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/ADT/DenseMap.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SetVector.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Analysis/TargetTransformInfo.h" 24 #include "llvm/Analysis/ValueTracking.h" 25 #include "llvm/IR/Constants.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/DerivedTypes.h" 28 #include "llvm/IR/GlobalVariable.h" 29 #include "llvm/IR/IRBuilder.h" 30 #include "llvm/IR/Instructions.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/LLVMContext.h" 33 #include "llvm/IR/MDBuilder.h" 34 #include "llvm/IR/Metadata.h" 35 #include "llvm/IR/Module.h" 36 #include "llvm/IR/Operator.h" 37 #include "llvm/IR/Type.h" 38 #include "llvm/Support/CFG.h" 39 #include "llvm/Support/CommandLine.h" 40 #include "llvm/Support/ConstantRange.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/NoFolder.h" 43 #include "llvm/Support/raw_ostream.h" 44 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 45 #include <algorithm> 46 #include <map> 47 #include <set> 48 using namespace llvm; 49 50 static cl::opt<unsigned> 51 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 52 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 53 54 static cl::opt<bool> 55 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 56 cl::desc("Duplicate return instructions into unconditional branches")); 57 58 static cl::opt<bool> 59 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 60 cl::desc("Sink common instructions down to the end block")); 61 62 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 63 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 64 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 65 STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 66 67 namespace { 68 /// ValueEqualityComparisonCase - Represents a case of a switch. 69 struct ValueEqualityComparisonCase { 70 ConstantInt *Value; 71 BasicBlock *Dest; 72 73 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 74 : Value(Value), Dest(Dest) {} 75 76 bool operator<(ValueEqualityComparisonCase RHS) const { 77 // Comparing pointers is ok as we only rely on the order for uniquing. 78 return Value < RHS.Value; 79 } 80 81 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 82 }; 83 84 class SimplifyCFGOpt { 85 const TargetTransformInfo &TTI; 86 const DataLayout *const TD; 87 88 Value *isValueEqualityComparison(TerminatorInst *TI); 89 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 90 std::vector<ValueEqualityComparisonCase> &Cases); 91 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 92 BasicBlock *Pred, 93 IRBuilder<> &Builder); 94 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 95 IRBuilder<> &Builder); 96 97 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 98 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 99 bool SimplifyUnreachable(UnreachableInst *UI); 100 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 101 bool SimplifyIndirectBr(IndirectBrInst *IBI); 102 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 103 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 104 105 public: 106 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD) 107 : TTI(TTI), TD(TD) {} 108 bool run(BasicBlock *BB); 109 }; 110 } 111 112 /// SafeToMergeTerminators - Return true if it is safe to merge these two 113 /// terminator instructions together. 114 /// 115 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 116 if (SI1 == SI2) return false; // Can't merge with self! 117 118 // It is not safe to merge these two switch instructions if they have a common 119 // successor, and if that successor has a PHI node, and if *that* PHI node has 120 // conflicting incoming values from the two switch blocks. 121 BasicBlock *SI1BB = SI1->getParent(); 122 BasicBlock *SI2BB = SI2->getParent(); 123 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 124 125 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 126 if (SI1Succs.count(*I)) 127 for (BasicBlock::iterator BBI = (*I)->begin(); 128 isa<PHINode>(BBI); ++BBI) { 129 PHINode *PN = cast<PHINode>(BBI); 130 if (PN->getIncomingValueForBlock(SI1BB) != 131 PN->getIncomingValueForBlock(SI2BB)) 132 return false; 133 } 134 135 return true; 136 } 137 138 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable 139 /// to merge these two terminator instructions together, where SI1 is an 140 /// unconditional branch. PhiNodes will store all PHI nodes in common 141 /// successors. 142 /// 143 static bool isProfitableToFoldUnconditional(BranchInst *SI1, 144 BranchInst *SI2, 145 Instruction *Cond, 146 SmallVectorImpl<PHINode*> &PhiNodes) { 147 if (SI1 == SI2) return false; // Can't merge with self! 148 assert(SI1->isUnconditional() && SI2->isConditional()); 149 150 // We fold the unconditional branch if we can easily update all PHI nodes in 151 // common successors: 152 // 1> We have a constant incoming value for the conditional branch; 153 // 2> We have "Cond" as the incoming value for the unconditional branch; 154 // 3> SI2->getCondition() and Cond have same operands. 155 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 156 if (!Ci2) return false; 157 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 158 Cond->getOperand(1) == Ci2->getOperand(1)) && 159 !(Cond->getOperand(0) == Ci2->getOperand(1) && 160 Cond->getOperand(1) == Ci2->getOperand(0))) 161 return false; 162 163 BasicBlock *SI1BB = SI1->getParent(); 164 BasicBlock *SI2BB = SI2->getParent(); 165 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 166 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 167 if (SI1Succs.count(*I)) 168 for (BasicBlock::iterator BBI = (*I)->begin(); 169 isa<PHINode>(BBI); ++BBI) { 170 PHINode *PN = cast<PHINode>(BBI); 171 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 172 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 173 return false; 174 PhiNodes.push_back(PN); 175 } 176 return true; 177 } 178 179 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 180 /// now be entries in it from the 'NewPred' block. The values that will be 181 /// flowing into the PHI nodes will be the same as those coming in from 182 /// ExistPred, an existing predecessor of Succ. 183 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 184 BasicBlock *ExistPred) { 185 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 186 187 PHINode *PN; 188 for (BasicBlock::iterator I = Succ->begin(); 189 (PN = dyn_cast<PHINode>(I)); ++I) 190 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 191 } 192 193 194 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at 195 /// least one PHI node in it), check to see if the merge at this block is due 196 /// to an "if condition". If so, return the boolean condition that determines 197 /// which entry into BB will be taken. Also, return by references the block 198 /// that will be entered from if the condition is true, and the block that will 199 /// be entered if the condition is false. 200 /// 201 /// This does no checking to see if the true/false blocks have large or unsavory 202 /// instructions in them. 203 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 204 BasicBlock *&IfFalse) { 205 PHINode *SomePHI = cast<PHINode>(BB->begin()); 206 assert(SomePHI->getNumIncomingValues() == 2 && 207 "Function can only handle blocks with 2 predecessors!"); 208 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 209 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 210 211 // We can only handle branches. Other control flow will be lowered to 212 // branches if possible anyway. 213 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 214 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 215 if (Pred1Br == 0 || Pred2Br == 0) 216 return 0; 217 218 // Eliminate code duplication by ensuring that Pred1Br is conditional if 219 // either are. 220 if (Pred2Br->isConditional()) { 221 // If both branches are conditional, we don't have an "if statement". In 222 // reality, we could transform this case, but since the condition will be 223 // required anyway, we stand no chance of eliminating it, so the xform is 224 // probably not profitable. 225 if (Pred1Br->isConditional()) 226 return 0; 227 228 std::swap(Pred1, Pred2); 229 std::swap(Pred1Br, Pred2Br); 230 } 231 232 if (Pred1Br->isConditional()) { 233 // The only thing we have to watch out for here is to make sure that Pred2 234 // doesn't have incoming edges from other blocks. If it does, the condition 235 // doesn't dominate BB. 236 if (Pred2->getSinglePredecessor() == 0) 237 return 0; 238 239 // If we found a conditional branch predecessor, make sure that it branches 240 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 241 if (Pred1Br->getSuccessor(0) == BB && 242 Pred1Br->getSuccessor(1) == Pred2) { 243 IfTrue = Pred1; 244 IfFalse = Pred2; 245 } else if (Pred1Br->getSuccessor(0) == Pred2 && 246 Pred1Br->getSuccessor(1) == BB) { 247 IfTrue = Pred2; 248 IfFalse = Pred1; 249 } else { 250 // We know that one arm of the conditional goes to BB, so the other must 251 // go somewhere unrelated, and this must not be an "if statement". 252 return 0; 253 } 254 255 return Pred1Br->getCondition(); 256 } 257 258 // Ok, if we got here, both predecessors end with an unconditional branch to 259 // BB. Don't panic! If both blocks only have a single (identical) 260 // predecessor, and THAT is a conditional branch, then we're all ok! 261 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 262 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 263 return 0; 264 265 // Otherwise, if this is a conditional branch, then we can use it! 266 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 267 if (BI == 0) return 0; 268 269 assert(BI->isConditional() && "Two successors but not conditional?"); 270 if (BI->getSuccessor(0) == Pred1) { 271 IfTrue = Pred1; 272 IfFalse = Pred2; 273 } else { 274 IfTrue = Pred2; 275 IfFalse = Pred1; 276 } 277 return BI->getCondition(); 278 } 279 280 /// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 281 /// given instruction, which is assumed to be safe to speculate. 1 means 282 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 283 static unsigned ComputeSpeculationCost(const User *I) { 284 assert(isSafeToSpeculativelyExecute(I) && 285 "Instruction is not safe to speculatively execute!"); 286 switch (Operator::getOpcode(I)) { 287 default: 288 // In doubt, be conservative. 289 return UINT_MAX; 290 case Instruction::GetElementPtr: 291 // GEPs are cheap if all indices are constant. 292 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 293 return UINT_MAX; 294 return 1; 295 case Instruction::Load: 296 case Instruction::Add: 297 case Instruction::Sub: 298 case Instruction::And: 299 case Instruction::Or: 300 case Instruction::Xor: 301 case Instruction::Shl: 302 case Instruction::LShr: 303 case Instruction::AShr: 304 case Instruction::ICmp: 305 case Instruction::Trunc: 306 case Instruction::ZExt: 307 case Instruction::SExt: 308 return 1; // These are all cheap. 309 310 case Instruction::Call: 311 case Instruction::Select: 312 return 2; 313 } 314 } 315 316 /// DominatesMergePoint - If we have a merge point of an "if condition" as 317 /// accepted above, return true if the specified value dominates the block. We 318 /// don't handle the true generality of domination here, just a special case 319 /// which works well enough for us. 320 /// 321 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 322 /// see if V (which must be an instruction) and its recursive operands 323 /// that do not dominate BB have a combined cost lower than CostRemaining and 324 /// are non-trapping. If both are true, the instruction is inserted into the 325 /// set and true is returned. 326 /// 327 /// The cost for most non-trapping instructions is defined as 1 except for 328 /// Select whose cost is 2. 329 /// 330 /// After this function returns, CostRemaining is decreased by the cost of 331 /// V plus its non-dominating operands. If that cost is greater than 332 /// CostRemaining, false is returned and CostRemaining is undefined. 333 static bool DominatesMergePoint(Value *V, BasicBlock *BB, 334 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 335 unsigned &CostRemaining) { 336 Instruction *I = dyn_cast<Instruction>(V); 337 if (!I) { 338 // Non-instructions all dominate instructions, but not all constantexprs 339 // can be executed unconditionally. 340 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 341 if (C->canTrap()) 342 return false; 343 return true; 344 } 345 BasicBlock *PBB = I->getParent(); 346 347 // We don't want to allow weird loops that might have the "if condition" in 348 // the bottom of this block. 349 if (PBB == BB) return false; 350 351 // If this instruction is defined in a block that contains an unconditional 352 // branch to BB, then it must be in the 'conditional' part of the "if 353 // statement". If not, it definitely dominates the region. 354 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 355 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 356 return true; 357 358 // If we aren't allowing aggressive promotion anymore, then don't consider 359 // instructions in the 'if region'. 360 if (AggressiveInsts == 0) return false; 361 362 // If we have seen this instruction before, don't count it again. 363 if (AggressiveInsts->count(I)) return true; 364 365 // Okay, it looks like the instruction IS in the "condition". Check to 366 // see if it's a cheap instruction to unconditionally compute, and if it 367 // only uses stuff defined outside of the condition. If so, hoist it out. 368 if (!isSafeToSpeculativelyExecute(I)) 369 return false; 370 371 unsigned Cost = ComputeSpeculationCost(I); 372 373 if (Cost > CostRemaining) 374 return false; 375 376 CostRemaining -= Cost; 377 378 // Okay, we can only really hoist these out if their operands do 379 // not take us over the cost threshold. 380 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 381 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 382 return false; 383 // Okay, it's safe to do this! Remember this instruction. 384 AggressiveInsts->insert(I); 385 return true; 386 } 387 388 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 389 /// and PointerNullValue. Return NULL if value is not a constant int. 390 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 391 // Normal constant int. 392 ConstantInt *CI = dyn_cast<ConstantInt>(V); 393 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 394 return CI; 395 396 // This is some kind of pointer constant. Turn it into a pointer-sized 397 // ConstantInt if possible. 398 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); 399 400 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 401 if (isa<ConstantPointerNull>(V)) 402 return ConstantInt::get(PtrTy, 0); 403 404 // IntToPtr const int. 405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 406 if (CE->getOpcode() == Instruction::IntToPtr) 407 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 408 // The constant is very likely to have the right type already. 409 if (CI->getType() == PtrTy) 410 return CI; 411 else 412 return cast<ConstantInt> 413 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 414 } 415 return 0; 416 } 417 418 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 419 /// collection of icmp eq/ne instructions that compare a value against a 420 /// constant, return the value being compared, and stick the constant into the 421 /// Values vector. 422 static Value * 423 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 424 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 425 Instruction *I = dyn_cast<Instruction>(V); 426 if (I == 0) return 0; 427 428 // If this is an icmp against a constant, handle this as one of the cases. 429 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 430 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 431 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 432 UsedICmps++; 433 Vals.push_back(C); 434 return I->getOperand(0); 435 } 436 437 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 438 // the set. 439 ConstantRange Span = 440 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 441 442 // If this is an and/!= check then we want to optimize "x ugt 2" into 443 // x != 0 && x != 1. 444 if (!isEQ) 445 Span = Span.inverse(); 446 447 // If there are a ton of values, we don't want to make a ginormous switch. 448 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 449 return 0; 450 451 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 452 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 453 UsedICmps++; 454 return I->getOperand(0); 455 } 456 return 0; 457 } 458 459 // Otherwise, we can only handle an | or &, depending on isEQ. 460 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 461 return 0; 462 463 unsigned NumValsBeforeLHS = Vals.size(); 464 unsigned UsedICmpsBeforeLHS = UsedICmps; 465 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 466 isEQ, UsedICmps)) { 467 unsigned NumVals = Vals.size(); 468 unsigned UsedICmpsBeforeRHS = UsedICmps; 469 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 470 isEQ, UsedICmps)) { 471 if (LHS == RHS) 472 return LHS; 473 Vals.resize(NumVals); 474 UsedICmps = UsedICmpsBeforeRHS; 475 } 476 477 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 478 // set it and return success. 479 if (Extra == 0 || Extra == I->getOperand(1)) { 480 Extra = I->getOperand(1); 481 return LHS; 482 } 483 484 Vals.resize(NumValsBeforeLHS); 485 UsedICmps = UsedICmpsBeforeLHS; 486 return 0; 487 } 488 489 // If the LHS can't be folded in, but Extra is available and RHS can, try to 490 // use LHS as Extra. 491 if (Extra == 0 || Extra == I->getOperand(0)) { 492 Value *OldExtra = Extra; 493 Extra = I->getOperand(0); 494 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 495 isEQ, UsedICmps)) 496 return RHS; 497 assert(Vals.size() == NumValsBeforeLHS); 498 Extra = OldExtra; 499 } 500 501 return 0; 502 } 503 504 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 505 Instruction *Cond = 0; 506 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 507 Cond = dyn_cast<Instruction>(SI->getCondition()); 508 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 509 if (BI->isConditional()) 510 Cond = dyn_cast<Instruction>(BI->getCondition()); 511 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 512 Cond = dyn_cast<Instruction>(IBI->getAddress()); 513 } 514 515 TI->eraseFromParent(); 516 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 517 } 518 519 /// isValueEqualityComparison - Return true if the specified terminator checks 520 /// to see if a value is equal to constant integer value. 521 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 522 Value *CV = 0; 523 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 524 // Do not permit merging of large switch instructions into their 525 // predecessors unless there is only one predecessor. 