SimplifyCFG.cpp revision c2aa0035a6c0bd69e4f46620a270451fe585f8cd
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/DataLayout.h" 18#include "llvm/DerivedTypes.h" 19#include "llvm/GlobalVariable.h" 20#include "llvm/IRBuilder.h" 21#include "llvm/Instructions.h" 22#include "llvm/IntrinsicInst.h" 23#include "llvm/LLVMContext.h" 24#include "llvm/MDBuilder.h" 25#include "llvm/Metadata.h" 26#include "llvm/Module.h" 27#include "llvm/Operator.h" 28#include "llvm/Type.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/STLExtras.h" 31#include "llvm/ADT/SetVector.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/SmallVector.h" 34#include "llvm/ADT/Statistic.h" 35#include "llvm/Analysis/InstructionSimplify.h" 36#include "llvm/Analysis/ValueTracking.h" 37#include "llvm/Support/CFG.h" 38#include "llvm/Support/CommandLine.h" 39#include "llvm/Support/ConstantRange.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/NoFolder.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/TargetTransformInfo.h" 44#include "llvm/Transforms/Utils/BasicBlockUtils.h" 45#include <algorithm> 46#include <set> 47#include <map> 48using namespace llvm; 49 50static cl::opt<unsigned> 51PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), 52 cl::desc("Control the amount of phi node folding to perform (default = 1)")); 53 54static cl::opt<bool> 55DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), 56 cl::desc("Duplicate return instructions into unconditional branches")); 57 58static cl::opt<bool> 59SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), 60 cl::desc("Sink common instructions down to the end block")); 61 62STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); 63STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); 64STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); 65STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 66 67namespace { 68 /// ValueEqualityComparisonCase - Represents a case of a switch. 69 struct ValueEqualityComparisonCase { 70 ConstantInt *Value; 71 BasicBlock *Dest; 72 73 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) 74 : Value(Value), Dest(Dest) {} 75 76 bool operator<(ValueEqualityComparisonCase RHS) const { 77 // Comparing pointers is ok as we only rely on the order for uniquing. 78 return Value < RHS.Value; 79 } 80 81 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } 82 }; 83 84class SimplifyCFGOpt { 85 const DataLayout *const TD; 86 const TargetTransformInfo *const TTI; 87 88 Value *isValueEqualityComparison(TerminatorInst *TI); 89 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, 90 std::vector<ValueEqualityComparisonCase> &Cases); 91 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 92 BasicBlock *Pred, 93 IRBuilder<> &Builder); 94 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 95 IRBuilder<> &Builder); 96 97 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder); 98 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder); 99 bool SimplifyUnreachable(UnreachableInst *UI); 100 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); 101 bool SimplifyIndirectBr(IndirectBrInst *IBI); 102 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder); 103 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); 104 105public: 106 SimplifyCFGOpt(const DataLayout *td, const TargetTransformInfo *tti) 107 : TD(td), TTI(tti) {} 108 bool run(BasicBlock *BB); 109}; 110} 111 112/// SafeToMergeTerminators - Return true if it is safe to merge these two 113/// terminator instructions together. 114/// 115static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 116 if (SI1 == SI2) return false; // Can't merge with self! 117 118 // It is not safe to merge these two switch instructions if they have a common 119 // successor, and if that successor has a PHI node, and if *that* PHI node has 120 // conflicting incoming values from the two switch blocks. 121 BasicBlock *SI1BB = SI1->getParent(); 122 BasicBlock *SI2BB = SI2->getParent(); 123 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 124 125 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 126 if (SI1Succs.count(*I)) 127 for (BasicBlock::iterator BBI = (*I)->begin(); 128 isa<PHINode>(BBI); ++BBI) { 129 PHINode *PN = cast<PHINode>(BBI); 130 if (PN->getIncomingValueForBlock(SI1BB) != 131 PN->getIncomingValueForBlock(SI2BB)) 132 return false; 133 } 134 135 return true; 136} 137 138/// isProfitableToFoldUnconditional - Return true if it is safe and profitable 139/// to merge these two terminator instructions together, where SI1 is an 140/// unconditional branch. PhiNodes will store all PHI nodes in common 141/// successors. 142/// 143static bool isProfitableToFoldUnconditional(BranchInst *SI1, 144 BranchInst *SI2, 145 Instruction *Cond, 146 SmallVectorImpl<PHINode*> &PhiNodes) { 147 if (SI1 == SI2) return false; // Can't merge with self! 148 assert(SI1->isUnconditional() && SI2->isConditional()); 149 150 // We fold the unconditional branch if we can easily update all PHI nodes in 151 // common successors: 152 // 1> We have a constant incoming value for the conditional branch; 153 // 2> We have "Cond" as the incoming value for the unconditional branch; 154 // 3> SI2->getCondition() and Cond have same operands. 155 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition()); 156 if (!Ci2) return false; 157 if (!(Cond->getOperand(0) == Ci2->getOperand(0) && 158 Cond->getOperand(1) == Ci2->getOperand(1)) && 159 !(Cond->getOperand(0) == Ci2->getOperand(1) && 160 Cond->getOperand(1) == Ci2->getOperand(0))) 161 return false; 162 163 BasicBlock *SI1BB = SI1->getParent(); 164 BasicBlock *SI2BB = SI2->getParent(); 165 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 166 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 167 if (SI1Succs.count(*I)) 168 for (BasicBlock::iterator BBI = (*I)->begin(); 169 isa<PHINode>(BBI); ++BBI) { 170 PHINode *PN = cast<PHINode>(BBI); 171 if (PN->getIncomingValueForBlock(SI1BB) != Cond || 172 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB))) 173 return false; 174 PhiNodes.push_back(PN); 175 } 176 return true; 177} 178 179/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 180/// now be entries in it from the 'NewPred' block. The values that will be 181/// flowing into the PHI nodes will be the same as those coming in from 182/// ExistPred, an existing predecessor of Succ. 183static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 184 BasicBlock *ExistPred) { 185 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 186 187 PHINode *PN; 188 for (BasicBlock::iterator I = Succ->begin(); 189 (PN = dyn_cast<PHINode>(I)); ++I) 190 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 191} 192 193 194/// GetIfCondition - Given a basic block (BB) with two predecessors (and at 195/// least one PHI node in it), check to see if the merge at this block is due 196/// to an "if condition". If so, return the boolean condition that determines 197/// which entry into BB will be taken. Also, return by references the block 198/// that will be entered from if the condition is true, and the block that will 199/// be entered if the condition is false. 