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> 48using namespace llvm; 49 50static cl::opt<unsigned> 51PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 52 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 53 54static cl::opt<bool> 55DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 56 cl::desc("Duplicate return instructions into unconditional branches")); 57 58static cl::opt<bool> 59SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 60 cl::desc("Sink common instructions down to the end block")); 61 62static cl::opt<bool> 63HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true), 64 cl::desc("Hoist conditional stores if an unconditional store preceeds")); 65 66STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 67STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 68STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 69STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 70 71namespace { 72 /// ValueEqualityComparisonCase - Represents a case of a switch. 73 struct ValueEqualityComparisonCase { 74 ConstantInt *Value; 75 BasicBlock *Dest; 76 77 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 78 : Value(Value), Dest(Dest) {} 79 80 bool operator<(ValueEqualityComparisonCase RHS) const { 81 // Comparing pointers is ok as we only rely on the order for uniquing. 82 return Value < RHS.Value; 83 } 84 85 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 86 }; 87 88class SimplifyCFGOpt { 89 const TargetTransformInfo &TTI; 90 const DataLayout *const TD; 91 92 Value *isValueEqualityComparison(TerminatorInst *TI); 93 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 94 std::vector<ValueEqualityComparisonCase> &Cases); 95 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 96 BasicBlock *Pred, 97 IRBuilder<> &Builder); 98 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 99 IRBuilder<> &Builder); 100 101 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 102 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 103 bool SimplifyUnreachable(UnreachableInst *UI); 104 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 105 bool SimplifyIndirectBr(IndirectBrInst *IBI); 106 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 107 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 108 109public: 110 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD) 111 : TTI(TTI), TD(TD) {} 112 bool run(BasicBlock *BB); 113}; 114} 115 116/// SafeToMergeTerminators - Return true if it is safe to merge these two 117/// terminator instructions together. 118/// 119static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 120 if (SI1 == SI2) return false; // Can't merge with self! 121 122 // It is not safe to merge these two switch instructions if they have a common 123 // successor, and if that successor has a PHI node, and if *that* PHI node has 124 // conflicting incoming values from the two switch blocks. 125 BasicBlock *SI1BB = SI1->getParent(); 126 BasicBlock *SI2BB = SI2->getParent(); 127 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 128 129 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 130 if (SI1Succs.count(*I)) 131 for (BasicBlock::iterator BBI = (*I)->begin(); 132 isa<PHINode>(BBI); ++BBI) { 133 PHINode *PN = cast<PHINode>(BBI); 134 if (PN->getIncomingValueForBlock(SI1BB) != 135 PN->getIncomingValueForBlock(SI2BB)) 136 return false; 137 } 138 139 return true; 140} 141 142/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 143/// to merge these two terminator instructions together, where SI1 is an 144/// unconditional branch. PhiNodes will store all PHI nodes in common 145/// successors. 146/// 147static bool isProfitableToFoldUnconditional(BranchInst *SI1, 148 BranchInst *SI2, 149 Instruction *Cond, 150 SmallVectorImpl<PHINode*> &PhiNodes) { 151 if (SI1 == SI2) return false; // Can't merge with self! 152 assert(SI1->isUnconditional() && SI2->isConditional()); 153 154 // We fold the unconditional branch if we can easily update all PHI nodes in 155 // common successors: 156 // 1> We have a constant incoming value for the conditional branch; 157 // 2> We have "Cond" as the incoming value for the unconditional branch; 158 // 3> SI2->getCondition() and Cond have same operands. 159 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 160 if (!Ci2) return false; 161 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 162 Cond->getOperand(1) == Ci2->getOperand(1)) && 163 !(Cond->getOperand(0) == Ci2->getOperand(1) && 164 Cond->getOperand(1) == Ci2->getOperand(0))) 165 return false; 166 167 BasicBlock *SI1BB = SI1->getParent(); 168 BasicBlock *SI2BB = SI2->getParent(); 169 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 170 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 171 if (SI1Succs.count(*I)) 172 for (BasicBlock::iterator BBI = (*I)->begin(); 173 isa<PHINode>(BBI); ++BBI) { 174 PHINode *PN = cast<PHINode>(BBI); 175 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 176 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 177 return false; 178 PhiNodes.push_back(PN); 179 } 180 return true; 181} 182 183/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 184/// now be entries in it from the 'NewPred' block. The values that will be 185/// flowing into the PHI nodes will be the same as those coming in from 186/// ExistPred, an existing predecessor of Succ. 187static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 188 BasicBlock *ExistPred) { 189 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 190 191 PHINode *PN; 192 for (BasicBlock::iterator I = Succ->begin(); 193 (PN = dyn_cast<PHINode>(I)); ++I) 194 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 195} 196 197 198/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 199/// least one PHI node in it), check to see if the merge at this block is due 200/// to an "if condition". If so, return the boolean condition that determines 201/// which entry into BB will be taken. Also, return by references the block 202/// that will be entered from if the condition is true, and the block that will 203/// be entered if the condition is false. 204/// 205/// This does no checking to see if the true/false blocks have large or unsavory 206/// instructions in them. 207static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 208 BasicBlock *&IfFalse) { 209 PHINode *SomePHI = cast<PHINode>(BB->begin()); 210 assert(SomePHI->getNumIncomingValues() == 2 && 211 "Function can only handle blocks with 2 predecessors!"); 212 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 213 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 214 215 // We can only handle branches. Other control flow will be lowered to 216 // branches if possible anyway. 217 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 218 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 219 if (Pred1Br == 0 || Pred2Br == 0) 220 return 0; 221 222 // Eliminate code duplication by ensuring that Pred1Br is conditional if 223 // either are. 224 if (Pred2Br->isConditional()) { 225 // If both branches are conditional, we don't have an "if statement". In 226 // reality, we could transform this case, but since the condition will be 227 // required anyway, we stand no chance of eliminating it, so the xform is 228 // probably not profitable. 229 if (Pred1Br->isConditional()) 230 return 0; 231 232 std::swap(Pred1, Pred2); 233 std::swap(Pred1Br, Pred2Br); 234 } 235 236 if (Pred1Br->isConditional()) { 237 // The only thing we have to watch out for here is to make sure that Pred2 238 // doesn't have incoming edges from other blocks. If it does, the condition 239 // doesn't dominate BB. 240 if (Pred2->getSinglePredecessor() == 0) 241 return 0; 242 243 // If we found a conditional branch predecessor, make sure that it branches 244 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 245 if (Pred1Br->getSuccessor(0) == BB && 246 Pred1Br->getSuccessor(1) == Pred2) { 247 IfTrue = Pred1; 248 IfFalse = Pred2; 249 } else if (Pred1Br->getSuccessor(0) == Pred2 && 250 Pred1Br->getSuccessor(1) == BB) { 251 IfTrue = Pred2; 252 IfFalse = Pred1; 253 } else { 254 // We know that one arm of the conditional goes to BB, so the other must 255 // go somewhere unrelated, and this must not be an "if statement". 256 return 0; 257 } 258 259 return Pred1Br->getCondition(); 260 } 261 262 // Ok, if we got here, both predecessors end with an unconditional branch to 263 // BB. Don't panic! If both blocks only have a single (identical) 264 // predecessor, and THAT is a conditional branch, then we're all ok! 265 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 266 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 267 return 0; 268 269 // Otherwise, if this is a conditional branch, then we can use it! 270 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 271 if (BI == 0) return 0; 272 273 assert(BI->isConditional() && "Two successors but not conditional?"); 274 if (BI->getSuccessor(0) == Pred1) { 275 IfTrue = Pred1; 276 IfFalse = Pred2; 277 } else { 278 IfTrue = Pred2; 279 IfFalse = Pred1; 280 } 281 return BI->getCondition(); 282} 283 284/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 285/// given instruction, which is assumed to be safe to speculate. 1 means 286/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 287static unsigned ComputeSpeculationCost(const User *I) { 288 assert(isSafeToSpeculativelyExecute(I) && 289 "Instruction is not safe to speculatively execute!"); 290 switch (Operator::getOpcode(I)) { 291 default: 292 // In doubt, be conservative. 293 return UINT_MAX; 294 case Instruction::GetElementPtr: 295 // GEPs are cheap if all indices are constant. 296 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 297 return UINT_MAX; 298 return 1; 299 case Instruction::Load: 300 case Instruction::Add: 301 case Instruction::Sub: 302 case Instruction::And: 303 case Instruction::Or: 304 case Instruction::Xor: 305 case Instruction::Shl: 306 case Instruction::LShr: 307 case Instruction::AShr: 308 case Instruction::ICmp: 309 case Instruction::Trunc: 310 case Instruction::ZExt: 311 case Instruction::SExt: 312 return 1; // These are all cheap. 313 314 case Instruction::Call: 315 case Instruction::Select: 316 return 2; 317 } 318} 319 320/// DominatesMergePoint - If we have a merge point of an "if condition" as 321/// accepted above, return true if the specified value dominates the block. We 322/// don't handle the true generality of domination here, just a special case 323/// which works well enough for us. 324/// 325/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 326/// see if V (which must be an instruction) and its recursive operands 327/// that do not dominate BB have a combined cost lower than CostRemaining and 328/// are non-trapping. If both are true, the instruction is inserted into the 329/// set and true is returned. 330/// 331/// The cost for most non-trapping instructions is defined as 1 except for 332/// Select whose cost is 2. 333/// 334/// After this function returns, CostRemaining is decreased by the cost of 335/// V plus its non-dominating operands. If that cost is greater than 336/// CostRemaining, false is returned and CostRemaining is undefined. 337static bool DominatesMergePoint(Value *V, BasicBlock *BB, 338 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 339 unsigned &CostRemaining) { 340 Instruction *I = dyn_cast<Instruction>(V); 341 if (!I) { 342 // Non-instructions all dominate instructions, but not all constantexprs 343 // can be executed unconditionally. 344 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 345 if (C->canTrap()) 346 return false; 347 return true; 348 } 349 BasicBlock *PBB = I->getParent(); 350 351 // We don't want to allow weird loops that might have the "if condition" in 352 // the bottom of this block. 353 if (PBB == BB) return false; 354 355 // If this instruction is defined in a block that contains an unconditional 356 // branch to BB, then it must be in the 'conditional' part of the "if 357 // statement". If not, it definitely dominates the region. 358 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 359 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 360 return true; 361 362 // If we aren't allowing aggressive promotion anymore, then don't consider 363 // instructions in the 'if region'. 364 if (AggressiveInsts == 0) return false; 365 366 // If we have seen this instruction before, don't count it again. 367 if (AggressiveInsts->count(I)) return true; 368 369 // Okay, it looks like the instruction IS in the "condition". Check to 370 // see if it's a cheap instruction to unconditionally compute, and if it 371 // only uses stuff defined outside of the condition. If so, hoist it out. 372 if (!isSafeToSpeculativelyExecute(I)) 373 return false; 374 375 unsigned Cost = ComputeSpeculationCost(I); 376 377 if (Cost > CostRemaining) 378 return false; 379 380 CostRemaining -= Cost; 381 382 // Okay, we can only really hoist these out if their operands do 383 // not take us over the cost threshold. 384 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 385 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 386 return false; 387 // Okay, it's safe to do this! Remember this instruction. 388 AggressiveInsts->insert(I); 389 return true; 390} 391 392/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 393/// and PointerNullValue. Return NULL if value is not a constant int. 394static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 395 // Normal constant int. 396 ConstantInt *CI = dyn_cast<ConstantInt>(V); 397 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 398 return CI; 399 400 // This is some kind of pointer constant. Turn it into a pointer-sized 401 // ConstantInt if possible. 402 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); 403 404 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 405 if (isa<ConstantPointerNull>(V)) 406 return ConstantInt::get(PtrTy, 0); 407 408 // IntToPtr const int. 409 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 410 if (CE->getOpcode() == Instruction::IntToPtr) 411 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 412 // The constant is very likely to have the right type already. 413 if (CI->getType() == PtrTy) 414 return CI; 415 else 416 return cast<ConstantInt> 417 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 418 } 419 return 0; 420} 421 422/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 423/// collection of icmp eq/ne instructions that compare a value against a 424/// constant, return the value being compared, and stick the constant into the 425/// Values vector. 426static Value * 427GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 428 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 429 Instruction *I = dyn_cast<Instruction>(V); 430 if (I == 0) return 0; 431 432 // If this is an icmp against a constant, handle this as one of the cases. 433 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 434 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 435 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 436 UsedICmps++; 437 Vals.push_back(C); 438 return I->getOperand(0); 439 } 440 441 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 442 // the set. 443 ConstantRange Span = 444 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 445 446 // If this is an and/!= check then we want to optimize "x ugt 2" into 447 // x != 0 && x != 1. 448 if (!isEQ) 449 Span = Span.inverse(); 450 451 // If there are a ton of values, we don't want to make a ginormous switch. 452 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 453 return 0; 454 455 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 456 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 457 UsedICmps++; 458 return I->getOperand(0); 459 } 460 return 0; 461 } 462 463 // Otherwise, we can only handle an | or &, depending on isEQ. 464 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 465 return 0; 466 467 unsigned NumValsBeforeLHS = Vals.size(); 468 unsigned UsedICmpsBeforeLHS = UsedICmps; 469 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 470 isEQ, UsedICmps)) { 471 unsigned NumVals = Vals.size(); 472 unsigned UsedICmpsBeforeRHS = UsedICmps; 473 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 474 isEQ, UsedICmps)) { 475 if (LHS == RHS) 476 return LHS; 477 Vals.resize(NumVals); 478 UsedICmps = UsedICmpsBeforeRHS; 479 } 480 481 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 482 // set it and return success. 483 if (Extra == 0 || Extra == I->getOperand(1)) { 484 Extra = I->getOperand(1); 485 return LHS; 486 } 487 488 Vals.resize(NumValsBeforeLHS); 489 UsedICmps = UsedICmpsBeforeLHS; 490 return 0; 491 } 492 493 // If the LHS can't be folded in, but Extra is available and RHS can, try to 494 // use LHS as Extra. 495 if (Extra == 0 || Extra == I->getOperand(0)) { 496 Value *OldExtra = Extra; 497 Extra = I->getOperand(0); 498 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 499 isEQ, UsedICmps)) 500 return RHS; 501 assert(Vals.size() == NumValsBeforeLHS); 502 Extra = OldExtra; 503 } 504 505 return 0; 506} 507 508static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 509 Instruction *Cond = 0; 510 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 511 Cond = dyn_cast<Instruction>(SI->getCondition()); 512 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 513 if (BI->isConditional()) 514 Cond = dyn_cast<Instruction>(BI->getCondition()); 515 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 516 Cond = dyn_cast<Instruction>(IBI->getAddress()); 517 } 518 519 TI->eraseFromParent(); 520 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 521} 522 523/// isValueEqualityComparison - Return true if the specified terminator checks 524/// to see if a value is equal to constant integer value. 525Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 526 Value *CV = 0; 527 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 528 // Do not permit merging of large switch instructions into their 529 // predecessors unless there is only one predecessor. 530 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 531 pred_end(SI->getParent())) <= 128) 532 CV = SI->getCondition(); 533 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 534 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 535 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 536 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 537 ICI->getPredicate() == ICmpInst::ICMP_NE) && 538 GetConstantInt(ICI->getOperand(1), TD)) 539 CV = ICI->getOperand(0); 540 541 // Unwrap any lossless ptrtoint cast. 542 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 543 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 544 CV = PTII->getOperand(0); 545 return CV; 546} 547 548/// GetValueEqualityComparisonCases - Given a value comparison instruction, 549/// decode all of the 'cases' that it represents and return the 'default' block. 550BasicBlock *SimplifyCFGOpt:: 551GetValueEqualityComparisonCases(TerminatorInst *TI, 552 std::vector<ValueEqualityComparisonCase> 553 &Cases) { 554 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 555 Cases.reserve(SI->getNumCases()); 556 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 557 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 558 i.getCaseSuccessor())); 559 return SI->getDefaultDest(); 560 } 561 562 BranchInst *BI = cast<BranchInst>(TI); 563 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 564 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 565 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 566 TD), 567 Succ)); 568 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 569} 570 571 572/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 573/// in the list that match the specified block. 