526 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 527 pred_end(SI->getParent())) <= 128) 528 CV = SI->getCondition(); 529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 530 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 531 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 532 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 533 ICI->getPredicate() == ICmpInst::ICMP_NE) && 534 GetConstantInt(ICI->getOperand(1), TD)) 535 CV = ICI->getOperand(0); 536 537 // Unwrap any lossless ptrtoint cast. 538 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 539 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 540 CV = PTII->getOperand(0); 541 return CV; 542 } 543 544 /// GetValueEqualityComparisonCases - Given a value comparison instruction, 545 /// decode all of the 'cases' that it represents and return the 'default' block. 546 BasicBlock *SimplifyCFGOpt:: 547 GetValueEqualityComparisonCases(TerminatorInst *TI, 548 std::vector<ValueEqualityComparisonCase> 549 &Cases) { 550 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 551 Cases.reserve(SI->getNumCases()); 552 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 553 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 554 i.getCaseSuccessor())); 555 return SI->getDefaultDest(); 556 } 557 558 BranchInst *BI = cast<BranchInst>(TI); 559 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 560 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 561 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 562 TD), 563 Succ)); 564 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 565 } 566 567 568 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 569 /// in the list that match the specified block. 570 static void EliminateBlockCases(BasicBlock *BB, 571 std::vector<ValueEqualityComparisonCase> &Cases) { 572 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 573 } 574 575 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 576 /// well. 577 static bool 578 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 579 std::vector<ValueEqualityComparisonCase > &C2) { 580 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 581 582 // Make V1 be smaller than V2. 583 if (V1->size() > V2->size()) 584 std::swap(V1, V2); 585 586 if (V1->size() == 0) return false; 587 if (V1->size() == 1) { 588 // Just scan V2. 589 ConstantInt *TheVal = (*V1)[0].Value; 590 for (unsigned i = 0, e = V2->size(); i != e; ++i) 591 if (TheVal == (*V2)[i].Value) 592 return true; 593 } 594 595 // Otherwise, just sort both lists and compare element by element. 596 array_pod_sort(V1->begin(), V1->end()); 597 array_pod_sort(V2->begin(), V2->end()); 598 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 599 while (i1 != e1 && i2 != e2) { 600 if ((*V1)[i1].Value == (*V2)[i2].Value) 601 return true; 602 if ((*V1)[i1].Value < (*V2)[i2].Value) 603 ++i1; 604 else 605 ++i2; 606 } 607 return false; 608 } 609 610 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 611 /// terminator instruction and its block is known to only have a single 612 /// predecessor block, check to see if that predecessor is also a value 613 /// comparison with the same value, and if that comparison determines the 614 /// outcome of this comparison. If so, simplify TI. This does a very limited 615 /// form of jump threading. 616 bool SimplifyCFGOpt:: 617 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 618 BasicBlock *Pred, 619 IRBuilder<> &Builder) { 620 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 621 if (!PredVal) return false; // Not a value comparison in predecessor. 622 623 Value *ThisVal = isValueEqualityComparison(TI); 624 assert(ThisVal && "This isn't a value comparison!!"); 625 if (ThisVal != PredVal) return false; // Different predicates. 626 627 // TODO: Preserve branch weight metadata, similarly to how 628 // FoldValueComparisonIntoPredecessors preserves it. 629 630 // Find out information about when control will move from Pred to TI's block. 631 std::vector<ValueEqualityComparisonCase> PredCases; 632 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 633 PredCases); 634 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 635 636 // Find information about how control leaves this block. 637 std::vector<ValueEqualityComparisonCase> ThisCases; 638 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 639 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 640 641 // If TI's block is the default block from Pred's comparison, potentially 642 // simplify TI based on this knowledge. 643 if (PredDef == TI->getParent()) { 644 // If we are here, we know that the value is none of those cases listed in 645 // PredCases. If there are any cases in ThisCases that are in PredCases, we 646 // can simplify TI. 647 if (!ValuesOverlap(PredCases, ThisCases)) 648 return false; 649 650 if (isa<BranchInst>(TI)) { 651 // Okay, one of the successors of this condbr is dead. Convert it to a 652 // uncond br. 653 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 654 // Insert the new branch. 655 Instruction *NI = Builder.CreateBr(ThisDef); 656 (void) NI; 657 658 // Remove PHI node entries for the dead edge. 659 ThisCases[0].Dest->removePredecessor(TI->getParent()); 660 661 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 662 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 663 664 EraseTerminatorInstAndDCECond(TI); 665 return true; 666 } 667 668 SwitchInst *SI = cast<SwitchInst>(TI); 669 // Okay, TI has cases that are statically dead, prune them away. 670 SmallPtrSet<Constant*, 16> DeadCases; 671 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 672 DeadCases.insert(PredCases[i].Value); 673 674 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 675 << "Through successor TI: " << *TI); 676 677 // Collect branch weights into a vector. 678 SmallVector<uint32_t, 8> Weights; 679 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 680 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 681 if (HasWeight) 682 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 683 ++MD_i) { 684 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 685 assert(CI); 686 Weights.push_back(CI->getValue().getZExtValue()); 687 } 688 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 689 --i; 690 if (DeadCases.count(i.getCaseValue())) { 691 if (HasWeight) { 692 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 693 Weights.pop_back(); 694 } 695 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 696 SI->removeCase(i); 697 } 698 } 699 if (HasWeight && Weights.size() >= 2) 700 SI->setMetadata(LLVMContext::MD_prof, 701 MDBuilder(SI->getParent()->getContext()). 702 createBranchWeights(Weights)); 703 704 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 705 return true; 706 } 707 708 // Otherwise, TI's block must correspond to some matched value. Find out 709 // which value (or set of values) this is. 710 ConstantInt *TIV = 0; 711 BasicBlock *TIBB = TI->getParent(); 712 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 713 if (PredCases[i].Dest == TIBB) { 714 if (TIV != 0) 715 return false; // Cannot handle multiple values coming to this block. 716 TIV = PredCases[i].Value; 717 } 718 assert(TIV && "No edge from pred to succ?"); 719 720 // Okay, we found the one constant that our value can be if we get into TI's 721 // BB. Find out which successor will unconditionally be branched to. 722 BasicBlock *TheRealDest = 0; 723 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 724 if (ThisCases[i].Value == TIV) { 725 TheRealDest = ThisCases[i].Dest; 726 break; 727 } 728 729 // If not handled by any explicit cases, it is handled by the default case. 730 if (TheRealDest == 0) TheRealDest = ThisDef; 731 732 // Remove PHI node entries for dead edges. 733 BasicBlock *CheckEdge = TheRealDest; 734 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 735 if (*SI != CheckEdge) 736 (*SI)->removePredecessor(TIBB); 737 else 738 CheckEdge = 0; 739 740 // Insert the new branch. 741 Instruction *NI = Builder.CreateBr(TheRealDest); 742 (void) NI; 743 744 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 745 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 746 747 EraseTerminatorInstAndDCECond(TI); 748 return true; 749 } 750 751 namespace { 752 /// ConstantIntOrdering - This class implements a stable ordering of constant 753 /// integers that does not depend on their address. This is important for 754 /// applications that sort ConstantInt's to ensure uniqueness. 755 struct ConstantIntOrdering { 756 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 757 return LHS->getValue().ult(RHS->getValue()); 758 } 759 }; 760 } 761 762 static int ConstantIntSortPredicate(const void *P1, const void *P2) { 763 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 764 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 765 if (LHS->getValue().ult(RHS->getValue())) 766 return 1; 767 if (LHS->getValue() == RHS->getValue()) 768 return 0; 769 return -1; 770 } 771 772 static inline bool HasBranchWeights(const Instruction* I) { 773 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 774 if (ProfMD && ProfMD->getOperand(0)) 775 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 776 return MDS->getString().equals("branch_weights"); 777 778 return false; 779 } 780 781 /// Get Weights of a given TerminatorInst, the default weight is at the front 782 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight 783 /// metadata. 784 static void GetBranchWeights(TerminatorInst *TI, 785 SmallVectorImpl<uint64_t> &Weights) { 786 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 787 assert(MD); 788 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 789 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 790 assert(CI); 791 Weights.push_back(CI->getValue().getZExtValue()); 792 } 793 794 // If TI is a conditional eq, the default case is the false case, 795 // and the corresponding branch-weight data is at index 2. We swap the 796 // default weight to be the first entry. 797 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 798 assert(Weights.size() == 2); 799 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 801 std::swap(Weights.front(), Weights.back()); 802 } 803 } 804 805 /// Sees if any of the weights are too big for a uint32_t, and halves all the 806 /// weights if any are. 807 static void FitWeights(MutableArrayRef<uint64_t> Weights) { 808 bool Halve = false; 809 for (unsigned i = 0; i < Weights.size(); ++i) 810 if (Weights[i] > UINT_MAX) { 811 Halve = true; 812 break; 813 } 814 815 if (! Halve) 816 return; 817 818 for (unsigned i = 0; i < Weights.size(); ++i) 819 Weights[i] /= 2; 820 } 821 822 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value 823 /// equality comparison instruction (either a switch or a branch on "X == c"). 824 /// See if any of the predecessors of the terminator block are value comparisons 825 /// on the same value. If so, and if safe to do so, fold them together. 826 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 827 IRBuilder<> &Builder) { 828 BasicBlock *BB = TI->getParent(); 829 Value *CV = isValueEqualityComparison(TI); // CondVal 830 assert(CV && "Not a comparison?"); 831 bool Changed = false; 832 833 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 834 while (!Preds.empty()) { 835 BasicBlock *Pred = Preds.pop_back_val(); 836 837 // See if the predecessor is a comparison with the same value. 838 TerminatorInst *PTI = Pred->getTerminator(); 839 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 840 841 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 842 // Figure out which 'cases' to copy from SI to PSI. 843 std::vector<ValueEqualityComparisonCase> BBCases; 844 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 845 846 std::vector<ValueEqualityComparisonCase> PredCases; 847 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 848 849 // Based on whether the default edge from PTI goes to BB or not, fill in 850 // PredCases and PredDefault with the new switch cases we would like to 851 // build. 852 SmallVector<BasicBlock*, 8> NewSuccessors; 853 854 // Update the branch weight metadata along the way 855 SmallVector<uint64_t, 8> Weights; 856 bool PredHasWeights = HasBranchWeights(PTI); 857 bool SuccHasWeights = HasBranchWeights(TI); 858 859 if (PredHasWeights) { 860 GetBranchWeights(PTI, Weights); 861 // branch-weight metadata is inconsistent here. 862 if (Weights.size() != 1 + PredCases.size()) 863 PredHasWeights = SuccHasWeights = false; 864 } else if (SuccHasWeights) 865 // If there are no predecessor weights but there are successor weights, 866 // populate Weights with 1, which will later be scaled to the sum of 867 // successor's weights 868 Weights.assign(1 + PredCases.size(), 1); 869 870 SmallVector<uint64_t, 8> SuccWeights; 871 if (SuccHasWeights) { 872 GetBranchWeights(TI, SuccWeights); 873 // branch-weight metadata is inconsistent here. 874 if (SuccWeights.size() != 1 + BBCases.size()) 875 PredHasWeights = SuccHasWeights = false; 876 } else if (PredHasWeights) 877 SuccWeights.assign(1 + BBCases.size(), 1); 878 879 if (PredDefault == BB) { 880 // If this is the default destination from PTI, only the edges in TI 881 // that don't occur in PTI, or that branch to BB will be activated. 882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 884 if (PredCases[i].Dest != BB) 885 PTIHandled.insert(PredCases[i].Value); 886 else { 887 // The default destination is BB, we don't need explicit targets. 888 std::swap(PredCases[i], PredCases.back()); 889 890 if (PredHasWeights || SuccHasWeights) { 891 // Increase weight for the default case. 892 Weights[0] += Weights[i+1]; 893 std::swap(Weights[i+1], Weights.back()); 894 Weights.pop_back(); 895 } 896 897 PredCases.pop_back(); 898 --i; --e; 899 } 900 901 // Reconstruct the new switch statement we will be building. 902 if (PredDefault != BBDefault) { 903 PredDefault->removePredecessor(Pred); 904 PredDefault = BBDefault; 905 NewSuccessors.push_back(BBDefault); 906 } 907 908 unsigned CasesFromPred = Weights.size(); 909 uint64_t ValidTotalSuccWeight = 0; 910 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 911 if (!PTIHandled.count(BBCases[i].Value) && 912 BBCases[i].Dest != BBDefault) { 913 PredCases.push_back(BBCases[i]); 914 NewSuccessors.push_back(BBCases[i].Dest); 915 if (SuccHasWeights || PredHasWeights) { 916 // The default weight is at index 0, so weight for the ith case 917 // should be at index i+1. Scale the cases from successor by 918 // PredDefaultWeight (Weights[0]). 919 Weights.push_back(Weights[0] * SuccWeights[i+1]); 920 ValidTotalSuccWeight += SuccWeights[i+1]; 921 } 922 } 923 924 if (SuccHasWeights || PredHasWeights) { 925 ValidTotalSuccWeight += SuccWeights[0]; 926 // Scale the cases from predecessor by ValidTotalSuccWeight. 927 for (unsigned i = 1; i < CasesFromPred; ++i) 928 Weights[i] *= ValidTotalSuccWeight; 929 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 930 Weights[0] *= SuccWeights[0]; 931 } 932 } else { 933 // If this is not the default destination from PSI, only the edges 934 // in SI that occur in PSI with a destination of BB will be 935 // activated. 936 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 937 std::map<ConstantInt*, uint64_t> WeightsForHandled; 938 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 939 if (PredCases[i].Dest == BB) { 940 PTIHandled.insert(PredCases[i].Value); 941 942 if (PredHasWeights || SuccHasWeights) { 943 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 944 std::swap(Weights[i+1], Weights.back()); 945 Weights.pop_back(); 946 } 947 948 std::swap(PredCases[i], PredCases.back()); 949 PredCases.pop_back(); 950 --i; --e; 951 } 952 953 // Okay, now we know which constants were sent to BB from the 954 // predecessor. Figure out where they will all go now. 955 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 956 if (PTIHandled.count(BBCases[i].Value)) { 957 // If this is one we are capable of getting... 958 if (PredHasWeights || SuccHasWeights) 959 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 960 PredCases.push_back(BBCases[i]); 961 NewSuccessors.push_back(BBCases[i].Dest); 962 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 963 } 964 965 // If there are any constants vectored to BB that TI doesn't handle, 966 // they must go to the default destination of TI. 967 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 968 PTIHandled.begin(), 969 E = PTIHandled.end(); I != E; ++I) { 970 if (PredHasWeights || SuccHasWeights) 971 Weights.push_back(WeightsForHandled[*I]); 972 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 973 NewSuccessors.push_back(BBDefault); 974 } 975 } 976 977 // Okay, at this point, we know which new successor Pred will get. Make 978 // sure we update the number of entries in the PHI nodes for these 979 // successors. 980 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 981 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 982 983 Builder.SetInsertPoint(PTI); 984 // Convert pointer to int before we switch. 985 if (CV->getType()->isPointerTy()) { 986 assert(TD && "Cannot switch on pointer without DataLayout"); 987 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 988 "magicptr"); 989 } 990 991 // Now that the successors are updated, create the new Switch instruction. 992 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 993 PredCases.size()); 994 NewSI->setDebugLoc(PTI->getDebugLoc()); 995 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 996 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 997 998 if (PredHasWeights || SuccHasWeights) { 999 // Halve the weights if any of them cannot fit in an uint32_t 1000 FitWeights(Weights); 1001 1002 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1003 1004 NewSI->setMetadata(LLVMContext::MD_prof, 1005 MDBuilder(BB->getContext()). 1006 createBranchWeights(MDWeights)); 1007 } 1008 1009 EraseTerminatorInstAndDCECond(PTI); 1010 1011 // Okay, last check. If BB is still a successor of PSI, then we must 1012 // have an infinite loop case. If so, add an infinitely looping block 1013 // to handle the case to preserve the behavior of the code. 1014 BasicBlock *InfLoopBlock = 0; 1015 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1016 if (NewSI->getSuccessor(i) == BB) { 1017 if (InfLoopBlock == 0) { 1018 // Insert it at the end of the function, because it's either code, 1019 // or it won't matter if it's hot. :) 1020 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1021 "infloop", BB->getParent()); 1022 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1023 } 1024 NewSI->setSuccessor(i, InfLoopBlock); 1025 } 1026 1027 Changed = true; 1028 } 1029 } 1030 return Changed; 1031 } 1032 1033 // isSafeToHoistInvoke - If we would need to insert a select that uses the 1034 // value of this invoke (comments in HoistThenElseCodeToIf explain why we 1035 // would need to do this), we can't hoist the invoke, as there is nowhere 1036 // to put the select in this case. 1037 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1038 Instruction *I1, Instruction *I2) { 1039 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1040 PHINode *PN; 1041 for (BasicBlock::iterator BBI = SI->begin(); 1042 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1043 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1044 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1045 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1046 return false; 1047 } 1048 } 1049 } 1050 return true; 1051 } 1052 1053 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1054 /// BB2, hoist any common code in the two blocks up into the branch block. The 1055 /// caller of this function guarantees that BI's block dominates BB1 and BB2. 