200/// 201/// This does no checking to see if the true/false blocks have large or unsavory 202/// instructions in them. 203static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 204 BasicBlock *&IfFalse) { 205 PHINode *SomePHI = cast<PHINode>(BB->begin()); 206 assert(SomePHI->getNumIncomingValues() == 2 && 207 "Function can only handle blocks with 2 predecessors!"); 208 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); 209 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); 210 211 // We can only handle branches. Other control flow will be lowered to 212 // branches if possible anyway. 213 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 214 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 215 if (Pred1Br == 0 || Pred2Br == 0) 216 return 0; 217 218 // Eliminate code duplication by ensuring that Pred1Br is conditional if 219 // either are. 220 if (Pred2Br->isConditional()) { 221 // If both branches are conditional, we don't have an "if statement". In 222 // reality, we could transform this case, but since the condition will be 223 // required anyway, we stand no chance of eliminating it, so the xform is 224 // probably not profitable. 225 if (Pred1Br->isConditional()) 226 return 0; 227 228 std::swap(Pred1, Pred2); 229 std::swap(Pred1Br, Pred2Br); 230 } 231 232 if (Pred1Br->isConditional()) { 233 // The only thing we have to watch out for here is to make sure that Pred2 234 // doesn't have incoming edges from other blocks. If it does, the condition 235 // doesn't dominate BB. 236 if (Pred2->getSinglePredecessor() == 0) 237 return 0; 238 239 // If we found a conditional branch predecessor, make sure that it branches 240 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 241 if (Pred1Br->getSuccessor(0) == BB && 242 Pred1Br->getSuccessor(1) == Pred2) { 243 IfTrue = Pred1; 244 IfFalse = Pred2; 245 } else if (Pred1Br->getSuccessor(0) == Pred2 && 246 Pred1Br->getSuccessor(1) == BB) { 247 IfTrue = Pred2; 248 IfFalse = Pred1; 249 } else { 250 // We know that one arm of the conditional goes to BB, so the other must 251 // go somewhere unrelated, and this must not be an "if statement". 252 return 0; 253 } 254 255 return Pred1Br->getCondition(); 256 } 257 258 // Ok, if we got here, both predecessors end with an unconditional branch to 259 // BB. Don't panic! If both blocks only have a single (identical) 260 // predecessor, and THAT is a conditional branch, then we're all ok! 261 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 262 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) 263 return 0; 264 265 // Otherwise, if this is a conditional branch, then we can use it! 266 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 267 if (BI == 0) return 0; 268 269 assert(BI->isConditional() && "Two successors but not conditional?"); 270 if (BI->getSuccessor(0) == Pred1) { 271 IfTrue = Pred1; 272 IfFalse = Pred2; 273 } else { 274 IfTrue = Pred2; 275 IfFalse = Pred1; 276 } 277 return BI->getCondition(); 278} 279 280/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the 281/// given instruction, which is assumed to be safe to speculate. 1 means 282/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. 283static unsigned ComputeSpeculationCost(const User *I) { 284 assert(isSafeToSpeculativelyExecute(I) && 285 "Instruction is not safe to speculatively execute!"); 286 switch (Operator::getOpcode(I)) { 287 default: 288 // In doubt, be conservative. 289 return UINT_MAX; 290 case Instruction::GetElementPtr: 291 // GEPs are cheap if all indices are constant. 292 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 293 return UINT_MAX; 294 return 1; 295 case Instruction::Load: 296 case Instruction::Add: 297 case Instruction::Sub: 298 case Instruction::And: 299 case Instruction::Or: 300 case Instruction::Xor: 301 case Instruction::Shl: 302 case Instruction::LShr: 303 case Instruction::AShr: 304 case Instruction::ICmp: 305 case Instruction::Trunc: 306 case Instruction::ZExt: 307 case Instruction::SExt: 308 return 1; // These are all cheap. 309 310 case Instruction::Call: 311 case Instruction::Select: 312 return 2; 313 } 314} 315 316/// DominatesMergePoint - If we have a merge point of an "if condition" as 317/// accepted above, return true if the specified value dominates the block. We 318/// don't handle the true generality of domination here, just a special case 319/// which works well enough for us. 320/// 321/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 322/// see if V (which must be an instruction) and its recursive operands 323/// that do not dominate BB have a combined cost lower than CostRemaining and 324/// are non-trapping. If both are true, the instruction is inserted into the 325/// set and true is returned. 326/// 327/// The cost for most non-trapping instructions is defined as 1 except for 328/// Select whose cost is 2. 329/// 330/// After this function returns, CostRemaining is decreased by the cost of 331/// V plus its non-dominating operands. If that cost is greater than 332/// CostRemaining, false is returned and CostRemaining is undefined. 333static bool DominatesMergePoint(Value *V, BasicBlock *BB, 334 SmallPtrSet<Instruction*, 4> *AggressiveInsts, 335 unsigned &CostRemaining) { 336 Instruction *I = dyn_cast<Instruction>(V); 337 if (!I) { 338 // Non-instructions all dominate instructions, but not all constantexprs 339 // can be executed unconditionally. 340 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 341 if (C->canTrap()) 342 return false; 343 return true; 344 } 345 BasicBlock *PBB = I->getParent(); 346 347 // We don't want to allow weird loops that might have the "if condition" in 348 // the bottom of this block. 349 if (PBB == BB) return false; 350 351 // If this instruction is defined in a block that contains an unconditional 352 // branch to BB, then it must be in the 'conditional' part of the "if 353 // statement". If not, it definitely dominates the region. 354 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()); 355 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB) 356 return true; 357 358 // If we aren't allowing aggressive promotion anymore, then don't consider 359 // instructions in the 'if region'. 360 if (AggressiveInsts == 0) return false; 361 362 // If we have seen this instruction before, don't count it again. 