SimplifyCFG.cpp revision 24473120a253a05f3601cd3373403b47e6d03d41
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/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/LLVMContext.h" 22#include "llvm/Metadata.h" 23#include "llvm/Operator.h" 24#include "llvm/Type.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Analysis/ValueTracking.h" 27#include "llvm/Target/TargetData.h" 28#include "llvm/Transforms/Utils/BasicBlockUtils.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/SetVector.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/ADT/STLExtras.h" 35#include "llvm/Support/CFG.h" 36#include "llvm/Support/CommandLine.h" 37#include "llvm/Support/ConstantRange.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/NoFolder.h" 41#include "llvm/Support/raw_ostream.h" 42#include <algorithm> 43#include <set> 44#include <map> 45using namespace llvm; 46 47static cl::opt<unsigned> 48PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 49 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 50 51static cl::opt<bool> 52DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 53 cl::desc("Duplicate return instructions into unconditional branches")); 54 55STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 56 57namespace { 58class SimplifyCFGOpt { 59 const TargetData *const TD; 60 61 Value *isValueEqualityComparison(TerminatorInst *TI); 62 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 63 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases); 64 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 65 BasicBlock *Pred, 66 IRBuilder<> &Builder); 67 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 68 IRBuilder<> &Builder); 69 70 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 71 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 72 bool SimplifyUnwind(UnwindInst *UI, IRBuilder<> &Builder); 73 bool SimplifyUnreachable(UnreachableInst *UI); 74 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 75 bool SimplifyIndirectBr(IndirectBrInst *IBI); 76 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 77 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 78 79public: 80 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} 81 bool run(BasicBlock *BB); 82}; 83} 84 85/// SafeToMergeTerminators - Return true if it is safe to merge these two 86/// terminator instructions together. 87/// 88static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 89 if (SI1 == SI2) return false; // Can't merge with self! 90 91 // It is not safe to merge these two switch instructions if they have a common 92 // successor, and if that successor has a PHI node, and if *that* PHI node has 93 // conflicting incoming values from the two switch blocks. 94 BasicBlock *SI1BB = SI1->getParent(); 95 BasicBlock *SI2BB = SI2->getParent(); 96 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 97 98 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 99 if (SI1Succs.count(*I)) 100 for (BasicBlock::iterator BBI = (*I)->begin(); 101 isa<PHINode>(BBI); ++BBI) { 102 PHINode *PN = cast<PHINode>(BBI); 103 if (PN->getIncomingValueForBlock(SI1BB) != 104 PN->getIncomingValueForBlock(SI2BB)) 105 return false; 106 } 107 108 return true; 109} 110 111/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 112/// now be entries in it from the 'NewPred' block. The values that will be 113/// flowing into the PHI nodes will be the same as those coming in from 114/// ExistPred, an existing predecessor of Succ. 115static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 116 BasicBlock *ExistPred) { 117 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 118 119 PHINode *PN; 120 for (BasicBlock::iterator I = Succ->begin(); 121 (PN = dyn_cast<PHINode>(I)); ++I) 122 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 123} 124 125 126/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 127/// least one PHI node in it), check to see if the merge at this block is due 128/// to an "if condition". If so, return the boolean condition that determines 129/// which entry into BB will be taken. Also, return by references the block 130/// that will be entered from if the condition is true, and the block that will 131/// be entered if the condition is false. 132/// 133/// This does no checking to see if the true/false blocks have large or unsavory 134/// instructions in them. 135static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 136 BasicBlock *&IfFalse) { 137 PHINode *SomePHI = cast<PHINode>(BB->begin()); 138 assert(SomePHI->getNumIncomingValues() == 2 && 139 "Function can only handle blocks with 2 predecessors!"); 140 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 141 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 142 143 // We can only handle branches. Other control flow will be lowered to 144 // branches if possible anyway. 145 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 146 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 147 if (Pred1Br == 0 || Pred2Br == 0) 148 return 0; 149 150 // Eliminate code duplication by ensuring that Pred1Br is conditional if 151 // either are. 152 if (Pred2Br->isConditional()) { 153 // If both branches are conditional, we don't have an "if statement". In 154 // reality, we could transform this case, but since the condition will be 155 // required anyway, we stand no chance of eliminating it, so the xform is 156 // probably not profitable. 157 if (Pred1Br->isConditional()) 158 return 0; 159 160 std::swap(Pred1, Pred2); 161 std::swap(Pred1Br, Pred2Br); 162 } 163 164 if (Pred1Br->isConditional()) { 165 // The only thing we have to watch out for here is to make sure that Pred2 166 // doesn't have incoming edges from other blocks. If it does, the condition 167 // doesn't dominate BB. 168 if (Pred2->getSinglePredecessor() == 0) 169 return 0; 170 171 // If we found a conditional branch predecessor, make sure that it branches 172 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 173 if (Pred1Br->getSuccessor(0) == BB && 174 Pred1Br->getSuccessor(1) == Pred2) { 175 IfTrue = Pred1; 176 IfFalse = Pred2; 177 } else if (Pred1Br->getSuccessor(0) == Pred2 && 178 Pred1Br->getSuccessor(1) == BB) { 179 IfTrue = Pred2; 180 IfFalse = Pred1; 181 } else { 182 // We know that one arm of the conditional goes to BB, so the other must 183 // go somewhere unrelated, and this must not be an "if statement". 184 return 0; 185 } 186 187 return Pred1Br->getCondition(); 188 } 189 190 // Ok, if we got here, both predecessors end with an unconditional branch to 191 // BB. Don't panic! If both blocks only have a single (identical) 192 // predecessor, and THAT is a conditional branch, then we're all ok! 193 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 194 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 195 return 0; 196 197 // Otherwise, if this is a conditional branch, then we can use it! 198 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 199 if (BI == 0) return 0; 200 201 assert(BI->isConditional() && "Two successors but not conditional?"); 202 if (BI->getSuccessor(0) == Pred1) { 203 IfTrue = Pred1; 204 IfFalse = Pred2; 205 } else { 206 IfTrue = Pred2; 207 IfFalse = Pred1; 208 } 209 return BI->getCondition(); 210} 211 212/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 213/// given instruction, which is assumed to be safe to speculate. 1 means 214/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 215static unsigned ComputeSpeculationCost(const User *I) { 216 assert(isSafeToSpeculativelyExecute(I) && 217 "Instruction is not safe to speculatively execute!"); 218 switch (Operator::getOpcode(I)) { 219 default: 220 // In doubt, be conservative. 221 return UINT_MAX; 222 case Instruction::GetElementPtr: 223 // GEPs are cheap if all indices are constant. 224 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 225 return UINT_MAX; 226 return 1; 227 case Instruction::Load: 228 case Instruction::Add: 229 case Instruction::Sub: 230 case Instruction::And: 231 case Instruction::Or: 232 case Instruction::Xor: 233 case Instruction::Shl: 234 case Instruction::LShr: 235 case Instruction::AShr: 236 case Instruction::ICmp: 237 case Instruction::Trunc: 238 case Instruction::ZExt: 239 case Instruction::SExt: 240 return 1; // These are all cheap. 241 242 case Instruction::Call: 243 case Instruction::Select: 244 return 2; 245 } 246} 247 248/// DominatesMergePoint - If we have a merge point of an "if condition" as 249/// accepted above, return true if the specified value dominates the block. We 250/// don't handle the true generality of domination here, just a special case 251/// which works well enough for us. 252/// 253/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 254/// see if V (which must be an instruction) and its recursive operands 255/// that do not dominate BB have a combined cost lower than CostRemaining and 256/// are non-trapping. If both are true, the instruction is inserted into the 257/// set and true is returned. 258/// 259/// The cost for most non-trapping instructions is defined as 1 except for 260/// Select whose cost is 2. 261/// 262/// After this function returns, CostRemaining is decreased by the cost of 263/// V plus its non-dominating operands. If that cost is greater than 264/// CostRemaining, false is returned and CostRemaining is undefined. 265static bool DominatesMergePoint(Value *V, BasicBlock *BB, 266 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 267 unsigned &CostRemaining) { 268 Instruction *I = dyn_cast<Instruction>(V); 269 if (!I) { 270 // Non-instructions all dominate instructions, but not all constantexprs 271 // can be executed unconditionally. 272 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 273 if (C->canTrap()) 274 return false; 275 return true; 276 } 277 BasicBlock *PBB = I->getParent(); 278 279 // We don't want to allow weird loops that might have the "if condition" in 280 // the bottom of this block. 281 if (PBB == BB) return false; 282 283 // If this instruction is defined in a block that contains an unconditional 284 // branch to BB, then it must be in the 'conditional' part of the "if 285 // statement". If not, it definitely dominates the region. 286 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 287 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 288 return true; 289 290 // If we aren't allowing aggressive promotion anymore, then don't consider 291 // instructions in the 'if region'. 292 if (AggressiveInsts == 0) return false; 293 294 // If we have seen this instruction before, don't count it again. 295 if (AggressiveInsts->count(I)) return true; 296 297 // Okay, it looks like the instruction IS in the "condition". Check to 298 // see if it's a cheap instruction to unconditionally compute, and if it 299 // only uses stuff defined outside of the condition. If so, hoist it out. 300 if (!isSafeToSpeculativelyExecute(I)) 301 return false; 302 303 unsigned Cost = ComputeSpeculationCost(I); 304 305 if (Cost > CostRemaining) 306 return false; 307 308 CostRemaining -= Cost; 309 310 // Okay, we can only really hoist these out if their operands do 311 // not take us over the cost threshold. 312 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 313 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 314 return false; 315 // Okay, it's safe to do this! Remember this instruction. 316 AggressiveInsts->insert(I); 317 return true; 318} 319 320/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 321/// and PointerNullValue. Return NULL if value is not a constant int. 322static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { 323 // Normal constant int. 324 ConstantInt *CI = dyn_cast<ConstantInt>(V); 325 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 326 return CI; 327 328 // This is some kind of pointer constant. Turn it into a pointer-sized 329 // ConstantInt if possible. 330 IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); 331 332 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 333 if (isa<ConstantPointerNull>(V)) 334 return ConstantInt::get(PtrTy, 0); 335 336 // IntToPtr const int. 337 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 338 if (CE->getOpcode() == Instruction::IntToPtr) 339 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 340 // The constant is very likely to have the right type already. 341 if (CI->getType() == PtrTy) 342 return CI; 343 else 344 return cast<ConstantInt> 345 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 346 } 347 return 0; 348} 349 350/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 351/// collection of icmp eq/ne instructions that compare a value against a 352/// constant, return the value being compared, and stick the constant into the 353/// Values vector. 354static Value * 355GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 356 const TargetData *TD, bool isEQ, unsigned &UsedICmps) { 357 Instruction *I = dyn_cast<Instruction>(V); 358 if (I == 0) return 0; 359 360 // If this is an icmp against a constant, handle this as one of the cases. 361 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 362 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 363 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 364 UsedICmps++; 365 Vals.push_back(C); 366 return I->getOperand(0); 367 } 368 369 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 370 // the set. 371 ConstantRange Span = 372 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 373 374 // If this is an and/!= check then we want to optimize "x ugt 2" into 375 // x != 0 && x != 1. 376 if (!isEQ) 377 Span = Span.inverse(); 378 379 // If there are a ton of values, we don't want to make a ginormous switch. 380 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 381 return 0; 382 383 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 384 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 385 UsedICmps++; 386 return I->getOperand(0); 387 } 388 return 0; 389 } 390 391 // Otherwise, we can only handle an | or &, depending on isEQ. 392 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 393 return 0; 394 395 unsigned NumValsBeforeLHS = Vals.size(); 396 unsigned UsedICmpsBeforeLHS = UsedICmps; 397 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 398 isEQ, UsedICmps)) { 399 unsigned NumVals = Vals.size(); 400 unsigned UsedICmpsBeforeRHS = UsedICmps; 401 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 402 isEQ, UsedICmps)) { 403 if (LHS == RHS) 404 return LHS; 405 Vals.resize(NumVals); 406 UsedICmps = UsedICmpsBeforeRHS; 407 } 408 409 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 410 // set it and return success. 