SimplifyCFG.cpp revision f7af1987ddcec768d8608d608ce187115510b21d
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 Value *StorePtr = StoreToHoist->getPointerOperand(); 1372 StoreInst *PreviousStore = 0; 1373 1374 // Look for a store to the same pointer in BrBB. 1375 unsigned MaxNumInstToLookAt = 10; 1376 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), 1377 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { 1378 Instruction *CurI = &*RI; 1379 1380 // Could be calling an instruction that effects memory like free(). 1381 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) 1382 return 0; 1383 1384 // Found the previous store make sure it stores to the same location. 1385 if ((PreviousStore = dyn_cast<StoreInst>(CurI))) { 1386 if (PreviousStore->getPointerOperand() == StorePtr) 1387 // Found the previous store, return its value operand. 1388 return PreviousStore->getValueOperand(); 1389 else 1390 return 0; // Unknown store. 1391 } 1392 } 1393 1394 return 0; 1395} 1396 1397/// \brief Speculate a conditional basic block flattening the CFG. 1398/// 1399/// Note that this is a very risky transform currently. Speculating 1400/// instructions like this is most often not desirable. Instead, there is an MI 1401/// pass which can do it with full awareness of the resource constraints. 1402/// However, some cases are "obvious" and we should do directly. An example of 1403/// this is speculating a single, reasonably cheap instruction. 1404/// 1405/// There is only one distinct advantage to flattening the CFG at the IR level: 1406/// it makes very common but simplistic optimizations such as are common in 1407/// instcombine and the DAG combiner more powerful by removing CFG edges and 1408/// modeling their effects with easier to reason about SSA value graphs. 1409/// 1410/// 1411/// An illustration of this transform is turning this IR: 1412/// \code 1413/// BB: 1414/// %cmp = icmp ult %x, %y 1415/// br i1 %cmp, label %EndBB, label %ThenBB 1416/// ThenBB: 1417/// %sub = sub %x, %y 1418/// br label BB2 1419/// EndBB: 1420/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1421/// ... 1422/// \endcode 1423/// 1424/// Into this IR: 1425/// \code 1426/// BB: 1427/// %cmp = icmp ult %x, %y 1428/// %sub = sub %x, %y 1429/// %cond = select i1 %cmp, 0, %sub 1430/// ... 1431/// \endcode 1432/// 1433/// \returns true if the conditional block is removed. 1434static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1435 // Be conservative for now. FP select instruction can often be expensive. 1436 Value *BrCond = BI->getCondition(); 1437 if (isa<FCmpInst>(BrCond)) 1438 return false; 1439 1440 BasicBlock *BB = BI->getParent(); 1441 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1442 1443 // If ThenBB is actually on the false edge of the conditional branch, remember 1444 // to swap the select operands later. 1445 bool Invert = false; 1446 if (ThenBB != BI->getSuccessor(0)) { 1447 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1448 Invert = true; 1449 } 1450 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1451 1452 // Keep a count of how many times instructions are used within CondBB when 1453 // they are candidates for sinking into CondBB. Specifically: 1454 // - They are defined in BB, and 1455 // - They have no side effects, and 1456 // - All of their uses are in CondBB. 1457 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1458 1459 unsigned SpeculationCost = 0; 1460 Value *SpeculatedStoreValue = 0; 1461 StoreInst *SpeculatedStore = 0; 1462 for (BasicBlock::iterator BBI = ThenBB->begin(), 1463 BBE = llvm::prior(ThenBB->end()); 1464 BBI != BBE; ++BBI) { 1465 Instruction *I = BBI; 1466 // Skip debug info. 1467 if (isa<DbgInfoIntrinsic>(I)) 1468 continue; 1469 1470 // Only speculatively execution a single instruction (not counting the 1471 // terminator) for now. 1472 ++SpeculationCost; 1473 if (SpeculationCost > 1) 1474 return false; 1475 1476 // Don't hoist the instruction if it's unsafe or expensive. 1477 if (!isSafeToSpeculativelyExecute(I) && 1478 !(HoistCondStores && 1479 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, 1480 EndBB)))) 1481 return false; 1482 if (!SpeculatedStoreValue && 1483 ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1484 return false; 1485 1486 // Store the store speculation candidate. 1487 if (SpeculatedStoreValue) 1488 SpeculatedStore = cast<StoreInst>(I); 1489 1490 // Do not hoist the instruction if any of its operands are defined but not 1491 // used in BB. The transformation will prevent the operand from 1492 // being sunk into the use block. 1493 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1494 i != e; ++i) { 1495 Instruction *OpI = dyn_cast<Instruction>(*i); 1496 if (!OpI || OpI->getParent() != BB || 1497 OpI->mayHaveSideEffects()) 1498 continue; // Not a candidate for sinking. 1499 1500 ++SinkCandidateUseCounts[OpI]; 1501 } 1502 } 1503 1504 // Consider any sink candidates which are only used in CondBB as costs for 1505 // speculation. Note, while we iterate over a DenseMap here, we are summing 1506 // and so iteration order isn't significant. 1507 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1508 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1509 I != E; ++I) 1510 if (I->first->getNumUses() == I->second) { 1511 ++SpeculationCost; 1512 if (SpeculationCost > 1) 1513 return false; 1514 } 1515 1516 // Check that the PHI nodes can be converted to selects. 1517 bool HaveRewritablePHIs = false; 1518 for (BasicBlock::iterator I = EndBB->begin(); 1519 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1520 Value *OrigV = PN->getIncomingValueForBlock(BB); 1521 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1522 1523 // Skip PHIs which are trivial. 1524 if (ThenV == OrigV) 1525 continue; 1526 1527 HaveRewritablePHIs = true; 1528 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV); 1529 if (!CE) 1530 continue; // Known safe and cheap. 1531 1532 if (!isSafeToSpeculativelyExecute(CE)) 1533 return false; 1534 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1535 return false; 1536 1537 // Account for the cost of an unfolded ConstantExpr which could end up 1538 // getting expanded into Instructions. 1539 // FIXME: This doesn't account for how many operations are combined in the 1540 // constant expression. 1541 ++SpeculationCost; 1542 if (SpeculationCost > 1) 1543 return false; 1544 } 1545 1546 // If there are no PHIs to process, bail early. This helps ensure idempotence 1547 // as well. 1548 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) 1549 return false; 1550 1551 // If we get here, we can hoist the instruction and if-convert. 1552 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1553 1554 // Insert a select of the value of the speculated store. 1555 if (SpeculatedStoreValue) { 1556 IRBuilder<true, NoFolder> Builder(BI); 1557 Value *TrueV = SpeculatedStore->getValueOperand(); 1558 Value *FalseV = SpeculatedStoreValue; 1559 if (Invert) 1560 std::swap(TrueV, FalseV); 1561 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + 1562 "." + FalseV->getName()); 1563 SpeculatedStore->setOperand(0, S); 1564 } 1565 1566 // Hoist the instructions. 1567 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1568 llvm::prior(ThenBB->end())); 1569 1570 // Insert selects and rewrite the PHI operands. 1571 IRBuilder<true, NoFolder> Builder(BI); 1572 for (BasicBlock::iterator I = EndBB->begin(); 1573 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1574 unsigned OrigI = PN->getBasicBlockIndex(BB); 1575 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1576 Value *OrigV = PN->getIncomingValue(OrigI); 1577 Value *ThenV = PN->getIncomingValue(ThenI); 1578 1579 // Skip PHIs which are trivial. 1580 if (OrigV == ThenV) 1581 continue; 1582 1583 // Create a select whose true value is the speculatively executed value and 1584 // false value is the preexisting value. Swap them if the branch 1585 // destinations were inverted. 1586 Value *TrueV = ThenV, *FalseV = OrigV; 1587 if (Invert) 1588 std::swap(TrueV, FalseV); 1589 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1590 TrueV->getName() + "." + FalseV->getName()); 1591 PN->setIncomingValue(OrigI, V); 1592 PN->setIncomingValue(ThenI, V); 1593 } 1594 1595 ++NumSpeculations; 1596 return true; 1597} 1598 1599/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1600/// across this block. 1601static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1602 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1603 unsigned Size = 0; 1604 1605 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1606 if (isa<DbgInfoIntrinsic>(BBI)) 1607 continue; 1608 if (Size > 10) return false; // Don't clone large BB's. 1609 ++Size; 1610 1611 // We can only support instructions that do not define values that are 1612 // live outside of the current basic block. 1613 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1614 UI != E; ++UI) { 1615 Instruction *U = cast<Instruction>(*UI); 1616 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1617 } 1618 1619 // Looks ok, continue checking. 1620 } 1621 1622 return true; 1623} 1624 1625/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1626/// that is defined in the same block as the branch and if any PHI entries are 1627/// constants, thread edges corresponding to that entry to be branches to their 1628/// ultimate destination. 1629static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1630 BasicBlock *BB = BI->getParent(); 1631 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1632 // NOTE: we currently cannot transform this case if the PHI node is used 1633 // outside of the block. 1634 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1635 return false; 1636 1637 // Degenerate case of a single entry PHI. 1638 if (PN->getNumIncomingValues() == 1) { 1639 FoldSingleEntryPHINodes(PN->getParent()); 1640 return true; 1641 } 1642 1643 // Now we know that this block has multiple preds and two succs. 1644 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1645 1646 // Okay, this is a simple enough basic block. See if any phi values are 1647 // constants. 1648 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1649 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1650 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1651 1652 // Okay, we now know that all edges from PredBB should be revectored to 1653 // branch to RealDest. 1654 BasicBlock *PredBB = PN->getIncomingBlock(i); 1655 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1656 1657 if (RealDest == BB) continue; // Skip self loops. 1658 // Skip if the predecessor's terminator is an indirect branch. 1659 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1660 1661 // The dest block might have PHI nodes, other predecessors and other 1662 // difficult cases. Instead of being smart about this, just insert a new 1663 // block that jumps to the destination block, effectively splitting 1664 // the edge we are about to create. 1665 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1666 RealDest->getName()+".critedge", 1667 RealDest->getParent(), RealDest); 1668 BranchInst::Create(RealDest, EdgeBB); 1669 1670 // Update PHI nodes. 1671 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1672 1673 // BB may have instructions that are being threaded over. Clone these 1674 // instructions into EdgeBB. We know that there will be no uses of the 1675 // cloned instructions outside of EdgeBB. 1676 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1677 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1678 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1679 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1680 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1681 continue; 1682 } 1683 // Clone the instruction. 1684 Instruction *N = BBI->clone(); 1685 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1686 1687 // Update operands due to translation. 1688 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1689 i != e; ++i) { 1690 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1691 if (PI != TranslateMap.end()) 1692 *i = PI->second; 1693 } 1694 1695 // Check for trivial simplification. 1696 if (Value *V = SimplifyInstruction(N, TD)) { 1697 TranslateMap[BBI] = V; 1698 delete N; // Instruction folded away, don't need actual inst 1699 } else { 1700 // Insert the new instruction into its new home. 1701 EdgeBB->getInstList().insert(InsertPt, N); 1702 if (!BBI->use_empty()) 1703 TranslateMap[BBI] = N; 1704 } 1705 } 1706 1707 // Loop over all of the edges from PredBB to BB, changing them to branch 1708 // to EdgeBB instead. 1709 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1710 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1711 if (PredBBTI->getSuccessor(i) == BB) { 1712 BB->removePredecessor(PredBB); 1713 PredBBTI->setSuccessor(i, EdgeBB); 1714 } 1715 1716 // Recurse, simplifying any other constants. 