from the two switch blocks. 108 BasicBlock *SI1BB = SI1->getParent(); 109 BasicBlock *SI2BB = SI2->getParent(); 110 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 111 112 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 113 if (SI1Succs.count(*I)) 114 for (BasicBlock::iterator BBI = (*I)->begin(); 115 isa<PHINode>(BBI); ++BBI) { 116 PHINode *PN = cast<PHINode>(BBI); 117 if (PN->getIncomingValueForBlock(SI1BB) != 118 PN->getIncomingValueForBlock(SI2BB)) 119 return false; 120 } 121 122 return true; 123} 124 125/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 126/// to merge these two terminator instructions together, where SI1 is an 127/// unconditional branch. PhiNodes will store all PHI nodes in common 128/// successors. 129/// 130static bool isProfitableToFoldUnconditional(BranchInst *SI1, 131 BranchInst *SI2, 132 Instruction* Cond, 133 SmallVectorImpl<PHINode*> &PhiNodes) { 134 if (SI1 == SI2) return false; // Can't merge with self! 135 assert(SI1->isUnconditional() && SI2->isConditional()); 136 137 // We fold the unconditional branch if we can easily update all PHI nodes in 138 // common successors: 139 // 1> We have a constant incoming value for the conditional branch; 140 // 2> We have "Cond" as the incoming value for the unconditional branch; 141 // 3> SI2->getCondition() and Cond have same operands. 142 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 143 if (!Ci2) return false; 144 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 145 Cond->getOperand(1) == Ci2->getOperand(1)) && 146 !(Cond->getOperand(0) == Ci2->getOperand(1) && 147 Cond->getOperand(1) == Ci2->getOperand(0))) 148 return false; 149 150 BasicBlock *SI1BB = SI1->getParent(); 151 BasicBlock *SI2BB = SI2->getParent(); 152 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 153 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 154 if (SI1Succs.count(*I)) 155 for (BasicBlock::iterator BBI = (*I)->begin(); 156 isa<PHINode>(BBI); ++BBI) { 157 PHINode *PN = cast<PHINode>(BBI); 158 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 159 !isa<Constant>(PN->getIncomingValueForBlock(SI2BB))) 160 return false; 161 PhiNodes.push_back(PN); 162 } 163 return true; 164} 165 166/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 167/// now be entries in it from the 'NewPred' block. The values that will be 168/// flowing into the PHI nodes will be the same as those coming in from 169/// ExistPred, an existing predecessor of Succ. 170static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 171 BasicBlock *ExistPred) { 172 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 173 174 PHINode *PN; 175 for (BasicBlock::iterator I = Succ->begin(); 176 (PN = dyn_cast<PHINode>(I)); ++I) 177 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 178} 179 180 181/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 182/// least one PHI node in it), check to see if the merge at this block is due 183/// to an "if condition". If so, return the boolean condition that determines 184/// which entry into BB will be taken. Also, return by references the block 185/// that will be entered from if the condition is true, and the block that will 186/// be entered if the condition is false. 187/// 188/// This does no checking to see if the true/false blocks have large or unsavory 189/// instructions in them. 190static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 191 BasicBlock *&IfFalse) { 192 PHINode *SomePHI = cast<PHINode>(BB->begin()); 193 assert(SomePHI->getNumIncomingValues() == 2 && 194 "Function can only handle blocks with 2 predecessors!"); 195 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 196 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 197 198 // We can only handle branches. Other control flow will be lowered to 199 // branches if possible anyway. 200 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 201 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 202 if (Pred1Br == 0 || Pred2Br == 0) 203 return 0; 204 205 // Eliminate code duplication by ensuring that Pred1Br is conditional if 206 // either are. 207 if (Pred2Br->isConditional()) { 208 // If both branches are conditional, we don't have an "if statement". In 209 // reality, we could transform this case, but since the condition will be 210 // required anyway, we stand no chance of eliminating it, so the xform is 211 // probably not profitable. 212 if (Pred1Br->isConditional()) 213 return 0; 214 215 std::swap(Pred1, Pred2); 216 std::swap(Pred1Br, Pred2Br); 217 } 218 219 if (Pred1Br->isConditional()) { 220 // The only thing we have to watch out for here is to make sure that Pred2 221 // doesn't have incoming edges from other blocks. If it does, the condition 222 // doesn't dominate BB. 223 if (Pred2->getSinglePredecessor() == 0) 224 return 0; 225 226 // If we found a conditional branch predecessor, make sure that it branches 227 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 228 if (Pred1Br->getSuccessor(0) == BB && 229 Pred1Br->getSuccessor(1) == Pred2) { 230 IfTrue = Pred1; 231 IfFalse = Pred2; 232 } else if (Pred1Br->getSuccessor(0) == Pred2 && 233 Pred1Br->getSuccessor(1) == BB) { 234 IfTrue = Pred2; 235 IfFalse = Pred1; 236 } else { 237 // We know that one arm of the conditional goes to BB, so the other must 238 // go somewhere unrelated, and this must not be an "if statement". 239 return 0; 240 } 241 242 return Pred1Br->getCondition(); 243 } 244 245 // Ok, if we got here, both predecessors end with an unconditional branch to 246 // BB. Don't panic! If both blocks only have a single (identical) 247 // predecessor, and THAT is a conditional branch, then we're all ok! 248 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 249 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 250 return 0; 251 252 // Otherwise, if this is a conditional branch, then we can use it! 253 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 254 if (BI == 0) return 0; 255 256 assert(BI->isConditional() && "Two successors but not conditional?"); 257 if (BI->getSuccessor(0) == Pred1) { 258 IfTrue = Pred1; 259 IfFalse = Pred2; 260 } else { 261 IfTrue = Pred2; 262 IfFalse = Pred1; 263 } 264 return BI->getCondition(); 265} 266 267/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 268/// given instruction, which is assumed to be safe to speculate. 1 means 269/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 270static unsigned ComputeSpeculationCost(const User *I) { 271 assert(isSafeToSpeculativelyExecute(I) && 272 "Instruction is not safe to speculatively execute!"); 273 switch (Operator::getOpcode(I)) { 274 default: 275 // In doubt, be conservative. 276 return UINT_MAX; 277 case Instruction::GetElementPtr: 278 // GEPs are cheap if all indices are constant. 279 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 280 return UINT_MAX; 281 return 1; 282 case Instruction::Load: 283 case Instruction::Add: 284 case Instruction::Sub: 285 case Instruction::And: 286 case Instruction::Or: 287 case Instruction::Xor: 288 case Instruction::Shl: 289 case Instruction::LShr: 290 case Instruction::AShr: 291 case Instruction::ICmp: 292 case Instruction::Trunc: 293 case Instruction::ZExt: 294 case Instruction::SExt: 295 return 1; // These are all cheap. 296 297 case Instruction::Call: 298 case Instruction::Select: 299 return 2; 300 } 301} 302 303/// DominatesMergePoint - If we have a merge point of an "if condition" as 304/// accepted above, return true if the specified value dominates the block. We 305/// don't handle the true generality of domination here, just a special case 306/// which works well enough for us. 307/// 308/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 309/// see if V (which must be an instruction) and its recursive operands 310/// that do not dominate BB have a combined cost lower than CostRemaining and 311/// are non-trapping. If both are true, the instruction is inserted into the 312/// set and true is returned. 313/// 314/// The cost for most non-trapping instructions is defined as 1 except for 315/// Select whose cost is 2. 316/// 317/// After this function returns, CostRemaining is decreased by the cost of 318/// V plus its non-dominating operands. If that cost is greater than 319/// CostRemaining, false is returned and CostRemaining is undefined. 320static bool DominatesMergePoint(Value *V, BasicBlock *BB, 321 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 322 unsigned &CostRemaining) { 323 Instruction *I = dyn_cast<Instruction>(V); 324 if (!I) { 325 // Non-instructions all dominate instructions, but not all constantexprs 326 // can be executed unconditionally. 327 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 328 if (C->canTrap()) 329 return false; 330 return true; 331 } 332 BasicBlock *PBB = I->getParent(); 333 334 // We don't want to allow weird loops that might have the "if condition" in 335 // the bottom of this block. 336 if (PBB == BB) return false; 337 338 // If this instruction is defined in a block that contains an unconditional 339 // branch to BB, then it must be in the 'conditional' part of the "if 340 // statement". If not, it definitely dominates the region. 341 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 342 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 343 return true; 344 345 // If we aren't allowing aggressive promotion anymore, then don't consider 346 // instructions in the 'if region'. 347 if (AggressiveInsts == 0) return false; 348 349 // If we have seen this instruction before, don't count it again. 350 if (AggressiveInsts->count(I)) return true; 351 352 // Okay, it looks like the instruction IS in the "condition". Check to 353 // see if it's a cheap instruction to unconditionally compute, and if it 354 // only uses stuff defined outside of the condition. If so, hoist it out. 355 if (!isSafeToSpeculativelyExecute(I)) 356 return false; 357 358 unsigned Cost = ComputeSpeculationCost(I); 359 360 if (Cost > CostRemaining) 361 return false; 362 363 CostRemaining -= Cost; 364 365 // Okay, we can only really hoist these out if their operands do 366 // not take us over the cost threshold. 367 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 368 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 369 return false; 370 // Okay, it's safe to do this! Remember this instruction. 371 AggressiveInsts->insert(I); 372 return true; 373} 374 375/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 376/// and PointerNullValue. Return NULL if value is not a constant int. 377static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { 378 // Normal constant int. 379 ConstantInt *CI = dyn_cast<ConstantInt>(V); 380 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 381 return CI; 382 383 // This is some kind of pointer constant. Turn it into a pointer-sized 384 // ConstantInt if possible. 385 IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); 386 387 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 388 if (isa<ConstantPointerNull>(V)) 389 return ConstantInt::get(PtrTy, 0); 390 391 // IntToPtr const int. 392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 393 if (CE->getOpcode() == Instruction::IntToPtr) 394 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 395 // The constant is very likely to have the right type already. 396 if (CI->getType() == PtrTy) 397 return CI; 398 else 399 return cast<ConstantInt> 400 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 401 } 402 return 0; 403} 404 405/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 406/// collection of icmp eq/ne instructions that compare a value against a 407/// constant, return the value being compared, and stick the constant into the 408/// Values vector. 409static Value * 410GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 411 const TargetData *TD, bool isEQ, unsigned &UsedICmps) { 412 Instruction *I = dyn_cast<Instruction>(V); 413 if (I == 0) return 0; 414 415 // If this is an icmp against a constant, handle this as one of the cases. 416 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 417 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 418 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 419 UsedICmps++; 420 Vals.push_back(C); 421 return I->getOperand(0); 422 } 423 424 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 425 // the set. 426 ConstantRange Span = 427 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 428 429 // If this is an and/!= check then we want to optimize "x ugt 2" into 430 // x != 0 && x != 1. 431 if (!isEQ) 432 Span = Span.inverse(); 433 434 // If there are a ton of values, we don't want to make a ginormous switch. 435 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 436 return 0; 437 438 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 439 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 440 UsedICmps++; 441 return I->getOperand(0); 442 } 443 return 0; 444 } 445 446 // Otherwise, we can only handle an | or &, depending on isEQ. 447 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 448 return 0; 449 450 unsigned NumValsBeforeLHS = Vals.size(); 451 unsigned UsedICmpsBeforeLHS = UsedICmps; 452 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 453 isEQ, UsedICmps)) { 454 unsigned NumVals = Vals.size(); 455 unsigned UsedICmpsBeforeRHS = UsedICmps; 456 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 457 isEQ, UsedICmps)) { 458 if (LHS == RHS) 459 return LHS; 460 Vals.resize(NumVals); 461 UsedICmps = UsedICmpsBeforeRHS; 462 } 463 464 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 465 // set it and return success. 466 if (Extra == 0 || Extra == I->getOperand(1)) { 467 Extra = I->getOperand(1); 468 return LHS; 469 } 470 471 Vals.resize(NumValsBeforeLHS); 472 UsedICmps = UsedICmpsBeforeLHS; 473 return 0; 474 } 475 476 // If the LHS can't be folded in, but Extra is available and RHS can, try to 477 // use LHS as Extra. 