411 if (Extra == 0 || Extra == I->getOperand(1)) { 412 Extra = I->getOperand(1); 413 return LHS; 414 } 415 416 Vals.resize(NumValsBeforeLHS); 417 UsedICmps = UsedICmpsBeforeLHS; 418 return 0; 419 } 420 421 // If the LHS can't be folded in, but Extra is available and RHS can, try to 422 // use LHS as Extra. 423 if (Extra == 0 || Extra == I->getOperand(0)) { 424 Value *OldExtra = Extra; 425 Extra = I->getOperand(0); 426 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 427 isEQ, UsedICmps)) 428 return RHS; 429 assert(Vals.size() == NumValsBeforeLHS); 430 Extra = OldExtra; 431 } 432 433 return 0; 434} 435 436static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 437 Instruction *Cond = 0; 438 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 439 Cond = dyn_cast<Instruction>(SI->getCondition()); 440 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 441 if (BI->isConditional()) 442 Cond = dyn_cast<Instruction>(BI->getCondition()); 443 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 444 Cond = dyn_cast<Instruction>(IBI->getAddress()); 445 } 446 447 TI->eraseFromParent(); 448 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 449} 450 451/// isValueEqualityComparison - Return true if the specified terminator checks 452/// to see if a value is equal to constant integer value. 453Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 454 Value *CV = 0; 455 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 456 // Do not permit merging of large switch instructions into their 457 // predecessors unless there is only one predecessor. 458 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 459 pred_end(SI->getParent())) <= 128) 460 CV = SI->getCondition(); 461 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 462 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 463 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 464 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 465 ICI->getPredicate() == ICmpInst::ICMP_NE) && 466 GetConstantInt(ICI->getOperand(1), TD)) 467 CV = ICI->getOperand(0); 468 469 // Unwrap any lossless ptrtoint cast. 470 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 471 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 472 CV = PTII->getOperand(0); 473 return CV; 474} 475 476/// GetValueEqualityComparisonCases - Given a value comparison instruction, 477/// decode all of the 'cases' that it represents and return the 'default' block. 478BasicBlock *SimplifyCFGOpt:: 479GetValueEqualityComparisonCases(TerminatorInst *TI, 480 std::vector<std::pair<ConstantInt*, 481 BasicBlock*> > &Cases) { 482 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 483 Cases.reserve(SI->getNumCases()); 484 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) 485 Cases.push_back(std::make_pair(SI->getCaseValue(i), 486 SI->getCaseSuccessor(i))); 487 return SI->getDefaultDest(); 488 } 489 490 BranchInst *BI = cast<BranchInst>(TI); 491 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 492 Cases.push_back(std::make_pair(GetConstantInt(ICI->getOperand(1), TD), 493 BI->getSuccessor(ICI->getPredicate() == 494 ICmpInst::ICMP_NE))); 495 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 496} 497 498 499/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 500/// in the list that match the specified block. 501static void EliminateBlockCases(BasicBlock *BB, 502 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 503 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 504 if (Cases[i].second == BB) { 505 Cases.erase(Cases.begin()+i); 506 --i; --e; 507 } 508} 509 510/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 511/// well. 512static bool 513ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 514 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 515 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 516 517 // Make V1 be smaller than V2. 518 if (V1->size() > V2->size()) 519 std::swap(V1, V2); 520 521 if (V1->size() == 0) return false; 522 if (V1->size() == 1) { 523 // Just scan V2. 524 ConstantInt *TheVal = (*V1)[0].first; 525 for (unsigned i = 0, e = V2->size(); i != e; ++i) 526 if (TheVal == (*V2)[i].first) 527 return true; 528 } 529 530 // Otherwise, just sort both lists and compare element by element. 531 array_pod_sort(V1->begin(), V1->end()); 532 array_pod_sort(V2->begin(), V2->end()); 533 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 534 while (i1 != e1 && i2 != e2) { 535 if ((*V1)[i1].first == (*V2)[i2].first) 536 return true; 537 if ((*V1)[i1].first < (*V2)[i2].first) 538 ++i1; 539 else 540 ++i2; 541 } 542 return false; 543} 544 545/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 546/// terminator instruction and its block is known to only have a single 547/// predecessor block, check to see if that predecessor is also a value 548/// comparison with the same value, and if that comparison determines the 549/// outcome of this comparison. If so, simplify TI. This does a very limited 550/// form of jump threading. 551bool SimplifyCFGOpt:: 552SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 553 BasicBlock *Pred, 554 IRBuilder<> &Builder) { 555 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 556 if (!PredVal) return false; // Not a value comparison in predecessor. 557 558 Value *ThisVal = isValueEqualityComparison(TI); 559 assert(ThisVal && "This isn't a value comparison!!"); 560 if (ThisVal != PredVal) return false; // Different predicates. 561 562 // Find out information about when control will move from Pred to TI's block. 563 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 564 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 565 PredCases); 566 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 567 568 // Find information about how control leaves this block. 569 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 570 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 571 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 572 573 // If TI's block is the default block from Pred's comparison, potentially 574 // simplify TI based on this knowledge. 575 if (PredDef == TI->getParent()) { 576 // If we are here, we know that the value is none of those cases listed in 577 // PredCases. If there are any cases in ThisCases that are in PredCases, we 578 // can simplify TI. 579 if (!ValuesOverlap(PredCases, ThisCases)) 580 return false; 581 582 if (isa<BranchInst>(TI)) { 583 // Okay, one of the successors of this condbr is dead. Convert it to a 584 // uncond br. 585 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 586 // Insert the new branch. 587 Instruction *NI = Builder.CreateBr(ThisDef); 588 (void) NI; 589 590 // Remove PHI node entries for the dead edge. 591 ThisCases[0].second->removePredecessor(TI->getParent()); 592 593 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 594 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 595 596 EraseTerminatorInstAndDCECond(TI); 597 return true; 598 } 599 600 SwitchInst *SI = cast<SwitchInst>(TI); 601 // Okay, TI has cases that are statically dead, prune them away. 602 SmallPtrSet<Constant*, 16> DeadCases; 603 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 604 DeadCases.insert(PredCases[i].first); 605 606 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 607 << "Through successor TI: " << *TI); 608 609 for (unsigned i = SI->getNumCases(); i != 0;) { 610 --i; 611 if (DeadCases.count(SI->getCaseValue(i))) { 612 SI->getCaseSuccessor(i)->removePredecessor(TI->getParent()); 613 SI->removeCase(i); 614 } 615 } 616 617 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 618 return true; 619 } 620 621 // Otherwise, TI's block must correspond to some matched value. Find out 622 // which value (or set of values) this is. 623 ConstantInt *TIV = 0; 624 BasicBlock *TIBB = TI->getParent(); 625 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 626 if (PredCases[i].second == TIBB) { 627 if (TIV != 0) 628 return false; // Cannot handle multiple values coming to this block. 629 TIV = PredCases[i].first; 630 } 631 assert(TIV && "No edge from pred to succ?"); 632 633 // Okay, we found the one constant that our value can be if we get into TI's 634 // BB. Find out which successor will unconditionally be branched to. 635 BasicBlock *TheRealDest = 0; 636 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 637 if (ThisCases[i].first == TIV) { 638 TheRealDest = ThisCases[i].second; 639 break; 640 } 641 642 // If not handled by any explicit cases, it is handled by the default case. 643 if (TheRealDest == 0) TheRealDest = ThisDef; 644 645 // Remove PHI node entries for dead edges. 646 BasicBlock *CheckEdge = TheRealDest; 647 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 648 if (*SI != CheckEdge) 649 (*SI)->removePredecessor(TIBB); 650 else 651 CheckEdge = 0; 652 653 // Insert the new branch. 654 Instruction *NI = Builder.CreateBr(TheRealDest); 655 (void) NI; 656 657 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 658 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 659 660 EraseTerminatorInstAndDCECond(TI); 661 return true; 662} 663 664namespace { 665 /// ConstantIntOrdering - This class implements a stable ordering of constant 666 /// integers that does not depend on their address. This is important for 667 /// applications that sort ConstantInt's to ensure uniqueness. 668 struct ConstantIntOrdering { 669 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 670 return LHS->getValue().ult(RHS->getValue()); 671 } 672 }; 673} 674 675static int ConstantIntSortPredicate(const void *P1, const void *P2) { 676 const ConstantInt *LHS = *(const ConstantInt**)P1; 677 const ConstantInt *RHS = *(const ConstantInt**)P2; 678 if (LHS->getValue().ult(RHS->getValue())) 679 return 1; 680 if (LHS->getValue() == RHS->getValue()) 681 return 0; 682 return -1; 683} 684 685/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 686/// equality comparison instruction (either a switch or a branch on "X == c"). 687/// See if any of the predecessors of the terminator block are value comparisons 688/// on the same value. If so, and if safe to do so, fold them together. 689bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 690 IRBuilder<> &Builder) { 691 BasicBlock *BB = TI->getParent(); 692 Value *CV = isValueEqualityComparison(TI); // CondVal 693 assert(CV && "Not a comparison?"); 694 bool Changed = false; 695 696 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 697 while (!Preds.empty()) { 698 BasicBlock *Pred = Preds.pop_back_val(); 699 700 // See if the predecessor is a comparison with the same value. 701 TerminatorInst *PTI = Pred->getTerminator(); 702 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 703 704 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 705 // Figure out which 'cases' to copy from SI to PSI. 706 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 707 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 708 709 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 710 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 711 712 // Based on whether the default edge from PTI goes to BB or not, fill in 713 // PredCases and PredDefault with the new switch cases we would like to 714 // build. 715 SmallVector<BasicBlock*, 8> NewSuccessors; 716 717 if (PredDefault == BB) { 718 // If this is the default destination from PTI, only the edges in TI 719 // that don't occur in PTI, or that branch to BB will be activated. 720 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 721 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 722 if (PredCases[i].second != BB) 723 PTIHandled.insert(PredCases[i].first); 724 else { 725 // The default destination is BB, we don't need explicit targets. 726 std::swap(PredCases[i], PredCases.back()); 727 PredCases.pop_back(); 728 --i; --e; 729 } 730 731 // Reconstruct the new switch statement we will be building. 732 if (PredDefault != BBDefault) { 733 PredDefault->removePredecessor(Pred); 734 PredDefault = BBDefault; 735 NewSuccessors.push_back(BBDefault); 736 } 737 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 738 if (!PTIHandled.count(BBCases[i].first) && 739 BBCases[i].second != BBDefault) { 740 PredCases.push_back(BBCases[i]); 741 NewSuccessors.push_back(BBCases[i].second); 742 } 743 744 } else { 745 // If this is not the default destination from PSI, only the edges 746 // in SI that occur in PSI with a destination of BB will be 747 // activated. 748 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 749 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 750 if (PredCases[i].second == BB) { 751 PTIHandled.insert(PredCases[i].first); 752 std::swap(PredCases[i], PredCases.back()); 753 PredCases.pop_back(); 754 --i; --e; 755 } 756 757 // Okay, now we know which constants were sent to BB from the 758 // predecessor. Figure out where they will all go now. 759 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 760 if (PTIHandled.count(BBCases[i].first)) { 761 // If this is one we are capable of getting... 762 PredCases.push_back(BBCases[i]); 763 NewSuccessors.push_back(BBCases[i].second); 764 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 765 } 766 767 // If there are any constants vectored to BB that TI doesn't handle, 768 // they must go to the default destination of TI. 769 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 770 PTIHandled.begin(), 771 E = PTIHandled.end(); I != E; ++I) { 772 PredCases.push_back(std::make_pair(*I, BBDefault)); 773 NewSuccessors.push_back(BBDefault); 774 } 775 } 776 777 // Okay, at this point, we know which new successor Pred will get. Make 778 // sure we update the number of entries in the PHI nodes for these 779 // successors. 780 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 781 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 782 783 Builder.SetInsertPoint(PTI); 784 // Convert pointer to int before we switch. 