1056 static bool HoistThenElseCodeToIf(BranchInst *BI) { 1057 // This does very trivial matching, with limited scanning, to find identical 1058 // instructions in the two blocks. In particular, we don't want to get into 1059 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1060 // such, we currently just scan for obviously identical instructions in an 1061 // identical order. 1062 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1063 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1064 1065 BasicBlock::iterator BB1_Itr = BB1->begin(); 1066 BasicBlock::iterator BB2_Itr = BB2->begin(); 1067 1068 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1069 // Skip debug info if it is not identical. 1070 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1071 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1072 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1073 while (isa<DbgInfoIntrinsic>(I1)) 1074 I1 = BB1_Itr++; 1075 while (isa<DbgInfoIntrinsic>(I2)) 1076 I2 = BB2_Itr++; 1077 } 1078 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1079 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1080 return false; 1081 1082 // If we get here, we can hoist at least one instruction. 1083 BasicBlock *BIParent = BI->getParent(); 1084 1085 do { 1086 // If we are hoisting the terminator instruction, don't move one (making a 1087 // broken BB), instead clone it, and remove BI. 1088 if (isa<TerminatorInst>(I1)) 1089 goto HoistTerminator; 1090 1091 // For a normal instruction, we just move one to right before the branch, 1092 // then replace all uses of the other with the first. Finally, we remove 1093 // the now redundant second instruction. 1094 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1095 if (!I2->use_empty()) 1096 I2->replaceAllUsesWith(I1); 1097 I1->intersectOptionalDataWith(I2); 1098 I2->eraseFromParent(); 1099 1100 I1 = BB1_Itr++; 1101 I2 = BB2_Itr++; 1102 // Skip debug info if it is not identical. 1103 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1104 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1105 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1106 while (isa<DbgInfoIntrinsic>(I1)) 1107 I1 = BB1_Itr++; 1108 while (isa<DbgInfoIntrinsic>(I2)) 1109 I2 = BB2_Itr++; 1110 } 1111 } while (I1->isIdenticalToWhenDefined(I2)); 1112 1113 return true; 1114 1115 HoistTerminator: 1116 // It may not be possible to hoist an invoke. 1117 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1118 return true; 1119 1120 // Okay, it is safe to hoist the terminator. 1121 Instruction *NT = I1->clone(); 1122 BIParent->getInstList().insert(BI, NT); 1123 if (!NT->getType()->isVoidTy()) { 1124 I1->replaceAllUsesWith(NT); 1125 I2->replaceAllUsesWith(NT); 1126 NT->takeName(I1); 1127 } 1128 1129 IRBuilder<true, NoFolder> Builder(NT); 1130 // Hoisting one of the terminators from our successor is a great thing. 1131 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1132 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1133 // nodes, so we insert select instruction to compute the final result. 1134 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1135 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1136 PHINode *PN; 1137 for (BasicBlock::iterator BBI = SI->begin(); 1138 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1139 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1140 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1141 if (BB1V == BB2V) continue; 1142 1143 // These values do not agree. Insert a select instruction before NT 1144 // that determines the right value. 1145 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1146 if (SI == 0) 1147 SI = cast<SelectInst> 1148 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1149 BB1V->getName()+"."+BB2V->getName())); 1150 1151 // Make the PHI node use the select for all incoming values for BB1/BB2 1152 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1153 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1154 PN->setIncomingValue(i, SI); 1155 } 1156 } 1157 1158 // Update any PHI nodes in our new successors. 1159 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1160 AddPredecessorToBlock(*SI, BIParent, BB1); 1161 1162 EraseTerminatorInstAndDCECond(BI); 1163 return true; 1164 } 1165 1166 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1167 /// check whether BBEnd has only two predecessors and the other predecessor 1168 /// ends with an unconditional branch. If it is true, sink any common code 1169 /// in the two predecessors to BBEnd. 1170 static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1171 assert(BI1->isUnconditional()); 1172 BasicBlock *BB1 = BI1->getParent(); 1173 BasicBlock *BBEnd = BI1->getSuccessor(0); 1174 1175 // Check that BBEnd has two predecessors and the other predecessor ends with 1176 // an unconditional branch. 1177 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1178 BasicBlock *Pred0 = *PI++; 1179 if (PI == PE) // Only one predecessor. 1180 return false; 1181 BasicBlock *Pred1 = *PI++; 1182 if (PI != PE) // More than two predecessors. 1183 return false; 1184 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1185 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1186 if (!BI2 || !BI2->isUnconditional()) 1187 return false; 1188 1189 // Gather the PHI nodes in BBEnd. 1190 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1191 Instruction *FirstNonPhiInBBEnd = 0; 1192 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1193 I != E; ++I) { 1194 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1195 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1196 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1197 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1198 } else { 1199 FirstNonPhiInBBEnd = &*I; 1200 break; 1201 } 1202 } 1203 if (!FirstNonPhiInBBEnd) 1204 return false; 1205 1206 1207 // This does very trivial matching, with limited scanning, to find identical 1208 // instructions in the two blocks. We scan backward for obviously identical 1209 // instructions in an identical order. 1210 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1211 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1212 RE2 = BB2->getInstList().rend(); 1213 // Skip debug info. 1214 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1215 if (RI1 == RE1) 1216 return false; 1217 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1218 if (RI2 == RE2) 1219 return false; 1220 // Skip the unconditional branches. 1221 ++RI1; 1222 ++RI2; 1223 1224 bool Changed = false; 1225 while (RI1 != RE1 && RI2 != RE2) { 1226 // Skip debug info. 1227 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1228 if (RI1 == RE1) 1229 return Changed; 1230 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1231 if (RI2 == RE2) 1232 return Changed; 1233 1234 Instruction *I1 = &*RI1, *I2 = &*RI2; 1235 // I1 and I2 should have a single use in the same PHI node, and they 1236 // perform the same operation. 1237 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1238 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1239 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1240 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1241 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1242 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1243 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1244 !I1->hasOneUse() || !I2->hasOneUse() || 1245 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1246 MapValueFromBB1ToBB2[I1].first != I2) 1247 return Changed; 1248 1249 // Check whether we should swap the operands of ICmpInst. 1250 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1251 bool SwapOpnds = false; 1252 if (ICmp1 && ICmp2 && 1253 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1254 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1255 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1256 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1257 ICmp2->swapOperands(); 1258 SwapOpnds = true; 1259 } 1260 if (!I1->isSameOperationAs(I2)) { 1261 if (SwapOpnds) 1262 ICmp2->swapOperands(); 1263 return Changed; 1264 } 1265 1266 // The operands should be either the same or they need to be generated 1267 // with a PHI node after sinking. We only handle the case where there is 1268 // a single pair of different operands. 1269 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1270 unsigned Op1Idx = 0; 1271 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1272 if (I1->getOperand(I) == I2->getOperand(I)) 1273 continue; 1274 // Early exit if we have more-than one pair of different operands or 1275 // the different operand is already in MapValueFromBB1ToBB2. 1276 // Early exit if we need a PHI node to replace a constant. 1277 if (DifferentOp1 || 1278 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1279 MapValueFromBB1ToBB2.end() || 1280 isa<Constant>(I1->getOperand(I)) || 1281 isa<Constant>(I2->getOperand(I))) { 1282 // If we can't sink the instructions, undo the swapping. 1283 if (SwapOpnds) 1284 ICmp2->swapOperands(); 1285 return Changed; 1286 } 1287 DifferentOp1 = I1->getOperand(I); 1288 Op1Idx = I; 1289 DifferentOp2 = I2->getOperand(I); 1290 } 1291 1292 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1293 // remove (I1, I2) from MapValueFromBB1ToBB2. 1294 if (DifferentOp1) { 1295 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1296 DifferentOp1->getName() + ".sink", 1297 BBEnd->begin()); 1298 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1299 // I1 should use NewPN instead of DifferentOp1. 1300 I1->setOperand(Op1Idx, NewPN); 1301 NewPN->addIncoming(DifferentOp1, BB1); 1302 NewPN->addIncoming(DifferentOp2, BB2); 1303 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1304 } 1305 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1306 MapValueFromBB1ToBB2.erase(I1); 1307 1308 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1309 DEBUG(dbgs() << " " << *I2 << "\n";); 1310 // We need to update RE1 and RE2 if we are going to sink the first 1311 // instruction in the basic block down. 1312 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1313 // Sink the instruction. 1314 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1315 if (!OldPN->use_empty()) 1316 OldPN->replaceAllUsesWith(I1); 1317 OldPN->eraseFromParent(); 1318 1319 if (!I2->use_empty()) 1320 I2->replaceAllUsesWith(I1); 1321 I1->intersectOptionalDataWith(I2); 1322 I2->eraseFromParent(); 1323 1324 if (UpdateRE1) 1325 RE1 = BB1->getInstList().rend(); 1326 if (UpdateRE2) 1327 RE2 = BB2->getInstList().rend(); 1328 FirstNonPhiInBBEnd = I1; 1329 NumSinkCommons++; 1330 Changed = true; 1331 } 1332 return Changed; 1333 } 1334 1335 /// \brief Speculate a conditional basic block flattening the CFG. 1336 /// 1337 /// Note that this is a very risky transform currently. Speculating 1338 /// instructions like this is most often not desirable. Instead, there is an MI 1339 /// pass which can do it with full awareness of the resource constraints. 1340 /// However, some cases are "obvious" and we should do directly. An example of 1341 /// this is speculating a single, reasonably cheap instruction. 1342 /// 1343 /// There is only one distinct advantage to flattening the CFG at the IR level: 1344 /// it makes very common but simplistic optimizations such as are common in 1345 /// instcombine and the DAG combiner more powerful by removing CFG edges and 1346 /// modeling their effects with easier to reason about SSA value graphs. 1347 /// 1348 /// 1349 /// An illustration of this transform is turning this IR: 1350 /// \code 1351 /// BB: 1352 /// %cmp = icmp ult %x, %y 1353 /// br i1 %cmp, label %EndBB, label %ThenBB 1354 /// ThenBB: 1355 /// %sub = sub %x, %y 1356 /// br label BB2 1357 /// EndBB: 1358 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1359 /// ... 1360 /// \endcode 1361 /// 1362 /// Into this IR: 1363 /// \code 1364 /// BB: 1365 /// %cmp = icmp ult %x, %y 1366 /// %sub = sub %x, %y 1367 /// %cond = select i1 %cmp, 0, %sub 1368 /// ... 1369 /// \endcode 1370 /// 1371 /// \returns true if the conditional block is removed. 1372 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1373 // Be conservative for now. FP select instruction can often be expensive. 1374 Value *BrCond = BI->getCondition(); 1375 if (isa<FCmpInst>(BrCond)) 1376 return false; 1377 1378 BasicBlock *BB = BI->getParent(); 1379 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1380 1381 // If ThenBB is actually on the false edge of the conditional branch, remember 1382 // to swap the select operands later. 1383 bool Invert = false; 1384 if (ThenBB != BI->getSuccessor(0)) { 1385 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1386 Invert = true; 1387 } 1388 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1389 1390 // Keep a count of how many times instructions are used within CondBB when 1391 // they are candidates for sinking into CondBB. Specifically: 1392 // - They are defined in BB, and 1393 // - They have no side effects, and 1394 // - All of their uses are in CondBB. 1395 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1396 1397 unsigned SpeculationCost = 0; 1398 for (BasicBlock::iterator BBI = ThenBB->begin(), 1399 BBE = llvm::prior(ThenBB->end()); 1400 BBI != BBE; ++BBI) { 1401 Instruction *I = BBI; 1402 // Skip debug info. 1403 if (isa<DbgInfoIntrinsic>(I)) 1404 continue; 1405 1406 // Only speculatively execution a single instruction (not counting the 1407 // terminator) for now. 1408 ++SpeculationCost; 1409 if (SpeculationCost > 1) 1410 return false; 1411 1412 // Don't hoist the instruction if it's unsafe or expensive. 1413 if (!isSafeToSpeculativelyExecute(I)) 1414 return false; 1415 if (ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1416 return false; 1417 1418 // Do not hoist the instruction if any of its operands are defined but not 1419 // used in this BB. The transformation will prevent the operand from 1420 // being sunk into the use block. 1421 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1422 i != e; ++i) { 1423 Instruction *OpI = dyn_cast<Instruction>(*i); 1424 if (!OpI || OpI->getParent() != BB || 1425 OpI->mayHaveSideEffects()) 1426 continue; // Not a candidate for sinking. 1427 1428 ++SinkCandidateUseCounts[OpI]; 1429 } 1430 } 1431 1432 // Consider any sink candidates which are only used in CondBB as costs for 1433 // speculation. Note, while we iterate over a DenseMap here, we are summing 1434 // and so iteration order isn't significant. 1435 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1436 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1437 I != E; ++I) 1438 if (I->first->getNumUses() == I->second) { 1439 ++SpeculationCost; 1440 if (SpeculationCost > 1) 1441 return false; 1442 } 1443 1444 // Check that the PHI nodes can be converted to selects. 1445 bool HaveRewritablePHIs = false; 1446 for (BasicBlock::iterator I = EndBB->begin(); 1447 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1448 Value *OrigV = PN->getIncomingValueForBlock(BB); 1449 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1450 1451 // Skip PHIs which are trivial. 1452 if (ThenV == OrigV) 1453 continue; 1454 1455 HaveRewritablePHIs = true; 1456 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV); 1457 if (!CE) 1458 continue; // Known safe and cheap. 1459 1460 if (!isSafeToSpeculativelyExecute(CE)) 1461 return false; 1462 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1463 return false; 1464 1465 // Account for the cost of an unfolded ConstantExpr which could end up 1466 // getting expanded into Instructions. 1467 // FIXME: This doesn't account for how many operations are combined in the 1468 // constant expression. 1469 ++SpeculationCost; 1470 if (SpeculationCost > 1) 1471 return false; 1472 } 1473 1474 // If there are no PHIs to process, bail early. This helps ensure idempotence 1475 // as well. 1476 if (!HaveRewritablePHIs) 1477 return false; 1478 1479 // If we get here, we can hoist the instruction and if-convert. 1480 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1481 1482 // Hoist the instructions. 1483 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1484 llvm::prior(ThenBB->end())); 1485 1486 // Insert selects and rewrite the PHI operands. 1487 IRBuilder<true, NoFolder> Builder(BI); 1488 for (BasicBlock::iterator I = EndBB->begin(); 1489 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1490 unsigned OrigI = PN->getBasicBlockIndex(BB); 1491 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1492 Value *OrigV = PN->getIncomingValue(OrigI); 1493 Value *ThenV = PN->getIncomingValue(ThenI); 1494 1495 // Skip PHIs which are trivial. 1496 if (OrigV == ThenV) 1497 continue; 1498 1499 // Create a select whose true value is the speculatively executed value and 1500 // false value is the preexisting value. Swap them if the branch 1501 // destinations were inverted. 1502 Value *TrueV = ThenV, *FalseV = OrigV; 1503 if (Invert) 1504 std::swap(TrueV, FalseV); 1505 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1506 TrueV->getName() + "." + FalseV->getName()); 1507 PN->setIncomingValue(OrigI, V); 1508 PN->setIncomingValue(ThenI, V); 1509 } 1510 1511 ++NumSpeculations; 1512 return true; 1513 } 1514 1515 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1516 /// across this block. 1517 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1518 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1519 unsigned Size = 0; 1520 1521 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1522 if (isa<DbgInfoIntrinsic>(BBI)) 1523 continue; 1524 if (Size > 10) return false; // Don't clone large BB's. 1525 ++Size; 1526 1527 // We can only support instructions that do not define values that are 1528 // live outside of the current basic block. 1529 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1530 UI != E; ++UI) { 1531 Instruction *U = cast<Instruction>(*UI); 1532 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1533 } 1534 1535 // Looks ok, continue checking. 1536 } 1537 1538 return true; 1539 } 1540 1541 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1542 /// that is defined in the same block as the branch and if any PHI entries are 1543 /// constants, thread edges corresponding to that entry to be branches to their 1544 /// ultimate destination. 1545 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1546 BasicBlock *BB = BI->getParent(); 1547 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1548 // NOTE: we currently cannot transform this case if the PHI node is used 1549 // outside of the block. 