478 if (Extra == 0 || Extra == I->getOperand(0)) { 479 Value *OldExtra = Extra; 480 Extra = I->getOperand(0); 481 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 482 isEQ, UsedICmps)) 483 return RHS; 484 assert(Vals.size() == NumValsBeforeLHS); 485 Extra = OldExtra; 486 } 487 488 return 0; 489} 490 491static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 492 Instruction *Cond = 0; 493 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 494 Cond = dyn_cast<Instruction>(SI->getCondition()); 495 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 496 if (BI->isConditional()) 497 Cond = dyn_cast<Instruction>(BI->getCondition()); 498 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 499 Cond = dyn_cast<Instruction>(IBI->getAddress()); 500 } 501 502 TI->eraseFromParent(); 503 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 504} 505 506/// isValueEqualityComparison - Return true if the specified terminator checks 507/// to see if a value is equal to constant integer value. 508Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 509 Value *CV = 0; 510 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 511 // Do not permit merging of large switch instructions into their 512 // predecessors unless there is only one predecessor. 513 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 514 pred_end(SI->getParent())) <= 128) 515 CV = SI->getCondition(); 516 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 517 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 518 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 519 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 520 ICI->getPredicate() == ICmpInst::ICMP_NE) && 521 GetConstantInt(ICI->getOperand(1), TD)) 522 CV = ICI->getOperand(0); 523 524 // Unwrap any lossless ptrtoint cast. 525 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 526 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 527 CV = PTII->getOperand(0); 528 return CV; 529} 530 531/// GetValueEqualityComparisonCases - Given a value comparison instruction, 532/// decode all of the 'cases' that it represents and return the 'default' block. 533BasicBlock *SimplifyCFGOpt:: 534GetValueEqualityComparisonCases(TerminatorInst *TI, 535 std::vector<ValueEqualityComparisonCase> 536 &Cases) { 537 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 538 Cases.reserve(SI->getNumCases()); 539 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 540 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 541 i.getCaseSuccessor())); 542 return SI->getDefaultDest(); 543 } 544 545 BranchInst *BI = cast<BranchInst>(TI); 546 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 547 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 548 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 549 TD), 550 Succ)); 551 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 552} 553 554 555/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 556/// in the list that match the specified block. 557static void EliminateBlockCases(BasicBlock *BB, 558 std::vector<ValueEqualityComparisonCase> &Cases) { 559 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 560 if (Cases[i].Dest == BB) { 561 Cases.erase(Cases.begin()+i); 562 --i; --e; 563 } 564} 565 566/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 567/// well. 568static bool 569ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 570 std::vector<ValueEqualityComparisonCase > &C2) { 571 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 572 573 // Make V1 be smaller than V2. 574 if (V1->size() > V2->size()) 575 std::swap(V1, V2); 576 577 if (V1->size() == 0) return false; 578 if (V1->size() == 1) { 579 // Just scan V2. 580 ConstantInt *TheVal = (*V1)[0].Value; 581 for (unsigned i = 0, e = V2->size(); i != e; ++i) 582 if (TheVal == (*V2)[i].Value) 583 return true; 584 } 585 586 // Otherwise, just sort both lists and compare element by element. 587 array_pod_sort(V1->begin(), V1->end()); 588 array_pod_sort(V2->begin(), V2->end()); 589 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 590 while (i1 != e1 && i2 != e2) { 591 if ((*V1)[i1].Value == (*V2)[i2].Value) 592 return true; 593 if ((*V1)[i1].Value < (*V2)[i2].Value) 594 ++i1; 595 else 596 ++i2; 597 } 598 return false; 599} 600 601/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 602/// terminator instruction and its block is known to only have a single 603/// predecessor block, check to see if that predecessor is also a value 604/// comparison with the same value, and if that comparison determines the 605/// outcome of this comparison. If so, simplify TI. This does a very limited 606/// form of jump threading. 607bool SimplifyCFGOpt:: 608SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 609 BasicBlock *Pred, 610 IRBuilder<> &Builder) { 611 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 612 if (!PredVal) return false; // Not a value comparison in predecessor. 613 614 Value *ThisVal = isValueEqualityComparison(TI); 615 assert(ThisVal && "This isn't a value comparison!!"); 616 if (ThisVal != PredVal) return false; // Different predicates. 617 618 // Find out information about when control will move from Pred to TI's block. 619 std::vector<ValueEqualityComparisonCase> PredCases; 620 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 621 PredCases); 622 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 623 624 // Find information about how control leaves this block. 625 std::vector<ValueEqualityComparisonCase> ThisCases; 626 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 627 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 628 629 // If TI's block is the default block from Pred's comparison, potentially 630 // simplify TI based on this knowledge. 631 if (PredDef == TI->getParent()) { 632 // If we are here, we know that the value is none of those cases listed in 633 // PredCases. If there are any cases in ThisCases that are in PredCases, we 634 // can simplify TI. 635 if (!ValuesOverlap(PredCases, ThisCases)) 636 return false; 637 638 if (isa<BranchInst>(TI)) { 639 // Okay, one of the successors of this condbr is dead. Convert it to a 640 // uncond br. 641 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 642 // Insert the new branch. 643 Instruction *NI = Builder.CreateBr(ThisDef); 644 (void) NI; 645 646 // Remove PHI node entries for the dead edge. 647 ThisCases[0].Dest->removePredecessor(TI->getParent()); 648 649 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 650 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 651 652 EraseTerminatorInstAndDCECond(TI); 653 return true; 654 } 655 656 SwitchInst *SI = cast<SwitchInst>(TI); 657 // Okay, TI has cases that are statically dead, prune them away. 658 SmallPtrSet<Constant*, 16> DeadCases; 659 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 660 DeadCases.insert(PredCases[i].Value); 661 662 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 663 << "Through successor TI: " << *TI); 664 665 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 666 --i; 667 if (DeadCases.count(i.getCaseValue())) { 668 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 669 SI->removeCase(i); 670 } 671 } 672 673 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 674 return true; 675 } 676 677 // Otherwise, TI's block must correspond to some matched value. Find out 678 // which value (or set of values) this is. 679 ConstantInt *TIV = 0; 680 BasicBlock *TIBB = TI->getParent(); 681 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 682 if (PredCases[i].Dest == TIBB) { 683 if (TIV != 0) 684 return false; // Cannot handle multiple values coming to this block. 685 TIV = PredCases[i].Value; 686 } 687 assert(TIV && "No edge from pred to succ?"); 688 689 // Okay, we found the one constant that our value can be if we get into TI's 690 // BB. Find out which successor will unconditionally be branched to. 691 BasicBlock *TheRealDest = 0; 692 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 693 if (ThisCases[i].Value == TIV) { 694 TheRealDest = ThisCases[i].Dest; 695 break; 696 } 697 698 // If not handled by any explicit cases, it is handled by the default case. 699 if (TheRealDest == 0) TheRealDest = ThisDef; 700 701 // Remove PHI node entries for dead edges. 702 BasicBlock *CheckEdge = TheRealDest; 703 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 704 if (*SI != CheckEdge) 705 (*SI)->removePredecessor(TIBB); 706 else 707 CheckEdge = 0; 708 709 // Insert the new branch. 710 Instruction *NI = Builder.CreateBr(TheRealDest); 711 (void) NI; 712 713 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 714 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 715 716 EraseTerminatorInstAndDCECond(TI); 717 return true; 718} 719 720namespace { 721 /// ConstantIntOrdering - This class implements a stable ordering of constant 722 /// integers that does not depend on their address. This is important for 723 /// applications that sort ConstantInt's to ensure uniqueness. 724 struct ConstantIntOrdering { 725 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 726 return LHS->getValue().ult(RHS->getValue()); 727 } 728 }; 729} 730 731static int ConstantIntSortPredicate(const void *P1, const void *P2) { 732 const ConstantInt *LHS = *(const ConstantInt**)P1; 733 const ConstantInt *RHS = *(const ConstantInt**)P2; 734 if (LHS->getValue().ult(RHS->getValue())) 735 return 1; 736 if (LHS->getValue() == RHS->getValue()) 737 return 0; 738 return -1; 739} 740 741/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 742/// equality comparison instruction (either a switch or a branch on "X == c"). 743/// See if any of the predecessors of the terminator block are value comparisons 744/// on the same value. If so, and if safe to do so, fold them together. 745bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 746 IRBuilder<> &Builder) { 747 BasicBlock *BB = TI->getParent(); 748 Value *CV = isValueEqualityComparison(TI); // CondVal 749 assert(CV && "Not a comparison?"); 750 bool Changed = false; 751 752 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 753 while (!Preds.empty()) { 754 BasicBlock *Pred = Preds.pop_back_val(); 755 756 // See if the predecessor is a comparison with the same value. 757 TerminatorInst *PTI = Pred->getTerminator(); 758 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 759 760 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 761 // Figure out which 'cases' to copy from SI to PSI. 762 std::vector<ValueEqualityComparisonCase> BBCases; 763 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 764 765 std::vector<ValueEqualityComparisonCase> PredCases; 766 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 767 768 // Based on whether the default edge from PTI goes to BB or not, fill in 769 // PredCases and PredDefault with the new switch cases we would like to 770 // build. 771 SmallVector<BasicBlock*, 8> NewSuccessors; 772 773 if (PredDefault == BB) { 774 // If this is the default destination from PTI, only the edges in TI 775 // that don't occur in PTI, or that branch to BB will be activated. 776 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 777 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 778 if (PredCases[i].Dest != BB) 779 PTIHandled.insert(PredCases[i].Value); 780 else { 781 // The default destination is BB, we don't need explicit targets. 782 std::swap(PredCases[i], PredCases.back()); 783 PredCases.pop_back(); 784 --i; --e; 785 } 786 787 // Reconstruct the new switch statement we will be building. 788 if (PredDefault != BBDefault) { 789 PredDefault->removePredecessor(Pred); 790 PredDefault = BBDefault; 791 NewSuccessors.push_back(BBDefault); 792 } 793 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 794 if (!PTIHandled.count(BBCases[i].Value) && 795 BBCases[i].Dest != BBDefault) { 796 PredCases.push_back(BBCases[i]); 797 NewSuccessors.push_back(BBCases[i].Dest); 798 } 799 800 } else { 801 // If this is not the default destination from PSI, only the edges 802 // in SI that occur in PSI with a destination of BB will be 803 // activated. 804 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 805 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 806 if (PredCases[i].Dest == BB) { 807 PTIHandled.insert(PredCases[i].Value); 808 std::swap(PredCases[i], PredCases.back()); 809 PredCases.pop_back(); 810 --i; --e; 811 } 812 813 // Okay, now we know which constants were sent to BB from the 814 // predecessor. Figure out where they will all go now. 815 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 816 if (PTIHandled.count(BBCases[i].Value)) { 817 // If this is one we are capable of getting... 818 PredCases.push_back(BBCases[i]); 819 NewSuccessors.push_back(BBCases[i].Dest); 820 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 821 } 822 823 // If there are any constants vectored to BB that TI doesn't handle, 824 // they must go to the default destination of TI. 825 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 826 PTIHandled.begin(), 827 E = PTIHandled.end(); I != E; ++I) { 828 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 829 NewSuccessors.push_back(BBDefault); 830 } 831 } 832 833 // Okay, at this point, we know which new successor Pred will get. Make 834 // sure we update the number of entries in the PHI nodes for these 835 // successors. 836 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 837 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 838 839 Builder.SetInsertPoint(PTI); 840 // Convert pointer to int before we switch. 841 if (CV->getType()->isPointerTy()) { 842 assert(TD && "Cannot switch on pointer without TargetData"); 843 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 844 "magicptr"); 845 } 846 847 // Now that the successors are updated, create the new Switch instruction. 