1717 return FoldCondBranchOnPHI(BI, TD) | true; 1718 } 1719 1720 return false; 1721} 1722 1723/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1724/// PHI node, see if we can eliminate it. 1725static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1726 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1727 // statement", which has a very simple dominance structure. Basically, we 1728 // are trying to find the condition that is being branched on, which 1729 // subsequently causes this merge to happen. We really want control 1730 // dependence information for this check, but simplifycfg can't keep it up 1731 // to date, and this catches most of the cases we care about anyway. 1732 BasicBlock *BB = PN->getParent(); 1733 BasicBlock *IfTrue, *IfFalse; 1734 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1735 if (!IfCond || 1736 // Don't bother if the branch will be constant folded trivially. 1737 isa<ConstantInt>(IfCond)) 1738 return false; 1739 1740 // Okay, we found that we can merge this two-entry phi node into a select. 1741 // Doing so would require us to fold *all* two entry phi nodes in this block. 1742 // At some point this becomes non-profitable (particularly if the target 1743 // doesn't support cmov's). Only do this transformation if there are two or 1744 // fewer PHI nodes in this block. 1745 unsigned NumPhis = 0; 1746 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1747 if (NumPhis > 2) 1748 return false; 1749 1750 // Loop over the PHI's seeing if we can promote them all to select 1751 // instructions. While we are at it, keep track of the instructions 1752 // that need to be moved to the dominating block. 1753 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1754 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1755 MaxCostVal1 = PHINodeFoldingThreshold; 1756 1757 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1758 PHINode *PN = cast<PHINode>(II++); 1759 if (Value *V = SimplifyInstruction(PN, TD)) { 1760 PN->replaceAllUsesWith(V); 1761 PN->eraseFromParent(); 1762 continue; 1763 } 1764 1765 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1766 MaxCostVal0) || 1767 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1768 MaxCostVal1)) 1769 return false; 1770 } 1771 1772 // If we folded the first phi, PN dangles at this point. Refresh it. If 1773 // we ran out of PHIs then we simplified them all. 1774 PN = dyn_cast<PHINode>(BB->begin()); 1775 if (PN == 0) return true; 1776 1777 // Don't fold i1 branches on PHIs which contain binary operators. These can 1778 // often be turned into switches and other things. 1779 if (PN->getType()->isIntegerTy(1) && 1780 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1781 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1782 isa<BinaryOperator>(IfCond))) 1783 return false; 1784 1785 // If we all PHI nodes are promotable, check to make sure that all 1786 // instructions in the predecessor blocks can be promoted as well. If 1787 // not, we won't be able to get rid of the control flow, so it's not 1788 // worth promoting to select instructions. 1789 BasicBlock *DomBlock = 0; 1790 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1791 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1792 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1793 IfBlock1 = 0; 1794 } else { 1795 DomBlock = *pred_begin(IfBlock1); 1796 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1797 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1798 // This is not an aggressive instruction that we can promote. 1799 // Because of this, we won't be able to get rid of the control 1800 // flow, so the xform is not worth it. 1801 return false; 1802 } 1803 } 1804 1805 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1806 IfBlock2 = 0; 1807 } else { 1808 DomBlock = *pred_begin(IfBlock2); 1809 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1810 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1811 // This is not an aggressive instruction that we can promote. 1812 // Because of this, we won't be able to get rid of the control 1813 // flow, so the xform is not worth it. 1814 return false; 1815 } 1816 } 1817 1818 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1819 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1820 1821 // If we can still promote the PHI nodes after this gauntlet of tests, 1822 // do all of the PHI's now. 1823 Instruction *InsertPt = DomBlock->getTerminator(); 1824 IRBuilder<true, NoFolder> Builder(InsertPt); 1825 1826 // Move all 'aggressive' instructions, which are defined in the 1827 // conditional parts of the if's up to the dominating block. 1828 if (IfBlock1) 1829 DomBlock->getInstList().splice(InsertPt, 1830 IfBlock1->getInstList(), IfBlock1->begin(), 1831 IfBlock1->getTerminator()); 1832 if (IfBlock2) 1833 DomBlock->getInstList().splice(InsertPt, 1834 IfBlock2->getInstList(), IfBlock2->begin(), 1835 IfBlock2->getTerminator()); 1836 1837 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1838 // Change the PHI node into a select instruction. 1839 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1840 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1841 1842 SelectInst *NV = 1843 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1844 PN->replaceAllUsesWith(NV); 1845 NV->takeName(PN); 1846 PN->eraseFromParent(); 1847 } 1848 1849 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1850 // has been flattened. Change DomBlock to jump directly to our new block to 1851 // avoid other simplifycfg's kicking in on the diamond. 1852 TerminatorInst *OldTI = DomBlock->getTerminator(); 1853 Builder.SetInsertPoint(OldTI); 1854 Builder.CreateBr(BB); 1855 OldTI->eraseFromParent(); 1856 return true; 1857} 1858 1859/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1860/// to two returning blocks, try to merge them together into one return, 1861/// introducing a select if the return values disagree. 1862static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1863 IRBuilder<> &Builder) { 1864 assert(BI->isConditional() && "Must be a conditional branch"); 1865 BasicBlock *TrueSucc = BI->getSuccessor(0); 1866 BasicBlock *FalseSucc = BI->getSuccessor(1); 1867 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1868 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1869 1870 // Check to ensure both blocks are empty (just a return) or optionally empty 1871 // with PHI nodes. If there are other instructions, merging would cause extra 1872 // computation on one path or the other. 1873 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1874 return false; 1875 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1876 return false; 1877 1878 Builder.SetInsertPoint(BI); 1879 // Okay, we found a branch that is going to two return nodes. If 1880 // there is no return value for this function, just change the 1881 // branch into a return. 1882 if (FalseRet->getNumOperands() == 0) { 1883 TrueSucc->removePredecessor(BI->getParent()); 1884 FalseSucc->removePredecessor(BI->getParent()); 1885 Builder.CreateRetVoid(); 1886 EraseTerminatorInstAndDCECond(BI); 1887 return true; 1888 } 1889 1890 // Otherwise, figure out what the true and false return values are 1891 // so we can insert a new select instruction. 1892 Value *TrueValue = TrueRet->getReturnValue(); 1893 Value *FalseValue = FalseRet->getReturnValue(); 1894 1895 // Unwrap any PHI nodes in the return blocks. 1896 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1897 if (TVPN->getParent() == TrueSucc) 1898 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1899 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1900 if (FVPN->getParent() == FalseSucc) 1901 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1902 1903 // In order for this transformation to be safe, we must be able to 1904 // unconditionally execute both operands to the return. This is 1905 // normally the case, but we could have a potentially-trapping 1906 // constant expression that prevents this transformation from being 1907 // safe. 1908 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1909 if (TCV->canTrap()) 1910 return false; 1911 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1912 if (FCV->canTrap()) 1913 return false; 1914 1915 // Okay, we collected all the mapped values and checked them for sanity, and 1916 // defined to really do this transformation. First, update the CFG. 1917 TrueSucc->removePredecessor(BI->getParent()); 1918 FalseSucc->removePredecessor(BI->getParent()); 1919 1920 // Insert select instructions where needed. 1921 Value *BrCond = BI->getCondition(); 1922 if (TrueValue) { 1923 // Insert a select if the results differ. 1924 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1925 } else if (isa<UndefValue>(TrueValue)) { 1926 TrueValue = FalseValue; 1927 } else { 1928 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1929 FalseValue, "retval"); 1930 } 1931 } 1932 1933 Value *RI = !TrueValue ? 1934 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1935 1936 (void) RI; 1937 1938 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1939 << "\n " << *BI << "NewRet = " << *RI 1940 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1941 1942 EraseTerminatorInstAndDCECond(BI); 1943 1944 return true; 1945} 1946 1947/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1948/// probabilities of the branch taking each edge. Fills in the two APInt 1949/// parameters and return true, or returns false if no or invalid metadata was 1950/// found. 1951static bool ExtractBranchMetadata(BranchInst *BI, 1952 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1953 assert(BI->isConditional() && 1954 "Looking for probabilities on unconditional branch?"); 1955 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1956 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1957 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1958 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1959 if (!CITrue || !CIFalse) return false; 1960 ProbTrue = CITrue->getValue().getZExtValue(); 1961 ProbFalse = CIFalse->getValue().getZExtValue(); 1962 return true; 1963} 1964 1965/// checkCSEInPredecessor - Return true if the given instruction is available 1966/// in its predecessor block. If yes, the instruction will be removed. 1967/// 1968static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1969 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1970 return false; 1971 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1972 Instruction *PBI = &*I; 1973 // Check whether Inst and PBI generate the same value. 1974 if (Inst->isIdenticalTo(PBI)) { 1975 Inst->replaceAllUsesWith(PBI); 1976 Inst->eraseFromParent(); 1977 return true; 1978 } 1979 } 1980 return false; 1981} 1982 1983/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1984/// predecessor branches to us and one of our successors, fold the block into 1985/// the predecessor and use logical operations to pick the right destination. 1986bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1987 BasicBlock *BB = BI->getParent(); 1988 1989 Instruction *Cond = 0; 1990 if (BI->isConditional()) 1991 Cond = dyn_cast<Instruction>(BI->getCondition()); 1992 else { 1993 // For unconditional branch, check for a simple CFG pattern, where 1994 // BB has a single predecessor and BB's successor is also its predecessor's 1995 // successor. If such pattern exisits, check for CSE between BB and its 1996 // predecessor. 1997 if (BasicBlock *PB = BB->getSinglePredecessor()) 1998 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1999 if (PBI->isConditional() && 2000 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 2001 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 2002 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 2003 I != E; ) { 2004 Instruction *Curr = I++; 2005 if (isa<CmpInst>(Curr)) { 2006 Cond = Curr; 2007 break; 2008 } 2009 // Quit if we can't remove this instruction. 2010 if (!checkCSEInPredecessor(Curr, PB)) 2011 return false; 2012 } 2013 } 2014 2015 if (Cond == 0) 2016 return false; 2017 } 2018 2019 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 2020 Cond->getParent() != BB || !Cond->hasOneUse()) 2021 return false; 2022 2023 // Only allow this if the condition is a simple instruction that can be 2024 // executed unconditionally. It must be in the same block as the branch, and 2025 // must be at the front of the block. 2026 BasicBlock::iterator FrontIt = BB->front(); 2027 2028 // Ignore dbg intrinsics. 2029 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2030 2031 // Allow a single instruction to be hoisted in addition to the compare 2032 // that feeds the branch. We later ensure that any values that _it_ uses 2033 // were also live in the predecessor, so that we don't unnecessarily create 2034 // register pressure or inhibit out-of-order execution. 2035 Instruction *BonusInst = 0; 2036 if (&*FrontIt != Cond && 2037 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 2038 isSafeToSpeculativelyExecute(FrontIt)) { 2039 BonusInst = &*FrontIt; 2040 ++FrontIt; 2041 2042 // Ignore dbg intrinsics. 2043 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2044 } 2045 2046 // Only a single bonus inst is allowed. 2047 if (&*FrontIt != Cond) 2048 return false; 2049 2050 // Make sure the instruction after the condition is the cond branch. 2051 BasicBlock::iterator CondIt = Cond; ++CondIt; 2052 2053 // Ingore dbg intrinsics. 