363 if (AggressiveInsts->count(I)) return true; 364 365 // Okay, it looks like the instruction IS in the "condition". Check to 366 // see if it's a cheap instruction to unconditionally compute, and if it 367 // only uses stuff defined outside of the condition. If so, hoist it out. 368 if (!isSafeToSpeculativelyExecute(I)) 369 return false; 370 371 unsigned Cost = ComputeSpeculationCost(I); 372 373 if (Cost > CostRemaining) 374 return false; 375 376 CostRemaining -= Cost; 377 378 // Okay, we can only really hoist these out if their operands do 379 // not take us over the cost threshold. 380 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 381 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining)) 382 return false; 383 // Okay, it's safe to do this! Remember this instruction. 384 AggressiveInsts->insert(I); 385 return true; 386} 387 388/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr 389/// and PointerNullValue. Return NULL if value is not a constant int. 390static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { 391 // Normal constant int. 392 ConstantInt *CI = dyn_cast<ConstantInt>(V); 393 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) 394 return CI; 395 396 // This is some kind of pointer constant. Turn it into a pointer-sized 397 // ConstantInt if possible. 398 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); 399 400 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). 401 if (isa<ConstantPointerNull>(V)) 402 return ConstantInt::get(PtrTy, 0); 403 404 // IntToPtr const int. 405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 406 if (CE->getOpcode() == Instruction::IntToPtr) 407 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { 408 // The constant is very likely to have the right type already. 409 if (CI->getType() == PtrTy) 410 return CI; 411 else 412 return cast<ConstantInt> 413 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); 414 } 415 return 0; 416} 417 418/// GatherConstantCompares - Given a potentially 'or'd or 'and'd together 419/// collection of icmp eq/ne instructions that compare a value against a 420/// constant, return the value being compared, and stick the constant into the 421/// Values vector. 422static Value * 423GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, 424 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { 425 Instruction *I = dyn_cast<Instruction>(V); 426 if (I == 0) return 0; 427 428 // If this is an icmp against a constant, handle this as one of the cases. 429 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { 430 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { 431 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { 432 UsedICmps++; 433 Vals.push_back(C); 434 return I->getOperand(0); 435 } 436 437 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to 438 // the set. 439 ConstantRange Span = 440 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); 441 442 // If this is an and/!= check then we want to optimize "x ugt 2" into 443 // x != 0 && x != 1. 444 if (!isEQ) 445 Span = Span.inverse(); 446 447 // If there are a ton of values, we don't want to make a ginormous switch. 448 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) 449 return 0; 450 451 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) 452 Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); 453 UsedICmps++; 454 return I->getOperand(0); 455 } 456 return 0; 457 } 458 459 // Otherwise, we can only handle an | or &, depending on isEQ. 460 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) 461 return 0; 462 463 unsigned NumValsBeforeLHS = Vals.size(); 464 unsigned UsedICmpsBeforeLHS = UsedICmps; 465 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, 466 isEQ, UsedICmps)) { 467 unsigned NumVals = Vals.size(); 468 unsigned UsedICmpsBeforeRHS = UsedICmps; 469 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 470 isEQ, UsedICmps)) { 471 if (LHS == RHS) 472 return LHS; 473 Vals.resize(NumVals); 474 UsedICmps = UsedICmpsBeforeRHS; 475 } 476 477 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet, 478 // set it and return success. 479 if (Extra == 0 || Extra == I->getOperand(1)) { 480 Extra = I->getOperand(1); 481 return LHS; 482 } 483 484 Vals.resize(NumValsBeforeLHS); 485 UsedICmps = UsedICmpsBeforeLHS; 486 return 0; 487 } 488 489 // If the LHS can't be folded in, but Extra is available and RHS can, try to 490 // use LHS as Extra. 491 if (Extra == 0 || Extra == I->getOperand(0)) { 492 Value *OldExtra = Extra; 493 Extra = I->getOperand(0); 494 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD, 495 isEQ, UsedICmps)) 496 return RHS; 497 assert(Vals.size() == NumValsBeforeLHS); 498 Extra = OldExtra; 499 } 500 501 return 0; 502} 503 504static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 505 Instruction *Cond = 0; 506 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 507 Cond = dyn_cast<Instruction>(SI->getCondition()); 508 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 509 if (BI->isConditional()) 510 Cond = dyn_cast<Instruction>(BI->getCondition()); 511 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) { 512 Cond = dyn_cast<Instruction>(IBI->getAddress()); 513 } 514 515 TI->eraseFromParent(); 516 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 517} 518 519/// isValueEqualityComparison - Return true if the specified terminator checks 520/// to see if a value is equal to constant integer value. 521Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { 522 Value *CV = 0; 523 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 524 // Do not permit merging of large switch instructions into their 525 // predecessors unless there is only one predecessor. 526 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), 527 pred_end(SI->getParent())) <= 128) 528 CV = SI->getCondition(); 529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 530 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 531 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 532 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 533 ICI->getPredicate() == ICmpInst::ICMP_NE) && 534 GetConstantInt(ICI->getOperand(1), TD)) 535 CV = ICI->getOperand(0); 536 537 // Unwrap any lossless ptrtoint cast. 538 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) 539 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) 540 CV = PTII->getOperand(0); 541 return CV; 542} 543 544/// GetValueEqualityComparisonCases - Given a value comparison instruction, 545/// decode all of the 'cases' that it represents and return the 'default' block. 