SimplifyCFG.cpp revision 3931bdb788774e33dd623d6b9142ae69542c33aa
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/// ComputeSpeculationCost - 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->isEquality() && GetConstantInt(ICI->getOperand(1), TD)) 537 CV = ICI->getOperand(0); 538 539 // Unwrap any lossless ptrtoint cast. 540 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 541 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 542 CV = PTII->getOperand(0); 543 return CV; 544} 545 546/// GetValueEqualityComparisonCases - Given a value comparison instruction, 547/// decode all of the 'cases' that it represents and return the 'default' block. 548BasicBlock *SimplifyCFGOpt:: 549GetValueEqualityComparisonCases(TerminatorInst *TI, 550 std::vector<ValueEqualityComparisonCase> 551 &Cases) { 552 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 553 Cases.reserve(SI->getNumCases()); 554 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 555 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 556 i.getCaseSuccessor())); 557 return SI->getDefaultDest(); 558 } 559 560 BranchInst *BI = cast<BranchInst>(TI); 561 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 562 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 563 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 564 TD), 565 Succ)); 566 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 567} 568 569 570/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 571/// in the list that match the specified block. 572static void EliminateBlockCases(BasicBlock *BB, 573 std::vector<ValueEqualityComparisonCase> &Cases) { 574 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 575} 576 577/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 578/// well. 579static bool 580ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 581 std::vector<ValueEqualityComparisonCase > &C2) { 582 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 583 584 // Make V1 be smaller than V2. 585 if (V1->size() > V2->size()) 586 std::swap(V1, V2); 587 588 if (V1->size() == 0) return false; 589 if (V1->size() == 1) { 590 // Just scan V2. 591 ConstantInt *TheVal = (*V1)[0].Value; 592 for (unsigned i = 0, e = V2->size(); i != e; ++i) 593 if (TheVal == (*V2)[i].Value) 594 return true; 595 } 596 597 // Otherwise, just sort both lists and compare element by element. 598 array_pod_sort(V1->begin(), V1->end()); 599 array_pod_sort(V2->begin(), V2->end()); 600 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 601 while (i1 != e1 && i2 != e2) { 602 if ((*V1)[i1].Value == (*V2)[i2].Value) 603 return true; 604 if ((*V1)[i1].Value < (*V2)[i2].Value) 605 ++i1; 606 else 607 ++i2; 608 } 609 return false; 610} 611 612/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 613/// terminator instruction and its block is known to only have a single 614/// predecessor block, check to see if that predecessor is also a value 615/// comparison with the same value, and if that comparison determines the 616/// outcome of this comparison. If so, simplify TI. This does a very limited 617/// form of jump threading. 618bool SimplifyCFGOpt:: 619SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 620 BasicBlock *Pred, 621 IRBuilder<> &Builder) { 622 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 623 if (!PredVal) return false; // Not a value comparison in predecessor. 624 625 Value *ThisVal = isValueEqualityComparison(TI); 626 assert(ThisVal && "This isn't a value comparison!!"); 627 if (ThisVal != PredVal) return false; // Different predicates. 628 629 // TODO: Preserve branch weight metadata, similarly to how 630 // FoldValueComparisonIntoPredecessors preserves it. 631 632 // Find out information about when control will move from Pred to TI's block. 633 std::vector<ValueEqualityComparisonCase> PredCases; 634 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 635 PredCases); 636 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 637 638 // Find information about how control leaves this block. 639 std::vector<ValueEqualityComparisonCase> ThisCases; 640 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 641 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 642 643 // If TI's block is the default block from Pred's comparison, potentially 644 // simplify TI based on this knowledge. 645 if (PredDef == TI->getParent()) { 646 // If we are here, we know that the value is none of those cases listed in 647 // PredCases. If there are any cases in ThisCases that are in PredCases, we 648 // can simplify TI. 649 if (!ValuesOverlap(PredCases, ThisCases)) 650 return false; 651 652 if (isa<BranchInst>(TI)) { 653 // Okay, one of the successors of this condbr is dead. Convert it to a 654 // uncond br. 655 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 656 // Insert the new branch. 657 Instruction *NI = Builder.CreateBr(ThisDef); 658 (void) NI; 659 660 // Remove PHI node entries for the dead edge. 661 ThisCases[0].Dest->removePredecessor(TI->getParent()); 662 663 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 664 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 665 666 EraseTerminatorInstAndDCECond(TI); 667 return true; 668 } 669 670 SwitchInst *SI = cast<SwitchInst>(TI); 671 // Okay, TI has cases that are statically dead, prune them away. 672 SmallPtrSet<Constant*, 16> DeadCases; 673 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 674 DeadCases.insert(PredCases[i].Value); 675 676 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 677 << "Through successor TI: " << *TI); 678 679 // Collect branch weights into a vector. 680 SmallVector<uint32_t, 8> Weights; 681 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 682 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 683 if (HasWeight) 684 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 685 ++MD_i) { 686 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 687 assert(CI); 688 Weights.push_back(CI->getValue().getZExtValue()); 689 } 690 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 691 --i; 692 if (DeadCases.count(i.getCaseValue())) { 693 if (HasWeight) { 694 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 695 Weights.pop_back(); 696 } 697 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 698 SI->removeCase(i); 699 } 700 } 701 if (HasWeight && Weights.size() >= 2) 702 SI->setMetadata(LLVMContext::MD_prof, 703 MDBuilder(SI->getParent()->getContext()). 704 createBranchWeights(Weights)); 705 706 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 707 return true; 708 } 709 710 // Otherwise, TI's block must correspond to some matched value. Find out 711 // which value (or set of values) this is. 712 ConstantInt *TIV = 0; 713 BasicBlock *TIBB = TI->getParent(); 714 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 715 if (PredCases[i].Dest == TIBB) { 716 if (TIV != 0) 717 return false; // Cannot handle multiple values coming to this block. 718 TIV = PredCases[i].Value; 719 } 720 assert(TIV && "No edge from pred to succ?"); 721 722 // Okay, we found the one constant that our value can be if we get into TI's 723 // BB. Find out which successor will unconditionally be branched to. 724 BasicBlock *TheRealDest = 0; 725 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 726 if (ThisCases[i].Value == TIV) { 727 TheRealDest = ThisCases[i].Dest; 728 break; 729 } 730 731 // If not handled by any explicit cases, it is handled by the default case. 732 if (TheRealDest == 0) TheRealDest = ThisDef; 733 734 // Remove PHI node entries for dead edges. 735 BasicBlock *CheckEdge = TheRealDest; 736 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 737 if (*SI != CheckEdge) 738 (*SI)->removePredecessor(TIBB); 739 else 740 CheckEdge = 0; 741 742 // Insert the new branch. 743 Instruction *NI = Builder.CreateBr(TheRealDest); 744 (void) NI; 745 746 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 747 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 748 749 EraseTerminatorInstAndDCECond(TI); 750 return true; 751} 752 753namespace { 754 /// ConstantIntOrdering - This class implements a stable ordering of constant 755 /// integers that does not depend on their address. This is important for 756 /// applications that sort ConstantInt's to ensure uniqueness. 757 struct ConstantIntOrdering { 758 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 759 return LHS->getValue().ult(RHS->getValue()); 760 } 761 }; 762} 763 764static int ConstantIntSortPredicate(const void *P1, const void *P2) { 765 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 766 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 767 if (LHS->getValue().ult(RHS->getValue())) 768 return 1; 769 if (LHS->getValue() == RHS->getValue()) 770 return 0; 771 return -1; 772} 773 774static inline bool HasBranchWeights(const Instruction* I) { 775 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 776 if (ProfMD && ProfMD->getOperand(0)) 777 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 778 return MDS->getString().equals("branch_weights"); 779 780 return false; 781} 782 783/// Get Weights of a given TerminatorInst, the default weight is at the front 784/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 785/// metadata. 786static void GetBranchWeights(TerminatorInst *TI, 787 SmallVectorImpl<uint64_t> &Weights) { 788 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 789 assert(MD); 790 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 791 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 792 assert(CI); 793 Weights.push_back(CI->getValue().getZExtValue()); 794 } 795 796 // If TI is a conditional eq, the default case is the false case, 797 // and the corresponding branch-weight data is at index 2. We swap the 798 // default weight to be the first entry. 799 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 800 assert(Weights.size() == 2); 801 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 802 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 803 std::swap(Weights.front(), Weights.back()); 804 } 805} 806 807/// Sees if any of the weights are too big for a uint32_t, and halves all the 808/// weights if any are. 809static void FitWeights(MutableArrayRef<uint64_t> Weights) { 810 bool Halve = false; 811 for (unsigned i = 0; i < Weights.size(); ++i) 812 if (Weights[i] > UINT_MAX) { 813 Halve = true; 814 break; 815 } 816 817 if (! Halve) 818 return; 819 820 for (unsigned i = 0; i < Weights.size(); ++i) 821 Weights[i] /= 2; 822} 823 824/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 825/// equality comparison instruction (either a switch or a branch on "X == c"). 826/// See if any of the predecessors of the terminator block are value comparisons 827/// on the same value. If so, and if safe to do so, fold them together. 828bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 829 IRBuilder<> &Builder) { 830 BasicBlock *BB = TI->getParent(); 831 Value *CV = isValueEqualityComparison(TI); // CondVal 832 assert(CV && "Not a comparison?"); 833 bool Changed = false; 834 835 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 836 while (!Preds.empty()) { 837 BasicBlock *Pred = Preds.pop_back_val(); 838 839 // See if the predecessor is a comparison with the same value. 840 TerminatorInst *PTI = Pred->getTerminator(); 841 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 842 843 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 844 // Figure out which 'cases' to copy from SI to PSI. 845 std::vector<ValueEqualityComparisonCase> BBCases; 846 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 847 848 std::vector<ValueEqualityComparisonCase> PredCases; 849 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 850 851 // Based on whether the default edge from PTI goes to BB or not, fill in 852 // PredCases and PredDefault with the new switch cases we would like to 853 // build. 854 SmallVector<BasicBlock*, 8> NewSuccessors; 855 856 // Update the branch weight metadata along the way 857 SmallVector<uint64_t, 8> Weights; 858 bool PredHasWeights = HasBranchWeights(PTI); 859 bool SuccHasWeights = HasBranchWeights(TI); 860 861 if (PredHasWeights) { 862 GetBranchWeights(PTI, Weights); 863 // branch-weight metadata is inconsistent here. 864 if (Weights.size() != 1 + PredCases.size()) 865 PredHasWeights = SuccHasWeights = false; 866 } else if (SuccHasWeights) 867 // If there are no predecessor weights but there are successor weights, 868 // populate Weights with 1, which will later be scaled to the sum of 869 // successor's weights 870 Weights.assign(1 + PredCases.size(), 1); 871 872 SmallVector<uint64_t, 8> SuccWeights; 873 if (SuccHasWeights) { 874 GetBranchWeights(TI, SuccWeights); 875 // branch-weight metadata is inconsistent here. 876 if (SuccWeights.size() != 1 + BBCases.size()) 877 PredHasWeights = SuccHasWeights = false; 878 } else if (PredHasWeights) 879 SuccWeights.assign(1 + BBCases.size(), 1); 880 881 if (PredDefault == BB) { 882 // If this is the default destination from PTI, only the edges in TI 883 // that don't occur in PTI, or that branch to BB will be activated. 884 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 885 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 886 if (PredCases[i].Dest != BB) 887 PTIHandled.insert(PredCases[i].Value); 888 else { 889 // The default destination is BB, we don't need explicit targets. 890 std::swap(PredCases[i], PredCases.back()); 891 892 if (PredHasWeights || SuccHasWeights) { 893 // Increase weight for the default case. 894 Weights[0] += Weights[i+1]; 895 std::swap(Weights[i+1], Weights.back()); 896 Weights.pop_back(); 897 } 898 899 PredCases.pop_back(); 900 --i; --e; 901 } 902 903 // Reconstruct the new switch statement we will be building. 904 if (PredDefault != BBDefault) { 905 PredDefault->removePredecessor(Pred); 906 PredDefault = BBDefault; 907 NewSuccessors.push_back(BBDefault); 908 } 909 910 unsigned CasesFromPred = Weights.size(); 911 uint64_t ValidTotalSuccWeight = 0; 912 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 913 if (!PTIHandled.count(BBCases[i].Value) && 914 BBCases[i].Dest != BBDefault) { 915 PredCases.push_back(BBCases[i]); 916 NewSuccessors.push_back(BBCases[i].Dest); 917 if (SuccHasWeights || PredHasWeights) { 918 // The default weight is at index 0, so weight for the ith case 919 // should be at index i+1. Scale the cases from successor by 920 // PredDefaultWeight (Weights[0]). 921 Weights.push_back(Weights[0] * SuccWeights[i+1]); 922 ValidTotalSuccWeight += SuccWeights[i+1]; 923 } 924 } 925 926 if (SuccHasWeights || PredHasWeights) { 927 ValidTotalSuccWeight += SuccWeights[0]; 928 // Scale the cases from predecessor by ValidTotalSuccWeight. 929 for (unsigned i = 1; i < CasesFromPred; ++i) 930 Weights[i] *= ValidTotalSuccWeight; 931 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 932 Weights[0] *= SuccWeights[0]; 933 } 934 } else { 935 // If this is not the default destination from PSI, only the edges 936 // in SI that occur in PSI with a destination of BB will be 937 // activated. 938 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 939 std::map<ConstantInt*, uint64_t> WeightsForHandled; 940 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 941 if (PredCases[i].Dest == BB) { 942 PTIHandled.insert(PredCases[i].Value); 943 944 if (PredHasWeights || SuccHasWeights) { 945 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 946 std::swap(Weights[i+1], Weights.back()); 947 Weights.pop_back(); 948 } 949 950 std::swap(PredCases[i], PredCases.back()); 951 PredCases.pop_back(); 952 --i; --e; 953 } 954 955 // Okay, now we know which constants were sent to BB from the 956 // predecessor. Figure out where they will all go now. 957 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 958 if (PTIHandled.count(BBCases[i].Value)) { 959 // If this is one we are capable of getting... 960 if (PredHasWeights || SuccHasWeights) 961 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 962 PredCases.push_back(BBCases[i]); 963 NewSuccessors.push_back(BBCases[i].Dest); 964 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 965 } 966 967 // If there are any constants vectored to BB that TI doesn't handle, 968 // they must go to the default destination of TI. 969 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 970 PTIHandled.begin(), 971 E = PTIHandled.end(); I != E; ++I) { 972 if (PredHasWeights || SuccHasWeights) 973 Weights.push_back(WeightsForHandled[*I]); 974 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 975 NewSuccessors.push_back(BBDefault); 976 } 977 } 978 979 // Okay, at this point, we know which new successor Pred will get. Make 980 // sure we update the number of entries in the PHI nodes for these 981 // successors. 982 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 983 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 984 985 Builder.SetInsertPoint(PTI); 986 // Convert pointer to int before we switch. 987 if (CV->getType()->isPointerTy()) { 988 assert(TD && "Cannot switch on pointer without DataLayout"); 989 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 990 "magicptr"); 991 } 992 993 // Now that the successors are updated, create the new Switch instruction. 