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