2054 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 2055 2056 if (&*CondIt != BI) 2057 return false; 2058 2059 // Cond is known to be a compare or binary operator. Check to make sure that 2060 // neither operand is a potentially-trapping constant expression. 2061 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 2062 if (CE->canTrap()) 2063 return false; 2064 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 2065 if (CE->canTrap()) 2066 return false; 2067 2068 // Finally, don't infinitely unroll conditional loops. 2069 BasicBlock *TrueDest = BI->getSuccessor(0); 2070 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 2071 if (TrueDest == BB || FalseDest == BB) 2072 return false; 2073 2074 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2075 BasicBlock *PredBlock = *PI; 2076 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 2077 2078 // Check that we have two conditional branches. If there is a PHI node in 2079 // the common successor, verify that the same value flows in from both 2080 // blocks. 2081 SmallVector<PHINode*, 4> PHIs; 2082 if (PBI == 0 || PBI->isUnconditional() || 2083 (BI->isConditional() && 2084 !SafeToMergeTerminators(BI, PBI)) || 2085 (!BI->isConditional() && 2086 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2087 continue; 2088 2089 // Determine if the two branches share a common destination. 2090 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2091 bool InvertPredCond = false; 2092 2093 if (BI->isConditional()) { 2094 if (PBI->getSuccessor(0) == TrueDest) 2095 Opc = Instruction::Or; 2096 else if (PBI->getSuccessor(1) == FalseDest) 2097 Opc = Instruction::And; 2098 else if (PBI->getSuccessor(0) == FalseDest) 2099 Opc = Instruction::And, InvertPredCond = true; 2100 else if (PBI->getSuccessor(1) == TrueDest) 2101 Opc = Instruction::Or, InvertPredCond = true; 2102 else 2103 continue; 2104 } else { 2105 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2106 continue; 2107 } 2108 2109 // Ensure that any values used in the bonus instruction are also used 2110 // by the terminator of the predecessor. This means that those values 2111 // must already have been resolved, so we won't be inhibiting the 2112 // out-of-order core by speculating them earlier. 2113 if (BonusInst) { 2114 // Collect the values used by the bonus inst 2115 SmallPtrSet<Value*, 4> UsedValues; 2116 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2117 OE = BonusInst->op_end(); OI != OE; ++OI) { 2118 Value *V = *OI; 2119 if (!isa<Constant>(V)) 2120 UsedValues.insert(V); 2121 } 2122 2123 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2124 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2125 2126 // Walk up to four levels back up the use-def chain of the predecessor's 2127 // terminator to see if all those values were used. The choice of four 2128 // levels is arbitrary, to provide a compile-time-cost bound. 2129 while (!Worklist.empty()) { 2130 std::pair<Value*, unsigned> Pair = Worklist.back(); 2131 Worklist.pop_back(); 2132 2133 if (Pair.second >= 4) continue; 2134 UsedValues.erase(Pair.first); 2135 if (UsedValues.empty()) break; 2136 2137 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2138 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2139 OI != OE; ++OI) 2140 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2141 } 2142 } 2143 2144 if (!UsedValues.empty()) return false; 2145 } 2146 2147 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2148 IRBuilder<> Builder(PBI); 2149 2150 // If we need to invert the condition in the pred block to match, do so now. 2151 if (InvertPredCond) { 2152 Value *NewCond = PBI->getCondition(); 2153 2154 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2155 CmpInst *CI = cast<CmpInst>(NewCond); 2156 CI->setPredicate(CI->getInversePredicate()); 2157 } else { 2158 NewCond = Builder.CreateNot(NewCond, 2159 PBI->getCondition()->getName()+".not"); 2160 } 2161 2162 PBI->setCondition(NewCond); 2163 PBI->swapSuccessors(); 2164 } 2165 2166 // If we have a bonus inst, clone it into the predecessor block. 2167 Instruction *NewBonus = 0; 2168 if (BonusInst) { 2169 NewBonus = BonusInst->clone(); 2170 PredBlock->getInstList().insert(PBI, NewBonus); 2171 NewBonus->takeName(BonusInst); 2172 BonusInst->setName(BonusInst->getName()+".old"); 2173 } 2174 2175 // Clone Cond into the predecessor basic block, and or/and the 2176 // two conditions together. 2177 Instruction *New = Cond->clone(); 2178 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2179 PredBlock->getInstList().insert(PBI, New); 2180 New->takeName(Cond); 2181 Cond->setName(New->getName()+".old"); 2182 2183 if (BI->isConditional()) { 2184 Instruction *NewCond = 2185 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2186 New, "or.cond")); 2187 PBI->setCondition(NewCond); 2188 2189 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2190 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2191 PredFalseWeight); 2192 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2193 SuccFalseWeight); 2194 SmallVector<uint64_t, 8> NewWeights; 2195 2196 if (PBI->getSuccessor(0) == BB) { 2197 if (PredHasWeights && SuccHasWeights) { 2198 // PBI: br i1 %x, BB, FalseDest 2199 // BI: br i1 %y, TrueDest, FalseDest 2200 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2201 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2202 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2203 // TrueWeight for PBI * FalseWeight for BI. 2204 // We assume that total weights of a BranchInst can fit into 32 bits. 2205 // Therefore, we will not have overflow using 64-bit arithmetic. 2206 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2207 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2208 } 2209 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2210 PBI->setSuccessor(0, TrueDest); 2211 } 2212 if (PBI->getSuccessor(1) == BB) { 2213 if (PredHasWeights && SuccHasWeights) { 2214 // PBI: br i1 %x, TrueDest, BB 2215 // BI: br i1 %y, TrueDest, FalseDest 2216 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2217 // FalseWeight for PBI * TrueWeight for BI. 2218 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2219 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2220 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2221 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2222 } 2223 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2224 PBI->setSuccessor(1, FalseDest); 2225 } 2226 if (NewWeights.size() == 2) { 2227 // Halve the weights if any of them cannot fit in an uint32_t 2228 FitWeights(NewWeights); 2229 2230 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2231 PBI->setMetadata(LLVMContext::MD_prof, 2232 MDBuilder(BI->getContext()). 2233 createBranchWeights(MDWeights)); 2234 } else 2235 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2236 } else { 2237 // Update PHI nodes in the common successors. 2238 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2239 ConstantInt *PBI_C = cast<ConstantInt>( 2240 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2241 assert(PBI_C->getType()->isIntegerTy(1)); 2242 Instruction *MergedCond = 0; 2243 if (PBI->getSuccessor(0) == TrueDest) { 2244 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2245 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2246 // is false: !PBI_Cond and BI_Value 2247 Instruction *NotCond = 2248 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2249 "not.cond")); 2250 MergedCond = 2251 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2252 NotCond, New, 2253 "and.cond")); 2254 if (PBI_C->isOne()) 2255 MergedCond = 2256 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2257 PBI->getCondition(), MergedCond, 2258 "or.cond")); 2259 } else { 2260 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2261 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2262 // is false: PBI_Cond and BI_Value 2263 MergedCond = 2264 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2265 PBI->getCondition(), New, 2266 "and.cond")); 2267 if (PBI_C->isOne()) { 2268 Instruction *NotCond = 2269 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2270 "not.cond")); 2271 MergedCond = 2272 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2273 NotCond, MergedCond, 2274 "or.cond")); 2275 } 2276 } 2277 // Update PHI Node. 2278 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2279 MergedCond); 2280 } 2281 // Change PBI from Conditional to Unconditional. 2282 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2283 EraseTerminatorInstAndDCECond(PBI); 2284 PBI = New_PBI; 2285 } 2286 2287 // TODO: If BB is reachable from all paths through PredBlock, then we 2288 // could replace PBI's branch probabilities with BI's. 2289 2290 // Copy any debug value intrinsics into the end of PredBlock. 2291 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2292 if (isa<DbgInfoIntrinsic>(*I)) 2293 I->clone()->insertBefore(PBI); 2294 2295 return true; 2296 } 2297 return false; 2298} 2299 2300/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2301/// predecessor of another block, this function tries to simplify it. We know 2302/// that PBI and BI are both conditional branches, and BI is in one of the 2303/// successor blocks of PBI - PBI branches to BI. 2304static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2305 assert(PBI->isConditional() && BI->isConditional()); 2306 BasicBlock *BB = BI->getParent(); 2307 2308 // If this block ends with a branch instruction, and if there is a 2309 // predecessor that ends on a branch of the same condition, make 2310 // this conditional branch redundant. 2311 if (PBI->getCondition() == BI->getCondition() && 2312 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2313 // Okay, the outcome of this conditional branch is statically 2314 // knowable. If this block had a single pred, handle specially. 2315 if (BB->getSinglePredecessor()) { 2316 // Turn this into a branch on constant. 2317 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2318 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2319 CondIsTrue)); 2320 return true; // Nuke the branch on constant. 2321 } 2322 2323 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2324 // in the constant and simplify the block result. Subsequent passes of 2325 // simplifycfg will thread the block. 2326 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2327 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2328 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2329 std::distance(PB, PE), 2330 BI->getCondition()->getName() + ".pr", 2331 BB->begin()); 2332 // Okay, we're going to insert the PHI node. Since PBI is not the only 2333 // predecessor, compute the PHI'd conditional value for all of the preds. 2334 // Any predecessor where the condition is not computable we keep symbolic. 2335 for (pred_iterator PI = PB; PI != PE; ++PI) { 2336 BasicBlock *P = *PI; 2337 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2338 PBI != BI && PBI->isConditional() && 2339 PBI->getCondition() == BI->getCondition() && 2340 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2341 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2342 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2343 CondIsTrue), P); 2344 } else { 2345 NewPN->addIncoming(BI->getCondition(), P); 2346 } 2347 } 2348 2349 BI->setCondition(NewPN); 2350 return true; 2351 } 2352 } 2353 2354 // If this is a conditional branch in an empty block, and if any 2355 // predecessors is a conditional branch to one of our destinations, 2356 // fold the conditions into logical ops and one cond br. 2357 BasicBlock::iterator BBI = BB->begin(); 2358 // Ignore dbg intrinsics. 2359 while (isa<DbgInfoIntrinsic>(BBI)) 2360 ++BBI; 2361 if (&*BBI != BI) 2362 return false; 2363 2364 2365 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2366 if (CE->canTrap()) 2367 return false; 2368 2369 int PBIOp, BIOp; 2370 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2371 PBIOp = BIOp = 0; 2372 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2373 PBIOp = 0, BIOp = 1; 2374 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2375 PBIOp = 1, BIOp = 0; 2376 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2377 PBIOp = BIOp = 1; 2378 else 2379 return false; 2380 2381 // Check to make sure that the other destination of this branch 2382 // isn't BB itself. If so, this is an infinite loop that will 2383 // keep getting unwound. 2384 if (PBI->getSuccessor(PBIOp) == BB) 2385 return false; 2386 2387 // Do not perform this transformation if it would require 2388 // insertion of a large number of select instructions. For targets 2389 // without predication/cmovs, this is a big pessimization. 2390 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2391 2392 unsigned NumPhis = 0; 2393 for (BasicBlock::iterator II = CommonDest->begin(); 2394 isa<PHINode>(II); ++II, ++NumPhis) 2395 if (NumPhis > 2) // Disable this xform. 