785 if (CV->getType()->isPointerTy()) { 786 assert(TD && "Cannot switch on pointer without TargetData"); 787 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 788 "magicptr"); 789 } 790 791 // Now that the successors are updated, create the new Switch instruction. 792 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 793 PredCases.size()); 794 NewSI->setDebugLoc(PTI->getDebugLoc()); 795 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 796 NewSI->addCase(PredCases[i].first, PredCases[i].second); 797 798 EraseTerminatorInstAndDCECond(PTI); 799 800 // Okay, last check. If BB is still a successor of PSI, then we must 801 // have an infinite loop case. If so, add an infinitely looping block 802 // to handle the case to preserve the behavior of the code. 803 BasicBlock *InfLoopBlock = 0; 804 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 805 if (NewSI->getSuccessor(i) == BB) { 806 if (InfLoopBlock == 0) { 807 // Insert it at the end of the function, because it's either code, 808 // or it won't matter if it's hot. :) 809 InfLoopBlock = BasicBlock::Create(BB->getContext(), 810 "infloop", BB->getParent()); 811 BranchInst::Create(InfLoopBlock, InfLoopBlock); 812 } 813 NewSI->setSuccessor(i, InfLoopBlock); 814 } 815 816 Changed = true; 817 } 818 } 819 return Changed; 820} 821 822// isSafeToHoistInvoke - If we would need to insert a select that uses the 823// value of this invoke (comments in HoistThenElseCodeToIf explain why we 824// would need to do this), we can't hoist the invoke, as there is nowhere 825// to put the select in this case. 826static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 827 Instruction *I1, Instruction *I2) { 828 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 829 PHINode *PN; 830 for (BasicBlock::iterator BBI = SI->begin(); 831 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 832 Value *BB1V = PN->getIncomingValueForBlock(BB1); 833 Value *BB2V = PN->getIncomingValueForBlock(BB2); 834 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 835 return false; 836 } 837 } 838 } 839 return true; 840} 841 842/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 843/// BB2, hoist any common code in the two blocks up into the branch block. The 844/// caller of this function guarantees that BI's block dominates BB1 and BB2. 845static bool HoistThenElseCodeToIf(BranchInst *BI) { 846 // This does very trivial matching, with limited scanning, to find identical 847 // instructions in the two blocks. In particular, we don't want to get into 848 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 849 // such, we currently just scan for obviously identical instructions in an 850 // identical order. 851 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 852 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 853 854 BasicBlock::iterator BB1_Itr = BB1->begin(); 855 BasicBlock::iterator BB2_Itr = BB2->begin(); 856 857 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 858 // Skip debug info if it is not identical. 859 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 860 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 861 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 862 while (isa<DbgInfoIntrinsic>(I1)) 863 I1 = BB1_Itr++; 864 while (isa<DbgInfoIntrinsic>(I2)) 865 I2 = BB2_Itr++; 866 } 867 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 868 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 869 return false; 870 871 // If we get here, we can hoist at least one instruction. 872 BasicBlock *BIParent = BI->getParent(); 873 874 do { 875 // If we are hoisting the terminator instruction, don't move one (making a 876 // broken BB), instead clone it, and remove BI. 877 if (isa<TerminatorInst>(I1)) 878 goto HoistTerminator; 879 880 // For a normal instruction, we just move one to right before the branch, 881 // then replace all uses of the other with the first. Finally, we remove 882 // the now redundant second instruction. 883 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 884 if (!I2->use_empty()) 885 I2->replaceAllUsesWith(I1); 886 I1->intersectOptionalDataWith(I2); 887 I2->eraseFromParent(); 888 889 I1 = BB1_Itr++; 890 I2 = BB2_Itr++; 891 // Skip debug info if it is not identical. 892 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 893 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 894 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 895 while (isa<DbgInfoIntrinsic>(I1)) 896 I1 = BB1_Itr++; 897 while (isa<DbgInfoIntrinsic>(I2)) 898 I2 = BB2_Itr++; 899 } 900 } while (I1->isIdenticalToWhenDefined(I2)); 901 902 return true; 903 904HoistTerminator: 905 // It may not be possible to hoist an invoke. 906 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 907 return true; 908 909 // Okay, it is safe to hoist the terminator. 910 Instruction *NT = I1->clone(); 911 BIParent->getInstList().insert(BI, NT); 912 if (!NT->getType()->isVoidTy()) { 913 I1->replaceAllUsesWith(NT); 914 I2->replaceAllUsesWith(NT); 915 NT->takeName(I1); 916 } 917 918 IRBuilder<true, NoFolder> Builder(NT); 919 // Hoisting one of the terminators from our successor is a great thing. 920 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 921 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 922 // nodes, so we insert select instruction to compute the final result. 923 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 924 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 925 PHINode *PN; 926 for (BasicBlock::iterator BBI = SI->begin(); 927 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 928 Value *BB1V = PN->getIncomingValueForBlock(BB1); 929 Value *BB2V = PN->getIncomingValueForBlock(BB2); 930 if (BB1V == BB2V) continue; 931 932 // These values do not agree. Insert a select instruction before NT 933 // that determines the right value. 934 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 935 if (SI == 0) 936 SI = cast<SelectInst> 937 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 938 BB1V->getName()+"."+BB2V->getName())); 939 940 // Make the PHI node use the select for all incoming values for BB1/BB2 941 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 942 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 943 PN->setIncomingValue(i, SI); 944 } 945 } 946 947 // Update any PHI nodes in our new successors. 948 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 949 AddPredecessorToBlock(*SI, BIParent, BB1); 950 951 EraseTerminatorInstAndDCECond(BI); 952 return true; 953} 954 955/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 956/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 957/// (for now, restricted to a single instruction that's side effect free) from 958/// the BB1 into the branch block to speculatively execute it. 959/// 960/// Turn 961/// BB: 962/// %t1 = icmp 963/// br i1 %t1, label %BB1, label %BB2 964/// BB1: 965/// %t3 = add %t2, c 966/// br label BB2 967/// BB2: 968/// => 969/// BB: 970/// %t1 = icmp 971/// %t4 = add %t2, c 972/// %t3 = select i1 %t1, %t2, %t3 973static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 974 // Only speculatively execution a single instruction (not counting the 975 // terminator) for now. 976 Instruction *HInst = NULL; 977 Instruction *Term = BB1->getTerminator(); 978 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 979 BBI != BBE; ++BBI) { 980 Instruction *I = BBI; 981 // Skip debug info. 982 if (isa<DbgInfoIntrinsic>(I)) continue; 983 if (I == Term) break; 984 985 if (HInst) 986 return false; 987 HInst = I; 988 } 989 990 BasicBlock *BIParent = BI->getParent(); 991 992 // Check the instruction to be hoisted, if there is one. 993 if (HInst) { 994 // Don't hoist the instruction if it's unsafe or expensive. 995 if (!isSafeToSpeculativelyExecute(HInst)) 996 return false; 997 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold) 998 return false; 999 1000 // Do not hoist the instruction if any of its operands are defined but not 1001 // used in this BB. The transformation will prevent the operand from 1002 // being sunk into the use block. 1003 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1004 i != e; ++i) { 1005 Instruction *OpI = dyn_cast<Instruction>(*i); 1006 if (OpI && OpI->getParent() == BIParent && 1007 !OpI->mayHaveSideEffects() && 1008 !OpI->isUsedInBasicBlock(BIParent)) 1009 return false; 1010 } 1011 } 1012 1013 // Be conservative for now. FP select instruction can often be expensive. 1014 Value *BrCond = BI->getCondition(); 1015 if (isa<FCmpInst>(BrCond)) 1016 return false; 1017 1018 // If BB1 is actually on the false edge of the conditional branch, remember 1019 // to swap the select operands later. 1020 bool Invert = false; 1021 if (BB1 != BI->getSuccessor(0)) { 1022 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1023 Invert = true; 1024 } 1025 1026 // Collect interesting PHIs, and scan for hazards. 1027 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs; 1028 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1029 for (BasicBlock::iterator I = BB2->begin(); 1030 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1031 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1032 Value *BIParentV = PN->getIncomingValueForBlock(BIParent); 1033 1034 // Skip PHIs which are trivial. 1035 if (BB1V == BIParentV) 1036 continue; 1037 1038 // Check for saftey. 1039 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) { 1040 // An unfolded ConstantExpr could end up getting expanded into 1041 // Instructions. Don't speculate this and another instruction at 1042 // the same time. 1043 if (HInst) 1044 return false; 1045 if (!isSafeToSpeculativelyExecute(CE)) 1046 return false; 1047 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1048 return false; 1049 } 1050 1051 // Ok, we may insert a select for this PHI. 1052 PHIs.insert(std::make_pair(BB1V, BIParentV)); 1053 } 1054 1055 // If there are no PHIs to process, bail early. This helps ensure idempotence 1056 // as well. 1057 if (PHIs.empty()) 1058 return false; 1059 1060 // If we get here, we can hoist the instruction and if-convert. 1061 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); 1062 1063 // Hoist the instruction. 1064 if (HInst) 1065 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst); 1066 1067 // Insert selects and rewrite the PHI operands. 1068 IRBuilder<true, NoFolder> Builder(BI); 1069 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1070 Value *TrueV = PHIs[i].first; 1071 Value *FalseV = PHIs[i].second; 1072 1073 // Create a select whose true value is the speculatively executed value and 1074 // false value is the previously determined FalseV. 1075 SelectInst *SI; 1076 if (Invert) 1077 SI = cast<SelectInst> 1078 (Builder.CreateSelect(BrCond, FalseV, TrueV, 1079 FalseV->getName() + "." + TrueV->getName())); 1080 else 1081 SI = cast<SelectInst> 1082 (Builder.CreateSelect(BrCond, TrueV, FalseV, 1083 TrueV->getName() + "." + FalseV->getName())); 1084 1085 // Make the PHI node use the select for all incoming values for "then" and 1086 // "if" blocks. 1087 for (BasicBlock::iterator I = BB2->begin(); 1088 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1089 unsigned BB1I = PN->getBasicBlockIndex(BB1); 1090 unsigned BIParentI = PN->getBasicBlockIndex(BIParent); 1091 Value *BB1V = PN->getIncomingValue(BB1I); 1092 Value *BIParentV = PN->getIncomingValue(BIParentI); 1093 if (TrueV == BB1V && FalseV == BIParentV) { 1094 PN->setIncomingValue(BB1I, SI); 1095 PN->setIncomingValue(BIParentI, SI); 1096 } 1097 } 1098 } 1099 1100 ++NumSpeculations; 1101 return true; 1102} 1103 1104/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1105/// across this block. 1106static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1107 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1108 unsigned Size = 0; 1109 1110 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1111 if (isa<DbgInfoIntrinsic>(BBI)) 1112 continue; 1113 if (Size > 10) return false; // Don't clone large BB's. 1114 ++Size; 1115 1116 // We can only support instructions that do not define values that are 1117 // live outside of the current basic block. 1118 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1119 UI != E; ++UI) { 1120 Instruction *U = cast<Instruction>(*UI); 1121 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1122 } 1123 1124 // Looks ok, continue checking. 1125 } 1126 1127 return true; 1128} 1129 1130/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1131/// that is defined in the same block as the branch and if any PHI entries are 1132/// constants, thread edges corresponding to that entry to be branches to their 1133/// ultimate destination. 1134static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { 1135 BasicBlock *BB = BI->getParent(); 1136 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1137 // NOTE: we currently cannot transform this case if the PHI node is used 1138 // outside of the block. 1139 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1140 return false; 1141 1142 // Degenerate case of a single entry PHI. 1143 if (PN->getNumIncomingValues() == 1) { 1144 FoldSingleEntryPHINodes(PN->getParent()); 1145 return true; 1146 } 1147 1148 // Now we know that this block has multiple preds and two succs. 1149 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1150 1151 // Okay, this is a simple enough basic block. See if any phi values are 1152 // constants. 1153 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1154 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1155 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1156 1157 // Okay, we now know that all edges from PredBB should be revectored to 1158 // branch to RealDest. 1159 BasicBlock *PredBB = PN->getIncomingBlock(i); 1160 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1161 1162 if (RealDest == BB) continue; // Skip self loops. 1163 // Skip if the predecessor's terminator is an indirect branch. 1164 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1165 1166 // The dest block might have PHI nodes, other predecessors and other 1167 // difficult cases. Instead of being smart about this, just insert a new 1168 // block that jumps to the destination block, effectively splitting 1169 // the edge we are about to create. 