SimplifyCFG.cpp revision aa76e9e2cf50af190de90bc778b7f7e42ef9ceff
1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Peephole optimize the CFG. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "simplifycfg" 15#include "llvm/Transforms/Utils/Local.h" 16#include "llvm/Constants.h" 17#include "llvm/DerivedTypes.h" 18#include "llvm/GlobalVariable.h" 19#include "llvm/IRBuilder.h" 20#include "llvm/Instructions.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/LLVMContext.h" 23#include "llvm/MDBuilder.h" 24#include "llvm/Metadata.h" 25#include "llvm/Module.h" 26#include "llvm/Operator.h" 27#include "llvm/Type.h" 28#include "llvm/ADT/DenseMap.h" 29#include "llvm/ADT/STLExtras.h" 30#include "llvm/ADT/SetVector.h" 31#include "llvm/ADT/SmallPtrSet.h" 32#include "llvm/ADT/SmallVector.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/Analysis/InstructionSimplify.h" 35#include "llvm/Analysis/ValueTracking.h" 36#include "llvm/Support/CFG.h" 37#include "llvm/Support/CommandLine.h" 38#include "llvm/Support/ConstantRange.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/NoFolder.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/DataLayout.h" 43#include "llvm/Transforms/Utils/BasicBlockUtils.h" 44#include <algorithm> 45#include <set> 46#include <map> 47using namespace llvm; 48 49static cl::opt<unsigned> 50PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 51 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 52 53static cl::opt<bool> 54DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 55 cl::desc("Duplicate return instructions into unconditional branches")); 56 57static cl::opt<bool> 58SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 59 cl::desc("Sink common instructions down to the end block")); 60 61STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 62STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 63STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 64STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 65 66namespace { 67 /// ValueEqualityComparisonCase - Represents a case of a switch. 68 struct ValueEqualityComparisonCase { 69 ConstantInt *Value; 70 BasicBlock *Dest; 71 72 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 73 : Value(Value), Dest(Dest) {} 74 75 bool operator<(ValueEqualityComparisonCase RHS) const { 76 // Comparing pointers is ok as we only rely on the order for uniquing. 77 return Value < RHS.Value; 78 } 79 80 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 81 }; 82 83class SimplifyCFGOpt { 84 const DataLayout *const TD; 85 86 Value *isValueEqualityComparison(TerminatorInst *TI); 87 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 88 std::vector<ValueEqualityComparisonCase> &Cases); 89 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 90 BasicBlock *Pred, 91 IRBuilder<> &Builder); 92 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 93 IRBuilder<> &Builder); 94 95 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 96 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 97 bool SimplifyUnreachable(UnreachableInst *UI); 98 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 99 bool SimplifyIndirectBr(IndirectBrInst *IBI); 100 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 101 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 102 103public: 104 explicit SimplifyCFGOpt(const DataLayout *td) : TD(td) {} 105 bool run(BasicBlock *BB); 106}; 107} 108 109/// SafeToMergeTerminators - Return true if it is safe to merge these two 110/// terminator instructions together. 111/// 112static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 113 if (SI1 == SI2) return false; // Can't merge with self! 114 115 // It is not safe to merge these two switch instructions if they have a common 116 // successor, and if that successor has a PHI node, and if *that* PHI node has 117 // conflicting incoming values from the two switch blocks. 118 BasicBlock *SI1BB = SI1->getParent(); 119 BasicBlock *SI2BB = SI2->getParent(); 120 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 121 122 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 123 if (SI1Succs.count(*I)) 124 for (BasicBlock::iterator BBI = (*I)->begin(); 125 isa<PHINode>(BBI); ++BBI) { 126 PHINode *PN = cast<PHINode>(BBI); 127 if (PN->getIncomingValueForBlock(SI1BB) != 128 PN->getIncomingValueForBlock(SI2BB)) 129 return false; 130 } 131 132 return true; 133} 134 135/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 136/// to merge these two terminator instructions together, where SI1 is an 137/// unconditional branch. PhiNodes will store all PHI nodes in common 138/// successors. 139/// 140static bool isProfitableToFoldUnconditional(BranchInst *SI1, 141 BranchInst *SI2, 142 Instruction *Cond, 143 SmallVectorImpl<PHINode*> &PhiNodes) { 144 if (SI1 == SI2) return false; // Can't merge with self! 145 assert(SI1->isUnconditional() && SI2->isConditional()); 146 147 // We fold the unconditional branch if we can easily update all PHI nodes in 148 // common successors: 149 // 1> We have a constant incoming value for the conditional branch; 150 // 2> We have "Cond" as the incoming value for the unconditional branch; 151 // 3> SI2->getCondition() and Cond have same operands. 152 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 153 if (!Ci2) return false; 154 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 155 Cond->getOperand(1) == Ci2->getOperand(1)) && 156 !(Cond->getOperand(0) == Ci2->getOperand(1) && 157 Cond->getOperand(1) == Ci2->getOperand(0))) 158 return false; 159 160 BasicBlock *SI1BB = SI1->getParent(); 161 BasicBlock *SI2BB = SI2->getParent(); 162 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 163 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 164 if (SI1Succs.count(*I)) 165 for (BasicBlock::iterator BBI = (*I)->begin(); 166 isa<PHINode>(BBI); ++BBI) { 167 PHINode *PN = cast<PHINode>(BBI); 168 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 169 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 170 return false; 171 PhiNodes.push_back(PN); 172 } 173 return true; 174} 175 176/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 177/// now be entries in it from the 'NewPred' block. The values that will be 178/// flowing into the PHI nodes will be the same as those coming in from 179/// ExistPred, an existing predecessor of Succ. 180static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 181 BasicBlock *ExistPred) { 182 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 183 184 PHINode *PN; 185 for (BasicBlock::iterator I = Succ->begin(); 186 (PN = dyn_cast<PHINode>(I)); ++I) 187 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 188} 189 190 191/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 192/// least one PHI node in it), check to see if the merge at this block is due 193/// to an "if condition". If so, return the boolean condition that determines 194/// which entry into BB will be taken. Also, return by references the block 195/// that will be entered from if the condition is true, and the block that will 196/// be entered if the condition is false. 197/// 198/// This does no checking to see if the true/false blocks have large or unsavory 199/// instructions in them. 200static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 201 BasicBlock *&IfFalse) { 202 PHINode *SomePHI = cast<PHINode>(BB->begin()); 203 assert(SomePHI->getNumIncomingValues() == 2 && 204 "Function can only handle blocks with 2 predecessors!"); 205 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 206 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 207 208 // We can only handle branches. Other control flow will be lowered to 209 // branches if possible anyway. 210 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 211 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 212 if (Pred1Br == 0 || Pred2Br == 0) 213 return 0; 214 215 // Eliminate code duplication by ensuring that Pred1Br is conditional if 216 // either are. 217 if (Pred2Br->isConditional()) { 218 // If both branches are conditional, we don't have an "if statement". In 219 // reality, we could transform this case, but since the condition will be 220 // required anyway, we stand no chance of eliminating it, so the xform is 221 // probably not profitable. 222 if (Pred1Br->isConditional()) 223 return 0; 224 225 std::swap(Pred1, Pred2); 226 std::swap(Pred1Br, Pred2Br); 227 } 228 229 if (Pred1Br->isConditional()) { 230 // The only thing we have to watch out for here is to make sure that Pred2 231 // doesn't have incoming edges from other blocks. If it does, the condition 232 // doesn't dominate BB. 233 if (Pred2->getSinglePredecessor() == 0) 234 return 0; 235 236 // If we found a conditional branch predecessor, make sure that it branches 237 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 238 if (Pred1Br->getSuccessor(0) == BB && 239 Pred1Br->getSuccessor(1) == Pred2) { 240 IfTrue = Pred1; 241 IfFalse = Pred2; 242 } else if (Pred1Br->getSuccessor(0) == Pred2 && 243 Pred1Br->getSuccessor(1) == BB) { 244 IfTrue = Pred2; 245 IfFalse = Pred1; 246 } else { 247 // We know that one arm of the conditional goes to BB, so the other must 248 // go somewhere unrelated, and this must not be an "if statement". 249 return 0; 250 } 251 252 return Pred1Br->getCondition(); 253 } 254 255 // Ok, if we got here, both predecessors end with an unconditional branch to 256 // BB. Don't panic! If both blocks only have a single (identical) 257 // predecessor, and THAT is a conditional branch, then we're all ok! 258 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 259 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 260 return 0; 261 262 // Otherwise, if this is a conditional branch, then we can use it! 263 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 264 if (BI == 0) return 0; 265 266 assert(BI->isConditional() && "Two successors but not conditional?"); 267 if (BI->getSuccessor(0) == Pred1) { 268 IfTrue = Pred1; 269 IfFalse = Pred2; 270 } else { 271 IfTrue = Pred2; 272 IfFalse = Pred1; 273 } 274 return BI->getCondition(); 275} 276 277/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 278/// given instruction, which is assumed to be safe to speculate. 1 means 279/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 280static unsigned ComputeSpeculationCost(const User *I) { 281 assert(isSafeToSpeculativelyExecute(I) && 282 "Instruction is not safe to speculatively execute!"); 283 switch (Operator::getOpcode(I)) { 284 default: 285 // In doubt, be conservative. 286 return UINT_MAX; 287 case Instruction::GetElementPtr: 288 // GEPs are cheap if all indices are constant. 289 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 290 return UINT_MAX; 291 return 1; 292 case Instruction::Load: 293 case Instruction::Add: 294 case Instruction::Sub: 295 case Instruction::And: 296 case Instruction::Or: 297 case Instruction::Xor: 298 case Instruction::Shl: 299 case Instruction::LShr: 300 case Instruction::AShr: 301 case Instruction::ICmp: 302 case Instruction::Trunc: 303 case Instruction::ZExt: 304 case Instruction::SExt: 305 return 1; // These are all cheap. 306 307 case Instruction::Call: 308 case Instruction::Select: 309 return 2; 310 } 311} 312 313/// DominatesMergePoint - If we have a merge point of an "if condition" as 314/// accepted above, return true if the specified value dominates the block. We 315/// don't handle the true generality of domination here, just a special case 316/// which works well enough for us. 317/// 318/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 319/// see if V (which must be an instruction) and its recursive operands 320/// that do not dominate BB have a combined cost lower than CostRemaining and 321/// are non-trapping. If both are true, the instruction is inserted into the 322/// set and true is returned. 323/// 324/// The cost for most non-trapping instructions is defined as 1 except for 325/// Select whose cost is 2. 326/// 327/// After this function returns, CostRemaining is decreased by the cost of 328/// V plus its non-dominating operands. If that cost is greater than 329/// CostRemaining, false is returned and CostRemaining is undefined. 330static bool DominatesMergePoint(Value *V, BasicBlock *BB, 331 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 332 unsigned &CostRemaining) { 333 Instruction *I = dyn_cast<Instruction>(V); 334 if (!I) { 335 // Non-instructions all dominate instructions, but not all constantexprs 336 // can be executed unconditionally. 337 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 338 if (C->canTrap()) 339 return false; 340 return true; 341 } 342 BasicBlock *PBB = I->getParent(); 343 344 // We don't want to allow weird loops that might have the "if condition" in 345 // the bottom of this block. 346 if (PBB == BB) return false; 347 348 // If this instruction is defined in a block that contains an unconditional 349 // branch to BB, then it must be in the 'conditional' part of the "if 350 // statement". If not, it definitely dominates the region. 351 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 352 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 353 return true; 354 355 // If we aren't allowing aggressive promotion anymore, then don't consider 356 // instructions in the 'if region'. 357 if (AggressiveInsts == 0) return false; 358 359 // If we have seen this instruction before, don't count it again. 360 if (AggressiveInsts->count(I)) return true; 361 362 // Okay, it looks like the instruction IS in the "condition". Check to 363 // see if it's a cheap instruction to unconditionally compute, and if it 364 // only uses stuff defined outside of the condition. If so, hoist it out. 365 if (!isSafeToSpeculativelyExecute(I)) 366 return false; 367 368 unsigned Cost = ComputeSpeculationCost(I); 369 370 if (Cost > CostRemaining) 371 return false; 372 373 CostRemaining -= Cost; 374 375 // Okay, we can only really hoist these out if their operands do 376 // not take us over the cost threshold. 377 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 378 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 379 return false; 380 // Okay, it's safe to do this! Remember this instruction. 381 AggressiveInsts->insert(I); 382 return true; 383} 384 385/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 386/// and PointerNullValue. Return NULL if value is not a constant int. 387static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 388 // Normal constant int. 389 ConstantInt *CI = dyn_cast<ConstantInt>(V); 390 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 391 return CI; 392 393 // This is some kind of pointer constant. Turn it into a pointer-sized 394 // ConstantInt if possible. 395 IntegerType *PtrTy = TD->getIntPtrType(V->getType()); 396 397 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 398 if (isa<ConstantPointerNull>(V)) 399 return ConstantInt::get(PtrTy, 0); 400 401 // IntToPtr const int. 402 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 403 if (CE->getOpcode() == Instruction::IntToPtr) 404 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 405 // The constant is very likely to have the right type already. 406 if (CI->getType() == PtrTy) 407 return CI; 408 else 409 return cast<ConstantInt> 410 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 411 } 412 return 0; 413} 414 415/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 416/// collection of icmp eq/ne instructions that compare a value against a 417/// constant, return the value being compared, and stick the constant into the 418/// Values vector. 419static Value * 420GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 421 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 422 Instruction *I = dyn_cast<Instruction>(V); 423 if (I == 0) return 0; 424 425 // If this is an icmp against a constant, handle this as one of the cases. 426 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 427 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 428 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 429 UsedICmps++; 430 Vals.push_back(C); 431 return I->getOperand(0); 432 } 433 434 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 435 // the set. 436 ConstantRange Span = 437 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 438 439 // If this is an and/!= check then we want to optimize "x ugt 2" into 440 // x != 0 && x != 1. 441 if (!isEQ) 442 Span = Span.inverse(); 443 444 // If there are a ton of values, we don't want to make a ginormous switch. 445 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 446 return 0; 447 448 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 449 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 450 UsedICmps++; 451 return I->getOperand(0); 452 } 453 return 0; 454 } 455 456 // Otherwise, we can only handle an | or &, depending on isEQ. 457 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 458 return 0; 459 460 unsigned NumValsBeforeLHS = Vals.size(); 461 unsigned UsedICmpsBeforeLHS = UsedICmps; 462 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 463 isEQ, UsedICmps)) { 464 unsigned NumVals = Vals.size(); 465 unsigned UsedICmpsBeforeRHS = UsedICmps; 466 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 467 isEQ, UsedICmps)) { 468 if (LHS == RHS) 469 return LHS; 470 Vals.resize(NumVals); 471 UsedICmps = UsedICmpsBeforeRHS; 472 } 473 474 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 475 // set it and return success. 476 if (Extra == 0 || Extra == I->getOperand(1)) { 477 Extra = I->getOperand(1); 478 return LHS; 479 } 480 481 Vals.resize(NumValsBeforeLHS); 482 UsedICmps = UsedICmpsBeforeLHS; 483 return 0; 484 } 485 486 // If the LHS can't be folded in, but Extra is available and RHS can, try to 487 // use LHS as Extra. 488 if (Extra == 0 || Extra == I->getOperand(0)) { 489 Value *OldExtra = Extra; 490 Extra = I->getOperand(0); 491 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 492 isEQ, UsedICmps)) 493 return RHS; 494 assert(Vals.size() == NumValsBeforeLHS); 495 Extra = OldExtra; 496 } 497 498 return 0; 499} 500 501static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 502 Instruction *Cond = 0; 503 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 504 Cond = dyn_cast<Instruction>(SI->getCondition()); 505 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 506 if (BI->isConditional()) 507 Cond = dyn_cast<Instruction>(BI->getCondition()); 508 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 509 Cond = dyn_cast<Instruction>(IBI->getAddress()); 510 } 511 512 TI->eraseFromParent(); 513 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 514} 515 516/// isValueEqualityComparison - Return true if the specified terminator checks 517/// to see if a value is equal to constant integer value. 518Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 519 Value *CV = 0; 520 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 521 // Do not permit merging of large switch instructions into their 522 // predecessors unless there is only one predecessor. 523 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 524 pred_end(SI->getParent())) <= 128) 525 CV = SI->getCondition(); 526 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 527 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 528 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 529 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 530 ICI->getPredicate() == ICmpInst::ICMP_NE) && 531 GetConstantInt(ICI->getOperand(1), TD)) 532 CV = ICI->getOperand(0); 533 534 // Unwrap any lossless ptrtoint cast. 535 if (TD && CV) { 536 PtrToIntInst *PTII = NULL; 537 if ((PTII = dyn_cast<PtrToIntInst>(CV)) && 538 CV->getType() == TD->getIntPtrType(CV->getContext(), 539 PTII->getPointerAddressSpace())) 540 CV = PTII->getOperand(0); 541 } 542 return CV; 543} 544 545/// GetValueEqualityComparisonCases - Given a value comparison instruction, 546/// decode all of the 'cases' that it represents and return the 'default' block. 547BasicBlock *SimplifyCFGOpt:: 548GetValueEqualityComparisonCases(TerminatorInst *TI, 549 std::vector<ValueEqualityComparisonCase> 550 &Cases) { 551 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 552 Cases.reserve(SI->getNumCases()); 553 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 554 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 555 i.getCaseSuccessor())); 556 return SI->getDefaultDest(); 557 } 558 559 BranchInst *BI = cast<BranchInst>(TI); 560 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 561 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 562 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 563 TD), 564 Succ)); 565 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 566} 567 568 569/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 570/// in the list that match the specified block. 571static void EliminateBlockCases(BasicBlock *BB, 572 std::vector<ValueEqualityComparisonCase> &Cases) { 573 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 574} 575 576/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 577/// well. 578static bool 579ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 580 std::vector<ValueEqualityComparisonCase > &C2) { 581 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 582 583 // Make V1 be smaller than V2. 584 if (V1->size() > V2->size()) 585 std::swap(V1, V2); 586 587 if (V1->size() == 0) return false; 588 if (V1->size() == 1) { 589 // Just scan V2. 590 ConstantInt *TheVal = (*V1)[0].Value; 591 for (unsigned i = 0, e = V2->size(); i != e; ++i) 592 if (TheVal == (*V2)[i].Value) 593 return true; 594 } 595 596 // Otherwise, just sort both lists and compare element by element. 597 array_pod_sort(V1->begin(), V1->end()); 598 array_pod_sort(V2->begin(), V2->end()); 599 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 600 while (i1 != e1 && i2 != e2) { 601 if ((*V1)[i1].Value == (*V2)[i2].Value) 602 return true; 603 if ((*V1)[i1].Value < (*V2)[i2].Value) 604 ++i1; 605 else 606 ++i2; 607 } 608 return false; 609} 610 611/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 612/// terminator instruction and its block is known to only have a single 613/// predecessor block, check to see if that predecessor is also a value 614/// comparison with the same value, and if that comparison determines the 615/// outcome of this comparison. If so, simplify TI. This does a very limited 616/// form of jump threading. 617bool SimplifyCFGOpt:: 618SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 619 BasicBlock *Pred, 620 IRBuilder<> &Builder) { 621 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 622 if (!PredVal) return false; // Not a value comparison in predecessor. 623 624 Value *ThisVal = isValueEqualityComparison(TI); 625 assert(ThisVal && "This isn't a value comparison!!"); 626 if (ThisVal != PredVal) return false; // Different predicates. 627 628 // TODO: Preserve branch weight metadata, similarly to how 629 // FoldValueComparisonIntoPredecessors preserves it. 630 631 // Find out information about when control will move from Pred to TI's block. 632 std::vector<ValueEqualityComparisonCase> PredCases; 633 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 634 PredCases); 635 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 636 637 // Find information about how control leaves this block. 638 std::vector<ValueEqualityComparisonCase> ThisCases; 639 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 640 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 641 642 // If TI's block is the default block from Pred's comparison, potentially 643 // simplify TI based on this knowledge. 644 if (PredDef == TI->getParent()) { 645 // If we are here, we know that the value is none of those cases listed in 646 // PredCases. If there are any cases in ThisCases that are in PredCases, we 647 // can simplify TI. 648 if (!ValuesOverlap(PredCases, ThisCases)) 649 return false; 650 651 if (isa<BranchInst>(TI)) { 652 // Okay, one of the successors of this condbr is dead. Convert it to a 653 // uncond br. 654 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 655 // Insert the new branch. 656 Instruction *NI = Builder.CreateBr(ThisDef); 657 (void) NI; 658 659 // Remove PHI node entries for the dead edge. 660 ThisCases[0].Dest->removePredecessor(TI->getParent()); 661 662 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 663 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 664 665 EraseTerminatorInstAndDCECond(TI); 666 return true; 667 } 668 669 SwitchInst *SI = cast<SwitchInst>(TI); 670 // Okay, TI has cases that are statically dead, prune them away. 671 SmallPtrSet<Constant*, 16> DeadCases; 672 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 673 DeadCases.insert(PredCases[i].Value); 674 675 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 676 << "Through successor TI: " << *TI); 677 678 // Collect branch weights into a vector. 679 SmallVector<uint32_t, 8> Weights; 680 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 681 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 682 if (HasWeight) 683 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 684 ++MD_i) { 685 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 686 assert(CI); 687 Weights.push_back(CI->getValue().getZExtValue()); 688 } 689 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 690 --i; 691 if (DeadCases.count(i.getCaseValue())) { 692 if (HasWeight) { 693 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 694 Weights.pop_back(); 695 } 696 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 697 SI->removeCase(i); 698 } 699 } 700 if (HasWeight && Weights.size() >= 2) 701 SI->setMetadata(LLVMContext::MD_prof, 702 MDBuilder(SI->getParent()->getContext()). 703 createBranchWeights(Weights)); 704 705 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 706 return true; 707 } 708 709 // Otherwise, TI's block must correspond to some matched value. Find out 710 // which value (or set of values) this is. 711 ConstantInt *TIV = 0; 712 BasicBlock *TIBB = TI->getParent(); 713 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 714 if (PredCases[i].Dest == TIBB) { 715 if (TIV != 0) 716 return false; // Cannot handle multiple values coming to this block. 717 TIV = PredCases[i].Value; 718 } 719 assert(TIV && "No edge from pred to succ?"); 720 721 // Okay, we found the one constant that our value can be if we get into TI's 722 // BB. Find out which successor will unconditionally be branched to. 723 BasicBlock *TheRealDest = 0; 724 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 725 if (ThisCases[i].Value == TIV) { 726 TheRealDest = ThisCases[i].Dest; 727 break; 728 } 729 730 // If not handled by any explicit cases, it is handled by the default case. 731 if (TheRealDest == 0) TheRealDest = ThisDef; 732 733 // Remove PHI node entries for dead edges. 734 BasicBlock *CheckEdge = TheRealDest; 735 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 736 if (*SI != CheckEdge) 737 (*SI)->removePredecessor(TIBB); 738 else 739 CheckEdge = 0; 740 741 // Insert the new branch. 742 Instruction *NI = Builder.CreateBr(TheRealDest); 743 (void) NI; 744 745 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 746 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 747 748 EraseTerminatorInstAndDCECond(TI); 749 return true; 750} 751 752namespace { 753 /// ConstantIntOrdering - This class implements a stable ordering of constant 754 /// integers that does not depend on their address. This is important for 755 /// applications that sort ConstantInt's to ensure uniqueness. 756 struct ConstantIntOrdering { 757 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 758 return LHS->getValue().ult(RHS->getValue()); 759 } 760 }; 761} 762 763static int ConstantIntSortPredicate(const void *P1, const void *P2) { 764 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 765 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 766 if (LHS->getValue().ult(RHS->getValue())) 767 return 1; 768 if (LHS->getValue() == RHS->getValue()) 769 return 0; 770 return -1; 771} 772 773static inline bool HasBranchWeights(const Instruction* I) { 774 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 775 if (ProfMD && ProfMD->getOperand(0)) 776 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 777 return MDS->getString().equals("branch_weights"); 778 779 return false; 780} 781 782/// Get Weights of a given TerminatorInst, the default weight is at the front 783/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 784/// metadata. 785static void GetBranchWeights(TerminatorInst *TI, 786 SmallVectorImpl<uint64_t> &Weights) { 787 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 788 assert(MD); 789 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 790 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 791 assert(CI); 792 Weights.push_back(CI->getValue().getZExtValue()); 793 } 794 795 // If TI is a conditional eq, the default case is the false case, 796 // and the corresponding branch-weight data is at index 2. We swap the 797 // default weight to be the first entry. 798 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 799 assert(Weights.size() == 2); 800 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 801 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 802 std::swap(Weights.front(), Weights.back()); 803 } 804} 805 806/// Sees if any of the weights are too big for a uint32_t, and halves all the 807/// weights if any are. 808static void FitWeights(MutableArrayRef<uint64_t> Weights) { 809 bool Halve = false; 810 for (unsigned i = 0; i < Weights.size(); ++i) 811 if (Weights[i] > UINT_MAX) { 812 Halve = true; 813 break; 814 } 815 816 if (! Halve) 817 return; 818 819 for (unsigned i = 0; i < Weights.size(); ++i) 820 Weights[i] /= 2; 821} 822 823/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 824/// equality comparison instruction (either a switch or a branch on "X == c"). 825/// See if any of the predecessors of the terminator block are value comparisons 826/// on the same value. If so, and if safe to do so, fold them together. 827bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 828 IRBuilder<> &Builder) { 829 BasicBlock *BB = TI->getParent(); 830 Value *CV = isValueEqualityComparison(TI); // CondVal 831 assert(CV && "Not a comparison?"); 832 bool Changed = false; 833 834 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 835 while (!Preds.empty()) { 836 BasicBlock *Pred = Preds.pop_back_val(); 837 838 // See if the predecessor is a comparison with the same value. 839 TerminatorInst *PTI = Pred->getTerminator(); 840 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 841 842 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 843 // Figure out which 'cases' to copy from SI to PSI. 844 std::vector<ValueEqualityComparisonCase> BBCases; 845 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 846 847 std::vector<ValueEqualityComparisonCase> PredCases; 848 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 849 850 // Based on whether the default edge from PTI goes to BB or not, fill in 851 // PredCases and PredDefault with the new switch cases we would like to 852 // build. 853 SmallVector<BasicBlock*, 8> NewSuccessors; 854 855 // Update the branch weight metadata along the way 856 SmallVector<uint64_t, 8> Weights; 857 bool PredHasWeights = HasBranchWeights(PTI); 858 bool SuccHasWeights = HasBranchWeights(TI); 859 860 if (PredHasWeights) { 861 GetBranchWeights(PTI, Weights); 862 // branch-weight metadata is inconsistant here. 863 if (Weights.size() != 1 + PredCases.size()) 864 PredHasWeights = SuccHasWeights = false; 865 } else if (SuccHasWeights) 866 // If there are no predecessor weights but there are successor weights, 867 // populate Weights with 1, which will later be scaled to the sum of 868 // successor's weights 869 Weights.assign(1 + PredCases.size(), 1); 870 871 SmallVector<uint64_t, 8> SuccWeights; 872 if (SuccHasWeights) { 873 GetBranchWeights(TI, SuccWeights); 874 // branch-weight metadata is inconsistant here. 875 if (SuccWeights.size() != 1 + BBCases.size()) 876 PredHasWeights = SuccHasWeights = false; 877 } else if (PredHasWeights) 878 SuccWeights.assign(1 + BBCases.size(), 1); 879 880 if (PredDefault == BB) { 881 // If this is the default destination from PTI, only the edges in TI 882 // that don't occur in PTI, or that branch to BB will be activated. 883 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 884 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 885 if (PredCases[i].Dest != BB) 886 PTIHandled.insert(PredCases[i].Value); 887 else { 888 // The default destination is BB, we don't need explicit targets. 889 std::swap(PredCases[i], PredCases.back()); 890 891 if (PredHasWeights || SuccHasWeights) { 892 // Increase weight for the default case. 893 Weights[0] += Weights[i+1]; 894 std::swap(Weights[i+1], Weights.back()); 895 Weights.pop_back(); 896 } 897 898 PredCases.pop_back(); 899 --i; --e; 900 } 901 902 // Reconstruct the new switch statement we will be building. 903 if (PredDefault != BBDefault) { 904 PredDefault->removePredecessor(Pred); 905 PredDefault = BBDefault; 906 NewSuccessors.push_back(BBDefault); 907 } 908 909 unsigned CasesFromPred = Weights.size(); 910 uint64_t ValidTotalSuccWeight = 0; 911 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 912 if (!PTIHandled.count(BBCases[i].Value) && 913 BBCases[i].Dest != BBDefault) { 914 PredCases.push_back(BBCases[i]); 915 NewSuccessors.push_back(BBCases[i].Dest); 916 if (SuccHasWeights || PredHasWeights) { 917 // The default weight is at index 0, so weight for the ith case 918 // should be at index i+1. Scale the cases from successor by 919 // PredDefaultWeight (Weights[0]). 920 Weights.push_back(Weights[0] * SuccWeights[i+1]); 921 ValidTotalSuccWeight += SuccWeights[i+1]; 922 } 923 } 924 925 if (SuccHasWeights || PredHasWeights) { 926 ValidTotalSuccWeight += SuccWeights[0]; 927 // Scale the cases from predecessor by ValidTotalSuccWeight. 928 for (unsigned i = 1; i < CasesFromPred; ++i) 929 Weights[i] *= ValidTotalSuccWeight; 930 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 931 Weights[0] *= SuccWeights[0]; 932 } 933 } else { 934 // If this is not the default destination from PSI, only the edges 935 // in SI that occur in PSI with a destination of BB will be 936 // activated. 937 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 938 std::map<ConstantInt*, uint64_t> WeightsForHandled; 939 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 940 if (PredCases[i].Dest == BB) { 941 PTIHandled.insert(PredCases[i].Value); 942 943 if (PredHasWeights || SuccHasWeights) { 944 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 945 std::swap(Weights[i+1], Weights.back()); 946 Weights.pop_back(); 947 } 948 949 std::swap(PredCases[i], PredCases.back()); 950 PredCases.pop_back(); 951 --i; --e; 952 } 953 954 // Okay, now we know which constants were sent to BB from the 955 // predecessor. Figure out where they will all go now. 956 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 957 if (PTIHandled.count(BBCases[i].Value)) { 958 // If this is one we are capable of getting... 959 if (PredHasWeights || SuccHasWeights) 960 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 961 PredCases.push_back(BBCases[i]); 962 NewSuccessors.push_back(BBCases[i].Dest); 963 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 964 } 965 966 // If there are any constants vectored to BB that TI doesn't handle, 967 // they must go to the default destination of TI. 968 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 969 PTIHandled.begin(), 970 E = PTIHandled.end(); I != E; ++I) { 971 if (PredHasWeights || SuccHasWeights) 972 Weights.push_back(WeightsForHandled[*I]); 973 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 974 NewSuccessors.push_back(BBDefault); 975 } 976 } 977 978 // Okay, at this point, we know which new successor Pred will get. Make 979 // sure we update the number of entries in the PHI nodes for these 980 // successors. 981 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 982 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 983 984 Builder.SetInsertPoint(PTI); 985 // Convert pointer to int before we switch. 986 if (CV->getType()->isPointerTy()) { 987 assert(TD && "Cannot switch on pointer without DataLayout"); 988 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getType()), 989 "magicptr"); 990 } 991 992 // Now that the successors are updated, create the new Switch instruction. 