994 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 995 PredCases.size()); 996 NewSI->setDebugLoc(PTI->getDebugLoc()); 997 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 998 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 999 1000 if (PredHasWeights || SuccHasWeights) { 1001 // Halve the weights if any of them cannot fit in an uint32_t 1002 FitWeights(Weights); 1003 1004 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1005 1006 NewSI->setMetadata(LLVMContext::MD_prof, 1007 MDBuilder(BB->getContext()). 1008 createBranchWeights(MDWeights)); 1009 } 1010 1011 EraseTerminatorInstAndDCECond(PTI); 1012 1013 // Okay, last check. If BB is still a successor of PSI, then we must 1014 // have an infinite loop case. If so, add an infinitely looping block 1015 // to handle the case to preserve the behavior of the code. 1016 BasicBlock *InfLoopBlock = 0; 1017 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1018 if (NewSI->getSuccessor(i) == BB) { 1019 if (InfLoopBlock == 0) { 1020 // Insert it at the end of the function, because it's either code, 1021 // or it won't matter if it's hot. :) 1022 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1023 "infloop", BB->getParent()); 1024 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1025 } 1026 NewSI->setSuccessor(i, InfLoopBlock); 1027 } 1028 1029 Changed = true; 1030 } 1031 } 1032 return Changed; 1033} 1034 1035// isSafeToHoistInvoke - If we would need to insert a select that uses the 1036// value of this invoke (comments in HoistThenElseCodeToIf explain why we 1037// would need to do this), we can't hoist the invoke, as there is nowhere 1038// to put the select in this case. 1039static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1040 Instruction *I1, Instruction *I2) { 1041 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1042 PHINode *PN; 1043 for (BasicBlock::iterator BBI = SI->begin(); 1044 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1045 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1046 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1047 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1048 return false; 1049 } 1050 } 1051 } 1052 return true; 1053} 1054 1055/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1056/// BB2, hoist any common code in the two blocks up into the branch block. The 1057/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1058static bool HoistThenElseCodeToIf(BranchInst *BI) { 1059 // This does very trivial matching, with limited scanning, to find identical 1060 // instructions in the two blocks. In particular, we don't want to get into 1061 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1062 // such, we currently just scan for obviously identical instructions in an 1063 // identical order. 1064 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1065 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1066 1067 BasicBlock::iterator BB1_Itr = BB1->begin(); 1068 BasicBlock::iterator BB2_Itr = BB2->begin(); 1069 1070 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1071 // Skip debug info if it is not identical. 1072 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1073 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1074 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1075 while (isa<DbgInfoIntrinsic>(I1)) 1076 I1 = BB1_Itr++; 1077 while (isa<DbgInfoIntrinsic>(I2)) 1078 I2 = BB2_Itr++; 1079 } 1080 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1081 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1082 return false; 1083 1084 BasicBlock *BIParent = BI->getParent(); 1085 1086 bool Changed = false; 1087 do { 1088 // If we are hoisting the terminator instruction, don't move one (making a 1089 // broken BB), instead clone it, and remove BI. 1090 if (isa<TerminatorInst>(I1)) 1091 goto HoistTerminator; 1092 1093 // For a normal instruction, we just move one to right before the branch, 1094 // then replace all uses of the other with the first. Finally, we remove 1095 // the now redundant second instruction. 1096 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1097 if (!I2->use_empty()) 1098 I2->replaceAllUsesWith(I1); 1099 I1->intersectOptionalDataWith(I2); 1100 I2->eraseFromParent(); 1101 Changed = true; 1102 1103 I1 = BB1_Itr++; 1104 I2 = BB2_Itr++; 1105 // Skip debug info if it is not identical. 1106 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1107 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1108 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1109 while (isa<DbgInfoIntrinsic>(I1)) 1110 I1 = BB1_Itr++; 1111 while (isa<DbgInfoIntrinsic>(I2)) 1112 I2 = BB2_Itr++; 1113 } 1114 } while (I1->isIdenticalToWhenDefined(I2)); 1115 1116 return true; 1117 1118HoistTerminator: 1119 // It may not be possible to hoist an invoke. 1120 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1121 return Changed; 1122 1123 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1124 PHINode *PN; 1125 for (BasicBlock::iterator BBI = SI->begin(); 1126 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1127 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1128 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1129 if (BB1V == BB2V) 1130 continue; 1131 1132 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V)) 1133 return Changed; 1134 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V)) 1135 return Changed; 1136 } 1137 } 1138 1139 // Okay, it is safe to hoist the terminator. 1140 Instruction *NT = I1->clone(); 1141 BIParent->getInstList().insert(BI, NT); 1142 if (!NT->getType()->isVoidTy()) { 1143 I1->replaceAllUsesWith(NT); 1144 I2->replaceAllUsesWith(NT); 1145 NT->takeName(I1); 1146 } 1147 1148 IRBuilder<true, NoFolder> Builder(NT); 1149 // Hoisting one of the terminators from our successor is a great thing. 1150 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1151 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1152 // nodes, so we insert select instruction to compute the final result. 1153 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1154 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1155 PHINode *PN; 1156 for (BasicBlock::iterator BBI = SI->begin(); 1157 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1158 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1159 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1160 if (BB1V == BB2V) continue; 1161 1162 // These values do not agree. Insert a select instruction before NT 1163 // that determines the right value. 1164 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1165 if (SI == 0) 1166 SI = cast<SelectInst> 1167 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1168 BB1V->getName()+"."+BB2V->getName())); 1169 1170 // Make the PHI node use the select for all incoming values for BB1/BB2 1171 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1172 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1173 PN->setIncomingValue(i, SI); 1174 } 1175 } 1176 1177 // Update any PHI nodes in our new successors. 1178 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1179 AddPredecessorToBlock(*SI, BIParent, BB1); 1180 1181 EraseTerminatorInstAndDCECond(BI); 1182 return true; 1183} 1184 1185/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1186/// check whether BBEnd has only two predecessors and the other predecessor 1187/// ends with an unconditional branch. If it is true, sink any common code 1188/// in the two predecessors to BBEnd. 1189static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1190 assert(BI1->isUnconditional()); 1191 BasicBlock *BB1 = BI1->getParent(); 1192 BasicBlock *BBEnd = BI1->getSuccessor(0); 1193 1194 // Check that BBEnd has two predecessors and the other predecessor ends with 1195 // an unconditional branch. 1196 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1197 BasicBlock *Pred0 = *PI++; 1198 if (PI == PE) // Only one predecessor. 1199 return false; 1200 BasicBlock *Pred1 = *PI++; 1201 if (PI != PE) // More than two predecessors. 1202 return false; 1203 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1204 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1205 if (!BI2 || !BI2->isUnconditional()) 1206 return false; 1207 1208 // Gather the PHI nodes in BBEnd. 1209 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1210 Instruction *FirstNonPhiInBBEnd = 0; 1211 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1212 I != E; ++I) { 1213 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1214 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1215 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1216 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1217 } else { 1218 FirstNonPhiInBBEnd = &*I; 1219 break; 1220 } 1221 } 1222 if (!FirstNonPhiInBBEnd) 1223 return false; 1224 1225 1226 // This does very trivial matching, with limited scanning, to find identical 1227 // instructions in the two blocks. We scan backward for obviously identical 1228 // instructions in an identical order. 1229 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1230 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1231 RE2 = BB2->getInstList().rend(); 1232 // Skip debug info. 1233 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1234 if (RI1 == RE1) 1235 return false; 1236 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1237 if (RI2 == RE2) 1238 return false; 1239 // Skip the unconditional branches. 1240 ++RI1; 1241 ++RI2; 1242 1243 bool Changed = false; 1244 while (RI1 != RE1 && RI2 != RE2) { 1245 // Skip debug info. 1246 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1247 if (RI1 == RE1) 1248 return Changed; 1249 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1250 if (RI2 == RE2) 1251 return Changed; 1252 1253 Instruction *I1 = &*RI1, *I2 = &*RI2; 1254 // I1 and I2 should have a single use in the same PHI node, and they 1255 // perform the same operation. 1256 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1257 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1258 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1259 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1260 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1261 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1262 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1263 !I1->hasOneUse() || !I2->hasOneUse() || 1264 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1265 MapValueFromBB1ToBB2[I1].first != I2) 1266 return Changed; 1267 1268 // Check whether we should swap the operands of ICmpInst. 1269 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1270 bool SwapOpnds = false; 1271 if (ICmp1 && ICmp2 && 1272 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1273 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1274 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1275 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1276 ICmp2->swapOperands(); 1277 SwapOpnds = true; 1278 } 1279 if (!I1->isSameOperationAs(I2)) { 1280 if (SwapOpnds) 1281 ICmp2->swapOperands(); 1282 return Changed; 1283 } 1284 1285 // The operands should be either the same or they need to be generated 1286 // with a PHI node after sinking. We only handle the case where there is 1287 // a single pair of different operands. 1288 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1289 unsigned Op1Idx = 0; 1290 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1291 if (I1->getOperand(I) == I2->getOperand(I)) 1292 continue; 1293 // Early exit if we have more-than one pair of different operands or 1294 // the different operand is already in MapValueFromBB1ToBB2. 1295 // Early exit if we need a PHI node to replace a constant. 1296 if (DifferentOp1 || 1297 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1298 MapValueFromBB1ToBB2.end() || 1299 isa<Constant>(I1->getOperand(I)) || 1300 isa<Constant>(I2->getOperand(I))) { 1301 // If we can't sink the instructions, undo the swapping. 1302 if (SwapOpnds) 1303 ICmp2->swapOperands(); 1304 return Changed; 1305 } 1306 DifferentOp1 = I1->getOperand(I); 1307 Op1Idx = I; 1308 DifferentOp2 = I2->getOperand(I); 1309 } 1310 1311 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1312 // remove (I1, I2) from MapValueFromBB1ToBB2. 1313 if (DifferentOp1) { 1314 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1315 DifferentOp1->getName() + ".sink", 1316 BBEnd->begin()); 1317 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1318 // I1 should use NewPN instead of DifferentOp1. 1319 I1->setOperand(Op1Idx, NewPN); 1320 NewPN->addIncoming(DifferentOp1, BB1); 1321 NewPN->addIncoming(DifferentOp2, BB2); 1322 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1323 } 1324 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1325 MapValueFromBB1ToBB2.erase(I1); 1326 1327 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1328 DEBUG(dbgs() << " " << *I2 << "\n";); 1329 // We need to update RE1 and RE2 if we are going to sink the first 1330 // instruction in the basic block down. 1331 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1332 // Sink the instruction. 1333 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1334 if (!OldPN->use_empty()) 1335 OldPN->replaceAllUsesWith(I1); 1336 OldPN->eraseFromParent(); 1337 1338 if (!I2->use_empty()) 1339 I2->replaceAllUsesWith(I1); 1340 I1->intersectOptionalDataWith(I2); 1341 I2->eraseFromParent(); 1342 1343 if (UpdateRE1) 1344 RE1 = BB1->getInstList().rend(); 1345 if (UpdateRE2) 1346 RE2 = BB2->getInstList().rend(); 1347 FirstNonPhiInBBEnd = I1; 1348 NumSinkCommons++; 1349 Changed = true; 1350 } 1351 return Changed; 1352} 1353 1354/// \brief Determine if we can hoist sink a sole store instruction out of a 1355/// conditional block. 1356/// 1357/// We are looking for code like the following: 1358/// BrBB: 1359/// store i32 %add, i32* %arrayidx2 1360/// ... // No other stores or function calls (we could be calling a memory 1361/// ... // function). 1362/// %cmp = icmp ult %x, %y 1363/// br i1 %cmp, label %EndBB, label %ThenBB 1364/// ThenBB: 1365/// store i32 %add5, i32* %arrayidx2 1366/// br label EndBB 1367/// EndBB: 1368/// ... 1369/// We are going to transform this into: 1370/// BrBB: 1371/// store i32 %add, i32* %arrayidx2 1372/// ... // 1373/// %cmp = icmp ult %x, %y 1374/// %add.add5 = select i1 %cmp, i32 %add, %add5 1375/// store i32 %add.add5, i32* %arrayidx2 1376/// ... 1377/// 1378/// \return The pointer to the value of the previous store if the store can be 1379/// hoisted into the predecessor block. 0 otherwise. 1380static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, 1381 BasicBlock *StoreBB, BasicBlock *EndBB) { 1382 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); 1383 if (!StoreToHoist) 1384 return 0; 1385 1386 // Volatile or atomic. 1387 if (!StoreToHoist->isSimple()) 1388 return 0; 1389 1390 Value *StorePtr = StoreToHoist->getPointerOperand(); 1391 1392 // Look for a store to the same pointer in BrBB. 1393 unsigned MaxNumInstToLookAt = 10; 1394 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), 1395 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { 1396 Instruction *CurI = &*RI; 1397 1398 // Could be calling an instruction that effects memory like free(). 1399 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) 1400 return 0; 1401 1402 StoreInst *SI = dyn_cast<StoreInst>(CurI); 1403 // Found the previous store make sure it stores to the same location. 1404 if (SI && SI->getPointerOperand() == StorePtr) 1405 // Found the previous store, return its value operand. 1406 return SI->getValueOperand(); 1407 else if (SI) 1408 return 0; // Unknown store. 1409 } 1410 1411 return 0; 1412} 1413 1414/// \brief Speculate a conditional basic block flattening the CFG. 1415/// 1416/// Note that this is a very risky transform currently. Speculating 1417/// instructions like this is most often not desirable. Instead, there is an MI 1418/// pass which can do it with full awareness of the resource constraints. 1419/// However, some cases are "obvious" and we should do directly. An example of 1420/// this is speculating a single, reasonably cheap instruction. 1421/// 1422/// There is only one distinct advantage to flattening the CFG at the IR level: 1423/// it makes very common but simplistic optimizations such as are common in 1424/// instcombine and the DAG combiner more powerful by removing CFG edges and 1425/// modeling their effects with easier to reason about SSA value graphs. 1426/// 1427/// 1428/// An illustration of this transform is turning this IR: 1429/// \code 1430/// BB: 1431/// %cmp = icmp ult %x, %y 1432/// br i1 %cmp, label %EndBB, label %ThenBB 1433/// ThenBB: 1434/// %sub = sub %x, %y 1435/// br label BB2 1436/// EndBB: 1437/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] 1438/// ... 1439/// \endcode 1440/// 1441/// Into this IR: 1442/// \code 1443/// BB: 1444/// %cmp = icmp ult %x, %y 1445/// %sub = sub %x, %y 1446/// %cond = select i1 %cmp, 0, %sub 1447/// ... 1448/// \endcode 1449/// 1450/// \returns true if the conditional block is removed. 1451static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { 1452 // Be conservative for now. FP select instruction can often be expensive. 