1550 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1551 return false; 1552 1553 // Degenerate case of a single entry PHI. 1554 if (PN->getNumIncomingValues() == 1) { 1555 FoldSingleEntryPHINodes(PN->getParent()); 1556 return true; 1557 } 1558 1559 // Now we know that this block has multiple preds and two succs. 1560 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1561 1562 // Okay, this is a simple enough basic block. See if any phi values are 1563 // constants. 1564 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1565 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1566 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1567 1568 // Okay, we now know that all edges from PredBB should be revectored to 1569 // branch to RealDest. 1570 BasicBlock *PredBB = PN->getIncomingBlock(i); 1571 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1572 1573 if (RealDest == BB) continue; // Skip self loops. 1574 // Skip if the predecessor's terminator is an indirect branch. 1575 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1576 1577 // The dest block might have PHI nodes, other predecessors and other 1578 // difficult cases. Instead of being smart about this, just insert a new 1579 // block that jumps to the destination block, effectively splitting 1580 // the edge we are about to create. 1581 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1582 RealDest->getName()+".critedge", 1583 RealDest->getParent(), RealDest); 1584 BranchInst::Create(RealDest, EdgeBB); 1585 1586 // Update PHI nodes. 1587 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1588 1589 // BB may have instructions that are being threaded over. Clone these 1590 // instructions into EdgeBB. We know that there will be no uses of the 1591 // cloned instructions outside of EdgeBB. 1592 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1593 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1594 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1595 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1596 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1597 continue; 1598 } 1599 // Clone the instruction. 1600 Instruction *N = BBI->clone(); 1601 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1602 1603 // Update operands due to translation. 1604 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1605 i != e; ++i) { 1606 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1607 if (PI != TranslateMap.end()) 1608 *i = PI->second; 1609 } 1610 1611 // Check for trivial simplification. 1612 if (Value *V = SimplifyInstruction(N, TD)) { 1613 TranslateMap[BBI] = V; 1614 delete N; // Instruction folded away, don't need actual inst 1615 } else { 1616 // Insert the new instruction into its new home. 1617 EdgeBB->getInstList().insert(InsertPt, N); 1618 if (!BBI->use_empty()) 1619 TranslateMap[BBI] = N; 1620 } 1621 } 1622 1623 // Loop over all of the edges from PredBB to BB, changing them to branch 1624 // to EdgeBB instead. 1625 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1626 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1627 if (PredBBTI->getSuccessor(i) == BB) { 1628 BB->removePredecessor(PredBB); 1629 PredBBTI->setSuccessor(i, EdgeBB); 1630 } 1631 1632 // Recurse, simplifying any other constants. 1633 return FoldCondBranchOnPHI(BI, TD) | true; 1634 } 1635 1636 return false; 1637 } 1638 1639 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1640 /// PHI node, see if we can eliminate it. 1641 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1642 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1643 // statement", which has a very simple dominance structure. Basically, we 1644 // are trying to find the condition that is being branched on, which 1645 // subsequently causes this merge to happen. We really want control 1646 // dependence information for this check, but simplifycfg can't keep it up 1647 // to date, and this catches most of the cases we care about anyway. 1648 BasicBlock *BB = PN->getParent(); 1649 BasicBlock *IfTrue, *IfFalse; 1650 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1651 if (!IfCond || 1652 // Don't bother if the branch will be constant folded trivially. 1653 isa<ConstantInt>(IfCond)) 1654 return false; 1655 1656 // Okay, we found that we can merge this two-entry phi node into a select. 1657 // Doing so would require us to fold *all* two entry phi nodes in this block. 1658 // At some point this becomes non-profitable (particularly if the target 1659 // doesn't support cmov's). Only do this transformation if there are two or 1660 // fewer PHI nodes in this block. 1661 unsigned NumPhis = 0; 1662 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1663 if (NumPhis > 2) 1664 return false; 1665 1666 // Loop over the PHI's seeing if we can promote them all to select 1667 // instructions. While we are at it, keep track of the instructions 1668 // that need to be moved to the dominating block. 1669 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1670 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1671 MaxCostVal1 = PHINodeFoldingThreshold; 1672 1673 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1674 PHINode *PN = cast<PHINode>(II++); 1675 if (Value *V = SimplifyInstruction(PN, TD)) { 1676 PN->replaceAllUsesWith(V); 1677 PN->eraseFromParent(); 1678 continue; 1679 } 1680 1681 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1682 MaxCostVal0) || 1683 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1684 MaxCostVal1)) 1685 return false; 1686 } 1687 1688 // If we folded the first phi, PN dangles at this point. Refresh it. If 1689 // we ran out of PHIs then we simplified them all. 1690 PN = dyn_cast<PHINode>(BB->begin()); 1691 if (PN == 0) return true; 1692 1693 // Don't fold i1 branches on PHIs which contain binary operators. These can 1694 // often be turned into switches and other things. 1695 if (PN->getType()->isIntegerTy(1) && 1696 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1697 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1698 isa<BinaryOperator>(IfCond))) 1699 return false; 1700 1701 // If we all PHI nodes are promotable, check to make sure that all 1702 // instructions in the predecessor blocks can be promoted as well. If 1703 // not, we won't be able to get rid of the control flow, so it's not 1704 // worth promoting to select instructions. 1705 BasicBlock *DomBlock = 0; 1706 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1707 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1708 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1709 IfBlock1 = 0; 1710 } else { 1711 DomBlock = *pred_begin(IfBlock1); 1712 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1713 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1714 // This is not an aggressive instruction that we can promote. 1715 // Because of this, we won't be able to get rid of the control 1716 // flow, so the xform is not worth it. 1717 return false; 1718 } 1719 } 1720 1721 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1722 IfBlock2 = 0; 1723 } else { 1724 DomBlock = *pred_begin(IfBlock2); 1725 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1726 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1727 // This is not an aggressive instruction that we can promote. 1728 // Because of this, we won't be able to get rid of the control 1729 // flow, so the xform is not worth it. 1730 return false; 1731 } 1732 } 1733 1734 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1735 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1736 1737 // If we can still promote the PHI nodes after this gauntlet of tests, 1738 // do all of the PHI's now. 1739 Instruction *InsertPt = DomBlock->getTerminator(); 1740 IRBuilder<true, NoFolder> Builder(InsertPt); 1741 1742 // Move all 'aggressive' instructions, which are defined in the 1743 // conditional parts of the if's up to the dominating block. 1744 if (IfBlock1) 1745 DomBlock->getInstList().splice(InsertPt, 1746 IfBlock1->getInstList(), IfBlock1->begin(), 1747 IfBlock1->getTerminator()); 1748 if (IfBlock2) 1749 DomBlock->getInstList().splice(InsertPt, 1750 IfBlock2->getInstList(), IfBlock2->begin(), 1751 IfBlock2->getTerminator()); 1752 1753 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1754 // Change the PHI node into a select instruction. 1755 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1756 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1757 1758 SelectInst *NV = 1759 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1760 PN->replaceAllUsesWith(NV); 1761 NV->takeName(PN); 1762 PN->eraseFromParent(); 1763 } 1764 1765 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1766 // has been flattened. Change DomBlock to jump directly to our new block to 1767 // avoid other simplifycfg's kicking in on the diamond. 1768 TerminatorInst *OldTI = DomBlock->getTerminator(); 1769 Builder.SetInsertPoint(OldTI); 1770 Builder.CreateBr(BB); 1771 OldTI->eraseFromParent(); 1772 return true; 1773 } 1774 1775 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1776 /// to two returning blocks, try to merge them together into one return, 1777 /// introducing a select if the return values disagree. 1778 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1779 IRBuilder<> &Builder) { 1780 assert(BI->isConditional() && "Must be a conditional branch"); 1781 BasicBlock *TrueSucc = BI->getSuccessor(0); 1782 BasicBlock *FalseSucc = BI->getSuccessor(1); 1783 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1784 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1785 1786 // Check to ensure both blocks are empty (just a return) or optionally empty 1787 // with PHI nodes. If there are other instructions, merging would cause extra 1788 // computation on one path or the other. 1789 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1790 return false; 1791 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1792 return false; 1793 1794 Builder.SetInsertPoint(BI); 1795 // Okay, we found a branch that is going to two return nodes. If 1796 // there is no return value for this function, just change the 1797 // branch into a return. 1798 if (FalseRet->getNumOperands() == 0) { 1799 TrueSucc->removePredecessor(BI->getParent()); 1800 FalseSucc->removePredecessor(BI->getParent()); 1801 Builder.CreateRetVoid(); 1802 EraseTerminatorInstAndDCECond(BI); 1803 return true; 1804 } 1805 1806 // Otherwise, figure out what the true and false return values are 1807 // so we can insert a new select instruction. 1808 Value *TrueValue = TrueRet->getReturnValue(); 1809 Value *FalseValue = FalseRet->getReturnValue(); 1810 1811 // Unwrap any PHI nodes in the return blocks. 1812 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1813 if (TVPN->getParent() == TrueSucc) 1814 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1815 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1816 if (FVPN->getParent() == FalseSucc) 1817 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1818 1819 // In order for this transformation to be safe, we must be able to 1820 // unconditionally execute both operands to the return. This is 1821 // normally the case, but we could have a potentially-trapping 1822 // constant expression that prevents this transformation from being 1823 // safe. 1824 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1825 if (TCV->canTrap()) 1826 return false; 1827 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1828 if (FCV->canTrap()) 1829 return false; 1830 1831 // Okay, we collected all the mapped values and checked them for sanity, and 1832 // defined to really do this transformation. First, update the CFG. 1833 TrueSucc->removePredecessor(BI->getParent()); 1834 FalseSucc->removePredecessor(BI->getParent()); 1835 1836 // Insert select instructions where needed. 1837 Value *BrCond = BI->getCondition(); 1838 if (TrueValue) { 1839 // Insert a select if the results differ. 1840 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1841 } else if (isa<UndefValue>(TrueValue)) { 1842 TrueValue = FalseValue; 1843 } else { 1844 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1845 FalseValue, "retval"); 1846 } 1847 } 1848 1849 Value *RI = !TrueValue ? 1850 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1851 1852 (void) RI; 1853 1854 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1855 << "\n " << *BI << "NewRet = " << *RI 1856 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1857 1858 EraseTerminatorInstAndDCECond(BI); 1859 1860 return true; 1861 } 1862 1863 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1864 /// probabilities of the branch taking each edge. Fills in the two APInt 1865 /// parameters and return true, or returns false if no or invalid metadata was 1866 /// found. 1867 static bool ExtractBranchMetadata(BranchInst *BI, 1868 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1869 assert(BI->isConditional() && 1870 "Looking for probabilities on unconditional branch?"); 1871 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1872 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1873 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1874 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1875 if (!CITrue || !CIFalse) return false; 1876 ProbTrue = CITrue->getValue().getZExtValue(); 1877 ProbFalse = CIFalse->getValue().getZExtValue(); 1878 return true; 1879 } 1880 1881 /// checkCSEInPredecessor - Return true if the given instruction is available 1882 /// in its predecessor block. If yes, the instruction will be removed. 1883 /// 1884 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1885 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1886 return false; 1887 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1888 Instruction *PBI = &*I; 1889 // Check whether Inst and PBI generate the same value. 1890 if (Inst->isIdenticalTo(PBI)) { 1891 Inst->replaceAllUsesWith(PBI); 1892 Inst->eraseFromParent(); 1893 return true; 1894 } 1895 } 1896 return false; 1897 } 1898 1899 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1900 /// predecessor branches to us and one of our successors, fold the block into 1901 /// the predecessor and use logical operations to pick the right destination. 1902 bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1903 BasicBlock *BB = BI->getParent(); 1904 1905 Instruction *Cond = 0; 1906 if (BI->isConditional()) 1907 Cond = dyn_cast<Instruction>(BI->getCondition()); 1908 else { 1909 // For unconditional branch, check for a simple CFG pattern, where 1910 // BB has a single predecessor and BB's successor is also its predecessor's 1911 // successor. If such pattern exisits, check for CSE between BB and its 1912 // predecessor. 1913 if (BasicBlock *PB = BB->getSinglePredecessor()) 1914 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1915 if (PBI->isConditional() && 1916 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1917 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1918 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1919 I != E; ) { 1920 Instruction *Curr = I++; 1921 if (isa<CmpInst>(Curr)) { 1922 Cond = Curr; 1923 break; 1924 } 1925 // Quit if we can't remove this instruction. 1926 if (!checkCSEInPredecessor(Curr, PB)) 1927 return false; 1928 } 1929 } 1930 1931 if (Cond == 0) 1932 return false; 1933 } 1934 1935 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1936 Cond->getParent() != BB || !Cond->hasOneUse()) 1937 return false; 1938 1939 // Only allow this if the condition is a simple instruction that can be 1940 // executed unconditionally. It must be in the same block as the branch, and 1941 // must be at the front of the block. 1942 BasicBlock::iterator FrontIt = BB->front(); 1943 1944 // Ignore dbg intrinsics. 1945 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1946 1947 // Allow a single instruction to be hoisted in addition to the compare 1948 // that feeds the branch. We later ensure that any values that _it_ uses 1949 // were also live in the predecessor, so that we don't unnecessarily create 1950 // register pressure or inhibit out-of-order execution. 1951 Instruction *BonusInst = 0; 1952 if (&*FrontIt != Cond && 1953 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1954 isSafeToSpeculativelyExecute(FrontIt)) { 1955 BonusInst = &*FrontIt; 1956 ++FrontIt; 1957 1958 // Ignore dbg intrinsics. 1959 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1960 } 1961 1962 // Only a single bonus inst is allowed. 1963 if (&*FrontIt != Cond) 1964 return false; 1965 1966 // Make sure the instruction after the condition is the cond branch. 1967 BasicBlock::iterator CondIt = Cond; ++CondIt; 1968 1969 // Ingore dbg intrinsics. 1970 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1971 1972 if (&*CondIt != BI) 1973 return false; 1974 1975 // Cond is known to be a compare or binary operator. Check to make sure that 1976 // neither operand is a potentially-trapping constant expression. 1977 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1978 if (CE->canTrap()) 1979 return false; 1980 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1981 if (CE->canTrap()) 1982 return false; 1983 1984 // Finally, don't infinitely unroll conditional loops. 1985 BasicBlock *TrueDest = BI->getSuccessor(0); 1986 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 1987 if (TrueDest == BB || FalseDest == BB) 1988 return false; 1989 1990 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1991 BasicBlock *PredBlock = *PI; 1992 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1993 1994 // Check that we have two conditional branches. If there is a PHI node in 1995 // the common successor, verify that the same value flows in from both 1996 // blocks. 1997 SmallVector<PHINode*, 4> PHIs; 1998 if (PBI == 0 || PBI->isUnconditional() || 1999 (BI->isConditional() && 2000 !SafeToMergeTerminators(BI, PBI)) || 2001 (!BI->isConditional() && 2002 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2003 continue; 2004 2005 // Determine if the two branches share a common destination. 2006 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2007 bool InvertPredCond = false; 2008 2009 if (BI->isConditional()) { 2010 if (PBI->getSuccessor(0) == TrueDest) 2011 Opc = Instruction::Or; 2012 else if (PBI->getSuccessor(1) == FalseDest) 2013 Opc = Instruction::And; 2014 else if (PBI->getSuccessor(0) == FalseDest) 2015 Opc = Instruction::And, InvertPredCond = true; 2016 else if (PBI->getSuccessor(1) == TrueDest) 2017 Opc = Instruction::Or, InvertPredCond = true; 2018 else 2019 continue; 2020 } else { 2021 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2022 continue; 2023 } 2024 2025 // Ensure that any values used in the bonus instruction are also used 2026 // by the terminator of the predecessor. This means that those values 2027 // must already have been resolved, so we won't be inhibiting the 2028 // out-of-order core by speculating them earlier. 