993 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 994 PredCases.size()); 995 NewSI->setDebugLoc(PTI->getDebugLoc()); 996 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 997 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 998 999 if (PredHasWeights || SuccHasWeights) { 1000 // Halve the weights if any of them cannot fit in an uint32_t 1001 FitWeights(Weights); 1002 1003 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1004 1005 NewSI->setMetadata(LLVMContext::MD_prof, 1006 MDBuilder(BB->getContext()). 1007 createBranchWeights(MDWeights)); 1008 } 1009 1010 EraseTerminatorInstAndDCECond(PTI); 1011 1012 // Okay, last check. If BB is still a successor of PSI, then we must 1013 // have an infinite loop case. If so, add an infinitely looping block 1014 // to handle the case to preserve the behavior of the code. 1015 BasicBlock *InfLoopBlock = 0; 1016 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1017 if (NewSI->getSuccessor(i) == BB) { 1018 if (InfLoopBlock == 0) { 1019 // Insert it at the end of the function, because it's either code, 1020 // or it won't matter if it's hot. :) 1021 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1022 "infloop", BB->getParent()); 1023 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1024 } 1025 NewSI->setSuccessor(i, InfLoopBlock); 1026 } 1027 1028 Changed = true; 1029 } 1030 } 1031 return Changed; 1032} 1033 1034// isSafeToHoistInvoke - If we would need to insert a select that uses the 1035// value of this invoke (comments in HoistThenElseCodeToIf explain why we 1036// would need to do this), we can't hoist the invoke, as there is nowhere 1037// to put the select in this case. 1038static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1039 Instruction *I1, Instruction *I2) { 1040 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1041 PHINode *PN; 1042 for (BasicBlock::iterator BBI = SI->begin(); 1043 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1044 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1045 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1046 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1047 return false; 1048 } 1049 } 1050 } 1051 return true; 1052} 1053 1054/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1055/// BB2, hoist any common code in the two blocks up into the branch block. The 1056/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1057static bool HoistThenElseCodeToIf(BranchInst *BI) { 1058 // This does very trivial matching, with limited scanning, to find identical 1059 // instructions in the two blocks. In particular, we don't want to get into 1060 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1061 // such, we currently just scan for obviously identical instructions in an 1062 // identical order. 1063 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1064 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1065 1066 BasicBlock::iterator BB1_Itr = BB1->begin(); 1067 BasicBlock::iterator BB2_Itr = BB2->begin(); 1068 1069 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1070 // Skip debug info if it is not identical. 1071 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1072 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1073 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1074 while (isa<DbgInfoIntrinsic>(I1)) 1075 I1 = BB1_Itr++; 1076 while (isa<DbgInfoIntrinsic>(I2)) 1077 I2 = BB2_Itr++; 1078 } 1079 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1080 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1081 return false; 1082 1083 // If we get here, we can hoist at least one instruction. 1084 BasicBlock *BIParent = BI->getParent(); 1085 1086 do { 1087 // If we are hoisting the terminator instruction, don't move one (making a 1088 // broken BB), instead clone it, and remove BI. 1089 if (isa<TerminatorInst>(I1)) 1090 goto HoistTerminator; 1091 1092 // For a normal instruction, we just move one to right before the branch, 1093 // then replace all uses of the other with the first. Finally, we remove 1094 // the now redundant second instruction. 1095 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1096 if (!I2->use_empty()) 1097 I2->replaceAllUsesWith(I1); 1098 I1->intersectOptionalDataWith(I2); 1099 I2->eraseFromParent(); 1100 1101 I1 = BB1_Itr++; 1102 I2 = BB2_Itr++; 1103 // Skip debug info if it is not identical. 1104 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1105 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1106 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1107 while (isa<DbgInfoIntrinsic>(I1)) 1108 I1 = BB1_Itr++; 1109 while (isa<DbgInfoIntrinsic>(I2)) 1110 I2 = BB2_Itr++; 1111 } 1112 } while (I1->isIdenticalToWhenDefined(I2)); 1113 1114 return true; 1115 1116HoistTerminator: 1117 // It may not be possible to hoist an invoke. 1118 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1119 return true; 1120 1121 // Okay, it is safe to hoist the terminator. 1122 Instruction *NT = I1->clone(); 1123 BIParent->getInstList().insert(BI, NT); 1124 if (!NT->getType()->isVoidTy()) { 1125 I1->replaceAllUsesWith(NT); 1126 I2->replaceAllUsesWith(NT); 1127 NT->takeName(I1); 1128 } 1129 1130 IRBuilder<true, NoFolder> Builder(NT); 1131 // Hoisting one of the terminators from our successor is a great thing. 1132 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1133 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1134 // nodes, so we insert select instruction to compute the final result. 1135 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1136 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1137 PHINode *PN; 1138 for (BasicBlock::iterator BBI = SI->begin(); 1139 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1140 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1141 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1142 if (BB1V == BB2V) continue; 1143 1144 // These values do not agree. Insert a select instruction before NT 1145 // that determines the right value. 1146 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1147 if (SI == 0) 1148 SI = cast<SelectInst> 1149 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1150 BB1V->getName()+"."+BB2V->getName())); 1151 1152 // Make the PHI node use the select for all incoming values for BB1/BB2 1153 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1154 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1155 PN->setIncomingValue(i, SI); 1156 } 1157 } 1158 1159 // Update any PHI nodes in our new successors. 1160 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1161 AddPredecessorToBlock(*SI, BIParent, BB1); 1162 1163 EraseTerminatorInstAndDCECond(BI); 1164 return true; 1165} 1166 1167/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1168/// check whether BBEnd has only two predecessors and the other predecessor 1169/// ends with an unconditional branch. If it is true, sink any common code 1170/// in the two predecessors to BBEnd. 1171static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1172 assert(BI1->isUnconditional()); 1173 BasicBlock *BB1 = BI1->getParent(); 1174 BasicBlock *BBEnd = BI1->getSuccessor(0); 1175 1176 // Check that BBEnd has two predecessors and the other predecessor ends with 1177 // an unconditional branch. 1178 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1179 BasicBlock *Pred0 = *PI++; 1180 if (PI == PE) // Only one predecessor. 1181 return false; 1182 BasicBlock *Pred1 = *PI++; 1183 if (PI != PE) // More than two predecessors. 1184 return false; 1185 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1186 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1187 if (!BI2 || !BI2->isUnconditional()) 1188 return false; 1189 1190 // Gather the PHI nodes in BBEnd. 1191 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1192 Instruction *FirstNonPhiInBBEnd = 0; 1193 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1194 I != E; ++I) { 1195 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1196 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1197 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1198 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1199 } else { 1200 FirstNonPhiInBBEnd = &*I; 1201 break; 1202 } 1203 } 1204 if (!FirstNonPhiInBBEnd) 1205 return false; 1206 1207 1208 // This does very trivial matching, with limited scanning, to find identical 1209 // instructions in the two blocks. We scan backward for obviously identical 1210 // instructions in an identical order. 1211 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1212 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1213 RE2 = BB2->getInstList().rend(); 1214 // Skip debug info. 1215 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1216 if (RI1 == RE1) 1217 return false; 1218 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1219 if (RI2 == RE2) 1220 return false; 1221 // Skip the unconditional branches. 1222 ++RI1; 1223 ++RI2; 1224 1225 bool Changed = false; 1226 while (RI1 != RE1 && RI2 != RE2) { 1227 // Skip debug info. 1228 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1229 if (RI1 == RE1) 1230 return Changed; 1231 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1232 if (RI2 == RE2) 1233 return Changed; 1234 1235 Instruction *I1 = &*RI1, *I2 = &*RI2; 1236 // I1 and I2 should have a single use in the same PHI node, and they 1237 // perform the same operation. 1238 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1239 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1240 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1241 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1242 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1243 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1244 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1245 !I1->hasOneUse() || !I2->hasOneUse() || 1246 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1247 MapValueFromBB1ToBB2[I1].first != I2) 1248 return Changed; 1249 1250 // Check whether we should swap the operands of ICmpInst. 1251 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1252 bool SwapOpnds = false; 1253 if (ICmp1 && ICmp2 && 1254 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1255 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1256 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1257 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1258 ICmp2->swapOperands(); 1259 SwapOpnds = true; 1260 } 1261 if (!I1->isSameOperationAs(I2)) { 1262 if (SwapOpnds) 1263 ICmp2->swapOperands(); 1264 return Changed; 1265 } 1266 1267 // The operands should be either the same or they need to be generated 1268 // with a PHI node after sinking. We only handle the case where there is 1269 // a single pair of different operands. 1270 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1271 unsigned Op1Idx = 0; 1272 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1273 if (I1->getOperand(I) == I2->getOperand(I)) 1274 continue; 1275 // Early exit if we have more-than one pair of different operands or 1276 // the different operand is already in MapValueFromBB1ToBB2. 1277 // Early exit if we need a PHI node to replace a constant. 1278 if (DifferentOp1 || 1279 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1280 MapValueFromBB1ToBB2.end() || 1281 isa<Constant>(I1->getOperand(I)) || 1282 isa<Constant>(I2->getOperand(I))) { 1283 // If we can't sink the instructions, undo the swapping. 1284 if (SwapOpnds) 1285 ICmp2->swapOperands(); 1286 return Changed; 1287 } 1288 DifferentOp1 = I1->getOperand(I); 1289 Op1Idx = I; 1290 DifferentOp2 = I2->getOperand(I); 1291 } 1292 1293 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1294 // remove (I1, I2) from MapValueFromBB1ToBB2. 1295 if (DifferentOp1) { 1296 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1297 DifferentOp1->getName() + ".sink", 1298 BBEnd->begin()); 1299 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1300 // I1 should use NewPN instead of DifferentOp1. 1301 I1->setOperand(Op1Idx, NewPN); 1302 NewPN->addIncoming(DifferentOp1, BB1); 1303 NewPN->addIncoming(DifferentOp2, BB2); 1304 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1305 } 1306 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1307 MapValueFromBB1ToBB2.erase(I1); 1308 1309 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1310 DEBUG(dbgs() << " " << *I2 << "\n";); 1311 // We need to update RE1 and RE2 if we are going to sink the first 1312 // instruction in the basic block down. 1313 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1314 // Sink the instruction. 1315 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1316 if (!OldPN->use_empty()) 1317 OldPN->replaceAllUsesWith(I1); 1318 OldPN->eraseFromParent(); 1319 1320 if (!I2->use_empty()) 1321 I2->replaceAllUsesWith(I1); 1322 I1->intersectOptionalDataWith(I2); 1323 I2->eraseFromParent(); 1324 1325 if (UpdateRE1) 1326 RE1 = BB1->getInstList().rend(); 1327 if (UpdateRE2) 1328 RE2 = BB2->getInstList().rend(); 1329 FirstNonPhiInBBEnd = I1; 1330 NumSinkCommons++; 1331 Changed = true; 1332 } 1333 return Changed; 1334} 1335 1336/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 1337/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 1338/// (for now, restricted to a single instruction that's side effect free) from 1339/// the BB1 into the branch block to speculatively execute it. 1340/// 1341/// Turn 1342/// BB: 1343/// %t1 = icmp 1344/// br i1 %t1, label %BB1, label %BB2 1345/// BB1: 1346/// %t3 = add %t2, c 1347/// br label BB2 1348/// BB2: 1349/// => 1350/// BB: 1351/// %t1 = icmp 1352/// %t4 = add %t2, c 1353/// %t3 = select i1 %t1, %t2, %t3 1354static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 1355 // Only speculatively execution a single instruction (not counting the 1356 // terminator) for now. 1357 Instruction *HInst = NULL; 1358 Instruction *Term = BB1->getTerminator(); 1359 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 1360 BBI != BBE; ++BBI) { 1361 Instruction *I = BBI; 1362 // Skip debug info. 1363 if (isa<DbgInfoIntrinsic>(I)) continue; 1364 if (I == Term) break; 1365 1366 if (HInst) 1367 return false; 1368 HInst = I; 1369 } 1370 1371 BasicBlock *BIParent = BI->getParent(); 1372 1373 // Check the instruction to be hoisted, if there is one. 1374 if (HInst) { 1375 // Don't hoist the instruction if it's unsafe or expensive. 1376 if (!isSafeToSpeculativelyExecute(HInst)) 1377 return false; 1378 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold) 1379 return false; 1380 1381 // Do not hoist the instruction if any of its operands are defined but not 1382 // used in this BB. The transformation will prevent the operand from 1383 // being sunk into the use block. 1384 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1385 i != e; ++i) { 1386 Instruction *OpI = dyn_cast<Instruction>(*i); 1387 if (OpI && OpI->getParent() == BIParent && 1388 !OpI->mayHaveSideEffects() && 1389 !OpI->isUsedInBasicBlock(BIParent)) 1390 return false; 1391 } 1392 } 1393 1394 // Be conservative for now. FP select instruction can often be expensive. 1395 Value *BrCond = BI->getCondition(); 1396 if (isa<FCmpInst>(BrCond)) 1397 return false; 1398 1399 // If BB1 is actually on the false edge of the conditional branch, remember 1400 // to swap the select operands later. 1401 bool Invert = false; 1402 if (BB1 != BI->getSuccessor(0)) { 1403 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1404 Invert = true; 1405 } 1406 1407 // Collect interesting PHIs, and scan for hazards. 1408 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs; 1409 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1410 for (BasicBlock::iterator I = BB2->begin(); 1411 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1412 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1413 Value *BIParentV = PN->getIncomingValueForBlock(BIParent); 1414 1415 // Skip PHIs which are trivial. 1416 if (BB1V == BIParentV) 1417 continue; 1418 1419 // Check for saftey. 1420 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) { 1421 // An unfolded ConstantExpr could end up getting expanded into 1422 // Instructions. Don't speculate this and another instruction at 1423 // the same time. 1424 if (HInst) 1425 return false; 1426 if (!isSafeToSpeculativelyExecute(CE)) 1427 return false; 1428 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1429 return false; 1430 } 1431 1432 // Ok, we may insert a select for this PHI. 1433 PHIs.insert(std::make_pair(BB1V, BIParentV)); 1434 } 1435 1436 // If there are no PHIs to process, bail early. This helps ensure idempotence 1437 // as well. 1438 if (PHIs.empty()) 1439 return false; 1440 1441 // If we get here, we can hoist the instruction and if-convert. 1442 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); 1443 1444 // Hoist the instruction. 1445 if (HInst) 1446 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst); 1447 1448 // Insert selects and rewrite the PHI operands. 1449 IRBuilder<true, NoFolder> Builder(BI); 1450 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1451 Value *TrueV = PHIs[i].first; 1452 Value *FalseV = PHIs[i].second; 1453 1454 // Create a select whose true value is the speculatively executed value and 1455 // false value is the previously determined FalseV. 1456 SelectInst *SI; 1457 if (Invert) 1458 SI = cast<SelectInst> 1459 (Builder.CreateSelect(BrCond, FalseV, TrueV, 1460 FalseV->getName() + "." + TrueV->getName())); 1461 else 1462 SI = cast<SelectInst> 1463 (Builder.CreateSelect(BrCond, TrueV, FalseV, 1464 TrueV->getName() + "." + FalseV->getName())); 1465 1466 // Make the PHI node use the select for all incoming values for "then" and 1467 // "if" blocks. 1468 for (BasicBlock::iterator I = BB2->begin(); 1469 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1470 unsigned BB1I = PN->getBasicBlockIndex(BB1); 1471 unsigned BIParentI = PN->getBasicBlockIndex(BIParent); 1472 Value *BB1V = PN->getIncomingValue(BB1I); 1473 Value *BIParentV = PN->getIncomingValue(BIParentI); 1474 if (TrueV == BB1V && FalseV == BIParentV) { 1475 PN->setIncomingValue(BB1I, SI); 1476 PN->setIncomingValue(BIParentI, SI); 1477 } 1478 } 1479 } 1480 1481 ++NumSpeculations; 1482 return true; 1483} 1484 1485/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1486/// across this block. 1487static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1488 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1489 unsigned Size = 0; 1490 1491 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1492 if (isa<DbgInfoIntrinsic>(BBI)) 1493 continue; 1494 if (Size > 10) return false; // Don't clone large BB's. 1495 ++Size; 1496 1497 // We can only support instructions that do not define values that are 1498 // live outside of the current basic block. 