1170 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1171 RealDest->getName()+".critedge", 1172 RealDest->getParent(), RealDest); 1173 BranchInst::Create(RealDest, EdgeBB); 1174 1175 // Update PHI nodes. 1176 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1177 1178 // BB may have instructions that are being threaded over. Clone these 1179 // instructions into EdgeBB. We know that there will be no uses of the 1180 // cloned instructions outside of EdgeBB. 1181 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1182 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1183 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1184 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1185 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1186 continue; 1187 } 1188 // Clone the instruction. 1189 Instruction *N = BBI->clone(); 1190 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1191 1192 // Update operands due to translation. 1193 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1194 i != e; ++i) { 1195 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1196 if (PI != TranslateMap.end()) 1197 *i = PI->second; 1198 } 1199 1200 // Check for trivial simplification. 1201 if (Value *V = SimplifyInstruction(N, TD)) { 1202 TranslateMap[BBI] = V; 1203 delete N; // Instruction folded away, don't need actual inst 1204 } else { 1205 // Insert the new instruction into its new home. 1206 EdgeBB->getInstList().insert(InsertPt, N); 1207 if (!BBI->use_empty()) 1208 TranslateMap[BBI] = N; 1209 } 1210 } 1211 1212 // Loop over all of the edges from PredBB to BB, changing them to branch 1213 // to EdgeBB instead. 1214 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1215 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1216 if (PredBBTI->getSuccessor(i) == BB) { 1217 BB->removePredecessor(PredBB); 1218 PredBBTI->setSuccessor(i, EdgeBB); 1219 } 1220 1221 // Recurse, simplifying any other constants. 1222 return FoldCondBranchOnPHI(BI, TD) | true; 1223 } 1224 1225 return false; 1226} 1227 1228/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1229/// PHI node, see if we can eliminate it. 1230static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { 1231 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1232 // statement", which has a very simple dominance structure. Basically, we 1233 // are trying to find the condition that is being branched on, which 1234 // subsequently causes this merge to happen. We really want control 1235 // dependence information for this check, but simplifycfg can't keep it up 1236 // to date, and this catches most of the cases we care about anyway. 1237 BasicBlock *BB = PN->getParent(); 1238 BasicBlock *IfTrue, *IfFalse; 1239 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1240 if (!IfCond || 1241 // Don't bother if the branch will be constant folded trivially. 1242 isa<ConstantInt>(IfCond)) 1243 return false; 1244 1245 // Okay, we found that we can merge this two-entry phi node into a select. 1246 // Doing so would require us to fold *all* two entry phi nodes in this block. 1247 // At some point this becomes non-profitable (particularly if the target 1248 // doesn't support cmov's). Only do this transformation if there are two or 1249 // fewer PHI nodes in this block. 1250 unsigned NumPhis = 0; 1251 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1252 if (NumPhis > 2) 1253 return false; 1254 1255 // Loop over the PHI's seeing if we can promote them all to select 1256 // instructions. While we are at it, keep track of the instructions 1257 // that need to be moved to the dominating block. 1258 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1259 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1260 MaxCostVal1 = PHINodeFoldingThreshold; 1261 1262 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1263 PHINode *PN = cast<PHINode>(II++); 1264 if (Value *V = SimplifyInstruction(PN, TD)) { 1265 PN->replaceAllUsesWith(V); 1266 PN->eraseFromParent(); 1267 continue; 1268 } 1269 1270 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1271 MaxCostVal0) || 1272 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1273 MaxCostVal1)) 1274 return false; 1275 } 1276 1277 // If we folded the the first phi, PN dangles at this point. Refresh it. If 1278 // we ran out of PHIs then we simplified them all. 1279 PN = dyn_cast<PHINode>(BB->begin()); 1280 if (PN == 0) return true; 1281 1282 // Don't fold i1 branches on PHIs which contain binary operators. These can 1283 // often be turned into switches and other things. 1284 if (PN->getType()->isIntegerTy(1) && 1285 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1286 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1287 isa<BinaryOperator>(IfCond))) 1288 return false; 1289 1290 // If we all PHI nodes are promotable, check to make sure that all 1291 // instructions in the predecessor blocks can be promoted as well. If 1292 // not, we won't be able to get rid of the control flow, so it's not 1293 // worth promoting to select instructions. 1294 BasicBlock *DomBlock = 0; 1295 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1296 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1297 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1298 IfBlock1 = 0; 1299 } else { 1300 DomBlock = *pred_begin(IfBlock1); 1301 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1302 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1303 // This is not an aggressive instruction that we can promote. 1304 // Because of this, we won't be able to get rid of the control 1305 // flow, so the xform is not worth it. 1306 return false; 1307 } 1308 } 1309 1310 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1311 IfBlock2 = 0; 1312 } else { 1313 DomBlock = *pred_begin(IfBlock2); 1314 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1315 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1316 // This is not an aggressive instruction that we can promote. 1317 // Because of this, we won't be able to get rid of the control 1318 // flow, so the xform is not worth it. 1319 return false; 1320 } 1321 } 1322 1323 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1324 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1325 1326 // If we can still promote the PHI nodes after this gauntlet of tests, 1327 // do all of the PHI's now. 1328 Instruction *InsertPt = DomBlock->getTerminator(); 1329 IRBuilder<true, NoFolder> Builder(InsertPt); 1330 1331 // Move all 'aggressive' instructions, which are defined in the 1332 // conditional parts of the if's up to the dominating block. 1333 if (IfBlock1) 1334 DomBlock->getInstList().splice(InsertPt, 1335 IfBlock1->getInstList(), IfBlock1->begin(), 1336 IfBlock1->getTerminator()); 1337 if (IfBlock2) 1338 DomBlock->getInstList().splice(InsertPt, 1339 IfBlock2->getInstList(), IfBlock2->begin(), 1340 IfBlock2->getTerminator()); 1341 1342 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1343 // Change the PHI node into a select instruction. 1344 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1345 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1346 1347 SelectInst *NV = 1348 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1349 PN->replaceAllUsesWith(NV); 1350 NV->takeName(PN); 1351 PN->eraseFromParent(); 1352 } 1353 1354 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1355 // has been flattened. Change DomBlock to jump directly to our new block to 1356 // avoid other simplifycfg's kicking in on the diamond. 1357 TerminatorInst *OldTI = DomBlock->getTerminator(); 1358 Builder.SetInsertPoint(OldTI); 1359 Builder.CreateBr(BB); 1360 OldTI->eraseFromParent(); 1361 return true; 1362} 1363 1364/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1365/// to two returning blocks, try to merge them together into one return, 1366/// introducing a select if the return values disagree. 1367static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1368 IRBuilder<> &Builder) { 1369 assert(BI->isConditional() && "Must be a conditional branch"); 1370 BasicBlock *TrueSucc = BI->getSuccessor(0); 1371 BasicBlock *FalseSucc = BI->getSuccessor(1); 1372 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1373 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1374 1375 // Check to ensure both blocks are empty (just a return) or optionally empty 1376 // with PHI nodes. If there are other instructions, merging would cause extra 1377 // computation on one path or the other. 1378 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1379 return false; 1380 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1381 return false; 1382 1383 Builder.SetInsertPoint(BI); 1384 // Okay, we found a branch that is going to two return nodes. If 1385 // there is no return value for this function, just change the 1386 // branch into a return. 1387 if (FalseRet->getNumOperands() == 0) { 1388 TrueSucc->removePredecessor(BI->getParent()); 1389 FalseSucc->removePredecessor(BI->getParent()); 1390 Builder.CreateRetVoid(); 1391 EraseTerminatorInstAndDCECond(BI); 1392 return true; 1393 } 1394 1395 // Otherwise, figure out what the true and false return values are 1396 // so we can insert a new select instruction. 1397 Value *TrueValue = TrueRet->getReturnValue(); 1398 Value *FalseValue = FalseRet->getReturnValue(); 1399 1400 // Unwrap any PHI nodes in the return blocks. 1401 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1402 if (TVPN->getParent() == TrueSucc) 1403 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1404 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1405 if (FVPN->getParent() == FalseSucc) 1406 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1407 1408 // In order for this transformation to be safe, we must be able to 1409 // unconditionally execute both operands to the return. This is 1410 // normally the case, but we could have a potentially-trapping 1411 // constant expression that prevents this transformation from being 1412 // safe. 1413 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1414 if (TCV->canTrap()) 1415 return false; 1416 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1417 if (FCV->canTrap()) 1418 return false; 1419 1420 // Okay, we collected all the mapped values and checked them for sanity, and 1421 // defined to really do this transformation. First, update the CFG. 1422 TrueSucc->removePredecessor(BI->getParent()); 1423 FalseSucc->removePredecessor(BI->getParent()); 1424 1425 // Insert select instructions where needed. 1426 Value *BrCond = BI->getCondition(); 1427 if (TrueValue) { 1428 // Insert a select if the results differ. 1429 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1430 } else if (isa<UndefValue>(TrueValue)) { 1431 TrueValue = FalseValue; 1432 } else { 1433 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1434 FalseValue, "retval"); 1435 } 1436 } 1437 1438 Value *RI = !TrueValue ? 1439 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1440 1441 (void) RI; 1442 1443 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1444 << "\n " << *BI << "NewRet = " << *RI 1445 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1446 1447 EraseTerminatorInstAndDCECond(BI); 1448 1449 return true; 1450} 1451 1452/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1453/// probabilities of the branch taking each edge. Fills in the two APInt 1454/// parameters and return true, or returns false if no or invalid metadata was 1455/// found. 1456static bool ExtractBranchMetadata(BranchInst *BI, 1457 APInt &ProbTrue, APInt &ProbFalse) { 1458 assert(BI->isConditional() && 1459 "Looking for probabilities on unconditional branch?"); 1460 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1461 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1462 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1463 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1464 if (!CITrue || !CIFalse) return false; 1465 ProbTrue = CITrue->getValue(); 1466 ProbFalse = CIFalse->getValue(); 1467 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 && 1468 "Branch probability metadata must be 32-bit integers"); 1469 return true; 1470} 1471 1472/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In 1473/// the event of overflow, logically-shifts all four inputs right until the 1474/// multiply fits. 1475static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D, 1476 unsigned &BitsLost) { 1477 BitsLost = 0; 1478 bool Overflow = false; 1479 APInt Result = A.umul_ov(B, Overflow); 1480 if (Overflow) { 1481 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A); 1482 do { 1483 B = B.lshr(1); 1484 ++BitsLost; 1485 } while (B.ugt(MaxB)); 1486 A = A.lshr(BitsLost); 1487 C = C.lshr(BitsLost); 1488 D = D.lshr(BitsLost); 1489 Result = A * B; 1490 } 1491 return Result; 1492} 1493 1494 1495/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1496/// predecessor branches to us and one of our successors, fold the block into 1497/// the predecessor and use logical operations to pick the right destination. 1498bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1499 BasicBlock *BB = BI->getParent(); 1500 1501 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1502 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1503 Cond->getParent() != BB || !Cond->hasOneUse()) 1504 return false; 1505 1506 // Only allow this if the condition is a simple instruction that can be 1507 // executed unconditionally. It must be in the same block as the branch, and 1508 // must be at the front of the block. 1509 BasicBlock::iterator FrontIt = BB->front(); 1510 1511 // Ignore dbg intrinsics. 1512 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1513 1514 // Allow a single instruction to be hoisted in addition to the compare 1515 // that feeds the branch. We later ensure that any values that _it_ uses 1516 // were also live in the predecessor, so that we don't unnecessarily create 1517 // register pressure or inhibit out-of-order execution. 1518 Instruction *BonusInst = 0; 1519 if (&*FrontIt != Cond && 1520 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1521 isSafeToSpeculativelyExecute(FrontIt)) { 1522 BonusInst = &*FrontIt; 1523 ++FrontIt; 1524 1525 // Ignore dbg intrinsics. 