848 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 849 PredCases.size()); 850 NewSI->setDebugLoc(PTI->getDebugLoc()); 851 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 852 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 853 854 EraseTerminatorInstAndDCECond(PTI); 855 856 // Okay, last check. If BB is still a successor of PSI, then we must 857 // have an infinite loop case. If so, add an infinitely looping block 858 // to handle the case to preserve the behavior of the code. 859 BasicBlock *InfLoopBlock = 0; 860 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 861 if (NewSI->getSuccessor(i) == BB) { 862 if (InfLoopBlock == 0) { 863 // Insert it at the end of the function, because it's either code, 864 // or it won't matter if it's hot. :) 865 InfLoopBlock = BasicBlock::Create(BB->getContext(), 866 "infloop", BB->getParent()); 867 BranchInst::Create(InfLoopBlock, InfLoopBlock); 868 } 869 NewSI->setSuccessor(i, InfLoopBlock); 870 } 871 872 Changed = true; 873 } 874 } 875 return Changed; 876} 877 878// isSafeToHoistInvoke - If we would need to insert a select that uses the 879// value of this invoke (comments in HoistThenElseCodeToIf explain why we 880// would need to do this), we can't hoist the invoke, as there is nowhere 881// to put the select in this case. 882static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 883 Instruction *I1, Instruction *I2) { 884 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 885 PHINode *PN; 886 for (BasicBlock::iterator BBI = SI->begin(); 887 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 888 Value *BB1V = PN->getIncomingValueForBlock(BB1); 889 Value *BB2V = PN->getIncomingValueForBlock(BB2); 890 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 891 return false; 892 } 893 } 894 } 895 return true; 896} 897 898/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 899/// BB2, hoist any common code in the two blocks up into the branch block. The 900/// caller of this function guarantees that BI's block dominates BB1 and BB2. 901static bool HoistThenElseCodeToIf(BranchInst *BI) { 902 // This does very trivial matching, with limited scanning, to find identical 903 // instructions in the two blocks. In particular, we don't want to get into 904 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 905 // such, we currently just scan for obviously identical instructions in an 906 // identical order. 907 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 908 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 909 910 BasicBlock::iterator BB1_Itr = BB1->begin(); 911 BasicBlock::iterator BB2_Itr = BB2->begin(); 912 913 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 914 // Skip debug info if it is not identical. 915 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 916 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 917 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 918 while (isa<DbgInfoIntrinsic>(I1)) 919 I1 = BB1_Itr++; 920 while (isa<DbgInfoIntrinsic>(I2)) 921 I2 = BB2_Itr++; 922 } 923 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 924 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 925 return false; 926 927 // If we get here, we can hoist at least one instruction. 928 BasicBlock *BIParent = BI->getParent(); 929 930 do { 931 // If we are hoisting the terminator instruction, don't move one (making a 932 // broken BB), instead clone it, and remove BI. 933 if (isa<TerminatorInst>(I1)) 934 goto HoistTerminator; 935 936 // For a normal instruction, we just move one to right before the branch, 937 // then replace all uses of the other with the first. Finally, we remove 938 // the now redundant second instruction. 939 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 940 if (!I2->use_empty()) 941 I2->replaceAllUsesWith(I1); 942 I1->intersectOptionalDataWith(I2); 943 I2->eraseFromParent(); 944 945 I1 = BB1_Itr++; 946 I2 = BB2_Itr++; 947 // Skip debug info if it is not identical. 948 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 949 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 950 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 951 while (isa<DbgInfoIntrinsic>(I1)) 952 I1 = BB1_Itr++; 953 while (isa<DbgInfoIntrinsic>(I2)) 954 I2 = BB2_Itr++; 955 } 956 } while (I1->isIdenticalToWhenDefined(I2)); 957 958 return true; 959 960HoistTerminator: 961 // It may not be possible to hoist an invoke. 962 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 963 return true; 964 965 // Okay, it is safe to hoist the terminator. 966 Instruction *NT = I1->clone(); 967 BIParent->getInstList().insert(BI, NT); 968 if (!NT->getType()->isVoidTy()) { 969 I1->replaceAllUsesWith(NT); 970 I2->replaceAllUsesWith(NT); 971 NT->takeName(I1); 972 } 973 974 IRBuilder<true, NoFolder> Builder(NT); 975 // Hoisting one of the terminators from our successor is a great thing. 976 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 977 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 978 // nodes, so we insert select instruction to compute the final result. 979 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 980 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 981 PHINode *PN; 982 for (BasicBlock::iterator BBI = SI->begin(); 983 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 984 Value *BB1V = PN->getIncomingValueForBlock(BB1); 985 Value *BB2V = PN->getIncomingValueForBlock(BB2); 986 if (BB1V == BB2V) continue; 987 988 // These values do not agree. Insert a select instruction before NT 989 // that determines the right value. 990 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 991 if (SI == 0) 992 SI = cast<SelectInst> 993 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 994 BB1V->getName()+"."+BB2V->getName())); 995 996 // Make the PHI node use the select for all incoming values for BB1/BB2 997 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 998 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 999 PN->setIncomingValue(i, SI); 1000 } 1001 } 1002 1003 // Update any PHI nodes in our new successors. 1004 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1005 AddPredecessorToBlock(*SI, BIParent, BB1); 1006 1007 EraseTerminatorInstAndDCECond(BI); 1008 return true; 1009} 1010 1011/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 1012/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 1013/// (for now, restricted to a single instruction that's side effect free) from 1014/// the BB1 into the branch block to speculatively execute it. 1015/// 1016/// Turn 1017/// BB: 1018/// %t1 = icmp 1019/// br i1 %t1, label %BB1, label %BB2 1020/// BB1: 1021/// %t3 = add %t2, c 1022/// br label BB2 1023/// BB2: 1024/// => 1025/// BB: 1026/// %t1 = icmp 1027/// %t4 = add %t2, c 1028/// %t3 = select i1 %t1, %t2, %t3 1029static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 1030 // Only speculatively execution a single instruction (not counting the 1031 // terminator) for now. 1032 Instruction *HInst = NULL; 1033 Instruction *Term = BB1->getTerminator(); 1034 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 1035 BBI != BBE; ++BBI) { 1036 Instruction *I = BBI; 1037 // Skip debug info. 1038 if (isa<DbgInfoIntrinsic>(I)) continue; 1039 if (I == Term) break; 1040 1041 if (HInst) 1042 return false; 1043 HInst = I; 1044 } 1045 1046 BasicBlock *BIParent = BI->getParent(); 1047 1048 // Check the instruction to be hoisted, if there is one. 1049 if (HInst) { 1050 // Don't hoist the instruction if it's unsafe or expensive. 1051 if (!isSafeToSpeculativelyExecute(HInst)) 1052 return false; 1053 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold) 1054 return false; 1055 1056 // Do not hoist the instruction if any of its operands are defined but not 1057 // used in this BB. The transformation will prevent the operand from 1058 // being sunk into the use block. 1059 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1060 i != e; ++i) { 1061 Instruction *OpI = dyn_cast<Instruction>(*i); 1062 if (OpI && OpI->getParent() == BIParent && 1063 !OpI->mayHaveSideEffects() && 1064 !OpI->isUsedInBasicBlock(BIParent)) 1065 return false; 1066 } 1067 } 1068 1069 // Be conservative for now. FP select instruction can often be expensive. 1070 Value *BrCond = BI->getCondition(); 1071 if (isa<FCmpInst>(BrCond)) 1072 return false; 1073 1074 // If BB1 is actually on the false edge of the conditional branch, remember 1075 // to swap the select operands later. 1076 bool Invert = false; 1077 if (BB1 != BI->getSuccessor(0)) { 1078 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1079 Invert = true; 1080 } 1081 1082 // Collect interesting PHIs, and scan for hazards. 1083 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs; 1084 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1085 for (BasicBlock::iterator I = BB2->begin(); 1086 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1087 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1088 Value *BIParentV = PN->getIncomingValueForBlock(BIParent); 1089 1090 // Skip PHIs which are trivial. 1091 if (BB1V == BIParentV) 1092 continue; 1093 1094 // Check for saftey. 1095 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) { 1096 // An unfolded ConstantExpr could end up getting expanded into 1097 // Instructions. Don't speculate this and another instruction at 1098 // the same time. 1099 if (HInst) 1100 return false; 1101 if (!isSafeToSpeculativelyExecute(CE)) 1102 return false; 1103 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1104 return false; 1105 } 1106 1107 // Ok, we may insert a select for this PHI. 1108 PHIs.insert(std::make_pair(BB1V, BIParentV)); 1109 } 1110 1111 // If there are no PHIs to process, bail early. This helps ensure idempotence 1112 // as well. 1113 if (PHIs.empty()) 1114 return false; 1115 1116 // If we get here, we can hoist the instruction and if-convert. 1117 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); 1118 1119 // Hoist the instruction. 1120 if (HInst) 1121 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst); 1122 1123 // Insert selects and rewrite the PHI operands. 1124 IRBuilder<true, NoFolder> Builder(BI); 1125 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1126 Value *TrueV = PHIs[i].first; 1127 Value *FalseV = PHIs[i].second; 1128 1129 // Create a select whose true value is the speculatively executed value and 1130 // false value is the previously determined FalseV. 1131 SelectInst *SI; 1132 if (Invert) 1133 SI = cast<SelectInst> 1134 (Builder.CreateSelect(BrCond, FalseV, TrueV, 1135 FalseV->getName() + "." + TrueV->getName())); 1136 else 1137 SI = cast<SelectInst> 1138 (Builder.CreateSelect(BrCond, TrueV, FalseV, 1139 TrueV->getName() + "." + FalseV->getName())); 1140 1141 // Make the PHI node use the select for all incoming values for "then" and 1142 // "if" blocks. 1143 for (BasicBlock::iterator I = BB2->begin(); 1144 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1145 unsigned BB1I = PN->getBasicBlockIndex(BB1); 1146 unsigned BIParentI = PN->getBasicBlockIndex(BIParent); 1147 Value *BB1V = PN->getIncomingValue(BB1I); 1148 Value *BIParentV = PN->getIncomingValue(BIParentI); 1149 if (TrueV == BB1V && FalseV == BIParentV) { 1150 PN->setIncomingValue(BB1I, SI); 1151 PN->setIncomingValue(BIParentI, SI); 1152 } 1153 } 1154 } 1155 1156 ++NumSpeculations; 1157 return true; 1158} 1159 1160/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1161/// across this block. 1162static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1163 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1164 unsigned Size = 0; 1165 1166 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1167 if (isa<DbgInfoIntrinsic>(BBI)) 1168 continue; 1169 if (Size > 10) return false; // Don't clone large BB's. 1170 ++Size; 1171 1172 // We can only support instructions that do not define values that are 1173 // live outside of the current basic block. 1174 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1175 UI != E; ++UI) { 1176 Instruction *U = cast<Instruction>(*UI); 1177 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1178 } 1179 1180 // Looks ok, continue checking. 1181 } 1182 1183 return true; 1184} 1185 1186/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1187/// that is defined in the same block as the branch and if any PHI entries are 1188/// constants, thread edges corresponding to that entry to be branches to their 1189/// ultimate destination. 1190static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { 1191 BasicBlock *BB = BI->getParent(); 1192 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1193 // NOTE: we currently cannot transform this case if the PHI node is used 1194 // outside of the block. 1195 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1196 return false; 1197 1198 // Degenerate case of a single entry PHI. 1199 if (PN->getNumIncomingValues() == 1) { 1200 FoldSingleEntryPHINodes(PN->getParent()); 1201 return true; 1202 } 1203 1204 // Now we know that this block has multiple preds and two succs. 1205 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1206 1207 // Okay, this is a simple enough basic block. See if any phi values are 1208 // constants. 1209 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1210 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1211 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1212 1213 // Okay, we now know that all edges from PredBB should be revectored to 1214 // branch to RealDest. 1215 BasicBlock *PredBB = PN->getIncomingBlock(i); 1216 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1217 1218 if (RealDest == BB) continue; // Skip self loops. 1219 // Skip if the predecessor's terminator is an indirect branch. 1220 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1221 1222 // The dest block might have PHI nodes, other predecessors and other 1223 // difficult cases. Instead of being smart about this, just insert a new 1224 // block that jumps to the destination block, effectively splitting 1225 // the edge we are about to create. 1226 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1227 RealDest->getName()+".critedge", 1228 RealDest->getParent(), RealDest); 1229 BranchInst::Create(RealDest, EdgeBB); 1230 1231 // Update PHI nodes. 1232 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1233 1234 // BB may have instructions that are being threaded over. Clone these 1235 // instructions into EdgeBB. We know that there will be no uses of the 1236 // cloned instructions outside of EdgeBB. 1237 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1238 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1239 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1240 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1241 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1242 continue; 1243 } 1244 // Clone the instruction. 1245 Instruction *N = BBI->clone(); 1246 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1247 1248 // Update operands due to translation. 1249 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1250 i != e; ++i) { 1251 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1252 if (PI != TranslateMap.end()) 1253 *i = PI->second; 1254 } 1255 1256 // Check for trivial simplification. 