2029 if (BonusInst) { 2030 // Collect the values used by the bonus inst 2031 SmallPtrSet<Value*, 4> UsedValues; 2032 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2033 OE = BonusInst->op_end(); OI != OE; ++OI) { 2034 Value *V = *OI; 2035 if (!isa<Constant>(V)) 2036 UsedValues.insert(V); 2037 } 2038 2039 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2040 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2041 2042 // Walk up to four levels back up the use-def chain of the predecessor's 2043 // terminator to see if all those values were used. The choice of four 2044 // levels is arbitrary, to provide a compile-time-cost bound. 2045 while (!Worklist.empty()) { 2046 std::pair<Value*, unsigned> Pair = Worklist.back(); 2047 Worklist.pop_back(); 2048 2049 if (Pair.second >= 4) continue; 2050 UsedValues.erase(Pair.first); 2051 if (UsedValues.empty()) break; 2052 2053 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2054 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2055 OI != OE; ++OI) 2056 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2057 } 2058 } 2059 2060 if (!UsedValues.empty()) return false; 2061 } 2062 2063 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2064 IRBuilder<> Builder(PBI); 2065 2066 // If we need to invert the condition in the pred block to match, do so now. 2067 if (InvertPredCond) { 2068 Value *NewCond = PBI->getCondition(); 2069 2070 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2071 CmpInst *CI = cast<CmpInst>(NewCond); 2072 CI->setPredicate(CI->getInversePredicate()); 2073 } else { 2074 NewCond = Builder.CreateNot(NewCond, 2075 PBI->getCondition()->getName()+".not"); 2076 } 2077 2078 PBI->setCondition(NewCond); 2079 PBI->swapSuccessors(); 2080 } 2081 2082 // If we have a bonus inst, clone it into the predecessor block. 2083 Instruction *NewBonus = 0; 2084 if (BonusInst) { 2085 NewBonus = BonusInst->clone(); 2086 PredBlock->getInstList().insert(PBI, NewBonus); 2087 NewBonus->takeName(BonusInst); 2088 BonusInst->setName(BonusInst->getName()+".old"); 2089 } 2090 2091 // Clone Cond into the predecessor basic block, and or/and the 2092 // two conditions together. 2093 Instruction *New = Cond->clone(); 2094 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2095 PredBlock->getInstList().insert(PBI, New); 2096 New->takeName(Cond); 2097 Cond->setName(New->getName()+".old"); 2098 2099 if (BI->isConditional()) { 2100 Instruction *NewCond = 2101 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2102 New, "or.cond")); 2103 PBI->setCondition(NewCond); 2104 2105 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2106 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2107 PredFalseWeight); 2108 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2109 SuccFalseWeight); 2110 SmallVector<uint64_t, 8> NewWeights; 2111 2112 if (PBI->getSuccessor(0) == BB) { 2113 if (PredHasWeights && SuccHasWeights) { 2114 // PBI: br i1 %x, BB, FalseDest 2115 // BI: br i1 %y, TrueDest, FalseDest 2116 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2117 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2118 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2119 // TrueWeight for PBI * FalseWeight for BI. 2120 // We assume that total weights of a BranchInst can fit into 32 bits. 2121 // Therefore, we will not have overflow using 64-bit arithmetic. 2122 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2123 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2124 } 2125 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2126 PBI->setSuccessor(0, TrueDest); 2127 } 2128 if (PBI->getSuccessor(1) == BB) { 2129 if (PredHasWeights && SuccHasWeights) { 2130 // PBI: br i1 %x, TrueDest, BB 2131 // BI: br i1 %y, TrueDest, FalseDest 2132 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2133 // FalseWeight for PBI * TrueWeight for BI. 2134 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2135 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2136 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2137 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2138 } 2139 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2140 PBI->setSuccessor(1, FalseDest); 2141 } 2142 if (NewWeights.size() == 2) { 2143 // Halve the weights if any of them cannot fit in an uint32_t 2144 FitWeights(NewWeights); 2145 2146 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2147 PBI->setMetadata(LLVMContext::MD_prof, 2148 MDBuilder(BI->getContext()). 2149 createBranchWeights(MDWeights)); 2150 } else 2151 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2152 } else { 2153 // Update PHI nodes in the common successors. 2154 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2155 ConstantInt *PBI_C = cast<ConstantInt>( 2156 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2157 assert(PBI_C->getType()->isIntegerTy(1)); 2158 Instruction *MergedCond = 0; 2159 if (PBI->getSuccessor(0) == TrueDest) { 2160 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2161 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2162 // is false: !PBI_Cond and BI_Value 2163 Instruction *NotCond = 2164 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2165 "not.cond")); 2166 MergedCond = 2167 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2168 NotCond, New, 2169 "and.cond")); 2170 if (PBI_C->isOne()) 2171 MergedCond = 2172 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2173 PBI->getCondition(), MergedCond, 2174 "or.cond")); 2175 } else { 2176 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2177 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2178 // is false: PBI_Cond and BI_Value 2179 MergedCond = 2180 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2181 PBI->getCondition(), New, 2182 "and.cond")); 2183 if (PBI_C->isOne()) { 2184 Instruction *NotCond = 2185 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2186 "not.cond")); 2187 MergedCond = 2188 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2189 NotCond, MergedCond, 2190 "or.cond")); 2191 } 2192 } 2193 // Update PHI Node. 2194 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2195 MergedCond); 2196 } 2197 // Change PBI from Conditional to Unconditional. 2198 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2199 EraseTerminatorInstAndDCECond(PBI); 2200 PBI = New_PBI; 2201 } 2202 2203 // TODO: If BB is reachable from all paths through PredBlock, then we 2204 // could replace PBI's branch probabilities with BI's. 2205 2206 // Copy any debug value intrinsics into the end of PredBlock. 2207 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2208 if (isa<DbgInfoIntrinsic>(*I)) 2209 I->clone()->insertBefore(PBI); 2210 2211 return true; 2212 } 2213 return false; 2214 } 2215 2216 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2217 /// predecessor of another block, this function tries to simplify it. We know 2218 /// that PBI and BI are both conditional branches, and BI is in one of the 2219 /// successor blocks of PBI - PBI branches to BI. 2220 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2221 assert(PBI->isConditional() && BI->isConditional()); 2222 BasicBlock *BB = BI->getParent(); 2223 2224 // If this block ends with a branch instruction, and if there is a 2225 // predecessor that ends on a branch of the same condition, make 2226 // this conditional branch redundant. 2227 if (PBI->getCondition() == BI->getCondition() && 2228 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2229 // Okay, the outcome of this conditional branch is statically 2230 // knowable. If this block had a single pred, handle specially. 2231 if (BB->getSinglePredecessor()) { 2232 // Turn this into a branch on constant. 2233 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2234 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2235 CondIsTrue)); 2236 return true; // Nuke the branch on constant. 2237 } 2238 2239 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2240 // in the constant and simplify the block result. Subsequent passes of 2241 // simplifycfg will thread the block. 2242 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2243 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2244 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2245 std::distance(PB, PE), 2246 BI->getCondition()->getName() + ".pr", 2247 BB->begin()); 2248 // Okay, we're going to insert the PHI node. Since PBI is not the only 2249 // predecessor, compute the PHI'd conditional value for all of the preds. 2250 // Any predecessor where the condition is not computable we keep symbolic. 2251 for (pred_iterator PI = PB; PI != PE; ++PI) { 2252 BasicBlock *P = *PI; 2253 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2254 PBI != BI && PBI->isConditional() && 2255 PBI->getCondition() == BI->getCondition() && 2256 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2257 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2258 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2259 CondIsTrue), P); 2260 } else { 2261 NewPN->addIncoming(BI->getCondition(), P); 2262 } 2263 } 2264 2265 BI->setCondition(NewPN); 2266 return true; 2267 } 2268 } 2269 2270 // If this is a conditional branch in an empty block, and if any 2271 // predecessors is a conditional branch to one of our destinations, 2272 // fold the conditions into logical ops and one cond br. 2273 BasicBlock::iterator BBI = BB->begin(); 2274 // Ignore dbg intrinsics. 2275 while (isa<DbgInfoIntrinsic>(BBI)) 2276 ++BBI; 2277 if (&*BBI != BI) 2278 return false; 2279 2280 2281 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2282 if (CE->canTrap()) 2283 return false; 2284 2285 int PBIOp, BIOp; 2286 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2287 PBIOp = BIOp = 0; 2288 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2289 PBIOp = 0, BIOp = 1; 2290 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2291 PBIOp = 1, BIOp = 0; 2292 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2293 PBIOp = BIOp = 1; 2294 else 2295 return false; 2296 2297 // Check to make sure that the other destination of this branch 2298 // isn't BB itself. If so, this is an infinite loop that will 2299 // keep getting unwound. 2300 if (PBI->getSuccessor(PBIOp) == BB) 2301 return false; 2302 2303 // Do not perform this transformation if it would require 2304 // insertion of a large number of select instructions. For targets 2305 // without predication/cmovs, this is a big pessimization. 2306 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2307 2308 unsigned NumPhis = 0; 2309 for (BasicBlock::iterator II = CommonDest->begin(); 2310 isa<PHINode>(II); ++II, ++NumPhis) 2311 if (NumPhis > 2) // Disable this xform. 2312 return false; 2313 2314 // Finally, if everything is ok, fold the branches to logical ops. 2315 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2316 2317 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2318 << "AND: " << *BI->getParent()); 2319 2320 2321 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2322 // branch in it, where one edge (OtherDest) goes back to itself but the other 2323 // exits. We don't *know* that the program avoids the infinite loop 2324 // (even though that seems likely). If we do this xform naively, we'll end up 2325 // recursively unpeeling the loop. Since we know that (after the xform is 2326 // done) that the block *is* infinite if reached, we just make it an obviously 2327 // infinite loop with no cond branch. 2328 if (OtherDest == BB) { 2329 // Insert it at the end of the function, because it's either code, 2330 // or it won't matter if it's hot. :) 2331 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2332 "infloop", BB->getParent()); 2333 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2334 OtherDest = InfLoopBlock; 2335 } 2336 2337 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2338 2339 // BI may have other predecessors. Because of this, we leave 2340 // it alone, but modify PBI. 2341 2342 // Make sure we get to CommonDest on True&True directions. 2343 Value *PBICond = PBI->getCondition(); 2344 IRBuilder<true, NoFolder> Builder(PBI); 2345 if (PBIOp) 2346 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2347 2348 Value *BICond = BI->getCondition(); 2349 if (BIOp) 2350 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2351 2352 // Merge the conditions. 2353 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2354 2355 // Modify PBI to branch on the new condition to the new dests. 2356 PBI->setCondition(Cond); 2357 PBI->setSuccessor(0, CommonDest); 2358 PBI->setSuccessor(1, OtherDest); 2359 2360 // Update branch weight for PBI. 2361 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2362 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2363 PredFalseWeight); 2364 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2365 SuccFalseWeight); 2366 if (PredHasWeights && SuccHasWeights) { 2367 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2368 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2369 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2370 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2371 // The weight to CommonDest should be PredCommon * SuccTotal + 2372 // PredOther * SuccCommon. 2373 // The weight to OtherDest should be PredOther * SuccOther. 2374 SmallVector<uint64_t, 2> NewWeights; 2375 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2376 PredOther * SuccCommon); 2377 NewWeights.push_back(PredOther * SuccOther); 2378 // Halve the weights if any of them cannot fit in an uint32_t 2379 FitWeights(NewWeights); 2380 2381 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2382 PBI->setMetadata(LLVMContext::MD_prof, 2383 MDBuilder(BI->getContext()). 2384 createBranchWeights(MDWeights)); 2385 } 2386 2387 // OtherDest may have phi nodes. If so, add an entry from PBI's 2388 // block that are identical to the entries for BI's block. 2389 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2390 2391 // We know that the CommonDest already had an edge from PBI to 2392 // it. If it has PHIs though, the PHIs may have different 2393 // entries for BB and PBI's BB. If so, insert a select to make 2394 // them agree. 2395 PHINode *PN; 2396 for (BasicBlock::iterator II = CommonDest->begin(); 2397 (PN = dyn_cast<PHINode>(II)); ++II) { 2398 Value *BIV = PN->getIncomingValueForBlock(BB); 2399 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2400 Value *PBIV = PN->getIncomingValue(PBBIdx); 2401 if (BIV != PBIV) { 2402 // Insert a select in PBI to pick the right value. 2403 Value *NV = cast<SelectInst> 2404 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2405 PN->setIncomingValue(PBBIdx, NV); 2406 } 2407 } 2408 2409 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2410 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2411 2412 // This basic block is probably dead. We know it has at least 2413 // one fewer predecessor. 2414 return true; 2415 } 2416 2417 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2418 // branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2419 // Takes care of updating the successors and removing the old terminator. 2420 // Also makes sure not to introduce new successors by assuming that edges to 2421 // non-successor TrueBBs and FalseBBs aren't reachable. 2422 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2423 BasicBlock *TrueBB, BasicBlock *FalseBB, 2424 uint32_t TrueWeight, 2425 uint32_t FalseWeight){ 2426 // Remove any superfluous successor edges from the CFG. 2427 // First, figure out which successors to preserve. 2428 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2429 // successor. 2430 BasicBlock *KeepEdge1 = TrueBB; 2431 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2432 2433 // Then remove the rest. 2434 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2435 BasicBlock *Succ = OldTerm->getSuccessor(I); 2436 // Make sure only to keep exactly one copy of each edge. 2437 if (Succ == KeepEdge1) 2438 KeepEdge1 = 0; 2439 else if (Succ == KeepEdge2) 2440 KeepEdge2 = 0; 2441 else 2442 Succ->removePredecessor(OldTerm->getParent()); 2443 } 2444 2445 IRBuilder<> Builder(OldTerm); 2446 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2447 2448 // Insert an appropriate new terminator. 2449 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2450 if (TrueBB == FalseBB) 2451 // We were only looking for one successor, and it was present. 2452 // Create an unconditional branch to it. 2453 Builder.CreateBr(TrueBB); 2454 else { 2455 // We found both of the successors we were looking for. 2456 // Create a conditional branch sharing the condition of the select. 2457 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2458 if (TrueWeight != FalseWeight) 2459 NewBI->setMetadata(LLVMContext::MD_prof, 2460 MDBuilder(OldTerm->getContext()). 2461 createBranchWeights(TrueWeight, FalseWeight)); 2462 } 2463 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2464 // Neither of the selected blocks were successors, so this 2465 // terminator must be unreachable. 2466 new UnreachableInst(OldTerm->getContext(), OldTerm); 2467 } else { 2468 // One of the selected values was a successor, but the other wasn't. 2469 // Insert an unconditional branch to the one that was found; 2470 // the edge to the one that wasn't must be unreachable. 2471 if (KeepEdge1 == 0) 2472 // Only TrueBB was found. 2473 Builder.CreateBr(TrueBB); 2474 else 2475 // Only FalseBB was found. 2476 Builder.CreateBr(FalseBB); 2477 } 2478 2479 EraseTerminatorInstAndDCECond(OldTerm); 2480 return true; 2481 } 2482 2483 // SimplifySwitchOnSelect - Replaces 2484 // (switch (select cond, X, Y)) on constant X, Y 2485 // with a branch - conditional if X and Y lead to distinct BBs, 2486 // unconditional otherwise. 2487 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2488 // Check for constant integer values in the select. 2489 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2490 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2491 if (!TrueVal || !FalseVal) 2492 return false; 2493 2494 // Find the relevant condition and destinations. 2495 Value *Condition = Select->getCondition(); 2496 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2497 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2498 2499 // Get weight for TrueBB and FalseBB. 2500 uint32_t TrueWeight = 0, FalseWeight = 0; 2501 SmallVector<uint64_t, 8> Weights; 2502 bool HasWeights = HasBranchWeights(SI); 2503 if (HasWeights) { 2504 GetBranchWeights(SI, Weights); 2505 if (Weights.size() == 1 + SI->getNumCases()) { 2506 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2507 getSuccessorIndex()]; 2508 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2509 getSuccessorIndex()]; 2510 } 2511 } 2512 2513 // Perform the actual simplification. 2514 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2515 TrueWeight, FalseWeight); 2516 } 2517 2518 // SimplifyIndirectBrOnSelect - Replaces 2519 // (indirectbr (select cond, blockaddress(@fn, BlockA), 2520 // blockaddress(@fn, BlockB))) 2521 // with 2522 // (br cond, BlockA, BlockB). 2523 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2524 // Check that both operands of the select are block addresses. 2525 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2526 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2527 if (!TBA || !FBA) 2528 return false; 2529 2530 // Extract the actual blocks. 