546BasicBlock *SimplifyCFGOpt:: 547GetValueEqualityComparisonCases(TerminatorInst *TI, 548 std::vector<ValueEqualityComparisonCase> 549 &Cases) { 550 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 551 Cases.reserve(SI->getNumCases()); 552 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) 553 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(), 554 i.getCaseSuccessor())); 555 return SI->getDefaultDest(); 556 } 557 558 BranchInst *BI = cast<BranchInst>(TI); 559 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 560 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE); 561 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1), 562 TD), 563 Succ)); 564 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 565} 566 567 568/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 569/// in the list that match the specified block. 570static void EliminateBlockCases(BasicBlock *BB, 571 std::vector<ValueEqualityComparisonCase> &Cases) { 572 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); 573} 574 575/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 576/// well. 577static bool 578ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, 579 std::vector<ValueEqualityComparisonCase > &C2) { 580 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; 581 582 // Make V1 be smaller than V2. 583 if (V1->size() > V2->size()) 584 std::swap(V1, V2); 585 586 if (V1->size() == 0) return false; 587 if (V1->size() == 1) { 588 // Just scan V2. 589 ConstantInt *TheVal = (*V1)[0].Value; 590 for (unsigned i = 0, e = V2->size(); i != e; ++i) 591 if (TheVal == (*V2)[i].Value) 592 return true; 593 } 594 595 // Otherwise, just sort both lists and compare element by element. 596 array_pod_sort(V1->begin(), V1->end()); 597 array_pod_sort(V2->begin(), V2->end()); 598 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 599 while (i1 != e1 && i2 != e2) { 600 if ((*V1)[i1].Value == (*V2)[i2].Value) 601 return true; 602 if ((*V1)[i1].Value < (*V2)[i2].Value) 603 ++i1; 604 else 605 ++i2; 606 } 607 return false; 608} 609 610/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 611/// terminator instruction and its block is known to only have a single 612/// predecessor block, check to see if that predecessor is also a value 613/// comparison with the same value, and if that comparison determines the 614/// outcome of this comparison. If so, simplify TI. This does a very limited 615/// form of jump threading. 616bool SimplifyCFGOpt:: 617SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 618 BasicBlock *Pred, 619 IRBuilder<> &Builder) { 620 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 621 if (!PredVal) return false; // Not a value comparison in predecessor. 622 623 Value *ThisVal = isValueEqualityComparison(TI); 624 assert(ThisVal && "This isn't a value comparison!!"); 625 if (ThisVal != PredVal) return false; // Different predicates. 626 627 // TODO: Preserve branch weight metadata, similarly to how 628 // FoldValueComparisonIntoPredecessors preserves it. 629 630 // Find out information about when control will move from Pred to TI's block. 631 std::vector<ValueEqualityComparisonCase> PredCases; 632 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 633 PredCases); 634 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 635 636 // Find information about how control leaves this block. 637 std::vector<ValueEqualityComparisonCase> ThisCases; 638 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 639 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 640 641 // If TI's block is the default block from Pred's comparison, potentially 642 // simplify TI based on this knowledge. 643 if (PredDef == TI->getParent()) { 644 // If we are here, we know that the value is none of those cases listed in 645 // PredCases. If there are any cases in ThisCases that are in PredCases, we 646 // can simplify TI. 647 if (!ValuesOverlap(PredCases, ThisCases)) 648 return false; 649 650 if (isa<BranchInst>(TI)) { 651 // Okay, one of the successors of this condbr is dead. Convert it to a 652 // uncond br. 653 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 654 // Insert the new branch. 655 Instruction *NI = Builder.CreateBr(ThisDef); 656 (void) NI; 657 658 // Remove PHI node entries for the dead edge. 659 ThisCases[0].Dest->removePredecessor(TI->getParent()); 660 661 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 662 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 663 664 EraseTerminatorInstAndDCECond(TI); 665 return true; 666 } 667 668 SwitchInst *SI = cast<SwitchInst>(TI); 669 // Okay, TI has cases that are statically dead, prune them away. 670 SmallPtrSet<Constant*, 16> DeadCases; 671 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 672 DeadCases.insert(PredCases[i].Value); 673 674 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 675 << "Through successor TI: " << *TI); 676 677 // Collect branch weights into a vector. 678 SmallVector<uint32_t, 8> Weights; 679 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 680 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); 681 if (HasWeight) 682 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 683 ++MD_i) { 684 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 685 assert(CI); 686 Weights.push_back(CI->getValue().getZExtValue()); 687 } 688 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { 689 --i; 690 if (DeadCases.count(i.getCaseValue())) { 691 if (HasWeight) { 692 std::swap(Weights[i.getCaseIndex()+1], Weights.back()); 693 Weights.pop_back(); 694 } 695 i.getCaseSuccessor()->removePredecessor(TI->getParent()); 696 SI->removeCase(i); 697 } 698 } 699 if (HasWeight && Weights.size() >= 2) 700 SI->setMetadata(LLVMContext::MD_prof, 701 MDBuilder(SI->getParent()->getContext()). 702 createBranchWeights(Weights)); 703 704 DEBUG(dbgs() << "Leaving: " << *TI << "\n"); 705 return true; 706 } 707 708 // Otherwise, TI's block must correspond to some matched value. Find out 709 // which value (or set of values) this is. 710 ConstantInt *TIV = 0; 711 BasicBlock *TIBB = TI->getParent(); 712 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 713 if (PredCases[i].Dest == TIBB) { 714 if (TIV != 0) 715 return false; // Cannot handle multiple values coming to this block. 716 TIV = PredCases[i].Value; 717 } 718 assert(TIV && "No edge from pred to succ?"); 719 720 // Okay, we found the one constant that our value can be if we get into TI's 721 // BB. Find out which successor will unconditionally be branched to. 722 BasicBlock *TheRealDest = 0; 723 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 724 if (ThisCases[i].Value == TIV) { 725 TheRealDest = ThisCases[i].Dest; 726 break; 727 } 728 729 // If not handled by any explicit cases, it is handled by the default case. 730 if (TheRealDest == 0) TheRealDest = ThisDef; 731 732 // Remove PHI node entries for dead edges. 733 BasicBlock *CheckEdge = TheRealDest; 734 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 735 if (*SI != CheckEdge) 736 (*SI)->removePredecessor(TIBB); 737 else 738 CheckEdge = 0; 739 740 // Insert the new branch. 