2396 return false; 2397 2398 // Finally, if everything is ok, fold the branches to logical ops. 2399 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2400 2401 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2402 << "AND: " << *BI->getParent()); 2403 2404 2405 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2406 // branch in it, where one edge (OtherDest) goes back to itself but the other 2407 // exits. We don't *know* that the program avoids the infinite loop 2408 // (even though that seems likely). If we do this xform naively, we'll end up 2409 // recursively unpeeling the loop. Since we know that (after the xform is 2410 // done) that the block *is* infinite if reached, we just make it an obviously 2411 // infinite loop with no cond branch. 2412 if (OtherDest == BB) { 2413 // Insert it at the end of the function, because it's either code, 2414 // or it won't matter if it's hot. :) 2415 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2416 "infloop", BB->getParent()); 2417 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2418 OtherDest = InfLoopBlock; 2419 } 2420 2421 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2422 2423 // BI may have other predecessors. Because of this, we leave 2424 // it alone, but modify PBI. 2425 2426 // Make sure we get to CommonDest on True&True directions. 2427 Value *PBICond = PBI->getCondition(); 2428 IRBuilder<true, NoFolder> Builder(PBI); 2429 if (PBIOp) 2430 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2431 2432 Value *BICond = BI->getCondition(); 2433 if (BIOp) 2434 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2435 2436 // Merge the conditions. 2437 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2438 2439 // Modify PBI to branch on the new condition to the new dests. 2440 PBI->setCondition(Cond); 2441 PBI->setSuccessor(0, CommonDest); 2442 PBI->setSuccessor(1, OtherDest); 2443 2444 // Update branch weight for PBI. 2445 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2446 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2447 PredFalseWeight); 2448 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2449 SuccFalseWeight); 2450 if (PredHasWeights && SuccHasWeights) { 2451 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2452 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2453 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2454 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2455 // The weight to CommonDest should be PredCommon * SuccTotal + 2456 // PredOther * SuccCommon. 2457 // The weight to OtherDest should be PredOther * SuccOther. 2458 SmallVector<uint64_t, 2> NewWeights; 2459 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2460 PredOther * SuccCommon); 2461 NewWeights.push_back(PredOther * SuccOther); 2462 // Halve the weights if any of them cannot fit in an uint32_t 2463 FitWeights(NewWeights); 2464 2465 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2466 PBI->setMetadata(LLVMContext::MD_prof, 2467 MDBuilder(BI->getContext()). 2468 createBranchWeights(MDWeights)); 2469 } 2470 2471 // OtherDest may have phi nodes. If so, add an entry from PBI's 2472 // block that are identical to the entries for BI's block. 2473 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2474 2475 // We know that the CommonDest already had an edge from PBI to 2476 // it. If it has PHIs though, the PHIs may have different 2477 // entries for BB and PBI's BB. If so, insert a select to make 2478 // them agree. 2479 PHINode *PN; 2480 for (BasicBlock::iterator II = CommonDest->begin(); 2481 (PN = dyn_cast<PHINode>(II)); ++II) { 2482 Value *BIV = PN->getIncomingValueForBlock(BB); 2483 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2484 Value *PBIV = PN->getIncomingValue(PBBIdx); 2485 if (BIV != PBIV) { 2486 // Insert a select in PBI to pick the right value. 2487 Value *NV = cast<SelectInst> 2488 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2489 PN->setIncomingValue(PBBIdx, NV); 2490 } 2491 } 2492 2493 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2494 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2495 2496 // This basic block is probably dead. We know it has at least 2497 // one fewer predecessor. 2498 return true; 2499} 2500 2501// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2502// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2503// Takes care of updating the successors and removing the old terminator. 2504// Also makes sure not to introduce new successors by assuming that edges to 2505// non-successor TrueBBs and FalseBBs aren't reachable. 2506static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2507 BasicBlock *TrueBB, BasicBlock *FalseBB, 2508 uint32_t TrueWeight, 2509 uint32_t FalseWeight){ 2510 // Remove any superfluous successor edges from the CFG. 2511 // First, figure out which successors to preserve. 2512 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2513 // successor. 2514 BasicBlock *KeepEdge1 = TrueBB; 2515 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2516 2517 // Then remove the rest. 2518 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2519 BasicBlock *Succ = OldTerm->getSuccessor(I); 2520 // Make sure only to keep exactly one copy of each edge. 2521 if (Succ == KeepEdge1) 2522 KeepEdge1 = 0; 2523 else if (Succ == KeepEdge2) 2524 KeepEdge2 = 0; 2525 else 2526 Succ->removePredecessor(OldTerm->getParent()); 2527 } 2528 2529 IRBuilder<> Builder(OldTerm); 2530 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2531 2532 // Insert an appropriate new terminator. 2533 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2534 if (TrueBB == FalseBB) 2535 // We were only looking for one successor, and it was present. 2536 // Create an unconditional branch to it. 2537 Builder.CreateBr(TrueBB); 2538 else { 2539 // We found both of the successors we were looking for. 2540 // Create a conditional branch sharing the condition of the select. 2541 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2542 if (TrueWeight != FalseWeight) 2543 NewBI->setMetadata(LLVMContext::MD_prof, 2544 MDBuilder(OldTerm->getContext()). 2545 createBranchWeights(TrueWeight, FalseWeight)); 2546 } 2547 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2548 // Neither of the selected blocks were successors, so this 2549 // terminator must be unreachable. 2550 new UnreachableInst(OldTerm->getContext(), OldTerm); 2551 } else { 2552 // One of the selected values was a successor, but the other wasn't. 2553 // Insert an unconditional branch to the one that was found; 2554 // the edge to the one that wasn't must be unreachable. 2555 if (KeepEdge1 == 0) 2556 // Only TrueBB was found. 2557 Builder.CreateBr(TrueBB); 2558 else 2559 // Only FalseBB was found. 2560 Builder.CreateBr(FalseBB); 2561 } 2562 2563 EraseTerminatorInstAndDCECond(OldTerm); 2564 return true; 2565} 2566 2567// SimplifySwitchOnSelect - Replaces 2568// (switch (select cond, X, Y)) on constant X, Y 2569// with a branch - conditional if X and Y lead to distinct BBs, 2570// unconditional otherwise. 2571static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2572 // Check for constant integer values in the select. 2573 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2574 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2575 if (!TrueVal || !FalseVal) 2576 return false; 2577 2578 // Find the relevant condition and destinations. 2579 Value *Condition = Select->getCondition(); 2580 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2581 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2582 2583 // Get weight for TrueBB and FalseBB. 2584 uint32_t TrueWeight = 0, FalseWeight = 0; 2585 SmallVector<uint64_t, 8> Weights; 2586 bool HasWeights = HasBranchWeights(SI); 2587 if (HasWeights) { 2588 GetBranchWeights(SI, Weights); 2589 if (Weights.size() == 1 + SI->getNumCases()) { 2590 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2591 getSuccessorIndex()]; 2592 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2593 getSuccessorIndex()]; 2594 } 2595 } 2596 2597 // Perform the actual simplification. 2598 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2599 TrueWeight, FalseWeight); 2600} 2601 2602// SimplifyIndirectBrOnSelect - Replaces 2603// (indirectbr (select cond, blockaddress(@fn, BlockA), 2604// blockaddress(@fn, BlockB))) 2605// with 2606// (br cond, BlockA, BlockB). 2607static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2608 // Check that both operands of the select are block addresses. 2609 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2610 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2611 if (!TBA || !FBA) 2612 return false; 2613 2614 // Extract the actual blocks. 2615 BasicBlock *TrueBB = TBA->getBasicBlock(); 2616 BasicBlock *FalseBB = FBA->getBasicBlock(); 2617 2618 // Perform the actual simplification. 2619 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2620 0, 0); 2621} 2622 2623/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2624/// instruction (a seteq/setne with a constant) as the only instruction in a 2625/// block that ends with an uncond branch. We are looking for a very specific 2626/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2627/// this case, we merge the first two "or's of icmp" into a switch, but then the 2628/// default value goes to an uncond block with a seteq in it, we get something 2629/// like: 2630/// 2631/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2632/// DEFAULT: 2633/// %tmp = icmp eq i8 %A, 92 2634/// br label %end 2635/// end: 2636/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2637/// 2638/// We prefer to split the edge to 'end' so that there is a true/false entry to 2639/// the PHI, merging the third icmp into the switch. 2640static bool TryToSimplifyUncondBranchWithICmpInIt( 2641 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2642 const DataLayout *TD) { 2643 BasicBlock *BB = ICI->getParent(); 2644 2645 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2646 // complex. 2647 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2648 2649 Value *V = ICI->getOperand(0); 2650 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2651 2652 // The pattern we're looking for is where our only predecessor is a switch on 2653 // 'V' and this block is the default case for the switch. In this case we can 2654 // fold the compared value into the switch to simplify things. 2655 BasicBlock *Pred = BB->getSinglePredecessor(); 2656 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2657 2658 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2659 if (SI->getCondition() != V) 2660 return false; 2661 2662 // If BB is reachable on a non-default case, then we simply know the value of 2663 // V in this block. Substitute it and constant fold the icmp instruction 2664 // away. 2665 if (SI->getDefaultDest() != BB) { 2666 ConstantInt *VVal = SI->findCaseDest(BB); 2667 assert(VVal && "Should have a unique destination value"); 2668 ICI->setOperand(0, VVal); 2669 2670 if (Value *V = SimplifyInstruction(ICI, TD)) { 2671 ICI->replaceAllUsesWith(V); 2672 ICI->eraseFromParent(); 2673 } 2674 // BB is now empty, so it is likely to simplify away. 2675 return SimplifyCFG(BB, TTI, TD) | true; 2676 } 2677 2678 // Ok, the block is reachable from the default dest. If the constant we're 2679 // comparing exists in one of the other edges, then we can constant fold ICI 2680 // and zap it. 2681 if (SI->findCaseValue(Cst) != SI->case_default()) { 2682 Value *V; 2683 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2684 V = ConstantInt::getFalse(BB->getContext()); 2685 else 2686 V = ConstantInt::getTrue(BB->getContext()); 2687 2688 ICI->replaceAllUsesWith(V); 2689 ICI->eraseFromParent(); 2690 // BB is now empty, so it is likely to simplify away. 2691 return SimplifyCFG(BB, TTI, TD) | true; 2692 } 2693 2694 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2695 // the block. 2696 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2697 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2698 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2699 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2700 return false; 2701 2702 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2703 // true in the PHI. 2704 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2705 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2706 2707 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2708 std::swap(DefaultCst, NewCst); 2709 2710 // Replace ICI (which is used by the PHI for the default value) with true or 2711 // false depending on if it is EQ or NE. 2712 ICI->replaceAllUsesWith(DefaultCst); 2713 ICI->eraseFromParent(); 2714 2715 // Okay, the switch goes to this block on a default value. Add an edge from 2716 // the switch to the merge point on the compared value. 2717 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2718 BB->getParent(), BB); 2719 SmallVector<uint64_t, 8> Weights; 2720 bool HasWeights = HasBranchWeights(SI); 2721 if (HasWeights) { 2722 GetBranchWeights(SI, Weights); 2723 if (Weights.size() == 1 + SI->getNumCases()) { 2724 // Split weight for default case to case for "Cst". 2725 Weights[0] = (Weights[0]+1) >> 1; 2726 Weights.push_back(Weights[0]); 2727 2728 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2729 SI->setMetadata(LLVMContext::MD_prof, 2730 MDBuilder(SI->getContext()). 