2531 BasicBlock *TrueBB = TBA->getBasicBlock(); 2532 BasicBlock *FalseBB = FBA->getBasicBlock(); 2533 2534 // Perform the actual simplification. 2535 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2536 0, 0); 2537 } 2538 2539 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2540 /// instruction (a seteq/setne with a constant) as the only instruction in a 2541 /// block that ends with an uncond branch. We are looking for a very specific 2542 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2543 /// this case, we merge the first two "or's of icmp" into a switch, but then the 2544 /// default value goes to an uncond block with a seteq in it, we get something 2545 /// like: 2546 /// 2547 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2548 /// DEFAULT: 2549 /// %tmp = icmp eq i8 %A, 92 2550 /// br label %end 2551 /// end: 2552 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2553 /// 2554 /// We prefer to split the edge to 'end' so that there is a true/false entry to 2555 /// the PHI, merging the third icmp into the switch. 2556 static bool TryToSimplifyUncondBranchWithICmpInIt( 2557 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2558 const DataLayout *TD) { 2559 BasicBlock *BB = ICI->getParent(); 2560 2561 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2562 // complex. 2563 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2564 2565 Value *V = ICI->getOperand(0); 2566 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2567 2568 // The pattern we're looking for is where our only predecessor is a switch on 2569 // 'V' and this block is the default case for the switch. In this case we can 2570 // fold the compared value into the switch to simplify things. 2571 BasicBlock *Pred = BB->getSinglePredecessor(); 2572 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2573 2574 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2575 if (SI->getCondition() != V) 2576 return false; 2577 2578 // If BB is reachable on a non-default case, then we simply know the value of 2579 // V in this block. Substitute it and constant fold the icmp instruction 2580 // away. 2581 if (SI->getDefaultDest() != BB) { 2582 ConstantInt *VVal = SI->findCaseDest(BB); 2583 assert(VVal && "Should have a unique destination value"); 2584 ICI->setOperand(0, VVal); 2585 2586 if (Value *V = SimplifyInstruction(ICI, TD)) { 2587 ICI->replaceAllUsesWith(V); 2588 ICI->eraseFromParent(); 2589 } 2590 // BB is now empty, so it is likely to simplify away. 2591 return SimplifyCFG(BB, TTI, TD) | true; 2592 } 2593 2594 // Ok, the block is reachable from the default dest. If the constant we're 2595 // comparing exists in one of the other edges, then we can constant fold ICI 2596 // and zap it. 2597 if (SI->findCaseValue(Cst) != SI->case_default()) { 2598 Value *V; 2599 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2600 V = ConstantInt::getFalse(BB->getContext()); 2601 else 2602 V = ConstantInt::getTrue(BB->getContext()); 2603 2604 ICI->replaceAllUsesWith(V); 2605 ICI->eraseFromParent(); 2606 // BB is now empty, so it is likely to simplify away. 2607 return SimplifyCFG(BB, TTI, TD) | true; 2608 } 2609 2610 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2611 // the block. 2612 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2613 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2614 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2615 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2616 return false; 2617 2618 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2619 // true in the PHI. 2620 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2621 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2622 2623 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2624 std::swap(DefaultCst, NewCst); 2625 2626 // Replace ICI (which is used by the PHI for the default value) with true or 2627 // false depending on if it is EQ or NE. 2628 ICI->replaceAllUsesWith(DefaultCst); 2629 ICI->eraseFromParent(); 2630 2631 // Okay, the switch goes to this block on a default value. Add an edge from 2632 // the switch to the merge point on the compared value. 2633 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2634 BB->getParent(), BB); 2635 SmallVector<uint64_t, 8> Weights; 2636 bool HasWeights = HasBranchWeights(SI); 2637 if (HasWeights) { 2638 GetBranchWeights(SI, Weights); 2639 if (Weights.size() == 1 + SI->getNumCases()) { 2640 // Split weight for default case to case for "Cst". 2641 Weights[0] = (Weights[0]+1) >> 1; 2642 Weights.push_back(Weights[0]); 2643 2644 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2645 SI->setMetadata(LLVMContext::MD_prof, 2646 MDBuilder(SI->getContext()). 2647 createBranchWeights(MDWeights)); 2648 } 2649 } 2650 SI->addCase(Cst, NewBB); 2651 2652 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2653 Builder.SetInsertPoint(NewBB); 2654 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2655 Builder.CreateBr(SuccBlock); 2656 PHIUse->addIncoming(NewCst, NewBB); 2657 return true; 2658 } 2659 2660 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2661 /// Check to see if it is branching on an or/and chain of icmp instructions, and 2662 /// fold it into a switch instruction if so. 2663 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2664 IRBuilder<> &Builder) { 2665 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2666 if (Cond == 0) return false; 2667 2668 2669 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2670 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2671 // 'setne's and'ed together, collect them. 2672 Value *CompVal = 0; 2673 std::vector<ConstantInt*> Values; 2674 bool TrueWhenEqual = true; 2675 Value *ExtraCase = 0; 2676 unsigned UsedICmps = 0; 2677 2678 if (Cond->getOpcode() == Instruction::Or) { 2679 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2680 UsedICmps); 2681 } else if (Cond->getOpcode() == Instruction::And) { 2682 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2683 UsedICmps); 2684 TrueWhenEqual = false; 2685 } 2686 2687 // If we didn't have a multiply compared value, fail. 2688 if (CompVal == 0) return false; 2689 2690 // Avoid turning single icmps into a switch. 2691 if (UsedICmps <= 1) 2692 return false; 2693 2694 // There might be duplicate constants in the list, which the switch 2695 // instruction can't handle, remove them now. 2696 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2697 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2698 2699 // If Extra was used, we require at least two switch values to do the 2700 // transformation. A switch with one value is just an cond branch. 2701 if (ExtraCase && Values.size() < 2) return false; 2702 2703 // TODO: Preserve branch weight metadata, similarly to how 2704 // FoldValueComparisonIntoPredecessors preserves it. 2705 2706 // Figure out which block is which destination. 2707 BasicBlock *DefaultBB = BI->getSuccessor(1); 2708 BasicBlock *EdgeBB = BI->getSuccessor(0); 2709 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2710 2711 BasicBlock *BB = BI->getParent(); 2712 2713 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2714 << " cases into SWITCH. BB is:\n" << *BB); 2715 2716 // If there are any extra values that couldn't be folded into the switch 2717 // then we evaluate them with an explicit branch first. Split the block 2718 // right before the condbr to handle it. 2719 if (ExtraCase) { 2720 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2721 // Remove the uncond branch added to the old block. 2722 TerminatorInst *OldTI = BB->getTerminator(); 2723 Builder.SetInsertPoint(OldTI); 2724 2725 if (TrueWhenEqual) 2726 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2727 else 2728 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2729 2730 OldTI->eraseFromParent(); 2731 2732 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2733 // for the edge we just added. 2734 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2735 2736 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2737 << "\nEXTRABB = " << *BB); 2738 BB = NewBB; 2739 } 2740 2741 Builder.SetInsertPoint(BI); 2742 // Convert pointer to int before we switch. 2743 if (CompVal->getType()->isPointerTy()) { 2744 assert(TD && "Cannot switch on pointer without DataLayout"); 2745 CompVal = Builder.CreatePtrToInt(CompVal, 2746 TD->getIntPtrType(CompVal->getContext()), 2747 "magicptr"); 2748 } 2749 2750 // Create the new switch instruction now. 2751 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2752 2753 // Add all of the 'cases' to the switch instruction. 2754 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2755 New->addCase(Values[i], EdgeBB); 2756 2757 // We added edges from PI to the EdgeBB. As such, if there were any 2758 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2759 // the number of edges added. 2760 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2761 isa<PHINode>(BBI); ++BBI) { 2762 PHINode *PN = cast<PHINode>(BBI); 2763 Value *InVal = PN->getIncomingValueForBlock(BB); 2764 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2765 PN->addIncoming(InVal, BB); 2766 } 2767 2768 // Erase the old branch instruction. 2769 EraseTerminatorInstAndDCECond(BI); 2770 2771 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2772 return true; 2773 } 2774 2775 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2776 // If this is a trivial landing pad that just continues unwinding the caught 2777 // exception then zap the landing pad, turning its invokes into calls. 2778 BasicBlock *BB = RI->getParent(); 2779 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2780 if (RI->getValue() != LPInst) 2781 // Not a landing pad, or the resume is not unwinding the exception that 2782 // caused control to branch here. 2783 return false; 2784 2785 // Check that there are no other instructions except for debug intrinsics. 2786 BasicBlock::iterator I = LPInst, E = RI; 2787 while (++I != E) 2788 if (!isa<DbgInfoIntrinsic>(I)) 2789 return false; 2790 2791 // Turn all invokes that unwind here into calls and delete the basic block. 2792 bool InvokeRequiresTableEntry = false; 2793 bool Changed = false; 2794 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2795 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2796 2797 if (II->hasFnAttr(Attribute::UWTable)) { 2798 // Don't remove an `invoke' instruction if the ABI requires an entry into 2799 // the table. 2800 InvokeRequiresTableEntry = true; 2801 continue; 2802 } 2803 2804 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2805 2806 // Insert a call instruction before the invoke. 2807 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2808 Call->takeName(II); 2809 Call->setCallingConv(II->getCallingConv()); 2810 Call->setAttributes(II->getAttributes()); 2811 Call->setDebugLoc(II->getDebugLoc()); 2812 2813 // Anything that used the value produced by the invoke instruction now uses 2814 // the value produced by the call instruction. Note that we do this even 2815 // for void functions and calls with no uses so that the callgraph edge is 2816 // updated. 2817 II->replaceAllUsesWith(Call); 2818 BB->removePredecessor(II->getParent()); 2819 2820 // Insert a branch to the normal destination right before the invoke. 2821 BranchInst::Create(II->getNormalDest(), II); 2822 2823 // Finally, delete the invoke instruction! 2824 II->eraseFromParent(); 2825 Changed = true; 2826 } 2827 2828 if (!InvokeRequiresTableEntry) 2829 // The landingpad is now unreachable. Zap it. 2830 BB->eraseFromParent(); 2831 2832 return Changed; 2833 } 2834 2835 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2836 BasicBlock *BB = RI->getParent(); 2837 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2838 2839 // Find predecessors that end with branches. 2840 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2841 SmallVector<BranchInst*, 8> CondBranchPreds; 2842 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2843 BasicBlock *P = *PI; 2844 TerminatorInst *PTI = P->getTerminator(); 2845 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2846 if (BI->isUnconditional()) 2847 UncondBranchPreds.push_back(P); 2848 else 2849 CondBranchPreds.push_back(BI); 2850 } 2851 } 2852 2853 // If we found some, do the transformation! 2854 if (!UncondBranchPreds.empty() && DupRet) { 2855 while (!UncondBranchPreds.empty()) { 2856 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2857 DEBUG(dbgs() << "FOLDING: " << *BB 2858 << "INTO UNCOND BRANCH PRED: " << *Pred); 2859 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2860 } 2861 2862 // If we eliminated all predecessors of the block, delete the block now. 2863 if (pred_begin(BB) == pred_end(BB)) 2864 // We know there are no successors, so just nuke the block. 2865 BB->eraseFromParent(); 2866 2867 return true; 2868 } 2869 2870 // Check out all of the conditional branches going to this return 2871 // instruction. If any of them just select between returns, change the 2872 // branch itself into a select/return pair. 2873 while (!CondBranchPreds.empty()) { 2874 BranchInst *BI = CondBranchPreds.pop_back_val(); 2875 2876 // Check to see if the non-BB successor is also a return block. 2877 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2878 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2879 SimplifyCondBranchToTwoReturns(BI, Builder)) 2880 return true; 2881 } 2882 return false; 2883 } 2884 2885 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2886 BasicBlock *BB = UI->getParent(); 2887 2888 bool Changed = false; 2889 2890 // If there are any instructions immediately before the unreachable that can 2891 // be removed, do so. 2892 while (UI != BB->begin()) { 2893 BasicBlock::iterator BBI = UI; 2894 --BBI; 2895 // Do not delete instructions that can have side effects which might cause 2896 // the unreachable to not be reachable; specifically, calls and volatile 2897 // operations may have this effect. 2898 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2899 2900 if (BBI->mayHaveSideEffects()) { 2901 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2902 if (SI->isVolatile()) 2903 break; 2904 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2905 if (LI->isVolatile()) 2906 break; 2907 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2908 if (RMWI->isVolatile()) 2909 break; 2910 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2911 if (CXI->isVolatile()) 2912 break; 2913 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2914 !isa<LandingPadInst>(BBI)) { 2915 break; 2916 } 2917 // Note that deleting LandingPad's here is in fact okay, although it 2918 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2919 // all the predecessors of this block will be the unwind edges of Invokes, 2920 // and we can therefore guarantee this block will be erased. 2921 } 2922 2923 // Delete this instruction (any uses are guaranteed to be dead) 2924 if (!BBI->use_empty()) 2925 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2926 BBI->eraseFromParent(); 2927 Changed = true; 2928 } 2929 2930 // If the unreachable instruction is the first in the block, take a gander 2931 // at all of the predecessors of this instruction, and simplify them. 2932 if (&BB->front() != UI) return Changed; 2933 2934 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2935 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2936 TerminatorInst *TI = Preds[i]->getTerminator(); 2937 IRBuilder<> Builder(TI); 2938 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2939 if (BI->isUnconditional()) { 2940 if (BI->getSuccessor(0) == BB) { 2941 new UnreachableInst(TI->getContext(), TI); 2942 TI->eraseFromParent(); 2943 Changed = true; 2944 } 2945 } else { 2946 if (BI->getSuccessor(0) == BB) { 2947 Builder.CreateBr(BI->getSuccessor(1)); 2948 EraseTerminatorInstAndDCECond(BI); 2949 } else if (BI->getSuccessor(1) == BB) { 2950 Builder.CreateBr(BI->getSuccessor(0)); 2951 EraseTerminatorInstAndDCECond(BI); 2952 Changed = true; 2953 } 2954 } 2955 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2956 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2957 i != e; ++i) 2958 if (i.getCaseSuccessor() == BB) { 2959 BB->removePredecessor(SI->getParent()); 2960 SI->removeCase(i); 2961 --i; --e; 2962 Changed = true; 2963 } 2964 // If the default value is unreachable, figure out the most popular 2965 // destination and make it the default. 2966 if (SI->getDefaultDest() == BB) { 2967 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2968 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2969 i != e; ++i) { 2970 std::pair<unsigned, unsigned> &entry = 2971 Popularity[i.getCaseSuccessor()]; 2972 if (entry.first == 0) { 2973 entry.first = 1; 2974 entry.second = i.getCaseIndex(); 2975 } else { 2976 entry.first++; 2977 } 2978 } 2979 2980 // Find the most popular block. 2981 unsigned MaxPop = 0; 2982 unsigned MaxIndex = 0; 2983 BasicBlock *MaxBlock = 0; 2984 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2985 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2986 if (I->second.first > MaxPop || 2987 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2988 MaxPop = I->second.first; 2989 MaxIndex = I->second.second; 2990 MaxBlock = I->first; 2991 } 2992 } 2993 if (MaxBlock) { 2994 // Make this the new default, allowing us to delete any explicit 2995 // edges to it. 2996 SI->setDefaultDest(MaxBlock); 2997 Changed = true; 2998 2999 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3000 // it. 3001 if (isa<PHINode>(MaxBlock->begin())) 3002 for (unsigned i = 0; i != MaxPop-1; ++i) 3003 MaxBlock->removePredecessor(SI->getParent()); 3004 3005 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3006 i != e; ++i) 3007 if (i.getCaseSuccessor() == MaxBlock) { 3008 SI->removeCase(i); 3009 --i; --e; 3010 } 3011 } 3012 } 3013 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3014 if (II->getUnwindDest() == BB) { 3015 // Convert the invoke to a call instruction. This would be a good 3016 // place to note that the call does not throw though. 3017 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3018 II->removeFromParent(); // Take out of symbol table 3019 3020 // Insert the call now... 3021 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3022 Builder.SetInsertPoint(BI); 3023 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3024 Args, II->getName()); 3025 CI->setCallingConv(II->getCallingConv()); 3026 CI->setAttributes(II->getAttributes()); 3027 // If the invoke produced a value, the call does now instead. 3028 II->replaceAllUsesWith(CI); 3029 delete II; 3030 Changed = true; 3031 } 3032 } 3033 } 3034 3035 // If this block is now dead, remove it. 3036 if (pred_begin(BB) == pred_end(BB) && 3037 BB != &BB->getParent()->getEntryBlock()) { 3038 // We know there are no successors, so just nuke the block. 3039 BB->eraseFromParent(); 3040 return true; 3041 } 3042 3043 return Changed; 3044 } 3045 3046 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3047 /// integer range comparison into a sub, an icmp and a branch. 