741 Instruction *NI = Builder.CreateBr(TheRealDest); 742 (void) NI; 743 744 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() 745 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); 746 747 EraseTerminatorInstAndDCECond(TI); 748 return true; 749} 750 751namespace { 752 /// ConstantIntOrdering - This class implements a stable ordering of constant 753 /// integers that does not depend on their address. This is important for 754 /// applications that sort ConstantInt's to ensure uniqueness. 755 struct ConstantIntOrdering { 756 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 757 return LHS->getValue().ult(RHS->getValue()); 758 } 759 }; 760} 761 762static int ConstantIntSortPredicate(const void *P1, const void *P2) { 763 const ConstantInt *LHS = *(const ConstantInt*const*)P1; 764 const ConstantInt *RHS = *(const ConstantInt*const*)P2; 765 if (LHS->getValue().ult(RHS->getValue())) 766 return 1; 767 if (LHS->getValue() == RHS->getValue()) 768 return 0; 769 return -1; 770} 771 772static inline bool HasBranchWeights(const Instruction* I) { 773 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); 774 if (ProfMD && ProfMD->getOperand(0)) 775 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) 776 return MDS->getString().equals("branch_weights"); 777 778 return false; 779} 780 781/// Get Weights of a given TerminatorInst, the default weight is at the front 782/// of the vector. If TI is a conditional eq, we need to swap the branch-weight 783/// metadata. 784static void GetBranchWeights(TerminatorInst *TI, 785 SmallVectorImpl<uint64_t> &Weights) { 786 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); 787 assert(MD); 788 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { 789 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); 790 assert(CI); 791 Weights.push_back(CI->getValue().getZExtValue()); 792 } 793 794 // If TI is a conditional eq, the default case is the false case, 795 // and the corresponding branch-weight data is at index 2. We swap the 796 // default weight to be the first entry. 797 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { 798 assert(Weights.size() == 2); 799 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 801 std::swap(Weights.front(), Weights.back()); 802 } 803} 804 805/// Sees if any of the weights are too big for a uint32_t, and halves all the 806/// weights if any are. 807static void FitWeights(MutableArrayRef<uint64_t> Weights) { 808 bool Halve = false; 809 for (unsigned i = 0; i < Weights.size(); ++i) 810 if (Weights[i] > UINT_MAX) { 811 Halve = true; 812 break; 813 } 814 815 if (! Halve) 816 return; 817 818 for (unsigned i = 0; i < Weights.size(); ++i) 819 Weights[i] /= 2; 820} 821 822/// FoldValueComparisonIntoPredecessors - The specified terminator is a value 823/// equality comparison instruction (either a switch or a branch on "X == c"). 824/// See if any of the predecessors of the terminator block are value comparisons 825/// on the same value. If so, and if safe to do so, fold them together. 826bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, 827 IRBuilder<> &Builder) { 828 BasicBlock *BB = TI->getParent(); 829 Value *CV = isValueEqualityComparison(TI); // CondVal 830 assert(CV && "Not a comparison?"); 831 bool Changed = false; 832 833 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 834 while (!Preds.empty()) { 835 BasicBlock *Pred = Preds.pop_back_val(); 836 837 // See if the predecessor is a comparison with the same value. 838 TerminatorInst *PTI = Pred->getTerminator(); 839 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 840 841 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 842 // Figure out which 'cases' to copy from SI to PSI. 843 std::vector<ValueEqualityComparisonCase> BBCases; 844 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 845 846 std::vector<ValueEqualityComparisonCase> PredCases; 847 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 848 849 // Based on whether the default edge from PTI goes to BB or not, fill in 850 // PredCases and PredDefault with the new switch cases we would like to 851 // build. 852 SmallVector<BasicBlock*, 8> NewSuccessors; 853 854 // Update the branch weight metadata along the way 855 SmallVector<uint64_t, 8> Weights; 856 bool PredHasWeights = HasBranchWeights(PTI); 857 bool SuccHasWeights = HasBranchWeights(TI); 858 859 if (PredHasWeights) { 860 GetBranchWeights(PTI, Weights); 861 // branch-weight metadata is inconsistent here. 862 if (Weights.size() != 1 + PredCases.size()) 863 PredHasWeights = SuccHasWeights = false; 864 } else if (SuccHasWeights) 865 // If there are no predecessor weights but there are successor weights, 866 // populate Weights with 1, which will later be scaled to the sum of 867 // successor's weights 868 Weights.assign(1 + PredCases.size(), 1); 869 870 SmallVector<uint64_t, 8> SuccWeights; 871 if (SuccHasWeights) { 872 GetBranchWeights(TI, SuccWeights); 873 // branch-weight metadata is inconsistent here. 874 if (SuccWeights.size() != 1 + BBCases.size()) 875 PredHasWeights = SuccHasWeights = false; 876 } else if (PredHasWeights) 877 SuccWeights.assign(1 + BBCases.size(), 1); 878 879 if (PredDefault == BB) { 880 // If this is the default destination from PTI, only the edges in TI 881 // that don't occur in PTI, or that branch to BB will be activated. 882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 884 if (PredCases[i].Dest != BB) 885 PTIHandled.insert(PredCases[i].Value); 886 else { 887 // The default destination is BB, we don't need explicit targets. 888 std::swap(PredCases[i], PredCases.back()); 889 890 if (PredHasWeights || SuccHasWeights) { 891 // Increase weight for the default case. 892 Weights[0] += Weights[i+1]; 893 std::swap(Weights[i+1], Weights.back()); 894 Weights.pop_back(); 895 } 896 897 PredCases.pop_back(); 898 --i; --e; 899 } 900 901 // Reconstruct the new switch statement we will be building. 902 if (PredDefault != BBDefault) { 903 PredDefault->removePredecessor(Pred); 904 PredDefault = BBDefault; 905 NewSuccessors.push_back(BBDefault); 906 } 907 908 unsigned CasesFromPred = Weights.size(); 909 uint64_t ValidTotalSuccWeight = 0; 910 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 911 if (!PTIHandled.count(BBCases[i].Value) && 912 BBCases[i].Dest != BBDefault) { 913 PredCases.push_back(BBCases[i]); 914 NewSuccessors.push_back(BBCases[i].Dest); 915 if (SuccHasWeights || PredHasWeights) { 916 // The default weight is at index 0, so weight for the ith case 917 // should be at index i+1. Scale the cases from successor by 918 // PredDefaultWeight (Weights[0]). 919 Weights.push_back(Weights[0] * SuccWeights[i+1]); 920 ValidTotalSuccWeight += SuccWeights[i+1]; 921 } 922 } 923 924 if (SuccHasWeights || PredHasWeights) { 925 ValidTotalSuccWeight += SuccWeights[0]; 926 // Scale the cases from predecessor by ValidTotalSuccWeight. 