1453 Value *BrCond = BI->getCondition(); 1454 if (isa<FCmpInst>(BrCond)) 1455 return false; 1456 1457 BasicBlock *BB = BI->getParent(); 1458 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0); 1459 1460 // If ThenBB is actually on the false edge of the conditional branch, remember 1461 // to swap the select operands later. 1462 bool Invert = false; 1463 if (ThenBB != BI->getSuccessor(0)) { 1464 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"); 1465 Invert = true; 1466 } 1467 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"); 1468 1469 // Keep a count of how many times instructions are used within CondBB when 1470 // they are candidates for sinking into CondBB. Specifically: 1471 // - They are defined in BB, and 1472 // - They have no side effects, and 1473 // - All of their uses are in CondBB. 1474 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; 1475 1476 unsigned SpeculationCost = 0; 1477 Value *SpeculatedStoreValue = 0; 1478 StoreInst *SpeculatedStore = 0; 1479 for (BasicBlock::iterator BBI = ThenBB->begin(), 1480 BBE = llvm::prior(ThenBB->end()); 1481 BBI != BBE; ++BBI) { 1482 Instruction *I = BBI; 1483 // Skip debug info. 1484 if (isa<DbgInfoIntrinsic>(I)) 1485 continue; 1486 1487 // Only speculatively execution a single instruction (not counting the 1488 // terminator) for now. 1489 ++SpeculationCost; 1490 if (SpeculationCost > 1) 1491 return false; 1492 1493 // Don't hoist the instruction if it's unsafe or expensive. 1494 if (!isSafeToSpeculativelyExecute(I) && 1495 !(HoistCondStores && 1496 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, 1497 EndBB)))) 1498 return false; 1499 if (!SpeculatedStoreValue && 1500 ComputeSpeculationCost(I) > PHINodeFoldingThreshold) 1501 return false; 1502 1503 // Store the store speculation candidate. 1504 if (SpeculatedStoreValue) 1505 SpeculatedStore = cast<StoreInst>(I); 1506 1507 // Do not hoist the instruction if any of its operands are defined but not 1508 // used in BB. The transformation will prevent the operand from 1509 // being sunk into the use block. 1510 for (User::op_iterator i = I->op_begin(), e = I->op_end(); 1511 i != e; ++i) { 1512 Instruction *OpI = dyn_cast<Instruction>(*i); 1513 if (!OpI || OpI->getParent() != BB || 1514 OpI->mayHaveSideEffects()) 1515 continue; // Not a candidate for sinking. 1516 1517 ++SinkCandidateUseCounts[OpI]; 1518 } 1519 } 1520 1521 // Consider any sink candidates which are only used in CondBB as costs for 1522 // speculation. Note, while we iterate over a DenseMap here, we are summing 1523 // and so iteration order isn't significant. 1524 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I = 1525 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end(); 1526 I != E; ++I) 1527 if (I->first->getNumUses() == I->second) { 1528 ++SpeculationCost; 1529 if (SpeculationCost > 1) 1530 return false; 1531 } 1532 1533 // Check that the PHI nodes can be converted to selects. 1534 bool HaveRewritablePHIs = false; 1535 for (BasicBlock::iterator I = EndBB->begin(); 1536 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1537 Value *OrigV = PN->getIncomingValueForBlock(BB); 1538 Value *ThenV = PN->getIncomingValueForBlock(ThenBB); 1539 1540 // Skip PHIs which are trivial. 1541 if (ThenV == OrigV) 1542 continue; 1543 1544 HaveRewritablePHIs = true; 1545 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV); 1546 if (!CE) 1547 continue; // Known safe and cheap. 1548 1549 if (!isSafeToSpeculativelyExecute(CE)) 1550 return false; 1551 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1552 return false; 1553 1554 // Account for the cost of an unfolded ConstantExpr which could end up 1555 // getting expanded into Instructions. 1556 // FIXME: This doesn't account for how many operations are combined in the 1557 // constant expression. 1558 ++SpeculationCost; 1559 if (SpeculationCost > 1) 1560 return false; 1561 } 1562 1563 // If there are no PHIs to process, bail early. This helps ensure idempotence 1564 // as well. 1565 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) 1566 return false; 1567 1568 // If we get here, we can hoist the instruction and if-convert. 1569 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); 1570 1571 // Insert a select of the value of the speculated store. 1572 if (SpeculatedStoreValue) { 1573 IRBuilder<true, NoFolder> Builder(BI); 1574 Value *TrueV = SpeculatedStore->getValueOperand(); 1575 Value *FalseV = SpeculatedStoreValue; 1576 if (Invert) 1577 std::swap(TrueV, FalseV); 1578 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + 1579 "." + FalseV->getName()); 1580 SpeculatedStore->setOperand(0, S); 1581 } 1582 1583 // Hoist the instructions. 1584 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), 1585 llvm::prior(ThenBB->end())); 1586 1587 // Insert selects and rewrite the PHI operands. 1588 IRBuilder<true, NoFolder> Builder(BI); 1589 for (BasicBlock::iterator I = EndBB->begin(); 1590 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1591 unsigned OrigI = PN->getBasicBlockIndex(BB); 1592 unsigned ThenI = PN->getBasicBlockIndex(ThenBB); 1593 Value *OrigV = PN->getIncomingValue(OrigI); 1594 Value *ThenV = PN->getIncomingValue(ThenI); 1595 1596 // Skip PHIs which are trivial. 1597 if (OrigV == ThenV) 1598 continue; 1599 1600 // Create a select whose true value is the speculatively executed value and 1601 // false value is the preexisting value. Swap them if the branch 1602 // destinations were inverted. 1603 Value *TrueV = ThenV, *FalseV = OrigV; 1604 if (Invert) 1605 std::swap(TrueV, FalseV); 1606 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, 1607 TrueV->getName() + "." + FalseV->getName()); 1608 PN->setIncomingValue(OrigI, V); 1609 PN->setIncomingValue(ThenI, V); 1610 } 1611 1612 ++NumSpeculations; 1613 return true; 1614} 1615 1616/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1617/// across this block. 1618static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1619 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1620 unsigned Size = 0; 1621 1622 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1623 if (isa<DbgInfoIntrinsic>(BBI)) 1624 continue; 1625 if (Size > 10) return false; // Don't clone large BB's. 1626 ++Size; 1627 1628 // We can only support instructions that do not define values that are 1629 // live outside of the current basic block. 1630 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1631 UI != E; ++UI) { 1632 Instruction *U = cast<Instruction>(*UI); 1633 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1634 } 1635 1636 // Looks ok, continue checking. 1637 } 1638 1639 return true; 1640} 1641 1642/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1643/// that is defined in the same block as the branch and if any PHI entries are 1644/// constants, thread edges corresponding to that entry to be branches to their 1645/// ultimate destination. 1646static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1647 BasicBlock *BB = BI->getParent(); 1648 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1649 // NOTE: we currently cannot transform this case if the PHI node is used 1650 // outside of the block. 1651 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1652 return false; 1653 1654 // Degenerate case of a single entry PHI. 1655 if (PN->getNumIncomingValues() == 1) { 1656 FoldSingleEntryPHINodes(PN->getParent()); 1657 return true; 1658 } 1659 1660 // Now we know that this block has multiple preds and two succs. 1661 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1662 1663 // Okay, this is a simple enough basic block. See if any phi values are 1664 // constants. 1665 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1666 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1667 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1668 1669 // Okay, we now know that all edges from PredBB should be revectored to 1670 // branch to RealDest. 1671 BasicBlock *PredBB = PN->getIncomingBlock(i); 1672 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1673 1674 if (RealDest == BB) continue; // Skip self loops. 1675 // Skip if the predecessor's terminator is an indirect branch. 1676 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1677 1678 // The dest block might have PHI nodes, other predecessors and other 1679 // difficult cases. Instead of being smart about this, just insert a new 1680 // block that jumps to the destination block, effectively splitting 1681 // the edge we are about to create. 1682 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1683 RealDest->getName()+".critedge", 1684 RealDest->getParent(), RealDest); 1685 BranchInst::Create(RealDest, EdgeBB); 1686 1687 // Update PHI nodes. 1688 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1689 1690 // BB may have instructions that are being threaded over. Clone these 1691 // instructions into EdgeBB. We know that there will be no uses of the 1692 // cloned instructions outside of EdgeBB. 1693 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1694 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1695 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1696 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1697 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1698 continue; 1699 } 1700 // Clone the instruction. 1701 Instruction *N = BBI->clone(); 1702 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1703 1704 // Update operands due to translation. 1705 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1706 i != e; ++i) { 1707 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1708 if (PI != TranslateMap.end()) 1709 *i = PI->second; 1710 } 1711 1712 // Check for trivial simplification. 1713 if (Value *V = SimplifyInstruction(N, TD)) { 1714 TranslateMap[BBI] = V; 1715 delete N; // Instruction folded away, don't need actual inst 1716 } else { 1717 // Insert the new instruction into its new home. 1718 EdgeBB->getInstList().insert(InsertPt, N); 1719 if (!BBI->use_empty()) 1720 TranslateMap[BBI] = N; 1721 } 1722 } 1723 1724 // Loop over all of the edges from PredBB to BB, changing them to branch 1725 // to EdgeBB instead. 1726 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1727 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1728 if (PredBBTI->getSuccessor(i) == BB) { 1729 BB->removePredecessor(PredBB); 1730 PredBBTI->setSuccessor(i, EdgeBB); 1731 } 1732 1733 // Recurse, simplifying any other constants. 1734 return FoldCondBranchOnPHI(BI, TD) | true; 1735 } 1736 1737 return false; 1738} 1739 1740/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1741/// PHI node, see if we can eliminate it. 1742static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1743 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1744 // statement", which has a very simple dominance structure. Basically, we 1745 // are trying to find the condition that is being branched on, which 1746 // subsequently causes this merge to happen. We really want control 1747 // dependence information for this check, but simplifycfg can't keep it up 1748 // to date, and this catches most of the cases we care about anyway. 1749 BasicBlock *BB = PN->getParent(); 1750 BasicBlock *IfTrue, *IfFalse; 1751 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1752 if (!IfCond || 1753 // Don't bother if the branch will be constant folded trivially. 1754 isa<ConstantInt>(IfCond)) 1755 return false; 1756 1757 // Okay, we found that we can merge this two-entry phi node into a select. 1758 // Doing so would require us to fold *all* two entry phi nodes in this block. 1759 // At some point this becomes non-profitable (particularly if the target 1760 // doesn't support cmov's). Only do this transformation if there are two or 1761 // fewer PHI nodes in this block. 1762 unsigned NumPhis = 0; 1763 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1764 if (NumPhis > 2) 1765 return false; 1766 1767 // Loop over the PHI's seeing if we can promote them all to select 1768 // instructions. While we are at it, keep track of the instructions 1769 // that need to be moved to the dominating block. 1770 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1771 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1772 MaxCostVal1 = PHINodeFoldingThreshold; 1773 1774 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1775 PHINode *PN = cast<PHINode>(II++); 1776 if (Value *V = SimplifyInstruction(PN, TD)) { 1777 PN->replaceAllUsesWith(V); 1778 PN->eraseFromParent(); 1779 continue; 1780 } 1781 1782 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1783 MaxCostVal0) || 1784 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1785 MaxCostVal1)) 1786 return false; 1787 } 1788 1789 // If we folded the first phi, PN dangles at this point. Refresh it. If 1790 // we ran out of PHIs then we simplified them all. 1791 PN = dyn_cast<PHINode>(BB->begin()); 1792 if (PN == 0) return true; 1793 1794 // Don't fold i1 branches on PHIs which contain binary operators. These can 1795 // often be turned into switches and other things. 1796 if (PN->getType()->isIntegerTy(1) && 1797 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1798 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1799 isa<BinaryOperator>(IfCond))) 1800 return false; 1801 1802 // If we all PHI nodes are promotable, check to make sure that all 1803 // instructions in the predecessor blocks can be promoted as well. If 1804 // not, we won't be able to get rid of the control flow, so it's not 1805 // worth promoting to select instructions. 1806 BasicBlock *DomBlock = 0; 1807 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1808 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1809 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1810 IfBlock1 = 0; 1811 } else { 1812 DomBlock = *pred_begin(IfBlock1); 1813 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1814 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1815 // This is not an aggressive instruction that we can promote. 1816 // Because of this, we won't be able to get rid of the control 1817 // flow, so the xform is not worth it. 1818 return false; 1819 } 1820 } 1821 1822 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1823 IfBlock2 = 0; 1824 } else { 1825 DomBlock = *pred_begin(IfBlock2); 1826 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1827 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1828 // This is not an aggressive instruction that we can promote. 1829 // Because of this, we won't be able to get rid of the control 1830 // flow, so the xform is not worth it. 1831 return false; 1832 } 1833 } 1834 1835 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1836 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1837 1838 // If we can still promote the PHI nodes after this gauntlet of tests, 1839 // do all of the PHI's now. 1840 Instruction *InsertPt = DomBlock->getTerminator(); 1841 IRBuilder<true, NoFolder> Builder(InsertPt); 1842 1843 // Move all 'aggressive' instructions, which are defined in the 1844 // conditional parts of the if's up to the dominating block. 1845 if (IfBlock1) 1846 DomBlock->getInstList().splice(InsertPt, 1847 IfBlock1->getInstList(), IfBlock1->begin(), 1848 IfBlock1->getTerminator()); 1849 if (IfBlock2) 1850 DomBlock->getInstList().splice(InsertPt, 1851 IfBlock2->getInstList(), IfBlock2->begin(), 1852 IfBlock2->getTerminator()); 1853 1854 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1855 // Change the PHI node into a select instruction. 1856 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1857 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1858 1859 SelectInst *NV = 1860 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1861 PN->replaceAllUsesWith(NV); 1862 NV->takeName(PN); 1863 PN->eraseFromParent(); 1864 } 1865 1866 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1867 // has been flattened. Change DomBlock to jump directly to our new block to 1868 // avoid other simplifycfg's kicking in on the diamond. 1869 TerminatorInst *OldTI = DomBlock->getTerminator(); 1870 Builder.SetInsertPoint(OldTI); 1871 Builder.CreateBr(BB); 1872 OldTI->eraseFromParent(); 1873 return true; 1874} 1875 1876/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1877/// to two returning blocks, try to merge them together into one return, 1878/// introducing a select if the return values disagree. 1879static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1880 IRBuilder<> &Builder) { 1881 assert(BI->isConditional() && "Must be a conditional branch"); 1882 BasicBlock *TrueSucc = BI->getSuccessor(0); 1883 BasicBlock *FalseSucc = BI->getSuccessor(1); 1884 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1885 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1886 1887 // Check to ensure both blocks are empty (just a return) or optionally empty 1888 // with PHI nodes. If there are other instructions, merging would cause extra 1889 // computation on one path or the other. 