1526 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1527 } 1528 1529 // Only a single bonus inst is allowed. 1530 if (&*FrontIt != Cond) 1531 return false; 1532 1533 // Make sure the instruction after the condition is the cond branch. 1534 BasicBlock::iterator CondIt = Cond; ++CondIt; 1535 1536 // Ingore dbg intrinsics. 1537 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1538 1539 if (&*CondIt != BI) 1540 return false; 1541 1542 // Cond is known to be a compare or binary operator. Check to make sure that 1543 // neither operand is a potentially-trapping constant expression. 1544 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1545 if (CE->canTrap()) 1546 return false; 1547 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1548 if (CE->canTrap()) 1549 return false; 1550 1551 // Finally, don't infinitely unroll conditional loops. 1552 BasicBlock *TrueDest = BI->getSuccessor(0); 1553 BasicBlock *FalseDest = BI->getSuccessor(1); 1554 if (TrueDest == BB || FalseDest == BB) 1555 return false; 1556 1557 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1558 BasicBlock *PredBlock = *PI; 1559 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1560 1561 // Check that we have two conditional branches. If there is a PHI node in 1562 // the common successor, verify that the same value flows in from both 1563 // blocks. 1564 if (PBI == 0 || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI)) 1565 continue; 1566 1567 // Determine if the two branches share a common destination. 1568 Instruction::BinaryOps Opc; 1569 bool InvertPredCond = false; 1570 1571 if (PBI->getSuccessor(0) == TrueDest) 1572 Opc = Instruction::Or; 1573 else if (PBI->getSuccessor(1) == FalseDest) 1574 Opc = Instruction::And; 1575 else if (PBI->getSuccessor(0) == FalseDest) 1576 Opc = Instruction::And, InvertPredCond = true; 1577 else if (PBI->getSuccessor(1) == TrueDest) 1578 Opc = Instruction::Or, InvertPredCond = true; 1579 else 1580 continue; 1581 1582 // Ensure that any values used in the bonus instruction are also used 1583 // by the terminator of the predecessor. This means that those values 1584 // must already have been resolved, so we won't be inhibiting the 1585 // out-of-order core by speculating them earlier. 1586 if (BonusInst) { 1587 // Collect the values used by the bonus inst 1588 SmallPtrSet<Value*, 4> UsedValues; 1589 for (Instruction::op_iterator OI = BonusInst->op_begin(), 1590 OE = BonusInst->op_end(); OI != OE; ++OI) { 1591 Value *V = *OI; 1592 if (!isa<Constant>(V)) 1593 UsedValues.insert(V); 1594 } 1595 1596 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 1597 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 1598 1599 // Walk up to four levels back up the use-def chain of the predecessor's 1600 // terminator to see if all those values were used. The choice of four 1601 // levels is arbitrary, to provide a compile-time-cost bound. 1602 while (!Worklist.empty()) { 1603 std::pair<Value*, unsigned> Pair = Worklist.back(); 1604 Worklist.pop_back(); 1605 1606 if (Pair.second >= 4) continue; 1607 UsedValues.erase(Pair.first); 1608 if (UsedValues.empty()) break; 1609 1610 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 1611 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 1612 OI != OE; ++OI) 1613 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 1614 } 1615 } 1616 1617 if (!UsedValues.empty()) return false; 1618 } 1619 1620 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 1621 IRBuilder<> Builder(PBI); 1622 1623 // If we need to invert the condition in the pred block to match, do so now. 1624 if (InvertPredCond) { 1625 Value *NewCond = PBI->getCondition(); 1626 1627 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1628 CmpInst *CI = cast<CmpInst>(NewCond); 1629 CI->setPredicate(CI->getInversePredicate()); 1630 } else { 1631 NewCond = Builder.CreateNot(NewCond, 1632 PBI->getCondition()->getName()+".not"); 1633 } 1634 1635 PBI->setCondition(NewCond); 1636 PBI->swapSuccessors(); 1637 } 1638 1639 // If we have a bonus inst, clone it into the predecessor block. 1640 Instruction *NewBonus = 0; 1641 if (BonusInst) { 1642 NewBonus = BonusInst->clone(); 1643 PredBlock->getInstList().insert(PBI, NewBonus); 1644 NewBonus->takeName(BonusInst); 1645 BonusInst->setName(BonusInst->getName()+".old"); 1646 } 1647 1648 // Clone Cond into the predecessor basic block, and or/and the 1649 // two conditions together. 1650 Instruction *New = Cond->clone(); 1651 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 1652 PredBlock->getInstList().insert(PBI, New); 1653 New->takeName(Cond); 1654 Cond->setName(New->getName()+".old"); 1655 1656 Instruction *NewCond = 1657 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 1658 New, "or.cond")); 1659 PBI->setCondition(NewCond); 1660 if (PBI->getSuccessor(0) == BB) { 1661 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1662 PBI->setSuccessor(0, TrueDest); 1663 } 1664 if (PBI->getSuccessor(1) == BB) { 1665 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1666 PBI->setSuccessor(1, FalseDest); 1667 } 1668 1669 // TODO: If BB is reachable from all paths through PredBlock, then we 1670 // could replace PBI's branch probabilities with BI's. 1671 1672 // Merge probability data into PredBlock's branch. 1673 APInt A, B, C, D; 1674 if (ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) { 1675 // Given IR which does: 1676 // bbA: 1677 // br i1 %x, label %bbB, label %bbC 1678 // bbB: 1679 // br i1 %y, label %bbD, label %bbC 1680 // Let's call the probability that we take the edge from %bbA to %bbB 1681 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to 1682 // %bbC probability 'd'. 1683 // 1684 // We transform the IR into: 1685 // bbA: 1686 // br i1 %z, label %bbD, label %bbC 1687 // where the probability of going to %bbD is (a*c) and going to bbC is 1688 // (b+a*d). 1689 // 1690 // Probabilities aren't stored as ratios directly. Using branch weights, 1691 // we get: 1692 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D. 1693 1694 // In the event of overflow, we want to drop the LSB of the input 1695 // probabilities. 1696 unsigned BitsLost; 1697 1698 // Ignore overflow result on ProbTrue. 1699 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost); 1700 1701 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost); 1702 if (BitsLost) { 1703 ProbTrue = ProbTrue.lshr(BitsLost*2); 1704 } 1705 1706 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost); 1707 if (BitsLost) { 1708 ProbTrue = ProbTrue.lshr(BitsLost*2); 1709 Tmp1 = Tmp1.lshr(BitsLost*2); 1710 } 1711 1712 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost); 1713 if (BitsLost) { 1714 ProbTrue = ProbTrue.lshr(BitsLost*2); 1715 Tmp1 = Tmp1.lshr(BitsLost*2); 1716 Tmp2 = Tmp2.lshr(BitsLost*2); 1717 } 1718 1719 bool Overflow1 = false, Overflow2 = false; 1720 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1); 1721 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2); 1722 1723 if (Overflow1 || Overflow2) { 1724 ProbTrue = ProbTrue.lshr(1); 1725 Tmp1 = Tmp1.lshr(1); 1726 Tmp2 = Tmp2.lshr(1); 1727 Tmp3 = Tmp3.lshr(1); 1728 Tmp4 = Tmp2 + Tmp3; 1729 ProbFalse = Tmp4 + Tmp1; 1730 } 1731 1732 // The sum of branch weights must fit in 32-bits. 1733 if (ProbTrue.isNegative() && ProbFalse.isNegative()) { 1734 ProbTrue = ProbTrue.lshr(1); 1735 ProbFalse = ProbFalse.lshr(1); 1736 } 1737 1738 if (ProbTrue != ProbFalse) { 1739 // Normalize the result. 1740 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse); 1741 ProbTrue = ProbTrue.udiv(GCD); 1742 ProbFalse = ProbFalse.udiv(GCD); 1743 1744 LLVMContext &Context = BI->getContext(); 1745 Value *Ops[3]; 1746 Ops[0] = BI->getMetadata(LLVMContext::MD_prof)->getOperand(0); 1747 Ops[1] = ConstantInt::get(Context, ProbTrue); 1748 Ops[2] = ConstantInt::get(Context, ProbFalse); 1749 PBI->setMetadata(LLVMContext::MD_prof, MDNode::get(Context, Ops)); 1750 } else { 1751 PBI->setMetadata(LLVMContext::MD_prof, NULL); 1752 } 1753 } else { 1754 PBI->setMetadata(LLVMContext::MD_prof, NULL); 1755 } 1756 1757 // Copy any debug value intrinsics into the end of PredBlock. 1758 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 1759 if (isa<DbgInfoIntrinsic>(*I)) 1760 I->clone()->insertBefore(PBI); 1761 1762 return true; 1763 } 1764 return false; 1765} 1766 1767/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1768/// predecessor of another block, this function tries to simplify it. We know 1769/// that PBI and BI are both conditional branches, and BI is in one of the 1770/// successor blocks of PBI - PBI branches to BI. 1771static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1772 assert(PBI->isConditional() && BI->isConditional()); 1773 BasicBlock *BB = BI->getParent(); 1774 1775 // If this block ends with a branch instruction, and if there is a 1776 // predecessor that ends on a branch of the same condition, make 1777 // this conditional branch redundant. 1778 if (PBI->getCondition() == BI->getCondition() && 1779 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1780 // Okay, the outcome of this conditional branch is statically 1781 // knowable. If this block had a single pred, handle specially. 1782 if (BB->getSinglePredecessor()) { 1783 // Turn this into a branch on constant. 1784 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1785 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1786 CondIsTrue)); 1787 return true; // Nuke the branch on constant. 1788 } 1789 1790 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1791 // in the constant and simplify the block result. Subsequent passes of 1792 // simplifycfg will thread the block. 1793 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1794 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 1795 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 1796 std::distance(PB, PE), 1797 BI->getCondition()->getName() + ".pr", 1798 BB->begin()); 1799 // Okay, we're going to insert the PHI node. Since PBI is not the only 1800 // predecessor, compute the PHI'd conditional value for all of the preds. 1801 // Any predecessor where the condition is not computable we keep symbolic. 1802 for (pred_iterator PI = PB; PI != PE; ++PI) { 1803 BasicBlock *P = *PI; 1804 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 1805 PBI != BI && PBI->isConditional() && 1806 PBI->getCondition() == BI->getCondition() && 1807 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1808 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1809 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 1810 CondIsTrue), P); 1811 } else { 1812 NewPN->addIncoming(BI->getCondition(), P); 1813 } 1814 } 1815 1816 BI->setCondition(NewPN); 1817 return true; 1818 } 1819 } 1820 1821 // If this is a conditional branch in an empty block, and if any 1822 // predecessors is a conditional branch to one of our destinations, 1823 // fold the conditions into logical ops and one cond br. 1824 BasicBlock::iterator BBI = BB->begin(); 1825 // Ignore dbg intrinsics. 1826 while (isa<DbgInfoIntrinsic>(BBI)) 1827 ++BBI; 1828 if (&*BBI != BI) 1829 return false; 1830 1831 1832 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1833 if (CE->canTrap()) 1834 return false; 1835 1836 int PBIOp, BIOp; 1837 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1838 PBIOp = BIOp = 0; 1839 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1840 PBIOp = 0, BIOp = 1; 1841 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1842 PBIOp = 1, BIOp = 0; 1843 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1844 PBIOp = BIOp = 1; 1845 else 1846 return false; 1847 1848 // Check to make sure that the other destination of this branch 1849 // isn't BB itself. If so, this is an infinite loop that will 1850 // keep getting unwound. 1851 if (PBI->getSuccessor(PBIOp) == BB) 1852 return false; 1853 1854 // Do not perform this transformation if it would require 1855 // insertion of a large number of select instructions. For targets 1856 // without predication/cmovs, this is a big pessimization. 1857 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1858 1859 unsigned NumPhis = 0; 1860 for (BasicBlock::iterator II = CommonDest->begin(); 1861 isa<PHINode>(II); ++II, ++NumPhis) 1862 if (NumPhis > 2) // Disable this xform. 1863 return false; 1864 1865 // Finally, if everything is ok, fold the branches to logical ops. 1866 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1867 1868 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 1869 << "AND: " << *BI->getParent()); 1870 1871 1872 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1873 // branch in it, where one edge (OtherDest) goes back to itself but the other 1874 // exits. We don't *know* that the program avoids the infinite loop 1875 // (even though that seems likely). If we do this xform naively, we'll end up 1876 // recursively unpeeling the loop. Since we know that (after the xform is 1877 // done) that the block *is* infinite if reached, we just make it an obviously 1878 // infinite loop with no cond branch. 1879 if (OtherDest == BB) { 1880 // Insert it at the end of the function, because it's either code, 1881 // or it won't matter if it's hot. :) 1882 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 1883 "infloop", BB->getParent()); 1884 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1885 OtherDest = InfLoopBlock; 1886 } 1887 1888 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1889 1890 // BI may have other predecessors. Because of this, we leave 1891 // it alone, but modify PBI. 1892 1893 // Make sure we get to CommonDest on True&True directions. 1894 Value *PBICond = PBI->getCondition(); 1895 IRBuilder<true, NoFolder> Builder(PBI); 1896 if (PBIOp) 1897 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 1898 1899 Value *BICond = BI->getCondition(); 1900 if (BIOp) 1901 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 1902 1903 // Merge the conditions. 1904 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 1905 1906 // Modify PBI to branch on the new condition to the new dests. 1907 PBI->setCondition(Cond); 1908 PBI->setSuccessor(0, CommonDest); 1909 PBI->setSuccessor(1, OtherDest); 1910 1911 // OtherDest may have phi nodes. If so, add an entry from PBI's 1912 // block that are identical to the entries for BI's block. 