927 for (unsigned i = 1; i < CasesFromPred; ++i) 928 Weights[i] *= ValidTotalSuccWeight; 929 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). 930 Weights[0] *= SuccWeights[0]; 931 } 932 } else { 933 // If this is not the default destination from PSI, only the edges 934 // in SI that occur in PSI with a destination of BB will be 935 // activated. 936 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 937 std::map<ConstantInt*, uint64_t> WeightsForHandled; 938 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 939 if (PredCases[i].Dest == BB) { 940 PTIHandled.insert(PredCases[i].Value); 941 942 if (PredHasWeights || SuccHasWeights) { 943 WeightsForHandled[PredCases[i].Value] = Weights[i+1]; 944 std::swap(Weights[i+1], Weights.back()); 945 Weights.pop_back(); 946 } 947 948 std::swap(PredCases[i], PredCases.back()); 949 PredCases.pop_back(); 950 --i; --e; 951 } 952 953 // Okay, now we know which constants were sent to BB from the 954 // predecessor. Figure out where they will all go now. 955 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 956 if (PTIHandled.count(BBCases[i].Value)) { 957 // If this is one we are capable of getting... 958 if (PredHasWeights || SuccHasWeights) 959 Weights.push_back(WeightsForHandled[BBCases[i].Value]); 960 PredCases.push_back(BBCases[i]); 961 NewSuccessors.push_back(BBCases[i].Dest); 962 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of 963 } 964 965 // If there are any constants vectored to BB that TI doesn't handle, 966 // they must go to the default destination of TI. 967 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 968 PTIHandled.begin(), 969 E = PTIHandled.end(); I != E; ++I) { 970 if (PredHasWeights || SuccHasWeights) 971 Weights.push_back(WeightsForHandled[*I]); 972 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); 973 NewSuccessors.push_back(BBDefault); 974 } 975 } 976 977 // Okay, at this point, we know which new successor Pred will get. Make 978 // sure we update the number of entries in the PHI nodes for these 979 // successors. 980 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 981 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 982 983 Builder.SetInsertPoint(PTI); 984 // Convert pointer to int before we switch. 985 if (CV->getType()->isPointerTy()) { 986 assert(TD && "Cannot switch on pointer without DataLayout"); 987 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), 988 "magicptr"); 989 } 990 991 // Now that the successors are updated, create the new Switch instruction. 992 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, 993 PredCases.size()); 994 NewSI->setDebugLoc(PTI->getDebugLoc()); 995 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 996 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); 997 998 if (PredHasWeights || SuccHasWeights) { 999 // Halve the weights if any of them cannot fit in an uint32_t 1000 FitWeights(Weights); 1001 1002 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 1003 1004 NewSI->setMetadata(LLVMContext::MD_prof, 1005 MDBuilder(BB->getContext()). 1006 createBranchWeights(MDWeights)); 1007 } 1008 1009 EraseTerminatorInstAndDCECond(PTI); 1010 1011 // Okay, last check. If BB is still a successor of PSI, then we must 1012 // have an infinite loop case. If so, add an infinitely looping block 1013 // to handle the case to preserve the behavior of the code. 1014 BasicBlock *InfLoopBlock = 0; 1015 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 1016 if (NewSI->getSuccessor(i) == BB) { 1017 if (InfLoopBlock == 0) { 1018 // Insert it at the end of the function, because it's either code, 1019 // or it won't matter if it's hot. :) 1020 InfLoopBlock = BasicBlock::Create(BB->getContext(), 1021 "infloop", BB->getParent()); 1022 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1023 } 1024 NewSI->setSuccessor(i, InfLoopBlock); 1025 } 1026 1027 Changed = true; 1028 } 1029 } 1030 return Changed; 1031} 1032 1033// isSafeToHoistInvoke - If we would need to insert a select that uses the 1034// value of this invoke (comments in HoistThenElseCodeToIf explain why we 1035// would need to do this), we can't hoist the invoke, as there is nowhere 1036// to put the select in this case. 1037static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 1038 Instruction *I1, Instruction *I2) { 1039 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1040 PHINode *PN; 1041 for (BasicBlock::iterator BBI = SI->begin(); 1042 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1043 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1044 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1045 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 1046 return false; 1047 } 1048 } 1049 } 1050 return true; 1051} 1052 1053/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 1054/// BB2, hoist any common code in the two blocks up into the branch block. The 1055/// caller of this function guarantees that BI's block dominates BB1 and BB2. 1056static bool HoistThenElseCodeToIf(BranchInst *BI) { 1057 // This does very trivial matching, with limited scanning, to find identical 1058 // instructions in the two blocks. In particular, we don't want to get into 1059 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 1060 // such, we currently just scan for obviously identical instructions in an 1061 // identical order. 1062 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 1063 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 1064 1065 BasicBlock::iterator BB1_Itr = BB1->begin(); 1066 BasicBlock::iterator BB2_Itr = BB2->begin(); 1067 1068 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 1069 // Skip debug info if it is not identical. 1070 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1071 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1072 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1073 while (isa<DbgInfoIntrinsic>(I1)) 1074 I1 = BB1_Itr++; 1075 while (isa<DbgInfoIntrinsic>(I2)) 1076 I2 = BB2_Itr++; 1077 } 1078 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) || 1079 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 1080 return false; 1081 1082 // If we get here, we can hoist at least one instruction. 1083 BasicBlock *BIParent = BI->getParent(); 1084 1085 do { 1086 // If we are hoisting the terminator instruction, don't move one (making a 1087 // broken BB), instead clone it, and remove BI. 1088 if (isa<TerminatorInst>(I1)) 1089 goto HoistTerminator; 1090 1091 // For a normal instruction, we just move one to right before the branch, 1092 // then replace all uses of the other with the first. Finally, we remove 1093 // the now redundant second instruction. 1094 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 1095 if (!I2->use_empty()) 1096 I2->replaceAllUsesWith(I1); 1097 I1->intersectOptionalDataWith(I2); 1098 I2->eraseFromParent(); 1099 1100 I1 = BB1_Itr++; 1101 I2 = BB2_Itr++; 1102 // Skip debug info if it is not identical. 