1890 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1891 return false; 1892 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1893 return false; 1894 1895 Builder.SetInsertPoint(BI); 1896 // Okay, we found a branch that is going to two return nodes. If 1897 // there is no return value for this function, just change the 1898 // branch into a return. 1899 if (FalseRet->getNumOperands() == 0) { 1900 TrueSucc->removePredecessor(BI->getParent()); 1901 FalseSucc->removePredecessor(BI->getParent()); 1902 Builder.CreateRetVoid(); 1903 EraseTerminatorInstAndDCECond(BI); 1904 return true; 1905 } 1906 1907 // Otherwise, figure out what the true and false return values are 1908 // so we can insert a new select instruction. 1909 Value *TrueValue = TrueRet->getReturnValue(); 1910 Value *FalseValue = FalseRet->getReturnValue(); 1911 1912 // Unwrap any PHI nodes in the return blocks. 1913 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1914 if (TVPN->getParent() == TrueSucc) 1915 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1916 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1917 if (FVPN->getParent() == FalseSucc) 1918 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1919 1920 // In order for this transformation to be safe, we must be able to 1921 // unconditionally execute both operands to the return. This is 1922 // normally the case, but we could have a potentially-trapping 1923 // constant expression that prevents this transformation from being 1924 // safe. 1925 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1926 if (TCV->canTrap()) 1927 return false; 1928 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1929 if (FCV->canTrap()) 1930 return false; 1931 1932 // Okay, we collected all the mapped values and checked them for sanity, and 1933 // defined to really do this transformation. First, update the CFG. 1934 TrueSucc->removePredecessor(BI->getParent()); 1935 FalseSucc->removePredecessor(BI->getParent()); 1936 1937 // Insert select instructions where needed. 1938 Value *BrCond = BI->getCondition(); 1939 if (TrueValue) { 1940 // Insert a select if the results differ. 1941 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1942 } else if (isa<UndefValue>(TrueValue)) { 1943 TrueValue = FalseValue; 1944 } else { 1945 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1946 FalseValue, "retval"); 1947 } 1948 } 1949 1950 Value *RI = !TrueValue ? 1951 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1952 1953 (void) RI; 1954 1955 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1956 << "\n " << *BI << "NewRet = " << *RI 1957 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1958 1959 EraseTerminatorInstAndDCECond(BI); 1960 1961 return true; 1962} 1963 1964/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1965/// probabilities of the branch taking each edge. Fills in the two APInt 1966/// parameters and return true, or returns false if no or invalid metadata was 1967/// found. 1968static bool ExtractBranchMetadata(BranchInst *BI, 1969 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1970 assert(BI->isConditional() && 1971 "Looking for probabilities on unconditional branch?"); 1972 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1973 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1974 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1975 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1976 if (!CITrue || !CIFalse) return false; 1977 ProbTrue = CITrue->getValue().getZExtValue(); 1978 ProbFalse = CIFalse->getValue().getZExtValue(); 1979 return true; 1980} 1981 1982/// checkCSEInPredecessor - Return true if the given instruction is available 1983/// in its predecessor block. If yes, the instruction will be removed. 1984/// 1985static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1986 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1987 return false; 1988 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1989 Instruction *PBI = &*I; 1990 // Check whether Inst and PBI generate the same value. 1991 if (Inst->isIdenticalTo(PBI)) { 1992 Inst->replaceAllUsesWith(PBI); 1993 Inst->eraseFromParent(); 1994 return true; 1995 } 1996 } 1997 return false; 1998} 1999 2000/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 2001/// predecessor branches to us and one of our successors, fold the block into 2002/// the predecessor and use logical operations to pick the right destination. 2003bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 2004 BasicBlock *BB = BI->getParent(); 2005 2006 Instruction *Cond = 0; 2007 if (BI->isConditional()) 2008 Cond = dyn_cast<Instruction>(BI->getCondition()); 2009 else { 2010 // For unconditional branch, check for a simple CFG pattern, where 2011 // BB has a single predecessor and BB's successor is also its predecessor's 2012 // successor. If such pattern exisits, check for CSE between BB and its 2013 // predecessor. 2014 if (BasicBlock *PB = BB->getSinglePredecessor()) 2015 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 2016 if (PBI->isConditional() && 2017 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 2018 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 2019 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 2020 I != E; ) { 2021 Instruction *Curr = I++; 2022 if (isa<CmpInst>(Curr)) { 2023 Cond = Curr; 2024 break; 2025 } 2026 // Quit if we can't remove this instruction. 2027 if (!checkCSEInPredecessor(Curr, PB)) 2028 return false; 2029 } 2030 } 2031 2032 if (Cond == 0) 2033 return false; 2034 } 2035 2036 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 2037 Cond->getParent() != BB || !Cond->hasOneUse()) 2038 return false; 2039 2040 // Only allow this if the condition is a simple instruction that can be 2041 // executed unconditionally. It must be in the same block as the branch, and 2042 // must be at the front of the block. 2043 BasicBlock::iterator FrontIt = BB->front(); 2044 2045 // Ignore dbg intrinsics. 2046 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2047 2048 // Allow a single instruction to be hoisted in addition to the compare 2049 // that feeds the branch. We later ensure that any values that _it_ uses 2050 // were also live in the predecessor, so that we don't unnecessarily create 2051 // register pressure or inhibit out-of-order execution. 2052 Instruction *BonusInst = 0; 2053 if (&*FrontIt != Cond && 2054 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 2055 isSafeToSpeculativelyExecute(FrontIt)) { 2056 BonusInst = &*FrontIt; 2057 ++FrontIt; 2058 2059 // Ignore dbg intrinsics. 2060 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 2061 } 2062 2063 // Only a single bonus inst is allowed. 2064 if (&*FrontIt != Cond) 2065 return false; 2066 2067 // Make sure the instruction after the condition is the cond branch. 2068 BasicBlock::iterator CondIt = Cond; ++CondIt; 2069 2070 // Ingore dbg intrinsics. 2071 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 2072 2073 if (&*CondIt != BI) 2074 return false; 2075 2076 // Cond is known to be a compare or binary operator. Check to make sure that 2077 // neither operand is a potentially-trapping constant expression. 2078 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 2079 if (CE->canTrap()) 2080 return false; 2081 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 2082 if (CE->canTrap()) 2083 return false; 2084 2085 // Finally, don't infinitely unroll conditional loops. 2086 BasicBlock *TrueDest = BI->getSuccessor(0); 2087 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 2088 if (TrueDest == BB || FalseDest == BB) 2089 return false; 2090 2091 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2092 BasicBlock *PredBlock = *PI; 2093 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 2094 2095 // Check that we have two conditional branches. If there is a PHI node in 2096 // the common successor, verify that the same value flows in from both 2097 // blocks. 2098 SmallVector<PHINode*, 4> PHIs; 2099 if (PBI == 0 || PBI->isUnconditional() || 2100 (BI->isConditional() && 2101 !SafeToMergeTerminators(BI, PBI)) || 2102 (!BI->isConditional() && 2103 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 2104 continue; 2105 2106 // Determine if the two branches share a common destination. 2107 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 2108 bool InvertPredCond = false; 2109 2110 if (BI->isConditional()) { 2111 if (PBI->getSuccessor(0) == TrueDest) 2112 Opc = Instruction::Or; 2113 else if (PBI->getSuccessor(1) == FalseDest) 2114 Opc = Instruction::And; 2115 else if (PBI->getSuccessor(0) == FalseDest) 2116 Opc = Instruction::And, InvertPredCond = true; 2117 else if (PBI->getSuccessor(1) == TrueDest) 2118 Opc = Instruction::Or, InvertPredCond = true; 2119 else 2120 continue; 2121 } else { 2122 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 2123 continue; 2124 } 2125 2126 // Ensure that any values used in the bonus instruction are also used 2127 // by the terminator of the predecessor. This means that those values 2128 // must already have been resolved, so we won't be inhibiting the 2129 // out-of-order core by speculating them earlier. 2130 if (BonusInst) { 2131 // Collect the values used by the bonus inst 2132 SmallPtrSet<Value*, 4> UsedValues; 2133 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2134 OE = BonusInst->op_end(); OI != OE; ++OI) { 2135 Value *V = *OI; 2136 if (!isa<Constant>(V)) 2137 UsedValues.insert(V); 2138 } 2139 2140 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2141 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2142 2143 // Walk up to four levels back up the use-def chain of the predecessor's 2144 // terminator to see if all those values were used. The choice of four 2145 // levels is arbitrary, to provide a compile-time-cost bound. 2146 while (!Worklist.empty()) { 2147 std::pair<Value*, unsigned> Pair = Worklist.back(); 2148 Worklist.pop_back(); 2149 2150 if (Pair.second >= 4) continue; 2151 UsedValues.erase(Pair.first); 2152 if (UsedValues.empty()) break; 2153 2154 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2155 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2156 OI != OE; ++OI) 2157 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2158 } 2159 } 2160 2161 if (!UsedValues.empty()) return false; 2162 } 2163 2164 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2165 IRBuilder<> Builder(PBI); 2166 2167 // If we need to invert the condition in the pred block to match, do so now. 2168 if (InvertPredCond) { 2169 Value *NewCond = PBI->getCondition(); 2170 2171 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2172 CmpInst *CI = cast<CmpInst>(NewCond); 2173 CI->setPredicate(CI->getInversePredicate()); 2174 } else { 2175 NewCond = Builder.CreateNot(NewCond, 2176 PBI->getCondition()->getName()+".not"); 2177 } 2178 2179 PBI->setCondition(NewCond); 2180 PBI->swapSuccessors(); 2181 } 2182 2183 // If we have a bonus inst, clone it into the predecessor block. 2184 Instruction *NewBonus = 0; 2185 if (BonusInst) { 2186 NewBonus = BonusInst->clone(); 2187 PredBlock->getInstList().insert(PBI, NewBonus); 2188 NewBonus->takeName(BonusInst); 2189 BonusInst->setName(BonusInst->getName()+".old"); 2190 } 2191 2192 // Clone Cond into the predecessor basic block, and or/and the 2193 // two conditions together. 2194 Instruction *New = Cond->clone(); 2195 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2196 PredBlock->getInstList().insert(PBI, New); 2197 New->takeName(Cond); 2198 Cond->setName(New->getName()+".old"); 2199 2200 if (BI->isConditional()) { 2201 Instruction *NewCond = 2202 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2203 New, "or.cond")); 2204 PBI->setCondition(NewCond); 2205 2206 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2207 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2208 PredFalseWeight); 2209 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2210 SuccFalseWeight); 2211 SmallVector<uint64_t, 8> NewWeights; 2212 2213 if (PBI->getSuccessor(0) == BB) { 2214 if (PredHasWeights && SuccHasWeights) { 2215 // PBI: br i1 %x, BB, FalseDest 2216 // BI: br i1 %y, TrueDest, FalseDest 2217 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2218 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2219 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2220 // TrueWeight for PBI * FalseWeight for BI. 2221 // We assume that total weights of a BranchInst can fit into 32 bits. 2222 // Therefore, we will not have overflow using 64-bit arithmetic. 2223 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2224 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2225 } 2226 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2227 PBI->setSuccessor(0, TrueDest); 2228 } 2229 if (PBI->getSuccessor(1) == BB) { 2230 if (PredHasWeights && SuccHasWeights) { 2231 // PBI: br i1 %x, TrueDest, BB 2232 // BI: br i1 %y, TrueDest, FalseDest 2233 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2234 // FalseWeight for PBI * TrueWeight for BI. 2235 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2236 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2237 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2238 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2239 } 2240 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2241 PBI->setSuccessor(1, FalseDest); 2242 } 2243 if (NewWeights.size() == 2) { 2244 // Halve the weights if any of them cannot fit in an uint32_t 2245 FitWeights(NewWeights); 2246 2247 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2248 PBI->setMetadata(LLVMContext::MD_prof, 2249 MDBuilder(BI->getContext()). 2250 createBranchWeights(MDWeights)); 2251 } else 2252 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2253 } else { 2254 // Update PHI nodes in the common successors. 2255 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2256 ConstantInt *PBI_C = cast<ConstantInt>( 2257 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2258 assert(PBI_C->getType()->isIntegerTy(1)); 2259 Instruction *MergedCond = 0; 2260 if (PBI->getSuccessor(0) == TrueDest) { 2261 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2262 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2263 // is false: !PBI_Cond and BI_Value 2264 Instruction *NotCond = 2265 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2266 "not.cond")); 2267 MergedCond = 2268 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2269 NotCond, New, 2270 "and.cond")); 2271 if (PBI_C->isOne()) 2272 MergedCond = 2273 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2274 PBI->getCondition(), MergedCond, 2275 "or.cond")); 2276 } else { 2277 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2278 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2279 // is false: PBI_Cond and BI_Value 2280 MergedCond = 2281 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2282 PBI->getCondition(), New, 2283 "and.cond")); 2284 if (PBI_C->isOne()) { 2285 Instruction *NotCond = 2286 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2287 "not.cond")); 2288 MergedCond = 2289 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2290 NotCond, MergedCond, 2291 "or.cond")); 2292 } 2293 } 2294 // Update PHI Node. 2295 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2296 MergedCond); 2297 } 2298 // Change PBI from Conditional to Unconditional. 2299 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2300 EraseTerminatorInstAndDCECond(PBI); 2301 PBI = New_PBI; 2302 } 2303 2304 // TODO: If BB is reachable from all paths through PredBlock, then we 2305 // could replace PBI's branch probabilities with BI's. 2306 2307 // Copy any debug value intrinsics into the end of PredBlock. 2308 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2309 if (isa<DbgInfoIntrinsic>(*I)) 2310 I->clone()->insertBefore(PBI); 2311 2312 return true; 2313 } 2314 return false; 2315} 2316 2317/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2318/// predecessor of another block, this function tries to simplify it. We know 2319/// that PBI and BI are both conditional branches, and BI is in one of the 2320/// successor blocks of PBI - PBI branches to BI. 2321static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2322 assert(PBI->isConditional() && BI->isConditional()); 2323 BasicBlock *BB = BI->getParent(); 2324 2325 // If this block ends with a branch instruction, and if there is a 2326 // predecessor that ends on a branch of the same condition, make 2327 // this conditional branch redundant. 2328 if (PBI->getCondition() == BI->getCondition() && 2329 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2330 // Okay, the outcome of this conditional branch is statically 2331 // knowable. If this block had a single pred, handle specially. 2332 if (BB->getSinglePredecessor()) { 2333 // Turn this into a branch on constant. 2334 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2335 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2336 CondIsTrue)); 2337 return true; // Nuke the branch on constant. 2338 } 2339 2340 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2341 // in the constant and simplify the block result. Subsequent passes of 2342 // simplifycfg will thread the block. 