1913 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 1914 1915 // We know that the CommonDest already had an edge from PBI to 1916 // it. If it has PHIs though, the PHIs may have different 1917 // entries for BB and PBI's BB. If so, insert a select to make 1918 // them agree. 1919 PHINode *PN; 1920 for (BasicBlock::iterator II = CommonDest->begin(); 1921 (PN = dyn_cast<PHINode>(II)); ++II) { 1922 Value *BIV = PN->getIncomingValueForBlock(BB); 1923 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1924 Value *PBIV = PN->getIncomingValue(PBBIdx); 1925 if (BIV != PBIV) { 1926 // Insert a select in PBI to pick the right value. 1927 Value *NV = cast<SelectInst> 1928 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 1929 PN->setIncomingValue(PBBIdx, NV); 1930 } 1931 } 1932 1933 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 1934 DEBUG(dbgs() << *PBI->getParent()->getParent()); 1935 1936 // This basic block is probably dead. We know it has at least 1937 // one fewer predecessor. 1938 return true; 1939} 1940 1941// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 1942// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 1943// Takes care of updating the successors and removing the old terminator. 1944// Also makes sure not to introduce new successors by assuming that edges to 1945// non-successor TrueBBs and FalseBBs aren't reachable. 1946static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 1947 BasicBlock *TrueBB, BasicBlock *FalseBB){ 1948 // Remove any superfluous successor edges from the CFG. 1949 // First, figure out which successors to preserve. 1950 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 1951 // successor. 1952 BasicBlock *KeepEdge1 = TrueBB; 1953 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 1954 1955 // Then remove the rest. 1956 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 1957 BasicBlock *Succ = OldTerm->getSuccessor(I); 1958 // Make sure only to keep exactly one copy of each edge. 1959 if (Succ == KeepEdge1) 1960 KeepEdge1 = 0; 1961 else if (Succ == KeepEdge2) 1962 KeepEdge2 = 0; 1963 else 1964 Succ->removePredecessor(OldTerm->getParent()); 1965 } 1966 1967 IRBuilder<> Builder(OldTerm); 1968 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 1969 1970 // Insert an appropriate new terminator. 1971 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 1972 if (TrueBB == FalseBB) 1973 // We were only looking for one successor, and it was present. 1974 // Create an unconditional branch to it. 1975 Builder.CreateBr(TrueBB); 1976 else 1977 // We found both of the successors we were looking for. 1978 // Create a conditional branch sharing the condition of the select. 1979 Builder.CreateCondBr(Cond, TrueBB, FalseBB); 1980 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 1981 // Neither of the selected blocks were successors, so this 1982 // terminator must be unreachable. 1983 new UnreachableInst(OldTerm->getContext(), OldTerm); 1984 } else { 1985 // One of the selected values was a successor, but the other wasn't. 1986 // Insert an unconditional branch to the one that was found; 1987 // the edge to the one that wasn't must be unreachable. 1988 if (KeepEdge1 == 0) 1989 // Only TrueBB was found. 1990 Builder.CreateBr(TrueBB); 1991 else 1992 // Only FalseBB was found. 1993 Builder.CreateBr(FalseBB); 1994 } 1995 1996 EraseTerminatorInstAndDCECond(OldTerm); 1997 return true; 1998} 1999 2000// SimplifySwitchOnSelect - Replaces 2001// (switch (select cond, X, Y)) on constant X, Y 2002// with a branch - conditional if X and Y lead to distinct BBs, 2003// unconditional otherwise. 2004static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2005 // Check for constant integer values in the select. 2006 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2007 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2008 if (!TrueVal || !FalseVal) 2009 return false; 2010 2011 // Find the relevant condition and destinations. 2012 Value *Condition = Select->getCondition(); 2013 unsigned TrueCase = SI->findCaseValue(TrueVal); 2014 unsigned FalseCase = SI->findCaseValue(FalseVal); 2015 BasicBlock *TrueBB = SI->getSuccessor(SI->resolveSuccessorIndex(TrueCase)); 2016 BasicBlock *FalseBB = SI->getSuccessor(SI->resolveSuccessorIndex(FalseCase)); 2017 2018 // Perform the actual simplification. 2019 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB); 2020} 2021 2022// SimplifyIndirectBrOnSelect - Replaces 2023// (indirectbr (select cond, blockaddress(@fn, BlockA), 2024// blockaddress(@fn, BlockB))) 2025// with 2026// (br cond, BlockA, BlockB). 2027static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2028 // Check that both operands of the select are block addresses. 2029 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2030 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2031 if (!TBA || !FBA) 2032 return false; 2033 2034 // Extract the actual blocks. 2035 BasicBlock *TrueBB = TBA->getBasicBlock(); 2036 BasicBlock *FalseBB = FBA->getBasicBlock(); 2037 2038 // Perform the actual simplification. 2039 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); 2040} 2041 2042/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2043/// instruction (a seteq/setne with a constant) as the only instruction in a 2044/// block that ends with an uncond branch. We are looking for a very specific 2045/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2046/// this case, we merge the first two "or's of icmp" into a switch, but then the 2047/// default value goes to an uncond block with a seteq in it, we get something 2048/// like: 2049/// 2050/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2051/// DEFAULT: 2052/// %tmp = icmp eq i8 %A, 92 2053/// br label %end 2054/// end: 2055/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2056/// 2057/// We prefer to split the edge to 'end' so that there is a true/false entry to 2058/// the PHI, merging the third icmp into the switch. 2059static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 2060 const TargetData *TD, 2061 IRBuilder<> &Builder) { 2062 BasicBlock *BB = ICI->getParent(); 2063 2064 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2065 // complex. 2066 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2067 2068 Value *V = ICI->getOperand(0); 2069 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2070 2071 // The pattern we're looking for is where our only predecessor is a switch on 2072 // 'V' and this block is the default case for the switch. In this case we can 2073 // fold the compared value into the switch to simplify things. 2074 BasicBlock *Pred = BB->getSinglePredecessor(); 2075 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2076 2077 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2078 if (SI->getCondition() != V) 2079 return false; 2080 2081 // If BB is reachable on a non-default case, then we simply know the value of 2082 // V in this block. Substitute it and constant fold the icmp instruction 2083 // away. 2084 if (SI->getDefaultDest() != BB) { 2085 ConstantInt *VVal = SI->findCaseDest(BB); 2086 assert(VVal && "Should have a unique destination value"); 2087 ICI->setOperand(0, VVal); 2088 2089 if (Value *V = SimplifyInstruction(ICI, TD)) { 2090 ICI->replaceAllUsesWith(V); 2091 ICI->eraseFromParent(); 2092 } 2093 // BB is now empty, so it is likely to simplify away. 2094 return SimplifyCFG(BB) | true; 2095 } 2096 2097 // Ok, the block is reachable from the default dest. If the constant we're 2098 // comparing exists in one of the other edges, then we can constant fold ICI 2099 // and zap it. 2100 if (SI->findCaseValue(Cst) != SwitchInst::ErrorIndex) { 2101 Value *V; 2102 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2103 V = ConstantInt::getFalse(BB->getContext()); 2104 else 2105 V = ConstantInt::getTrue(BB->getContext()); 2106 2107 ICI->replaceAllUsesWith(V); 2108 ICI->eraseFromParent(); 2109 // BB is now empty, so it is likely to simplify away. 2110 return SimplifyCFG(BB) | true; 2111 } 2112 2113 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2114 // the block. 2115 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2116 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2117 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2118 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2119 return false; 2120 2121 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2122 // true in the PHI. 2123 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2124 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2125 2126 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2127 std::swap(DefaultCst, NewCst); 2128 2129 // Replace ICI (which is used by the PHI for the default value) with true or 2130 // false depending on if it is EQ or NE. 2131 ICI->replaceAllUsesWith(DefaultCst); 2132 ICI->eraseFromParent(); 2133 2134 // Okay, the switch goes to this block on a default value. Add an edge from 2135 // the switch to the merge point on the compared value. 2136 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2137 BB->getParent(), BB); 2138 SI->addCase(Cst, NewBB); 2139 2140 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2141 Builder.SetInsertPoint(NewBB); 2142 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2143 Builder.CreateBr(SuccBlock); 2144 PHIUse->addIncoming(NewCst, NewBB); 2145 return true; 2146} 2147 2148/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2149/// Check to see if it is branching on an or/and chain of icmp instructions, and 2150/// fold it into a switch instruction if so. 2151static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, 2152 IRBuilder<> &Builder) { 2153 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2154 if (Cond == 0) return false; 2155 2156 2157 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2158 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2159 // 'setne's and'ed together, collect them. 2160 Value *CompVal = 0; 2161 std::vector<ConstantInt*> Values; 2162 bool TrueWhenEqual = true; 2163 Value *ExtraCase = 0; 2164 unsigned UsedICmps = 0; 2165 2166 if (Cond->getOpcode() == Instruction::Or) { 2167 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2168 UsedICmps); 2169 } else if (Cond->getOpcode() == Instruction::And) { 2170 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2171 UsedICmps); 2172 TrueWhenEqual = false; 2173 } 2174 2175 // If we didn't have a multiply compared value, fail. 2176 if (CompVal == 0) return false; 2177 2178 // Avoid turning single icmps into a switch. 2179 if (UsedICmps <= 1) 2180 return false; 2181 2182 // There might be duplicate constants in the list, which the switch 2183 // instruction can't handle, remove them now. 2184 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2185 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2186 2187 // If Extra was used, we require at least two switch values to do the 2188 // transformation. A switch with one value is just an cond branch. 2189 if (ExtraCase && Values.size() < 2) return false; 2190 2191 // Figure out which block is which destination. 2192 BasicBlock *DefaultBB = BI->getSuccessor(1); 2193 BasicBlock *EdgeBB = BI->getSuccessor(0); 2194 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2195 2196 BasicBlock *BB = BI->getParent(); 2197 2198 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2199 << " cases into SWITCH. BB is:\n" << *BB); 2200 2201 // If there are any extra values that couldn't be folded into the switch 2202 // then we evaluate them with an explicit branch first. Split the block 2203 // right before the condbr to handle it. 2204 if (ExtraCase) { 2205 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2206 // Remove the uncond branch added to the old block. 2207 TerminatorInst *OldTI = BB->getTerminator(); 2208 Builder.SetInsertPoint(OldTI); 2209 2210 if (TrueWhenEqual) 2211 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2212 else 2213 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2214 2215 OldTI->eraseFromParent(); 2216 2217 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2218 // for the edge we just added. 2219 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2220 2221 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2222 << "\nEXTRABB = " << *BB); 2223 BB = NewBB; 2224 } 2225 2226 Builder.SetInsertPoint(BI); 2227 // Convert pointer to int before we switch. 2228 if (CompVal->getType()->isPointerTy()) { 2229 assert(TD && "Cannot switch on pointer without TargetData"); 2230 CompVal = Builder.CreatePtrToInt(CompVal, 2231 TD->getIntPtrType(CompVal->getContext()), 2232 "magicptr"); 2233 } 2234 2235 // Create the new switch instruction now. 2236 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2237 2238 // Add all of the 'cases' to the switch instruction. 2239 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2240 New->addCase(Values[i], EdgeBB); 2241 2242 // We added edges from PI to the EdgeBB. As such, if there were any 2243 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2244 // the number of edges added. 2245 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2246 isa<PHINode>(BBI); ++BBI) { 2247 PHINode *PN = cast<PHINode>(BBI); 2248 Value *InVal = PN->getIncomingValueForBlock(BB); 2249 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2250 PN->addIncoming(InVal, BB); 2251 } 2252 2253 // Erase the old branch instruction. 2254 EraseTerminatorInstAndDCECond(BI); 2255 2256 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2257 return true; 2258} 2259 2260bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2261 // If this is a trivial landing pad that just continues unwinding the caught 2262 // exception then zap the landing pad, turning its invokes into calls. 2263 BasicBlock *BB = RI->getParent(); 2264 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2265 if (RI->getValue() != LPInst) 2266 // Not a landing pad, or the resume is not unwinding the exception that 2267 // caused control to branch here. 2268 return false; 2269 2270 // Check that there are no other instructions except for debug intrinsics. 