1103 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1); 1104 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2); 1105 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) { 1106 while (isa<DbgInfoIntrinsic>(I1)) 1107 I1 = BB1_Itr++; 1108 while (isa<DbgInfoIntrinsic>(I2)) 1109 I2 = BB2_Itr++; 1110 } 1111 } while (I1->isIdenticalToWhenDefined(I2)); 1112 1113 return true; 1114 1115HoistTerminator: 1116 // It may not be possible to hoist an invoke. 1117 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 1118 return true; 1119 1120 // Okay, it is safe to hoist the terminator. 1121 Instruction *NT = I1->clone(); 1122 BIParent->getInstList().insert(BI, NT); 1123 if (!NT->getType()->isVoidTy()) { 1124 I1->replaceAllUsesWith(NT); 1125 I2->replaceAllUsesWith(NT); 1126 NT->takeName(I1); 1127 } 1128 1129 IRBuilder<true, NoFolder> Builder(NT); 1130 // Hoisting one of the terminators from our successor is a great thing. 1131 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 1132 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 1133 // nodes, so we insert select instruction to compute the final result. 1134 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 1135 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 1136 PHINode *PN; 1137 for (BasicBlock::iterator BBI = SI->begin(); 1138 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1139 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1140 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1141 if (BB1V == BB2V) continue; 1142 1143 // These values do not agree. Insert a select instruction before NT 1144 // that determines the right value. 1145 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 1146 if (SI == 0) 1147 SI = cast<SelectInst> 1148 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, 1149 BB1V->getName()+"."+BB2V->getName())); 1150 1151 // Make the PHI node use the select for all incoming values for BB1/BB2 1152 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1153 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 1154 PN->setIncomingValue(i, SI); 1155 } 1156 } 1157 1158 // Update any PHI nodes in our new successors. 1159 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 1160 AddPredecessorToBlock(*SI, BIParent, BB1); 1161 1162 EraseTerminatorInstAndDCECond(BI); 1163 return true; 1164} 1165 1166/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, 1167/// check whether BBEnd has only two predecessors and the other predecessor 1168/// ends with an unconditional branch. If it is true, sink any common code 1169/// in the two predecessors to BBEnd. 1170static bool SinkThenElseCodeToEnd(BranchInst *BI1) { 1171 assert(BI1->isUnconditional()); 1172 BasicBlock *BB1 = BI1->getParent(); 1173 BasicBlock *BBEnd = BI1->getSuccessor(0); 1174 1175 // Check that BBEnd has two predecessors and the other predecessor ends with 1176 // an unconditional branch. 1177 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); 1178 BasicBlock *Pred0 = *PI++; 1179 if (PI == PE) // Only one predecessor. 1180 return false; 1181 BasicBlock *Pred1 = *PI++; 1182 if (PI != PE) // More than two predecessors. 1183 return false; 1184 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; 1185 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); 1186 if (!BI2 || !BI2->isUnconditional()) 1187 return false; 1188 1189 // Gather the PHI nodes in BBEnd. 1190 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; 1191 Instruction *FirstNonPhiInBBEnd = 0; 1192 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); 1193 I != E; ++I) { 1194 if (PHINode *PN = dyn_cast<PHINode>(I)) { 1195 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1196 Value *BB2V = PN->getIncomingValueForBlock(BB2); 1197 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); 1198 } else { 1199 FirstNonPhiInBBEnd = &*I; 1200 break; 1201 } 1202 } 1203 if (!FirstNonPhiInBBEnd) 1204 return false; 1205 1206 1207 // This does very trivial matching, with limited scanning, to find identical 1208 // instructions in the two blocks. We scan backward for obviously identical 1209 // instructions in an identical order. 1210 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), 1211 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), 1212 RE2 = BB2->getInstList().rend(); 1213 // Skip debug info. 1214 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1215 if (RI1 == RE1) 1216 return false; 1217 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1218 if (RI2 == RE2) 1219 return false; 1220 // Skip the unconditional branches. 1221 ++RI1; 1222 ++RI2; 1223 1224 bool Changed = false; 1225 while (RI1 != RE1 && RI2 != RE2) { 1226 // Skip debug info. 1227 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; 1228 if (RI1 == RE1) 1229 return Changed; 1230 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; 1231 if (RI2 == RE2) 1232 return Changed; 1233 1234 Instruction *I1 = &*RI1, *I2 = &*RI2; 1235 // I1 and I2 should have a single use in the same PHI node, and they 1236 // perform the same operation. 1237 // Cannot move control-flow-involving, volatile loads, vaarg, etc. 1238 if (isa<PHINode>(I1) || isa<PHINode>(I2) || 1239 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || 1240 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || 1241 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || 1242 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || 1243 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || 1244 !I1->hasOneUse() || !I2->hasOneUse() || 1245 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || 1246 MapValueFromBB1ToBB2[I1].first != I2) 1247 return Changed; 1248 1249 // Check whether we should swap the operands of ICmpInst. 1250 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); 1251 bool SwapOpnds = false; 1252 if (ICmp1 && ICmp2 && 1253 ICmp1->getOperand(0) != ICmp2->getOperand(0) && 1254 ICmp1->getOperand(1) != ICmp2->getOperand(1) && 1255 (ICmp1->getOperand(0) == ICmp2->getOperand(1) || 1256 ICmp1->getOperand(1) == ICmp2->getOperand(0))) { 1257 ICmp2->swapOperands(); 1258 SwapOpnds = true; 1259 } 1260 if (!I1->isSameOperationAs(I2)) { 1261 if (SwapOpnds) 1262 ICmp2->swapOperands(); 1263 return Changed; 1264 } 1265 1266 // The operands should be either the same or they need to be generated 1267 // with a PHI node after sinking. We only handle the case where there is 1268 // a single pair of different operands. 1269 Value *DifferentOp1 = 0, *DifferentOp2 = 0; 1270 unsigned Op1Idx = 0; 1271 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { 1272 if (I1->getOperand(I) == I2->getOperand(I)) 1273 continue; 1274 // Early exit if we have more-than one pair of different operands or 1275 // the different operand is already in MapValueFromBB1ToBB2. 