1499 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1500 UI != E; ++UI) { 1501 Instruction *U = cast<Instruction>(*UI); 1502 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1503 } 1504 1505 // Looks ok, continue checking. 1506 } 1507 1508 return true; 1509} 1510 1511/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1512/// that is defined in the same block as the branch and if any PHI entries are 1513/// constants, thread edges corresponding to that entry to be branches to their 1514/// ultimate destination. 1515static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1516 BasicBlock *BB = BI->getParent(); 1517 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1518 // NOTE: we currently cannot transform this case if the PHI node is used 1519 // outside of the block. 1520 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1521 return false; 1522 1523 // Degenerate case of a single entry PHI. 1524 if (PN->getNumIncomingValues() == 1) { 1525 FoldSingleEntryPHINodes(PN->getParent()); 1526 return true; 1527 } 1528 1529 // Now we know that this block has multiple preds and two succs. 1530 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1531 1532 // Okay, this is a simple enough basic block. See if any phi values are 1533 // constants. 1534 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1535 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1536 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1537 1538 // Okay, we now know that all edges from PredBB should be revectored to 1539 // branch to RealDest. 1540 BasicBlock *PredBB = PN->getIncomingBlock(i); 1541 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1542 1543 if (RealDest == BB) continue; // Skip self loops. 1544 // Skip if the predecessor's terminator is an indirect branch. 1545 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1546 1547 // The dest block might have PHI nodes, other predecessors and other 1548 // difficult cases. Instead of being smart about this, just insert a new 1549 // block that jumps to the destination block, effectively splitting 1550 // the edge we are about to create. 1551 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1552 RealDest->getName()+".critedge", 1553 RealDest->getParent(), RealDest); 1554 BranchInst::Create(RealDest, EdgeBB); 1555 1556 // Update PHI nodes. 1557 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1558 1559 // BB may have instructions that are being threaded over. Clone these 1560 // instructions into EdgeBB. We know that there will be no uses of the 1561 // cloned instructions outside of EdgeBB. 1562 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1563 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1564 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1565 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1566 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1567 continue; 1568 } 1569 // Clone the instruction. 1570 Instruction *N = BBI->clone(); 1571 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1572 1573 // Update operands due to translation. 1574 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1575 i != e; ++i) { 1576 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1577 if (PI != TranslateMap.end()) 1578 *i = PI->second; 1579 } 1580 1581 // Check for trivial simplification. 1582 if (Value *V = SimplifyInstruction(N, TD)) { 1583 TranslateMap[BBI] = V; 1584 delete N; // Instruction folded away, don't need actual inst 1585 } else { 1586 // Insert the new instruction into its new home. 1587 EdgeBB->getInstList().insert(InsertPt, N); 1588 if (!BBI->use_empty()) 1589 TranslateMap[BBI] = N; 1590 } 1591 } 1592 1593 // Loop over all of the edges from PredBB to BB, changing them to branch 1594 // to EdgeBB instead. 1595 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1596 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1597 if (PredBBTI->getSuccessor(i) == BB) { 1598 BB->removePredecessor(PredBB); 1599 PredBBTI->setSuccessor(i, EdgeBB); 1600 } 1601 1602 // Recurse, simplifying any other constants. 1603 return FoldCondBranchOnPHI(BI, TD) | true; 1604 } 1605 1606 return false; 1607} 1608 1609/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1610/// PHI node, see if we can eliminate it. 1611static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1612 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1613 // statement", which has a very simple dominance structure. Basically, we 1614 // are trying to find the condition that is being branched on, which 1615 // subsequently causes this merge to happen. We really want control 1616 // dependence information for this check, but simplifycfg can't keep it up 1617 // to date, and this catches most of the cases we care about anyway. 1618 BasicBlock *BB = PN->getParent(); 1619 BasicBlock *IfTrue, *IfFalse; 1620 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1621 if (!IfCond || 1622 // Don't bother if the branch will be constant folded trivially. 1623 isa<ConstantInt>(IfCond)) 1624 return false; 1625 1626 // Okay, we found that we can merge this two-entry phi node into a select. 1627 // Doing so would require us to fold *all* two entry phi nodes in this block. 1628 // At some point this becomes non-profitable (particularly if the target 1629 // doesn't support cmov's). Only do this transformation if there are two or 1630 // fewer PHI nodes in this block. 1631 unsigned NumPhis = 0; 1632 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1633 if (NumPhis > 2) 1634 return false; 1635 1636 // Loop over the PHI's seeing if we can promote them all to select 1637 // instructions. While we are at it, keep track of the instructions 1638 // that need to be moved to the dominating block. 1639 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1640 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1641 MaxCostVal1 = PHINodeFoldingThreshold; 1642 1643 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1644 PHINode *PN = cast<PHINode>(II++); 1645 if (Value *V = SimplifyInstruction(PN, TD)) { 1646 PN->replaceAllUsesWith(V); 1647 PN->eraseFromParent(); 1648 continue; 1649 } 1650 1651 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1652 MaxCostVal0) || 1653 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1654 MaxCostVal1)) 1655 return false; 1656 } 1657 1658 // If we folded the first phi, PN dangles at this point. Refresh it. If 1659 // we ran out of PHIs then we simplified them all. 1660 PN = dyn_cast<PHINode>(BB->begin()); 1661 if (PN == 0) return true; 1662 1663 // Don't fold i1 branches on PHIs which contain binary operators. These can 1664 // often be turned into switches and other things. 1665 if (PN->getType()->isIntegerTy(1) && 1666 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1667 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1668 isa<BinaryOperator>(IfCond))) 1669 return false; 1670 1671 // If we all PHI nodes are promotable, check to make sure that all 1672 // instructions in the predecessor blocks can be promoted as well. If 1673 // not, we won't be able to get rid of the control flow, so it's not 1674 // worth promoting to select instructions. 1675 BasicBlock *DomBlock = 0; 1676 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1677 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1678 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1679 IfBlock1 = 0; 1680 } else { 1681 DomBlock = *pred_begin(IfBlock1); 1682 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1683 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1684 // This is not an aggressive instruction that we can promote. 1685 // Because of this, we won't be able to get rid of the control 1686 // flow, so the xform is not worth it. 1687 return false; 1688 } 1689 } 1690 1691 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1692 IfBlock2 = 0; 1693 } else { 1694 DomBlock = *pred_begin(IfBlock2); 1695 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1696 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1697 // This is not an aggressive instruction that we can promote. 1698 // Because of this, we won't be able to get rid of the control 1699 // flow, so the xform is not worth it. 1700 return false; 1701 } 1702 } 1703 1704 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1705 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1706 1707 // If we can still promote the PHI nodes after this gauntlet of tests, 1708 // do all of the PHI's now. 1709 Instruction *InsertPt = DomBlock->getTerminator(); 1710 IRBuilder<true, NoFolder> Builder(InsertPt); 1711 1712 // Move all 'aggressive' instructions, which are defined in the 1713 // conditional parts of the if's up to the dominating block. 1714 if (IfBlock1) 1715 DomBlock->getInstList().splice(InsertPt, 1716 IfBlock1->getInstList(), IfBlock1->begin(), 1717 IfBlock1->getTerminator()); 1718 if (IfBlock2) 1719 DomBlock->getInstList().splice(InsertPt, 1720 IfBlock2->getInstList(), IfBlock2->begin(), 1721 IfBlock2->getTerminator()); 1722 1723 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1724 // Change the PHI node into a select instruction. 1725 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1726 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1727 1728 SelectInst *NV = 1729 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1730 PN->replaceAllUsesWith(NV); 1731 NV->takeName(PN); 1732 PN->eraseFromParent(); 1733 } 1734 1735 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1736 // has been flattened. Change DomBlock to jump directly to our new block to 1737 // avoid other simplifycfg's kicking in on the diamond. 1738 TerminatorInst *OldTI = DomBlock->getTerminator(); 1739 Builder.SetInsertPoint(OldTI); 1740 Builder.CreateBr(BB); 1741 OldTI->eraseFromParent(); 1742 return true; 1743} 1744 1745/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1746/// to two returning blocks, try to merge them together into one return, 1747/// introducing a select if the return values disagree. 1748static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1749 IRBuilder<> &Builder) { 1750 assert(BI->isConditional() && "Must be a conditional branch"); 1751 BasicBlock *TrueSucc = BI->getSuccessor(0); 1752 BasicBlock *FalseSucc = BI->getSuccessor(1); 1753 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1754 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1755 1756 // Check to ensure both blocks are empty (just a return) or optionally empty 1757 // with PHI nodes. If there are other instructions, merging would cause extra 1758 // computation on one path or the other. 1759 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1760 return false; 1761 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1762 return false; 1763 1764 Builder.SetInsertPoint(BI); 1765 // Okay, we found a branch that is going to two return nodes. If 1766 // there is no return value for this function, just change the 1767 // branch into a return. 1768 if (FalseRet->getNumOperands() == 0) { 1769 TrueSucc->removePredecessor(BI->getParent()); 1770 FalseSucc->removePredecessor(BI->getParent()); 1771 Builder.CreateRetVoid(); 1772 EraseTerminatorInstAndDCECond(BI); 1773 return true; 1774 } 1775 1776 // Otherwise, figure out what the true and false return values are 1777 // so we can insert a new select instruction. 1778 Value *TrueValue = TrueRet->getReturnValue(); 1779 Value *FalseValue = FalseRet->getReturnValue(); 1780 1781 // Unwrap any PHI nodes in the return blocks. 1782 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1783 if (TVPN->getParent() == TrueSucc) 1784 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1785 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1786 if (FVPN->getParent() == FalseSucc) 1787 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1788 1789 // In order for this transformation to be safe, we must be able to 1790 // unconditionally execute both operands to the return. This is 1791 // normally the case, but we could have a potentially-trapping 1792 // constant expression that prevents this transformation from being 1793 // safe. 1794 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1795 if (TCV->canTrap()) 1796 return false; 1797 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1798 if (FCV->canTrap()) 1799 return false; 1800 1801 // Okay, we collected all the mapped values and checked them for sanity, and 1802 // defined to really do this transformation. First, update the CFG. 1803 TrueSucc->removePredecessor(BI->getParent()); 1804 FalseSucc->removePredecessor(BI->getParent()); 1805 1806 // Insert select instructions where needed. 1807 Value *BrCond = BI->getCondition(); 1808 if (TrueValue) { 1809 // Insert a select if the results differ. 1810 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1811 } else if (isa<UndefValue>(TrueValue)) { 1812 TrueValue = FalseValue; 1813 } else { 1814 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1815 FalseValue, "retval"); 1816 } 1817 } 1818 1819 Value *RI = !TrueValue ? 1820 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1821 1822 (void) RI; 1823 1824 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1825 << "\n " << *BI << "NewRet = " << *RI 1826 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1827 1828 EraseTerminatorInstAndDCECond(BI); 1829 1830 return true; 1831} 1832 1833/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1834/// probabilities of the branch taking each edge. Fills in the two APInt 1835/// parameters and return true, or returns false if no or invalid metadata was 1836/// found. 1837static bool ExtractBranchMetadata(BranchInst *BI, 1838 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1839 assert(BI->isConditional() && 1840 "Looking for probabilities on unconditional branch?"); 1841 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1842 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1843 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1844 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1845 if (!CITrue || !CIFalse) return false; 1846 ProbTrue = CITrue->getValue().getZExtValue(); 1847 ProbFalse = CIFalse->getValue().getZExtValue(); 1848 return true; 1849} 1850 1851/// checkCSEInPredecessor - Return true if the given instruction is available 1852/// in its predecessor block. If yes, the instruction will be removed. 1853/// 1854static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1855 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1856 return false; 1857 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1858 Instruction *PBI = &*I; 1859 // Check whether Inst and PBI generate the same value. 1860 if (Inst->isIdenticalTo(PBI)) { 1861 Inst->replaceAllUsesWith(PBI); 1862 Inst->eraseFromParent(); 1863 return true; 1864 } 1865 } 1866 return false; 1867} 1868 1869/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1870/// predecessor branches to us and one of our successors, fold the block into 1871/// the predecessor and use logical operations to pick the right destination. 1872bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1873 BasicBlock *BB = BI->getParent(); 1874 1875 Instruction *Cond = 0; 1876 if (BI->isConditional()) 1877 Cond = dyn_cast<Instruction>(BI->getCondition()); 1878 else { 1879 // For unconditional branch, check for a simple CFG pattern, where 1880 // BB has a single predecessor and BB's successor is also its predecessor's 1881 // successor. If such pattern exisits, check for CSE between BB and its 1882 // predecessor. 1883 if (BasicBlock *PB = BB->getSinglePredecessor()) 1884 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1885 if (PBI->isConditional() && 1886 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1887 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1888 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1889 I != E; ) { 1890 Instruction *Curr = I++; 1891 if (isa<CmpInst>(Curr)) { 1892 Cond = Curr; 1893 break; 1894 } 1895 // Quit if we can't remove this instruction. 1896 if (!checkCSEInPredecessor(Curr, PB)) 1897 return false; 1898 } 1899 } 1900 1901 if (Cond == 0) 1902 return false; 1903 } 1904 1905 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1906 Cond->getParent() != BB || !Cond->hasOneUse()) 1907 return false; 1908 1909 // Only allow this if the condition is a simple instruction that can be 1910 // executed unconditionally. It must be in the same block as the branch, and 1911 // must be at the front of the block. 1912 BasicBlock::iterator FrontIt = BB->front(); 1913 1914 // Ignore dbg intrinsics. 1915 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1916 1917 // Allow a single instruction to be hoisted in addition to the compare 1918 // that feeds the branch. We later ensure that any values that _it_ uses 1919 // were also live in the predecessor, so that we don't unnecessarily create 1920 // register pressure or inhibit out-of-order execution. 1921 Instruction *BonusInst = 0; 1922 if (&*FrontIt != Cond && 1923 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1924 isSafeToSpeculativelyExecute(FrontIt)) { 1925 BonusInst = &*FrontIt; 1926 ++FrontIt; 1927 1928 // Ignore dbg intrinsics. 1929 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1930 } 1931 1932 // Only a single bonus inst is allowed. 1933 if (&*FrontIt != Cond) 1934 return false; 1935 1936 // Make sure the instruction after the condition is the cond branch. 1937 BasicBlock::iterator CondIt = Cond; ++CondIt; 1938 1939 // Ingore dbg intrinsics. 1940 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1941 1942 if (&*CondIt != BI) 1943 return false; 1944 1945 // Cond is known to be a compare or binary operator. Check to make sure that 1946 // neither operand is a potentially-trapping constant expression. 1947 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1948 if (CE->canTrap()) 1949 return false; 1950 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1951 if (CE->canTrap()) 1952 return false; 1953 1954 // Finally, don't infinitely unroll conditional loops. 