2731 createBranchWeights(MDWeights)); 2732 } 2733 } 2734 SI->addCase(Cst, NewBB); 2735 2736 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2737 Builder.SetInsertPoint(NewBB); 2738 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2739 Builder.CreateBr(SuccBlock); 2740 PHIUse->addIncoming(NewCst, NewBB); 2741 return true; 2742} 2743 2744/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2745/// Check to see if it is branching on an or/and chain of icmp instructions, and 2746/// fold it into a switch instruction if so. 2747static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2748 IRBuilder<> &Builder) { 2749 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2750 if (Cond == 0) return false; 2751 2752 2753 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2754 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2755 // 'setne's and'ed together, collect them. 2756 Value *CompVal = 0; 2757 std::vector<ConstantInt*> Values; 2758 bool TrueWhenEqual = true; 2759 Value *ExtraCase = 0; 2760 unsigned UsedICmps = 0; 2761 2762 if (Cond->getOpcode() == Instruction::Or) { 2763 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2764 UsedICmps); 2765 } else if (Cond->getOpcode() == Instruction::And) { 2766 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2767 UsedICmps); 2768 TrueWhenEqual = false; 2769 } 2770 2771 // If we didn't have a multiply compared value, fail. 2772 if (CompVal == 0) return false; 2773 2774 // Avoid turning single icmps into a switch. 2775 if (UsedICmps <= 1) 2776 return false; 2777 2778 // There might be duplicate constants in the list, which the switch 2779 // instruction can't handle, remove them now. 2780 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2781 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2782 2783 // If Extra was used, we require at least two switch values to do the 2784 // transformation. A switch with one value is just an cond branch. 2785 if (ExtraCase && Values.size() < 2) return false; 2786 2787 // TODO: Preserve branch weight metadata, similarly to how 2788 // FoldValueComparisonIntoPredecessors preserves it. 2789 2790 // Figure out which block is which destination. 2791 BasicBlock *DefaultBB = BI->getSuccessor(1); 2792 BasicBlock *EdgeBB = BI->getSuccessor(0); 2793 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2794 2795 BasicBlock *BB = BI->getParent(); 2796 2797 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2798 << " cases into SWITCH. BB is:\n" << *BB); 2799 2800 // If there are any extra values that couldn't be folded into the switch 2801 // then we evaluate them with an explicit branch first. Split the block 2802 // right before the condbr to handle it. 2803 if (ExtraCase) { 2804 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2805 // Remove the uncond branch added to the old block. 2806 TerminatorInst *OldTI = BB->getTerminator(); 2807 Builder.SetInsertPoint(OldTI); 2808 2809 if (TrueWhenEqual) 2810 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2811 else 2812 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2813 2814 OldTI->eraseFromParent(); 2815 2816 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2817 // for the edge we just added. 2818 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2819 2820 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2821 << "\nEXTRABB = " << *BB); 2822 BB = NewBB; 2823 } 2824 2825 Builder.SetInsertPoint(BI); 2826 // Convert pointer to int before we switch. 2827 if (CompVal->getType()->isPointerTy()) { 2828 assert(TD && "Cannot switch on pointer without DataLayout"); 2829 CompVal = Builder.CreatePtrToInt(CompVal, 2830 TD->getIntPtrType(CompVal->getContext()), 2831 "magicptr"); 2832 } 2833 2834 // Create the new switch instruction now. 2835 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2836 2837 // Add all of the 'cases' to the switch instruction. 2838 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2839 New->addCase(Values[i], EdgeBB); 2840 2841 // We added edges from PI to the EdgeBB. As such, if there were any 2842 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2843 // the number of edges added. 2844 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2845 isa<PHINode>(BBI); ++BBI) { 2846 PHINode *PN = cast<PHINode>(BBI); 2847 Value *InVal = PN->getIncomingValueForBlock(BB); 2848 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2849 PN->addIncoming(InVal, BB); 2850 } 2851 2852 // Erase the old branch instruction. 2853 EraseTerminatorInstAndDCECond(BI); 2854 2855 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2856 return true; 2857} 2858 2859bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2860 // If this is a trivial landing pad that just continues unwinding the caught 2861 // exception then zap the landing pad, turning its invokes into calls. 2862 BasicBlock *BB = RI->getParent(); 2863 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2864 if (RI->getValue() != LPInst) 2865 // Not a landing pad, or the resume is not unwinding the exception that 2866 // caused control to branch here. 2867 return false; 2868 2869 // Check that there are no other instructions except for debug intrinsics. 2870 BasicBlock::iterator I = LPInst, E = RI; 2871 while (++I != E) 2872 if (!isa<DbgInfoIntrinsic>(I)) 2873 return false; 2874 2875 // Turn all invokes that unwind here into calls and delete the basic block. 2876 bool InvokeRequiresTableEntry = false; 2877 bool Changed = false; 2878 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2879 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2880 2881 if (II->hasFnAttr(Attribute::UWTable)) { 2882 // Don't remove an `invoke' instruction if the ABI requires an entry into 2883 // the table. 2884 InvokeRequiresTableEntry = true; 2885 continue; 2886 } 2887 2888 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2889 2890 // Insert a call instruction before the invoke. 2891 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2892 Call->takeName(II); 2893 Call->setCallingConv(II->getCallingConv()); 2894 Call->setAttributes(II->getAttributes()); 2895 Call->setDebugLoc(II->getDebugLoc()); 2896 2897 // Anything that used the value produced by the invoke instruction now uses 2898 // the value produced by the call instruction. Note that we do this even 2899 // for void functions and calls with no uses so that the callgraph edge is 2900 // updated. 2901 II->replaceAllUsesWith(Call); 2902 BB->removePredecessor(II->getParent()); 2903 2904 // Insert a branch to the normal destination right before the invoke. 2905 BranchInst::Create(II->getNormalDest(), II); 2906 2907 // Finally, delete the invoke instruction! 2908 II->eraseFromParent(); 2909 Changed = true; 2910 } 2911 2912 if (!InvokeRequiresTableEntry) 2913 // The landingpad is now unreachable. Zap it. 2914 BB->eraseFromParent(); 2915 2916 return Changed; 2917} 2918 2919bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2920 BasicBlock *BB = RI->getParent(); 2921 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2922 2923 // Find predecessors that end with branches. 2924 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2925 SmallVector<BranchInst*, 8> CondBranchPreds; 2926 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2927 BasicBlock *P = *PI; 2928 TerminatorInst *PTI = P->getTerminator(); 2929 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2930 if (BI->isUnconditional()) 2931 UncondBranchPreds.push_back(P); 2932 else 2933 CondBranchPreds.push_back(BI); 2934 } 2935 } 2936 2937 // If we found some, do the transformation! 2938 if (!UncondBranchPreds.empty() && DupRet) { 2939 while (!UncondBranchPreds.empty()) { 2940 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2941 DEBUG(dbgs() << "FOLDING: " << *BB 2942 << "INTO UNCOND BRANCH PRED: " << *Pred); 2943 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2944 } 2945 2946 // If we eliminated all predecessors of the block, delete the block now. 2947 if (pred_begin(BB) == pred_end(BB)) 2948 // We know there are no successors, so just nuke the block. 2949 BB->eraseFromParent(); 2950 2951 return true; 2952 } 2953 2954 // Check out all of the conditional branches going to this return 2955 // instruction. If any of them just select between returns, change the 2956 // branch itself into a select/return pair. 2957 while (!CondBranchPreds.empty()) { 2958 BranchInst *BI = CondBranchPreds.pop_back_val(); 2959 2960 // Check to see if the non-BB successor is also a return block. 2961 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2962 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2963 SimplifyCondBranchToTwoReturns(BI, Builder)) 2964 return true; 2965 } 2966 return false; 2967} 2968 2969bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2970 BasicBlock *BB = UI->getParent(); 2971 2972 bool Changed = false; 2973 2974 // If there are any instructions immediately before the unreachable that can 2975 // be removed, do so. 2976 while (UI != BB->begin()) { 2977 BasicBlock::iterator BBI = UI; 2978 --BBI; 2979 // Do not delete instructions that can have side effects which might cause 2980 // the unreachable to not be reachable; specifically, calls and volatile 2981 // operations may have this effect. 2982 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2983 2984 if (BBI->mayHaveSideEffects()) { 2985 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2986 if (SI->isVolatile()) 2987 break; 2988 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2989 if (LI->isVolatile()) 2990 break; 2991 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2992 if (RMWI->isVolatile()) 2993 break; 2994 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2995 if (CXI->isVolatile()) 2996 break; 2997 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2998 !isa<LandingPadInst>(BBI)) { 2999 break; 3000 } 3001 // Note that deleting LandingPad's here is in fact okay, although it 3002 // involves a bit of subtle reasoning. If this inst is a LandingPad, 3003 // all the predecessors of this block will be the unwind edges of Invokes, 3004 // and we can therefore guarantee this block will be erased. 3005 } 3006 3007 // Delete this instruction (any uses are guaranteed to be dead) 3008 if (!BBI->use_empty()) 3009 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 3010 BBI->eraseFromParent(); 3011 Changed = true; 3012 } 3013 3014 // If the unreachable instruction is the first in the block, take a gander 3015 // at all of the predecessors of this instruction, and simplify them. 3016 if (&BB->front() != UI) return Changed; 3017 3018 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 3019 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 3020 TerminatorInst *TI = Preds[i]->getTerminator(); 3021 IRBuilder<> Builder(TI); 3022 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 3023 if (BI->isUnconditional()) { 3024 if (BI->getSuccessor(0) == BB) { 3025 new UnreachableInst(TI->getContext(), TI); 3026 TI->eraseFromParent(); 3027 Changed = true; 3028 } 3029 } else { 3030 if (BI->getSuccessor(0) == BB) { 3031 Builder.CreateBr(BI->getSuccessor(1)); 3032 EraseTerminatorInstAndDCECond(BI); 3033 } else if (BI->getSuccessor(1) == BB) { 3034 Builder.CreateBr(BI->getSuccessor(0)); 3035 EraseTerminatorInstAndDCECond(BI); 3036 Changed = true; 3037 } 3038 } 3039 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 3040 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3041 i != e; ++i) 3042 if (i.getCaseSuccessor() == BB) { 3043 BB->removePredecessor(SI->getParent()); 3044 SI->removeCase(i); 3045 --i; --e; 3046 Changed = true; 3047 } 3048 // If the default value is unreachable, figure out the most popular 3049 // destination and make it the default. 3050 if (SI->getDefaultDest() == BB) { 3051 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 3052 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3053 i != e; ++i) { 3054 std::pair<unsigned, unsigned> &entry = 3055 Popularity[i.getCaseSuccessor()]; 3056 if (entry.first == 0) { 3057 entry.first = 1; 3058 entry.second = i.getCaseIndex(); 3059 } else { 3060 entry.first++; 3061 } 3062 } 3063 3064 // Find the most popular block. 3065 unsigned MaxPop = 0; 3066 unsigned MaxIndex = 0; 3067 BasicBlock *MaxBlock = 0; 3068 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 3069 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 3070 if (I->second.first > MaxPop || 3071 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 3072 MaxPop = I->second.first; 3073 MaxIndex = I->second.second; 3074 MaxBlock = I->first; 3075 } 3076 } 3077 if (MaxBlock) { 3078 // Make this the new default, allowing us to delete any explicit 3079 // edges to it. 3080 SI->setDefaultDest(MaxBlock); 3081 Changed = true; 3082 3083 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3084 // it. 3085 if (isa<PHINode>(MaxBlock->begin())) 3086 for (unsigned i = 0; i != MaxPop-1; ++i) 3087 MaxBlock->removePredecessor(SI->getParent()); 3088 3089 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3090 i != e; ++i) 3091 if (i.getCaseSuccessor() == MaxBlock) { 3092 SI->removeCase(i); 3093 --i; --e; 3094 } 3095 } 3096 } 3097 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3098 if (II->getUnwindDest() == BB) { 3099 // Convert the invoke to a call instruction. This would be a good 3100 // place to note that the call does not throw though. 