1257 if (Value *V = SimplifyInstruction(N, TD)) { 1258 TranslateMap[BBI] = V; 1259 delete N; // Instruction folded away, don't need actual inst 1260 } else { 1261 // Insert the new instruction into its new home. 1262 EdgeBB->getInstList().insert(InsertPt, N); 1263 if (!BBI->use_empty()) 1264 TranslateMap[BBI] = N; 1265 } 1266 } 1267 1268 // Loop over all of the edges from PredBB to BB, changing them to branch 1269 // to EdgeBB instead. 1270 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1271 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1272 if (PredBBTI->getSuccessor(i) == BB) { 1273 BB->removePredecessor(PredBB); 1274 PredBBTI->setSuccessor(i, EdgeBB); 1275 } 1276 1277 // Recurse, simplifying any other constants. 1278 return FoldCondBranchOnPHI(BI, TD) | true; 1279 } 1280 1281 return false; 1282} 1283 1284/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1285/// PHI node, see if we can eliminate it. 1286static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { 1287 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1288 // statement", which has a very simple dominance structure. Basically, we 1289 // are trying to find the condition that is being branched on, which 1290 // subsequently causes this merge to happen. We really want control 1291 // dependence information for this check, but simplifycfg can't keep it up 1292 // to date, and this catches most of the cases we care about anyway. 1293 BasicBlock *BB = PN->getParent(); 1294 BasicBlock *IfTrue, *IfFalse; 1295 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1296 if (!IfCond || 1297 // Don't bother if the branch will be constant folded trivially. 1298 isa<ConstantInt>(IfCond)) 1299 return false; 1300 1301 // Okay, we found that we can merge this two-entry phi node into a select. 1302 // Doing so would require us to fold *all* two entry phi nodes in this block. 1303 // At some point this becomes non-profitable (particularly if the target 1304 // doesn't support cmov's). Only do this transformation if there are two or 1305 // fewer PHI nodes in this block. 1306 unsigned NumPhis = 0; 1307 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1308 if (NumPhis > 2) 1309 return false; 1310 1311 // Loop over the PHI's seeing if we can promote them all to select 1312 // instructions. While we are at it, keep track of the instructions 1313 // that need to be moved to the dominating block. 1314 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1315 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1316 MaxCostVal1 = PHINodeFoldingThreshold; 1317 1318 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1319 PHINode *PN = cast<PHINode>(II++); 1320 if (Value *V = SimplifyInstruction(PN, TD)) { 1321 PN->replaceAllUsesWith(V); 1322 PN->eraseFromParent(); 1323 continue; 1324 } 1325 1326 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1327 MaxCostVal0) || 1328 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1329 MaxCostVal1)) 1330 return false; 1331 } 1332 1333 // If we folded the the first phi, PN dangles at this point. Refresh it. If 1334 // we ran out of PHIs then we simplified them all. 1335 PN = dyn_cast<PHINode>(BB->begin()); 1336 if (PN == 0) return true; 1337 1338 // Don't fold i1 branches on PHIs which contain binary operators. These can 1339 // often be turned into switches and other things. 1340 if (PN->getType()->isIntegerTy(1) && 1341 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1342 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1343 isa<BinaryOperator>(IfCond))) 1344 return false; 1345 1346 // If we all PHI nodes are promotable, check to make sure that all 1347 // instructions in the predecessor blocks can be promoted as well. If 1348 // not, we won't be able to get rid of the control flow, so it's not 1349 // worth promoting to select instructions. 1350 BasicBlock *DomBlock = 0; 1351 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1352 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1353 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1354 IfBlock1 = 0; 1355 } else { 1356 DomBlock = *pred_begin(IfBlock1); 1357 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1358 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1359 // This is not an aggressive instruction that we can promote. 1360 // Because of this, we won't be able to get rid of the control 1361 // flow, so the xform is not worth it. 1362 return false; 1363 } 1364 } 1365 1366 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1367 IfBlock2 = 0; 1368 } else { 1369 DomBlock = *pred_begin(IfBlock2); 1370 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1371 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1372 // This is not an aggressive instruction that we can promote. 1373 // Because of this, we won't be able to get rid of the control 1374 // flow, so the xform is not worth it. 1375 return false; 1376 } 1377 } 1378 1379 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1380 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1381 1382 // If we can still promote the PHI nodes after this gauntlet of tests, 1383 // do all of the PHI's now. 1384 Instruction *InsertPt = DomBlock->getTerminator(); 1385 IRBuilder<true, NoFolder> Builder(InsertPt); 1386 1387 // Move all 'aggressive' instructions, which are defined in the 1388 // conditional parts of the if's up to the dominating block. 1389 if (IfBlock1) 1390 DomBlock->getInstList().splice(InsertPt, 1391 IfBlock1->getInstList(), IfBlock1->begin(), 1392 IfBlock1->getTerminator()); 1393 if (IfBlock2) 1394 DomBlock->getInstList().splice(InsertPt, 1395 IfBlock2->getInstList(), IfBlock2->begin(), 1396 IfBlock2->getTerminator()); 1397 1398 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1399 // Change the PHI node into a select instruction. 1400 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1401 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1402 1403 SelectInst *NV = 1404 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1405 PN->replaceAllUsesWith(NV); 1406 NV->takeName(PN); 1407 PN->eraseFromParent(); 1408 } 1409 1410 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1411 // has been flattened. Change DomBlock to jump directly to our new block to 1412 // avoid other simplifycfg's kicking in on the diamond. 1413 TerminatorInst *OldTI = DomBlock->getTerminator(); 1414 Builder.SetInsertPoint(OldTI); 1415 Builder.CreateBr(BB); 1416 OldTI->eraseFromParent(); 1417 return true; 1418} 1419 1420/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1421/// to two returning blocks, try to merge them together into one return, 1422/// introducing a select if the return values disagree. 1423static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1424 IRBuilder<> &Builder) { 1425 assert(BI->isConditional() && "Must be a conditional branch"); 1426 BasicBlock *TrueSucc = BI->getSuccessor(0); 1427 BasicBlock *FalseSucc = BI->getSuccessor(1); 1428 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1429 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1430 1431 // Check to ensure both blocks are empty (just a return) or optionally empty 1432 // with PHI nodes. If there are other instructions, merging would cause extra 1433 // computation on one path or the other. 1434 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1435 return false; 1436 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1437 return false; 1438 1439 Builder.SetInsertPoint(BI); 1440 // Okay, we found a branch that is going to two return nodes. If 1441 // there is no return value for this function, just change the 1442 // branch into a return. 1443 if (FalseRet->getNumOperands() == 0) { 1444 TrueSucc->removePredecessor(BI->getParent()); 1445 FalseSucc->removePredecessor(BI->getParent()); 1446 Builder.CreateRetVoid(); 1447 EraseTerminatorInstAndDCECond(BI); 1448 return true; 1449 } 1450 1451 // Otherwise, figure out what the true and false return values are 1452 // so we can insert a new select instruction. 1453 Value *TrueValue = TrueRet->getReturnValue(); 1454 Value *FalseValue = FalseRet->getReturnValue(); 1455 1456 // Unwrap any PHI nodes in the return blocks. 1457 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1458 if (TVPN->getParent() == TrueSucc) 1459 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1460 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1461 if (FVPN->getParent() == FalseSucc) 1462 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1463 1464 // In order for this transformation to be safe, we must be able to 1465 // unconditionally execute both operands to the return. This is 1466 // normally the case, but we could have a potentially-trapping 1467 // constant expression that prevents this transformation from being 1468 // safe. 1469 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1470 if (TCV->canTrap()) 1471 return false; 1472 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1473 if (FCV->canTrap()) 1474 return false; 1475 1476 // Okay, we collected all the mapped values and checked them for sanity, and 1477 // defined to really do this transformation. First, update the CFG. 1478 TrueSucc->removePredecessor(BI->getParent()); 1479 FalseSucc->removePredecessor(BI->getParent()); 1480 1481 // Insert select instructions where needed. 1482 Value *BrCond = BI->getCondition(); 1483 if (TrueValue) { 1484 // Insert a select if the results differ. 1485 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1486 } else if (isa<UndefValue>(TrueValue)) { 1487 TrueValue = FalseValue; 1488 } else { 1489 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1490 FalseValue, "retval"); 1491 } 1492 } 1493 1494 Value *RI = !TrueValue ? 1495 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1496 1497 (void) RI; 1498 1499 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1500 << "\n " << *BI << "NewRet = " << *RI 1501 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1502 1503 EraseTerminatorInstAndDCECond(BI); 1504 1505 return true; 1506} 1507 1508/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1509/// probabilities of the branch taking each edge. Fills in the two APInt 1510/// parameters and return true, or returns false if no or invalid metadata was 1511/// found. 1512static bool ExtractBranchMetadata(BranchInst *BI, 1513 APInt &ProbTrue, APInt &ProbFalse) { 1514 assert(BI->isConditional() && 1515 "Looking for probabilities on unconditional branch?"); 1516 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1517 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1518 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1519 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1520 if (!CITrue || !CIFalse) return false; 1521 ProbTrue = CITrue->getValue(); 1522 ProbFalse = CIFalse->getValue(); 1523 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 && 1524 "Branch probability metadata must be 32-bit integers"); 1525 return true; 1526} 1527 1528/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In 1529/// the event of overflow, logically-shifts all four inputs right until the 1530/// multiply fits. 1531static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D, 1532 unsigned &BitsLost) { 1533 BitsLost = 0; 1534 bool Overflow = false; 1535 APInt Result = A.umul_ov(B, Overflow); 1536 if (Overflow) { 1537 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A); 1538 do { 1539 B = B.lshr(1); 1540 ++BitsLost; 1541 } while (B.ugt(MaxB)); 1542 A = A.lshr(BitsLost); 1543 C = C.lshr(BitsLost); 1544 D = D.lshr(BitsLost); 1545 Result = A * B; 1546 } 1547 return Result; 1548} 1549 1550/// checkCSEInPredecessor - Return true if the given instruction is available 1551/// in its predecessor block. If yes, the instruction will be removed. 1552/// 1553bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1554 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1555 return false; 1556 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1557 Instruction *PBI = &*I; 1558 // Check whether Inst and PBI generate the same value. 1559 if (Inst->isIdenticalTo(PBI)) { 1560 Inst->replaceAllUsesWith(PBI); 1561 Inst->eraseFromParent(); 1562 return true; 1563 } 1564 } 1565 return false; 1566} 1567 1568/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1569/// predecessor branches to us and one of our successors, fold the block into 1570/// the predecessor and use logical operations to pick the right destination. 1571bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1572 BasicBlock *BB = BI->getParent(); 1573 1574 Instruction *Cond = 0; 1575 if (BI->isConditional()) 1576 Cond = dyn_cast<Instruction>(BI->getCondition()); 1577 else { 1578 // For unconditional branch, check for a simple CFG pattern, where 1579 // BB has a single predecessor and BB's successor is also its predecessor's 1580 // successor. If such pattern exisits, check for CSE between BB and its 1581 // predecessor. 1582 if (BasicBlock *PB = BB->getSinglePredecessor()) 1583 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1584 if (PBI->isConditional() && 1585 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1586 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1587 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1588 I != E; ) { 1589 Instruction *Curr = I++; 1590 if (isa<CmpInst>(Curr)) { 1591 Cond = Curr; 1592 break; 1593 } 1594 // Quit if we can't remove this instruction. 1595 if (!checkCSEInPredecessor(Curr, PB)) 1596 return false; 1597 } 1598 } 1599 1600 if (Cond == 0) 1601 return false; 1602 } 1603 1604 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1605 Cond->getParent() != BB || !Cond->hasOneUse()) 1606 return false; 1607 1608 // Only allow this if the condition is a simple instruction that can be 1609 // executed unconditionally. It must be in the same block as the branch, and 1610 // must be at the front of the block. 1611 BasicBlock::iterator FrontIt = BB->front(); 1612 1613 // Ignore dbg intrinsics. 1614 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1615 1616 // Allow a single instruction to be hoisted in addition to the compare 1617 // that feeds the branch. We later ensure that any values that _it_ uses 1618 // were also live in the predecessor, so that we don't unnecessarily create 1619 // register pressure or inhibit out-of-order execution. 1620 Instruction *BonusInst = 0; 1621 if (&*FrontIt != Cond && 1622 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1623 isSafeToSpeculativelyExecute(FrontIt)) { 1624 BonusInst = &*FrontIt; 1625 ++FrontIt; 1626 1627 // Ignore dbg intrinsics. 