1955 BasicBlock *TrueDest = BI->getSuccessor(0); 1956 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 1957 if (TrueDest == BB || FalseDest == BB) 1958 return false; 1959 1960 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1961 BasicBlock *PredBlock = *PI; 1962 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1963 1964 // Check that we have two conditional branches. If there is a PHI node in 1965 // the common successor, verify that the same value flows in from both 1966 // blocks. 1967 SmallVector<PHINode*, 4> PHIs; 1968 if (PBI == 0 || PBI->isUnconditional() || 1969 (BI->isConditional() && 1970 !SafeToMergeTerminators(BI, PBI)) || 1971 (!BI->isConditional() && 1972 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 1973 continue; 1974 1975 // Determine if the two branches share a common destination. 1976 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 1977 bool InvertPredCond = false; 1978 1979 if (BI->isConditional()) { 1980 if (PBI->getSuccessor(0) == TrueDest) 1981 Opc = Instruction::Or; 1982 else if (PBI->getSuccessor(1) == FalseDest) 1983 Opc = Instruction::And; 1984 else if (PBI->getSuccessor(0) == FalseDest) 1985 Opc = Instruction::And, InvertPredCond = true; 1986 else if (PBI->getSuccessor(1) == TrueDest) 1987 Opc = Instruction::Or, InvertPredCond = true; 1988 else 1989 continue; 1990 } else { 1991 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 1992 continue; 1993 } 1994 1995 // Ensure that any values used in the bonus instruction are also used 1996 // by the terminator of the predecessor. This means that those values 1997 // must already have been resolved, so we won't be inhibiting the 1998 // out-of-order core by speculating them earlier. 1999 if (BonusInst) { 2000 // Collect the values used by the bonus inst 2001 SmallPtrSet<Value*, 4> UsedValues; 2002 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2003 OE = BonusInst->op_end(); OI != OE; ++OI) { 2004 Value *V = *OI; 2005 if (!isa<Constant>(V)) 2006 UsedValues.insert(V); 2007 } 2008 2009 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2010 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2011 2012 // Walk up to four levels back up the use-def chain of the predecessor's 2013 // terminator to see if all those values were used. The choice of four 2014 // levels is arbitrary, to provide a compile-time-cost bound. 2015 while (!Worklist.empty()) { 2016 std::pair<Value*, unsigned> Pair = Worklist.back(); 2017 Worklist.pop_back(); 2018 2019 if (Pair.second >= 4) continue; 2020 UsedValues.erase(Pair.first); 2021 if (UsedValues.empty()) break; 2022 2023 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2024 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2025 OI != OE; ++OI) 2026 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2027 } 2028 } 2029 2030 if (!UsedValues.empty()) return false; 2031 } 2032 2033 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2034 IRBuilder<> Builder(PBI); 2035 2036 // If we need to invert the condition in the pred block to match, do so now. 2037 if (InvertPredCond) { 2038 Value *NewCond = PBI->getCondition(); 2039 2040 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2041 CmpInst *CI = cast<CmpInst>(NewCond); 2042 CI->setPredicate(CI->getInversePredicate()); 2043 } else { 2044 NewCond = Builder.CreateNot(NewCond, 2045 PBI->getCondition()->getName()+".not"); 2046 } 2047 2048 PBI->setCondition(NewCond); 2049 PBI->swapSuccessors(); 2050 } 2051 2052 // If we have a bonus inst, clone it into the predecessor block. 2053 Instruction *NewBonus = 0; 2054 if (BonusInst) { 2055 NewBonus = BonusInst->clone(); 2056 PredBlock->getInstList().insert(PBI, NewBonus); 2057 NewBonus->takeName(BonusInst); 2058 BonusInst->setName(BonusInst->getName()+".old"); 2059 } 2060 2061 // Clone Cond into the predecessor basic block, and or/and the 2062 // two conditions together. 2063 Instruction *New = Cond->clone(); 2064 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2065 PredBlock->getInstList().insert(PBI, New); 2066 New->takeName(Cond); 2067 Cond->setName(New->getName()+".old"); 2068 2069 if (BI->isConditional()) { 2070 Instruction *NewCond = 2071 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2072 New, "or.cond")); 2073 PBI->setCondition(NewCond); 2074 2075 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2076 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2077 PredFalseWeight); 2078 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2079 SuccFalseWeight); 2080 SmallVector<uint64_t, 8> NewWeights; 2081 2082 if (PBI->getSuccessor(0) == BB) { 2083 if (PredHasWeights && SuccHasWeights) { 2084 // PBI: br i1 %x, BB, FalseDest 2085 // BI: br i1 %y, TrueDest, FalseDest 2086 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2087 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2088 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2089 // TrueWeight for PBI * FalseWeight for BI. 2090 // We assume that total weights of a BranchInst can fit into 32 bits. 2091 // Therefore, we will not have overflow using 64-bit arithmetic. 2092 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2093 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2094 } 2095 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2096 PBI->setSuccessor(0, TrueDest); 2097 } 2098 if (PBI->getSuccessor(1) == BB) { 2099 if (PredHasWeights && SuccHasWeights) { 2100 // PBI: br i1 %x, TrueDest, BB 2101 // BI: br i1 %y, TrueDest, FalseDest 2102 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2103 // FalseWeight for PBI * TrueWeight for BI. 2104 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2105 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2106 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2107 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2108 } 2109 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2110 PBI->setSuccessor(1, FalseDest); 2111 } 2112 if (NewWeights.size() == 2) { 2113 // Halve the weights if any of them cannot fit in an uint32_t 2114 FitWeights(NewWeights); 2115 2116 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2117 PBI->setMetadata(LLVMContext::MD_prof, 2118 MDBuilder(BI->getContext()). 2119 createBranchWeights(MDWeights)); 2120 } else 2121 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2122 } else { 2123 // Update PHI nodes in the common successors. 2124 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2125 ConstantInt *PBI_C = cast<ConstantInt>( 2126 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2127 assert(PBI_C->getType()->isIntegerTy(1)); 2128 Instruction *MergedCond = 0; 2129 if (PBI->getSuccessor(0) == TrueDest) { 2130 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2131 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2132 // is false: !PBI_Cond and BI_Value 2133 Instruction *NotCond = 2134 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2135 "not.cond")); 2136 MergedCond = 2137 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2138 NotCond, New, 2139 "and.cond")); 2140 if (PBI_C->isOne()) 2141 MergedCond = 2142 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2143 PBI->getCondition(), MergedCond, 2144 "or.cond")); 2145 } else { 2146 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2147 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2148 // is false: PBI_Cond and BI_Value 2149 MergedCond = 2150 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2151 PBI->getCondition(), New, 2152 "and.cond")); 2153 if (PBI_C->isOne()) { 2154 Instruction *NotCond = 2155 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2156 "not.cond")); 2157 MergedCond = 2158 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2159 NotCond, MergedCond, 2160 "or.cond")); 2161 } 2162 } 2163 // Update PHI Node. 2164 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2165 MergedCond); 2166 } 2167 // Change PBI from Conditional to Unconditional. 2168 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2169 EraseTerminatorInstAndDCECond(PBI); 2170 PBI = New_PBI; 2171 } 2172 2173 // TODO: If BB is reachable from all paths through PredBlock, then we 2174 // could replace PBI's branch probabilities with BI's. 2175 2176 // Copy any debug value intrinsics into the end of PredBlock. 2177 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2178 if (isa<DbgInfoIntrinsic>(*I)) 2179 I->clone()->insertBefore(PBI); 2180 2181 return true; 2182 } 2183 return false; 2184} 2185 2186/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2187/// predecessor of another block, this function tries to simplify it. We know 2188/// that PBI and BI are both conditional branches, and BI is in one of the 2189/// successor blocks of PBI - PBI branches to BI. 2190static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2191 assert(PBI->isConditional() && BI->isConditional()); 2192 BasicBlock *BB = BI->getParent(); 2193 2194 // If this block ends with a branch instruction, and if there is a 2195 // predecessor that ends on a branch of the same condition, make 2196 // this conditional branch redundant. 2197 if (PBI->getCondition() == BI->getCondition() && 2198 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2199 // Okay, the outcome of this conditional branch is statically 2200 // knowable. If this block had a single pred, handle specially. 2201 if (BB->getSinglePredecessor()) { 2202 // Turn this into a branch on constant. 2203 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2204 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2205 CondIsTrue)); 2206 return true; // Nuke the branch on constant. 2207 } 2208 2209 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2210 // in the constant and simplify the block result. Subsequent passes of 2211 // simplifycfg will thread the block. 2212 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2213 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2214 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2215 std::distance(PB, PE), 2216 BI->getCondition()->getName() + ".pr", 2217 BB->begin()); 2218 // Okay, we're going to insert the PHI node. Since PBI is not the only 2219 // predecessor, compute the PHI'd conditional value for all of the preds. 2220 // Any predecessor where the condition is not computable we keep symbolic. 2221 for (pred_iterator PI = PB; PI != PE; ++PI) { 2222 BasicBlock *P = *PI; 2223 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2224 PBI != BI && PBI->isConditional() && 2225 PBI->getCondition() == BI->getCondition() && 2226 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2227 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2228 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2229 CondIsTrue), P); 2230 } else { 2231 NewPN->addIncoming(BI->getCondition(), P); 2232 } 2233 } 2234 2235 BI->setCondition(NewPN); 2236 return true; 2237 } 2238 } 2239 2240 // If this is a conditional branch in an empty block, and if any 2241 // predecessors is a conditional branch to one of our destinations, 2242 // fold the conditions into logical ops and one cond br. 2243 BasicBlock::iterator BBI = BB->begin(); 2244 // Ignore dbg intrinsics. 2245 while (isa<DbgInfoIntrinsic>(BBI)) 2246 ++BBI; 2247 if (&*BBI != BI) 2248 return false; 2249 2250 2251 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2252 if (CE->canTrap()) 2253 return false; 2254 2255 int PBIOp, BIOp; 2256 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2257 PBIOp = BIOp = 0; 2258 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2259 PBIOp = 0, BIOp = 1; 2260 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2261 PBIOp = 1, BIOp = 0; 2262 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2263 PBIOp = BIOp = 1; 2264 else 2265 return false; 2266 2267 // Check to make sure that the other destination of this branch 2268 // isn't BB itself. If so, this is an infinite loop that will 2269 // keep getting unwound. 2270 if (PBI->getSuccessor(PBIOp) == BB) 2271 return false; 2272 2273 // Do not perform this transformation if it would require 2274 // insertion of a large number of select instructions. For targets 2275 // without predication/cmovs, this is a big pessimization. 2276 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2277 2278 unsigned NumPhis = 0; 2279 for (BasicBlock::iterator II = CommonDest->begin(); 2280 isa<PHINode>(II); ++II, ++NumPhis) 2281 if (NumPhis > 2) // Disable this xform. 2282 return false; 2283 2284 // Finally, if everything is ok, fold the branches to logical ops. 2285 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2286 2287 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2288 << "AND: " << *BI->getParent()); 2289 2290 2291 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2292 // branch in it, where one edge (OtherDest) goes back to itself but the other 2293 // exits. We don't *know* that the program avoids the infinite loop 2294 // (even though that seems likely). If we do this xform naively, we'll end up 2295 // recursively unpeeling the loop. Since we know that (after the xform is 2296 // done) that the block *is* infinite if reached, we just make it an obviously 2297 // infinite loop with no cond branch. 2298 if (OtherDest == BB) { 2299 // Insert it at the end of the function, because it's either code, 2300 // or it won't matter if it's hot. :) 2301 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2302 "infloop", BB->getParent()); 2303 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2304 OtherDest = InfLoopBlock; 2305 } 2306 2307 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2308 2309 // BI may have other predecessors. Because of this, we leave 2310 // it alone, but modify PBI. 2311 2312 // Make sure we get to CommonDest on True&True directions. 2313 Value *PBICond = PBI->getCondition(); 2314 IRBuilder<true, NoFolder> Builder(PBI); 2315 if (PBIOp) 2316 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2317 2318 Value *BICond = BI->getCondition(); 2319 if (BIOp) 2320 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2321 2322 // Merge the conditions. 2323 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2324 2325 // Modify PBI to branch on the new condition to the new dests. 2326 PBI->setCondition(Cond); 2327 PBI->setSuccessor(0, CommonDest); 2328 PBI->setSuccessor(1, OtherDest); 2329 2330 // Update branch weight for PBI. 2331 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2332 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2333 PredFalseWeight); 2334 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2335 SuccFalseWeight); 2336 if (PredHasWeights && SuccHasWeights) { 2337 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2338 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2339 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2340 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2341 // The weight to CommonDest should be PredCommon * SuccTotal + 2342 // PredOther * SuccCommon. 2343 // The weight to OtherDest should be PredOther * SuccOther. 2344 SmallVector<uint64_t, 2> NewWeights; 2345 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2346 PredOther * SuccCommon); 2347 NewWeights.push_back(PredOther * SuccOther); 2348 // Halve the weights if any of them cannot fit in an uint32_t 2349 FitWeights(NewWeights); 2350 2351 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2352 PBI->setMetadata(LLVMContext::MD_prof, 2353 MDBuilder(BI->getContext()). 2354 createBranchWeights(MDWeights)); 2355 } 2356 2357 // OtherDest may have phi nodes. If so, add an entry from PBI's 2358 // block that are identical to the entries for BI's block. 2359 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2360 2361 // We know that the CommonDest already had an edge from PBI to 2362 // it. If it has PHIs though, the PHIs may have different 2363 // entries for BB and PBI's BB. If so, insert a select to make 2364 // them agree. 2365 PHINode *PN; 2366 for (BasicBlock::iterator II = CommonDest->begin(); 2367 (PN = dyn_cast<PHINode>(II)); ++II) { 2368 Value *BIV = PN->getIncomingValueForBlock(BB); 2369 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2370 Value *PBIV = PN->getIncomingValue(PBBIdx); 2371 if (BIV != PBIV) { 2372 // Insert a select in PBI to pick the right value. 2373 Value *NV = cast<SelectInst> 2374 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2375 PN->setIncomingValue(PBBIdx, NV); 2376 } 2377 } 2378 2379 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2380 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2381 2382 // This basic block is probably dead. We know it has at least 2383 // one fewer predecessor. 2384 return true; 2385} 2386 2387// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2388// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2389// Takes care of updating the successors and removing the old terminator. 2390// Also makes sure not to introduce new successors by assuming that edges to 2391// non-successor TrueBBs and FalseBBs aren't reachable. 2392static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2393 BasicBlock *TrueBB, BasicBlock *FalseBB, 2394 uint32_t TrueWeight, 2395 uint32_t FalseWeight){ 2396 // Remove any superfluous successor edges from the CFG. 2397 // First, figure out which successors to preserve. 2398 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2399 // successor. 2400 BasicBlock *KeepEdge1 = TrueBB; 2401 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2402 2403 // Then remove the rest. 2404 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2405 BasicBlock *Succ = OldTerm->getSuccessor(I); 2406 // Make sure only to keep exactly one copy of each edge. 2407 if (Succ == KeepEdge1) 2408 KeepEdge1 = 0; 2409 else if (Succ == KeepEdge2) 2410 KeepEdge2 = 0; 2411 else 2412 Succ->removePredecessor(OldTerm->getParent()); 2413 } 2414 2415 IRBuilder<> Builder(OldTerm); 2416 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2417 2418 // Insert an appropriate new terminator. 2419 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2420 if (TrueBB == FalseBB) 2421 // We were only looking for one successor, and it was present. 2422 // Create an unconditional branch to it. 2423 Builder.CreateBr(TrueBB); 2424 else { 2425 // We found both of the successors we were looking for. 2426 // Create a conditional branch sharing the condition of the select. 2427 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2428 if (TrueWeight != FalseWeight) 2429 NewBI->setMetadata(LLVMContext::MD_prof, 2430 MDBuilder(OldTerm->getContext()). 