2343 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2344 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2345 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2346 std::distance(PB, PE), 2347 BI->getCondition()->getName() + ".pr", 2348 BB->begin()); 2349 // Okay, we're going to insert the PHI node. Since PBI is not the only 2350 // predecessor, compute the PHI'd conditional value for all of the preds. 2351 // Any predecessor where the condition is not computable we keep symbolic. 2352 for (pred_iterator PI = PB; PI != PE; ++PI) { 2353 BasicBlock *P = *PI; 2354 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2355 PBI != BI && PBI->isConditional() && 2356 PBI->getCondition() == BI->getCondition() && 2357 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2358 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2359 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2360 CondIsTrue), P); 2361 } else { 2362 NewPN->addIncoming(BI->getCondition(), P); 2363 } 2364 } 2365 2366 BI->setCondition(NewPN); 2367 return true; 2368 } 2369 } 2370 2371 // If this is a conditional branch in an empty block, and if any 2372 // predecessors is a conditional branch to one of our destinations, 2373 // fold the conditions into logical ops and one cond br. 2374 BasicBlock::iterator BBI = BB->begin(); 2375 // Ignore dbg intrinsics. 2376 while (isa<DbgInfoIntrinsic>(BBI)) 2377 ++BBI; 2378 if (&*BBI != BI) 2379 return false; 2380 2381 2382 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2383 if (CE->canTrap()) 2384 return false; 2385 2386 int PBIOp, BIOp; 2387 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2388 PBIOp = BIOp = 0; 2389 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2390 PBIOp = 0, BIOp = 1; 2391 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2392 PBIOp = 1, BIOp = 0; 2393 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2394 PBIOp = BIOp = 1; 2395 else 2396 return false; 2397 2398 // Check to make sure that the other destination of this branch 2399 // isn't BB itself. If so, this is an infinite loop that will 2400 // keep getting unwound. 2401 if (PBI->getSuccessor(PBIOp) == BB) 2402 return false; 2403 2404 // Do not perform this transformation if it would require 2405 // insertion of a large number of select instructions. For targets 2406 // without predication/cmovs, this is a big pessimization. 2407 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2408 2409 unsigned NumPhis = 0; 2410 for (BasicBlock::iterator II = CommonDest->begin(); 2411 isa<PHINode>(II); ++II, ++NumPhis) 2412 if (NumPhis > 2) // Disable this xform. 2413 return false; 2414 2415 // Finally, if everything is ok, fold the branches to logical ops. 2416 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2417 2418 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2419 << "AND: " << *BI->getParent()); 2420 2421 2422 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2423 // branch in it, where one edge (OtherDest) goes back to itself but the other 2424 // exits. We don't *know* that the program avoids the infinite loop 2425 // (even though that seems likely). If we do this xform naively, we'll end up 2426 // recursively unpeeling the loop. Since we know that (after the xform is 2427 // done) that the block *is* infinite if reached, we just make it an obviously 2428 // infinite loop with no cond branch. 2429 if (OtherDest == BB) { 2430 // Insert it at the end of the function, because it's either code, 2431 // or it won't matter if it's hot. :) 2432 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2433 "infloop", BB->getParent()); 2434 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2435 OtherDest = InfLoopBlock; 2436 } 2437 2438 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2439 2440 // BI may have other predecessors. Because of this, we leave 2441 // it alone, but modify PBI. 2442 2443 // Make sure we get to CommonDest on True&True directions. 2444 Value *PBICond = PBI->getCondition(); 2445 IRBuilder<true, NoFolder> Builder(PBI); 2446 if (PBIOp) 2447 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2448 2449 Value *BICond = BI->getCondition(); 2450 if (BIOp) 2451 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2452 2453 // Merge the conditions. 2454 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2455 2456 // Modify PBI to branch on the new condition to the new dests. 2457 PBI->setCondition(Cond); 2458 PBI->setSuccessor(0, CommonDest); 2459 PBI->setSuccessor(1, OtherDest); 2460 2461 // Update branch weight for PBI. 2462 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2463 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2464 PredFalseWeight); 2465 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2466 SuccFalseWeight); 2467 if (PredHasWeights && SuccHasWeights) { 2468 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2469 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2470 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2471 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2472 // The weight to CommonDest should be PredCommon * SuccTotal + 2473 // PredOther * SuccCommon. 2474 // The weight to OtherDest should be PredOther * SuccOther. 2475 SmallVector<uint64_t, 2> NewWeights; 2476 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2477 PredOther * SuccCommon); 2478 NewWeights.push_back(PredOther * SuccOther); 2479 // Halve the weights if any of them cannot fit in an uint32_t 2480 FitWeights(NewWeights); 2481 2482 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2483 PBI->setMetadata(LLVMContext::MD_prof, 2484 MDBuilder(BI->getContext()). 2485 createBranchWeights(MDWeights)); 2486 } 2487 2488 // OtherDest may have phi nodes. If so, add an entry from PBI's 2489 // block that are identical to the entries for BI's block. 2490 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2491 2492 // We know that the CommonDest already had an edge from PBI to 2493 // it. If it has PHIs though, the PHIs may have different 2494 // entries for BB and PBI's BB. If so, insert a select to make 2495 // them agree. 2496 PHINode *PN; 2497 for (BasicBlock::iterator II = CommonDest->begin(); 2498 (PN = dyn_cast<PHINode>(II)); ++II) { 2499 Value *BIV = PN->getIncomingValueForBlock(BB); 2500 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2501 Value *PBIV = PN->getIncomingValue(PBBIdx); 2502 if (BIV != PBIV) { 2503 // Insert a select in PBI to pick the right value. 2504 Value *NV = cast<SelectInst> 2505 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2506 PN->setIncomingValue(PBBIdx, NV); 2507 } 2508 } 2509 2510 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2511 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2512 2513 // This basic block is probably dead. We know it has at least 2514 // one fewer predecessor. 2515 return true; 2516} 2517 2518// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2519// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2520// Takes care of updating the successors and removing the old terminator. 2521// Also makes sure not to introduce new successors by assuming that edges to 2522// non-successor TrueBBs and FalseBBs aren't reachable. 2523static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2524 BasicBlock *TrueBB, BasicBlock *FalseBB, 2525 uint32_t TrueWeight, 2526 uint32_t FalseWeight){ 2527 // Remove any superfluous successor edges from the CFG. 2528 // First, figure out which successors to preserve. 2529 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2530 // successor. 2531 BasicBlock *KeepEdge1 = TrueBB; 2532 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2533 2534 // Then remove the rest. 2535 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2536 BasicBlock *Succ = OldTerm->getSuccessor(I); 2537 // Make sure only to keep exactly one copy of each edge. 2538 if (Succ == KeepEdge1) 2539 KeepEdge1 = 0; 2540 else if (Succ == KeepEdge2) 2541 KeepEdge2 = 0; 2542 else 2543 Succ->removePredecessor(OldTerm->getParent()); 2544 } 2545 2546 IRBuilder<> Builder(OldTerm); 2547 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2548 2549 // Insert an appropriate new terminator. 2550 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2551 if (TrueBB == FalseBB) 2552 // We were only looking for one successor, and it was present. 2553 // Create an unconditional branch to it. 2554 Builder.CreateBr(TrueBB); 2555 else { 2556 // We found both of the successors we were looking for. 2557 // Create a conditional branch sharing the condition of the select. 2558 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2559 if (TrueWeight != FalseWeight) 2560 NewBI->setMetadata(LLVMContext::MD_prof, 2561 MDBuilder(OldTerm->getContext()). 2562 createBranchWeights(TrueWeight, FalseWeight)); 2563 } 2564 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2565 // Neither of the selected blocks were successors, so this 2566 // terminator must be unreachable. 2567 new UnreachableInst(OldTerm->getContext(), OldTerm); 2568 } else { 2569 // One of the selected values was a successor, but the other wasn't. 2570 // Insert an unconditional branch to the one that was found; 2571 // the edge to the one that wasn't must be unreachable. 2572 if (KeepEdge1 == 0) 2573 // Only TrueBB was found. 2574 Builder.CreateBr(TrueBB); 2575 else 2576 // Only FalseBB was found. 2577 Builder.CreateBr(FalseBB); 2578 } 2579 2580 EraseTerminatorInstAndDCECond(OldTerm); 2581 return true; 2582} 2583 2584// SimplifySwitchOnSelect - Replaces 2585// (switch (select cond, X, Y)) on constant X, Y 2586// with a branch - conditional if X and Y lead to distinct BBs, 2587// unconditional otherwise. 2588static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2589 // Check for constant integer values in the select. 2590 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2591 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2592 if (!TrueVal || !FalseVal) 2593 return false; 2594 2595 // Find the relevant condition and destinations. 2596 Value *Condition = Select->getCondition(); 2597 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2598 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2599 2600 // Get weight for TrueBB and FalseBB. 2601 uint32_t TrueWeight = 0, FalseWeight = 0; 2602 SmallVector<uint64_t, 8> Weights; 2603 bool HasWeights = HasBranchWeights(SI); 2604 if (HasWeights) { 2605 GetBranchWeights(SI, Weights); 2606 if (Weights.size() == 1 + SI->getNumCases()) { 2607 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2608 getSuccessorIndex()]; 2609 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2610 getSuccessorIndex()]; 2611 } 2612 } 2613 2614 // Perform the actual simplification. 2615 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2616 TrueWeight, FalseWeight); 2617} 2618 2619// SimplifyIndirectBrOnSelect - Replaces 2620// (indirectbr (select cond, blockaddress(@fn, BlockA), 2621// blockaddress(@fn, BlockB))) 2622// with 2623// (br cond, BlockA, BlockB). 2624static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2625 // Check that both operands of the select are block addresses. 2626 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2627 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2628 if (!TBA || !FBA) 2629 return false; 2630 2631 // Extract the actual blocks. 2632 BasicBlock *TrueBB = TBA->getBasicBlock(); 2633 BasicBlock *FalseBB = FBA->getBasicBlock(); 2634 2635 // Perform the actual simplification. 2636 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2637 0, 0); 2638} 2639 2640/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2641/// instruction (a seteq/setne with a constant) as the only instruction in a 2642/// block that ends with an uncond branch. We are looking for a very specific 2643/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2644/// this case, we merge the first two "or's of icmp" into a switch, but then the 2645/// default value goes to an uncond block with a seteq in it, we get something 2646/// like: 2647/// 2648/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2649/// DEFAULT: 2650/// %tmp = icmp eq i8 %A, 92 2651/// br label %end 2652/// end: 2653/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2654/// 2655/// We prefer to split the edge to 'end' so that there is a true/false entry to 2656/// the PHI, merging the third icmp into the switch. 2657static bool TryToSimplifyUncondBranchWithICmpInIt( 2658 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI, 2659 const DataLayout *TD) { 2660 BasicBlock *BB = ICI->getParent(); 2661 2662 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2663 // complex. 2664 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2665 2666 Value *V = ICI->getOperand(0); 2667 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2668 2669 // The pattern we're looking for is where our only predecessor is a switch on 2670 // 'V' and this block is the default case for the switch. In this case we can 2671 // fold the compared value into the switch to simplify things. 2672 BasicBlock *Pred = BB->getSinglePredecessor(); 2673 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2674 2675 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2676 if (SI->getCondition() != V) 2677 return false; 2678 2679 // If BB is reachable on a non-default case, then we simply know the value of 2680 // V in this block. Substitute it and constant fold the icmp instruction 2681 // away. 2682 if (SI->getDefaultDest() != BB) { 2683 ConstantInt *VVal = SI->findCaseDest(BB); 2684 assert(VVal && "Should have a unique destination value"); 2685 ICI->setOperand(0, VVal); 2686 2687 if (Value *V = SimplifyInstruction(ICI, TD)) { 2688 ICI->replaceAllUsesWith(V); 2689 ICI->eraseFromParent(); 2690 } 2691 // BB is now empty, so it is likely to simplify away. 2692 return SimplifyCFG(BB, TTI, TD) | true; 2693 } 2694 2695 // Ok, the block is reachable from the default dest. If the constant we're 2696 // comparing exists in one of the other edges, then we can constant fold ICI 2697 // and zap it. 2698 if (SI->findCaseValue(Cst) != SI->case_default()) { 2699 Value *V; 2700 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2701 V = ConstantInt::getFalse(BB->getContext()); 2702 else 2703 V = ConstantInt::getTrue(BB->getContext()); 2704 2705 ICI->replaceAllUsesWith(V); 2706 ICI->eraseFromParent(); 2707 // BB is now empty, so it is likely to simplify away. 2708 return SimplifyCFG(BB, TTI, TD) | true; 2709 } 2710 2711 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2712 // the block. 2713 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2714 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2715 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2716 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2717 return false; 2718 2719 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2720 // true in the PHI. 2721 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2722 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2723 2724 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2725 std::swap(DefaultCst, NewCst); 2726 2727 // Replace ICI (which is used by the PHI for the default value) with true or 2728 // false depending on if it is EQ or NE. 2729 ICI->replaceAllUsesWith(DefaultCst); 2730 ICI->eraseFromParent(); 2731 2732 // Okay, the switch goes to this block on a default value. Add an edge from 2733 // the switch to the merge point on the compared value. 2734 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2735 BB->getParent(), BB); 2736 SmallVector<uint64_t, 8> Weights; 2737 bool HasWeights = HasBranchWeights(SI); 2738 if (HasWeights) { 2739 GetBranchWeights(SI, Weights); 2740 if (Weights.size() == 1 + SI->getNumCases()) { 2741 // Split weight for default case to case for "Cst". 2742 Weights[0] = (Weights[0]+1) >> 1; 2743 Weights.push_back(Weights[0]); 2744 2745 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2746 SI->setMetadata(LLVMContext::MD_prof, 2747 MDBuilder(SI->getContext()). 2748 createBranchWeights(MDWeights)); 2749 } 2750 } 2751 SI->addCase(Cst, NewBB); 2752 2753 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2754 Builder.SetInsertPoint(NewBB); 2755 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2756 Builder.CreateBr(SuccBlock); 2757 PHIUse->addIncoming(NewCst, NewBB); 2758 return true; 2759} 2760 2761/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2762/// Check to see if it is branching on an or/and chain of icmp instructions, and 2763/// fold it into a switch instruction if so. 2764static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2765 IRBuilder<> &Builder) { 2766 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2767 if (Cond == 0) return false; 2768 2769 2770 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2771 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2772 // 'setne's and'ed together, collect them. 2773 Value *CompVal = 0; 2774 std::vector<ConstantInt*> Values; 2775 bool TrueWhenEqual = true; 2776 Value *ExtraCase = 0; 2777 unsigned UsedICmps = 0; 2778 2779 if (Cond->getOpcode() == Instruction::Or) { 2780 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2781 UsedICmps); 2782 } else if (Cond->getOpcode() == Instruction::And) { 2783 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2784 UsedICmps); 2785 TrueWhenEqual = false; 2786 } 2787 2788 // If we didn't have a multiply compared value, fail. 2789 if (CompVal == 0) return false; 2790 2791 // Avoid turning single icmps into a switch. 