574static void EliminateBlockCases(BasicBlock *BB, 575 std::vector<ValueEqualityComparisonCase> &Cases) { 576 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 577} 578 579/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 580/// well. 581static bool 582ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 583 std::vector<ValueEqualityComparisonCase > &C2) { 584 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 585 586 // Make V1 be smaller than V2. 587 if (V1->size() > V2->size()) 588 std::swap(V1, V2); 589 590 if (V1->size() == 0) return false; 591 if (V1->size() == 1) { 592 // Just scan V2. 593 ConstantInt *TheVal = (*V1)[0].Value; 594 for (unsigned i = 0, e = V2->size(); i != e; ++i) 595 if (TheVal == (*V2)[i].Value) 596 return true; 597 } 598 599 // Otherwise, just sort both lists and compare element by element. 600 array_pod_sort(V1->begin(), V1->end()); 601 array_pod_sort(V2->begin(), V2->end()); 602 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 603 while (i1 != e1 && i2 != e2) { 604 if ((*V1)[i1].Value == (*V2)[i2].Value) 605 return true; 606 if ((*V1)[i1].Value < (*V2)[i2].Value) 607 ++i1; 608 else 609 ++i2; 610 } 611 return false; 612} 613 614/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 615/// terminator instruction and its block is known to only have a single 616/// predecessor block, check to see if that predecessor is also a value 617/// comparison with the same value, and if that comparison determines the 618/// outcome of this comparison. If so, simplify TI. This does a very limited 619/// form of jump threading. 620bool SimplifyCFGOpt:: 621SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 622 BasicBlock *Pred, 623 IRBuilder<> &Builder) { 624 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 625 if (!PredVal) return false; // Not a value comparison in predecessor. 626 627 Value *ThisVal = isValueEqualityComparison(TI); 628 assert(ThisVal && "This isn't a value comparison!!"); 629 if (ThisVal != PredVal) return false; // Different predicates. 630 631 // TODO: Preserve branch weight metadata, similarly to how 632 // FoldValueComparisonIntoPredecessors preserves it. 633 634 // Find out information about when control will move from Pred to TI's block. 635 std::vector<ValueEqualityComparisonCase> PredCases; 636 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 637 PredCases); 638 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 639 640 // Find information about how control leaves this block. 641 std::vector<ValueEqualityComparisonCase> ThisCases; 642 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 643 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 644 645 // If TI's block is the default block from Pred's comparison, potentially 646 // simplify TI based on this knowledge. 647 if (PredDef == TI->getParent()) { 648 // If we are here, we know that the value is none of those cases listed in 649 // PredCases. If there are any cases in ThisCases that are in PredCases, we 650 // can simplify TI. 651 if (!ValuesOverlap(PredCases, ThisCases)) 652 return false; 653 654 if (isa<BranchInst>(TI)) { 655 // Okay, one of the successors of this condbr is dead. Convert it to a 656 // uncond br. 657 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 658 // Insert the new branch. 659 Instruction *NI = Builder.CreateBr(ThisDef); 660 (void) NI; 661 662 // Remove PHI node entries for the dead edge. 663 ThisCases[0].Dest->removePredecessor(TI->getParent()); 664 665 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 666 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 667 668 EraseTerminatorInstAndDCECond(TI); 669 return true; 670 } 671 672 SwitchInst *SI = cast<SwitchInst>(TI); 673 // Okay, TI has cases that are statically dead, prune them away. 674 SmallPtrSet<Constant*, 16> DeadCases; 675 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 676 DeadCases.insert(PredCases[i].Value); 677 678 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 679 << "Through successor TI: " << *TI); 680 681 // Collect branch weights into a vector. 682 SmallVector<uint32_t, 8> Weights; 683 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 684 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 685 if (HasWeight) 686 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 687 ++MD_i) { 688 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 689 assert(CI); 690 Weights.push_back(CI->getValue().getZExtValue()); 691 } 692 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 693 --i; 694 if (DeadCases.count(i.getCaseValue())) { 695 if (HasWeight) { 696 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 697 Weights.pop_back(); 698 } 699 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 700 SI->removeCase(i); 701 } 702 } 703 if (HasWeight && Weights.size() >= 2) 704 SI->setMetadata(LLVMContext::MD_prof, 705 MDBuilder(SI->getParent()->getContext()). 706 createBranchWeights(Weights)); 707 708 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 709 return true; 710 } 711 712 // Otherwise, TI's block must correspond to some matched value. Find out 713 // which value (or set of values) this is. 714 ConstantInt *TIV = 0; 715 BasicBlock *TIBB = TI->getParent(); 716 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 717 if (PredCases[i].Dest == TIBB) { 718 if (TIV != 0) 719 return false; // Cannot handle multiple values coming to this block. 720 TIV = PredCases[i].Value; 721 } 722 assert(TIV && "No edge from pred to succ?"); 723 724 // Okay, we found the one constant that our value can be if we get into TI's 725 // BB. Find out which successor will unconditionally be branched to. 726 BasicBlock *TheRealDest = 0; 727 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 728 if (ThisCases[i].Value == TIV) { 729 TheRealDest = ThisCases[i].Dest; 730 break; 731 } 732 733 // If not handled by any explicit cases, it is handled by the default case. 734 if (TheRealDest == 0) TheRealDest = ThisDef; 735 736 // Remove PHI node entries for dead edges. 737 BasicBlock *CheckEdge = TheRealDest; 738 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 739 if (*SI != CheckEdge) 740 (*SI)->removePredecessor(TIBB); 741 else 742 CheckEdge = 0; 743 744 // Insert the new branch. 745 Instruction *NI = Builder.CreateBr(TheRealDest); 746 (void) NI; 747 748 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 749 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 750 751 EraseTerminatorInstAndDCECond(TI); 752 return true; 753} 754 755namespace { 756 /// ConstantIntOrdering - This class implements a stable ordering of constant 757 /// integers that does not depend on their address. This is important for 758 /// applications that sort ConstantInt's to ensure uniqueness. 759 struct ConstantIntOrdering { 760 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 761 return LHS->getValue().ult(RHS->getValue()); 762 } 763 }; 764} 765 766static int ConstantIntSortPredicate(const void *P1, const void *P2) { 767 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 768 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 769 if (LHS->getValue().ult(RHS->getValue())) 770 return 1; 771 if (LHS->getValue() == RHS->getValue()) 772 return 0; 773 return -1; 774} 775 776static inline bool HasBranchWeights(const Instruction* I) { 777 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 778 if (ProfMD && ProfMD->getOperand(0)) 779 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 780 return MDS->getString().equals("branch_weights"); 781 782 return false; 783} 784 785/// Get Weights of a given TerminatorInst, the default weight is at the front 786/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 787/// metadata. 788static void GetBranchWeights(TerminatorInst *TI, 789 SmallVectorImpl<uint64_t> &Weights) { 790 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 791 assert(MD); 792 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 793 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 794 assert(CI); 795 Weights.push_back(CI->getValue().getZExtValue()); 796 } 797 798 // If TI is a conditional eq, the default case is the false case, 799 // and the corresponding branch-weight data is at index 2. We swap the 800 // default weight to be the first entry. 801 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 802 assert(Weights.size() == 2); 803 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 804 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 805 std::swap(Weights.front(), Weights.back()); 806 } 807} 808 809/// Sees if any of the weights are too big for a uint32_t, and halves all the 810/// weights if any are. 811static void FitWeights(MutableArrayRef<uint64_t> Weights) { 812 bool Halve = false; 813 for (unsigned i = 0; i < Weights.size(); ++i) 814 if (Weights[i] > UINT_MAX) { 815 Halve = true; 816 break; 817 } 818 819 if (! Halve) 820 return; 821 822 for (unsigned i = 0; i < Weights.size(); ++i) 823 Weights[i] /= 2; 824} 825 826/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 827/// equality comparison instruction (either a switch or a branch on "X == c"). 828/// See if any of the predecessors of the terminator block are value comparisons 829/// on the same value. If so, and if safe to do so, fold them together. 830bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 831 IRBuilder<> &Builder) { 832 BasicBlock *BB = TI->getParent(); 833 Value *CV = isValueEqualityComparison(TI); // CondVal 834 assert(CV && "Not a comparison?"); 835 bool Changed = false; 836 837 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 838 while (!Preds.empty()) { 839 BasicBlock *Pred = Preds.pop_back_val(); 840 841 // See if the predecessor is a comparison with the same value. 842 TerminatorInst *PTI = Pred->getTerminator(); 843 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 844 845 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 846 // Figure out which 'cases' to copy from SI to PSI. 847 std::vector<ValueEqualityComparisonCase> BBCases; 848 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 849 850 std::vector<ValueEqualityComparisonCase> PredCases; 851 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 852 853 // Based on whether the default edge from PTI goes to BB or not, fill in 854 // PredCases and PredDefault with the new switch cases we would like to 855 // build. 856 SmallVector<BasicBlock*, 8> NewSuccessors; 857 858 // Update the branch weight metadata along the way 859 SmallVector<uint64_t, 8> Weights; 860 bool PredHasWeights = HasBranchWeights(PTI); 861 bool SuccHasWeights = HasBranchWeights(TI); 862 863 if (PredHasWeights) { 864 GetBranchWeights(PTI, Weights); 865 // branch-weight metadata is inconsistent here. 866 if (Weights.size() != 1 + PredCases.size()) 867 PredHasWeights = SuccHasWeights = false; 868 } else if (SuccHasWeights) 869 // If there are no predecessor weights but there are successor weights, 870 // populate Weights with 1, which will later be scaled to the sum of 871 // successor's weights 872 Weights.assign(1 + PredCases.size(), 1); 873 874 SmallVector<uint64_t, 8> SuccWeights; 875 if (SuccHasWeights) { 876 GetBranchWeights(TI, SuccWeights); 877 // branch-weight metadata is inconsistent here. 878 if (SuccWeights.size() != 1 + BBCases.size()) 879 PredHasWeights = SuccHasWeights = false; 880 } else if (PredHasWeights) 881 SuccWeights.assign(1 + BBCases.size(), 1); 882 883 if (PredDefault == BB) { 884 // If this is the default destination from PTI, only the edges in TI 885 // that don't occur in PTI, or that branch to BB will be activated. 886 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 887 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 888 if (PredCases[i].Dest != BB) 889 PTIHandled.insert(PredCases[i].Value); 890 else { 891 // The default destination is BB, we don't need explicit targets. 892 std::swap(PredCases[i], PredCases.back()); 893 894 if (PredHasWeights || SuccHasWeights) { 895 // Increase weight for the default case. 896 Weights[0] += Weights[i+1]; 897 std::swap(Weights[i+1], Weights.back()); 898 Weights.pop_back(); 899 } 900 901 PredCases.pop_back(); 902 --i; --e; 903 } 904 905 // Reconstruct the new switch statement we will be building. 906 if (PredDefault != BBDefault) { 907 PredDefault->removePredecessor(Pred); 908 PredDefault = BBDefault; 909 NewSuccessors.push_back(BBDefault); 910 } 911 912 unsigned CasesFromPred = Weights.size(); 913 uint64_t ValidTotalSuccWeight = 0; 914 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 915 if (!PTIHandled.count(BBCases[i].Value) && 916 BBCases[i].Dest != BBDefault) { 917 PredCases.push_back(BBCases[i]); 918 NewSuccessors.push_back(BBCases[i].Dest); 919 if (SuccHasWeights || PredHasWeights) { 920 // The default weight is at index 0, so weight for the ith case 921 // should be at index i+1. Scale the cases from successor by 922 // PredDefaultWeight (Weights[0]). 923 Weights.push_back(Weights[0] * SuccWeights[i+1]); 924 ValidTotalSuccWeight += SuccWeights[i+1]; 925 } 926 } 927 928 if (SuccHasWeights || PredHasWeights) { 929 ValidTotalSuccWeight += SuccWeights[0]; 930 // Scale the cases from predecessor by ValidTotalSuccWeight. 931 for (unsigned i = 1; i < CasesFromPred; ++i) 932 Weights[i] *= ValidTotalSuccWeight; 933 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 934 Weights[0] *= SuccWeights[0]; 935 } 936 } else { 937 // If this is not the default destination from PSI, only the edges 938 // in SI that occur in PSI with a destination of BB will be 939 // activated. 940 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 941 std::map<ConstantInt*, uint64_t> WeightsForHandled; 942 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 943 if (PredCases[i].Dest == BB) { 944 PTIHandled.insert(PredCases[i].Value); 945 946 if (PredHasWeights || SuccHasWeights) { 947 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 948 std::swap(Weights[i+1], Weights.back()); 949 Weights.pop_back(); 950 } 951 952 std::swap(PredCases[i], PredCases.back()); 953 PredCases.pop_back(); 954 --i; --e; 955 } 956 957 // Okay, now we know which constants were sent to BB from the 958 // predecessor. Figure out where they will all go now. 959 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 960 if (PTIHandled.count(BBCases[i].Value)) { 961 // If this is one we are capable of getting... 962 if (PredHasWeights || SuccHasWeights) 963 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 964 PredCases.push_back(BBCases[i]); 965 NewSuccessors.push_back(BBCases[i].Dest); 966 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 967 } 968 969 // If there are any constants vectored to BB that TI doesn't handle, 970 // they must go to the default destination of TI. 971 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 972 PTIHandled.begin(), 973 E = PTIHandled.end(); I != E; ++I) { 974 if (PredHasWeights || SuccHasWeights) 975 Weights.push_back(WeightsForHandled[*I]); 976 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 977 NewSuccessors.push_back(BBDefault); 978 } 979 } 980 981 // Okay, at this point, we know which new successor Pred will get. Make 982 // sure we update the number of entries in the PHI nodes for these 983 // successors. 984 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 985 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 986 987 Builder.SetInsertPoint(PTI); 988 // Convert pointer to int before we switch. 989 if (CV->getType()->isPointerTy()) { 990 assert(TD && "Cannot switch on pointer without DataLayout"); 991 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 992 "magicptr"); 993 } 994 995 // Now that the successors are updated, create the new Switch instruction. 996 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 997 PredCases.size()); 998 NewSI->setDebugLoc(PTI->getDebugLoc()); 999 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 1000 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 1001 1002 if (PredHasWeights || SuccHasWeights) { 1003 // Halve the weights if any of them cannot fit in an uint32_t 1004 FitWeights(Weights); 1005 1006 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1007 1008 NewSI->setMetadata(LLVMContext::MD_prof, 1009 MDBuilder(BB->getContext()). 1010 createBranchWeights(MDWeights)); 1011 } 1012 1013 EraseTerminatorInstAndDCECond(PTI); 1014 1015 // Okay, last check. If BB is still a successor of PSI, then we must 1016 // have an infinite loop case. If so, add an infinitely looping block 1017 // to handle the case to preserve the behavior of the code. 1018 BasicBlock *InfLoopBlock = 0; 1019 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1020 if (NewSI->getSuccessor(i) == BB) { 1021 if (InfLoopBlock == 0) { 1022 // Insert it at the end of the function, because it's either code, 1023 // or it won't matter if it's hot. :) 1024 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1025 "infloop", BB->getParent()); 1026 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1027 } 1028 NewSI->setSuccessor(i, InfLoopBlock); 1029 } 1030 1031 Changed = true; 1032 } 1033 } 1034 return Changed; 1035} 1036 1037// isSafeToHoistInvoke - If we would need to insert a select that uses the 1038// value of this invoke (comments in HoistThenElseCodeToIf explain why we 1039// would need to do this), we can't hoist the invoke, as there is nowhere 1040// to put the select in this case. 1041static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1042 Instruction *I1, Instruction *I2) { 1043 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1044 PHINode *PN; 1045 for (BasicBlock::iterator BBI = SI->begin(); 1046 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1047 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1048 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1049 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1050 return false; 1051 } 1052 } 1053 } 1054 return true; 1055} 1056 1057/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1058/// BB2, hoist any common code in the two blocks up into the branch block. The 1059/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1060static bool HoistThenElseCodeToIf(BranchInst *BI) { 1061 // This does very trivial matching, with limited scanning, to find identical 1062 // instructions in the two blocks. In particular, we don't want to get into 1063 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1064 // such, we currently just scan for obviously identical instructions in an 1065 // identical order. 1066 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1067 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1068 1069 BasicBlock::iterator BB1_Itr = BB1->begin(); 1070 BasicBlock::iterator BB2_Itr = BB2->begin(); 1071 1072 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1073 // Skip debug info if it is not identical. 