1628 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1629 } 1630 1631 // Only a single bonus inst is allowed. 1632 if (&*FrontIt != Cond) 1633 return false; 1634 1635 // Make sure the instruction after the condition is the cond branch. 1636 BasicBlock::iterator CondIt = Cond; ++CondIt; 1637 1638 // Ingore dbg intrinsics. 1639 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1640 1641 if (&*CondIt != BI) 1642 return false; 1643 1644 // Cond is known to be a compare or binary operator. Check to make sure that 1645 // neither operand is a potentially-trapping constant expression. 1646 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1647 if (CE->canTrap()) 1648 return false; 1649 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1650 if (CE->canTrap()) 1651 return false; 1652 1653 // Finally, don't infinitely unroll conditional loops. 1654 BasicBlock *TrueDest = BI->getSuccessor(0); 1655 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 1656 if (TrueDest == BB || FalseDest == BB) 1657 return false; 1658 1659 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1660 BasicBlock *PredBlock = *PI; 1661 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1662 1663 // Check that we have two conditional branches. If there is a PHI node in 1664 // the common successor, verify that the same value flows in from both 1665 // blocks. 1666 SmallVector<PHINode*, 4> PHIs; 1667 if (PBI == 0 || PBI->isUnconditional() || 1668 (BI->isConditional() && 1669 !SafeToMergeTerminators(BI, PBI)) || 1670 (!BI->isConditional() && 1671 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 1672 continue; 1673 1674 // Determine if the two branches share a common destination. 1675 Instruction::BinaryOps Opc; 1676 bool InvertPredCond = false; 1677 1678 if (BI->isConditional()) { 1679 if (PBI->getSuccessor(0) == TrueDest) 1680 Opc = Instruction::Or; 1681 else if (PBI->getSuccessor(1) == FalseDest) 1682 Opc = Instruction::And; 1683 else if (PBI->getSuccessor(0) == FalseDest) 1684 Opc = Instruction::And, InvertPredCond = true; 1685 else if (PBI->getSuccessor(1) == TrueDest) 1686 Opc = Instruction::Or, InvertPredCond = true; 1687 else 1688 continue; 1689 } else { 1690 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 1691 continue; 1692 } 1693 1694 // Ensure that any values used in the bonus instruction are also used 1695 // by the terminator of the predecessor. This means that those values 1696 // must already have been resolved, so we won't be inhibiting the 1697 // out-of-order core by speculating them earlier. 1698 if (BonusInst) { 1699 // Collect the values used by the bonus inst 1700 SmallPtrSet<Value*, 4> UsedValues; 1701 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1702 OE = BonusInst->op_end(); OI != OE; ++OI) { 1703 Value *V = *OI; 1704 if (!isa<Constant>(V)) 1705 UsedValues.insert(V); 1706 } 1707 1708 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1709 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1710 1711 // Walk up to four levels back up the use-def chain of the predecessor's 1712 // terminator to see if all those values were used. The choice of four 1713 // levels is arbitrary, to provide a compile-time-cost bound. 1714 while (!Worklist.empty()) { 1715 std::pair<Value*, unsigned> Pair = Worklist.back(); 1716 Worklist.pop_back(); 1717 1718 if (Pair.second >= 4) continue; 1719 UsedValues.erase(Pair.first); 1720 if (UsedValues.empty()) break; 1721 1722 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 1723 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1724 OI != OE; ++OI) 1725 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1726 } 1727 } 1728 1729 if (!UsedValues.empty()) return false; 1730 } 1731 1732 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1733 IRBuilder<> Builder(PBI); 1734 1735 // If we need to invert the condition in the pred block to match, do so now. 1736 if (InvertPredCond) { 1737 Value *NewCond = PBI->getCondition(); 1738 1739 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1740 CmpInst *CI = cast<CmpInst>(NewCond); 1741 CI->setPredicate(CI->getInversePredicate()); 1742 } else { 1743 NewCond = Builder.CreateNot(NewCond, 1744 PBI->getCondition()->getName()+".not"); 1745 } 1746 1747 PBI->setCondition(NewCond); 1748 PBI->swapSuccessors(); 1749 } 1750 1751 // If we have a bonus inst, clone it into the predecessor block. 1752 Instruction *NewBonus = 0; 1753 if (BonusInst) { 1754 NewBonus = BonusInst->clone(); 1755 PredBlock->getInstList().insert(PBI, NewBonus); 1756 NewBonus->takeName(BonusInst); 1757 BonusInst->setName(BonusInst->getName()+".old"); 1758 } 1759 1760 // Clone Cond into the predecessor basic block, and or/and the 1761 // two conditions together. 1762 Instruction *New = Cond->clone(); 1763 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1764 PredBlock->getInstList().insert(PBI, New); 1765 New->takeName(Cond); 1766 Cond->setName(New->getName()+".old"); 1767 1768 if (BI->isConditional()) { 1769 Instruction *NewCond = 1770 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 1771 New, "or.cond")); 1772 PBI->setCondition(NewCond); 1773 1774 if (PBI->getSuccessor(0) == BB) { 1775 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1776 PBI->setSuccessor(0, TrueDest); 1777 } 1778 if (PBI->getSuccessor(1) == BB) { 1779 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1780 PBI->setSuccessor(1, FalseDest); 1781 } 1782 } else { 1783 // Update PHI nodes in the common successors. 1784 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1785 ConstantInt *PBI_C = dyn_cast<ConstantInt>( 1786 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 1787 assert(PBI_C->getType()->isIntegerTy(1)); 1788 Instruction *MergedCond = 0; 1789 if (PBI->getSuccessor(0) == TrueDest) { 1790 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 1791 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 1792 // is false: !PBI_Cond and BI_Value 1793 Instruction *NotCond = 1794 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 1795 "not.cond")); 1796 MergedCond = 1797 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 1798 NotCond, New, 1799 "and.cond")); 1800 if (PBI_C->isOne()) 1801 MergedCond = 1802 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 1803 PBI->getCondition(), MergedCond, 1804 "or.cond")); 1805 } else { 1806 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 1807 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 1808 // is false: PBI_Cond and BI_Value 1809 MergedCond = 1810 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 1811 PBI->getCondition(), New, 1812 "and.cond")); 1813 if (PBI_C->isOne()) { 1814 Instruction *NotCond = 1815 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 1816 "not.cond")); 1817 MergedCond = 1818 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 1819 NotCond, MergedCond, 1820 "or.cond")); 1821 } 1822 } 1823 // Update PHI Node. 1824 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 1825 MergedCond); 1826 } 1827 // Change PBI from Conditional to Unconditional. 1828 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 1829 EraseTerminatorInstAndDCECond(PBI); 1830 PBI = New_PBI; 1831 } 1832 1833 // TODO: If BB is reachable from all paths through PredBlock, then we 1834 // could replace PBI's branch probabilities with BI's. 1835 1836 // Merge probability data into PredBlock's branch. 1837 APInt A, B, C, D; 1838 if (PBI->isConditional() && BI->isConditional() && 1839 ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) { 1840 // Given IR which does: 1841 // bbA: 1842 // br i1 %x, label %bbB, label %bbC 1843 // bbB: 1844 // br i1 %y, label %bbD, label %bbC 1845 // Let's call the probability that we take the edge from %bbA to %bbB 1846 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to 1847 // %bbC probability 'd'. 1848 // 1849 // We transform the IR into: 1850 // bbA: 1851 // br i1 %z, label %bbD, label %bbC 1852 // where the probability of going to %bbD is (a*c) and going to bbC is 1853 // (b+a*d). 1854 // 1855 // Probabilities aren't stored as ratios directly. Using branch weights, 1856 // we get: 1857 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D. 1858 1859 // In the event of overflow, we want to drop the LSB of the input 1860 // probabilities. 1861 unsigned BitsLost; 1862 1863 // Ignore overflow result on ProbTrue. 1864 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost); 1865 1866 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost); 1867 if (BitsLost) { 1868 ProbTrue = ProbTrue.lshr(BitsLost*2); 1869 } 1870 1871 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost); 1872 if (BitsLost) { 1873 ProbTrue = ProbTrue.lshr(BitsLost*2); 1874 Tmp1 = Tmp1.lshr(BitsLost*2); 1875 } 1876 1877 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost); 1878 if (BitsLost) { 1879 ProbTrue = ProbTrue.lshr(BitsLost*2); 1880 Tmp1 = Tmp1.lshr(BitsLost*2); 1881 Tmp2 = Tmp2.lshr(BitsLost*2); 1882 } 1883 1884 bool Overflow1 = false, Overflow2 = false; 1885 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1); 1886 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2); 1887 1888 if (Overflow1 || Overflow2) { 1889 ProbTrue = ProbTrue.lshr(1); 1890 Tmp1 = Tmp1.lshr(1); 1891 Tmp2 = Tmp2.lshr(1); 1892 Tmp3 = Tmp3.lshr(1); 1893 Tmp4 = Tmp2 + Tmp3; 1894 ProbFalse = Tmp4 + Tmp1; 1895 } 1896 1897 // The sum of branch weights must fit in 32-bits. 1898 if (ProbTrue.isNegative() && ProbFalse.isNegative()) { 1899 ProbTrue = ProbTrue.lshr(1); 1900 ProbFalse = ProbFalse.lshr(1); 1901 } 1902 1903 if (ProbTrue != ProbFalse) { 1904 // Normalize the result. 1905 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse); 1906 ProbTrue = ProbTrue.udiv(GCD); 1907 ProbFalse = ProbFalse.udiv(GCD); 1908 1909 MDBuilder MDB(BI->getContext()); 1910 MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(), 1911 ProbFalse.getZExtValue()); 1912 PBI->setMetadata(LLVMContext::MD_prof, N); 1913 } else { 1914 PBI->setMetadata(LLVMContext::MD_prof, NULL); 1915 } 1916 } else { 1917 PBI->setMetadata(LLVMContext::MD_prof, NULL); 1918 } 1919 1920 // Copy any debug value intrinsics into the end of PredBlock. 1921 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 1922 if (isa<DbgInfoIntrinsic>(*I)) 1923 I->clone()->insertBefore(PBI); 1924 1925 return true; 1926 } 1927 return false; 1928} 1929 1930/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1931/// predecessor of another block, this function tries to simplify it. We know 1932/// that PBI and BI are both conditional branches, and BI is in one of the 1933/// successor blocks of PBI - PBI branches to BI. 1934static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1935 assert(PBI->isConditional() && BI->isConditional()); 1936 BasicBlock *BB = BI->getParent(); 1937 1938 // If this block ends with a branch instruction, and if there is a 1939 // predecessor that ends on a branch of the same condition, make 1940 // this conditional branch redundant. 1941 if (PBI->getCondition() == BI->getCondition() && 1942 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1943 // Okay, the outcome of this conditional branch is statically 1944 // knowable. If this block had a single pred, handle specially. 1945 if (BB->getSinglePredecessor()) { 1946 // Turn this into a branch on constant. 1947 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1948 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1949 CondIsTrue)); 1950 return true; // Nuke the branch on constant. 1951 } 1952 1953 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1954 // in the constant and simplify the block result. Subsequent passes of 1955 // simplifycfg will thread the block. 1956 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1957 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 1958 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 1959 std::distance(PB, PE), 1960 BI->getCondition()->getName() + ".pr", 1961 BB->begin()); 1962 // Okay, we're going to insert the PHI node. Since PBI is not the only 1963 // predecessor, compute the PHI'd conditional value for all of the preds. 1964 // Any predecessor where the condition is not computable we keep symbolic. 1965 for (pred_iterator PI = PB; PI != PE; ++PI) { 1966 BasicBlock *P = *PI; 1967 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 1968 PBI != BI && PBI->isConditional() && 1969 PBI->getCondition() == BI->getCondition() && 1970 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1971 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1972 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1973 CondIsTrue), P); 1974 } else { 1975 NewPN->addIncoming(BI->getCondition(), P); 1976 } 1977 } 1978 1979 BI->setCondition(NewPN); 1980 return true; 1981 } 1982 } 1983 1984 // If this is a conditional branch in an empty block, and if any 1985 // predecessors is a conditional branch to one of our destinations, 1986 // fold the conditions into logical ops and one cond br. 1987 BasicBlock::iterator BBI = BB->begin(); 1988 // Ignore dbg intrinsics. 1989 while (isa<DbgInfoIntrinsic>(BBI)) 1990 ++BBI; 1991 if (&*BBI != BI) 1992 return false; 1993 1994 1995 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1996 if (CE->canTrap()) 1997 return false; 1998 1999 int PBIOp, BIOp; 2000 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2001 PBIOp = BIOp = 0; 2002 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2003 PBIOp = 0, BIOp = 1; 2004 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2005 PBIOp = 1, BIOp = 0; 2006 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2007 PBIOp = BIOp = 1; 2008 else 2009 return false; 2010 2011 // Check to make sure that the other destination of this branch 2012 // isn't BB itself. If so, this is an infinite loop that will 2013 // keep getting unwound. 2014 if (PBI->getSuccessor(PBIOp) == BB) 2015 return false; 2016 2017 // Do not perform this transformation if it would require 2018 // insertion of a large number of select instructions. For targets 2019 // without predication/cmovs, this is a big pessimization. 2020 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2021 2022 unsigned NumPhis = 0; 2023 for (BasicBlock::iterator II = CommonDest->begin(); 2024 isa<PHINode>(II); ++II, ++NumPhis) 2025 if (NumPhis > 2) // Disable this xform. 