3101 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3102 II->removeFromParent(); // Take out of symbol table 3103 3104 // Insert the call now... 3105 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3106 Builder.SetInsertPoint(BI); 3107 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3108 Args, II->getName()); 3109 CI->setCallingConv(II->getCallingConv()); 3110 CI->setAttributes(II->getAttributes()); 3111 // If the invoke produced a value, the call does now instead. 3112 II->replaceAllUsesWith(CI); 3113 delete II; 3114 Changed = true; 3115 } 3116 } 3117 } 3118 3119 // If this block is now dead, remove it. 3120 if (pred_begin(BB) == pred_end(BB) && 3121 BB != &BB->getParent()->getEntryBlock()) { 3122 // We know there are no successors, so just nuke the block. 3123 BB->eraseFromParent(); 3124 return true; 3125 } 3126 3127 return Changed; 3128} 3129 3130/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3131/// integer range comparison into a sub, an icmp and a branch. 3132static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3133 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3134 3135 // Make sure all cases point to the same destination and gather the values. 3136 SmallVector<ConstantInt *, 16> Cases; 3137 SwitchInst::CaseIt I = SI->case_begin(); 3138 Cases.push_back(I.getCaseValue()); 3139 SwitchInst::CaseIt PrevI = I++; 3140 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3141 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3142 return false; 3143 Cases.push_back(I.getCaseValue()); 3144 } 3145 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3146 3147 // Sort the case values, then check if they form a range we can transform. 3148 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3149 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3150 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3151 return false; 3152 } 3153 3154 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3155 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3156 3157 Value *Sub = SI->getCondition(); 3158 if (!Offset->isNullValue()) 3159 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3160 Value *Cmp; 3161 // If NumCases overflowed, then all possible values jump to the successor. 3162 if (NumCases->isNullValue() && SI->getNumCases() != 0) 3163 Cmp = ConstantInt::getTrue(SI->getContext()); 3164 else 3165 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3166 BranchInst *NewBI = Builder.CreateCondBr( 3167 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3168 3169 // Update weight for the newly-created conditional branch. 3170 SmallVector<uint64_t, 8> Weights; 3171 bool HasWeights = HasBranchWeights(SI); 3172 if (HasWeights) { 3173 GetBranchWeights(SI, Weights); 3174 if (Weights.size() == 1 + SI->getNumCases()) { 3175 // Combine all weights for the cases to be the true weight of NewBI. 3176 // We assume that the sum of all weights for a Terminator can fit into 32 3177 // bits. 3178 uint32_t NewTrueWeight = 0; 3179 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3180 NewTrueWeight += (uint32_t)Weights[I]; 3181 NewBI->setMetadata(LLVMContext::MD_prof, 3182 MDBuilder(SI->getContext()). 3183 createBranchWeights(NewTrueWeight, 3184 (uint32_t)Weights[0])); 3185 } 3186 } 3187 3188 // Prune obsolete incoming values off the successor's PHI nodes. 3189 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3190 isa<PHINode>(BBI); ++BBI) { 3191 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3192 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3193 } 3194 SI->eraseFromParent(); 3195 3196 return true; 3197} 3198 3199/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3200/// and use it to remove dead cases. 3201static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3202 Value *Cond = SI->getCondition(); 3203 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3204 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3205 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3206 3207 // Gather dead cases. 3208 SmallVector<ConstantInt*, 8> DeadCases; 3209 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3210 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3211 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3212 DeadCases.push_back(I.getCaseValue()); 3213 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3214 << I.getCaseValue() << "' is dead.\n"); 3215 } 3216 } 3217 3218 SmallVector<uint64_t, 8> Weights; 3219 bool HasWeight = HasBranchWeights(SI); 3220 if (HasWeight) { 3221 GetBranchWeights(SI, Weights); 3222 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3223 } 3224 3225 // Remove dead cases from the switch. 3226 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3227 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3228 assert(Case != SI->case_default() && 3229 "Case was not found. Probably mistake in DeadCases forming."); 3230 if (HasWeight) { 3231 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3232 Weights.pop_back(); 3233 } 3234 3235 // Prune unused values from PHI nodes. 3236 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3237 SI->removeCase(Case); 3238 } 3239 if (HasWeight) { 3240 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3241 SI->setMetadata(LLVMContext::MD_prof, 3242 MDBuilder(SI->getParent()->getContext()). 3243 createBranchWeights(MDWeights)); 3244 } 3245 3246 return !DeadCases.empty(); 3247} 3248 3249/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3250/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3251/// by an unconditional branch), look at the phi node for BB in the successor 3252/// block and see if the incoming value is equal to CaseValue. If so, return 3253/// the phi node, and set PhiIndex to BB's index in the phi node. 3254static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3255 BasicBlock *BB, 3256 int *PhiIndex) { 3257 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3258 return NULL; // BB must be empty to be a candidate for simplification. 3259 if (!BB->getSinglePredecessor()) 3260 return NULL; // BB must be dominated by the switch. 3261 3262 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3263 if (!Branch || !Branch->isUnconditional()) 3264 return NULL; // Terminator must be unconditional branch. 3265 3266 BasicBlock *Succ = Branch->getSuccessor(0); 3267 3268 BasicBlock::iterator I = Succ->begin(); 3269 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3270 int Idx = PHI->getBasicBlockIndex(BB); 3271 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3272 3273 Value *InValue = PHI->getIncomingValue(Idx); 3274 if (InValue != CaseValue) continue; 3275 3276 *PhiIndex = Idx; 3277 return PHI; 3278 } 3279 3280 return NULL; 3281} 3282 3283/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3284/// instruction to a phi node dominated by the switch, if that would mean that 3285/// some of the destination blocks of the switch can be folded away. 3286/// Returns true if a change is made. 3287static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3288 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3289 ForwardingNodesMap ForwardingNodes; 3290 3291 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3292 ConstantInt *CaseValue = I.getCaseValue(); 3293 BasicBlock *CaseDest = I.getCaseSuccessor(); 3294 3295 int PhiIndex; 3296 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3297 &PhiIndex); 3298 if (!PHI) continue; 3299 3300 ForwardingNodes[PHI].push_back(PhiIndex); 3301 } 3302 3303 bool Changed = false; 3304 3305 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3306 E = ForwardingNodes.end(); I != E; ++I) { 3307 PHINode *Phi = I->first; 3308 SmallVector<int,4> &Indexes = I->second; 3309 3310 if (Indexes.size() < 2) continue; 3311 3312 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3313 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3314 Changed = true; 3315 } 3316 3317 return Changed; 3318} 3319 3320/// ValidLookupTableConstant - Return true if the backend will be able to handle 3321/// initializing an array of constants like C. 3322static bool ValidLookupTableConstant(Constant *C) { 3323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3324 return CE->isGEPWithNoNotionalOverIndexing(); 3325 3326 return isa<ConstantFP>(C) || 3327 isa<ConstantInt>(C) || 3328 isa<ConstantPointerNull>(C) || 3329 isa<GlobalValue>(C) || 3330 isa<UndefValue>(C); 3331} 3332 3333/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3334/// its constant value in ConstantPool, returning 0 if it's not there. 3335static Constant *LookupConstant(Value *V, 3336 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3337 if (Constant *C = dyn_cast<Constant>(V)) 3338 return C; 3339 return ConstantPool.lookup(V); 3340} 3341 3342/// ConstantFold - Try to fold instruction I into a constant. This works for 3343/// simple instructions such as binary operations where both operands are 3344/// constant or can be replaced by constants from the ConstantPool. Returns the 3345/// resulting constant on success, 0 otherwise. 3346static Constant *ConstantFold(Instruction *I, 3347 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3348 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3349 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3350 if (!A) 3351 return 0; 3352 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3353 if (!B) 3354 return 0; 3355 return ConstantExpr::get(BO->getOpcode(), A, B); 3356 } 3357 3358 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3359 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3360 if (!A) 3361 return 0; 3362 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3363 if (!B) 3364 return 0; 3365 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3366 } 3367 3368 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3369 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3370 if (!A) 3371 return 0; 3372 if (A->isAllOnesValue()) 3373 return LookupConstant(Select->getTrueValue(), ConstantPool); 3374 if (A->isNullValue()) 3375 return LookupConstant(Select->getFalseValue(), ConstantPool); 3376 return 0; 3377 } 3378 3379 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3380 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3381 if (!A) 3382 return 0; 3383 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3384 } 3385 3386 return 0; 3387} 3388 3389/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3390/// at the common destination basic block, *CommonDest, for one of the case 3391/// destionations CaseDest corresponding to value CaseVal (0 for the default 3392/// case), of a switch instruction SI. 3393static bool GetCaseResults(SwitchInst *SI, 3394 ConstantInt *CaseVal, 3395 BasicBlock *CaseDest, 3396 BasicBlock **CommonDest, 3397 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3398 // The block from which we enter the common destination. 3399 BasicBlock *Pred = SI->getParent(); 3400 3401 // If CaseDest is empty except for some side-effect free instructions through 3402 // which we can constant-propagate the CaseVal, continue to its successor. 3403 SmallDenseMap<Value*, Constant*> ConstantPool; 3404 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3405 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3406 ++I) { 3407 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3408 // If the terminator is a simple branch, continue to the next block. 3409 if (T->getNumSuccessors() != 1) 3410 return false; 3411 Pred = CaseDest; 3412 CaseDest = T->getSuccessor(0); 3413 } else if (isa<DbgInfoIntrinsic>(I)) { 3414 // Skip debug intrinsic. 3415 continue; 3416 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3417 // Instruction is side-effect free and constant. 3418 ConstantPool.insert(std::make_pair(I, C)); 3419 } else { 3420 break; 3421 } 3422 } 3423 3424 // If we did not have a CommonDest before, use the current one. 3425 if (!*CommonDest) 3426 *CommonDest = CaseDest; 3427 // If the destination isn't the common one, abort. 3428 if (CaseDest != *CommonDest) 3429 return false; 3430 3431 // Get the values for this case from phi nodes in the destination block. 3432 BasicBlock::iterator I = (*CommonDest)->begin(); 3433 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3434 int Idx = PHI->getBasicBlockIndex(Pred); 3435 if (Idx == -1) 3436 continue; 3437 3438 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3439 ConstantPool); 3440 if (!ConstVal) 3441 return false; 3442 3443 // Note: If the constant comes from constant-propagating the case value 3444 // through the CaseDest basic block, it will be safe to remove the 3445 // instructions in that block. They cannot be used (except in the phi nodes 3446 // we visit) outside CaseDest, because that block does not dominate its 3447 // successor. If it did, we would not be in this phi node. 3448 3449 // Be conservative about which kinds of constants we support. 