2792 if (UsedICmps <= 1) 2793 return false; 2794 2795 // There might be duplicate constants in the list, which the switch 2796 // instruction can't handle, remove them now. 2797 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2798 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2799 2800 // If Extra was used, we require at least two switch values to do the 2801 // transformation. A switch with one value is just an cond branch. 2802 if (ExtraCase && Values.size() < 2) return false; 2803 2804 // TODO: Preserve branch weight metadata, similarly to how 2805 // FoldValueComparisonIntoPredecessors preserves it. 2806 2807 // Figure out which block is which destination. 2808 BasicBlock *DefaultBB = BI->getSuccessor(1); 2809 BasicBlock *EdgeBB = BI->getSuccessor(0); 2810 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2811 2812 BasicBlock *BB = BI->getParent(); 2813 2814 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2815 << " cases into SWITCH. BB is:\n" << *BB); 2816 2817 // If there are any extra values that couldn't be folded into the switch 2818 // then we evaluate them with an explicit branch first. Split the block 2819 // right before the condbr to handle it. 2820 if (ExtraCase) { 2821 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2822 // Remove the uncond branch added to the old block. 2823 TerminatorInst *OldTI = BB->getTerminator(); 2824 Builder.SetInsertPoint(OldTI); 2825 2826 if (TrueWhenEqual) 2827 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2828 else 2829 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2830 2831 OldTI->eraseFromParent(); 2832 2833 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2834 // for the edge we just added. 2835 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2836 2837 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2838 << "\nEXTRABB = " << *BB); 2839 BB = NewBB; 2840 } 2841 2842 Builder.SetInsertPoint(BI); 2843 // Convert pointer to int before we switch. 2844 if (CompVal->getType()->isPointerTy()) { 2845 assert(TD && "Cannot switch on pointer without DataLayout"); 2846 CompVal = Builder.CreatePtrToInt(CompVal, 2847 TD->getIntPtrType(CompVal->getContext()), 2848 "magicptr"); 2849 } 2850 2851 // Create the new switch instruction now. 2852 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2853 2854 // Add all of the 'cases' to the switch instruction. 2855 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2856 New->addCase(Values[i], EdgeBB); 2857 2858 // We added edges from PI to the EdgeBB. As such, if there were any 2859 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2860 // the number of edges added. 2861 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2862 isa<PHINode>(BBI); ++BBI) { 2863 PHINode *PN = cast<PHINode>(BBI); 2864 Value *InVal = PN->getIncomingValueForBlock(BB); 2865 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2866 PN->addIncoming(InVal, BB); 2867 } 2868 2869 // Erase the old branch instruction. 2870 EraseTerminatorInstAndDCECond(BI); 2871 2872 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2873 return true; 2874} 2875 2876bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2877 // If this is a trivial landing pad that just continues unwinding the caught 2878 // exception then zap the landing pad, turning its invokes into calls. 2879 BasicBlock *BB = RI->getParent(); 2880 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2881 if (RI->getValue() != LPInst) 2882 // Not a landing pad, or the resume is not unwinding the exception that 2883 // caused control to branch here. 2884 return false; 2885 2886 // Check that there are no other instructions except for debug intrinsics. 2887 BasicBlock::iterator I = LPInst, E = RI; 2888 while (++I != E) 2889 if (!isa<DbgInfoIntrinsic>(I)) 2890 return false; 2891 2892 // Turn all invokes that unwind here into calls and delete the basic block. 2893 bool InvokeRequiresTableEntry = false; 2894 bool Changed = false; 2895 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2896 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2897 2898 if (II->hasFnAttr(Attribute::UWTable)) { 2899 // Don't remove an `invoke' instruction if the ABI requires an entry into 2900 // the table. 2901 InvokeRequiresTableEntry = true; 2902 continue; 2903 } 2904 2905 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2906 2907 // Insert a call instruction before the invoke. 2908 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2909 Call->takeName(II); 2910 Call->setCallingConv(II->getCallingConv()); 2911 Call->setAttributes(II->getAttributes()); 2912 Call->setDebugLoc(II->getDebugLoc()); 2913 2914 // Anything that used the value produced by the invoke instruction now uses 2915 // the value produced by the call instruction. Note that we do this even 2916 // for void functions and calls with no uses so that the callgraph edge is 2917 // updated. 2918 II->replaceAllUsesWith(Call); 2919 BB->removePredecessor(II->getParent()); 2920 2921 // Insert a branch to the normal destination right before the invoke. 2922 BranchInst::Create(II->getNormalDest(), II); 2923 2924 // Finally, delete the invoke instruction! 2925 II->eraseFromParent(); 2926 Changed = true; 2927 } 2928 2929 if (!InvokeRequiresTableEntry) 2930 // The landingpad is now unreachable. Zap it. 2931 BB->eraseFromParent(); 2932 2933 return Changed; 2934} 2935 2936bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2937 BasicBlock *BB = RI->getParent(); 2938 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2939 2940 // Find predecessors that end with branches. 2941 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2942 SmallVector<BranchInst*, 8> CondBranchPreds; 2943 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2944 BasicBlock *P = *PI; 2945 TerminatorInst *PTI = P->getTerminator(); 2946 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2947 if (BI->isUnconditional()) 2948 UncondBranchPreds.push_back(P); 2949 else 2950 CondBranchPreds.push_back(BI); 2951 } 2952 } 2953 2954 // If we found some, do the transformation! 2955 if (!UncondBranchPreds.empty() && DupRet) { 2956 while (!UncondBranchPreds.empty()) { 2957 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2958 DEBUG(dbgs() << "FOLDING: " << *BB 2959 << "INTO UNCOND BRANCH PRED: " << *Pred); 2960 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2961 } 2962 2963 // If we eliminated all predecessors of the block, delete the block now. 2964 if (pred_begin(BB) == pred_end(BB)) 2965 // We know there are no successors, so just nuke the block. 2966 BB->eraseFromParent(); 2967 2968 return true; 2969 } 2970 2971 // Check out all of the conditional branches going to this return 2972 // instruction. If any of them just select between returns, change the 2973 // branch itself into a select/return pair. 2974 while (!CondBranchPreds.empty()) { 2975 BranchInst *BI = CondBranchPreds.pop_back_val(); 2976 2977 // Check to see if the non-BB successor is also a return block. 2978 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2979 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2980 SimplifyCondBranchToTwoReturns(BI, Builder)) 2981 return true; 2982 } 2983 return false; 2984} 2985 2986bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2987 BasicBlock *BB = UI->getParent(); 2988 2989 bool Changed = false; 2990 2991 // If there are any instructions immediately before the unreachable that can 2992 // be removed, do so. 2993 while (UI != BB->begin()) { 2994 BasicBlock::iterator BBI = UI; 2995 --BBI; 2996 // Do not delete instructions that can have side effects which might cause 2997 // the unreachable to not be reachable; specifically, calls and volatile 2998 // operations may have this effect. 2999 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 3000 3001 if (BBI->mayHaveSideEffects()) { 3002 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 3003 if (SI->isVolatile()) 3004 break; 3005 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 3006 if (LI->isVolatile()) 3007 break; 3008 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 3009 if (RMWI->isVolatile()) 3010 break; 3011 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 3012 if (CXI->isVolatile()) 3013 break; 3014 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 3015 !isa<LandingPadInst>(BBI)) { 3016 break; 3017 } 3018 // Note that deleting LandingPad's here is in fact okay, although it 3019 // involves a bit of subtle reasoning. If this inst is a LandingPad, 3020 // all the predecessors of this block will be the unwind edges of Invokes, 3021 // and we can therefore guarantee this block will be erased. 3022 } 3023 3024 // Delete this instruction (any uses are guaranteed to be dead) 3025 if (!BBI->use_empty()) 3026 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 3027 BBI->eraseFromParent(); 3028 Changed = true; 3029 } 3030 3031 // If the unreachable instruction is the first in the block, take a gander 3032 // at all of the predecessors of this instruction, and simplify them. 3033 if (&BB->front() != UI) return Changed; 3034 3035 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 3036 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 3037 TerminatorInst *TI = Preds[i]->getTerminator(); 3038 IRBuilder<> Builder(TI); 3039 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 3040 if (BI->isUnconditional()) { 3041 if (BI->getSuccessor(0) == BB) { 3042 new UnreachableInst(TI->getContext(), TI); 3043 TI->eraseFromParent(); 3044 Changed = true; 3045 } 3046 } else { 3047 if (BI->getSuccessor(0) == BB) { 3048 Builder.CreateBr(BI->getSuccessor(1)); 3049 EraseTerminatorInstAndDCECond(BI); 3050 } else if (BI->getSuccessor(1) == BB) { 3051 Builder.CreateBr(BI->getSuccessor(0)); 3052 EraseTerminatorInstAndDCECond(BI); 3053 Changed = true; 3054 } 3055 } 3056 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 3057 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3058 i != e; ++i) 3059 if (i.getCaseSuccessor() == BB) { 3060 BB->removePredecessor(SI->getParent()); 3061 SI->removeCase(i); 3062 --i; --e; 3063 Changed = true; 3064 } 3065 // If the default value is unreachable, figure out the most popular 3066 // destination and make it the default. 3067 if (SI->getDefaultDest() == BB) { 3068 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 3069 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3070 i != e; ++i) { 3071 std::pair<unsigned, unsigned> &entry = 3072 Popularity[i.getCaseSuccessor()]; 3073 if (entry.first == 0) { 3074 entry.first = 1; 3075 entry.second = i.getCaseIndex(); 3076 } else { 3077 entry.first++; 3078 } 3079 } 3080 3081 // Find the most popular block. 3082 unsigned MaxPop = 0; 3083 unsigned MaxIndex = 0; 3084 BasicBlock *MaxBlock = 0; 3085 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 3086 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 3087 if (I->second.first > MaxPop || 3088 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 3089 MaxPop = I->second.first; 3090 MaxIndex = I->second.second; 3091 MaxBlock = I->first; 3092 } 3093 } 3094 if (MaxBlock) { 3095 // Make this the new default, allowing us to delete any explicit 3096 // edges to it. 3097 SI->setDefaultDest(MaxBlock); 3098 Changed = true; 3099 3100 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 3101 // it. 3102 if (isa<PHINode>(MaxBlock->begin())) 3103 for (unsigned i = 0; i != MaxPop-1; ++i) 3104 MaxBlock->removePredecessor(SI->getParent()); 3105 3106 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 3107 i != e; ++i) 3108 if (i.getCaseSuccessor() == MaxBlock) { 3109 SI->removeCase(i); 3110 --i; --e; 3111 } 3112 } 3113 } 3114 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 3115 if (II->getUnwindDest() == BB) { 3116 // Convert the invoke to a call instruction. This would be a good 3117 // place to note that the call does not throw though. 3118 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 3119 II->removeFromParent(); // Take out of symbol table 3120 3121 // Insert the call now... 3122 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 3123 Builder.SetInsertPoint(BI); 3124 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 3125 Args, II->getName()); 3126 CI->setCallingConv(II->getCallingConv()); 3127 CI->setAttributes(II->getAttributes()); 3128 // If the invoke produced a value, the call does now instead. 3129 II->replaceAllUsesWith(CI); 3130 delete II; 3131 Changed = true; 3132 } 3133 } 3134 } 3135 3136 // If this block is now dead, remove it. 3137 if (pred_begin(BB) == pred_end(BB) && 3138 BB != &BB->getParent()->getEntryBlock()) { 3139 // We know there are no successors, so just nuke the block. 3140 BB->eraseFromParent(); 3141 return true; 3142 } 3143 3144 return Changed; 3145} 3146 3147/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3148/// integer range comparison into a sub, an icmp and a branch. 3149static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3150 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3151 3152 // Make sure all cases point to the same destination and gather the values. 3153 SmallVector<ConstantInt *, 16> Cases; 3154 SwitchInst::CaseIt I = SI->case_begin(); 3155 Cases.push_back(I.getCaseValue()); 3156 SwitchInst::CaseIt PrevI = I++; 3157 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3158 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3159 return false; 3160 Cases.push_back(I.getCaseValue()); 3161 } 3162 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3163 3164 // Sort the case values, then check if they form a range we can transform. 3165 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3166 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3167 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3168 return false; 3169 } 3170 3171 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3172 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3173 3174 Value *Sub = SI->getCondition(); 3175 if (!Offset->isNullValue()) 3176 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3177 Value *Cmp; 3178 // If NumCases overflowed, then all possible values jump to the successor. 3179 if (NumCases->isNullValue() && SI->getNumCases() != 0) 3180 Cmp = ConstantInt::getTrue(SI->getContext()); 3181 else 3182 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3183 BranchInst *NewBI = Builder.CreateCondBr( 3184 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3185 3186 // Update weight for the newly-created conditional branch. 3187 SmallVector<uint64_t, 8> Weights; 3188 bool HasWeights = HasBranchWeights(SI); 3189 if (HasWeights) { 3190 GetBranchWeights(SI, Weights); 3191 if (Weights.size() == 1 + SI->getNumCases()) { 3192 // Combine all weights for the cases to be the true weight of NewBI. 3193 // We assume that the sum of all weights for a Terminator can fit into 32 3194 // bits. 3195 uint32_t NewTrueWeight = 0; 3196 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3197 NewTrueWeight += (uint32_t)Weights[I]; 3198 NewBI->setMetadata(LLVMContext::MD_prof, 3199 MDBuilder(SI->getContext()). 3200 createBranchWeights(NewTrueWeight, 3201 (uint32_t)Weights[0])); 3202 } 3203 } 3204 3205 // Prune obsolete incoming values off the successor's PHI nodes. 3206 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3207 isa<PHINode>(BBI); ++BBI) { 3208 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3209 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3210 } 3211 SI->eraseFromParent(); 3212 3213 return true; 3214} 3215 3216/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3217/// and use it to remove dead cases. 3218static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3219 Value *Cond = SI->getCondition(); 3220 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3221 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3222 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3223 3224 // Gather dead cases. 3225 SmallVector<ConstantInt*, 8> DeadCases; 3226 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3227 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3228 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3229 DeadCases.push_back(I.getCaseValue()); 3230 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3231 << I.getCaseValue() << "' is dead.\n"); 3232 } 3233 } 3234 3235 SmallVector<uint64_t, 8> Weights; 3236 bool HasWeight = HasBranchWeights(SI); 3237 if (HasWeight) { 3238 GetBranchWeights(SI, Weights); 3239 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3240 } 3241 3242 // Remove dead cases from the switch. 3243 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3244 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3245 assert(Case != SI->case_default() && 3246 "Case was not found. Probably mistake in DeadCases forming."); 3247 if (HasWeight) { 3248 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3249 Weights.pop_back(); 3250 } 3251 3252 // Prune unused values from PHI nodes. 