3048 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3049 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3050 3051 // Make sure all cases point to the same destination and gather the values. 3052 SmallVector<ConstantInt *, 16> Cases; 3053 SwitchInst::CaseIt I = SI->case_begin(); 3054 Cases.push_back(I.getCaseValue()); 3055 SwitchInst::CaseIt PrevI = I++; 3056 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3057 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3058 return false; 3059 Cases.push_back(I.getCaseValue()); 3060 } 3061 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3062 3063 // Sort the case values, then check if they form a range we can transform. 3064 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3065 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3066 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3067 return false; 3068 } 3069 3070 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3071 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3072 3073 Value *Sub = SI->getCondition(); 3074 if (!Offset->isNullValue()) 3075 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3076 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3077 BranchInst *NewBI = Builder.CreateCondBr( 3078 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3079 3080 // Update weight for the newly-created conditional branch. 3081 SmallVector<uint64_t, 8> Weights; 3082 bool HasWeights = HasBranchWeights(SI); 3083 if (HasWeights) { 3084 GetBranchWeights(SI, Weights); 3085 if (Weights.size() == 1 + SI->getNumCases()) { 3086 // Combine all weights for the cases to be the true weight of NewBI. 3087 // We assume that the sum of all weights for a Terminator can fit into 32 3088 // bits. 3089 uint32_t NewTrueWeight = 0; 3090 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3091 NewTrueWeight += (uint32_t)Weights[I]; 3092 NewBI->setMetadata(LLVMContext::MD_prof, 3093 MDBuilder(SI->getContext()). 3094 createBranchWeights(NewTrueWeight, 3095 (uint32_t)Weights[0])); 3096 } 3097 } 3098 3099 // Prune obsolete incoming values off the successor's PHI nodes. 3100 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3101 isa<PHINode>(BBI); ++BBI) { 3102 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3103 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3104 } 3105 SI->eraseFromParent(); 3106 3107 return true; 3108 } 3109 3110 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3111 /// and use it to remove dead cases. 3112 static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3113 Value *Cond = SI->getCondition(); 3114 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3115 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3116 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3117 3118 // Gather dead cases. 3119 SmallVector<ConstantInt*, 8> DeadCases; 3120 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3121 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3122 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3123 DeadCases.push_back(I.getCaseValue()); 3124 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3125 << I.getCaseValue() << "' is dead.\n"); 3126 } 3127 } 3128 3129 SmallVector<uint64_t, 8> Weights; 3130 bool HasWeight = HasBranchWeights(SI); 3131 if (HasWeight) { 3132 GetBranchWeights(SI, Weights); 3133 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3134 } 3135 3136 // Remove dead cases from the switch. 3137 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3138 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3139 assert(Case != SI->case_default() && 3140 "Case was not found. Probably mistake in DeadCases forming."); 3141 if (HasWeight) { 3142 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3143 Weights.pop_back(); 3144 } 3145 3146 // Prune unused values from PHI nodes. 3147 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3148 SI->removeCase(Case); 3149 } 3150 if (HasWeight) { 3151 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3152 SI->setMetadata(LLVMContext::MD_prof, 3153 MDBuilder(SI->getParent()->getContext()). 3154 createBranchWeights(MDWeights)); 3155 } 3156 3157 return !DeadCases.empty(); 3158 } 3159 3160 /// FindPHIForConditionForwarding - If BB would be eligible for simplification 3161 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3162 /// by an unconditional branch), look at the phi node for BB in the successor 3163 /// block and see if the incoming value is equal to CaseValue. If so, return 3164 /// the phi node, and set PhiIndex to BB's index in the phi node. 3165 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3166 BasicBlock *BB, 3167 int *PhiIndex) { 3168 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3169 return NULL; // BB must be empty to be a candidate for simplification. 3170 if (!BB->getSinglePredecessor()) 3171 return NULL; // BB must be dominated by the switch. 3172 3173 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3174 if (!Branch || !Branch->isUnconditional()) 3175 return NULL; // Terminator must be unconditional branch. 3176 3177 BasicBlock *Succ = Branch->getSuccessor(0); 3178 3179 BasicBlock::iterator I = Succ->begin(); 3180 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3181 int Idx = PHI->getBasicBlockIndex(BB); 3182 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3183 3184 Value *InValue = PHI->getIncomingValue(Idx); 3185 if (InValue != CaseValue) continue; 3186 3187 *PhiIndex = Idx; 3188 return PHI; 3189 } 3190 3191 return NULL; 3192 } 3193 3194 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3195 /// instruction to a phi node dominated by the switch, if that would mean that 3196 /// some of the destination blocks of the switch can be folded away. 3197 /// Returns true if a change is made. 3198 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3199 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3200 ForwardingNodesMap ForwardingNodes; 3201 3202 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3203 ConstantInt *CaseValue = I.getCaseValue(); 3204 BasicBlock *CaseDest = I.getCaseSuccessor(); 3205 3206 int PhiIndex; 3207 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3208 &PhiIndex); 3209 if (!PHI) continue; 3210 3211 ForwardingNodes[PHI].push_back(PhiIndex); 3212 } 3213 3214 bool Changed = false; 3215 3216 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3217 E = ForwardingNodes.end(); I != E; ++I) { 3218 PHINode *Phi = I->first; 3219 SmallVector<int,4> &Indexes = I->second; 3220 3221 if (Indexes.size() < 2) continue; 3222 3223 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3224 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3225 Changed = true; 3226 } 3227 3228 return Changed; 3229 } 3230 3231 /// ValidLookupTableConstant - Return true if the backend will be able to handle 3232 /// initializing an array of constants like C. 3233 static bool ValidLookupTableConstant(Constant *C) { 3234 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3235 return CE->isGEPWithNoNotionalOverIndexing(); 3236 3237 return isa<ConstantFP>(C) || 3238 isa<ConstantInt>(C) || 3239 isa<ConstantPointerNull>(C) || 3240 isa<GlobalValue>(C) || 3241 isa<UndefValue>(C); 3242 } 3243 3244 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3245 /// its constant value in ConstantPool, returning 0 if it's not there. 3246 static Constant *LookupConstant(Value *V, 3247 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3248 if (Constant *C = dyn_cast<Constant>(V)) 3249 return C; 3250 return ConstantPool.lookup(V); 3251 } 3252 3253 /// ConstantFold - Try to fold instruction I into a constant. This works for 3254 /// simple instructions such as binary operations where both operands are 3255 /// constant or can be replaced by constants from the ConstantPool. Returns the 3256 /// resulting constant on success, 0 otherwise. 3257 static Constant *ConstantFold(Instruction *I, 3258 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3259 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3260 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3261 if (!A) 3262 return 0; 3263 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3264 if (!B) 3265 return 0; 3266 return ConstantExpr::get(BO->getOpcode(), A, B); 3267 } 3268 3269 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3270 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3271 if (!A) 3272 return 0; 3273 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3274 if (!B) 3275 return 0; 3276 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3277 } 3278 3279 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3280 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3281 if (!A) 3282 return 0; 3283 if (A->isAllOnesValue()) 3284 return LookupConstant(Select->getTrueValue(), ConstantPool); 3285 if (A->isNullValue()) 3286 return LookupConstant(Select->getFalseValue(), ConstantPool); 3287 return 0; 3288 } 3289 3290 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3291 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3292 if (!A) 3293 return 0; 3294 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3295 } 3296 3297 return 0; 3298 } 3299 3300 /// GetCaseResults - Try to determine the resulting constant values in phi nodes 3301 /// at the common destination basic block, *CommonDest, for one of the case 3302 /// destionations CaseDest corresponding to value CaseVal (0 for the default 3303 /// case), of a switch instruction SI. 3304 static bool GetCaseResults(SwitchInst *SI, 3305 ConstantInt *CaseVal, 3306 BasicBlock *CaseDest, 3307 BasicBlock **CommonDest, 3308 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3309 // The block from which we enter the common destination. 3310 BasicBlock *Pred = SI->getParent(); 3311 3312 // If CaseDest is empty except for some side-effect free instructions through 3313 // which we can constant-propagate the CaseVal, continue to its successor. 3314 SmallDenseMap<Value*, Constant*> ConstantPool; 3315 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3316 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3317 ++I) { 3318 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3319 // If the terminator is a simple branch, continue to the next block. 3320 if (T->getNumSuccessors() != 1) 3321 return false; 3322 Pred = CaseDest; 3323 CaseDest = T->getSuccessor(0); 3324 } else if (isa<DbgInfoIntrinsic>(I)) { 3325 // Skip debug intrinsic. 3326 continue; 3327 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3328 // Instruction is side-effect free and constant. 3329 ConstantPool.insert(std::make_pair(I, C)); 3330 } else { 3331 break; 3332 } 3333 } 3334 3335 // If we did not have a CommonDest before, use the current one. 3336 if (!*CommonDest) 3337 *CommonDest = CaseDest; 3338 // If the destination isn't the common one, abort. 3339 if (CaseDest != *CommonDest) 3340 return false; 3341 3342 // Get the values for this case from phi nodes in the destination block. 3343 BasicBlock::iterator I = (*CommonDest)->begin(); 3344 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3345 int Idx = PHI->getBasicBlockIndex(Pred); 3346 if (Idx == -1) 3347 continue; 3348 3349 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3350 ConstantPool); 3351 if (!ConstVal) 3352 return false; 3353 3354 // Note: If the constant comes from constant-propagating the case value 3355 // through the CaseDest basic block, it will be safe to remove the 3356 // instructions in that block. They cannot be used (except in the phi nodes 3357 // we visit) outside CaseDest, because that block does not dominate its 3358 // successor. If it did, we would not be in this phi node. 3359 3360 // Be conservative about which kinds of constants we support. 3361 if (!ValidLookupTableConstant(ConstVal)) 3362 return false; 3363 3364 Res.push_back(std::make_pair(PHI, ConstVal)); 3365 } 3366 3367 return true; 3368 } 3369 3370 namespace { 3371 /// SwitchLookupTable - This class represents a lookup table that can be used 3372 /// to replace a switch. 3373 class SwitchLookupTable { 3374 public: 3375 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3376 /// with the contents of Values, using DefaultValue to fill any holes in the 3377 /// table. 3378 SwitchLookupTable(Module &M, 3379 uint64_t TableSize, 3380 ConstantInt *Offset, 3381 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3382 Constant *DefaultValue, 3383 const DataLayout *TD); 3384 3385 /// BuildLookup - Build instructions with Builder to retrieve the value at 3386 /// the position given by Index in the lookup table. 3387 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3388 3389 /// WouldFitInRegister - Return true if a table with TableSize elements of 3390 /// type ElementType would fit in a target-legal register. 3391 static bool WouldFitInRegister(const DataLayout *TD, 3392 uint64_t TableSize, 3393 const Type *ElementType); 3394 3395 private: 3396 // Depending on the contents of the table, it can be represented in 3397 // different ways. 3398 enum { 3399 // For tables where each element contains the same value, we just have to 3400 // store that single value and return it for each lookup. 3401 SingleValueKind, 3402 3403 // For small tables with integer elements, we can pack them into a bitmap 3404 // that fits into a target-legal register. Values are retrieved by 3405 // shift and mask operations. 3406 BitMapKind, 3407 3408 // The table is stored as an array of values. Values are retrieved by load 3409 // instructions from the table. 3410 ArrayKind 3411 } Kind; 3412 3413 // For SingleValueKind, this is the single value. 3414 Constant *SingleValue; 3415 3416 // For BitMapKind, this is the bitmap. 3417 ConstantInt *BitMap; 3418 IntegerType *BitMapElementTy; 3419 3420 // For ArrayKind, this is the array. 3421 GlobalVariable *Array; 3422 }; 3423 } 3424 3425 SwitchLookupTable::SwitchLookupTable(Module &M, 3426 uint64_t TableSize, 3427 ConstantInt *Offset, 3428 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3429 Constant *DefaultValue, 3430 const DataLayout *TD) 3431 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3432 assert(Values.size() && "Can't build lookup table without values!"); 3433 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3434 3435 // If all values in the table are equal, this is that value. 3436 SingleValue = Values.begin()->second; 3437 3438 // Build up the table contents. 3439 SmallVector<Constant*, 64> TableContents(TableSize); 3440 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3441 ConstantInt *CaseVal = Values[I].first; 3442 Constant *CaseRes = Values[I].second; 3443 assert(CaseRes->getType() == DefaultValue->getType()); 3444 3445 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3446 .getLimitedValue(); 3447 TableContents[Idx] = CaseRes; 3448 3449 if (CaseRes != SingleValue) 3450 SingleValue = 0; 3451 } 3452 3453 // Fill in any holes in the table with the default result. 3454 if (Values.size() < TableSize) { 3455 for (uint64_t I = 0; I < TableSize; ++I) { 3456 if (!TableContents[I]) 3457 TableContents[I] = DefaultValue; 3458 } 3459 3460 if (DefaultValue != SingleValue) 3461 SingleValue = 0; 3462 } 3463 3464 // If each element in the table contains the same value, we only need to store 3465 // that single value. 3466 if (SingleValue) { 3467 Kind = SingleValueKind; 3468 return; 3469 } 3470 3471 // If the type is integer and the table fits in a register, build a bitmap. 3472 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3473 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3474 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3475 for (uint64_t I = TableSize; I > 0; --I) { 3476 TableInt <<= IT->getBitWidth(); 3477 // Insert values into the bitmap. Undef values are set to zero. 3478 if (!isa<UndefValue>(TableContents[I - 1])) { 3479 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3480 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3481 } 3482 } 3483 BitMap = ConstantInt::get(M.getContext(), TableInt); 3484 BitMapElementTy = IT; 3485 Kind = BitMapKind; 3486 ++NumBitMaps; 3487 return; 3488 } 3489 3490 // Store the table in an array. 3491 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3492 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3493 3494 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3495 GlobalVariable::PrivateLinkage, 3496 Initializer, 3497 "switch.table"); 3498 Array->setUnnamedAddr(true); 3499 Kind = ArrayKind; 3500 } 3501 3502 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3503 switch (Kind) { 3504 case SingleValueKind: 3505 return SingleValue; 3506 case BitMapKind: { 3507 // Type of the bitmap (e.g. i59). 3508 IntegerType *MapTy = BitMap->getType(); 3509 3510 // Cast Index to the same type as the bitmap. 3511 // Note: The Index is <= the number of elements in the table, so 3512 // truncating it to the width of the bitmask is safe. 3513 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3514 3515 // Multiply the shift amount by the element width. 3516 ShiftAmt = Builder.CreateMul(ShiftAmt, 3517 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3518 "switch.shiftamt"); 3519 3520 // Shift down. 3521 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3522 "switch.downshift"); 3523 // Mask off. 3524 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3525 "switch.masked"); 3526 } 3527 case ArrayKind: { 3528 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3529 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3530 "switch.gep"); 3531 return Builder.CreateLoad(GEP, "switch.load"); 3532 } 3533 } 3534 llvm_unreachable("Unknown lookup table kind!"); 3535 } 3536 3537 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3538 uint64_t TableSize, 3539 const Type *ElementType) { 3540 if (!TD) 3541 return false; 3542 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3543 if (!IT) 3544 return false; 3545 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3546 // are <= 15, we could try to narrow the type. 3547 3548 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3549 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3550 return false; 3551 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3552 } 3553 3554 /// ShouldBuildLookupTable - Determine whether a lookup table should be built 3555 /// for this switch, based on the number of caes, size of the table and the 3556 /// types of the results. 3557 static bool ShouldBuildLookupTable(SwitchInst *SI, 3558 uint64_t TableSize, 3559 const TargetTransformInfo &TTI, 3560 const DataLayout *TD, 3561 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3562 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3563 return false; // TableSize overflowed, or mul below might overflow. 3564 3565 bool AllTablesFitInRegister = true; 3566 bool HasIllegalType = false; 3567 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3568 E = ResultTypes.end(); I != E; ++I) { 3569 Type *Ty = I->second; 3570 3571 // Saturate this flag to true. 3572 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3573 3574 // Saturate this flag to false. 