2026 return false; 2027 2028 // Finally, if everything is ok, fold the branches to logical ops. 2029 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2030 2031 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2032 << "AND: " << *BI->getParent()); 2033 2034 2035 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2036 // branch in it, where one edge (OtherDest) goes back to itself but the other 2037 // exits. We don't *know* that the program avoids the infinite loop 2038 // (even though that seems likely). If we do this xform naively, we'll end up 2039 // recursively unpeeling the loop. Since we know that (after the xform is 2040 // done) that the block *is* infinite if reached, we just make it an obviously 2041 // infinite loop with no cond branch. 2042 if (OtherDest == BB) { 2043 // Insert it at the end of the function, because it's either code, 2044 // or it won't matter if it's hot. :) 2045 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2046 "infloop", BB->getParent()); 2047 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2048 OtherDest = InfLoopBlock; 2049 } 2050 2051 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2052 2053 // BI may have other predecessors. Because of this, we leave 2054 // it alone, but modify PBI. 2055 2056 // Make sure we get to CommonDest on True&True directions. 2057 Value *PBICond = PBI->getCondition(); 2058 IRBuilder<true, NoFolder> Builder(PBI); 2059 if (PBIOp) 2060 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2061 2062 Value *BICond = BI->getCondition(); 2063 if (BIOp) 2064 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2065 2066 // Merge the conditions. 2067 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2068 2069 // Modify PBI to branch on the new condition to the new dests. 2070 PBI->setCondition(Cond); 2071 PBI->setSuccessor(0, CommonDest); 2072 PBI->setSuccessor(1, OtherDest); 2073 2074 // OtherDest may have phi nodes. If so, add an entry from PBI's 2075 // block that are identical to the entries for BI's block. 2076 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2077 2078 // We know that the CommonDest already had an edge from PBI to 2079 // it. If it has PHIs though, the PHIs may have different 2080 // entries for BB and PBI's BB. If so, insert a select to make 2081 // them agree. 2082 PHINode *PN; 2083 for (BasicBlock::iterator II = CommonDest->begin(); 2084 (PN = dyn_cast<PHINode>(II)); ++II) { 2085 Value *BIV = PN->getIncomingValueForBlock(BB); 2086 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2087 Value *PBIV = PN->getIncomingValue(PBBIdx); 2088 if (BIV != PBIV) { 2089 // Insert a select in PBI to pick the right value. 2090 Value *NV = cast<SelectInst> 2091 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2092 PN->setIncomingValue(PBBIdx, NV); 2093 } 2094 } 2095 2096 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2097 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2098 2099 // This basic block is probably dead. We know it has at least 2100 // one fewer predecessor. 2101 return true; 2102} 2103 2104// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2105// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2106// Takes care of updating the successors and removing the old terminator. 2107// Also makes sure not to introduce new successors by assuming that edges to 2108// non-successor TrueBBs and FalseBBs aren't reachable. 2109static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2110 BasicBlock *TrueBB, BasicBlock *FalseBB){ 2111 // Remove any superfluous successor edges from the CFG. 2112 // First, figure out which successors to preserve. 2113 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2114 // successor. 2115 BasicBlock *KeepEdge1 = TrueBB; 2116 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2117 2118 // Then remove the rest. 2119 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2120 BasicBlock *Succ = OldTerm->getSuccessor(I); 2121 // Make sure only to keep exactly one copy of each edge. 2122 if (Succ == KeepEdge1) 2123 KeepEdge1 = 0; 2124 else if (Succ == KeepEdge2) 2125 KeepEdge2 = 0; 2126 else 2127 Succ->removePredecessor(OldTerm->getParent()); 2128 } 2129 2130 IRBuilder<> Builder(OldTerm); 2131 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2132 2133 // Insert an appropriate new terminator. 2134 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2135 if (TrueBB == FalseBB) 2136 // We were only looking for one successor, and it was present. 2137 // Create an unconditional branch to it. 2138 Builder.CreateBr(TrueBB); 2139 else 2140 // We found both of the successors we were looking for. 2141 // Create a conditional branch sharing the condition of the select. 2142 Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2143 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2144 // Neither of the selected blocks were successors, so this 2145 // terminator must be unreachable. 2146 new UnreachableInst(OldTerm->getContext(), OldTerm); 2147 } else { 2148 // One of the selected values was a successor, but the other wasn't. 2149 // Insert an unconditional branch to the one that was found; 2150 // the edge to the one that wasn't must be unreachable. 2151 if (KeepEdge1 == 0) 2152 // Only TrueBB was found. 2153 Builder.CreateBr(TrueBB); 2154 else 2155 // Only FalseBB was found. 2156 Builder.CreateBr(FalseBB); 2157 } 2158 2159 EraseTerminatorInstAndDCECond(OldTerm); 2160 return true; 2161} 2162 2163// SimplifySwitchOnSelect - Replaces 2164// (switch (select cond, X, Y)) on constant X, Y 2165// with a branch - conditional if X and Y lead to distinct BBs, 2166// unconditional otherwise. 2167static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2168 // Check for constant integer values in the select. 2169 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2170 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2171 if (!TrueVal || !FalseVal) 2172 return false; 2173 2174 // Find the relevant condition and destinations. 2175 Value *Condition = Select->getCondition(); 2176 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2177 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2178 2179 // Perform the actual simplification. 2180 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB); 2181} 2182 2183// SimplifyIndirectBrOnSelect - Replaces 2184// (indirectbr (select cond, blockaddress(@fn, BlockA), 2185// blockaddress(@fn, BlockB))) 2186// with 2187// (br cond, BlockA, BlockB). 2188static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2189 // Check that both operands of the select are block addresses. 2190 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2191 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2192 if (!TBA || !FBA) 2193 return false; 2194 2195 // Extract the actual blocks. 2196 BasicBlock *TrueBB = TBA->getBasicBlock(); 2197 BasicBlock *FalseBB = FBA->getBasicBlock(); 2198 2199 // Perform the actual simplification. 2200 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); 2201} 2202 2203/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2204/// instruction (a seteq/setne with a constant) as the only instruction in a 2205/// block that ends with an uncond branch. We are looking for a very specific 2206/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2207/// this case, we merge the first two "or's of icmp" into a switch, but then the 2208/// default value goes to an uncond block with a seteq in it, we get something 2209/// like: 2210/// 2211/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2212/// DEFAULT: 2213/// %tmp = icmp eq i8 %A, 92 2214/// br label %end 2215/// end: 2216/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2217/// 2218/// We prefer to split the edge to 'end' so that there is a true/false entry to 2219/// the PHI, merging the third icmp into the switch. 2220static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 2221 const TargetData *TD, 2222 IRBuilder<> &Builder) { 2223 BasicBlock *BB = ICI->getParent(); 2224 2225 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2226 // complex. 2227 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2228 2229 Value *V = ICI->getOperand(0); 2230 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2231 2232 // The pattern we're looking for is where our only predecessor is a switch on 2233 // 'V' and this block is the default case for the switch. In this case we can 2234 // fold the compared value into the switch to simplify things. 2235 BasicBlock *Pred = BB->getSinglePredecessor(); 2236 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2237 2238 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2239 if (SI->getCondition() != V) 2240 return false; 2241 2242 // If BB is reachable on a non-default case, then we simply know the value of 2243 // V in this block. Substitute it and constant fold the icmp instruction 2244 // away. 2245 if (SI->getDefaultDest() != BB) { 2246 ConstantInt *VVal = SI->findCaseDest(BB); 2247 assert(VVal && "Should have a unique destination value"); 2248 ICI->setOperand(0, VVal); 2249 2250 if (Value *V = SimplifyInstruction(ICI, TD)) { 2251 ICI->replaceAllUsesWith(V); 2252 ICI->eraseFromParent(); 2253 } 2254 // BB is now empty, so it is likely to simplify away. 2255 return SimplifyCFG(BB) | true; 2256 } 2257 2258 // Ok, the block is reachable from the default dest. If the constant we're 2259 // comparing exists in one of the other edges, then we can constant fold ICI 2260 // and zap it. 2261 if (SI->findCaseValue(Cst) != SI->case_default()) { 2262 Value *V; 2263 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2264 V = ConstantInt::getFalse(BB->getContext()); 2265 else 2266 V = ConstantInt::getTrue(BB->getContext()); 2267 2268 ICI->replaceAllUsesWith(V); 2269 ICI->eraseFromParent(); 2270 // BB is now empty, so it is likely to simplify away. 2271 return SimplifyCFG(BB) | true; 2272 } 2273 2274 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2275 // the block. 2276 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2277 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2278 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2279 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2280 return false; 2281 2282 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2283 // true in the PHI. 2284 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2285 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2286 2287 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2288 std::swap(DefaultCst, NewCst); 2289 2290 // Replace ICI (which is used by the PHI for the default value) with true or 2291 // false depending on if it is EQ or NE. 2292 ICI->replaceAllUsesWith(DefaultCst); 2293 ICI->eraseFromParent(); 2294 2295 // Okay, the switch goes to this block on a default value. Add an edge from 2296 // the switch to the merge point on the compared value. 2297 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2298 BB->getParent(), BB); 2299 SI->addCase(Cst, NewBB); 2300 2301 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2302 Builder.SetInsertPoint(NewBB); 2303 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2304 Builder.CreateBr(SuccBlock); 2305 PHIUse->addIncoming(NewCst, NewBB); 2306 return true; 2307} 2308 2309/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2310/// Check to see if it is branching on an or/and chain of icmp instructions, and 2311/// fold it into a switch instruction if so. 2312static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, 2313 IRBuilder<> &Builder) { 2314 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2315 if (Cond == 0) return false; 2316 2317 2318 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2319 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2320 // 'setne's and'ed together, collect them. 2321 Value *CompVal = 0; 2322 std::vector<ConstantInt*> Values; 2323 bool TrueWhenEqual = true; 2324 Value *ExtraCase = 0; 2325 unsigned UsedICmps = 0; 2326 2327 if (Cond->getOpcode() == Instruction::Or) { 2328 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2329 UsedICmps); 2330 } else if (Cond->getOpcode() == Instruction::And) { 2331 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2332 UsedICmps); 2333 TrueWhenEqual = false; 2334 } 2335 2336 // If we didn't have a multiply compared value, fail. 2337 if (CompVal == 0) return false; 2338 2339 // Avoid turning single icmps into a switch. 2340 if (UsedICmps <= 1) 2341 return false; 2342 2343 // There might be duplicate constants in the list, which the switch 2344 // instruction can't handle, remove them now. 2345 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2346 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2347 2348 // If Extra was used, we require at least two switch values to do the 2349 // transformation. A switch with one value is just an cond branch. 2350 if (ExtraCase && Values.size() < 2) return false; 2351 2352 // Figure out which block is which destination. 2353 BasicBlock *DefaultBB = BI->getSuccessor(1); 2354 BasicBlock *EdgeBB = BI->getSuccessor(0); 2355 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2356 2357 BasicBlock *BB = BI->getParent(); 2358 2359 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2360 << " cases into SWITCH. BB is:\n" << *BB); 2361 2362 // If there are any extra values that couldn't be folded into the switch 2363 // then we evaluate them with an explicit branch first. Split the block 2364 // right before the condbr to handle it. 2365 if (ExtraCase) { 2366 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2367 // Remove the uncond branch added to the old block. 2368 TerminatorInst *OldTI = BB->getTerminator(); 2369 Builder.SetInsertPoint(OldTI); 2370 2371 if (TrueWhenEqual) 2372 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2373 else 2374 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2375 2376 OldTI->eraseFromParent(); 2377 2378 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2379 // for the edge we just added. 2380 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2381 2382 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2383 << "\nEXTRABB = " << *BB); 2384 BB = NewBB; 2385 } 2386 2387 Builder.SetInsertPoint(BI); 2388 // Convert pointer to int before we switch. 