3253 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3254 SI->removeCase(Case); 3255 } 3256 if (HasWeight) { 3257 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3258 SI->setMetadata(LLVMContext::MD_prof, 3259 MDBuilder(SI->getParent()->getContext()). 3260 createBranchWeights(MDWeights)); 3261 } 3262 3263 return !DeadCases.empty(); 3264} 3265 3266/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3267/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3268/// by an unconditional branch), look at the phi node for BB in the successor 3269/// block and see if the incoming value is equal to CaseValue. If so, return 3270/// the phi node, and set PhiIndex to BB's index in the phi node. 3271static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3272 BasicBlock *BB, 3273 int *PhiIndex) { 3274 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3275 return NULL; // BB must be empty to be a candidate for simplification. 3276 if (!BB->getSinglePredecessor()) 3277 return NULL; // BB must be dominated by the switch. 3278 3279 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3280 if (!Branch || !Branch->isUnconditional()) 3281 return NULL; // Terminator must be unconditional branch. 3282 3283 BasicBlock *Succ = Branch->getSuccessor(0); 3284 3285 BasicBlock::iterator I = Succ->begin(); 3286 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3287 int Idx = PHI->getBasicBlockIndex(BB); 3288 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3289 3290 Value *InValue = PHI->getIncomingValue(Idx); 3291 if (InValue != CaseValue) continue; 3292 3293 *PhiIndex = Idx; 3294 return PHI; 3295 } 3296 3297 return NULL; 3298} 3299 3300/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3301/// instruction to a phi node dominated by the switch, if that would mean that 3302/// some of the destination blocks of the switch can be folded away. 3303/// Returns true if a change is made. 3304static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3305 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3306 ForwardingNodesMap ForwardingNodes; 3307 3308 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3309 ConstantInt *CaseValue = I.getCaseValue(); 3310 BasicBlock *CaseDest = I.getCaseSuccessor(); 3311 3312 int PhiIndex; 3313 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3314 &PhiIndex); 3315 if (!PHI) continue; 3316 3317 ForwardingNodes[PHI].push_back(PhiIndex); 3318 } 3319 3320 bool Changed = false; 3321 3322 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3323 E = ForwardingNodes.end(); I != E; ++I) { 3324 PHINode *Phi = I->first; 3325 SmallVector<int,4> &Indexes = I->second; 3326 3327 if (Indexes.size() < 2) continue; 3328 3329 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3330 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3331 Changed = true; 3332 } 3333 3334 return Changed; 3335} 3336 3337/// ValidLookupTableConstant - Return true if the backend will be able to handle 3338/// initializing an array of constants like C. 3339static bool ValidLookupTableConstant(Constant *C) { 3340 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3341 return CE->isGEPWithNoNotionalOverIndexing(); 3342 3343 return isa<ConstantFP>(C) || 3344 isa<ConstantInt>(C) || 3345 isa<ConstantPointerNull>(C) || 3346 isa<GlobalValue>(C) || 3347 isa<UndefValue>(C); 3348} 3349 3350/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3351/// its constant value in ConstantPool, returning 0 if it's not there. 3352static Constant *LookupConstant(Value *V, 3353 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3354 if (Constant *C = dyn_cast<Constant>(V)) 3355 return C; 3356 return ConstantPool.lookup(V); 3357} 3358 3359/// ConstantFold - Try to fold instruction I into a constant. This works for 3360/// simple instructions such as binary operations where both operands are 3361/// constant or can be replaced by constants from the ConstantPool. Returns the 3362/// resulting constant on success, 0 otherwise. 3363static Constant *ConstantFold(Instruction *I, 3364 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3365 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3366 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3367 if (!A) 3368 return 0; 3369 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3370 if (!B) 3371 return 0; 3372 return ConstantExpr::get(BO->getOpcode(), A, B); 3373 } 3374 3375 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3376 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3377 if (!A) 3378 return 0; 3379 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3380 if (!B) 3381 return 0; 3382 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3383 } 3384 3385 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3386 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3387 if (!A) 3388 return 0; 3389 if (A->isAllOnesValue()) 3390 return LookupConstant(Select->getTrueValue(), ConstantPool); 3391 if (A->isNullValue()) 3392 return LookupConstant(Select->getFalseValue(), ConstantPool); 3393 return 0; 3394 } 3395 3396 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3397 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3398 if (!A) 3399 return 0; 3400 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3401 } 3402 3403 return 0; 3404} 3405 3406/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3407/// at the common destination basic block, *CommonDest, for one of the case 3408/// destionations CaseDest corresponding to value CaseVal (0 for the default 3409/// case), of a switch instruction SI. 3410static bool GetCaseResults(SwitchInst *SI, 3411 ConstantInt *CaseVal, 3412 BasicBlock *CaseDest, 3413 BasicBlock **CommonDest, 3414 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3415 // The block from which we enter the common destination. 3416 BasicBlock *Pred = SI->getParent(); 3417 3418 // If CaseDest is empty except for some side-effect free instructions through 3419 // which we can constant-propagate the CaseVal, continue to its successor. 3420 SmallDenseMap<Value*, Constant*> ConstantPool; 3421 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3422 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3423 ++I) { 3424 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3425 // If the terminator is a simple branch, continue to the next block. 3426 if (T->getNumSuccessors() != 1) 3427 return false; 3428 Pred = CaseDest; 3429 CaseDest = T->getSuccessor(0); 3430 } else if (isa<DbgInfoIntrinsic>(I)) { 3431 // Skip debug intrinsic. 3432 continue; 3433 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3434 // Instruction is side-effect free and constant. 3435 ConstantPool.insert(std::make_pair(I, C)); 3436 } else { 3437 break; 3438 } 3439 } 3440 3441 // If we did not have a CommonDest before, use the current one. 3442 if (!*CommonDest) 3443 *CommonDest = CaseDest; 3444 // If the destination isn't the common one, abort. 3445 if (CaseDest != *CommonDest) 3446 return false; 3447 3448 // Get the values for this case from phi nodes in the destination block. 3449 BasicBlock::iterator I = (*CommonDest)->begin(); 3450 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3451 int Idx = PHI->getBasicBlockIndex(Pred); 3452 if (Idx == -1) 3453 continue; 3454 3455 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3456 ConstantPool); 3457 if (!ConstVal) 3458 return false; 3459 3460 // Note: If the constant comes from constant-propagating the case value 3461 // through the CaseDest basic block, it will be safe to remove the 3462 // instructions in that block. They cannot be used (except in the phi nodes 3463 // we visit) outside CaseDest, because that block does not dominate its 3464 // successor. If it did, we would not be in this phi node. 3465 3466 // Be conservative about which kinds of constants we support. 3467 if (!ValidLookupTableConstant(ConstVal)) 3468 return false; 3469 3470 Res.push_back(std::make_pair(PHI, ConstVal)); 3471 } 3472 3473 return true; 3474} 3475 3476namespace { 3477 /// SwitchLookupTable - This class represents a lookup table that can be used 3478 /// to replace a switch. 3479 class SwitchLookupTable { 3480 public: 3481 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3482 /// with the contents of Values, using DefaultValue to fill any holes in the 3483 /// table. 3484 SwitchLookupTable(Module &M, 3485 uint64_t TableSize, 3486 ConstantInt *Offset, 3487 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3488 Constant *DefaultValue, 3489 const DataLayout *TD); 3490 3491 /// BuildLookup - Build instructions with Builder to retrieve the value at 3492 /// the position given by Index in the lookup table. 3493 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3494 3495 /// WouldFitInRegister - Return true if a table with TableSize elements of 3496 /// type ElementType would fit in a target-legal register. 3497 static bool WouldFitInRegister(const DataLayout *TD, 3498 uint64_t TableSize, 3499 const Type *ElementType); 3500 3501 private: 3502 // Depending on the contents of the table, it can be represented in 3503 // different ways. 3504 enum { 3505 // For tables where each element contains the same value, we just have to 3506 // store that single value and return it for each lookup. 3507 SingleValueKind, 3508 3509 // For small tables with integer elements, we can pack them into a bitmap 3510 // that fits into a target-legal register. Values are retrieved by 3511 // shift and mask operations. 3512 BitMapKind, 3513 3514 // The table is stored as an array of values. Values are retrieved by load 3515 // instructions from the table. 3516 ArrayKind 3517 } Kind; 3518 3519 // For SingleValueKind, this is the single value. 3520 Constant *SingleValue; 3521 3522 // For BitMapKind, this is the bitmap. 3523 ConstantInt *BitMap; 3524 IntegerType *BitMapElementTy; 3525 3526 // For ArrayKind, this is the array. 3527 GlobalVariable *Array; 3528 }; 3529} 3530 3531SwitchLookupTable::SwitchLookupTable(Module &M, 3532 uint64_t TableSize, 3533 ConstantInt *Offset, 3534 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3535 Constant *DefaultValue, 3536 const DataLayout *TD) 3537 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { 3538 assert(Values.size() && "Can't build lookup table without values!"); 3539 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3540 3541 // If all values in the table are equal, this is that value. 3542 SingleValue = Values.begin()->second; 3543 3544 // Build up the table contents. 3545 SmallVector<Constant*, 64> TableContents(TableSize); 3546 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3547 ConstantInt *CaseVal = Values[I].first; 3548 Constant *CaseRes = Values[I].second; 3549 assert(CaseRes->getType() == DefaultValue->getType()); 3550 3551 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3552 .getLimitedValue(); 3553 TableContents[Idx] = CaseRes; 3554 3555 if (CaseRes != SingleValue) 3556 SingleValue = 0; 3557 } 3558 3559 // Fill in any holes in the table with the default result. 3560 if (Values.size() < TableSize) { 3561 for (uint64_t I = 0; I < TableSize; ++I) { 3562 if (!TableContents[I]) 3563 TableContents[I] = DefaultValue; 3564 } 3565 3566 if (DefaultValue != SingleValue) 3567 SingleValue = 0; 3568 } 3569 3570 // If each element in the table contains the same value, we only need to store 3571 // that single value. 3572 if (SingleValue) { 3573 Kind = SingleValueKind; 3574 return; 3575 } 3576 3577 // If the type is integer and the table fits in a register, build a bitmap. 3578 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3579 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3580 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3581 for (uint64_t I = TableSize; I > 0; --I) { 3582 TableInt <<= IT->getBitWidth(); 3583 // Insert values into the bitmap. Undef values are set to zero. 3584 if (!isa<UndefValue>(TableContents[I - 1])) { 3585 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3586 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3587 } 3588 } 3589 BitMap = ConstantInt::get(M.getContext(), TableInt); 3590 BitMapElementTy = IT; 3591 Kind = BitMapKind; 3592 ++NumBitMaps; 3593 return; 3594 } 3595 3596 // Store the table in an array. 3597 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3598 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3599 3600 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3601 GlobalVariable::PrivateLinkage, 3602 Initializer, 3603 "switch.table"); 3604 Array->setUnnamedAddr(true); 3605 Kind = ArrayKind; 3606} 3607 3608Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3609 switch (Kind) { 3610 case SingleValueKind: 3611 return SingleValue; 3612 case BitMapKind: { 3613 // Type of the bitmap (e.g. i59). 3614 IntegerType *MapTy = BitMap->getType(); 3615 3616 // Cast Index to the same type as the bitmap. 3617 // Note: The Index is <= the number of elements in the table, so 3618 // truncating it to the width of the bitmask is safe. 3619 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3620 3621 // Multiply the shift amount by the element width. 3622 ShiftAmt = Builder.CreateMul(ShiftAmt, 3623 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3624 "switch.shiftamt"); 3625 3626 // Shift down. 3627 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3628 "switch.downshift"); 3629 // Mask off. 3630 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3631 "switch.masked"); 3632 } 3633 case ArrayKind: { 3634 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3635 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3636 "switch.gep"); 3637 return Builder.CreateLoad(GEP, "switch.load"); 3638 } 3639 } 3640 llvm_unreachable("Unknown lookup table kind!"); 3641} 3642 3643bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3644 uint64_t TableSize, 3645 const Type *ElementType) { 3646 if (!TD) 3647 return false; 3648 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3649 if (!IT) 3650 return false; 3651 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3652 // are <= 15, we could try to narrow the type. 3653 3654 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3655 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3656 return false; 3657 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3658} 3659 3660/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3661/// for this switch, based on the number of caes, size of the table and the 3662/// types of the results. 3663static bool ShouldBuildLookupTable(SwitchInst *SI, 3664 uint64_t TableSize, 3665 const TargetTransformInfo &TTI, 3666 const DataLayout *TD, 3667 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3668 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3669 return false; // TableSize overflowed, or mul below might overflow. 3670 3671 bool AllTablesFitInRegister = true; 3672 bool HasIllegalType = false; 3673 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3674 E = ResultTypes.end(); I != E; ++I) { 3675 Type *Ty = I->second; 3676 3677 // Saturate this flag to true. 3678 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty); 3679 3680 // Saturate this flag to false. 3681 AllTablesFitInRegister = AllTablesFitInRegister && 3682 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty); 3683 3684 // If both flags saturate, we're done. NOTE: This *only* works with 3685 // saturating flags, and all flags have to saturate first due to the 3686 // non-deterministic behavior of iterating over a dense map. 3687 if (HasIllegalType && !AllTablesFitInRegister) 3688 break; 3689 } 3690 3691 // If each table would fit in a register, we should build it anyway. 3692 if (AllTablesFitInRegister) 3693 return true; 3694 3695 // Don't build a table that doesn't fit in-register if it has illegal types. 3696 if (HasIllegalType) 3697 return false; 3698 3699 // The table density should be at least 40%. This is the same criterion as for 3700 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3701 // FIXME: Find the best cut-off. 3702 return SI->getNumCases() * 10 >= TableSize * 4; 3703} 3704 3705/// SwitchToLookupTable - If the switch is only used to initialize one or more 3706/// phi nodes in a common successor block with different constant values, 3707/// replace the switch with lookup tables. 3708static bool SwitchToLookupTable(SwitchInst *SI, 3709 IRBuilder<> &Builder, 3710 const TargetTransformInfo &TTI, 3711 const DataLayout* TD) { 3712 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3713 3714 // Only build lookup table when we have a target that supports it. 3715 if (!TTI.shouldBuildLookupTables()) 3716 return false; 3717 3718 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3719 // split off a dense part and build a lookup table for that. 3720 3721 // FIXME: This creates arrays of GEPs to constant strings, which means each 3722 // GEP needs a runtime relocation in PIC code. We should just build one big 3723 // string and lookup indices into that. 3724 3725 // Ignore the switch if the number of cases is too small. 3726 // This is similar to the check when building jump tables in 3727 // SelectionDAGBuilder::handleJTSwitchCase. 