2431 createBranchWeights(TrueWeight, FalseWeight)); 2432 } 2433 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2434 // Neither of the selected blocks were successors, so this 2435 // terminator must be unreachable. 2436 new UnreachableInst(OldTerm->getContext(), OldTerm); 2437 } else { 2438 // One of the selected values was a successor, but the other wasn't. 2439 // Insert an unconditional branch to the one that was found; 2440 // the edge to the one that wasn't must be unreachable. 2441 if (KeepEdge1 == 0) 2442 // Only TrueBB was found. 2443 Builder.CreateBr(TrueBB); 2444 else 2445 // Only FalseBB was found. 2446 Builder.CreateBr(FalseBB); 2447 } 2448 2449 EraseTerminatorInstAndDCECond(OldTerm); 2450 return true; 2451} 2452 2453// SimplifySwitchOnSelect - Replaces 2454// (switch (select cond, X, Y)) on constant X, Y 2455// with a branch - conditional if X and Y lead to distinct BBs, 2456// unconditional otherwise. 2457static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2458 // Check for constant integer values in the select. 2459 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2460 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2461 if (!TrueVal || !FalseVal) 2462 return false; 2463 2464 // Find the relevant condition and destinations. 2465 Value *Condition = Select->getCondition(); 2466 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2467 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2468 2469 // Get weight for TrueBB and FalseBB. 2470 uint32_t TrueWeight = 0, FalseWeight = 0; 2471 SmallVector<uint64_t, 8> Weights; 2472 bool HasWeights = HasBranchWeights(SI); 2473 if (HasWeights) { 2474 GetBranchWeights(SI, Weights); 2475 if (Weights.size() == 1 + SI->getNumCases()) { 2476 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2477 getSuccessorIndex()]; 2478 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2479 getSuccessorIndex()]; 2480 } 2481 } 2482 2483 // Perform the actual simplification. 2484 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2485 TrueWeight, FalseWeight); 2486} 2487 2488// SimplifyIndirectBrOnSelect - Replaces 2489// (indirectbr (select cond, blockaddress(@fn, BlockA), 2490// blockaddress(@fn, BlockB))) 2491// with 2492// (br cond, BlockA, BlockB). 2493static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2494 // Check that both operands of the select are block addresses. 2495 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2496 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2497 if (!TBA || !FBA) 2498 return false; 2499 2500 // Extract the actual blocks. 2501 BasicBlock *TrueBB = TBA->getBasicBlock(); 2502 BasicBlock *FalseBB = FBA->getBasicBlock(); 2503 2504 // Perform the actual simplification. 2505 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2506 0, 0); 2507} 2508 2509/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2510/// instruction (a seteq/setne with a constant) as the only instruction in a 2511/// block that ends with an uncond branch. We are looking for a very specific 2512/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2513/// this case, we merge the first two "or's of icmp" into a switch, but then the 2514/// default value goes to an uncond block with a seteq in it, we get something 2515/// like: 2516/// 2517/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2518/// DEFAULT: 2519/// %tmp = icmp eq i8 %A, 92 2520/// br label %end 2521/// end: 2522/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2523/// 2524/// We prefer to split the edge to 'end' so that there is a true/false entry to 2525/// the PHI, merging the third icmp into the switch. 2526static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 2527 const DataLayout *TD, 2528 IRBuilder<> &Builder) { 2529 BasicBlock *BB = ICI->getParent(); 2530 2531 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2532 // complex. 2533 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2534 2535 Value *V = ICI->getOperand(0); 2536 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2537 2538 // The pattern we're looking for is where our only predecessor is a switch on 2539 // 'V' and this block is the default case for the switch. In this case we can 2540 // fold the compared value into the switch to simplify things. 2541 BasicBlock *Pred = BB->getSinglePredecessor(); 2542 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2543 2544 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2545 if (SI->getCondition() != V) 2546 return false; 2547 2548 // If BB is reachable on a non-default case, then we simply know the value of 2549 // V in this block. Substitute it and constant fold the icmp instruction 2550 // away. 2551 if (SI->getDefaultDest() != BB) { 2552 ConstantInt *VVal = SI->findCaseDest(BB); 2553 assert(VVal && "Should have a unique destination value"); 2554 ICI->setOperand(0, VVal); 2555 2556 if (Value *V = SimplifyInstruction(ICI, TD)) { 2557 ICI->replaceAllUsesWith(V); 2558 ICI->eraseFromParent(); 2559 } 2560 // BB is now empty, so it is likely to simplify away. 2561 return SimplifyCFG(BB) | true; 2562 } 2563 2564 // Ok, the block is reachable from the default dest. If the constant we're 2565 // comparing exists in one of the other edges, then we can constant fold ICI 2566 // and zap it. 2567 if (SI->findCaseValue(Cst) != SI->case_default()) { 2568 Value *V; 2569 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2570 V = ConstantInt::getFalse(BB->getContext()); 2571 else 2572 V = ConstantInt::getTrue(BB->getContext()); 2573 2574 ICI->replaceAllUsesWith(V); 2575 ICI->eraseFromParent(); 2576 // BB is now empty, so it is likely to simplify away. 2577 return SimplifyCFG(BB) | true; 2578 } 2579 2580 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2581 // the block. 2582 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2583 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2584 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2585 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2586 return false; 2587 2588 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2589 // true in the PHI. 2590 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2591 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2592 2593 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2594 std::swap(DefaultCst, NewCst); 2595 2596 // Replace ICI (which is used by the PHI for the default value) with true or 2597 // false depending on if it is EQ or NE. 2598 ICI->replaceAllUsesWith(DefaultCst); 2599 ICI->eraseFromParent(); 2600 2601 // Okay, the switch goes to this block on a default value. Add an edge from 2602 // the switch to the merge point on the compared value. 2603 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2604 BB->getParent(), BB); 2605 SmallVector<uint64_t, 8> Weights; 2606 bool HasWeights = HasBranchWeights(SI); 2607 if (HasWeights) { 2608 GetBranchWeights(SI, Weights); 2609 if (Weights.size() == 1 + SI->getNumCases()) { 2610 // Split weight for default case to case for "Cst". 2611 Weights[0] = (Weights[0]+1) >> 1; 2612 Weights.push_back(Weights[0]); 2613 2614 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2615 SI->setMetadata(LLVMContext::MD_prof, 2616 MDBuilder(SI->getContext()). 2617 createBranchWeights(MDWeights)); 2618 } 2619 } 2620 SI->addCase(Cst, NewBB); 2621 2622 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2623 Builder.SetInsertPoint(NewBB); 2624 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2625 Builder.CreateBr(SuccBlock); 2626 PHIUse->addIncoming(NewCst, NewBB); 2627 return true; 2628} 2629 2630/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2631/// Check to see if it is branching on an or/and chain of icmp instructions, and 2632/// fold it into a switch instruction if so. 2633static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2634 IRBuilder<> &Builder) { 2635 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2636 if (Cond == 0) return false; 2637 2638 2639 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2640 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2641 // 'setne's and'ed together, collect them. 2642 Value *CompVal = 0; 2643 std::vector<ConstantInt*> Values; 2644 bool TrueWhenEqual = true; 2645 Value *ExtraCase = 0; 2646 unsigned UsedICmps = 0; 2647 2648 if (Cond->getOpcode() == Instruction::Or) { 2649 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2650 UsedICmps); 2651 } else if (Cond->getOpcode() == Instruction::And) { 2652 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2653 UsedICmps); 2654 TrueWhenEqual = false; 2655 } 2656 2657 // If we didn't have a multiply compared value, fail. 2658 if (CompVal == 0) return false; 2659 2660 // Avoid turning single icmps into a switch. 2661 if (UsedICmps <= 1) 2662 return false; 2663 2664 // There might be duplicate constants in the list, which the switch 2665 // instruction can't handle, remove them now. 2666 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2667 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2668 2669 // If Extra was used, we require at least two switch values to do the 2670 // transformation. A switch with one value is just an cond branch. 2671 if (ExtraCase && Values.size() < 2) return false; 2672 2673 // TODO: Preserve branch weight metadata, similarly to how 2674 // FoldValueComparisonIntoPredecessors preserves it. 2675 2676 // Figure out which block is which destination. 2677 BasicBlock *DefaultBB = BI->getSuccessor(1); 2678 BasicBlock *EdgeBB = BI->getSuccessor(0); 2679 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2680 2681 BasicBlock *BB = BI->getParent(); 2682 2683 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2684 << " cases into SWITCH. BB is:\n" << *BB); 2685 2686 // If there are any extra values that couldn't be folded into the switch 2687 // then we evaluate them with an explicit branch first. Split the block 2688 // right before the condbr to handle it. 2689 if (ExtraCase) { 2690 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2691 // Remove the uncond branch added to the old block. 2692 TerminatorInst *OldTI = BB->getTerminator(); 2693 Builder.SetInsertPoint(OldTI); 2694 2695 if (TrueWhenEqual) 2696 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2697 else 2698 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2699 2700 OldTI->eraseFromParent(); 2701 2702 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2703 // for the edge we just added. 2704 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2705 2706 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2707 << "\nEXTRABB = " << *BB); 2708 BB = NewBB; 2709 } 2710 2711 Builder.SetInsertPoint(BI); 2712 // Convert pointer to int before we switch. 2713 if (CompVal->getType()->isPointerTy()) { 2714 assert(TD && "Cannot switch on pointer without DataLayout"); 2715 CompVal = Builder.CreatePtrToInt(CompVal, 2716 TD->getIntPtrType(CompVal->getType()), 2717 "magicptr"); 2718 } 2719 2720 // Create the new switch instruction now. 2721 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2722 2723 // Add all of the 'cases' to the switch instruction. 2724 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2725 New->addCase(Values[i], EdgeBB); 2726 2727 // We added edges from PI to the EdgeBB. As such, if there were any 2728 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2729 // the number of edges added. 2730 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2731 isa<PHINode>(BBI); ++BBI) { 2732 PHINode *PN = cast<PHINode>(BBI); 2733 Value *InVal = PN->getIncomingValueForBlock(BB); 2734 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2735 PN->addIncoming(InVal, BB); 2736 } 2737 2738 // Erase the old branch instruction. 2739 EraseTerminatorInstAndDCECond(BI); 2740 2741 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2742 return true; 2743} 2744 2745bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2746 // If this is a trivial landing pad that just continues unwinding the caught 2747 // exception then zap the landing pad, turning its invokes into calls. 2748 BasicBlock *BB = RI->getParent(); 2749 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2750 if (RI->getValue() != LPInst) 2751 // Not a landing pad, or the resume is not unwinding the exception that 2752 // caused control to branch here. 2753 return false; 2754 2755 // Check that there are no other instructions except for debug intrinsics. 2756 BasicBlock::iterator I = LPInst, E = RI; 2757 while (++I != E) 2758 if (!isa<DbgInfoIntrinsic>(I)) 2759 return false; 2760 2761 // Turn all invokes that unwind here into calls and delete the basic block. 2762 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2763 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2764 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2765 // Insert a call instruction before the invoke. 2766 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2767 Call->takeName(II); 2768 Call->setCallingConv(II->getCallingConv()); 2769 Call->setAttributes(II->getAttributes()); 2770 Call->setDebugLoc(II->getDebugLoc()); 2771 2772 // Anything that used the value produced by the invoke instruction now uses 2773 // the value produced by the call instruction. Note that we do this even 2774 // for void functions and calls with no uses so that the callgraph edge is 2775 // updated. 2776 II->replaceAllUsesWith(Call); 2777 BB->removePredecessor(II->getParent()); 2778 2779 // Insert a branch to the normal destination right before the invoke. 2780 BranchInst::Create(II->getNormalDest(), II); 2781 2782 // Finally, delete the invoke instruction! 2783 II->eraseFromParent(); 2784 } 2785 2786 // The landingpad is now unreachable. Zap it. 2787 BB->eraseFromParent(); 2788 return true; 2789} 2790 2791bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2792 BasicBlock *BB = RI->getParent(); 2793 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2794 2795 // Find predecessors that end with branches. 2796 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2797 SmallVector<BranchInst*, 8> CondBranchPreds; 2798 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2799 BasicBlock *P = *PI; 2800 TerminatorInst *PTI = P->getTerminator(); 2801 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2802 if (BI->isUnconditional()) 2803 UncondBranchPreds.push_back(P); 2804 else 2805 CondBranchPreds.push_back(BI); 2806 } 2807 } 2808 2809 // If we found some, do the transformation! 2810 if (!UncondBranchPreds.empty() && DupRet) { 2811 while (!UncondBranchPreds.empty()) { 2812 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2813 DEBUG(dbgs() << "FOLDING: " << *BB 2814 << "INTO UNCOND BRANCH PRED: " << *Pred); 2815 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2816 } 2817 2818 // If we eliminated all predecessors of the block, delete the block now. 2819 if (pred_begin(BB) == pred_end(BB)) 2820 // We know there are no successors, so just nuke the block. 2821 BB->eraseFromParent(); 2822 2823 return true; 2824 } 2825 2826 // Check out all of the conditional branches going to this return 2827 // instruction. If any of them just select between returns, change the 2828 // branch itself into a select/return pair. 2829 while (!CondBranchPreds.empty()) { 2830 BranchInst *BI = CondBranchPreds.pop_back_val(); 2831 2832 // Check to see if the non-BB successor is also a return block. 2833 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2834 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2835 SimplifyCondBranchToTwoReturns(BI, Builder)) 2836 return true; 2837 } 2838 return false; 2839} 2840 2841bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2842 BasicBlock *BB = UI->getParent(); 2843 2844 bool Changed = false; 2845 2846 // If there are any instructions immediately before the unreachable that can 2847 // be removed, do so. 2848 while (UI != BB->begin()) { 2849 BasicBlock::iterator BBI = UI; 2850 --BBI; 2851 // Do not delete instructions that can have side effects which might cause 2852 // the unreachable to not be reachable; specifically, calls and volatile 2853 // operations may have this effect. 2854 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2855 2856 if (BBI->mayHaveSideEffects()) { 2857 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2858 if (SI->isVolatile()) 2859 break; 2860 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2861 if (LI->isVolatile()) 2862 break; 2863 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2864 if (RMWI->isVolatile()) 2865 break; 2866 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2867 if (CXI->isVolatile()) 2868 break; 2869 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2870 !isa<LandingPadInst>(BBI)) { 2871 break; 2872 } 2873 // Note that deleting LandingPad's here is in fact okay, although it 2874 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2875 // all the predecessors of this block will be the unwind edges of Invokes, 2876 // and we can therefore guarantee this block will be erased. 2877 } 2878 2879 // Delete this instruction (any uses are guaranteed to be dead) 2880 if (!BBI->use_empty()) 2881 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2882 BBI->eraseFromParent(); 2883 Changed = true; 2884 } 2885 2886 // If the unreachable instruction is the first in the block, take a gander 2887 // at all of the predecessors of this instruction, and simplify them. 2888 if (&BB->front() != UI) return Changed; 2889 2890 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2891 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2892 TerminatorInst *TI = Preds[i]->getTerminator(); 2893 IRBuilder<> Builder(TI); 2894 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2895 if (BI->isUnconditional()) { 2896 if (BI->getSuccessor(0) == BB) { 2897 new UnreachableInst(TI->getContext(), TI); 2898 TI->eraseFromParent(); 2899 Changed = true; 2900 } 2901 } else { 2902 if (BI->getSuccessor(0) == BB) { 2903 Builder.CreateBr(BI->getSuccessor(1)); 2904 EraseTerminatorInstAndDCECond(BI); 2905 } else if (BI->getSuccessor(1) == BB) { 2906 Builder.CreateBr(BI->getSuccessor(0)); 2907 EraseTerminatorInstAndDCECond(BI); 2908 Changed = true; 2909 } 2910 } 2911 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2912 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2913 i != e; ++i) 2914 if (i.getCaseSuccessor() == BB) { 2915 BB->removePredecessor(SI->getParent()); 2916 SI->removeCase(i); 2917 --i; --e; 2918 Changed = true; 2919 } 2920 // If the default value is unreachable, figure out the most popular 2921 // destination and make it the default. 