1074 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1075 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1076 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1077 while (isa<DbgInfoIntrinsic>(I1)) 1078 I1 = BB1_Itr++; 1079 while (isa<DbgInfoIntrinsic>(I2)) 1080 I2 = BB2_Itr++; 1081 } 1082 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1083 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1084 return false; 1085 1086 // If we get here, we can hoist at least one instruction. 1087 BasicBlock *BIParent = BI->getParent(); 1088 1089 do { 1090 // If we are hoisting the terminator instruction, don't move one (making a 1091 // broken BB), instead clone it, and remove BI. 1092 if (isa<TerminatorInst>(I1)) 1093 goto HoistTerminator; 1094 1095 // For a normal instruction, we just move one to right before the branch, 1096 // then replace all uses of the other with the first. Finally, we remove 1097 // the now redundant second instruction. 1098 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1099 if (!I2->use_empty()) 1100 I2->replaceAllUsesWith(I1); 1101 I1->intersectOptionalDataWith(I2); 1102 I2->eraseFromParent(); 1103 1104 I1 = BB1_Itr++; 1105 I2 = BB2_Itr++; 1106 // Skip debug info if it is not identical. 1107 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1108 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1109 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1110 while (isa<DbgInfoIntrinsic>(I1)) 1111 I1 = BB1_Itr++; 1112 while (isa<DbgInfoIntrinsic>(I2)) 1113 I2 = BB2_Itr++; 1114 } 1115 } while (I1->isIdenticalToWhenDefined(I2)); 1116 1117 return true; 1118 1119HoistTerminator: 1120 // It may not be possible to hoist an invoke. 1121 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1122 return true; 1123 1124 // Okay, it is safe to hoist the terminator. 1125 Instruction *NT = I1->clone(); 1126 BIParent->getInstList().insert(BI, NT); 1127 if (!NT->getType()->isVoidTy()) { 1128 I1->replaceAllUsesWith(NT); 1129 I2->replaceAllUsesWith(NT); 1130 NT->takeName(I1); 1131 } 1132 1133 IRBuilder<true, NoFolder> Builder(NT); 1134 // Hoisting one of the terminators from our successor is a great thing. 1135 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1136 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1137 // nodes, so we insert select instruction to compute the final result. 1138 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1139 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1140 PHINode *PN; 1141 for (BasicBlock::iterator BBI = SI->begin(); 1142 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1143 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1144 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1145 if (BB1V == BB2V) continue; 1146 1147 // These values do not agree. Insert a select instruction before NT 1148 // that determines the right value. 1149 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1150 if (SI == 0) 1151 SI = cast<SelectInst> 1152 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1153 BB1V->getName()+"."+BB2V->getName())); 1154 1155 // Make the PHI node use the select for all incoming values for BB1/BB2 1156 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1157 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1158 PN->setIncomingValue(i, SI); 1159 } 1160 } 1161 1162 // Update any PHI nodes in our new successors. 1163 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1164 AddPredecessorToBlock(*SI, BIParent, BB1); 1165 1166 EraseTerminatorInstAndDCECond(BI); 1167 return true; 1168} 1169 1170/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1171/// check whether BBEnd has only two predecessors and the other predecessor 1172/// ends with an unconditional branch. If it is true, sink any common code 1173/// in the two predecessors to BBEnd. 1174static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1175 assert(BI1->isUnconditional()); 1176 BasicBlock *BB1 = BI1->getParent(); 1177 BasicBlock *BBEnd = BI1->getSuccessor(0); 1178 1179 // Check that BBEnd has two predecessors and the other predecessor ends with 1180 // an unconditional branch. 1181 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1182 BasicBlock *Pred0 = *PI++; 1183 if (PI == PE) // Only one predecessor. 1184 return false; 1185 BasicBlock *Pred1 = *PI++; 1186 if (PI != PE) // More than two predecessors. 1187 return false; 1188 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1189 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1190 if (!BI2 || !BI2->isUnconditional()) 1191 return false; 1192 1193 // Gather the PHI nodes in BBEnd. 1194 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1195 Instruction *FirstNonPhiInBBEnd = 0; 1196 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1197 I != E; ++I) { 1198 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1199 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1200 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1201 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1202 } else { 1203 FirstNonPhiInBBEnd = &*I; 1204 break; 1205 } 1206 } 1207 if (!FirstNonPhiInBBEnd) 1208 return false; 1209 1210 1211 // This does very trivial matching, with limited scanning, to find identical 1212 // instructions in the two blocks. We scan backward for obviously identical 1213 // instructions in an identical order. 1214 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1215 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1216 RE2 = BB2->getInstList().rend(); 1217 // Skip debug info. 1218 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1219 if (RI1 == RE1) 1220 return false; 1221 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1222 if (RI2 == RE2) 1223 return false; 1224 // Skip the unconditional branches. 1225 ++RI1; 1226 ++RI2; 1227 1228 bool Changed = false; 1229 while (RI1 != RE1 && RI2 != RE2) { 1230 // Skip debug info. 1231 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1232 if (RI1 == RE1) 1233 return Changed; 1234 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1235 if (RI2 == RE2) 1236 return Changed; 1237 1238 Instruction *I1 = &*RI1, *I2 = &*RI2; 1239 // I1 and I2 should have a single use in the same PHI node, and they 1240 // perform the same operation. 1241 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1242 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1243 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1244 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1245 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1246 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1247 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1248 !I1->hasOneUse() || !I2->hasOneUse() || 1249 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1250 MapValueFromBB1ToBB2[I1].first != I2) 1251 return Changed; 1252 1253 // Check whether we should swap the operands of ICmpInst. 1254 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1255 bool SwapOpnds = false; 1256 if (ICmp1 && ICmp2 && 1257 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1258 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1259 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1260 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1261 ICmp2->swapOperands(); 1262 SwapOpnds = true; 1263 } 1264 if (!I1->isSameOperationAs(I2)) { 1265 if (SwapOpnds) 1266 ICmp2->swapOperands(); 1267 return Changed; 1268 } 1269 1270 // The operands should be either the same or they need to be generated 1271 // with a PHI node after sinking. We only handle the case where there is 1272 // a single pair of different operands. 1273 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1274 unsigned Op1Idx = 0; 1275 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1276 if (I1->getOperand(I) == I2->getOperand(I)) 1277 continue; 1278 // Early exit if we have more-than one pair of different operands or 1279 // the different operand is already in MapValueFromBB1ToBB2. 1280 // Early exit if we need a PHI node to replace a constant. 1281 if (DifferentOp1 || 1282 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1283 MapValueFromBB1ToBB2.end() || 1284 isa<Constant>(I1->getOperand(I)) || 1285 isa<Constant>(I2->getOperand(I))) { 1286 // If we can't sink the instructions, undo the swapping. 1287 if (SwapOpnds) 1288 ICmp2->swapOperands(); 1289 return Changed; 1290 } 1291 DifferentOp1 = I1->getOperand(I); 1292 Op1Idx = I; 1293 DifferentOp2 = I2->getOperand(I); 1294 } 1295 1296 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1297 // remove (I1, I2) from MapValueFromBB1ToBB2. 1298 if (DifferentOp1) { 1299 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1300 DifferentOp1->getName() + ".sink", 1301 BBEnd->begin()); 1302 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1303 // I1 should use NewPN instead of DifferentOp1. 1304 I1->setOperand(Op1Idx, NewPN); 1305 NewPN->addIncoming(DifferentOp1, BB1); 1306 NewPN->addIncoming(DifferentOp2, BB2); 1307 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1308 } 1309 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1310 MapValueFromBB1ToBB2.erase(I1); 1311 1312 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1313 DEBUG(dbgs() << " " << *I2 << "\n";); 1314 // We need to update RE1 and RE2 if we are going to sink the first 1315 // instruction in the basic block down. 1316 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1317 // Sink the instruction. 1318 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1319 if (!OldPN->use_empty()) 1320 OldPN->replaceAllUsesWith(I1); 1321 OldPN->eraseFromParent(); 1322 1323 if (!I2->use_empty()) 1324 I2->replaceAllUsesWith(I1); 1325 I1->intersectOptionalDataWith(I2); 1326 I2->eraseFromParent(); 1327 1328 if (UpdateRE1) 1329 RE1 = BB1->getInstList().rend(); 1330 if (UpdateRE2) 1331 RE2 = BB2->getInstList().rend(); 1332 FirstNonPhiInBBEnd = I1; 1333 NumSinkCommons++; 1334 Changed = true; 1335 } 1336 return Changed; 1337} 1338 1339/// \brief Determine if we can hoist sink a sole store instruction out of a 1340/// conditional block. 1341/// 1342/// We are looking for code like the following: 1343/// BrBB: 1344/// store i32 %add, i32* %arrayidx2 1345/// ... // No other stores or function calls (we could be calling a memory 1346/// ... // function). 1347/// %cmp = icmp ult %x, %y 1348/// br i1 %cmp, label %EndBB, label %ThenBB 1349/// ThenBB: 1350/// store i32 %add5, i32* %arrayidx2 1351/// br label EndBB 1352/// EndBB: 1353/// ... 1354/// We are going to transform this into: 1355/// BrBB: 1356/// store i32 %add, i32* %arrayidx2 1357/// ... // 1358/// %cmp = icmp ult %x, %y 1359/// %add.add5 = select i1 %cmp, i32 %add, %add5 1360/// store i32 %add.add5, i32* %arrayidx2 1361/// ... 1362/// 1363/// \return The pointer to the value of the previous store if the store can be 1364/// hoisted into the predecessor block. 0 otherwise. 1365Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, 1366 BasicBlock *StoreBB, BasicBlock *EndBB) { 1367 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); 1368 if (!StoreToHoist) 1369 return 0; 1370 1371 // Volatile or atomic. 1372 if (!StoreToHoist->isSimple()) 1373 return 0; 1374 1375 Value *StorePtr = StoreToHoist->getPointerOperand(); 1376 1377 // Look for a store to the same pointer in BrBB. 1378 unsigned MaxNumInstToLookAt = 10; 1379 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), 1380 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { 1381 Instruction *CurI = &*RI; 1382 1383 // Could be calling an instruction that effects memory like free(). 1384 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) 1385 return 0; 1386 1387 StoreInst *SI = dyn_cast<StoreInst>(CurI); 1388 // Found the previous store make sure it stores to the same location. 1389 if (SI && SI->getPointerOperand() == StorePtr) 1390 // Found the previous store, return its value operand. 1391 return SI->getValueOperand(); 1392 else if (SI) 1393 return 0; // Unknown store. 1394 } 1395 1396 return 0; 1397} 1398 1399/// \brief Speculate a conditional basic block flattening the CFG. 1400/// 1401/// Note that this is a very risky transform currently. Speculating 1402/// instructions like this is most often not desirable. Instead, there is an MI 1403/// pass which can do it with full awareness of the resource constraints. 1404/// However, some cases are "obvious" and we should do directly. An example of 1405/// this is speculating a single, reasonably cheap instruction. 1406/// 1407/// There is only one distinct advantage to flattening the CFG at the IR level: 1408/// it makes very common but simplistic optimizations such as are common in 1409/// instcombine and the DAG combiner more powerful by removing CFG edges and 1410/// modeling their effects with easier to reason about SSA value graphs. 1411/// 1412/// 1413/// An illustration of this transform is turning this IR: 1414/// \code 1415/// BB: 1416/// %cmp = icmp ult %x, %y 1417/// br i1 %cmp, label %EndBB, label %ThenBB 1418/// ThenBB: 1419/// %sub = sub %x, %y 1420/// br label BB2 1421/// EndBB: 1422/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1423/// ... 1424/// \endcode 1425/// 1426/// Into this IR: 1427/// \code 1428/// BB: 1429/// %cmp = icmp ult %x, %y 1430/// %sub = sub %x, %y 1431/// %cond = select i1 %cmp, 0, %sub 1432/// ... 1433/// \endcode 1434/// 1435/// \returns true if the conditional block is removed. 1436static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1437 // Be conservative for now. FP select instruction can often be expensive. 1438 Value *BrCond = BI->getCondition(); 1439 if (isa<FCmpInst>(BrCond)) 1440 return false; 1441 1442 BasicBlock *BB = BI->getParent(); 1443 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1444 1445 // If ThenBB is actually on the false edge of the conditional branch, remember 1446 // to swap the select operands later. 1447 bool Invert = false; 1448 if (ThenBB != BI->getSuccessor(0)) { 1449 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1450 Invert = true; 1451 } 1452 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1453 1454 // Keep a count of how many times instructions are used within CondBB when 1455 // they are candidates for sinking into CondBB. Specifically: 1456 // - They are defined in BB, and 1457 // - They have no side effects, and 1458 // - All of their uses are in CondBB. 1459 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1460 1461 unsigned SpeculationCost = 0; 1462 Value *SpeculatedStoreValue = 0; 1463 StoreInst *SpeculatedStore = 0; 1464 for (BasicBlock::iterator BBI = ThenBB->begin(), 1465 BBE = llvm::prior(ThenBB->end()); 1466 BBI != BBE; ++BBI) { 1467 Instruction *I = BBI; 1468 // Skip debug info. 1469 if (isa<DbgInfoIntrinsic>(I)) 1470 continue; 1471 1472 // Only speculatively execution a single instruction (not counting the 1473 // terminator) for now. 1474 ++SpeculationCost; 1475 if (SpeculationCost > 1) 1476 return false; 1477 1478 // Don't hoist the instruction if it's unsafe or expensive. 1479 if (!isSafeToSpeculativelyExecute(I) && 1480 !(HoistCondStores && 1481 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, 1482 EndBB)))) 1483 return false; 1484 if (!SpeculatedStoreValue && 1485 ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1486 return false; 1487 1488 // Store the store speculation candidate. 1489 if (SpeculatedStoreValue) 1490 SpeculatedStore = cast<StoreInst>(I); 1491 1492 // Do not hoist the instruction if any of its operands are defined but not 1493 // used in BB. The transformation will prevent the operand from 1494 // being sunk into the use block. 1495 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1496 i != e; ++i) { 1497 Instruction *OpI = dyn_cast<Instruction>(*i); 1498 if (!OpI || OpI->getParent() != BB || 1499 OpI->mayHaveSideEffects()) 1500 continue; // Not a candidate for sinking. 1501 1502 ++SinkCandidateUseCounts[OpI]; 1503 } 1504 } 1505 1506 // Consider any sink candidates which are only used in CondBB as costs for 1507 // speculation. Note, while we iterate over a DenseMap here, we are summing 1508 // and so iteration order isn't significant. 1509 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1510 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1511 I != E; ++I) 1512 if (I->first->getNumUses() == I->second) { 1513 ++SpeculationCost; 1514 if (SpeculationCost > 1) 1515 return false; 1516 } 1517 1518 // Check that the PHI nodes can be converted to selects. 1519 bool HaveRewritablePHIs = false; 1520 for (BasicBlock::iterator I = EndBB->begin(); 1521 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1522 Value *OrigV = PN->getIncomingValueForBlock(BB); 1523 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1524 1525 // Skip PHIs which are trivial. 1526 if (ThenV == OrigV) 1527 continue; 1528 1529 HaveRewritablePHIs = true; 1530 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV); 1531 if (!CE) 1532 continue; // Known safe and cheap. 1533 1534 if (!isSafeToSpeculativelyExecute(CE)) 1535 return false; 1536 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1537 return false; 1538 1539 // Account for the cost of an unfolded ConstantExpr which could end up 1540 // getting expanded into Instructions. 1541 // FIXME: This doesn't account for how many operations are combined in the 1542 // constant expression. 1543 ++SpeculationCost; 1544 if (SpeculationCost > 1) 1545 return false; 1546 } 1547 1548 // If there are no PHIs to process, bail early. This helps ensure idempotence 1549 // as well. 1550 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) 1551 return false; 1552 1553 // If we get here, we can hoist the instruction and if-convert. 1554 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1555 1556 // Insert a select of the value of the speculated store. 1557 if (SpeculatedStoreValue) { 1558 IRBuilder<true, NoFolder> Builder(BI); 1559 Value *TrueV = SpeculatedStore->getValueOperand(); 1560 Value *FalseV = SpeculatedStoreValue; 1561 if (Invert) 1562 std::swap(TrueV, FalseV); 1563 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + 1564 "." + FalseV->getName()); 1565 SpeculatedStore->setOperand(0, S); 1566 } 1567 1568 // Hoist the instructions. 1569 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1570 llvm::prior(ThenBB->end())); 1571 1572 // Insert selects and rewrite the PHI operands. 1573 IRBuilder<true, NoFolder> Builder(BI); 1574 for (BasicBlock::iterator I = EndBB->begin(); 1575 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1576 unsigned OrigI = PN->getBasicBlockIndex(BB); 1577 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1578 Value *OrigV = PN->getIncomingValue(OrigI); 1579 Value *ThenV = PN->getIncomingValue(ThenI); 1580 1581 // Skip PHIs which are trivial. 1582 if (OrigV == ThenV) 1583 continue; 1584 1585 // Create a select whose true value is the speculatively executed value and 1586 // false value is the preexisting value. Swap them if the branch 1587 // destinations were inverted. 1588 Value *TrueV = ThenV, *FalseV = OrigV; 1589 if (Invert) 1590 std::swap(TrueV, FalseV); 1591 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1592 TrueV->getName() + "." + FalseV->getName()); 1593 PN->setIncomingValue(OrigI, V); 1594 PN->setIncomingValue(ThenI, V); 1595 } 1596 1597 ++NumSpeculations; 1598 return true; 1599} 1600 1601/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1602/// across this block. 