3728 // FIXME: Determine the best cut-off. 3729 if (SI->getNumCases() < 4) 3730 return false; 3731 3732 // Figure out the corresponding result for each case value and phi node in the 3733 // common destination, as well as the the min and max case values. 3734 assert(SI->case_begin() != SI->case_end()); 3735 SwitchInst::CaseIt CI = SI->case_begin(); 3736 ConstantInt *MinCaseVal = CI.getCaseValue(); 3737 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3738 3739 BasicBlock *CommonDest = 0; 3740 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3741 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3742 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3743 SmallDenseMap<PHINode*, Type*> ResultTypes; 3744 SmallVector<PHINode*, 4> PHIs; 3745 3746 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3747 ConstantInt *CaseVal = CI.getCaseValue(); 3748 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3749 MinCaseVal = CaseVal; 3750 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3751 MaxCaseVal = CaseVal; 3752 3753 // Resulting value at phi nodes for this case value. 3754 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3755 ResultsTy Results; 3756 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3757 Results)) 3758 return false; 3759 3760 // Append the result from this case to the list for each phi. 3761 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3762 if (!ResultLists.count(I->first)) 3763 PHIs.push_back(I->first); 3764 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3765 } 3766 } 3767 3768 // Get the resulting values for the default case. 3769 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3770 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3771 DefaultResultsList)) 3772 return false; 3773 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3774 PHINode *PHI = DefaultResultsList[I].first; 3775 Constant *Result = DefaultResultsList[I].second; 3776 DefaultResults[PHI] = Result; 3777 ResultTypes[PHI] = Result->getType(); 3778 } 3779 3780 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3781 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3782 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes)) 3783 return false; 3784 3785 // Create the BB that does the lookups. 3786 Module &Mod = *CommonDest->getParent()->getParent(); 3787 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3788 "switch.lookup", 3789 CommonDest->getParent(), 3790 CommonDest); 3791 3792 // Check whether the condition value is within the case range, and branch to 3793 // the new BB. 3794 Builder.SetInsertPoint(SI); 3795 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3796 "switch.tableidx"); 3797 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3798 MinCaseVal->getType(), TableSize)); 3799 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3800 3801 // Populate the BB that does the lookups. 3802 Builder.SetInsertPoint(LookupBB); 3803 bool ReturnedEarly = false; 3804 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3805 PHINode *PHI = PHIs[I]; 3806 3807 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3808 DefaultResults[PHI], TD); 3809 3810 Value *Result = Table.BuildLookup(TableIndex, Builder); 3811 3812 // If the result is used to return immediately from the function, we want to 3813 // do that right here. 3814 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3815 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3816 Builder.CreateRet(Result); 3817 ReturnedEarly = true; 3818 break; 3819 } 3820 3821 PHI->addIncoming(Result, LookupBB); 3822 } 3823 3824 if (!ReturnedEarly) 3825 Builder.CreateBr(CommonDest); 3826 3827 // Remove the switch. 3828 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3829 BasicBlock *Succ = SI->getSuccessor(i); 3830 if (Succ == SI->getDefaultDest()) continue; 3831 Succ->removePredecessor(SI->getParent()); 3832 } 3833 SI->eraseFromParent(); 3834 3835 ++NumLookupTables; 3836 return true; 3837} 3838 3839bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3840 BasicBlock *BB = SI->getParent(); 3841 3842 if (isValueEqualityComparison(SI)) { 3843 // If we only have one predecessor, and if it is a branch on this value, 3844 // see if that predecessor totally determines the outcome of this switch. 3845 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3846 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3847 return SimplifyCFG(BB, TTI, TD) | true; 3848 3849 Value *Cond = SI->getCondition(); 3850 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3851 if (SimplifySwitchOnSelect(SI, Select)) 3852 return SimplifyCFG(BB, TTI, TD) | true; 3853 3854 // If the block only contains the switch, see if we can fold the block 3855 // away into any preds. 3856 BasicBlock::iterator BBI = BB->begin(); 3857 // Ignore dbg intrinsics. 3858 while (isa<DbgInfoIntrinsic>(BBI)) 3859 ++BBI; 3860 if (SI == &*BBI) 3861 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3862 return SimplifyCFG(BB, TTI, TD) | true; 3863 } 3864 3865 // Try to transform the switch into an icmp and a branch. 3866 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3867 return SimplifyCFG(BB, TTI, TD) | true; 3868 3869 // Remove unreachable cases. 3870 if (EliminateDeadSwitchCases(SI)) 3871 return SimplifyCFG(BB, TTI, TD) | true; 3872 3873 if (ForwardSwitchConditionToPHI(SI)) 3874 return SimplifyCFG(BB, TTI, TD) | true; 3875 3876 if (SwitchToLookupTable(SI, Builder, TTI, TD)) 3877 return SimplifyCFG(BB, TTI, TD) | true; 3878 3879 return false; 3880} 3881 3882bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3883 BasicBlock *BB = IBI->getParent(); 3884 bool Changed = false; 3885 3886 // Eliminate redundant destinations. 3887 SmallPtrSet<Value *, 8> Succs; 3888 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3889 BasicBlock *Dest = IBI->getDestination(i); 3890 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3891 Dest->removePredecessor(BB); 3892 IBI->removeDestination(i); 3893 --i; --e; 3894 Changed = true; 3895 } 3896 } 3897 3898 if (IBI->getNumDestinations() == 0) { 3899 // If the indirectbr has no successors, change it to unreachable. 3900 new UnreachableInst(IBI->getContext(), IBI); 3901 EraseTerminatorInstAndDCECond(IBI); 3902 return true; 3903 } 3904 3905 if (IBI->getNumDestinations() == 1) { 3906 // If the indirectbr has one successor, change it to a direct branch. 3907 BranchInst::Create(IBI->getDestination(0), IBI); 3908 EraseTerminatorInstAndDCECond(IBI); 3909 return true; 3910 } 3911 3912 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3913 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3914 return SimplifyCFG(BB, TTI, TD) | true; 3915 } 3916 return Changed; 3917} 3918 3919bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3920 BasicBlock *BB = BI->getParent(); 3921 3922 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3923 return true; 3924 3925 // If the Terminator is the only non-phi instruction, simplify the block. 3926 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3927 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3928 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3929 return true; 3930 3931 // If the only instruction in the block is a seteq/setne comparison 3932 // against a constant, try to simplify the block. 3933 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3934 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3935 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3936 ; 3937 if (I->isTerminator() && 3938 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD)) 3939 return true; 3940 } 3941 3942 // If this basic block is ONLY a compare and a branch, and if a predecessor 3943 // branches to us and our successor, fold the comparison into the 3944 // predecessor and use logical operations to update the incoming value 3945 // for PHI nodes in common successor. 3946 if (FoldBranchToCommonDest(BI)) 3947 return SimplifyCFG(BB, TTI, TD) | true; 3948 return false; 3949} 3950 3951 3952bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3953 BasicBlock *BB = BI->getParent(); 3954 3955 // Conditional branch 3956 if (isValueEqualityComparison(BI)) { 3957 // If we only have one predecessor, and if it is a branch on this value, 3958 // see if that predecessor totally determines the outcome of this 3959 // switch. 3960 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3961 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3962 return SimplifyCFG(BB, TTI, TD) | true; 3963 3964 // This block must be empty, except for the setcond inst, if it exists. 3965 // Ignore dbg intrinsics. 3966 BasicBlock::iterator I = BB->begin(); 3967 // Ignore dbg intrinsics. 3968 while (isa<DbgInfoIntrinsic>(I)) 3969 ++I; 3970 if (&*I == BI) { 3971 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3972 return SimplifyCFG(BB, TTI, TD) | true; 3973 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3974 ++I; 3975 // Ignore dbg intrinsics. 3976 while (isa<DbgInfoIntrinsic>(I)) 3977 ++I; 3978 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3979 return SimplifyCFG(BB, TTI, TD) | true; 3980 } 3981 } 3982 3983 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3984 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3985 return true; 3986 3987 // If this basic block is ONLY a compare and a branch, and if a predecessor 3988 // branches to us and one of our successors, fold the comparison into the 3989 // predecessor and use logical operations to pick the right destination. 3990 if (FoldBranchToCommonDest(BI)) 3991 return SimplifyCFG(BB, TTI, TD) | true; 3992 3993 // We have a conditional branch to two blocks that are only reachable 3994 // from BI. We know that the condbr dominates the two blocks, so see if 3995 // there is any identical code in the "then" and "else" blocks. If so, we 3996 // can hoist it up to the branching block. 3997 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3998 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3999 if (HoistThenElseCodeToIf(BI)) 4000 return SimplifyCFG(BB, TTI, TD) | true; 4001 } else { 4002 // If Successor #1 has multiple preds, we may be able to conditionally 4003 // execute Successor #0 if it branches to successor #1. 4004 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 4005 if (Succ0TI->getNumSuccessors() == 1 && 4006 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 4007 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 4008 return SimplifyCFG(BB, TTI, TD) | true; 4009 } 4010 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 4011 // If Successor #0 has multiple preds, we may be able to conditionally 4012 // execute Successor #1 if it branches to successor #0. 4013 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 4014 if (Succ1TI->getNumSuccessors() == 1 && 4015 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 4016 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 4017 return SimplifyCFG(BB, TTI, TD) | true; 4018 } 4019 4020 // If this is a branch on a phi node in the current block, thread control 4021 // through this block if any PHI node entries are constants. 4022 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 4023 if (PN->getParent() == BI->getParent()) 4024 if (FoldCondBranchOnPHI(BI, TD)) 4025 return SimplifyCFG(BB, TTI, TD) | true; 4026 4027 // Scan predecessor blocks for conditional branches. 4028 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 4029 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 4030 if (PBI != BI && PBI->isConditional()) 4031 if (SimplifyCondBranchToCondBranch(PBI, BI)) 4032 return SimplifyCFG(BB, TTI, TD) | true; 4033 4034 return false; 4035} 4036 4037/// Check if passing a value to an instruction will cause undefined behavior. 4038static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 4039 Constant *C = dyn_cast<Constant>(V); 4040 if (!C) 4041 return false; 4042 4043 if (I->use_empty()) 4044 return false; 4045 4046 if (C->isNullValue()) { 4047 // Only look at the first use, avoid hurting compile time with long uselists 4048 User *Use = *I->use_begin(); 4049 4050 // Now make sure that there are no instructions in between that can alter 4051 // control flow (eg. calls) 4052 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 4053 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 4054 return false; 4055 4056 // Look through GEPs. A load from a GEP derived from NULL is still undefined 4057 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 4058 if (GEP->getPointerOperand() == I) 4059 return passingValueIsAlwaysUndefined(V, GEP); 4060 4061 // Look through bitcasts. 4062 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 4063 return passingValueIsAlwaysUndefined(V, BC); 4064 4065 // Load from null is undefined. 4066 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 4067 if (!LI->isVolatile()) 4068 return LI->getPointerAddressSpace() == 0; 4069 4070 // Store to null is undefined. 4071 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 4072 if (!SI->isVolatile()) 4073 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 4074 } 4075 return false; 4076} 4077 4078/// If BB has an incoming value that will always trigger undefined behavior 4079/// (eg. null pointer dereference), remove the branch leading here. 4080static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 4081 for (BasicBlock::iterator i = BB->begin(); 4082 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 4083 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 4084 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 4085 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 4086 IRBuilder<> Builder(T); 4087 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 4088 BB->removePredecessor(PHI->getIncomingBlock(i)); 4089 // Turn uncoditional branches into unreachables and remove the dead 4090 // destination from conditional branches. 4091 if (BI->isUnconditional()) 4092 Builder.CreateUnreachable(); 4093 else 4094 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 4095 BI->getSuccessor(0)); 4096 BI->eraseFromParent(); 4097 return true; 4098 } 4099 // TODO: SwitchInst. 4100 } 4101 4102 return false; 4103} 4104 4105bool SimplifyCFGOpt::run(BasicBlock *BB) { 4106 bool Changed = false; 4107 4108 assert(BB && BB->getParent() && "Block not embedded in function!"); 4109 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 4110 4111 // Remove basic blocks that have no predecessors (except the entry block)... 4112 // or that just have themself as a predecessor. These are unreachable. 4113 if ((pred_begin(BB) == pred_end(BB) && 4114 BB != &BB->getParent()->getEntryBlock()) || 4115 BB->getSinglePredecessor() == BB) { 4116 DEBUG(dbgs() << "Removing BB: \n" << *BB); 4117 DeleteDeadBlock(BB); 4118 return true; 4119 } 4120 4121 // Check to see if we can constant propagate this terminator instruction 4122 // away... 4123 Changed |= ConstantFoldTerminator(BB, true); 4124 4125 // Check for and eliminate duplicate PHI nodes in this block. 4126 Changed |= EliminateDuplicatePHINodes(BB); 4127 4128 // Check for and remove branches that will always cause undefined behavior. 4129 Changed |= removeUndefIntroducingPredecessor(BB); 4130 4131 // Merge basic blocks into their predecessor if there is only one distinct 4132 // pred, and if there is only one distinct successor of the predecessor, and 4133 // if there are no PHI nodes. 4134 // 4135 if (MergeBlockIntoPredecessor(BB)) 4136 return true; 4137 4138 IRBuilder<> Builder(BB); 4139 4140 // If there is a trivial two-entry PHI node in this basic block, and we can 4141 // eliminate it, do so now. 4142 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 4143 if (PN->getNumIncomingValues() == 2) 4144 Changed |= FoldTwoEntryPHINode(PN, TD); 4145 4146 Builder.SetInsertPoint(BB->getTerminator()); 4147 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 4148 if (BI->isUnconditional()) { 4149 if (SimplifyUncondBranch(BI, Builder)) return true; 4150 } else { 4151 if (SimplifyCondBranch(BI, Builder)) return true; 4152 } 4153 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 4154 if (SimplifyReturn(RI, Builder)) return true; 4155 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 4156 if (SimplifyResume(RI, Builder)) return true; 4157 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 4158 if (SimplifySwitch(SI, Builder)) return true; 4159 } else if (UnreachableInst *UI = 4160 dyn_cast<UnreachableInst>(BB->getTerminator())) { 4161 if (SimplifyUnreachable(UI)) return true; 4162 } else if (IndirectBrInst *IBI = 4163 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 4164 if (SimplifyIndirectBr(IBI)) return true; 4165 } 4166 4167 return Changed; 4168} 4169 4170/// SimplifyCFG - This function is used to do simplification of a CFG. For 4171/// example, it adjusts branches to branches to eliminate the extra hop, it 4172/// eliminates unreachable basic blocks, and does other "peephole" optimization 4173/// of the CFG. It returns true if a modification was made. 4174/// 4175bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, 4176 const DataLayout *TD) { 4177 return SimplifyCFGOpt(TTI, TD).run(BB); 4178} 4179