2271 BasicBlock::iterator I = LPInst, E = RI; 2272 while (++I != E) 2273 if (!isa<DbgInfoIntrinsic>(I)) 2274 return false; 2275 2276 // Turn all invokes that unwind here into calls and delete the basic block. 2277 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2278 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2279 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2280 // Insert a call instruction before the invoke. 2281 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2282 Call->takeName(II); 2283 Call->setCallingConv(II->getCallingConv()); 2284 Call->setAttributes(II->getAttributes()); 2285 Call->setDebugLoc(II->getDebugLoc()); 2286 2287 // Anything that used the value produced by the invoke instruction now uses 2288 // the value produced by the call instruction. Note that we do this even 2289 // for void functions and calls with no uses so that the callgraph edge is 2290 // updated. 2291 II->replaceAllUsesWith(Call); 2292 BB->removePredecessor(II->getParent()); 2293 2294 // Insert a branch to the normal destination right before the invoke. 2295 BranchInst::Create(II->getNormalDest(), II); 2296 2297 // Finally, delete the invoke instruction! 2298 II->eraseFromParent(); 2299 } 2300 2301 // The landingpad is now unreachable. Zap it. 2302 BB->eraseFromParent(); 2303 return true; 2304} 2305 2306bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2307 BasicBlock *BB = RI->getParent(); 2308 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2309 2310 // Find predecessors that end with branches. 2311 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2312 SmallVector<BranchInst*, 8> CondBranchPreds; 2313 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2314 BasicBlock *P = *PI; 2315 TerminatorInst *PTI = P->getTerminator(); 2316 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2317 if (BI->isUnconditional()) 2318 UncondBranchPreds.push_back(P); 2319 else 2320 CondBranchPreds.push_back(BI); 2321 } 2322 } 2323 2324 // If we found some, do the transformation! 2325 if (!UncondBranchPreds.empty() && DupRet) { 2326 while (!UncondBranchPreds.empty()) { 2327 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2328 DEBUG(dbgs() << "FOLDING: " << *BB 2329 << "INTO UNCOND BRANCH PRED: " << *Pred); 2330 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2331 } 2332 2333 // If we eliminated all predecessors of the block, delete the block now. 2334 if (pred_begin(BB) == pred_end(BB)) 2335 // We know there are no successors, so just nuke the block. 2336 BB->eraseFromParent(); 2337 2338 return true; 2339 } 2340 2341 // Check out all of the conditional branches going to this return 2342 // instruction. If any of them just select between returns, change the 2343 // branch itself into a select/return pair. 2344 while (!CondBranchPreds.empty()) { 2345 BranchInst *BI = CondBranchPreds.pop_back_val(); 2346 2347 // Check to see if the non-BB successor is also a return block. 2348 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2349 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2350 SimplifyCondBranchToTwoReturns(BI, Builder)) 2351 return true; 2352 } 2353 return false; 2354} 2355 2356bool SimplifyCFGOpt::SimplifyUnwind(UnwindInst *UI, IRBuilder<> &Builder) { 2357 // Check to see if the first instruction in this block is just an unwind. 2358 // If so, replace any invoke instructions which use this as an exception 2359 // destination with call instructions. 2360 BasicBlock *BB = UI->getParent(); 2361 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2362 2363 bool Changed = false; 2364 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2365 while (!Preds.empty()) { 2366 BasicBlock *Pred = Preds.back(); 2367 InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator()); 2368 if (II && II->getUnwindDest() == BB) { 2369 // Insert a new branch instruction before the invoke, because this 2370 // is now a fall through. 2371 Builder.SetInsertPoint(II); 2372 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2373 Pred->getInstList().remove(II); // Take out of symbol table 2374 2375 // Insert the call now. 2376 SmallVector<Value*,8> Args(II->op_begin(), II->op_end()-3); 2377 Builder.SetInsertPoint(BI); 2378 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2379 Args, II->getName()); 2380 CI->setCallingConv(II->getCallingConv()); 2381 CI->setAttributes(II->getAttributes()); 2382 // If the invoke produced a value, the Call now does instead. 2383 II->replaceAllUsesWith(CI); 2384 delete II; 2385 Changed = true; 2386 } 2387 2388 Preds.pop_back(); 2389 } 2390 2391 // If this block is now dead (and isn't the entry block), remove it. 2392 if (pred_begin(BB) == pred_end(BB) && 2393 BB != &BB->getParent()->getEntryBlock()) { 2394 // We know there are no successors, so just nuke the block. 2395 BB->eraseFromParent(); 2396 return true; 2397 } 2398 2399 return Changed; 2400} 2401 2402bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2403 BasicBlock *BB = UI->getParent(); 2404 2405 bool Changed = false; 2406 2407 // If there are any instructions immediately before the unreachable that can 2408 // be removed, do so. 2409 while (UI != BB->begin()) { 2410 BasicBlock::iterator BBI = UI; 2411 --BBI; 2412 // Do not delete instructions that can have side effects which might cause 2413 // the unreachable to not be reachable; specifically, calls and volatile 2414 // operations may have this effect. 2415 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2416 2417 if (BBI->mayHaveSideEffects()) { 2418 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2419 if (SI->isVolatile()) 2420 break; 2421 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2422 if (LI->isVolatile()) 2423 break; 2424 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2425 if (RMWI->isVolatile()) 2426 break; 2427 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2428 if (CXI->isVolatile()) 2429 break; 2430 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2431 !isa<LandingPadInst>(BBI)) { 2432 break; 2433 } 2434 // Note that deleting LandingPad's here is in fact okay, although it 2435 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2436 // all the predecessors of this block will be the unwind edges of Invokes, 2437 // and we can therefore guarantee this block will be erased. 2438 } 2439 2440 // Delete this instruction (any uses are guaranteed to be dead) 2441 if (!BBI->use_empty()) 2442 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2443 BBI->eraseFromParent(); 2444 Changed = true; 2445 } 2446 2447 // If the unreachable instruction is the first in the block, take a gander 2448 // at all of the predecessors of this instruction, and simplify them. 2449 if (&BB->front() != UI) return Changed; 2450 2451 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2452 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2453 TerminatorInst *TI = Preds[i]->getTerminator(); 2454 IRBuilder<> Builder(TI); 2455 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2456 if (BI->isUnconditional()) { 2457 if (BI->getSuccessor(0) == BB) { 2458 new UnreachableInst(TI->getContext(), TI); 2459 TI->eraseFromParent(); 2460 Changed = true; 2461 } 2462 } else { 2463 if (BI->getSuccessor(0) == BB) { 2464 Builder.CreateBr(BI->getSuccessor(1)); 2465 EraseTerminatorInstAndDCECond(BI); 2466 } else if (BI->getSuccessor(1) == BB) { 2467 Builder.CreateBr(BI->getSuccessor(0)); 2468 EraseTerminatorInstAndDCECond(BI); 2469 Changed = true; 2470 } 2471 } 2472 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2473 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) 2474 if (SI->getCaseSuccessor(i) == BB) { 2475 BB->removePredecessor(SI->getParent()); 2476 SI->removeCase(i); 2477 --i; --e; 2478 Changed = true; 2479 } 2480 // If the default value is unreachable, figure out the most popular 2481 // destination and make it the default. 2482 if (SI->getDefaultDest() == BB) { 2483 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2484 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) { 2485 std::pair<unsigned, unsigned> &entry = 2486 Popularity[SI->getCaseSuccessor(i)]; 2487 if (entry.first == 0) { 2488 entry.first = 1; 2489 entry.second = i; 2490 } else { 2491 entry.first++; 2492 } 2493 } 2494 2495 // Find the most popular block. 2496 unsigned MaxPop = 0; 2497 unsigned MaxIndex = 0; 2498 BasicBlock *MaxBlock = 0; 2499 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2500 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2501 if (I->second.first > MaxPop || 2502 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2503 MaxPop = I->second.first; 2504 MaxIndex = I->second.second; 2505 MaxBlock = I->first; 2506 } 2507 } 2508 if (MaxBlock) { 2509 // Make this the new default, allowing us to delete any explicit 2510 // edges to it. 2511 SI->setDefaultDest(MaxBlock); 2512 Changed = true; 2513 2514 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2515 // it. 2516 if (isa<PHINode>(MaxBlock->begin())) 2517 for (unsigned i = 0; i != MaxPop-1; ++i) 2518 MaxBlock->removePredecessor(SI->getParent()); 2519 2520 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) 2521 if (SI->getCaseSuccessor(i) == MaxBlock) { 2522 SI->removeCase(i); 2523 --i; --e; 2524 } 2525 } 2526 } 2527 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2528 if (II->getUnwindDest() == BB) { 2529 // Convert the invoke to a call instruction. This would be a good 2530 // place to note that the call does not throw though. 2531 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2532 II->removeFromParent(); // Take out of symbol table 2533 2534 // Insert the call now... 2535 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2536 Builder.SetInsertPoint(BI); 2537 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2538 Args, II->getName()); 2539 CI->setCallingConv(II->getCallingConv()); 2540 CI->setAttributes(II->getAttributes()); 2541 // If the invoke produced a value, the call does now instead. 2542 II->replaceAllUsesWith(CI); 2543 delete II; 2544 Changed = true; 2545 } 2546 } 2547 } 2548 2549 // If this block is now dead, remove it. 2550 if (pred_begin(BB) == pred_end(BB) && 2551 BB != &BB->getParent()->getEntryBlock()) { 2552 // We know there are no successors, so just nuke the block. 2553 BB->eraseFromParent(); 2554 return true; 2555 } 2556 2557 return Changed; 2558} 2559 2560/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 2561/// integer range comparison into a sub, an icmp and a branch. 2562static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 2563 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 2564 2565 // Make sure all cases point to the same destination and gather the values. 2566 SmallVector<ConstantInt *, 16> Cases; 2567 Cases.push_back(SI->getCaseValue(0)); 2568 for (unsigned I = 1, E = SI->getNumCases(); I != E; ++I) { 2569 if (SI->getCaseSuccessor(I-1) != SI->getCaseSuccessor(I)) 2570 return false; 2571 Cases.push_back(SI->getCaseValue(I)); 2572 } 2573 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 2574 2575 // Sort the case values, then check if they form a range we can transform. 2576 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 2577 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 2578 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 2579 return false; 2580 } 2581 2582 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 2583 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 2584 2585 Value *Sub = SI->getCondition(); 2586 if (!Offset->isNullValue()) 2587 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 2588 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 2589 Builder.CreateCondBr(Cmp, SI->getCaseSuccessor(0), SI->getDefaultDest()); 2590 2591 // Prune obsolete incoming values off the successor's PHI nodes. 2592 for (BasicBlock::iterator BBI = SI->getCaseSuccessor(0)->begin(); 2593 isa<PHINode>(BBI); ++BBI) { 2594 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 2595 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 2596 } 2597 SI->eraseFromParent(); 2598 2599 return true; 2600} 2601 2602/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 2603/// and use it to remove dead cases. 2604static bool EliminateDeadSwitchCases(SwitchInst *SI) { 2605 Value *Cond = SI->getCondition(); 2606 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 2607 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 2608 ComputeMaskedBits(Cond, APInt::getAllOnesValue(Bits), KnownZero, KnownOne); 2609 2610 // Gather dead cases. 2611 SmallVector<ConstantInt*, 8> DeadCases; 2612 for (unsigned I = 0, E = SI->getNumCases(); I != E; ++I) { 2613 if ((SI->getCaseValue(I)->getValue() & KnownZero) != 0 || 2614 (SI->getCaseValue(I)->getValue() & KnownOne) != KnownOne) { 2615 DeadCases.push_back(SI->getCaseValue(I)); 2616 DEBUG(dbgs() << "SimplifyCFG: switch case '" 2617 << SI->getCaseValue(I)->getValue() << "' is dead.\n"); 2618 } 2619 } 2620 2621 // Remove dead cases from the switch. 2622 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 2623 unsigned Case = SI->findCaseValue(DeadCases[I]); 2624 assert(Case != SwitchInst::ErrorIndex && 2625 "Case was not found. Probably mistake in DeadCases forming."); 2626 // Prune unused values from PHI nodes. 2627 SI->getCaseSuccessor(Case)->removePredecessor(SI->getParent()); 2628 SI->removeCase(Case); 2629 } 2630 2631 return !DeadCases.empty(); 2632} 2633 2634/// FindPHIForConditionForwarding - If BB would be eligible for simplification 2635/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 2636/// by an unconditional branch), look at the phi node for BB in the successor 2637/// block and see if the incoming value is equal to CaseValue. If so, return 2638/// the phi node, and set PhiIndex to BB's index in the phi node. 2639static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 2640 BasicBlock *BB, 2641 int *PhiIndex) { 2642 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 2643 return NULL; // BB must be empty to be a candidate for simplification. 2644 if (!BB->getSinglePredecessor()) 2645 return NULL; // BB must be dominated by the switch. 2646 2647 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 2648 if (!Branch || !Branch->isUnconditional()) 2649 return NULL; // Terminator must be unconditional branch. 