1603static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1604 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1605 unsigned Size = 0; 1606 1607 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1608 if (isa<DbgInfoIntrinsic>(BBI)) 1609 continue; 1610 if (Size > 10) return false; // Don't clone large BB's. 1611 ++Size; 1612 1613 // We can only support instructions that do not define values that are 1614 // live outside of the current basic block. 1615 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1616 UI != E; ++UI) { 1617 Instruction *U = cast<Instruction>(*UI); 1618 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1619 } 1620 1621 // Looks ok, continue checking. 1622 } 1623 1624 return true; 1625} 1626 1627/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1628/// that is defined in the same block as the branch and if any PHI entries are 1629/// constants, thread edges corresponding to that entry to be branches to their 1630/// ultimate destination. 1631static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1632 BasicBlock *BB = BI->getParent(); 1633 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1634 // NOTE: we currently cannot transform this case if the PHI node is used 1635 // outside of the block. 1636 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1637 return false; 1638 1639 // Degenerate case of a single entry PHI. 1640 if (PN->getNumIncomingValues() == 1) { 1641 FoldSingleEntryPHINodes(PN->getParent()); 1642 return true; 1643 } 1644 1645 // Now we know that this block has multiple preds and two succs. 1646 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1647 1648 // Okay, this is a simple enough basic block. See if any phi values are 1649 // constants. 1650 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1651 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1652 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1653 1654 // Okay, we now know that all edges from PredBB should be revectored to 1655 // branch to RealDest. 1656 BasicBlock *PredBB = PN->getIncomingBlock(i); 1657 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1658 1659 if (RealDest == BB) continue; // Skip self loops. 1660 // Skip if the predecessor's terminator is an indirect branch. 1661 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1662 1663 // The dest block might have PHI nodes, other predecessors and other 1664 // difficult cases. Instead of being smart about this, just insert a new 1665 // block that jumps to the destination block, effectively splitting 1666 // the edge we are about to create. 1667 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1668 RealDest->getName()+".critedge", 1669 RealDest->getParent(), RealDest); 1670 BranchInst::Create(RealDest, EdgeBB); 1671 1672 // Update PHI nodes. 1673 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1674 1675 // BB may have instructions that are being threaded over. Clone these 1676 // instructions into EdgeBB. We know that there will be no uses of the 1677 // cloned instructions outside of EdgeBB. 1678 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1679 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1680 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1681 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1682 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1683 continue; 1684 } 1685 // Clone the instruction. 1686 Instruction *N = BBI->clone(); 1687 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1688 1689 // Update operands due to translation. 1690 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1691 i != e; ++i) { 1692 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1693 if (PI != TranslateMap.end()) 1694 *i = PI->second; 1695 } 1696 1697 // Check for trivial simplification. 1698 if (Value *V = SimplifyInstruction(N, TD)) { 1699 TranslateMap[BBI] = V; 1700 delete N; // Instruction folded away, don't need actual inst 1701 } else { 1702 // Insert the new instruction into its new home. 1703 EdgeBB->getInstList().insert(InsertPt, N); 1704 if (!BBI->use_empty()) 1705 TranslateMap[BBI] = N; 1706 } 1707 } 1708 1709 // Loop over all of the edges from PredBB to BB, changing them to branch 1710 // to EdgeBB instead. 1711 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1712 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1713 if (PredBBTI->getSuccessor(i) == BB) { 1714 BB->removePredecessor(PredBB); 1715 PredBBTI->setSuccessor(i, EdgeBB); 1716 } 1717 1718 // Recurse, simplifying any other constants. 1719 return FoldCondBranchOnPHI(BI, TD) | true; 1720 } 1721 1722 return false; 1723} 1724 1725/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1726/// PHI node, see if we can eliminate it. 1727static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1728 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1729 // statement", which has a very simple dominance structure. Basically, we 1730 // are trying to find the condition that is being branched on, which 1731 // subsequently causes this merge to happen. We really want control 1732 // dependence information for this check, but simplifycfg can't keep it up 1733 // to date, and this catches most of the cases we care about anyway. 1734 BasicBlock *BB = PN->getParent(); 1735 BasicBlock *IfTrue, *IfFalse; 1736 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1737 if (!IfCond || 1738 // Don't bother if the branch will be constant folded trivially. 1739 isa<ConstantInt>(IfCond)) 1740 return false; 1741 1742 // Okay, we found that we can merge this two-entry phi node into a select. 1743 // Doing so would require us to fold *all* two entry phi nodes in this block. 1744 // At some point this becomes non-profitable (particularly if the target 1745 // doesn't support cmov's). Only do this transformation if there are two or 1746 // fewer PHI nodes in this block. 1747 unsigned NumPhis = 0; 1748 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1749 if (NumPhis > 2) 1750 return false; 1751 1752 // Loop over the PHI's seeing if we can promote them all to select 1753 // instructions. While we are at it, keep track of the instructions 1754 // that need to be moved to the dominating block. 1755 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1756 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1757 MaxCostVal1 = PHINodeFoldingThreshold; 1758 1759 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1760 PHINode *PN = cast<PHINode>(II++); 1761 if (Value *V = SimplifyInstruction(PN, TD)) { 1762 PN->replaceAllUsesWith(V); 1763 PN->eraseFromParent(); 1764 continue; 1765 } 1766 1767 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1768 MaxCostVal0) || 1769 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1770 MaxCostVal1)) 1771 return false; 1772 } 1773 1774 // If we folded the first phi, PN dangles at this point. Refresh it. If 1775 // we ran out of PHIs then we simplified them all. 1776 PN = dyn_cast<PHINode>(BB->begin()); 1777 if (PN == 0) return true; 1778 1779 // Don't fold i1 branches on PHIs which contain binary operators. These can 1780 // often be turned into switches and other things. 1781 if (PN->getType()->isIntegerTy(1) && 1782 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1783 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1784 isa<BinaryOperator>(IfCond))) 1785 return false; 1786 1787 // If we all PHI nodes are promotable, check to make sure that all 1788 // instructions in the predecessor blocks can be promoted as well. If 1789 // not, we won't be able to get rid of the control flow, so it's not 1790 // worth promoting to select instructions. 1791 BasicBlock *DomBlock = 0; 1792 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1793 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1794 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1795 IfBlock1 = 0; 1796 } else { 1797 DomBlock = *pred_begin(IfBlock1); 1798 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1799 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1800 // This is not an aggressive instruction that we can promote. 1801 // Because of this, we won't be able to get rid of the control 1802 // flow, so the xform is not worth it. 1803 return false; 1804 } 1805 } 1806 1807 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1808 IfBlock2 = 0; 1809 } else { 1810 DomBlock = *pred_begin(IfBlock2); 1811 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1812 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1813 // This is not an aggressive instruction that we can promote. 1814 // Because of this, we won't be able to get rid of the control 1815 // flow, so the xform is not worth it. 1816 return false; 1817 } 1818 } 1819 1820 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1821 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1822 1823 // If we can still promote the PHI nodes after this gauntlet of tests, 1824 // do all of the PHI's now. 1825 Instruction *InsertPt = DomBlock->getTerminator(); 1826 IRBuilder<true, NoFolder> Builder(InsertPt); 1827 1828 // Move all 'aggressive' instructions, which are defined in the 1829 // conditional parts of the if's up to the dominating block. 1830 if (IfBlock1) 1831 DomBlock->getInstList().splice(InsertPt, 1832 IfBlock1->getInstList(), IfBlock1->begin(), 1833 IfBlock1->getTerminator()); 1834 if (IfBlock2) 1835 DomBlock->getInstList().splice(InsertPt, 1836 IfBlock2->getInstList(), IfBlock2->begin(), 1837 IfBlock2->getTerminator()); 1838 1839 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1840 // Change the PHI node into a select instruction. 1841 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1842 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1843 1844 SelectInst *NV = 1845 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1846 PN->replaceAllUsesWith(NV); 1847 NV->takeName(PN); 1848 PN->eraseFromParent(); 1849 } 1850 1851 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1852 // has been flattened. Change DomBlock to jump directly to our new block to 1853 // avoid other simplifycfg's kicking in on the diamond. 1854 TerminatorInst *OldTI = DomBlock->getTerminator(); 1855 Builder.SetInsertPoint(OldTI); 1856 Builder.CreateBr(BB); 1857 OldTI->eraseFromParent(); 1858 return true; 1859} 1860 1861/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1862/// to two returning blocks, try to merge them together into one return, 1863/// introducing a select if the return values disagree. 1864static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1865 IRBuilder<> &Builder) { 1866 assert(BI->isConditional() && "Must be a conditional branch"); 1867 BasicBlock *TrueSucc = BI->getSuccessor(0); 1868 BasicBlock *FalseSucc = BI->getSuccessor(1); 1869 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1870 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1871 1872 // Check to ensure both blocks are empty (just a return) or optionally empty 1873 // with PHI nodes. If there are other instructions, merging would cause extra 1874 // computation on one path or the other. 1875 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1876 return false; 1877 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1878 return false; 1879 1880 Builder.SetInsertPoint(BI); 1881 // Okay, we found a branch that is going to two return nodes. If 1882 // there is no return value for this function, just change the 1883 // branch into a return. 1884 if (FalseRet->getNumOperands() == 0) { 1885 TrueSucc->removePredecessor(BI->getParent()); 1886 FalseSucc->removePredecessor(BI->getParent()); 1887 Builder.CreateRetVoid(); 1888 EraseTerminatorInstAndDCECond(BI); 1889 return true; 1890 } 1891 1892 // Otherwise, figure out what the true and false return values are 1893 // so we can insert a new select instruction. 1894 Value *TrueValue = TrueRet->getReturnValue(); 1895 Value *FalseValue = FalseRet->getReturnValue(); 1896 1897 // Unwrap any PHI nodes in the return blocks. 1898 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1899 if (TVPN->getParent() == TrueSucc) 1900 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1901 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1902 if (FVPN->getParent() == FalseSucc) 1903 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1904 1905 // In order for this transformation to be safe, we must be able to 1906 // unconditionally execute both operands to the return. This is 1907 // normally the case, but we could have a potentially-trapping 1908 // constant expression that prevents this transformation from being 1909 // safe. 1910 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1911 if (TCV->canTrap()) 1912 return false; 1913 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1914 if (FCV->canTrap()) 1915 return false; 1916 1917 // Okay, we collected all the mapped values and checked them for sanity, and 1918 // defined to really do this transformation. First, update the CFG. 1919 TrueSucc->removePredecessor(BI->getParent()); 1920 FalseSucc->removePredecessor(BI->getParent()); 1921 1922 // Insert select instructions where needed. 1923 Value *BrCond = BI->getCondition(); 1924 if (TrueValue) { 1925 // Insert a select if the results differ. 1926 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1927 } else if (isa<UndefValue>(TrueValue)) { 1928 TrueValue = FalseValue; 1929 } else { 1930 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1931 FalseValue, "retval"); 1932 } 1933 } 1934 1935 Value *RI = !TrueValue ? 1936 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1937 1938 (void) RI; 1939 1940 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1941 << "\n " << *BI << "NewRet = " << *RI 1942 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1943 1944 EraseTerminatorInstAndDCECond(BI); 1945 1946 return true; 1947} 1948 1949/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1950/// probabilities of the branch taking each edge. Fills in the two APInt 1951/// parameters and return true, or returns false if no or invalid metadata was 1952/// found. 1953static bool ExtractBranchMetadata(BranchInst *BI, 1954 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1955 assert(BI->isConditional() && 1956 "Looking for probabilities on unconditional branch?"); 1957 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1958 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1959 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1960 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1961 if (!CITrue || !CIFalse) return false; 1962 ProbTrue = CITrue->getValue().getZExtValue(); 1963 ProbFalse = CIFalse->getValue().getZExtValue(); 1964 return true; 1965} 1966 1967/// checkCSEInPredecessor - Return true if the given instruction is available 1968/// in its predecessor block. If yes, the instruction will be removed. 1969/// 1970static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1971 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1972 return false; 1973 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1974 Instruction *PBI = &*I; 1975 // Check whether Inst and PBI generate the same value. 1976 if (Inst->isIdenticalTo(PBI)) { 1977 Inst->replaceAllUsesWith(PBI); 1978 Inst->eraseFromParent(); 1979 return true; 1980 } 1981 } 1982 return false; 1983} 1984 1985/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1986/// predecessor branches to us and one of our successors, fold the block into 1987/// the predecessor and use logical operations to pick the right destination. 1988bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1989 BasicBlock *BB = BI->getParent(); 1990 1991 Instruction *Cond = 0; 1992 if (BI->isConditional()) 1993 Cond = dyn_cast<Instruction>(BI->getCondition()); 1994 else { 1995 // For unconditional branch, check for a simple CFG pattern, where 1996 // BB has a single predecessor and BB's successor is also its predecessor's 1997 // successor. If such pattern exisits, check for CSE between BB and its 1998 // predecessor. 1999 if (BasicBlock *PB = BB->getSinglePredecessor()) 2000 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 2001 if (PBI->isConditional() && 2002 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 2003 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 2004 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 2005 I != E; ) { 2006 Instruction *Curr = I++; 2007 if (isa<CmpInst>(Curr)) { 2008 Cond = Curr; 2009 break; 2010 } 2011 // Quit if we can't remove this instruction. 2012 if (!checkCSEInPredecessor(Curr, PB)) 2013 return false; 2014 } 2015 } 2016 2017 if (Cond == 0) 2018 return false; 2019 } 2020 2021 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 2022 Cond->getParent() != BB || !Cond->hasOneUse()) 2023 return false; 2024 2025 // Only allow this if the condition is a simple instruction that can be 2026 // executed unconditionally. It must be in the same block as the branch, and 2027 // must be at the front of the block. 2028 BasicBlock::iterator FrontIt = BB->front(); 2029 2030 // Ignore dbg intrinsics. 2031 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2032 2033 // Allow a single instruction to be hoisted in addition to the compare 2034 // that feeds the branch. We later ensure that any values that _it_ uses 2035 // were also live in the predecessor, so that we don't unnecessarily create 2036 // register pressure or inhibit out-of-order execution. 2037 Instruction *BonusInst = 0; 2038 if (&*FrontIt != Cond && 2039 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 2040 isSafeToSpeculativelyExecute(FrontIt)) { 2041 BonusInst = &*FrontIt; 2042 ++FrontIt; 2043 2044 // Ignore dbg intrinsics. 2045 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2046 } 2047 2048 // Only a single bonus inst is allowed. 2049 if (&*FrontIt != Cond) 2050 return false; 2051 2052 // Make sure the instruction after the condition is the cond branch. 2053 BasicBlock::iterator CondIt = Cond; ++CondIt; 2054 2055 // Ingore dbg intrinsics. 2056 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 2057 2058 if (&*CondIt != BI) 2059 return false; 2060 2061 // Cond is known to be a compare or binary operator. Check to make sure that 2062 // neither operand is a potentially-trapping constant expression. 2063 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 2064 if (CE->canTrap()) 2065 return false; 2066 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 2067 if (CE->canTrap()) 2068 return false; 2069 2070 // Finally, don't infinitely unroll conditional loops. 2071 BasicBlock *TrueDest = BI->getSuccessor(0); 2072 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 2073 if (TrueDest == BB || FalseDest == BB) 2074 return false; 2075 2076 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2077 BasicBlock *PredBlock = *PI; 2078 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 2079 2080 // Check that we have two conditional branches. If there is a PHI node in 2081 // the common successor, verify that the same value flows in from both 2082 // blocks. 2083 SmallVector<PHINode*, 4> PHIs; 2084 if (PBI == 0 || PBI->isUnconditional() || 2085 (BI->isConditional() && 2086 !SafeToMergeTerminators(BI, PBI)) || 2087 (!BI->isConditional() && 2088 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2089 continue; 2090 2091 // Determine if the two branches share a common destination. 