3575 AllTablesFitInRegister = AllTablesFitInRegister && 3576 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3577 3578 // If both flags saturate, we're done. NOTE: This *only* works with 3579 // saturating flags, and all flags have to saturate first due to the 3580 // non-deterministic behavior of iterating over a dense map. 3581 if (HasIllegalType && !AllTablesFitInRegister) 3582 break; 3583 } 3584 3585 // If each table would fit in a register, we should build it anyway. 3586 if (AllTablesFitInRegister) 3587 return true; 3588 3589 // Don't build a table that doesn't fit in-register if it has illegal types. 3590 if (HasIllegalType) 3591 return false; 3592 3593 // The table density should be at least 40%. This is the same criterion as for 3594 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3595 // FIXME: Find the best cut-off. 3596 return SI->getNumCases() * 10 >= TableSize * 4; 3597 } 3598 3599 /// SwitchToLookupTable - If the switch is only used to initialize one or more 3600 /// phi nodes in a common successor block with different constant values, 3601 /// replace the switch with lookup tables. 3602 static bool SwitchToLookupTable(SwitchInst *SI, 3603 IRBuilder<> &Builder, 3604 const TargetTransformInfo &TTI, 3605 const DataLayout* TD) { 3606 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3607 3608 // Only build lookup table when we have a target that supports it. 3609 if (!TTI.shouldBuildLookupTables()) 3610 return false; 3611 3612 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3613 // split off a dense part and build a lookup table for that. 3614 3615 // FIXME: This creates arrays of GEPs to constant strings, which means each 3616 // GEP needs a runtime relocation in PIC code. We should just build one big 3617 // string and lookup indices into that. 3618 3619 // Ignore the switch if the number of cases is too small. 3620 // This is similar to the check when building jump tables in 3621 // SelectionDAGBuilder::handleJTSwitchCase. 3622 // FIXME: Determine the best cut-off. 3623 if (SI->getNumCases() < 4) 3624 return false; 3625 3626 // Figure out the corresponding result for each case value and phi node in the 3627 // common destination, as well as the the min and max case values. 3628 assert(SI->case_begin() != SI->case_end()); 3629 SwitchInst::CaseIt CI = SI->case_begin(); 3630 ConstantInt *MinCaseVal = CI.getCaseValue(); 3631 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3632 3633 BasicBlock *CommonDest = 0; 3634 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3635 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3636 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3637 SmallDenseMap<PHINode*, Type*> ResultTypes; 3638 SmallVector<PHINode*, 4> PHIs; 3639 3640 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3641 ConstantInt *CaseVal = CI.getCaseValue(); 3642 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3643 MinCaseVal = CaseVal; 3644 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3645 MaxCaseVal = CaseVal; 3646 3647 // Resulting value at phi nodes for this case value. 3648 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3649 ResultsTy Results; 3650 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3651 Results)) 3652 return false; 3653 3654 // Append the result from this case to the list for each phi. 3655 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3656 if (!ResultLists.count(I->first)) 3657 PHIs.push_back(I->first); 3658 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3659 } 3660 } 3661 3662 // Get the resulting values for the default case. 3663 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3664 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3665 DefaultResultsList)) 3666 return false; 3667 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3668 PHINode *PHI = DefaultResultsList[I].first; 3669 Constant *Result = DefaultResultsList[I].second; 3670 DefaultResults[PHI] = Result; 3671 ResultTypes[PHI] = Result->getType(); 3672 } 3673 3674 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3675 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3676 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3677 return false; 3678 3679 // Create the BB that does the lookups. 3680 Module &Mod = *CommonDest->getParent()->getParent(); 3681 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3682 "switch.lookup", 3683 CommonDest->getParent(), 3684 CommonDest); 3685 3686 // Check whether the condition value is within the case range, and branch to 3687 // the new BB. 3688 Builder.SetInsertPoint(SI); 3689 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3690 "switch.tableidx"); 3691 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3692 MinCaseVal->getType(), TableSize)); 3693 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3694 3695 // Populate the BB that does the lookups. 3696 Builder.SetInsertPoint(LookupBB); 3697 bool ReturnedEarly = false; 3698 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3699 PHINode *PHI = PHIs[I]; 3700 3701 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3702 DefaultResults[PHI], TD); 3703 3704 Value *Result = Table.BuildLookup(TableIndex, Builder); 3705 3706 // If the result is used to return immediately from the function, we want to 3707 // do that right here. 3708 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3709 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3710 Builder.CreateRet(Result); 3711 ReturnedEarly = true; 3712 break; 3713 } 3714 3715 PHI->addIncoming(Result, LookupBB); 3716 } 3717 3718 if (!ReturnedEarly) 3719 Builder.CreateBr(CommonDest); 3720 3721 // Remove the switch. 3722 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3723 BasicBlock *Succ = SI->getSuccessor(i); 3724 if (Succ == SI->getDefaultDest()) continue; 3725 Succ->removePredecessor(SI->getParent()); 3726 } 3727 SI->eraseFromParent(); 3728 3729 ++NumLookupTables; 3730 return true; 3731 } 3732 3733 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3734 BasicBlock *BB = SI->getParent(); 3735 3736 if (isValueEqualityComparison(SI)) { 3737 // If we only have one predecessor, and if it is a branch on this value, 3738 // see if that predecessor totally determines the outcome of this switch. 3739 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3740 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3741 return SimplifyCFG(BB, TTI, TD) | true; 3742 3743 Value *Cond = SI->getCondition(); 3744 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3745 if (SimplifySwitchOnSelect(SI, Select)) 3746 return SimplifyCFG(BB, TTI, TD) | true; 3747 3748 // If the block only contains the switch, see if we can fold the block 3749 // away into any preds. 3750 BasicBlock::iterator BBI = BB->begin(); 3751 // Ignore dbg intrinsics. 3752 while (isa<DbgInfoIntrinsic>(BBI)) 3753 ++BBI; 3754 if (SI == &*BBI) 3755 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3756 return SimplifyCFG(BB, TTI, TD) | true; 3757 } 3758 3759 // Try to transform the switch into an icmp and a branch. 3760 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3761 return SimplifyCFG(BB, TTI, TD) | true; 3762 3763 // Remove unreachable cases. 3764 if (EliminateDeadSwitchCases(SI)) 3765 return SimplifyCFG(BB, TTI, TD) | true; 3766 3767 if (ForwardSwitchConditionToPHI(SI)) 3768 return SimplifyCFG(BB, TTI, TD) | true; 3769 3770 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3771 return SimplifyCFG(BB, TTI, TD) | true; 3772 3773 return false; 3774 } 3775 3776 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3777 BasicBlock *BB = IBI->getParent(); 3778 bool Changed = false; 3779 3780 // Eliminate redundant destinations. 3781 SmallPtrSet<Value *, 8> Succs; 3782 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3783 BasicBlock *Dest = IBI->getDestination(i); 3784 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3785 Dest->removePredecessor(BB); 3786 IBI->removeDestination(i); 3787 --i; --e; 3788 Changed = true; 3789 } 3790 } 3791 3792 if (IBI->getNumDestinations() == 0) { 3793 // If the indirectbr has no successors, change it to unreachable. 3794 new UnreachableInst(IBI->getContext(), IBI); 3795 EraseTerminatorInstAndDCECond(IBI); 3796 return true; 3797 } 3798 3799 if (IBI->getNumDestinations() == 1) { 3800 // If the indirectbr has one successor, change it to a direct branch. 3801 BranchInst::Create(IBI->getDestination(0), IBI); 3802 EraseTerminatorInstAndDCECond(IBI); 3803 return true; 3804 } 3805 3806 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3807 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3808 return SimplifyCFG(BB, TTI, TD) | true; 3809 } 3810 return Changed; 3811 } 3812 3813 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3814 BasicBlock *BB = BI->getParent(); 3815 3816 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3817 return true; 3818 3819 // If the Terminator is the only non-phi instruction, simplify the block. 3820 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3821 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3822 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3823 return true; 3824 3825 // If the only instruction in the block is a seteq/setne comparison 3826 // against a constant, try to simplify the block. 3827 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3828 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3829 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3830 ; 3831 if (I->isTerminator() && 3832 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3833 return true; 3834 } 3835 3836 // If this basic block is ONLY a compare and a branch, and if a predecessor 3837 // branches to us and our successor, fold the comparison into the 3838 // predecessor and use logical operations to update the incoming value 3839 // for PHI nodes in common successor. 3840 if (FoldBranchToCommonDest(BI)) 3841 return SimplifyCFG(BB, TTI, TD) | true; 3842 return false; 3843 } 3844 3845 3846 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3847 BasicBlock *BB = BI->getParent(); 3848 3849 // Conditional branch 3850 if (isValueEqualityComparison(BI)) { 3851 // If we only have one predecessor, and if it is a branch on this value, 3852 // see if that predecessor totally determines the outcome of this 3853 // switch. 3854 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3855 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3856 return SimplifyCFG(BB, TTI, TD) | true; 3857 3858 // This block must be empty, except for the setcond inst, if it exists. 3859 // Ignore dbg intrinsics. 3860 BasicBlock::iterator I = BB->begin(); 3861 // Ignore dbg intrinsics. 3862 while (isa<DbgInfoIntrinsic>(I)) 3863 ++I; 3864 if (&*I == BI) { 3865 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3866 return SimplifyCFG(BB, TTI, TD) | true; 3867 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3868 ++I; 3869 // Ignore dbg intrinsics. 3870 while (isa<DbgInfoIntrinsic>(I)) 3871 ++I; 3872 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3873 return SimplifyCFG(BB, TTI, TD) | true; 3874 } 3875 } 3876 3877 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3878 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3879 return true; 3880 3881 // If this basic block is ONLY a compare and a branch, and if a predecessor 3882 // branches to us and one of our successors, fold the comparison into the 3883 // predecessor and use logical operations to pick the right destination. 3884 if (FoldBranchToCommonDest(BI)) 3885 return SimplifyCFG(BB, TTI, TD) | true; 3886 3887 // We have a conditional branch to two blocks that are only reachable 3888 // from BI. We know that the condbr dominates the two blocks, so see if 3889 // there is any identical code in the "then" and "else" blocks. If so, we 3890 // can hoist it up to the branching block. 3891 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3892 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3893 if (HoistThenElseCodeToIf(BI)) 3894 return SimplifyCFG(BB, TTI, TD) | true; 3895 } else { 3896 // If Successor #1 has multiple preds, we may be able to conditionally 3897 // execute Successor #0 if it branches to successor #1. 3898 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3899 if (Succ0TI->getNumSuccessors() == 1 && 3900 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3901 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3902 return SimplifyCFG(BB, TTI, TD) | true; 3903 } 3904 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3905 // If Successor #0 has multiple preds, we may be able to conditionally 3906 // execute Successor #1 if it branches to successor #0. 3907 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3908 if (Succ1TI->getNumSuccessors() == 1 && 3909 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3910 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3911 return SimplifyCFG(BB, TTI, TD) | true; 3912 } 3913 3914 // If this is a branch on a phi node in the current block, thread control 3915 // through this block if any PHI node entries are constants. 3916 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3917 if (PN->getParent() == BI->getParent()) 3918 if (FoldCondBranchOnPHI(BI, TD)) 3919 return SimplifyCFG(BB, TTI, TD) | true; 3920 3921 // Scan predecessor blocks for conditional branches. 3922 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3923 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3924 if (PBI != BI && PBI->isConditional()) 3925 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3926 return SimplifyCFG(BB, TTI, TD) | true; 3927 3928 return false; 3929 } 3930 3931 /// Check if passing a value to an instruction will cause undefined behavior. 3932 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3933 Constant *C = dyn_cast<Constant>(V); 3934 if (!C) 3935 return false; 3936 3937 if (I->use_empty()) 3938 return false; 3939 3940 if (C->isNullValue()) { 3941 // Only look at the first use, avoid hurting compile time with long uselists 3942 User *Use = *I->use_begin(); 3943 3944 // Now make sure that there are no instructions in between that can alter 3945 // control flow (eg. calls) 3946 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3947 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3948 return false; 3949 3950 // Look through GEPs. A load from a GEP derived from NULL is still undefined 3951 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 3952 if (GEP->getPointerOperand() == I) 3953 return passingValueIsAlwaysUndefined(V, GEP); 3954 3955 // Look through bitcasts. 3956 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 3957 return passingValueIsAlwaysUndefined(V, BC); 3958 3959 // Load from null is undefined. 3960 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 3961 if (!LI->isVolatile()) 3962 return LI->getPointerAddressSpace() == 0; 3963 3964 // Store to null is undefined. 3965 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 3966 if (!SI->isVolatile()) 3967 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 3968 } 3969 return false; 3970 } 3971 3972 /// If BB has an incoming value that will always trigger undefined behavior 3973 /// (eg. null pointer dereference), remove the branch leading here. 3974 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 3975 for (BasicBlock::iterator i = BB->begin(); 3976 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 3977 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 3978 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 3979 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 3980 IRBuilder<> Builder(T); 3981 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 3982 BB->removePredecessor(PHI->getIncomingBlock(i)); 3983 // Turn uncoditional branches into unreachables and remove the dead 3984 // destination from conditional branches. 3985 if (BI->isUnconditional()) 3986 Builder.CreateUnreachable(); 3987 else 3988 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 3989 BI->getSuccessor(0)); 3990 BI->eraseFromParent(); 3991 return true; 3992 } 3993 // TODO: SwitchInst. 3994 } 3995 3996 return false; 3997 } 3998 3999 bool SimplifyCFGOpt::run(BasicBlock *BB) { 4000 bool Changed = false; 4001 4002 assert(BB && BB->getParent() && "Block not embedded in function!"); 4003 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4004 4005 // Remove basic blocks that have no predecessors (except the entry block)... 4006 // or that just have themself as a predecessor. These are unreachable. 4007 if ((pred_begin(BB) == pred_end(BB) && 4008 BB != &BB->getParent()->getEntryBlock()) || 4009 BB->getSinglePredecessor() == BB) { 4010 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4011 DeleteDeadBlock(BB); 4012 return true; 4013 } 4014 4015 // Check to see if we can constant propagate this terminator instruction 4016 // away... 4017 Changed |= ConstantFoldTerminator(BB, true); 4018 4019 // Check for and eliminate duplicate PHI nodes in this block. 4020 Changed |= EliminateDuplicatePHINodes(BB); 4021 4022 // Check for and remove branches that will always cause undefined behavior. 4023 Changed |= removeUndefIntroducingPredecessor(BB); 4024 4025 // Merge basic blocks into their predecessor if there is only one distinct 4026 // pred, and if there is only one distinct successor of the predecessor, and 4027 // if there are no PHI nodes. 4028 // 4029 if (MergeBlockIntoPredecessor(BB)) 4030 return true; 4031 4032 IRBuilder<> Builder(BB); 4033 4034 // If there is a trivial two-entry PHI node in this basic block, and we can 4035 // eliminate it, do so now. 4036 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4037 if (PN->getNumIncomingValues() == 2) 4038 Changed |= FoldTwoEntryPHINode(PN, TD); 4039 4040 Builder.SetInsertPoint(BB->getTerminator()); 4041 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4042 if (BI->isUnconditional()) { 4043 if (SimplifyUncondBranch(BI, Builder)) return true; 4044 } else { 4045 if (SimplifyCondBranch(BI, Builder)) return true; 4046 } 4047 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4048 if (SimplifyReturn(RI, Builder)) return true; 4049 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4050 if (SimplifyResume(RI, Builder)) return true; 4051 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4052 if (SimplifySwitch(SI, Builder)) return true; 4053 } else if (UnreachableInst *UI = 4054 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4055 if (SimplifyUnreachable(UI)) return true; 4056 } else if (IndirectBrInst *IBI = 4057 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4058 if (SimplifyIndirectBr(IBI)) return true; 4059 } 4060 4061 return Changed; 4062 } 4063 4064 /// SimplifyCFG - This function is used to do simplification of a CFG. For 4065 /// example, it adjusts branches to branches to eliminate the extra hop, it 4066 /// eliminates unreachable basic blocks, and does other "peephole" optimization 4067 /// of the CFG. It returns true if a modification was made. 4068 /// 4069 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4070 const DataLayout *TD) { 4071 return SimplifyCFGOpt(TTI, TD).run(BB); 4072 } 4073