2650 2651 BasicBlock *Succ = Branch->getSuccessor(0); 2652 2653 BasicBlock::iterator I = Succ->begin(); 2654 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 2655 int Idx = PHI->getBasicBlockIndex(BB); 2656 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 2657 2658 Value *InValue = PHI->getIncomingValue(Idx); 2659 if (InValue != CaseValue) continue; 2660 2661 *PhiIndex = Idx; 2662 return PHI; 2663 } 2664 2665 return NULL; 2666} 2667 2668/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 2669/// instruction to a phi node dominated by the switch, if that would mean that 2670/// some of the destination blocks of the switch can be folded away. 2671/// Returns true if a change is made. 2672static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 2673 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 2674 ForwardingNodesMap ForwardingNodes; 2675 2676 for (unsigned I = 0; I < SI->getNumCases(); ++I) { // 0 is the default case. 2677 ConstantInt *CaseValue = SI->getCaseValue(I); 2678 BasicBlock *CaseDest = SI->getCaseSuccessor(I); 2679 2680 int PhiIndex; 2681 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 2682 &PhiIndex); 2683 if (!PHI) continue; 2684 2685 ForwardingNodes[PHI].push_back(PhiIndex); 2686 } 2687 2688 bool Changed = false; 2689 2690 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 2691 E = ForwardingNodes.end(); I != E; ++I) { 2692 PHINode *Phi = I->first; 2693 SmallVector<int,4> &Indexes = I->second; 2694 2695 if (Indexes.size() < 2) continue; 2696 2697 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 2698 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 2699 Changed = true; 2700 } 2701 2702 return Changed; 2703} 2704 2705bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 2706 // If this switch is too complex to want to look at, ignore it. 2707 if (!isValueEqualityComparison(SI)) 2708 return false; 2709 2710 BasicBlock *BB = SI->getParent(); 2711 2712 // If we only have one predecessor, and if it is a branch on this value, 2713 // see if that predecessor totally determines the outcome of this switch. 2714 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2715 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 2716 return SimplifyCFG(BB) | true; 2717 2718 Value *Cond = SI->getCondition(); 2719 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 2720 if (SimplifySwitchOnSelect(SI, Select)) 2721 return SimplifyCFG(BB) | true; 2722 2723 // If the block only contains the switch, see if we can fold the block 2724 // away into any preds. 2725 BasicBlock::iterator BBI = BB->begin(); 2726 // Ignore dbg intrinsics. 2727 while (isa<DbgInfoIntrinsic>(BBI)) 2728 ++BBI; 2729 if (SI == &*BBI) 2730 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 2731 return SimplifyCFG(BB) | true; 2732 2733 // Try to transform the switch into an icmp and a branch. 2734 if (TurnSwitchRangeIntoICmp(SI, Builder)) 2735 return SimplifyCFG(BB) | true; 2736 2737 // Remove unreachable cases. 2738 if (EliminateDeadSwitchCases(SI)) 2739 return SimplifyCFG(BB) | true; 2740 2741 if (ForwardSwitchConditionToPHI(SI)) 2742 return SimplifyCFG(BB) | true; 2743 2744 return false; 2745} 2746 2747bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 2748 BasicBlock *BB = IBI->getParent(); 2749 bool Changed = false; 2750 2751 // Eliminate redundant destinations. 2752 SmallPtrSet<Value *, 8> Succs; 2753 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 2754 BasicBlock *Dest = IBI->getDestination(i); 2755 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 2756 Dest->removePredecessor(BB); 2757 IBI->removeDestination(i); 2758 --i; --e; 2759 Changed = true; 2760 } 2761 } 2762 2763 if (IBI->getNumDestinations() == 0) { 2764 // If the indirectbr has no successors, change it to unreachable. 2765 new UnreachableInst(IBI->getContext(), IBI); 2766 EraseTerminatorInstAndDCECond(IBI); 2767 return true; 2768 } 2769 2770 if (IBI->getNumDestinations() == 1) { 2771 // If the indirectbr has one successor, change it to a direct branch. 2772 BranchInst::Create(IBI->getDestination(0), IBI); 2773 EraseTerminatorInstAndDCECond(IBI); 2774 return true; 2775 } 2776 2777 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 2778 if (SimplifyIndirectBrOnSelect(IBI, SI)) 2779 return SimplifyCFG(BB) | true; 2780 } 2781 return Changed; 2782} 2783 2784bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 2785 BasicBlock *BB = BI->getParent(); 2786 2787 // If the Terminator is the only non-phi instruction, simplify the block. 2788 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 2789 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 2790 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 2791 return true; 2792 2793 // If the only instruction in the block is a seteq/setne comparison 2794 // against a constant, try to simplify the block. 2795 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 2796 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 2797 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 2798 ; 2799 if (I->isTerminator() && 2800 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) 2801 return true; 2802 } 2803 2804 return false; 2805} 2806 2807 2808bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 2809 BasicBlock *BB = BI->getParent(); 2810 2811 // Conditional branch 2812 if (isValueEqualityComparison(BI)) { 2813 // If we only have one predecessor, and if it is a branch on this value, 2814 // see if that predecessor totally determines the outcome of this 2815 // switch. 2816 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 2817 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 2818 return SimplifyCFG(BB) | true; 2819 2820 // This block must be empty, except for the setcond inst, if it exists. 2821 // Ignore dbg intrinsics. 2822 BasicBlock::iterator I = BB->begin(); 2823 // Ignore dbg intrinsics. 2824 while (isa<DbgInfoIntrinsic>(I)) 2825 ++I; 2826 if (&*I == BI) { 2827 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 2828 return SimplifyCFG(BB) | true; 2829 } else if (&*I == cast<Instruction>(BI->getCondition())){ 2830 ++I; 2831 // Ignore dbg intrinsics. 2832 while (isa<DbgInfoIntrinsic>(I)) 2833 ++I; 2834 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 2835 return SimplifyCFG(BB) | true; 2836 } 2837 } 2838 2839 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 2840 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 2841 return true; 2842 2843 // If this basic block is ONLY a compare and a branch, and if a predecessor 2844 // branches to us and one of our successors, fold the comparison into the 2845 // predecessor and use logical operations to pick the right destination. 2846 if (FoldBranchToCommonDest(BI)) 2847 return SimplifyCFG(BB) | true; 2848 2849 // We have a conditional branch to two blocks that are only reachable 2850 // from BI. We know that the condbr dominates the two blocks, so see if 2851 // there is any identical code in the "then" and "else" blocks. If so, we 2852 // can hoist it up to the branching block. 2853 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 2854 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2855 if (HoistThenElseCodeToIf(BI)) 2856 return SimplifyCFG(BB) | true; 2857 } else { 2858 // If Successor #1 has multiple preds, we may be able to conditionally 2859 // execute Successor #0 if it branches to successor #1. 2860 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 2861 if (Succ0TI->getNumSuccessors() == 1 && 2862 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 2863 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 2864 return SimplifyCFG(BB) | true; 2865 } 2866 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 2867 // If Successor #0 has multiple preds, we may be able to conditionally 2868 // execute Successor #1 if it branches to successor #0. 2869 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 2870 if (Succ1TI->getNumSuccessors() == 1 && 2871 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 2872 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 2873 return SimplifyCFG(BB) | true; 2874 } 2875 2876 // If this is a branch on a phi node in the current block, thread control 2877 // through this block if any PHI node entries are constants. 2878 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 2879 if (PN->getParent() == BI->getParent()) 2880 if (FoldCondBranchOnPHI(BI, TD)) 2881 return SimplifyCFG(BB) | true; 2882 2883 // Scan predecessor blocks for conditional branches. 2884 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2885 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2886 if (PBI != BI && PBI->isConditional()) 2887 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2888 return SimplifyCFG(BB) | true; 2889 2890 return false; 2891} 2892 2893/// Check if passing a value to an instruction will cause undefined behavior. 2894static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 2895 Constant *C = dyn_cast<Constant>(V); 2896 if (!C) 2897 return false; 2898 2899 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time 2900 return false; 2901 2902 if (C->isNullValue()) { 2903 Instruction *Use = I->use_back(); 2904 2905 // Now make sure that there are no instructions in between that can alter 2906 // control flow (eg. calls) 2907 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 2908 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 2909 return false; 2910 2911 // Look through GEPs. A load from a GEP derived from NULL is still undefined 2912 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 2913 if (GEP->getPointerOperand() == I) 2914 return passingValueIsAlwaysUndefined(V, GEP); 2915 2916 // Look through bitcasts. 2917 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 2918 return passingValueIsAlwaysUndefined(V, BC); 2919 2920 // Load from null is undefined. 2921 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 2922 return LI->getPointerAddressSpace() == 0; 2923 2924 // Store to null is undefined. 2925 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 2926 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 2927 } 2928 return false; 2929} 2930 2931/// If BB has an incoming value that will always trigger undefined behavior 2932/// (eg. null pointer dereference), remove the branch leading here. 2933static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 2934 for (BasicBlock::iterator i = BB->begin(); 2935 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 2936 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 2937 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 2938 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 2939 IRBuilder<> Builder(T); 2940 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 2941 BB->removePredecessor(PHI->getIncomingBlock(i)); 2942 // Turn uncoditional branches into unreachables and remove the dead 2943 // destination from conditional branches. 2944 if (BI->isUnconditional()) 2945 Builder.CreateUnreachable(); 2946 else 2947 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 2948 BI->getSuccessor(0)); 2949 BI->eraseFromParent(); 2950 return true; 2951 } 2952 // TODO: SwitchInst. 2953 } 2954 2955 return false; 2956} 2957 2958bool SimplifyCFGOpt::run(BasicBlock *BB) { 2959 bool Changed = false; 2960 2961 assert(BB && BB->getParent() && "Block not embedded in function!"); 2962 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 2963 2964 // Remove basic blocks that have no predecessors (except the entry block)... 2965 // or that just have themself as a predecessor. These are unreachable. 2966 if ((pred_begin(BB) == pred_end(BB) && 2967 BB != &BB->getParent()->getEntryBlock()) || 2968 BB->getSinglePredecessor() == BB) { 2969 DEBUG(dbgs() << "Removing BB: \n" << *BB); 2970 DeleteDeadBlock(BB); 2971 return true; 2972 } 2973 2974 // Check to see if we can constant propagate this terminator instruction 2975 // away... 2976 Changed |= ConstantFoldTerminator(BB, true); 2977 2978 // Check for and eliminate duplicate PHI nodes in this block. 2979 Changed |= EliminateDuplicatePHINodes(BB); 2980 2981 // Check for and remove branches that will always cause undefined behavior. 2982 Changed |= removeUndefIntroducingPredecessor(BB); 2983 2984 // Merge basic blocks into their predecessor if there is only one distinct 2985 // pred, and if there is only one distinct successor of the predecessor, and 2986 // if there are no PHI nodes. 2987 // 2988 if (MergeBlockIntoPredecessor(BB)) 2989 return true; 2990 2991 IRBuilder<> Builder(BB); 2992 2993 // If there is a trivial two-entry PHI node in this basic block, and we can 2994 // eliminate it, do so now. 2995 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 2996 if (PN->getNumIncomingValues() == 2) 2997 Changed |= FoldTwoEntryPHINode(PN, TD); 2998 2999 Builder.SetInsertPoint(BB->getTerminator()); 3000 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 3001 if (BI->isUnconditional()) { 3002 if (SimplifyUncondBranch(BI, Builder)) return true; 3003 } else { 3004 if (SimplifyCondBranch(BI, Builder)) return true; 3005 } 3006 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 3007 if (SimplifyResume(RI, Builder)) return true; 3008 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 3009 if (SimplifyReturn(RI, Builder)) return true; 3010 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 3011 if (SimplifySwitch(SI, Builder)) return true; 3012 } else if (UnreachableInst *UI = 3013 dyn_cast<UnreachableInst>(BB->getTerminator())) { 3014 if (SimplifyUnreachable(UI)) return true; 3015 } else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 3016 if (SimplifyUnwind(UI, Builder)) return true; 3017 } else if (IndirectBrInst *IBI = 3018 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 3019 if (SimplifyIndirectBr(IBI)) return true; 3020 } 3021 3022 return Changed; 3023} 3024 3025/// SimplifyCFG - This function is used to do simplification of a CFG. For 3026/// example, it adjusts branches to branches to eliminate the extra hop, it 3027/// eliminates unreachable basic blocks, and does other "peephole" optimization 3028/// of the CFG. It returns true if a modification was made. 3029/// 3030bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { 3031 return SimplifyCFGOpt(TD).run(BB); 3032} 3033