2092 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2093 bool InvertPredCond = false; 2094 2095 if (BI->isConditional()) { 2096 if (PBI->getSuccessor(0) == TrueDest) 2097 Opc = Instruction::Or; 2098 else if (PBI->getSuccessor(1) == FalseDest) 2099 Opc = Instruction::And; 2100 else if (PBI->getSuccessor(0) == FalseDest) 2101 Opc = Instruction::And, InvertPredCond = true; 2102 else if (PBI->getSuccessor(1) == TrueDest) 2103 Opc = Instruction::Or, InvertPredCond = true; 2104 else 2105 continue; 2106 } else { 2107 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2108 continue; 2109 } 2110 2111 // Ensure that any values used in the bonus instruction are also used 2112 // by the terminator of the predecessor. This means that those values 2113 // must already have been resolved, so we won't be inhibiting the 2114 // out-of-order core by speculating them earlier. 2115 if (BonusInst) { 2116 // Collect the values used by the bonus inst 2117 SmallPtrSet<Value*, 4> UsedValues; 2118 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2119 OE = BonusInst->op_end(); OI != OE; ++OI) { 2120 Value *V = *OI; 2121 if (!isa<Constant>(V)) 2122 UsedValues.insert(V); 2123 } 2124 2125 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2126 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2127 2128 // Walk up to four levels back up the use-def chain of the predecessor's 2129 // terminator to see if all those values were used. The choice of four 2130 // levels is arbitrary, to provide a compile-time-cost bound. 2131 while (!Worklist.empty()) { 2132 std::pair<Value*, unsigned> Pair = Worklist.back(); 2133 Worklist.pop_back(); 2134 2135 if (Pair.second >= 4) continue; 2136 UsedValues.erase(Pair.first); 2137 if (UsedValues.empty()) break; 2138 2139 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2140 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2141 OI != OE; ++OI) 2142 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2143 } 2144 } 2145 2146 if (!UsedValues.empty()) return false; 2147 } 2148 2149 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2150 IRBuilder<> Builder(PBI); 2151 2152 // If we need to invert the condition in the pred block to match, do so now. 2153 if (InvertPredCond) { 2154 Value *NewCond = PBI->getCondition(); 2155 2156 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2157 CmpInst *CI = cast<CmpInst>(NewCond); 2158 CI->setPredicate(CI->getInversePredicate()); 2159 } else { 2160 NewCond = Builder.CreateNot(NewCond, 2161 PBI->getCondition()->getName()+".not"); 2162 } 2163 2164 PBI->setCondition(NewCond); 2165 PBI->swapSuccessors(); 2166 } 2167 2168 // If we have a bonus inst, clone it into the predecessor block. 2169 Instruction *NewBonus = 0; 2170 if (BonusInst) { 2171 NewBonus = BonusInst->clone(); 2172 PredBlock->getInstList().insert(PBI, NewBonus); 2173 NewBonus->takeName(BonusInst); 2174 BonusInst->setName(BonusInst->getName()+".old"); 2175 } 2176 2177 // Clone Cond into the predecessor basic block, and or/and the 2178 // two conditions together. 2179 Instruction *New = Cond->clone(); 2180 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2181 PredBlock->getInstList().insert(PBI, New); 2182 New->takeName(Cond); 2183 Cond->setName(New->getName()+".old"); 2184 2185 if (BI->isConditional()) { 2186 Instruction *NewCond = 2187 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2188 New, "or.cond")); 2189 PBI->setCondition(NewCond); 2190 2191 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2192 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2193 PredFalseWeight); 2194 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2195 SuccFalseWeight); 2196 SmallVector<uint64_t, 8> NewWeights; 2197 2198 if (PBI->getSuccessor(0) == BB) { 2199 if (PredHasWeights && SuccHasWeights) { 2200 // PBI: br i1 %x, BB, FalseDest 2201 // BI: br i1 %y, TrueDest, FalseDest 2202 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2203 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2204 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2205 // TrueWeight for PBI * FalseWeight for BI. 2206 // We assume that total weights of a BranchInst can fit into 32 bits. 2207 // Therefore, we will not have overflow using 64-bit arithmetic. 2208 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2209 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2210 } 2211 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2212 PBI->setSuccessor(0, TrueDest); 2213 } 2214 if (PBI->getSuccessor(1) == BB) { 2215 if (PredHasWeights && SuccHasWeights) { 2216 // PBI: br i1 %x, TrueDest, BB 2217 // BI: br i1 %y, TrueDest, FalseDest 2218 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2219 // FalseWeight for PBI * TrueWeight for BI. 2220 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2221 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2222 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2223 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2224 } 2225 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2226 PBI->setSuccessor(1, FalseDest); 2227 } 2228 if (NewWeights.size() == 2) { 2229 // Halve the weights if any of them cannot fit in an uint32_t 2230 FitWeights(NewWeights); 2231 2232 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2233 PBI->setMetadata(LLVMContext::MD_prof, 2234 MDBuilder(BI->getContext()). 2235 createBranchWeights(MDWeights)); 2236 } else 2237 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2238 } else { 2239 // Update PHI nodes in the common successors. 2240 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2241 ConstantInt *PBI_C = cast<ConstantInt>( 2242 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2243 assert(PBI_C->getType()->isIntegerTy(1)); 2244 Instruction *MergedCond = 0; 2245 if (PBI->getSuccessor(0) == TrueDest) { 2246 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2247 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2248 // is false: !PBI_Cond and BI_Value 2249 Instruction *NotCond = 2250 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2251 "not.cond")); 2252 MergedCond = 2253 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2254 NotCond, New, 2255 "and.cond")); 2256 if (PBI_C->isOne()) 2257 MergedCond = 2258 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2259 PBI->getCondition(), MergedCond, 2260 "or.cond")); 2261 } else { 2262 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2263 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2264 // is false: PBI_Cond and BI_Value 2265 MergedCond = 2266 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2267 PBI->getCondition(), New, 2268 "and.cond")); 2269 if (PBI_C->isOne()) { 2270 Instruction *NotCond = 2271 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2272 "not.cond")); 2273 MergedCond = 2274 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2275 NotCond, MergedCond, 2276 "or.cond")); 2277 } 2278 } 2279 // Update PHI Node. 2280 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2281 MergedCond); 2282 } 2283 // Change PBI from Conditional to Unconditional. 2284 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2285 EraseTerminatorInstAndDCECond(PBI); 2286 PBI = New_PBI; 2287 } 2288 2289 // TODO: If BB is reachable from all paths through PredBlock, then we 2290 // could replace PBI's branch probabilities with BI's. 2291 2292 // Copy any debug value intrinsics into the end of PredBlock. 2293 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2294 if (isa<DbgInfoIntrinsic>(*I)) 2295 I->clone()->insertBefore(PBI); 2296 2297 return true; 2298 } 2299 return false; 2300} 2301 2302/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2303/// predecessor of another block, this function tries to simplify it. We know 2304/// that PBI and BI are both conditional branches, and BI is in one of the 2305/// successor blocks of PBI - PBI branches to BI. 2306static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2307 assert(PBI->isConditional() && BI->isConditional()); 2308 BasicBlock *BB = BI->getParent(); 2309 2310 // If this block ends with a branch instruction, and if there is a 2311 // predecessor that ends on a branch of the same condition, make 2312 // this conditional branch redundant. 2313 if (PBI->getCondition() == BI->getCondition() && 2314 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2315 // Okay, the outcome of this conditional branch is statically 2316 // knowable. If this block had a single pred, handle specially. 2317 if (BB->getSinglePredecessor()) { 2318 // Turn this into a branch on constant. 2319 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2320 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2321 CondIsTrue)); 2322 return true; // Nuke the branch on constant. 2323 } 2324 2325 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2326 // in the constant and simplify the block result. Subsequent passes of 2327 // simplifycfg will thread the block. 2328 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2329 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2330 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2331 std::distance(PB, PE), 2332 BI->getCondition()->getName() + ".pr", 2333 BB->begin()); 2334 // Okay, we're going to insert the PHI node. Since PBI is not the only 2335 // predecessor, compute the PHI'd conditional value for all of the preds. 2336 // Any predecessor where the condition is not computable we keep symbolic. 2337 for (pred_iterator PI = PB; PI != PE; ++PI) { 2338 BasicBlock *P = *PI; 2339 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2340 PBI != BI && PBI->isConditional() && 2341 PBI->getCondition() == BI->getCondition() && 2342 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2343 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2344 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2345 CondIsTrue), P); 2346 } else { 2347 NewPN->addIncoming(BI->getCondition(), P); 2348 } 2349 } 2350 2351 BI->setCondition(NewPN); 2352 return true; 2353 } 2354 } 2355 2356 // If this is a conditional branch in an empty block, and if any 2357 // predecessors is a conditional branch to one of our destinations, 2358 // fold the conditions into logical ops and one cond br. 2359 BasicBlock::iterator BBI = BB->begin(); 2360 // Ignore dbg intrinsics. 2361 while (isa<DbgInfoIntrinsic>(BBI)) 2362 ++BBI; 2363 if (&*BBI != BI) 2364 return false; 2365 2366 2367 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2368 if (CE->canTrap()) 2369 return false; 2370 2371 int PBIOp, BIOp; 2372 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2373 PBIOp = BIOp = 0; 2374 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2375 PBIOp = 0, BIOp = 1; 2376 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2377 PBIOp = 1, BIOp = 0; 2378 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2379 PBIOp = BIOp = 1; 2380 else 2381 return false; 2382 2383 // Check to make sure that the other destination of this branch 2384 // isn't BB itself. If so, this is an infinite loop that will 2385 // keep getting unwound. 2386 if (PBI->getSuccessor(PBIOp) == BB) 2387 return false; 2388 2389 // Do not perform this transformation if it would require 2390 // insertion of a large number of select instructions. For targets 2391 // without predication/cmovs, this is a big pessimization. 2392 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2393 2394 unsigned NumPhis = 0; 2395 for (BasicBlock::iterator II = CommonDest->begin(); 2396 isa<PHINode>(II); ++II, ++NumPhis) 2397 if (NumPhis > 2) // Disable this xform. 2398 return false; 2399 2400 // Finally, if everything is ok, fold the branches to logical ops. 2401 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2402 2403 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2404 << "AND: " << *BI->getParent()); 2405 2406 2407 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2408 // branch in it, where one edge (OtherDest) goes back to itself but the other 2409 // exits. We don't *know* that the program avoids the infinite loop 2410 // (even though that seems likely). If we do this xform naively, we'll end up 2411 // recursively unpeeling the loop. Since we know that (after the xform is 2412 // done) that the block *is* infinite if reached, we just make it an obviously 2413 // infinite loop with no cond branch. 2414 if (OtherDest == BB) { 2415 // Insert it at the end of the function, because it's either code, 2416 // or it won't matter if it's hot. :) 2417 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2418 "infloop", BB->getParent()); 2419 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2420 OtherDest = InfLoopBlock; 2421 } 2422 2423 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2424 2425 // BI may have other predecessors. Because of this, we leave 2426 // it alone, but modify PBI. 2427 2428 // Make sure we get to CommonDest on True&True directions. 2429 Value *PBICond = PBI->getCondition(); 2430 IRBuilder<true, NoFolder> Builder(PBI); 2431 if (PBIOp) 2432 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2433 2434 Value *BICond = BI->getCondition(); 2435 if (BIOp) 2436 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2437 2438 // Merge the conditions. 2439 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2440 2441 // Modify PBI to branch on the new condition to the new dests. 2442 PBI->setCondition(Cond); 2443 PBI->setSuccessor(0, CommonDest); 2444 PBI->setSuccessor(1, OtherDest); 2445 2446 // Update branch weight for PBI. 2447 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2448 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2449 PredFalseWeight); 2450 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2451 SuccFalseWeight); 2452 if (PredHasWeights && SuccHasWeights) { 2453 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2454 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2455 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2456 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2457 // The weight to CommonDest should be PredCommon * SuccTotal + 2458 // PredOther * SuccCommon. 2459 // The weight to OtherDest should be PredOther * SuccOther. 2460 SmallVector<uint64_t, 2> NewWeights; 2461 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2462 PredOther * SuccCommon); 2463 NewWeights.push_back(PredOther * SuccOther); 2464 // Halve the weights if any of them cannot fit in an uint32_t 2465 FitWeights(NewWeights); 2466 2467 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2468 PBI->setMetadata(LLVMContext::MD_prof, 2469 MDBuilder(BI->getContext()). 2470 createBranchWeights(MDWeights)); 2471 } 2472 2473 // OtherDest may have phi nodes. If so, add an entry from PBI's 2474 // block that are identical to the entries for BI's block. 2475 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2476 2477 // We know that the CommonDest already had an edge from PBI to 2478 // it. If it has PHIs though, the PHIs may have different 2479 // entries for BB and PBI's BB. If so, insert a select to make 2480 // them agree. 2481 PHINode *PN; 2482 for (BasicBlock::iterator II = CommonDest->begin(); 2483 (PN = dyn_cast<PHINode>(II)); ++II) { 2484 Value *BIV = PN->getIncomingValueForBlock(BB); 2485 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2486 Value *PBIV = PN->getIncomingValue(PBBIdx); 2487 if (BIV != PBIV) { 2488 // Insert a select in PBI to pick the right value. 2489 Value *NV = cast<SelectInst> 2490 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2491 PN->setIncomingValue(PBBIdx, NV); 2492 } 2493 } 2494 2495 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2496 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2497 2498 // This basic block is probably dead. We know it has at least 2499 // one fewer predecessor. 2500 return true; 2501} 2502 2503// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2504// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2505// Takes care of updating the successors and removing the old terminator. 2506// Also makes sure not to introduce new successors by assuming that edges to 2507// non-successor TrueBBs and FalseBBs aren't reachable. 2508static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2509 BasicBlock *TrueBB, BasicBlock *FalseBB, 2510 uint32_t TrueWeight, 2511 uint32_t FalseWeight){ 2512 // Remove any superfluous successor edges from the CFG. 2513 // First, figure out which successors to preserve. 2514 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2515 // successor. 2516 BasicBlock *KeepEdge1 = TrueBB; 2517 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2518 2519 // Then remove the rest. 2520 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2521 BasicBlock *Succ = OldTerm->getSuccessor(I); 2522 // Make sure only to keep exactly one copy of each edge. 2523 if (Succ == KeepEdge1) 2524 KeepEdge1 = 0; 2525 else if (Succ == KeepEdge2) 2526 KeepEdge2 = 0; 2527 else 2528 Succ->removePredecessor(OldTerm->getParent()); 2529 } 2530 2531 IRBuilder<> Builder(OldTerm); 2532 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2533 2534 // Insert an appropriate new terminator. 2535 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2536 if (TrueBB == FalseBB) 2537 // We were only looking for one successor, and it was present. 2538 // Create an unconditional branch to it. 2539 Builder.CreateBr(TrueBB); 2540 else { 2541 // We found both of the successors we were looking for. 2542 // Create a conditional branch sharing the condition of the select. 2543 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2544 if (TrueWeight != FalseWeight) 2545 NewBI->setMetadata(LLVMContext::MD_prof, 2546 MDBuilder(OldTerm->getContext()). 2547 createBranchWeights(TrueWeight, FalseWeight)); 2548 } 2549 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2550 // Neither of the selected blocks were successors, so this 2551 // terminator must be unreachable. 2552 new UnreachableInst(OldTerm->getContext(), OldTerm); 2553 } else { 2554 // One of the selected values was a successor, but the other wasn't. 2555 // Insert an unconditional branch to the one that was found; 2556 // the edge to the one that wasn't must be unreachable. 2557 if (KeepEdge1 == 0) 2558 // Only TrueBB was found. 2559 Builder.CreateBr(TrueBB); 2560 else 2561 // Only FalseBB was found. 2562 Builder.CreateBr(FalseBB); 2563 } 2564 2565 EraseTerminatorInstAndDCECond(OldTerm); 2566 return true; 2567} 2568 2569// SimplifySwitchOnSelect - Replaces 2570// (switch (select cond, X, Y)) on constant X, Y 2571// with a branch - conditional if X and Y lead to distinct BBs, 2572// unconditional otherwise. 2573static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2574 // Check for constant integer values in the select. 2575 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2576 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2577 if (!TrueVal || !FalseVal) 2578 return false; 2579 2580 // Find the relevant condition and destinations. 2581 Value *Condition = Select->getCondition(); 2582 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2583 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2584 2585 // Get weight for TrueBB and FalseBB. 2586 uint32_t TrueWeight = 0, FalseWeight = 0; 2587 SmallVector<uint64_t, 8> Weights; 2588 bool HasWeights = HasBranchWeights(SI); 2589 if (HasWeights) { 2590 GetBranchWeights(SI, Weights); 2591 if (Weights.size() == 1 + SI->getNumCases()) { 2592 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2593 getSuccessorIndex()]; 2594 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2595 getSuccessorIndex()]; 2596 } 2597 } 2598 2599 // Perform the actual simplification. 