3450 if (!ValidLookupTableConstant(ConstVal)) 3451 return false; 3452 3453 Res.push_back(std::make_pair(PHI, ConstVal)); 3454 } 3455 3456 return true; 3457} 3458 3459namespace { 3460 /// SwitchLookupTable - This class represents a lookup table that can be used 3461 /// to replace a switch. 3462 class SwitchLookupTable { 3463 public: 3464 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3465 /// with the contents of Values, using DefaultValue to fill any holes in the 3466 /// table. 3467 SwitchLookupTable(Module &M, 3468 uint64_t TableSize, 3469 ConstantInt *Offset, 3470 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3471 Constant *DefaultValue, 3472 const DataLayout *TD); 3473 3474 /// BuildLookup - Build instructions with Builder to retrieve the value at 3475 /// the position given by Index in the lookup table. 3476 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3477 3478 /// WouldFitInRegister - Return true if a table with TableSize elements of 3479 /// type ElementType would fit in a target-legal register. 3480 static bool WouldFitInRegister(const DataLayout *TD, 3481 uint64_t TableSize, 3482 const Type *ElementType); 3483 3484 private: 3485 // Depending on the contents of the table, it can be represented in 3486 // different ways. 3487 enum { 3488 // For tables where each element contains the same value, we just have to 3489 // store that single value and return it for each lookup. 3490 SingleValueKind, 3491 3492 // For small tables with integer elements, we can pack them into a bitmap 3493 // that fits into a target-legal register. Values are retrieved by 3494 // shift and mask operations. 3495 BitMapKind, 3496 3497 // The table is stored as an array of values. Values are retrieved by load 3498 // instructions from the table. 3499 ArrayKind 3500 } Kind; 3501 3502 // For SingleValueKind, this is the single value. 3503 Constant *SingleValue; 3504 3505 // For BitMapKind, this is the bitmap. 3506 ConstantInt *BitMap; 3507 IntegerType *BitMapElementTy; 3508 3509 // For ArrayKind, this is the array. 3510 GlobalVariable *Array; 3511 }; 3512} 3513 3514SwitchLookupTable::SwitchLookupTable(Module &M, 3515 uint64_t TableSize, 3516 ConstantInt *Offset, 3517 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3518 Constant *DefaultValue, 3519 const DataLayout *TD) 3520 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3521 assert(Values.size() && "Can't build lookup table without values!"); 3522 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3523 3524 // If all values in the table are equal, this is that value. 3525 SingleValue = Values.begin()->second; 3526 3527 // Build up the table contents. 3528 SmallVector<Constant*, 64> TableContents(TableSize); 3529 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3530 ConstantInt *CaseVal = Values[I].first; 3531 Constant *CaseRes = Values[I].second; 3532 assert(CaseRes->getType() == DefaultValue->getType()); 3533 3534 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3535 .getLimitedValue(); 3536 TableContents[Idx] = CaseRes; 3537 3538 if (CaseRes != SingleValue) 3539 SingleValue = 0; 3540 } 3541 3542 // Fill in any holes in the table with the default result. 3543 if (Values.size() < TableSize) { 3544 for (uint64_t I = 0; I < TableSize; ++I) { 3545 if (!TableContents[I]) 3546 TableContents[I] = DefaultValue; 3547 } 3548 3549 if (DefaultValue != SingleValue) 3550 SingleValue = 0; 3551 } 3552 3553 // If each element in the table contains the same value, we only need to store 3554 // that single value. 3555 if (SingleValue) { 3556 Kind = SingleValueKind; 3557 return; 3558 } 3559 3560 // If the type is integer and the table fits in a register, build a bitmap. 3561 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3562 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3563 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3564 for (uint64_t I = TableSize; I > 0; --I) { 3565 TableInt <<= IT->getBitWidth(); 3566 // Insert values into the bitmap. Undef values are set to zero. 3567 if (!isa<UndefValue>(TableContents[I - 1])) { 3568 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3569 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3570 } 3571 } 3572 BitMap = ConstantInt::get(M.getContext(), TableInt); 3573 BitMapElementTy = IT; 3574 Kind = BitMapKind; 3575 ++NumBitMaps; 3576 return; 3577 } 3578 3579 // Store the table in an array. 3580 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3581 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3582 3583 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3584 GlobalVariable::PrivateLinkage, 3585 Initializer, 3586 "switch.table"); 3587 Array->setUnnamedAddr(true); 3588 Kind = ArrayKind; 3589} 3590 3591Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3592 switch (Kind) { 3593 case SingleValueKind: 3594 return SingleValue; 3595 case BitMapKind: { 3596 // Type of the bitmap (e.g. i59). 3597 IntegerType *MapTy = BitMap->getType(); 3598 3599 // Cast Index to the same type as the bitmap. 3600 // Note: The Index is <= the number of elements in the table, so 3601 // truncating it to the width of the bitmask is safe. 3602 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3603 3604 // Multiply the shift amount by the element width. 3605 ShiftAmt = Builder.CreateMul(ShiftAmt, 3606 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3607 "switch.shiftamt"); 3608 3609 // Shift down. 3610 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3611 "switch.downshift"); 3612 // Mask off. 3613 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3614 "switch.masked"); 3615 } 3616 case ArrayKind: { 3617 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3618 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3619 "switch.gep"); 3620 return Builder.CreateLoad(GEP, "switch.load"); 3621 } 3622 } 3623 llvm_unreachable("Unknown lookup table kind!"); 3624} 3625 3626bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3627 uint64_t TableSize, 3628 const Type *ElementType) { 3629 if (!TD) 3630 return false; 3631 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3632 if (!IT) 3633 return false; 3634 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3635 // are <= 15, we could try to narrow the type. 3636 3637 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3638 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3639 return false; 3640 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3641} 3642 3643/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3644/// for this switch, based on the number of caes, size of the table and the 3645/// types of the results. 3646static bool ShouldBuildLookupTable(SwitchInst *SI, 3647 uint64_t TableSize, 3648 const TargetTransformInfo &TTI, 3649 const DataLayout *TD, 3650 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3651 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3652 return false; // TableSize overflowed, or mul below might overflow. 3653 3654 bool AllTablesFitInRegister = true; 3655 bool HasIllegalType = false; 3656 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3657 E = ResultTypes.end(); I != E; ++I) { 3658 Type *Ty = I->second; 3659 3660 // Saturate this flag to true. 3661 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3662 3663 // Saturate this flag to false. 3664 AllTablesFitInRegister = AllTablesFitInRegister && 3665 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3666 3667 // If both flags saturate, we're done. NOTE: This *only* works with 3668 // saturating flags, and all flags have to saturate first due to the 3669 // non-deterministic behavior of iterating over a dense map. 3670 if (HasIllegalType && !AllTablesFitInRegister) 3671 break; 3672 } 3673 3674 // If each table would fit in a register, we should build it anyway. 3675 if (AllTablesFitInRegister) 3676 return true; 3677 3678 // Don't build a table that doesn't fit in-register if it has illegal types. 3679 if (HasIllegalType) 3680 return false; 3681 3682 // The table density should be at least 40%. This is the same criterion as for 3683 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3684 // FIXME: Find the best cut-off. 3685 return SI->getNumCases() * 10 >= TableSize * 4; 3686} 3687 3688/// SwitchToLookupTable - If the switch is only used to initialize one or more 3689/// phi nodes in a common successor block with different constant values, 3690/// replace the switch with lookup tables. 3691static bool SwitchToLookupTable(SwitchInst *SI, 3692 IRBuilder<> &Builder, 3693 const TargetTransformInfo &TTI, 3694 const DataLayout* TD) { 3695 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3696 3697 // Only build lookup table when we have a target that supports it. 3698 if (!TTI.shouldBuildLookupTables()) 3699 return false; 3700 3701 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3702 // split off a dense part and build a lookup table for that. 3703 3704 // FIXME: This creates arrays of GEPs to constant strings, which means each 3705 // GEP needs a runtime relocation in PIC code. We should just build one big 3706 // string and lookup indices into that. 3707 3708 // Ignore the switch if the number of cases is too small. 3709 // This is similar to the check when building jump tables in 3710 // SelectionDAGBuilder::handleJTSwitchCase. 3711 // FIXME: Determine the best cut-off. 3712 if (SI->getNumCases() < 4) 3713 return false; 3714 3715 // Figure out the corresponding result for each case value and phi node in the 3716 // common destination, as well as the the min and max case values. 3717 assert(SI->case_begin() != SI->case_end()); 3718 SwitchInst::CaseIt CI = SI->case_begin(); 3719 ConstantInt *MinCaseVal = CI.getCaseValue(); 3720 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3721 3722 BasicBlock *CommonDest = 0; 3723 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3724 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3725 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3726 SmallDenseMap<PHINode*, Type*> ResultTypes; 3727 SmallVector<PHINode*, 4> PHIs; 3728 3729 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3730 ConstantInt *CaseVal = CI.getCaseValue(); 3731 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3732 MinCaseVal = CaseVal; 3733 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3734 MaxCaseVal = CaseVal; 3735 3736 // Resulting value at phi nodes for this case value. 3737 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3738 ResultsTy Results; 3739 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3740 Results)) 3741 return false; 3742 3743 // Append the result from this case to the list for each phi. 3744 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3745 if (!ResultLists.count(I->first)) 3746 PHIs.push_back(I->first); 3747 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3748 } 3749 } 3750 3751 // Get the resulting values for the default case. 3752 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3753 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3754 DefaultResultsList)) 3755 return false; 3756 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3757 PHINode *PHI = DefaultResultsList[I].first; 3758 Constant *Result = DefaultResultsList[I].second; 3759 DefaultResults[PHI] = Result; 3760 ResultTypes[PHI] = Result->getType(); 3761 } 3762 3763 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3764 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3765 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3766 return false; 3767 3768 // Create the BB that does the lookups. 3769 Module &Mod = *CommonDest->getParent()->getParent(); 3770 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3771 "switch.lookup", 3772 CommonDest->getParent(), 3773 CommonDest); 3774 3775 // Check whether the condition value is within the case range, and branch to 3776 // the new BB. 3777 Builder.SetInsertPoint(SI); 3778 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3779 "switch.tableidx"); 3780 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3781 MinCaseVal->getType(), TableSize)); 3782 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3783 3784 // Populate the BB that does the lookups. 3785 Builder.SetInsertPoint(LookupBB); 3786 bool ReturnedEarly = false; 3787 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3788 PHINode *PHI = PHIs[I]; 3789 3790 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3791 DefaultResults[PHI], TD); 3792 3793 Value *Result = Table.BuildLookup(TableIndex, Builder); 3794 3795 // If the result is used to return immediately from the function, we want to 3796 // do that right here. 3797 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3798 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3799 Builder.CreateRet(Result); 3800 ReturnedEarly = true; 3801 break; 3802 } 3803 3804 PHI->addIncoming(Result, LookupBB); 3805 } 3806 3807 if (!ReturnedEarly) 3808 Builder.CreateBr(CommonDest); 3809 3810 // Remove the switch. 