2389 if (CompVal->getType()->isPointerTy()) { 2390 assert(TD && "Cannot switch on pointer without TargetData"); 2391 CompVal = Builder.CreatePtrToInt(CompVal, 2392 TD->getIntPtrType(CompVal->getContext()), 2393 "magicptr"); 2394 } 2395 2396 // Create the new switch instruction now. 2397 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2398 2399 // Add all of the 'cases' to the switch instruction. 2400 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2401 New->addCase(Values[i], EdgeBB); 2402 2403 // We added edges from PI to the EdgeBB. As such, if there were any 2404 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2405 // the number of edges added. 2406 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2407 isa<PHINode>(BBI); ++BBI) { 2408 PHINode *PN = cast<PHINode>(BBI); 2409 Value *InVal = PN->getIncomingValueForBlock(BB); 2410 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2411 PN->addIncoming(InVal, BB); 2412 } 2413 2414 // Erase the old branch instruction. 2415 EraseTerminatorInstAndDCECond(BI); 2416 2417 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2418 return true; 2419} 2420 2421bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2422 // If this is a trivial landing pad that just continues unwinding the caught 2423 // exception then zap the landing pad, turning its invokes into calls. 2424 BasicBlock *BB = RI->getParent(); 2425 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2426 if (RI->getValue() != LPInst) 2427 // Not a landing pad, or the resume is not unwinding the exception that 2428 // caused control to branch here. 2429 return false; 2430 2431 // Check that there are no other instructions except for debug intrinsics. 2432 BasicBlock::iterator I = LPInst, E = RI; 2433 while (++I != E) 2434 if (!isa<DbgInfoIntrinsic>(I)) 2435 return false; 2436 2437 // Turn all invokes that unwind here into calls and delete the basic block. 2438 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2439 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2440 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2441 // Insert a call instruction before the invoke. 2442 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2443 Call->takeName(II); 2444 Call->setCallingConv(II->getCallingConv()); 2445 Call->setAttributes(II->getAttributes()); 2446 Call->setDebugLoc(II->getDebugLoc()); 2447 2448 // Anything that used the value produced by the invoke instruction now uses 2449 // the value produced by the call instruction. Note that we do this even 2450 // for void functions and calls with no uses so that the callgraph edge is 2451 // updated. 2452 II->replaceAllUsesWith(Call); 2453 BB->removePredecessor(II->getParent()); 2454 2455 // Insert a branch to the normal destination right before the invoke. 2456 BranchInst::Create(II->getNormalDest(), II); 2457 2458 // Finally, delete the invoke instruction! 2459 II->eraseFromParent(); 2460 } 2461 2462 // The landingpad is now unreachable. Zap it. 2463 BB->eraseFromParent(); 2464 return true; 2465} 2466 2467bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2468 BasicBlock *BB = RI->getParent(); 2469 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2470 2471 // Find predecessors that end with branches. 2472 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2473 SmallVector<BranchInst*, 8> CondBranchPreds; 2474 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2475 BasicBlock *P = *PI; 2476 TerminatorInst *PTI = P->getTerminator(); 2477 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2478 if (BI->isUnconditional()) 2479 UncondBranchPreds.push_back(P); 2480 else 2481 CondBranchPreds.push_back(BI); 2482 } 2483 } 2484 2485 // If we found some, do the transformation! 2486 if (!UncondBranchPreds.empty() && DupRet) { 2487 while (!UncondBranchPreds.empty()) { 2488 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2489 DEBUG(dbgs() << "FOLDING: " << *BB 2490 << "INTO UNCOND BRANCH PRED: " << *Pred); 2491 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2492 } 2493 2494 // If we eliminated all predecessors of the block, delete the block now. 2495 if (pred_begin(BB) == pred_end(BB)) 2496 // We know there are no successors, so just nuke the block. 2497 BB->eraseFromParent(); 2498 2499 return true; 2500 } 2501 2502 // Check out all of the conditional branches going to this return 2503 // instruction. If any of them just select between returns, change the 2504 // branch itself into a select/return pair. 2505 while (!CondBranchPreds.empty()) { 2506 BranchInst *BI = CondBranchPreds.pop_back_val(); 2507 2508 // Check to see if the non-BB successor is also a return block. 2509 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2510 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2511 SimplifyCondBranchToTwoReturns(BI, Builder)) 2512 return true; 2513 } 2514 return false; 2515} 2516 2517bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2518 BasicBlock *BB = UI->getParent(); 2519 2520 bool Changed = false; 2521 2522 // If there are any instructions immediately before the unreachable that can 2523 // be removed, do so. 2524 while (UI != BB->begin()) { 2525 BasicBlock::iterator BBI = UI; 2526 --BBI; 2527 // Do not delete instructions that can have side effects which might cause 2528 // the unreachable to not be reachable; specifically, calls and volatile 2529 // operations may have this effect. 2530 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2531 2532 if (BBI->mayHaveSideEffects()) { 2533 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2534 if (SI->isVolatile()) 2535 break; 2536 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2537 if (LI->isVolatile()) 2538 break; 2539 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2540 if (RMWI->isVolatile()) 2541 break; 2542 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2543 if (CXI->isVolatile()) 2544 break; 2545 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2546 !isa<LandingPadInst>(BBI)) { 2547 break; 2548 } 2549 // Note that deleting LandingPad's here is in fact okay, although it 2550 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2551 // all the predecessors of this block will be the unwind edges of Invokes, 2552 // and we can therefore guarantee this block will be erased. 2553 } 2554 2555 // Delete this instruction (any uses are guaranteed to be dead) 2556 if (!BBI->use_empty()) 2557 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2558 BBI->eraseFromParent(); 2559 Changed = true; 2560 } 2561 2562 // If the unreachable instruction is the first in the block, take a gander 2563 // at all of the predecessors of this instruction, and simplify them. 2564 if (&BB->front() != UI) return Changed; 2565 2566 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2567 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2568 TerminatorInst *TI = Preds[i]->getTerminator(); 2569 IRBuilder<> Builder(TI); 2570 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2571 if (BI->isUnconditional()) { 2572 if (BI->getSuccessor(0) == BB) { 2573 new UnreachableInst(TI->getContext(), TI); 2574 TI->eraseFromParent(); 2575 Changed = true; 2576 } 2577 } else { 2578 if (BI->getSuccessor(0) == BB) { 2579 Builder.CreateBr(BI->getSuccessor(1)); 2580 EraseTerminatorInstAndDCECond(BI); 2581 } else if (BI->getSuccessor(1) == BB) { 2582 Builder.CreateBr(BI->getSuccessor(0)); 2583 EraseTerminatorInstAndDCECond(BI); 2584 Changed = true; 2585 } 2586 } 2587 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2588 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2589 i != e; ++i) 2590 if (i.getCaseSuccessor() == BB) { 2591 BB->removePredecessor(SI->getParent()); 2592 SI->removeCase(i); 2593 --i; --e; 2594 Changed = true; 2595 } 2596 // If the default value is unreachable, figure out the most popular 2597 // destination and make it the default. 2598 if (SI->getDefaultDest() == BB) { 2599 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2600 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2601 i != e; ++i) { 2602 std::pair<unsigned, unsigned> &entry = 2603 Popularity[i.getCaseSuccessor()]; 2604 if (entry.first == 0) { 2605 entry.first = 1; 2606 entry.second = i.getCaseIndex(); 2607 } else { 2608 entry.first++; 2609 } 2610 } 2611 2612 // Find the most popular block. 2613 unsigned MaxPop = 0; 2614 unsigned MaxIndex = 0; 2615 BasicBlock *MaxBlock = 0; 2616 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2617 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2618 if (I->second.first > MaxPop || 2619 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2620 MaxPop = I->second.first; 2621 MaxIndex = I->second.second; 2622 MaxBlock = I->first; 2623 } 2624 } 2625 if (MaxBlock) { 2626 // Make this the new default, allowing us to delete any explicit 2627 // edges to it. 2628 SI->setDefaultDest(MaxBlock); 2629 Changed = true; 2630 2631 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2632 // it. 2633 if (isa<PHINode>(MaxBlock->begin())) 2634 for (unsigned i = 0; i != MaxPop-1; ++i) 2635 MaxBlock->removePredecessor(SI->getParent()); 2636 2637 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2638 i != e; ++i) 2639 if (i.getCaseSuccessor() == MaxBlock) { 2640 SI->removeCase(i); 2641 --i; --e; 2642 } 2643 } 2644 } 2645 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2646 if (II->getUnwindDest() == BB) { 2647 // Convert the invoke to a call instruction. This would be a good 2648 // place to note that the call does not throw though. 2649 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2650 II->removeFromParent(); // Take out of symbol table 2651 2652 // Insert the call now... 2653 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2654 Builder.SetInsertPoint(BI); 2655 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2656 Args, II->getName()); 2657 CI->setCallingConv(II->getCallingConv()); 2658 CI->setAttributes(II->getAttributes()); 2659 // If the invoke produced a value, the call does now instead. 2660 II->replaceAllUsesWith(CI); 2661 delete II; 2662 Changed = true; 2663 } 2664 } 2665 } 2666 2667 // If this block is now dead, remove it. 2668 if (pred_begin(BB) == pred_end(BB) && 2669 BB != &BB->getParent()->getEntryBlock()) { 2670 // We know there are no successors, so just nuke the block. 2671 BB->eraseFromParent(); 2672 return true; 2673 } 2674 2675 return Changed; 2676} 2677 2678/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 2679/// integer range comparison into a sub, an icmp and a branch. 2680static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 2681 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 2682 2683 // Make sure all cases point to the same destination and gather the values. 2684 SmallVector<ConstantInt *, 16> Cases; 2685 SwitchInst::CaseIt I = SI->case_begin(); 2686 Cases.push_back(I.getCaseValue()); 2687 SwitchInst::CaseIt PrevI = I++; 2688 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 2689 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 2690 return false; 2691 Cases.push_back(I.getCaseValue()); 2692 } 2693 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 2694 2695 // Sort the case values, then check if they form a range we can transform. 2696 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 2697 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 2698 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 2699 return false; 2700 } 2701 2702 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 2703 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 2704 2705 Value *Sub = SI->getCondition(); 2706 if (!Offset->isNullValue()) 2707 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 2708 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 2709 Builder.CreateCondBr( 2710 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 2711 2712 // Prune obsolete incoming values off the successor's PHI nodes. 2713 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 2714 isa<PHINode>(BBI); ++BBI) { 2715 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 2716 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 2717 } 2718 SI->eraseFromParent(); 2719 2720 return true; 2721} 2722 2723/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 2724/// and use it to remove dead cases. 2725static bool EliminateDeadSwitchCases(SwitchInst *SI) { 2726 Value *Cond = SI->getCondition(); 2727 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 2728 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 2729 ComputeMaskedBits(Cond, KnownZero, KnownOne); 2730 2731 // Gather dead cases. 2732 SmallVector<ConstantInt*, 8> DeadCases; 2733 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 2734 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 2735 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 2736 DeadCases.push_back(I.getCaseValue()); 2737 DEBUG(dbgs() << "SimplifyCFG: switch case '" 2738 << I.getCaseValue() << "' is dead.\n"); 2739 } 2740 } 2741 2742 // Remove dead cases from the switch. 2743 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 2744 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 2745 assert(Case != SI->case_default() && 2746 "Case was not found. Probably mistake in DeadCases forming."); 2747 // Prune unused values from PHI nodes. 2748 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 2749 SI->removeCase(Case); 2750 } 2751 2752 return !DeadCases.empty(); 2753} 2754 2755/// FindPHIForConditionForwarding - If BB would be eligible for simplification 2756/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 2757/// by an unconditional branch), look at the phi node for BB in the successor 2758/// block and see if the incoming value is equal to CaseValue. If so, return 2759/// the phi node, and set PhiIndex to BB's index in the phi node. 2760static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 2761 BasicBlock *BB, 2762 int *PhiIndex) { 2763 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 2764 return NULL; // BB must be empty to be a candidate for simplification. 2765 if (!BB->getSinglePredecessor()) 2766 return NULL; // BB must be dominated by the switch. 2767 2768 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 2769 if (!Branch || !Branch->isUnconditional()) 2770 return NULL; // Terminator must be unconditional branch. 