2600 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2601 TrueWeight, FalseWeight); 2602} 2603 2604// SimplifyIndirectBrOnSelect - Replaces 2605// (indirectbr (select cond, blockaddress(@fn, BlockA), 2606// blockaddress(@fn, BlockB))) 2607// with 2608// (br cond, BlockA, BlockB). 2609static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2610 // Check that both operands of the select are block addresses. 2611 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2612 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2613 if (!TBA || !FBA) 2614 return false; 2615 2616 // Extract the actual blocks. 2617 BasicBlock *TrueBB = TBA->getBasicBlock(); 2618 BasicBlock *FalseBB = FBA->getBasicBlock(); 2619 2620 // Perform the actual simplification. 2621 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2622 0, 0); 2623} 2624 2625/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2626/// instruction (a seteq/setne with a constant) as the only instruction in a 2627/// block that ends with an uncond branch. We are looking for a very specific 2628/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2629/// this case, we merge the first two "or's of icmp" into a switch, but then the 2630/// default value goes to an uncond block with a seteq in it, we get something 2631/// like: 2632/// 2633/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2634/// DEFAULT: 2635/// %tmp = icmp eq i8 %A, 92 2636/// br label %end 2637/// end: 2638/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2639/// 2640/// We prefer to split the edge to 'end' so that there is a true/false entry to 2641/// the PHI, merging the third icmp into the switch. 2642static bool TryToSimplifyUncondBranchWithICmpInIt( 2643 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2644 const DataLayout *TD) { 2645 BasicBlock *BB = ICI->getParent(); 2646 2647 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2648 // complex. 2649 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2650 2651 Value *V = ICI->getOperand(0); 2652 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2653 2654 // The pattern we're looking for is where our only predecessor is a switch on 2655 // 'V' and this block is the default case for the switch. In this case we can 2656 // fold the compared value into the switch to simplify things. 2657 BasicBlock *Pred = BB->getSinglePredecessor(); 2658 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2659 2660 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2661 if (SI->getCondition() != V) 2662 return false; 2663 2664 // If BB is reachable on a non-default case, then we simply know the value of 2665 // V in this block. Substitute it and constant fold the icmp instruction 2666 // away. 2667 if (SI->getDefaultDest() != BB) { 2668 ConstantInt *VVal = SI->findCaseDest(BB); 2669 assert(VVal && "Should have a unique destination value"); 2670 ICI->setOperand(0, VVal); 2671 2672 if (Value *V = SimplifyInstruction(ICI, TD)) { 2673 ICI->replaceAllUsesWith(V); 2674 ICI->eraseFromParent(); 2675 } 2676 // BB is now empty, so it is likely to simplify away. 2677 return SimplifyCFG(BB, TTI, TD) | true; 2678 } 2679 2680 // Ok, the block is reachable from the default dest. If the constant we're 2681 // comparing exists in one of the other edges, then we can constant fold ICI 2682 // and zap it. 2683 if (SI->findCaseValue(Cst) != SI->case_default()) { 2684 Value *V; 2685 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2686 V = ConstantInt::getFalse(BB->getContext()); 2687 else 2688 V = ConstantInt::getTrue(BB->getContext()); 2689 2690 ICI->replaceAllUsesWith(V); 2691 ICI->eraseFromParent(); 2692 // BB is now empty, so it is likely to simplify away. 2693 return SimplifyCFG(BB, TTI, TD) | true; 2694 } 2695 2696 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2697 // the block. 2698 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2699 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2700 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2701 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2702 return false; 2703 2704 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2705 // true in the PHI. 2706 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2707 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2708 2709 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2710 std::swap(DefaultCst, NewCst); 2711 2712 // Replace ICI (which is used by the PHI for the default value) with true or 2713 // false depending on if it is EQ or NE. 2714 ICI->replaceAllUsesWith(DefaultCst); 2715 ICI->eraseFromParent(); 2716 2717 // Okay, the switch goes to this block on a default value. Add an edge from 2718 // the switch to the merge point on the compared value. 2719 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2720 BB->getParent(), BB); 2721 SmallVector<uint64_t, 8> Weights; 2722 bool HasWeights = HasBranchWeights(SI); 2723 if (HasWeights) { 2724 GetBranchWeights(SI, Weights); 2725 if (Weights.size() == 1 + SI->getNumCases()) { 2726 // Split weight for default case to case for "Cst". 2727 Weights[0] = (Weights[0]+1) >> 1; 2728 Weights.push_back(Weights[0]); 2729 2730 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2731 SI->setMetadata(LLVMContext::MD_prof, 2732 MDBuilder(SI->getContext()). 2733 createBranchWeights(MDWeights)); 2734 } 2735 } 2736 SI->addCase(Cst, NewBB); 2737 2738 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2739 Builder.SetInsertPoint(NewBB); 2740 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2741 Builder.CreateBr(SuccBlock); 2742 PHIUse->addIncoming(NewCst, NewBB); 2743 return true; 2744} 2745 2746/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2747/// Check to see if it is branching on an or/and chain of icmp instructions, and 2748/// fold it into a switch instruction if so. 2749static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2750 IRBuilder<> &Builder) { 2751 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2752 if (Cond == 0) return false; 2753 2754 2755 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2756 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2757 // 'setne's and'ed together, collect them. 2758 Value *CompVal = 0; 2759 std::vector<ConstantInt*> Values; 2760 bool TrueWhenEqual = true; 2761 Value *ExtraCase = 0; 2762 unsigned UsedICmps = 0; 2763 2764 if (Cond->getOpcode() == Instruction::Or) { 2765 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2766 UsedICmps); 2767 } else if (Cond->getOpcode() == Instruction::And) { 2768 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2769 UsedICmps); 2770 TrueWhenEqual = false; 2771 } 2772 2773 // If we didn't have a multiply compared value, fail. 2774 if (CompVal == 0) return false; 2775 2776 // Avoid turning single icmps into a switch. 2777 if (UsedICmps <= 1) 2778 return false; 2779 2780 // There might be duplicate constants in the list, which the switch 2781 // instruction can't handle, remove them now. 2782 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2783 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2784 2785 // If Extra was used, we require at least two switch values to do the 2786 // transformation. A switch with one value is just an cond branch. 2787 if (ExtraCase && Values.size() < 2) return false; 2788 2789 // TODO: Preserve branch weight metadata, similarly to how 2790 // FoldValueComparisonIntoPredecessors preserves it. 2791 2792 // Figure out which block is which destination. 2793 BasicBlock *DefaultBB = BI->getSuccessor(1); 2794 BasicBlock *EdgeBB = BI->getSuccessor(0); 2795 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2796 2797 BasicBlock *BB = BI->getParent(); 2798 2799 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2800 << " cases into SWITCH. BB is:\n" << *BB); 2801 2802 // If there are any extra values that couldn't be folded into the switch 2803 // then we evaluate them with an explicit branch first. Split the block 2804 // right before the condbr to handle it. 2805 if (ExtraCase) { 2806 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2807 // Remove the uncond branch added to the old block. 2808 TerminatorInst *OldTI = BB->getTerminator(); 2809 Builder.SetInsertPoint(OldTI); 2810 2811 if (TrueWhenEqual) 2812 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2813 else 2814 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2815 2816 OldTI->eraseFromParent(); 2817 2818 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2819 // for the edge we just added. 2820 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2821 2822 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2823 << "\nEXTRABB = " << *BB); 2824 BB = NewBB; 2825 } 2826 2827 Builder.SetInsertPoint(BI); 2828 // Convert pointer to int before we switch. 2829 if (CompVal->getType()->isPointerTy()) { 2830 assert(TD && "Cannot switch on pointer without DataLayout"); 2831 CompVal = Builder.CreatePtrToInt(CompVal, 2832 TD->getIntPtrType(CompVal->getContext()), 2833 "magicptr"); 2834 } 2835 2836 // Create the new switch instruction now. 2837 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2838 2839 // Add all of the 'cases' to the switch instruction. 2840 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2841 New->addCase(Values[i], EdgeBB); 2842 2843 // We added edges from PI to the EdgeBB. As such, if there were any 2844 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2845 // the number of edges added. 2846 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2847 isa<PHINode>(BBI); ++BBI) { 2848 PHINode *PN = cast<PHINode>(BBI); 2849 Value *InVal = PN->getIncomingValueForBlock(BB); 2850 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2851 PN->addIncoming(InVal, BB); 2852 } 2853 2854 // Erase the old branch instruction. 2855 EraseTerminatorInstAndDCECond(BI); 2856 2857 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2858 return true; 2859} 2860 2861bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2862 // If this is a trivial landing pad that just continues unwinding the caught 2863 // exception then zap the landing pad, turning its invokes into calls. 2864 BasicBlock *BB = RI->getParent(); 2865 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2866 if (RI->getValue() != LPInst) 2867 // Not a landing pad, or the resume is not unwinding the exception that 2868 // caused control to branch here. 2869 return false; 2870 2871 // Check that there are no other instructions except for debug intrinsics. 2872 BasicBlock::iterator I = LPInst, E = RI; 2873 while (++I != E) 2874 if (!isa<DbgInfoIntrinsic>(I)) 2875 return false; 2876 2877 // Turn all invokes that unwind here into calls and delete the basic block. 2878 bool InvokeRequiresTableEntry = false; 2879 bool Changed = false; 2880 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2881 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2882 2883 if (II->hasFnAttr(Attribute::UWTable)) { 2884 // Don't remove an `invoke' instruction if the ABI requires an entry into 2885 // the table. 2886 InvokeRequiresTableEntry = true; 2887 continue; 2888 } 2889 2890 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2891 2892 // Insert a call instruction before the invoke. 2893 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2894 Call->takeName(II); 2895 Call->setCallingConv(II->getCallingConv()); 2896 Call->setAttributes(II->getAttributes()); 2897 Call->setDebugLoc(II->getDebugLoc()); 2898 2899 // Anything that used the value produced by the invoke instruction now uses 2900 // the value produced by the call instruction. Note that we do this even 2901 // for void functions and calls with no uses so that the callgraph edge is 2902 // updated. 2903 II->replaceAllUsesWith(Call); 2904 BB->removePredecessor(II->getParent()); 2905 2906 // Insert a branch to the normal destination right before the invoke. 2907 BranchInst::Create(II->getNormalDest(), II); 2908 2909 // Finally, delete the invoke instruction! 2910 II->eraseFromParent(); 2911 Changed = true; 2912 } 2913 2914 if (!InvokeRequiresTableEntry) 2915 // The landingpad is now unreachable. Zap it. 2916 BB->eraseFromParent(); 2917 2918 return Changed; 2919} 2920 2921bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2922 BasicBlock *BB = RI->getParent(); 2923 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2924 2925 // Find predecessors that end with branches. 2926 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2927 SmallVector<BranchInst*, 8> CondBranchPreds; 2928 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2929 BasicBlock *P = *PI; 2930 TerminatorInst *PTI = P->getTerminator(); 2931 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2932 if (BI->isUnconditional()) 2933 UncondBranchPreds.push_back(P); 2934 else 2935 CondBranchPreds.push_back(BI); 2936 } 2937 } 2938 2939 // If we found some, do the transformation! 2940 if (!UncondBranchPreds.empty() && DupRet) { 2941 while (!UncondBranchPreds.empty()) { 2942 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2943 DEBUG(dbgs() << "FOLDING: " << *BB 2944 << "INTO UNCOND BRANCH PRED: " << *Pred); 2945 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2946 } 2947 2948 // If we eliminated all predecessors of the block, delete the block now. 2949 if (pred_begin(BB) == pred_end(BB)) 2950 // We know there are no successors, so just nuke the block. 2951 BB->eraseFromParent(); 2952 2953 return true; 2954 } 2955 2956 // Check out all of the conditional branches going to this return 2957 // instruction. If any of them just select between returns, change the 2958 // branch itself into a select/return pair. 2959 while (!CondBranchPreds.empty()) { 2960 BranchInst *BI = CondBranchPreds.pop_back_val(); 2961 2962 // Check to see if the non-BB successor is also a return block. 2963 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2964 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2965 SimplifyCondBranchToTwoReturns(BI, Builder)) 2966 return true; 2967 } 2968 return false; 2969} 2970 2971bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2972 BasicBlock *BB = UI->getParent(); 2973 2974 bool Changed = false; 2975 2976 // If there are any instructions immediately before the unreachable that can 2977 // be removed, do so. 2978 while (UI != BB->begin()) { 2979 BasicBlock::iterator BBI = UI; 2980 --BBI; 2981 // Do not delete instructions that can have side effects which might cause 2982 // the unreachable to not be reachable; specifically, calls and volatile 2983 // operations may have this effect. 2984 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2985 2986 if (BBI->mayHaveSideEffects()) { 2987 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2988 if (SI->isVolatile()) 2989 break; 2990 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2991 if (LI->isVolatile()) 2992 break; 2993 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2994 if (RMWI->isVolatile()) 2995 break; 2996 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2997 if (CXI->isVolatile()) 2998 break; 2999 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 3000 !isa<LandingPadInst>(BBI)) { 3001 break; 3002 } 3003 // Note that deleting LandingPad's here is in fact okay, although it 3004 // involves a bit of subtle reasoning. If this inst is a LandingPad, 3005 // all the predecessors of this block will be the unwind edges of Invokes, 3006 // and we can therefore guarantee this block will be erased. 3007 } 3008 3009 // Delete this instruction (any uses are guaranteed to be dead) 3010 if (!BBI->use_empty()) 3011 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 3012 BBI->eraseFromParent(); 3013 Changed = true; 3014 } 3015 3016 // If the unreachable instruction is the first in the block, take a gander 3017 // at all of the predecessors of this instruction, and simplify them. 3018 if (&BB->front() != UI) return Changed; 3019 3020 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 3021 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 3022 TerminatorInst *TI = Preds[i]->getTerminator(); 3023 IRBuilder<> Builder(TI); 3024 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 3025 if (BI->isUnconditional()) { 3026 if (BI->getSuccessor(0) == BB) { 3027 new UnreachableInst(TI->getContext(), TI); 3028 TI->eraseFromParent(); 3029 Changed = true; 3030 } 3031 } else { 3032 if (BI->getSuccessor(0) == BB) { 3033 Builder.CreateBr(BI->getSuccessor(1)); 3034 EraseTerminatorInstAndDCECond(BI); 3035 } else if (BI->getSuccessor(1) == BB) { 3036 Builder.CreateBr(BI->getSuccessor(0)); 3037 EraseTerminatorInstAndDCECond(BI); 3038 Changed = true; 3039 } 3040 } 3041 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 3042 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3043 i != e; ++i) 3044 if (i.getCaseSuccessor() == BB) { 3045 BB->removePredecessor(SI->getParent()); 3046 SI->removeCase(i); 3047 --i; --e; 3048 Changed = true; 3049 } 3050 // If the default value is unreachable, figure out the most popular 3051 // destination and make it the default. 3052 if (SI->getDefaultDest() == BB) { 3053 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 3054 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3055 i != e; ++i) { 3056 std::pair<unsigned, unsigned> &entry = 3057 Popularity[i.getCaseSuccessor()]; 3058 if (entry.first == 0) { 3059 entry.first = 1; 3060 entry.second = i.getCaseIndex(); 3061 } else { 3062 entry.first++; 3063 } 3064 } 3065 3066 // Find the most popular block. 3067 unsigned MaxPop = 0; 3068 unsigned MaxIndex = 0; 3069 BasicBlock *MaxBlock = 0; 3070 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 3071 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 3072 if (I->second.first > MaxPop || 3073 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 3074 MaxPop = I->second.first; 3075 MaxIndex = I->second.second; 3076 MaxBlock = I->first; 3077 } 3078 } 3079 if (MaxBlock) { 3080 // Make this the new default, allowing us to delete any explicit 3081 // edges to it. 3082 SI->setDefaultDest(MaxBlock); 3083 Changed = true; 3084 3085 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3086 // it. 3087 if (isa<PHINode>(MaxBlock->begin())) 3088 for (unsigned i = 0; i != MaxPop-1; ++i) 3089 MaxBlock->removePredecessor(SI->getParent()); 3090 3091 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3092 i != e; ++i) 3093 if (i.getCaseSuccessor() == MaxBlock) { 3094 SI->removeCase(i); 3095 --i; --e; 3096 } 3097 } 3098 } 3099 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3100 if (II->getUnwindDest() == BB) { 3101 // Convert the invoke to a call instruction. This would be a good 3102 // place to note that the call does not throw though. 3103 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3104 II->removeFromParent(); // Take out of symbol table 3105 3106 // Insert the call now... 3107 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3108 Builder.SetInsertPoint(BI); 3109 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3110 Args, II->getName()); 3111 CI->setCallingConv(II->getCallingConv()); 3112 CI->setAttributes(II->getAttributes()); 3113 // If the invoke produced a value, the call does now instead. 