3811 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3812 BasicBlock *Succ = SI->getSuccessor(i); 3813 if (Succ == SI->getDefaultDest()) continue; 3814 Succ->removePredecessor(SI->getParent()); 3815 } 3816 SI->eraseFromParent(); 3817 3818 ++NumLookupTables; 3819 return true; 3820} 3821 3822bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3823 BasicBlock *BB = SI->getParent(); 3824 3825 if (isValueEqualityComparison(SI)) { 3826 // If we only have one predecessor, and if it is a branch on this value, 3827 // see if that predecessor totally determines the outcome of this switch. 3828 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3829 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3830 return SimplifyCFG(BB, TTI, TD) | true; 3831 3832 Value *Cond = SI->getCondition(); 3833 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3834 if (SimplifySwitchOnSelect(SI, Select)) 3835 return SimplifyCFG(BB, TTI, TD) | true; 3836 3837 // If the block only contains the switch, see if we can fold the block 3838 // away into any preds. 3839 BasicBlock::iterator BBI = BB->begin(); 3840 // Ignore dbg intrinsics. 3841 while (isa<DbgInfoIntrinsic>(BBI)) 3842 ++BBI; 3843 if (SI == &*BBI) 3844 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3845 return SimplifyCFG(BB, TTI, TD) | true; 3846 } 3847 3848 // Try to transform the switch into an icmp and a branch. 3849 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3850 return SimplifyCFG(BB, TTI, TD) | true; 3851 3852 // Remove unreachable cases. 3853 if (EliminateDeadSwitchCases(SI)) 3854 return SimplifyCFG(BB, TTI, TD) | true; 3855 3856 if (ForwardSwitchConditionToPHI(SI)) 3857 return SimplifyCFG(BB, TTI, TD) | true; 3858 3859 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3860 return SimplifyCFG(BB, TTI, TD) | true; 3861 3862 return false; 3863} 3864 3865bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3866 BasicBlock *BB = IBI->getParent(); 3867 bool Changed = false; 3868 3869 // Eliminate redundant destinations. 3870 SmallPtrSet<Value *, 8> Succs; 3871 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3872 BasicBlock *Dest = IBI->getDestination(i); 3873 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3874 Dest->removePredecessor(BB); 3875 IBI->removeDestination(i); 3876 --i; --e; 3877 Changed = true; 3878 } 3879 } 3880 3881 if (IBI->getNumDestinations() == 0) { 3882 // If the indirectbr has no successors, change it to unreachable. 3883 new UnreachableInst(IBI->getContext(), IBI); 3884 EraseTerminatorInstAndDCECond(IBI); 3885 return true; 3886 } 3887 3888 if (IBI->getNumDestinations() == 1) { 3889 // If the indirectbr has one successor, change it to a direct branch. 3890 BranchInst::Create(IBI->getDestination(0), IBI); 3891 EraseTerminatorInstAndDCECond(IBI); 3892 return true; 3893 } 3894 3895 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3896 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3897 return SimplifyCFG(BB, TTI, TD) | true; 3898 } 3899 return Changed; 3900} 3901 3902bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3903 BasicBlock *BB = BI->getParent(); 3904 3905 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3906 return true; 3907 3908 // If the Terminator is the only non-phi instruction, simplify the block. 3909 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3910 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3911 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3912 return true; 3913 3914 // If the only instruction in the block is a seteq/setne comparison 3915 // against a constant, try to simplify the block. 3916 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3917 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3918 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3919 ; 3920 if (I->isTerminator() && 3921 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3922 return true; 3923 } 3924 3925 // If this basic block is ONLY a compare and a branch, and if a predecessor 3926 // branches to us and our successor, fold the comparison into the 3927 // predecessor and use logical operations to update the incoming value 3928 // for PHI nodes in common successor. 3929 if (FoldBranchToCommonDest(BI)) 3930 return SimplifyCFG(BB, TTI, TD) | true; 3931 return false; 3932} 3933 3934 3935bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3936 BasicBlock *BB = BI->getParent(); 3937 3938 // Conditional branch 3939 if (isValueEqualityComparison(BI)) { 3940 // If we only have one predecessor, and if it is a branch on this value, 3941 // see if that predecessor totally determines the outcome of this 3942 // switch. 3943 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3944 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3945 return SimplifyCFG(BB, TTI, TD) | true; 3946 3947 // This block must be empty, except for the setcond inst, if it exists. 3948 // Ignore dbg intrinsics. 3949 BasicBlock::iterator I = BB->begin(); 3950 // Ignore dbg intrinsics. 3951 while (isa<DbgInfoIntrinsic>(I)) 3952 ++I; 3953 if (&*I == BI) { 3954 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3955 return SimplifyCFG(BB, TTI, TD) | true; 3956 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3957 ++I; 3958 // Ignore dbg intrinsics. 3959 while (isa<DbgInfoIntrinsic>(I)) 3960 ++I; 3961 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3962 return SimplifyCFG(BB, TTI, TD) | true; 3963 } 3964 } 3965 3966 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3967 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3968 return true; 3969 3970 // If this basic block is ONLY a compare and a branch, and if a predecessor 3971 // branches to us and one of our successors, fold the comparison into the 3972 // predecessor and use logical operations to pick the right destination. 3973 if (FoldBranchToCommonDest(BI)) 3974 return SimplifyCFG(BB, TTI, TD) | true; 3975 3976 // We have a conditional branch to two blocks that are only reachable 3977 // from BI. We know that the condbr dominates the two blocks, so see if 3978 // there is any identical code in the "then" and "else" blocks. If so, we 3979 // can hoist it up to the branching block. 3980 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3981 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3982 if (HoistThenElseCodeToIf(BI)) 3983 return SimplifyCFG(BB, TTI, TD) | true; 3984 } else { 3985 // If Successor #1 has multiple preds, we may be able to conditionally 3986 // execute Successor #0 if it branches to successor #1. 3987 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3988 if (Succ0TI->getNumSuccessors() == 1 && 3989 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3990 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3991 return SimplifyCFG(BB, TTI, TD) | true; 3992 } 3993 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3994 // If Successor #0 has multiple preds, we may be able to conditionally 3995 // execute Successor #1 if it branches to successor #0. 3996 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3997 if (Succ1TI->getNumSuccessors() == 1 && 3998 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3999 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 4000 return SimplifyCFG(BB, TTI, TD) | true; 4001 } 4002 4003 // If this is a branch on a phi node in the current block, thread control 4004 // through this block if any PHI node entries are constants. 4005 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 4006 if (PN->getParent() == BI->getParent()) 4007 if (FoldCondBranchOnPHI(BI, TD)) 4008 return SimplifyCFG(BB, TTI, TD) | true; 4009 4010 // Scan predecessor blocks for conditional branches. 4011 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 4012 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 4013 if (PBI != BI && PBI->isConditional()) 4014 if (SimplifyCondBranchToCondBranch(PBI, BI)) 4015 return SimplifyCFG(BB, TTI, TD) | true; 4016 4017 return false; 4018} 4019 4020/// Check if passing a value to an instruction will cause undefined behavior. 4021static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 4022 Constant *C = dyn_cast<Constant>(V); 4023 if (!C) 4024 return false; 4025 4026 if (I->use_empty()) 4027 return false; 4028 4029 if (C->isNullValue()) { 4030 // Only look at the first use, avoid hurting compile time with long uselists 4031 User *Use = *I->use_begin(); 4032 4033 // Now make sure that there are no instructions in between that can alter 4034 // control flow (eg. calls) 4035 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 4036 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 4037 return false; 4038 4039 // Look through GEPs. A load from a GEP derived from NULL is still undefined 4040 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 4041 if (GEP->getPointerOperand() == I) 4042 return passingValueIsAlwaysUndefined(V, GEP); 4043 4044 // Look through bitcasts. 4045 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 4046 return passingValueIsAlwaysUndefined(V, BC); 4047 4048 // Load from null is undefined. 4049 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 4050 if (!LI->isVolatile()) 4051 return LI->getPointerAddressSpace() == 0; 4052 4053 // Store to null is undefined. 4054 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 4055 if (!SI->isVolatile()) 4056 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 4057 } 4058 return false; 4059} 4060 4061/// If BB has an incoming value that will always trigger undefined behavior 4062/// (eg. null pointer dereference), remove the branch leading here. 4063static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 4064 for (BasicBlock::iterator i = BB->begin(); 4065 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 4066 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 4067 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 4068 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 4069 IRBuilder<> Builder(T); 4070 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 4071 BB->removePredecessor(PHI->getIncomingBlock(i)); 4072 // Turn uncoditional branches into unreachables and remove the dead 4073 // destination from conditional branches. 4074 if (BI->isUnconditional()) 4075 Builder.CreateUnreachable(); 4076 else 4077 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 4078 BI->getSuccessor(0)); 4079 BI->eraseFromParent(); 4080 return true; 4081 } 4082 // TODO: SwitchInst. 4083 } 4084 4085 return false; 4086} 4087 4088bool SimplifyCFGOpt::run(BasicBlock *BB) { 4089 bool Changed = false; 4090 4091 assert(BB && BB->getParent() && "Block not embedded in function!"); 4092 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4093 4094 // Remove basic blocks that have no predecessors (except the entry block)... 4095 // or that just have themself as a predecessor. These are unreachable. 4096 if ((pred_begin(BB) == pred_end(BB) && 4097 BB != &BB->getParent()->getEntryBlock()) || 4098 BB->getSinglePredecessor() == BB) { 4099 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4100 DeleteDeadBlock(BB); 4101 return true; 4102 } 4103 4104 // Check to see if we can constant propagate this terminator instruction 4105 // away... 4106 Changed |= ConstantFoldTerminator(BB, true); 4107 4108 // Check for and eliminate duplicate PHI nodes in this block. 4109 Changed |= EliminateDuplicatePHINodes(BB); 4110 4111 // Check for and remove branches that will always cause undefined behavior. 4112 Changed |= removeUndefIntroducingPredecessor(BB); 4113 4114 // Merge basic blocks into their predecessor if there is only one distinct 4115 // pred, and if there is only one distinct successor of the predecessor, and 4116 // if there are no PHI nodes. 4117 // 4118 if (MergeBlockIntoPredecessor(BB)) 4119 return true; 4120 4121 IRBuilder<> Builder(BB); 4122 4123 // If there is a trivial two-entry PHI node in this basic block, and we can 4124 // eliminate it, do so now. 4125 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4126 if (PN->getNumIncomingValues() == 2) 4127 Changed |= FoldTwoEntryPHINode(PN, TD); 4128 4129 Builder.SetInsertPoint(BB->getTerminator()); 4130 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4131 if (BI->isUnconditional()) { 4132 if (SimplifyUncondBranch(BI, Builder)) return true; 4133 } else { 4134 if (SimplifyCondBranch(BI, Builder)) return true; 4135 } 4136 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4137 if (SimplifyReturn(RI, Builder)) return true; 4138 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4139 if (SimplifyResume(RI, Builder)) return true; 4140 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4141 if (SimplifySwitch(SI, Builder)) return true; 4142 } else if (UnreachableInst *UI = 4143 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4144 if (SimplifyUnreachable(UI)) return true; 4145 } else if (IndirectBrInst *IBI = 4146 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4147 if (SimplifyIndirectBr(IBI)) return true; 4148 } 4149 4150 return Changed; 4151} 4152 4153/// SimplifyCFG - This function is used to do simplification of a CFG. For 4154/// example, it adjusts branches to branches to eliminate the extra hop, it 4155/// eliminates unreachable basic blocks, and does other "peephole" optimization 4156/// of the CFG. It returns true if a modification was made. 4157/// 4158bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4159 const DataLayout *TD) { 4160 return SimplifyCFGOpt(TTI, TD).run(BB); 4161} 4162