2922 if (SI->getDefaultDest() == BB) { 2923 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2924 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2925 i != e; ++i) { 2926 std::pair<unsigned, unsigned> &entry = 2927 Popularity[i.getCaseSuccessor()]; 2928 if (entry.first == 0) { 2929 entry.first = 1; 2930 entry.second = i.getCaseIndex(); 2931 } else { 2932 entry.first++; 2933 } 2934 } 2935 2936 // Find the most popular block. 2937 unsigned MaxPop = 0; 2938 unsigned MaxIndex = 0; 2939 BasicBlock *MaxBlock = 0; 2940 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2941 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2942 if (I->second.first > MaxPop || 2943 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2944 MaxPop = I->second.first; 2945 MaxIndex = I->second.second; 2946 MaxBlock = I->first; 2947 } 2948 } 2949 if (MaxBlock) { 2950 // Make this the new default, allowing us to delete any explicit 2951 // edges to it. 2952 SI->setDefaultDest(MaxBlock); 2953 Changed = true; 2954 2955 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2956 // it. 2957 if (isa<PHINode>(MaxBlock->begin())) 2958 for (unsigned i = 0; i != MaxPop-1; ++i) 2959 MaxBlock->removePredecessor(SI->getParent()); 2960 2961 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2962 i != e; ++i) 2963 if (i.getCaseSuccessor() == MaxBlock) { 2964 SI->removeCase(i); 2965 --i; --e; 2966 } 2967 } 2968 } 2969 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2970 if (II->getUnwindDest() == BB) { 2971 // Convert the invoke to a call instruction. This would be a good 2972 // place to note that the call does not throw though. 2973 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2974 II->removeFromParent(); // Take out of symbol table 2975 2976 // Insert the call now... 2977 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2978 Builder.SetInsertPoint(BI); 2979 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2980 Args, II->getName()); 2981 CI->setCallingConv(II->getCallingConv()); 2982 CI->setAttributes(II->getAttributes()); 2983 // If the invoke produced a value, the call does now instead. 2984 II->replaceAllUsesWith(CI); 2985 delete II; 2986 Changed = true; 2987 } 2988 } 2989 } 2990 2991 // If this block is now dead, remove it. 2992 if (pred_begin(BB) == pred_end(BB) && 2993 BB != &BB->getParent()->getEntryBlock()) { 2994 // We know there are no successors, so just nuke the block. 2995 BB->eraseFromParent(); 2996 return true; 2997 } 2998 2999 return Changed; 3000} 3001 3002/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3003/// integer range comparison into a sub, an icmp and a branch. 3004static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3005 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3006 3007 // Make sure all cases point to the same destination and gather the values. 3008 SmallVector<ConstantInt *, 16> Cases; 3009 SwitchInst::CaseIt I = SI->case_begin(); 3010 Cases.push_back(I.getCaseValue()); 3011 SwitchInst::CaseIt PrevI = I++; 3012 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3013 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3014 return false; 3015 Cases.push_back(I.getCaseValue()); 3016 } 3017 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3018 3019 // Sort the case values, then check if they form a range we can transform. 3020 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3021 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3022 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3023 return false; 3024 } 3025 3026 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3027 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3028 3029 Value *Sub = SI->getCondition(); 3030 if (!Offset->isNullValue()) 3031 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3032 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3033 BranchInst *NewBI = Builder.CreateCondBr( 3034 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3035 3036 // Update weight for the newly-created conditional branch. 3037 SmallVector<uint64_t, 8> Weights; 3038 bool HasWeights = HasBranchWeights(SI); 3039 if (HasWeights) { 3040 GetBranchWeights(SI, Weights); 3041 if (Weights.size() == 1 + SI->getNumCases()) { 3042 // Combine all weights for the cases to be the true weight of NewBI. 3043 // We assume that the sum of all weights for a Terminator can fit into 32 3044 // bits. 3045 uint32_t NewTrueWeight = 0; 3046 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3047 NewTrueWeight += (uint32_t)Weights[I]; 3048 NewBI->setMetadata(LLVMContext::MD_prof, 3049 MDBuilder(SI->getContext()). 3050 createBranchWeights(NewTrueWeight, 3051 (uint32_t)Weights[0])); 3052 } 3053 } 3054 3055 // Prune obsolete incoming values off the successor's PHI nodes. 3056 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3057 isa<PHINode>(BBI); ++BBI) { 3058 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3059 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3060 } 3061 SI->eraseFromParent(); 3062 3063 return true; 3064} 3065 3066/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3067/// and use it to remove dead cases. 3068static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3069 Value *Cond = SI->getCondition(); 3070 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3071 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3072 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3073 3074 // Gather dead cases. 3075 SmallVector<ConstantInt*, 8> DeadCases; 3076 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3077 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3078 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3079 DeadCases.push_back(I.getCaseValue()); 3080 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3081 << I.getCaseValue() << "' is dead.\n"); 3082 } 3083 } 3084 3085 SmallVector<uint64_t, 8> Weights; 3086 bool HasWeight = HasBranchWeights(SI); 3087 if (HasWeight) { 3088 GetBranchWeights(SI, Weights); 3089 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3090 } 3091 3092 // Remove dead cases from the switch. 3093 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3094 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3095 assert(Case != SI->case_default() && 3096 "Case was not found. Probably mistake in DeadCases forming."); 3097 if (HasWeight) { 3098 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3099 Weights.pop_back(); 3100 } 3101 3102 // Prune unused values from PHI nodes. 3103 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3104 SI->removeCase(Case); 3105 } 3106 if (HasWeight) { 3107 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3108 SI->setMetadata(LLVMContext::MD_prof, 3109 MDBuilder(SI->getParent()->getContext()). 3110 createBranchWeights(MDWeights)); 3111 } 3112 3113 return !DeadCases.empty(); 3114} 3115 3116/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3117/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3118/// by an unconditional branch), look at the phi node for BB in the successor 3119/// block and see if the incoming value is equal to CaseValue. If so, return 3120/// the phi node, and set PhiIndex to BB's index in the phi node. 3121static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3122 BasicBlock *BB, 3123 int *PhiIndex) { 3124 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3125 return NULL; // BB must be empty to be a candidate for simplification. 3126 if (!BB->getSinglePredecessor()) 3127 return NULL; // BB must be dominated by the switch. 3128 3129 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3130 if (!Branch || !Branch->isUnconditional()) 3131 return NULL; // Terminator must be unconditional branch. 3132 3133 BasicBlock *Succ = Branch->getSuccessor(0); 3134 3135 BasicBlock::iterator I = Succ->begin(); 3136 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3137 int Idx = PHI->getBasicBlockIndex(BB); 3138 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3139 3140 Value *InValue = PHI->getIncomingValue(Idx); 3141 if (InValue != CaseValue) continue; 3142 3143 *PhiIndex = Idx; 3144 return PHI; 3145 } 3146 3147 return NULL; 3148} 3149 3150/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3151/// instruction to a phi node dominated by the switch, if that would mean that 3152/// some of the destination blocks of the switch can be folded away. 3153/// Returns true if a change is made. 3154static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3155 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3156 ForwardingNodesMap ForwardingNodes; 3157 3158 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3159 ConstantInt *CaseValue = I.getCaseValue(); 3160 BasicBlock *CaseDest = I.getCaseSuccessor(); 3161 3162 int PhiIndex; 3163 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3164 &PhiIndex); 3165 if (!PHI) continue; 3166 3167 ForwardingNodes[PHI].push_back(PhiIndex); 3168 } 3169 3170 bool Changed = false; 3171 3172 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3173 E = ForwardingNodes.end(); I != E; ++I) { 3174 PHINode *Phi = I->first; 3175 SmallVector<int,4> &Indexes = I->second; 3176 3177 if (Indexes.size() < 2) continue; 3178 3179 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3180 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3181 Changed = true; 3182 } 3183 3184 return Changed; 3185} 3186 3187/// ValidLookupTableConstant - Return true if the backend will be able to handle 3188/// initializing an array of constants like C. 3189static bool ValidLookupTableConstant(Constant *C) { 3190 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3191 return CE->isGEPWithNoNotionalOverIndexing(); 3192 3193 return isa<ConstantFP>(C) || 3194 isa<ConstantInt>(C) || 3195 isa<ConstantPointerNull>(C) || 3196 isa<GlobalValue>(C) || 3197 isa<UndefValue>(C); 3198} 3199 3200/// GetCaseResulsts - Try to determine the resulting constant values in phi 3201/// nodes at the common destination basic block for one of the case 3202/// destinations of a switch instruction. 3203static bool GetCaseResults(SwitchInst *SI, 3204 BasicBlock *CaseDest, 3205 BasicBlock **CommonDest, 3206 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3207 // The block from which we enter the common destination. 3208 BasicBlock *Pred = SI->getParent(); 3209 3210 // If CaseDest is empty, continue to its successor. 3211 if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() && 3212 !isa<PHINode>(CaseDest->begin())) { 3213 3214 TerminatorInst *Terminator = CaseDest->getTerminator(); 3215 if (Terminator->getNumSuccessors() != 1) 3216 return false; 3217 3218 Pred = CaseDest; 3219 CaseDest = Terminator->getSuccessor(0); 3220 } 3221 3222 // If we did not have a CommonDest before, use the current one. 3223 if (!*CommonDest) 3224 *CommonDest = CaseDest; 3225 // If the destination isn't the common one, abort. 3226 if (CaseDest != *CommonDest) 3227 return false; 3228 3229 // Get the values for this case from phi nodes in the destination block. 3230 BasicBlock::iterator I = (*CommonDest)->begin(); 3231 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3232 int Idx = PHI->getBasicBlockIndex(Pred); 3233 if (Idx == -1) 3234 continue; 3235 3236 Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx)); 3237 if (!ConstVal) 3238 return false; 3239 3240 // Be conservative about which kinds of constants we support. 3241 if (!ValidLookupTableConstant(ConstVal)) 3242 return false; 3243 3244 Res.push_back(std::make_pair(PHI, ConstVal)); 3245 } 3246 3247 return true; 3248} 3249 3250namespace { 3251 /// SwitchLookupTable - This class represents a lookup table that can be used 3252 /// to replace a switch. 3253 class SwitchLookupTable { 3254 public: 3255 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3256 /// with the contents of Values, using DefaultValue to fill any holes in the 3257 /// table. 3258 SwitchLookupTable(Module &M, 3259 uint64_t TableSize, 3260 ConstantInt *Offset, 3261 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3262 Constant *DefaultValue, 3263 const DataLayout *TD); 3264 3265 /// BuildLookup - Build instructions with Builder to retrieve the value at 3266 /// the position given by Index in the lookup table. 3267 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3268 3269 /// WouldFitInRegister - Return true if a table with TableSize elements of 3270 /// type ElementType would fit in a target-legal register. 3271 static bool WouldFitInRegister(const DataLayout *TD, 3272 uint64_t TableSize, 3273 const Type *ElementType); 3274 3275 private: 3276 // Depending on the contents of the table, it can be represented in 3277 // different ways. 3278 enum { 3279 // For tables where each element contains the same value, we just have to 3280 // store that single value and return it for each lookup. 3281 SingleValueKind, 3282 3283 // For small tables with integer elements, we can pack them into a bitmap 3284 // that fits into a target-legal register. Values are retrieved by 3285 // shift and mask operations. 3286 BitMapKind, 3287 3288 // The table is stored as an array of values. Values are retrieved by load 3289 // instructions from the table. 3290 ArrayKind 3291 } Kind; 3292 3293 // For SingleValueKind, this is the single value. 3294 Constant *SingleValue; 3295 3296 // For BitMapKind, this is the bitmap. 3297 ConstantInt *BitMap; 3298 IntegerType *BitMapElementTy; 3299 3300 // For ArrayKind, this is the array. 3301 GlobalVariable *Array; 3302 }; 3303} 3304 3305SwitchLookupTable::SwitchLookupTable(Module &M, 3306 uint64_t TableSize, 3307 ConstantInt *Offset, 3308 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3309 Constant *DefaultValue, 3310 const DataLayout *TD) { 3311 assert(Values.size() && "Can't build lookup table without values!"); 3312 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3313 3314 // If all values in the table are equal, this is that value. 3315 SingleValue = Values.begin()->second; 3316 3317 // Build up the table contents. 3318 SmallVector<Constant*, 64> TableContents(TableSize); 3319 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3320 ConstantInt *CaseVal = Values[I].first; 3321 Constant *CaseRes = Values[I].second; 3322 assert(CaseRes->getType() == DefaultValue->getType()); 3323 3324 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3325 .getLimitedValue(); 3326 TableContents[Idx] = CaseRes; 3327 3328 if (CaseRes != SingleValue) 3329 SingleValue = NULL; 3330 } 3331 3332 // Fill in any holes in the table with the default result. 3333 if (Values.size() < TableSize) { 3334 for (uint64_t I = 0; I < TableSize; ++I) { 3335 if (!TableContents[I]) 3336 TableContents[I] = DefaultValue; 3337 } 3338 3339 if (DefaultValue != SingleValue) 3340 SingleValue = NULL; 3341 } 3342 3343 // If each element in the table contains the same value, we only need to store 3344 // that single value. 3345 if (SingleValue) { 3346 Kind = SingleValueKind; 3347 return; 3348 } 3349 3350 // If the type is integer and the table fits in a register, build a bitmap. 3351 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3352 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3353 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3354 for (uint64_t I = TableSize; I > 0; --I) { 3355 TableInt <<= IT->getBitWidth(); 3356 // Insert values into the bitmap. Undef values are set to zero. 3357 if (!isa<UndefValue>(TableContents[I - 1])) { 3358 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3359 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3360 } 3361 } 3362 BitMap = ConstantInt::get(M.getContext(), TableInt); 3363 BitMapElementTy = IT; 3364 Kind = BitMapKind; 3365 ++NumBitMaps; 3366 return; 3367 } 3368 3369 // Store the table in an array. 3370 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3371 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3372 3373 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3374 GlobalVariable::PrivateLinkage, 3375 Initializer, 3376 "switch.table"); 3377 Array->setUnnamedAddr(true); 3378 Kind = ArrayKind; 3379} 3380 3381Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3382 switch (Kind) { 3383 case SingleValueKind: 3384 return SingleValue; 3385 case BitMapKind: { 3386 // Type of the bitmap (e.g. i59). 3387 IntegerType *MapTy = BitMap->getType(); 3388 3389 // Cast Index to the same type as the bitmap. 3390 // Note: The Index is <= the number of elements in the table, so 3391 // truncating it to the width of the bitmask is safe. 3392 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3393 3394 // Multiply the shift amount by the element width. 3395 ShiftAmt = Builder.CreateMul(ShiftAmt, 3396 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3397 "switch.shiftamt"); 3398 3399 // Shift down. 3400 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3401 "switch.downshift"); 3402 // Mask off. 3403 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3404 "switch.masked"); 3405 } 3406 case ArrayKind: { 3407 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3408 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3409 "switch.gep"); 3410 return Builder.CreateLoad(GEP, "switch.load"); 3411 } 3412 } 3413 llvm_unreachable("Unknown lookup table kind!"); 3414} 3415 3416bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3417 uint64_t TableSize, 3418 const Type *ElementType) { 3419 if (!TD) 3420 return false; 3421 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3422 if (!IT) 3423 return false; 3424 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3425 // are <= 15, we could try to narrow the type. 3426 3427 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3428 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3429 return false; 3430 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3431} 3432 3433/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3434/// for this switch, based on the number of caes, size of the table and the 3435/// types of the results. 3436static bool ShouldBuildLookupTable(SwitchInst *SI, 3437 uint64_t TableSize, 3438 const DataLayout *TD, 3439 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3440 // The table density should be at least 40%. This is the same criterion as for 3441 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3442 // FIXME: Find the best cut-off. 3443 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3444 return false; // TableSize overflowed, or mul below might overflow. 