3114 II->replaceAllUsesWith(CI); 3115 delete II; 3116 Changed = true; 3117 } 3118 } 3119 } 3120 3121 // If this block is now dead, remove it. 3122 if (pred_begin(BB) == pred_end(BB) && 3123 BB != &BB->getParent()->getEntryBlock()) { 3124 // We know there are no successors, so just nuke the block. 3125 BB->eraseFromParent(); 3126 return true; 3127 } 3128 3129 return Changed; 3130} 3131 3132/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3133/// integer range comparison into a sub, an icmp and a branch. 3134static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3135 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3136 3137 // Make sure all cases point to the same destination and gather the values. 3138 SmallVector<ConstantInt *, 16> Cases; 3139 SwitchInst::CaseIt I = SI->case_begin(); 3140 Cases.push_back(I.getCaseValue()); 3141 SwitchInst::CaseIt PrevI = I++; 3142 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3143 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3144 return false; 3145 Cases.push_back(I.getCaseValue()); 3146 } 3147 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3148 3149 // Sort the case values, then check if they form a range we can transform. 3150 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3151 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3152 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3153 return false; 3154 } 3155 3156 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3157 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3158 3159 Value *Sub = SI->getCondition(); 3160 if (!Offset->isNullValue()) 3161 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3162 Value *Cmp; 3163 // If NumCases overflowed, then all possible values jump to the successor. 3164 if (NumCases->isNullValue() && SI->getNumCases() != 0) 3165 Cmp = ConstantInt::getTrue(SI->getContext()); 3166 else 3167 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3168 BranchInst *NewBI = Builder.CreateCondBr( 3169 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3170 3171 // Update weight for the newly-created conditional branch. 3172 SmallVector<uint64_t, 8> Weights; 3173 bool HasWeights = HasBranchWeights(SI); 3174 if (HasWeights) { 3175 GetBranchWeights(SI, Weights); 3176 if (Weights.size() == 1 + SI->getNumCases()) { 3177 // Combine all weights for the cases to be the true weight of NewBI. 3178 // We assume that the sum of all weights for a Terminator can fit into 32 3179 // bits. 3180 uint32_t NewTrueWeight = 0; 3181 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3182 NewTrueWeight += (uint32_t)Weights[I]; 3183 NewBI->setMetadata(LLVMContext::MD_prof, 3184 MDBuilder(SI->getContext()). 3185 createBranchWeights(NewTrueWeight, 3186 (uint32_t)Weights[0])); 3187 } 3188 } 3189 3190 // Prune obsolete incoming values off the successor's PHI nodes. 3191 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3192 isa<PHINode>(BBI); ++BBI) { 3193 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3194 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3195 } 3196 SI->eraseFromParent(); 3197 3198 return true; 3199} 3200 3201/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3202/// and use it to remove dead cases. 3203static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3204 Value *Cond = SI->getCondition(); 3205 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3206 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3207 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3208 3209 // Gather dead cases. 3210 SmallVector<ConstantInt*, 8> DeadCases; 3211 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3212 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3213 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3214 DeadCases.push_back(I.getCaseValue()); 3215 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3216 << I.getCaseValue() << "' is dead.\n"); 3217 } 3218 } 3219 3220 SmallVector<uint64_t, 8> Weights; 3221 bool HasWeight = HasBranchWeights(SI); 3222 if (HasWeight) { 3223 GetBranchWeights(SI, Weights); 3224 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3225 } 3226 3227 // Remove dead cases from the switch. 3228 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3229 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3230 assert(Case != SI->case_default() && 3231 "Case was not found. Probably mistake in DeadCases forming."); 3232 if (HasWeight) { 3233 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3234 Weights.pop_back(); 3235 } 3236 3237 // Prune unused values from PHI nodes. 3238 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3239 SI->removeCase(Case); 3240 } 3241 if (HasWeight) { 3242 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3243 SI->setMetadata(LLVMContext::MD_prof, 3244 MDBuilder(SI->getParent()->getContext()). 3245 createBranchWeights(MDWeights)); 3246 } 3247 3248 return !DeadCases.empty(); 3249} 3250 3251/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3252/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3253/// by an unconditional branch), look at the phi node for BB in the successor 3254/// block and see if the incoming value is equal to CaseValue. If so, return 3255/// the phi node, and set PhiIndex to BB's index in the phi node. 3256static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3257 BasicBlock *BB, 3258 int *PhiIndex) { 3259 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3260 return NULL; // BB must be empty to be a candidate for simplification. 3261 if (!BB->getSinglePredecessor()) 3262 return NULL; // BB must be dominated by the switch. 3263 3264 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3265 if (!Branch || !Branch->isUnconditional()) 3266 return NULL; // Terminator must be unconditional branch. 3267 3268 BasicBlock *Succ = Branch->getSuccessor(0); 3269 3270 BasicBlock::iterator I = Succ->begin(); 3271 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3272 int Idx = PHI->getBasicBlockIndex(BB); 3273 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3274 3275 Value *InValue = PHI->getIncomingValue(Idx); 3276 if (InValue != CaseValue) continue; 3277 3278 *PhiIndex = Idx; 3279 return PHI; 3280 } 3281 3282 return NULL; 3283} 3284 3285/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3286/// instruction to a phi node dominated by the switch, if that would mean that 3287/// some of the destination blocks of the switch can be folded away. 3288/// Returns true if a change is made. 3289static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3290 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3291 ForwardingNodesMap ForwardingNodes; 3292 3293 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3294 ConstantInt *CaseValue = I.getCaseValue(); 3295 BasicBlock *CaseDest = I.getCaseSuccessor(); 3296 3297 int PhiIndex; 3298 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3299 &PhiIndex); 3300 if (!PHI) continue; 3301 3302 ForwardingNodes[PHI].push_back(PhiIndex); 3303 } 3304 3305 bool Changed = false; 3306 3307 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3308 E = ForwardingNodes.end(); I != E; ++I) { 3309 PHINode *Phi = I->first; 3310 SmallVector<int,4> &Indexes = I->second; 3311 3312 if (Indexes.size() < 2) continue; 3313 3314 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3315 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3316 Changed = true; 3317 } 3318 3319 return Changed; 3320} 3321 3322/// ValidLookupTableConstant - Return true if the backend will be able to handle 3323/// initializing an array of constants like C. 3324static bool ValidLookupTableConstant(Constant *C) { 3325 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3326 return CE->isGEPWithNoNotionalOverIndexing(); 3327 3328 return isa<ConstantFP>(C) || 3329 isa<ConstantInt>(C) || 3330 isa<ConstantPointerNull>(C) || 3331 isa<GlobalValue>(C) || 3332 isa<UndefValue>(C); 3333} 3334 3335/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3336/// its constant value in ConstantPool, returning 0 if it's not there. 3337static Constant *LookupConstant(Value *V, 3338 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3339 if (Constant *C = dyn_cast<Constant>(V)) 3340 return C; 3341 return ConstantPool.lookup(V); 3342} 3343 3344/// ConstantFold - Try to fold instruction I into a constant. This works for 3345/// simple instructions such as binary operations where both operands are 3346/// constant or can be replaced by constants from the ConstantPool. Returns the 3347/// resulting constant on success, 0 otherwise. 3348static Constant *ConstantFold(Instruction *I, 3349 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3350 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3351 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3352 if (!A) 3353 return 0; 3354 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3355 if (!B) 3356 return 0; 3357 return ConstantExpr::get(BO->getOpcode(), A, B); 3358 } 3359 3360 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3361 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3362 if (!A) 3363 return 0; 3364 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3365 if (!B) 3366 return 0; 3367 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3368 } 3369 3370 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3371 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3372 if (!A) 3373 return 0; 3374 if (A->isAllOnesValue()) 3375 return LookupConstant(Select->getTrueValue(), ConstantPool); 3376 if (A->isNullValue()) 3377 return LookupConstant(Select->getFalseValue(), ConstantPool); 3378 return 0; 3379 } 3380 3381 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3382 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3383 if (!A) 3384 return 0; 3385 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3386 } 3387 3388 return 0; 3389} 3390 3391/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3392/// at the common destination basic block, *CommonDest, for one of the case 3393/// destionations CaseDest corresponding to value CaseVal (0 for the default 3394/// case), of a switch instruction SI. 3395static bool GetCaseResults(SwitchInst *SI, 3396 ConstantInt *CaseVal, 3397 BasicBlock *CaseDest, 3398 BasicBlock **CommonDest, 3399 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3400 // The block from which we enter the common destination. 3401 BasicBlock *Pred = SI->getParent(); 3402 3403 // If CaseDest is empty except for some side-effect free instructions through 3404 // which we can constant-propagate the CaseVal, continue to its successor. 3405 SmallDenseMap<Value*, Constant*> ConstantPool; 3406 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3407 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3408 ++I) { 3409 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3410 // If the terminator is a simple branch, continue to the next block. 3411 if (T->getNumSuccessors() != 1) 3412 return false; 3413 Pred = CaseDest; 3414 CaseDest = T->getSuccessor(0); 3415 } else if (isa<DbgInfoIntrinsic>(I)) { 3416 // Skip debug intrinsic. 3417 continue; 3418 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3419 // Instruction is side-effect free and constant. 3420 ConstantPool.insert(std::make_pair(I, C)); 3421 } else { 3422 break; 3423 } 3424 } 3425 3426 // If we did not have a CommonDest before, use the current one. 3427 if (!*CommonDest) 3428 *CommonDest = CaseDest; 3429 // If the destination isn't the common one, abort. 3430 if (CaseDest != *CommonDest) 3431 return false; 3432 3433 // Get the values for this case from phi nodes in the destination block. 3434 BasicBlock::iterator I = (*CommonDest)->begin(); 3435 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3436 int Idx = PHI->getBasicBlockIndex(Pred); 3437 if (Idx == -1) 3438 continue; 3439 3440 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3441 ConstantPool); 3442 if (!ConstVal) 3443 return false; 3444 3445 // Note: If the constant comes from constant-propagating the case value 3446 // through the CaseDest basic block, it will be safe to remove the 3447 // instructions in that block. They cannot be used (except in the phi nodes 3448 // we visit) outside CaseDest, because that block does not dominate its 3449 // successor. If it did, we would not be in this phi node. 3450 3451 // Be conservative about which kinds of constants we support. 3452 if (!ValidLookupTableConstant(ConstVal)) 3453 return false; 3454 3455 Res.push_back(std::make_pair(PHI, ConstVal)); 3456 } 3457 3458 return true; 3459} 3460 3461namespace { 3462 /// SwitchLookupTable - This class represents a lookup table that can be used 3463 /// to replace a switch. 3464 class SwitchLookupTable { 3465 public: 3466 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3467 /// with the contents of Values, using DefaultValue to fill any holes in the 3468 /// table. 3469 SwitchLookupTable(Module &M, 3470 uint64_t TableSize, 3471 ConstantInt *Offset, 3472 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3473 Constant *DefaultValue, 3474 const DataLayout *TD); 3475 3476 /// BuildLookup - Build instructions with Builder to retrieve the value at 3477 /// the position given by Index in the lookup table. 3478 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3479 3480 /// WouldFitInRegister - Return true if a table with TableSize elements of 3481 /// type ElementType would fit in a target-legal register. 3482 static bool WouldFitInRegister(const DataLayout *TD, 3483 uint64_t TableSize, 3484 const Type *ElementType); 3485 3486 private: 3487 // Depending on the contents of the table, it can be represented in 3488 // different ways. 3489 enum { 3490 // For tables where each element contains the same value, we just have to 3491 // store that single value and return it for each lookup. 3492 SingleValueKind, 3493 3494 // For small tables with integer elements, we can pack them into a bitmap 3495 // that fits into a target-legal register. Values are retrieved by 3496 // shift and mask operations. 3497 BitMapKind, 3498 3499 // The table is stored as an array of values. Values are retrieved by load 3500 // instructions from the table. 3501 ArrayKind 3502 } Kind; 3503 3504 // For SingleValueKind, this is the single value. 3505 Constant *SingleValue; 3506 3507 // For BitMapKind, this is the bitmap. 3508 ConstantInt *BitMap; 3509 IntegerType *BitMapElementTy; 3510 3511 // For ArrayKind, this is the array. 3512 GlobalVariable *Array; 3513 }; 3514} 3515 3516SwitchLookupTable::SwitchLookupTable(Module &M, 3517 uint64_t TableSize, 3518 ConstantInt *Offset, 3519 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3520 Constant *DefaultValue, 3521 const DataLayout *TD) 3522 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3523 assert(Values.size() && "Can't build lookup table without values!"); 3524 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3525 3526 // If all values in the table are equal, this is that value. 3527 SingleValue = Values.begin()->second; 3528 3529 // Build up the table contents. 3530 SmallVector<Constant*, 64> TableContents(TableSize); 3531 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3532 ConstantInt *CaseVal = Values[I].first; 3533 Constant *CaseRes = Values[I].second; 3534 assert(CaseRes->getType() == DefaultValue->getType()); 3535 3536 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3537 .getLimitedValue(); 3538 TableContents[Idx] = CaseRes; 3539 3540 if (CaseRes != SingleValue) 3541 SingleValue = 0; 3542 } 3543 3544 // Fill in any holes in the table with the default result. 3545 if (Values.size() < TableSize) { 3546 for (uint64_t I = 0; I < TableSize; ++I) { 3547 if (!TableContents[I]) 3548 TableContents[I] = DefaultValue; 3549 } 3550 3551 if (DefaultValue != SingleValue) 3552 SingleValue = 0; 3553 } 3554 3555 // If each element in the table contains the same value, we only need to store 3556 // that single value. 3557 if (SingleValue) { 3558 Kind = SingleValueKind; 3559 return; 3560 } 3561 3562 // If the type is integer and the table fits in a register, build a bitmap. 3563 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3564 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3565 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3566 for (uint64_t I = TableSize; I > 0; --I) { 3567 TableInt <<= IT->getBitWidth(); 3568 // Insert values into the bitmap. Undef values are set to zero. 3569 if (!isa<UndefValue>(TableContents[I - 1])) { 3570 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3571 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3572 } 3573 } 3574 BitMap = ConstantInt::get(M.getContext(), TableInt); 3575 BitMapElementTy = IT; 3576 Kind = BitMapKind; 3577 ++NumBitMaps; 3578 return; 3579 } 3580 3581 // Store the table in an array. 3582 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3583 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3584 3585 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3586 GlobalVariable::PrivateLinkage, 3587 Initializer, 3588 "switch.table"); 3589 Array->setUnnamedAddr(true); 3590 Kind = ArrayKind; 3591} 3592 3593Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3594 switch (Kind) { 3595 case SingleValueKind: 3596 return SingleValue; 3597 case BitMapKind: { 3598 // Type of the bitmap (e.g. i59). 3599 IntegerType *MapTy = BitMap->getType(); 3600 3601 // Cast Index to the same type as the bitmap. 3602 // Note: The Index is <= the number of elements in the table, so 3603 // truncating it to the width of the bitmask is safe. 3604 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3605 3606 // Multiply the shift amount by the element width. 3607 ShiftAmt = Builder.CreateMul(ShiftAmt, 3608 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3609 "switch.shiftamt"); 3610 3611 // Shift down. 3612 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3613 "switch.downshift"); 3614 // Mask off. 3615 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3616 "switch.masked"); 3617 } 3618 case ArrayKind: { 3619 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3620 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3621 "switch.gep"); 3622 return Builder.CreateLoad(GEP, "switch.load"); 3623 } 3624 } 3625 llvm_unreachable("Unknown lookup table kind!"); 3626} 3627 3628bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3629 uint64_t TableSize, 3630 const Type *ElementType) { 3631 if (!TD) 3632 return false; 3633 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3634 if (!IT) 3635 return false; 3636 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3637 // are <= 15, we could try to narrow the type. 3638 3639 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3640 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3641 return false; 3642 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3643} 3644 3645/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3646/// for this switch, based on the number of caes, size of the table and the 3647/// types of the results. 3648static bool ShouldBuildLookupTable(SwitchInst *SI, 3649 uint64_t TableSize, 3650 const TargetTransformInfo &TTI, 3651 const DataLayout *TD, 3652 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3653 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3654 return false; // TableSize overflowed, or mul below might overflow. 