2771 2772 BasicBlock *Succ = Branch->getSuccessor(0); 2773 2774 BasicBlock::iterator I = Succ->begin(); 2775 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 2776 int Idx = PHI->getBasicBlockIndex(BB); 2777 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 2778 2779 Value *InValue = PHI->getIncomingValue(Idx); 2780 if (InValue != CaseValue) continue; 2781 2782 *PhiIndex = Idx; 2783 return PHI; 2784 } 2785 2786 return NULL; 2787} 2788 2789/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 2790/// instruction to a phi node dominated by the switch, if that would mean that 2791/// some of the destination blocks of the switch can be folded away. 2792/// Returns true if a change is made. 2793static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 2794 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 2795 ForwardingNodesMap ForwardingNodes; 2796 2797 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 2798 ConstantInt *CaseValue = I.getCaseValue(); 2799 BasicBlock *CaseDest = I.getCaseSuccessor(); 2800 2801 int PhiIndex; 2802 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 2803 &PhiIndex); 2804 if (!PHI) continue; 2805 2806 ForwardingNodes[PHI].push_back(PhiIndex); 2807 } 2808 2809 bool Changed = false; 2810 2811 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 2812 E = ForwardingNodes.end(); I != E; ++I) { 2813 PHINode *Phi = I->first; 2814 SmallVector<int,4> &Indexes = I->second; 2815 2816 if (Indexes.size() < 2) continue; 2817 2818 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 2819 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 2820 Changed = true; 2821 } 2822 2823 return Changed; 2824} 2825 2826bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 2827 // If this switch is too complex to want to look at, ignore it. 2828 if (!isValueEqualityComparison(SI)) 2829 return false; 2830 2831 BasicBlock *BB = SI->getParent(); 2832 2833 // If we only have one predecessor, and if it is a branch on this value, 2834 // see if that predecessor totally determines the outcome of this switch. 2835 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2836 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 2837 return SimplifyCFG(BB) | true; 2838 2839 Value *Cond = SI->getCondition(); 2840 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 2841 if (SimplifySwitchOnSelect(SI, Select)) 2842 return SimplifyCFG(BB) | true; 2843 2844 // If the block only contains the switch, see if we can fold the block 2845 // away into any preds. 2846 BasicBlock::iterator BBI = BB->begin(); 2847 // Ignore dbg intrinsics. 2848 while (isa<DbgInfoIntrinsic>(BBI)) 2849 ++BBI; 2850 if (SI == &*BBI) 2851 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 2852 return SimplifyCFG(BB) | true; 2853 2854 // Try to transform the switch into an icmp and a branch. 2855 if (TurnSwitchRangeIntoICmp(SI, Builder)) 2856 return SimplifyCFG(BB) | true; 2857 2858 // Remove unreachable cases. 2859 if (EliminateDeadSwitchCases(SI)) 2860 return SimplifyCFG(BB) | true; 2861 2862 if (ForwardSwitchConditionToPHI(SI)) 2863 return SimplifyCFG(BB) | true; 2864 2865 return false; 2866} 2867 2868bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 2869 BasicBlock *BB = IBI->getParent(); 2870 bool Changed = false; 2871 2872 // Eliminate redundant destinations. 2873 SmallPtrSet<Value *, 8> Succs; 2874 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 2875 BasicBlock *Dest = IBI->getDestination(i); 2876 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 2877 Dest->removePredecessor(BB); 2878 IBI->removeDestination(i); 2879 --i; --e; 2880 Changed = true; 2881 } 2882 } 2883 2884 if (IBI->getNumDestinations() == 0) { 2885 // If the indirectbr has no successors, change it to unreachable. 2886 new UnreachableInst(IBI->getContext(), IBI); 2887 EraseTerminatorInstAndDCECond(IBI); 2888 return true; 2889 } 2890 2891 if (IBI->getNumDestinations() == 1) { 2892 // If the indirectbr has one successor, change it to a direct branch. 2893 BranchInst::Create(IBI->getDestination(0), IBI); 2894 EraseTerminatorInstAndDCECond(IBI); 2895 return true; 2896 } 2897 2898 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 2899 if (SimplifyIndirectBrOnSelect(IBI, SI)) 2900 return SimplifyCFG(BB) | true; 2901 } 2902 return Changed; 2903} 2904 2905bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 2906 BasicBlock *BB = BI->getParent(); 2907 2908 // If the Terminator is the only non-phi instruction, simplify the block. 2909 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 2910 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 2911 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 2912 return true; 2913 2914 // If the only instruction in the block is a seteq/setne comparison 2915 // against a constant, try to simplify the block. 2916 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 2917 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 2918 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 2919 ; 2920 if (I->isTerminator() && 2921 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) 2922 return true; 2923 } 2924 2925 // If this basic block is ONLY a compare and a branch, and if a predecessor 2926 // branches to us and our successor, fold the comparison into the 2927 // predecessor and use logical operations to update the incoming value 2928 // for PHI nodes in common successor. 2929 if (FoldBranchToCommonDest(BI)) 2930 return SimplifyCFG(BB) | true; 2931 return false; 2932} 2933 2934 2935bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 2936 BasicBlock *BB = BI->getParent(); 2937 2938 // Conditional branch 2939 if (isValueEqualityComparison(BI)) { 2940 // If we only have one predecessor, and if it is a branch on this value, 2941 // see if that predecessor totally determines the outcome of this 2942 // switch. 2943 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2944 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 2945 return SimplifyCFG(BB) | true; 2946 2947 // This block must be empty, except for the setcond inst, if it exists. 2948 // Ignore dbg intrinsics. 2949 BasicBlock::iterator I = BB->begin(); 2950 // Ignore dbg intrinsics. 2951 while (isa<DbgInfoIntrinsic>(I)) 2952 ++I; 2953 if (&*I == BI) { 2954 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 2955 return SimplifyCFG(BB) | true; 2956 } else if (&*I == cast<Instruction>(BI->getCondition())){ 2957 ++I; 2958 // Ignore dbg intrinsics. 2959 while (isa<DbgInfoIntrinsic>(I)) 2960 ++I; 2961 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 2962 return SimplifyCFG(BB) | true; 2963 } 2964 } 2965 2966 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 2967 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 2968 return true; 2969 2970 // If this basic block is ONLY a compare and a branch, and if a predecessor 2971 // branches to us and one of our successors, fold the comparison into the 2972 // predecessor and use logical operations to pick the right destination. 2973 if (FoldBranchToCommonDest(BI)) 2974 return SimplifyCFG(BB) | true; 2975 2976 // We have a conditional branch to two blocks that are only reachable 2977 // from BI. We know that the condbr dominates the two blocks, so see if 2978 // there is any identical code in the "then" and "else" blocks. If so, we 2979 // can hoist it up to the branching block. 2980 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 2981 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2982 if (HoistThenElseCodeToIf(BI)) 2983 return SimplifyCFG(BB) | true; 2984 } else { 2985 // If Successor #1 has multiple preds, we may be able to conditionally 2986 // execute Successor #0 if it branches to successor #1. 2987 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 2988 if (Succ0TI->getNumSuccessors() == 1 && 2989 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 2990 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 2991 return SimplifyCFG(BB) | true; 2992 } 2993 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2994 // If Successor #0 has multiple preds, we may be able to conditionally 2995 // execute Successor #1 if it branches to successor #0. 2996 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 2997 if (Succ1TI->getNumSuccessors() == 1 && 2998 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 2999 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3000 return SimplifyCFG(BB) | true; 3001 } 3002 3003 // If this is a branch on a phi node in the current block, thread control 3004 // through this block if any PHI node entries are constants. 3005 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3006 if (PN->getParent() == BI->getParent()) 3007 if (FoldCondBranchOnPHI(BI, TD)) 3008 return SimplifyCFG(BB) | true; 3009 3010 // Scan predecessor blocks for conditional branches. 3011 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3012 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3013 if (PBI != BI && PBI->isConditional()) 3014 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3015 return SimplifyCFG(BB) | true; 3016 3017 return false; 3018} 3019 3020/// Check if passing a value to an instruction will cause undefined behavior. 3021static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3022 Constant *C = dyn_cast<Constant>(V); 3023 if (!C) 3024 return false; 3025 3026 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time 3027 return false; 3028 3029 if (C->isNullValue()) { 3030 Instruction *Use = I->use_back(); 3031 3032 // Now make sure that there are no instructions in between that can alter 3033 // control flow (eg. calls) 3034 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3035 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3036 return false; 3037 3038 // Look through GEPs. A load from a GEP derived from NULL is still undefined 3039 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 3040 if (GEP->getPointerOperand() == I) 3041 return passingValueIsAlwaysUndefined(V, GEP); 3042 3043 // Look through bitcasts. 3044 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 3045 return passingValueIsAlwaysUndefined(V, BC); 3046 3047 // Load from null is undefined. 3048 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 3049 return LI->getPointerAddressSpace() == 0; 3050 3051 // Store to null is undefined. 3052 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 3053 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 3054 } 3055 return false; 3056} 3057 3058/// If BB has an incoming value that will always trigger undefined behavior 3059/// (eg. null pointer dereference), remove the branch leading here. 3060static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 3061 for (BasicBlock::iterator i = BB->begin(); 3062 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 3063 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 3064 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 3065 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 3066 IRBuilder<> Builder(T); 3067 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 3068 BB->removePredecessor(PHI->getIncomingBlock(i)); 3069 // Turn uncoditional branches into unreachables and remove the dead 3070 // destination from conditional branches. 3071 if (BI->isUnconditional()) 3072 Builder.CreateUnreachable(); 3073 else 3074 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 3075 BI->getSuccessor(0)); 3076 BI->eraseFromParent(); 3077 return true; 3078 } 3079 // TODO: SwitchInst. 3080 } 3081 3082 return false; 3083} 3084 3085bool SimplifyCFGOpt::run(BasicBlock *BB) { 3086 bool Changed = false; 3087 3088 assert(BB && BB->getParent() && "Block not embedded in function!"); 3089 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 3090 3091 // Remove basic blocks that have no predecessors (except the entry block)... 3092 // or that just have themself as a predecessor. These are unreachable. 3093 if ((pred_begin(BB) == pred_end(BB) && 3094 BB != &BB->getParent()->getEntryBlock()) || 3095 BB->getSinglePredecessor() == BB) { 3096 DEBUG(dbgs() << "Removing BB: \n" << *BB); 3097 DeleteDeadBlock(BB); 3098 return true; 3099 } 3100 3101 // Check to see if we can constant propagate this terminator instruction 3102 // away... 3103 Changed |= ConstantFoldTerminator(BB, true); 3104 3105 // Check for and eliminate duplicate PHI nodes in this block. 3106 Changed |= EliminateDuplicatePHINodes(BB); 3107 3108 // Check for and remove branches that will always cause undefined behavior. 3109 Changed |= removeUndefIntroducingPredecessor(BB); 3110 3111 // Merge basic blocks into their predecessor if there is only one distinct 3112 // pred, and if there is only one distinct successor of the predecessor, and 3113 // if there are no PHI nodes. 3114 // 3115 if (MergeBlockIntoPredecessor(BB)) 3116 return true; 3117 3118 IRBuilder<> Builder(BB); 3119 3120 // If there is a trivial two-entry PHI node in this basic block, and we can 3121 // eliminate it, do so now. 3122 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 3123 if (PN->getNumIncomingValues() == 2) 3124 Changed |= FoldTwoEntryPHINode(PN, TD); 3125 3126 Builder.SetInsertPoint(BB->getTerminator()); 3127 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 3128 if (BI->isUnconditional()) { 3129 if (SimplifyUncondBranch(BI, Builder)) return true; 3130 } else { 3131 if (SimplifyCondBranch(BI, Builder)) return true; 3132 } 3133 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 3134 if (SimplifyReturn(RI, Builder)) return true; 3135 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 3136 if (SimplifyResume(RI, Builder)) return true; 3137 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 3138 if (SimplifySwitch(SI, Builder)) return true; 3139 } else if (UnreachableInst *UI = 3140 dyn_cast<UnreachableInst>(BB->getTerminator())) { 3141 if (SimplifyUnreachable(UI)) return true; 3142 } else if (IndirectBrInst *IBI = 3143 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 3144 if (SimplifyIndirectBr(IBI)) return true; 3145 } 3146 3147 return Changed; 3148} 3149 3150/// SimplifyCFG - This function is used to do simplification of a CFG. For 3151/// example, it adjusts branches to branches to eliminate the extra hop, it 3152/// eliminates unreachable basic blocks, and does other "peephole" optimization 3153/// of the CFG. It returns true if a modification was made. 3154/// 3155bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 3156 return SimplifyCFGOpt(TD).run(BB); 3157} 3158