3445 if (SI->getNumCases() * 10 >= TableSize * 4) 3446 return true; 3447 3448 // If each table would fit in a register, we should build it anyway. 3449 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3450 E = ResultTypes.end(); I != E; ++I) { 3451 if (!SwitchLookupTable::WouldFitInRegister(TD, TableSize, I->second)) 3452 return false; 3453 } 3454 return true; 3455} 3456 3457/// SwitchToLookupTable - If the switch is only used to initialize one or more 3458/// phi nodes in a common successor block with different constant values, 3459/// replace the switch with lookup tables. 3460static bool SwitchToLookupTable(SwitchInst *SI, 3461 IRBuilder<> &Builder, 3462 const DataLayout* TD) { 3463 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3464 // FIXME: Handle unreachable cases. 3465 3466 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3467 // split off a dense part and build a lookup table for that. 3468 3469 // FIXME: This creates arrays of GEPs to constant strings, which means each 3470 // GEP needs a runtime relocation in PIC code. We should just build one big 3471 // string and lookup indices into that. 3472 3473 // Ignore the switch if the number of cases is too small. 3474 // This is similar to the check when building jump tables in 3475 // SelectionDAGBuilder::handleJTSwitchCase. 3476 // FIXME: Determine the best cut-off. 3477 if (SI->getNumCases() < 4) 3478 return false; 3479 3480 // Figure out the corresponding result for each case value and phi node in the 3481 // common destination, as well as the the min and max case values. 3482 assert(SI->case_begin() != SI->case_end()); 3483 SwitchInst::CaseIt CI = SI->case_begin(); 3484 ConstantInt *MinCaseVal = CI.getCaseValue(); 3485 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3486 3487 BasicBlock *CommonDest = NULL; 3488 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3489 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3490 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3491 SmallDenseMap<PHINode*, Type*> ResultTypes; 3492 SmallVector<PHINode*, 4> PHIs; 3493 3494 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3495 ConstantInt *CaseVal = CI.getCaseValue(); 3496 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3497 MinCaseVal = CaseVal; 3498 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3499 MaxCaseVal = CaseVal; 3500 3501 // Resulting value at phi nodes for this case value. 3502 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3503 ResultsTy Results; 3504 if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results)) 3505 return false; 3506 3507 // Append the result from this case to the list for each phi. 3508 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3509 if (!ResultLists.count(I->first)) 3510 PHIs.push_back(I->first); 3511 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3512 } 3513 } 3514 3515 // Get the resulting values for the default case. 3516 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3517 if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList)) 3518 return false; 3519 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3520 PHINode *PHI = DefaultResultsList[I].first; 3521 Constant *Result = DefaultResultsList[I].second; 3522 DefaultResults[PHI] = Result; 3523 ResultTypes[PHI] = Result->getType(); 3524 } 3525 3526 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3527 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3528 if (!ShouldBuildLookupTable(SI, TableSize, TD, ResultTypes)) 3529 return false; 3530 3531 // Create the BB that does the lookups. 3532 Module &Mod = *CommonDest->getParent()->getParent(); 3533 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3534 "switch.lookup", 3535 CommonDest->getParent(), 3536 CommonDest); 3537 3538 // Check whether the condition value is within the case range, and branch to 3539 // the new BB. 3540 Builder.SetInsertPoint(SI); 3541 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3542 "switch.tableidx"); 3543 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3544 MinCaseVal->getType(), TableSize)); 3545 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3546 3547 // Populate the BB that does the lookups. 3548 Builder.SetInsertPoint(LookupBB); 3549 bool ReturnedEarly = false; 3550 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3551 PHINode *PHI = PHIs[I]; 3552 3553 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3554 DefaultResults[PHI], TD); 3555 3556 Value *Result = Table.BuildLookup(TableIndex, Builder); 3557 3558 // If the result is used to return immediately from the function, we want to 3559 // do that right here. 3560 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3561 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3562 Builder.CreateRet(Result); 3563 ReturnedEarly = true; 3564 break; 3565 } 3566 3567 PHI->addIncoming(Result, LookupBB); 3568 } 3569 3570 if (!ReturnedEarly) 3571 Builder.CreateBr(CommonDest); 3572 3573 // Remove the switch. 3574 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3575 BasicBlock *Succ = SI->getSuccessor(i); 3576 if (Succ == SI->getDefaultDest()) continue; 3577 Succ->removePredecessor(SI->getParent()); 3578 } 3579 SI->eraseFromParent(); 3580 3581 ++NumLookupTables; 3582 return true; 3583} 3584 3585bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3586 // If this switch is too complex to want to look at, ignore it. 3587 if (!isValueEqualityComparison(SI)) 3588 return false; 3589 3590 BasicBlock *BB = SI->getParent(); 3591 3592 // If we only have one predecessor, and if it is a branch on this value, 3593 // see if that predecessor totally determines the outcome of this switch. 3594 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3595 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3596 return SimplifyCFG(BB) | true; 3597 3598 Value *Cond = SI->getCondition(); 3599 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3600 if (SimplifySwitchOnSelect(SI, Select)) 3601 return SimplifyCFG(BB) | true; 3602 3603 // If the block only contains the switch, see if we can fold the block 3604 // away into any preds. 3605 BasicBlock::iterator BBI = BB->begin(); 3606 // Ignore dbg intrinsics. 3607 while (isa<DbgInfoIntrinsic>(BBI)) 3608 ++BBI; 3609 if (SI == &*BBI) 3610 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3611 return SimplifyCFG(BB) | true; 3612 3613 // Try to transform the switch into an icmp and a branch. 3614 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3615 return SimplifyCFG(BB) | true; 3616 3617 // Remove unreachable cases. 3618 if (EliminateDeadSwitchCases(SI)) 3619 return SimplifyCFG(BB) | true; 3620 3621 if (ForwardSwitchConditionToPHI(SI)) 3622 return SimplifyCFG(BB) | true; 3623 3624 if (SwitchToLookupTable(SI, Builder, TD)) 3625 return SimplifyCFG(BB) | true; 3626 3627 return false; 3628} 3629 3630bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3631 BasicBlock *BB = IBI->getParent(); 3632 bool Changed = false; 3633 3634 // Eliminate redundant destinations. 3635 SmallPtrSet<Value *, 8> Succs; 3636 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3637 BasicBlock *Dest = IBI->getDestination(i); 3638 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3639 Dest->removePredecessor(BB); 3640 IBI->removeDestination(i); 3641 --i; --e; 3642 Changed = true; 3643 } 3644 } 3645 3646 if (IBI->getNumDestinations() == 0) { 3647 // If the indirectbr has no successors, change it to unreachable. 3648 new UnreachableInst(IBI->getContext(), IBI); 3649 EraseTerminatorInstAndDCECond(IBI); 3650 return true; 3651 } 3652 3653 if (IBI->getNumDestinations() == 1) { 3654 // If the indirectbr has one successor, change it to a direct branch. 3655 BranchInst::Create(IBI->getDestination(0), IBI); 3656 EraseTerminatorInstAndDCECond(IBI); 3657 return true; 3658 } 3659 3660 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3661 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3662 return SimplifyCFG(BB) | true; 3663 } 3664 return Changed; 3665} 3666 3667bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3668 BasicBlock *BB = BI->getParent(); 3669 3670 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3671 return true; 3672 3673 // If the Terminator is the only non-phi instruction, simplify the block. 3674 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3675 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3676 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3677 return true; 3678 3679 // If the only instruction in the block is a seteq/setne comparison 3680 // against a constant, try to simplify the block. 3681 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3682 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3683 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3684 ; 3685 if (I->isTerminator() && 3686 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) 3687 return true; 3688 } 3689 3690 // If this basic block is ONLY a compare and a branch, and if a predecessor 3691 // branches to us and our successor, fold the comparison into the 3692 // predecessor and use logical operations to update the incoming value 3693 // for PHI nodes in common successor. 3694 if (FoldBranchToCommonDest(BI)) 3695 return SimplifyCFG(BB) | true; 3696 return false; 3697} 3698 3699 3700bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3701 BasicBlock *BB = BI->getParent(); 3702 3703 // Conditional branch 3704 if (isValueEqualityComparison(BI)) { 3705 // If we only have one predecessor, and if it is a branch on this value, 3706 // see if that predecessor totally determines the outcome of this 3707 // switch. 3708 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3709 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3710 return SimplifyCFG(BB) | true; 3711 3712 // This block must be empty, except for the setcond inst, if it exists. 3713 // Ignore dbg intrinsics. 3714 BasicBlock::iterator I = BB->begin(); 3715 // Ignore dbg intrinsics. 3716 while (isa<DbgInfoIntrinsic>(I)) 3717 ++I; 3718 if (&*I == BI) { 3719 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3720 return SimplifyCFG(BB) | true; 3721 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3722 ++I; 3723 // Ignore dbg intrinsics. 3724 while (isa<DbgInfoIntrinsic>(I)) 3725 ++I; 3726 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3727 return SimplifyCFG(BB) | true; 3728 } 3729 } 3730 3731 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3732 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3733 return true; 3734 3735 // If this basic block is ONLY a compare and a branch, and if a predecessor 3736 // branches to us and one of our successors, fold the comparison into the 3737 // predecessor and use logical operations to pick the right destination. 3738 if (FoldBranchToCommonDest(BI)) 3739 return SimplifyCFG(BB) | true; 3740 3741 // We have a conditional branch to two blocks that are only reachable 3742 // from BI. We know that the condbr dominates the two blocks, so see if 3743 // there is any identical code in the "then" and "else" blocks. If so, we 3744 // can hoist it up to the branching block. 3745 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3746 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3747 if (HoistThenElseCodeToIf(BI)) 3748 return SimplifyCFG(BB) | true; 3749 } else { 3750 // If Successor #1 has multiple preds, we may be able to conditionally 3751 // execute Successor #0 if it branches to successor #1. 3752 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3753 if (Succ0TI->getNumSuccessors() == 1 && 3754 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3755 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3756 return SimplifyCFG(BB) | true; 3757 } 3758 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3759 // If Successor #0 has multiple preds, we may be able to conditionally 3760 // execute Successor #1 if it branches to successor #0. 3761 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3762 if (Succ1TI->getNumSuccessors() == 1 && 3763 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3764 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3765 return SimplifyCFG(BB) | true; 3766 } 3767 3768 // If this is a branch on a phi node in the current block, thread control 3769 // through this block if any PHI node entries are constants. 3770 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3771 if (PN->getParent() == BI->getParent()) 3772 if (FoldCondBranchOnPHI(BI, TD)) 3773 return SimplifyCFG(BB) | true; 3774 3775 // Scan predecessor blocks for conditional branches. 3776 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3777 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3778 if (PBI != BI && PBI->isConditional()) 3779 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3780 return SimplifyCFG(BB) | true; 3781 3782 return false; 3783} 3784 3785/// Check if passing a value to an instruction will cause undefined behavior. 3786static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3787 Constant *C = dyn_cast<Constant>(V); 3788 if (!C) 3789 return false; 3790 3791 if (I->use_empty()) 3792 return false; 3793 3794 if (C->isNullValue()) { 3795 // Only look at the first use, avoid hurting compile time with long uselists 3796 User *Use = *I->use_begin(); 3797 3798 // Now make sure that there are no instructions in between that can alter 3799 // control flow (eg. calls) 3800 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3801 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3802 return false; 3803 3804 // Look through GEPs. A load from a GEP derived from NULL is still undefined 3805 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 3806 if (GEP->getPointerOperand() == I) 3807 return passingValueIsAlwaysUndefined(V, GEP); 3808 3809 // Look through bitcasts. 3810 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 3811 return passingValueIsAlwaysUndefined(V, BC); 3812 3813 // Load from null is undefined. 3814 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 3815 return LI->getPointerAddressSpace() == 0; 3816 3817 // Store to null is undefined. 3818 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 3819 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 3820 } 3821 return false; 3822} 3823 3824/// If BB has an incoming value that will always trigger undefined behavior 3825/// (eg. null pointer dereference), remove the branch leading here. 3826static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 3827 for (BasicBlock::iterator i = BB->begin(); 3828 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 3829 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 3830 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 3831 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 3832 IRBuilder<> Builder(T); 3833 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 3834 BB->removePredecessor(PHI->getIncomingBlock(i)); 3835 // Turn uncoditional branches into unreachables and remove the dead 3836 // destination from conditional branches. 3837 if (BI->isUnconditional()) 3838 Builder.CreateUnreachable(); 3839 else 3840 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 3841 BI->getSuccessor(0)); 3842 BI->eraseFromParent(); 3843 return true; 3844 } 3845 // TODO: SwitchInst. 3846 } 3847 3848 return false; 3849} 3850 3851bool SimplifyCFGOpt::run(BasicBlock *BB) { 3852 bool Changed = false; 3853 3854 assert(BB && BB->getParent() && "Block not embedded in function!"); 3855 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 3856 3857 // Remove basic blocks that have no predecessors (except the entry block)... 3858 // or that just have themself as a predecessor. These are unreachable. 3859 if ((pred_begin(BB) == pred_end(BB) && 3860 BB != &BB->getParent()->getEntryBlock()) || 3861 BB->getSinglePredecessor() == BB) { 3862 DEBUG(dbgs() << "Removing BB: \n" << *BB); 3863 DeleteDeadBlock(BB); 3864 return true; 3865 } 3866 3867 // Check to see if we can constant propagate this terminator instruction 3868 // away... 3869 Changed |= ConstantFoldTerminator(BB, true); 3870 3871 // Check for and eliminate duplicate PHI nodes in this block. 3872 Changed |= EliminateDuplicatePHINodes(BB); 3873 3874 // Check for and remove branches that will always cause undefined behavior. 3875 Changed |= removeUndefIntroducingPredecessor(BB); 3876 3877 // Merge basic blocks into their predecessor if there is only one distinct 3878 // pred, and if there is only one distinct successor of the predecessor, and 3879 // if there are no PHI nodes. 3880 // 3881 if (MergeBlockIntoPredecessor(BB)) 3882 return true; 3883 3884 IRBuilder<> Builder(BB); 3885 3886 // If there is a trivial two-entry PHI node in this basic block, and we can 3887 // eliminate it, do so now. 3888 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 3889 if (PN->getNumIncomingValues() == 2) 3890 Changed |= FoldTwoEntryPHINode(PN, TD); 3891 3892 Builder.SetInsertPoint(BB->getTerminator()); 3893 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 3894 if (BI->isUnconditional()) { 3895 if (SimplifyUncondBranch(BI, Builder)) return true; 3896 } else { 3897 if (SimplifyCondBranch(BI, Builder)) return true; 3898 } 3899 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 3900 if (SimplifyReturn(RI, Builder)) return true; 3901 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 3902 if (SimplifyResume(RI, Builder)) return true; 3903 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 3904 if (SimplifySwitch(SI, Builder)) return true; 3905 } else if (UnreachableInst *UI = 3906 dyn_cast<UnreachableInst>(BB->getTerminator())) { 3907 if (SimplifyUnreachable(UI)) return true; 3908 } else if (IndirectBrInst *IBI = 3909 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 3910 if (SimplifyIndirectBr(IBI)) return true; 3911 } 3912 3913 return Changed; 3914} 3915 3916/// SimplifyCFG - This function is used to do simplification of a CFG. For 3917/// example, it adjusts branches to branches to eliminate the extra hop, it 3918/// eliminates unreachable basic blocks, and does other "peephole" optimization 3919/// of the CFG. It returns true if a modification was made. 3920/// 3921bool llvm::SimplifyCFG(BasicBlock *BB, const DataLayout *TD) { 3922 return SimplifyCFGOpt(TD).run(BB); 3923} 3924