3655 3656 bool AllTablesFitInRegister = true; 3657 bool HasIllegalType = false; 3658 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3659 E = ResultTypes.end(); I != E; ++I) { 3660 Type *Ty = I->second; 3661 3662 // Saturate this flag to true. 3663 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3664 3665 // Saturate this flag to false. 3666 AllTablesFitInRegister = AllTablesFitInRegister && 3667 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3668 3669 // If both flags saturate, we're done. NOTE: This *only* works with 3670 // saturating flags, and all flags have to saturate first due to the 3671 // non-deterministic behavior of iterating over a dense map. 3672 if (HasIllegalType && !AllTablesFitInRegister) 3673 break; 3674 } 3675 3676 // If each table would fit in a register, we should build it anyway. 3677 if (AllTablesFitInRegister) 3678 return true; 3679 3680 // Don't build a table that doesn't fit in-register if it has illegal types. 3681 if (HasIllegalType) 3682 return false; 3683 3684 // The table density should be at least 40%. This is the same criterion as for 3685 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3686 // FIXME: Find the best cut-off. 3687 return SI->getNumCases() * 10 >= TableSize * 4; 3688} 3689 3690/// SwitchToLookupTable - If the switch is only used to initialize one or more 3691/// phi nodes in a common successor block with different constant values, 3692/// replace the switch with lookup tables. 3693static bool SwitchToLookupTable(SwitchInst *SI, 3694 IRBuilder<> &Builder, 3695 const TargetTransformInfo &TTI, 3696 const DataLayout* TD) { 3697 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3698 3699 // Only build lookup table when we have a target that supports it. 3700 if (!TTI.shouldBuildLookupTables()) 3701 return false; 3702 3703 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3704 // split off a dense part and build a lookup table for that. 3705 3706 // FIXME: This creates arrays of GEPs to constant strings, which means each 3707 // GEP needs a runtime relocation in PIC code. We should just build one big 3708 // string and lookup indices into that. 3709 3710 // Ignore the switch if the number of cases is too small. 3711 // This is similar to the check when building jump tables in 3712 // SelectionDAGBuilder::handleJTSwitchCase. 3713 // FIXME: Determine the best cut-off. 3714 if (SI->getNumCases() < 4) 3715 return false; 3716 3717 // Figure out the corresponding result for each case value and phi node in the 3718 // common destination, as well as the the min and max case values. 3719 assert(SI->case_begin() != SI->case_end()); 3720 SwitchInst::CaseIt CI = SI->case_begin(); 3721 ConstantInt *MinCaseVal = CI.getCaseValue(); 3722 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3723 3724 BasicBlock *CommonDest = 0; 3725 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3726 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3727 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3728 SmallDenseMap<PHINode*, Type*> ResultTypes; 3729 SmallVector<PHINode*, 4> PHIs; 3730 3731 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3732 ConstantInt *CaseVal = CI.getCaseValue(); 3733 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3734 MinCaseVal = CaseVal; 3735 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3736 MaxCaseVal = CaseVal; 3737 3738 // Resulting value at phi nodes for this case value. 3739 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3740 ResultsTy Results; 3741 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3742 Results)) 3743 return false; 3744 3745 // Append the result from this case to the list for each phi. 3746 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3747 if (!ResultLists.count(I->first)) 3748 PHIs.push_back(I->first); 3749 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3750 } 3751 } 3752 3753 // Get the resulting values for the default case. 3754 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3755 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3756 DefaultResultsList)) 3757 return false; 3758 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3759 PHINode *PHI = DefaultResultsList[I].first; 3760 Constant *Result = DefaultResultsList[I].second; 3761 DefaultResults[PHI] = Result; 3762 ResultTypes[PHI] = Result->getType(); 3763 } 3764 3765 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3766 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3767 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3768 return false; 3769 3770 // Create the BB that does the lookups. 3771 Module &Mod = *CommonDest->getParent()->getParent(); 3772 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3773 "switch.lookup", 3774 CommonDest->getParent(), 3775 CommonDest); 3776 3777 // Check whether the condition value is within the case range, and branch to 3778 // the new BB. 3779 Builder.SetInsertPoint(SI); 3780 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3781 "switch.tableidx"); 3782 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3783 MinCaseVal->getType(), TableSize)); 3784 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3785 3786 // Populate the BB that does the lookups. 3787 Builder.SetInsertPoint(LookupBB); 3788 bool ReturnedEarly = false; 3789 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3790 PHINode *PHI = PHIs[I]; 3791 3792 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3793 DefaultResults[PHI], TD); 3794 3795 Value *Result = Table.BuildLookup(TableIndex, Builder); 3796 3797 // If the result is used to return immediately from the function, we want to 3798 // do that right here. 3799 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3800 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3801 Builder.CreateRet(Result); 3802 ReturnedEarly = true; 3803 break; 3804 } 3805 3806 PHI->addIncoming(Result, LookupBB); 3807 } 3808 3809 if (!ReturnedEarly) 3810 Builder.CreateBr(CommonDest); 3811 3812 // Remove the switch. 3813 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3814 BasicBlock *Succ = SI->getSuccessor(i); 3815 if (Succ == SI->getDefaultDest()) continue; 3816 Succ->removePredecessor(SI->getParent()); 3817 } 3818 SI->eraseFromParent(); 3819 3820 ++NumLookupTables; 3821 return true; 3822} 3823 3824bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3825 BasicBlock *BB = SI->getParent(); 3826 3827 if (isValueEqualityComparison(SI)) { 3828 // If we only have one predecessor, and if it is a branch on this value, 3829 // see if that predecessor totally determines the outcome of this switch. 3830 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3831 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3832 return SimplifyCFG(BB, TTI, TD) | true; 3833 3834 Value *Cond = SI->getCondition(); 3835 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3836 if (SimplifySwitchOnSelect(SI, Select)) 3837 return SimplifyCFG(BB, TTI, TD) | true; 3838 3839 // If the block only contains the switch, see if we can fold the block 3840 // away into any preds. 3841 BasicBlock::iterator BBI = BB->begin(); 3842 // Ignore dbg intrinsics. 3843 while (isa<DbgInfoIntrinsic>(BBI)) 3844 ++BBI; 3845 if (SI == &*BBI) 3846 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3847 return SimplifyCFG(BB, TTI, TD) | true; 3848 } 3849 3850 // Try to transform the switch into an icmp and a branch. 3851 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3852 return SimplifyCFG(BB, TTI, TD) | true; 3853 3854 // Remove unreachable cases. 3855 if (EliminateDeadSwitchCases(SI)) 3856 return SimplifyCFG(BB, TTI, TD) | true; 3857 3858 if (ForwardSwitchConditionToPHI(SI)) 3859 return SimplifyCFG(BB, TTI, TD) | true; 3860 3861 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3862 return SimplifyCFG(BB, TTI, TD) | true; 3863 3864 return false; 3865} 3866 3867bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3868 BasicBlock *BB = IBI->getParent(); 3869 bool Changed = false; 3870 3871 // Eliminate redundant destinations. 3872 SmallPtrSet<Value *, 8> Succs; 3873 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3874 BasicBlock *Dest = IBI->getDestination(i); 3875 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3876 Dest->removePredecessor(BB); 3877 IBI->removeDestination(i); 3878 --i; --e; 3879 Changed = true; 3880 } 3881 } 3882 3883 if (IBI->getNumDestinations() == 0) { 3884 // If the indirectbr has no successors, change it to unreachable. 3885 new UnreachableInst(IBI->getContext(), IBI); 3886 EraseTerminatorInstAndDCECond(IBI); 3887 return true; 3888 } 3889 3890 if (IBI->getNumDestinations() == 1) { 3891 // If the indirectbr has one successor, change it to a direct branch. 3892 BranchInst::Create(IBI->getDestination(0), IBI); 3893 EraseTerminatorInstAndDCECond(IBI); 3894 return true; 3895 } 3896 3897 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3898 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3899 return SimplifyCFG(BB, TTI, TD) | true; 3900 } 3901 return Changed; 3902} 3903 3904bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3905 BasicBlock *BB = BI->getParent(); 3906 3907 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3908 return true; 3909 3910 // If the Terminator is the only non-phi instruction, simplify the block. 3911 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3912 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3913 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3914 return true; 3915 3916 // If the only instruction in the block is a seteq/setne comparison 3917 // against a constant, try to simplify the block. 3918 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3919 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3920 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3921 ; 3922 if (I->isTerminator() && 3923 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3924 return true; 3925 } 3926 3927 // If this basic block is ONLY a compare and a branch, and if a predecessor 3928 // branches to us and our successor, fold the comparison into the 3929 // predecessor and use logical operations to update the incoming value 3930 // for PHI nodes in common successor. 3931 if (FoldBranchToCommonDest(BI)) 3932 return SimplifyCFG(BB, TTI, TD) | true; 3933 return false; 3934} 3935 3936 3937bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3938 BasicBlock *BB = BI->getParent(); 3939 3940 // Conditional branch 3941 if (isValueEqualityComparison(BI)) { 3942 // If we only have one predecessor, and if it is a branch on this value, 3943 // see if that predecessor totally determines the outcome of this 3944 // switch. 3945 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3946 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3947 return SimplifyCFG(BB, TTI, TD) | true; 3948 3949 // This block must be empty, except for the setcond inst, if it exists. 3950 // Ignore dbg intrinsics. 3951 BasicBlock::iterator I = BB->begin(); 3952 // Ignore dbg intrinsics. 3953 while (isa<DbgInfoIntrinsic>(I)) 3954 ++I; 3955 if (&*I == BI) { 3956 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3957 return SimplifyCFG(BB, TTI, TD) | true; 3958 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3959 ++I; 3960 // Ignore dbg intrinsics. 3961 while (isa<DbgInfoIntrinsic>(I)) 3962 ++I; 3963 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3964 return SimplifyCFG(BB, TTI, TD) | true; 3965 } 3966 } 3967 3968 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3969 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3970 return true; 3971 3972 // If this basic block is ONLY a compare and a branch, and if a predecessor 3973 // branches to us and one of our successors, fold the comparison into the 3974 // predecessor and use logical operations to pick the right destination. 3975 if (FoldBranchToCommonDest(BI)) 3976 return SimplifyCFG(BB, TTI, TD) | true; 3977 3978 // We have a conditional branch to two blocks that are only reachable 3979 // from BI. We know that the condbr dominates the two blocks, so see if 3980 // there is any identical code in the "then" and "else" blocks. If so, we 3981 // can hoist it up to the branching block. 3982 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3983 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3984 if (HoistThenElseCodeToIf(BI)) 3985 return SimplifyCFG(BB, TTI, TD) | true; 3986 } else { 3987 // If Successor #1 has multiple preds, we may be able to conditionally 3988 // execute Successor #0 if it branches to successor #1. 3989 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3990 if (Succ0TI->getNumSuccessors() == 1 && 3991 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3992 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3993 return SimplifyCFG(BB, TTI, TD) | true; 3994 } 3995 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3996 // If Successor #0 has multiple preds, we may be able to conditionally 3997 // execute Successor #1 if it branches to successor #0. 3998 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3999 if (Succ1TI->getNumSuccessors() == 1 && 4000 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 4001 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 4002 return SimplifyCFG(BB, TTI, TD) | true; 4003 } 4004 4005 // If this is a branch on a phi node in the current block, thread control 4006 // through this block if any PHI node entries are constants. 4007 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 4008 if (PN->getParent() == BI->getParent()) 4009 if (FoldCondBranchOnPHI(BI, TD)) 4010 return SimplifyCFG(BB, TTI, TD) | true; 4011 4012 // Scan predecessor blocks for conditional branches. 4013 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 4014 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 4015 if (PBI != BI && PBI->isConditional()) 4016 if (SimplifyCondBranchToCondBranch(PBI, BI)) 4017 return SimplifyCFG(BB, TTI, TD) | true; 4018 4019 return false; 4020} 4021 4022/// Check if passing a value to an instruction will cause undefined behavior. 4023static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 4024 Constant *C = dyn_cast<Constant>(V); 4025 if (!C) 4026 return false; 4027 4028 if (I->use_empty()) 4029 return false; 4030 4031 if (C->isNullValue()) { 4032 // Only look at the first use, avoid hurting compile time with long uselists 4033 User *Use = *I->use_begin(); 4034 4035 // Now make sure that there are no instructions in between that can alter 4036 // control flow (eg. calls) 4037 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 4038 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 4039 return false; 4040 4041 // Look through GEPs. A load from a GEP derived from NULL is still undefined 4042 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 4043 if (GEP->getPointerOperand() == I) 4044 return passingValueIsAlwaysUndefined(V, GEP); 4045 4046 // Look through bitcasts. 4047 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 4048 return passingValueIsAlwaysUndefined(V, BC); 4049 4050 // Load from null is undefined. 4051 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 4052 if (!LI->isVolatile()) 4053 return LI->getPointerAddressSpace() == 0; 4054 4055 // Store to null is undefined. 4056 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 4057 if (!SI->isVolatile()) 4058 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 4059 } 4060 return false; 4061} 4062 4063/// If BB has an incoming value that will always trigger undefined behavior 4064/// (eg. null pointer dereference), remove the branch leading here. 4065static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 4066 for (BasicBlock::iterator i = BB->begin(); 4067 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 4068 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 4069 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 4070 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 4071 IRBuilder<> Builder(T); 4072 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 4073 BB->removePredecessor(PHI->getIncomingBlock(i)); 4074 // Turn uncoditional branches into unreachables and remove the dead 4075 // destination from conditional branches. 4076 if (BI->isUnconditional()) 4077 Builder.CreateUnreachable(); 4078 else 4079 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 4080 BI->getSuccessor(0)); 4081 BI->eraseFromParent(); 4082 return true; 4083 } 4084 // TODO: SwitchInst. 4085 } 4086 4087 return false; 4088} 4089 4090bool SimplifyCFGOpt::run(BasicBlock *BB) { 4091 bool Changed = false; 4092 4093 assert(BB && BB->getParent() && "Block not embedded in function!"); 4094 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4095 4096 // Remove basic blocks that have no predecessors (except the entry block)... 4097 // or that just have themself as a predecessor. These are unreachable. 4098 if ((pred_begin(BB) == pred_end(BB) && 4099 BB != &BB->getParent()->getEntryBlock()) || 4100 BB->getSinglePredecessor() == BB) { 4101 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4102 DeleteDeadBlock(BB); 4103 return true; 4104 } 4105 4106 // Check to see if we can constant propagate this terminator instruction 4107 // away... 4108 Changed |= ConstantFoldTerminator(BB, true); 4109 4110 // Check for and eliminate duplicate PHI nodes in this block. 4111 Changed |= EliminateDuplicatePHINodes(BB); 4112 4113 // Check for and remove branches that will always cause undefined behavior. 4114 Changed |= removeUndefIntroducingPredecessor(BB); 4115 4116 // Merge basic blocks into their predecessor if there is only one distinct 4117 // pred, and if there is only one distinct successor of the predecessor, and 4118 // if there are no PHI nodes. 4119 // 4120 if (MergeBlockIntoPredecessor(BB)) 4121 return true; 4122 4123 IRBuilder<> Builder(BB); 4124 4125 // If there is a trivial two-entry PHI node in this basic block, and we can 4126 // eliminate it, do so now. 4127 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4128 if (PN->getNumIncomingValues() == 2) 4129 Changed |= FoldTwoEntryPHINode(PN, TD); 4130 4131 Builder.SetInsertPoint(BB->getTerminator()); 4132 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4133 if (BI->isUnconditional()) { 4134 if (SimplifyUncondBranch(BI, Builder)) return true; 4135 } else { 4136 if (SimplifyCondBranch(BI, Builder)) return true; 4137 } 4138 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4139 if (SimplifyReturn(RI, Builder)) return true; 4140 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4141 if (SimplifyResume(RI, Builder)) return true; 4142 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4143 if (SimplifySwitch(SI, Builder)) return true; 4144 } else if (UnreachableInst *UI = 4145 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4146 if (SimplifyUnreachable(UI)) return true; 4147 } else if (IndirectBrInst *IBI = 4148 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4149 if (SimplifyIndirectBr(IBI)) return true; 4150 } 4151 4152 return Changed; 4153} 4154 4155/// SimplifyCFG - This function is used to do simplification of a CFG. For 4156/// example, it adjusts branches to branches to eliminate the extra hop, it 4157/// eliminates unreachable basic blocks, and does other "peephole" optimization 4158/// of the CFG. It returns true if a modification was made. 4159/// 4160bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4161 const DataLayout *TD) { 4162 return SimplifyCFGOpt(TTI, TD).run(BB); 4163} 4164