1276 // Early exit if we need a PHI node to replace a constant. 1277 if (DifferentOp1 || 1278 MapValueFromBB1ToBB2.find(I1->getOperand(I)) != 1279 MapValueFromBB1ToBB2.end() || 1280 isa<Constant>(I1->getOperand(I)) || 1281 isa<Constant>(I2->getOperand(I))) { 1282 // If we can't sink the instructions, undo the swapping. 1283 if (SwapOpnds) 1284 ICmp2->swapOperands(); 1285 return Changed; 1286 } 1287 DifferentOp1 = I1->getOperand(I); 1288 Op1Idx = I; 1289 DifferentOp2 = I2->getOperand(I); 1290 } 1291 1292 // We insert the pair of different operands to MapValueFromBB1ToBB2 and 1293 // remove (I1, I2) from MapValueFromBB1ToBB2. 1294 if (DifferentOp1) { 1295 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, 1296 DifferentOp1->getName() + ".sink", 1297 BBEnd->begin()); 1298 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); 1299 // I1 should use NewPN instead of DifferentOp1. 1300 I1->setOperand(Op1Idx, NewPN); 1301 NewPN->addIncoming(DifferentOp1, BB1); 1302 NewPN->addIncoming(DifferentOp2, BB2); 1303 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); 1304 } 1305 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; 1306 MapValueFromBB1ToBB2.erase(I1); 1307 1308 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); 1309 DEBUG(dbgs() << " " << *I2 << "\n";); 1310 // We need to update RE1 and RE2 if we are going to sink the first 1311 // instruction in the basic block down. 1312 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); 1313 // Sink the instruction. 1314 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); 1315 if (!OldPN->use_empty()) 1316 OldPN->replaceAllUsesWith(I1); 1317 OldPN->eraseFromParent(); 1318 1319 if (!I2->use_empty()) 1320 I2->replaceAllUsesWith(I1); 1321 I1->intersectOptionalDataWith(I2); 1322 I2->eraseFromParent(); 1323 1324 if (UpdateRE1) 1325 RE1 = BB1->getInstList().rend(); 1326 if (UpdateRE2) 1327 RE2 = BB2->getInstList().rend(); 1328 FirstNonPhiInBBEnd = I1; 1329 NumSinkCommons++; 1330 Changed = true; 1331 } 1332 return Changed; 1333} 1334 1335/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 1336/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 1337/// (for now, restricted to a single instruction that's side effect free) from 1338/// the BB1 into the branch block to speculatively execute it. 1339/// 1340/// Turn 1341/// BB: 1342/// %t1 = icmp 1343/// br i1 %t1, label %BB1, label %BB2 1344/// BB1: 1345/// %t3 = add %t2, c 1346/// br label BB2 1347/// BB2: 1348/// => 1349/// BB: 1350/// %t1 = icmp 1351/// %t4 = add %t2, c 1352/// %t3 = select i1 %t1, %t2, %t3 1353static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 1354 // Only speculatively execution a single instruction (not counting the 1355 // terminator) for now. 1356 Instruction *HInst = NULL; 1357 Instruction *Term = BB1->getTerminator(); 1358 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 1359 BBI != BBE; ++BBI) { 1360 Instruction *I = BBI; 1361 // Skip debug info. 1362 if (isa<DbgInfoIntrinsic>(I)) continue; 1363 if (I == Term) break; 1364 1365 if (HInst) 1366 return false; 1367 HInst = I; 1368 } 1369 1370 BasicBlock *BIParent = BI->getParent(); 1371 1372 // Check the instruction to be hoisted, if there is one. 1373 if (HInst) { 1374 // Don't hoist the instruction if it's unsafe or expensive. 1375 if (!isSafeToSpeculativelyExecute(HInst)) 1376 return false; 1377 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold) 1378 return false; 1379 1380 // Do not hoist the instruction if any of its operands are defined but not 1381 // used in this BB. The transformation will prevent the operand from 1382 // being sunk into the use block. 1383 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1384 i != e; ++i) { 1385 Instruction *OpI = dyn_cast<Instruction>(*i); 1386 if (OpI && OpI->getParent() == BIParent && 1387 !OpI->mayHaveSideEffects() && 1388 !OpI->isUsedInBasicBlock(BIParent)) 1389 return false; 1390 } 1391 } 1392 1393 // Be conservative for now. FP select instruction can often be expensive. 1394 Value *BrCond = BI->getCondition(); 1395 if (isa<FCmpInst>(BrCond)) 1396 return false; 1397 1398 // If BB1 is actually on the false edge of the conditional branch, remember 1399 // to swap the select operands later. 1400 bool Invert = false; 1401 if (BB1 != BI->getSuccessor(0)) { 1402 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1403 Invert = true; 1404 } 1405 1406 // Collect interesting PHIs, and scan for hazards. 1407 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs; 1408 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1409 for (BasicBlock::iterator I = BB2->begin(); 1410 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1411 Value *BB1V = PN->getIncomingValueForBlock(BB1); 1412 Value *BIParentV = PN->getIncomingValueForBlock(BIParent); 1413 1414 // Skip PHIs which are trivial. 1415 if (BB1V == BIParentV) 1416 continue; 1417 1418 // Check for safety. 1419 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) { 1420 // An unfolded ConstantExpr could end up getting expanded into 1421 // Instructions. Don't speculate this and another instruction at 1422 // the same time. 1423 if (HInst) 1424 return false; 1425 if (!isSafeToSpeculativelyExecute(CE)) 1426 return false; 1427 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) 1428 return false; 1429 } 1430 1431 // Ok, we may insert a select for this PHI. 1432 PHIs.insert(std::make_pair(BB1V, BIParentV)); 1433 } 1434 1435 // If there are no PHIs to process, bail early. This helps ensure idempotence 1436 // as well. 1437 if (PHIs.empty()) 1438 return false; 1439 1440 // If we get here, we can hoist the instruction and if-convert. 1441 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); 1442 1443 // Hoist the instruction. 1444 if (HInst) 1445 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst); 1446 1447 // Insert selects and rewrite the PHI operands. 1448 IRBuilder<true, NoFolder> Builder(BI); 1449 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 1450 Value *TrueV = PHIs[i].first; 1451 Value *FalseV = PHIs[i].second; 1452 1453 // Create a select whose true value is the speculatively executed value and 1454 // false value is the previously determined FalseV. 1455 SelectInst *SI; 1456 if (Invert) 1457 SI = cast<SelectInst> 1458 (Builder.CreateSelect(BrCond, FalseV, TrueV, 1459 FalseV->getName() + "." + TrueV->getName())); 1460 else 1461 SI = cast<SelectInst> 1462 (Builder.CreateSelect(BrCond, TrueV, FalseV, 1463 TrueV->getName() + "." + FalseV->getName())); 1464 1465 // Make the PHI node use the select for all incoming values for "then" and 1466 // "if" blocks. 1467 for (BasicBlock::iterator I = BB2->begin(); 1468 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1469 unsigned BB1I = PN->getBasicBlockIndex(BB1); 1470 unsigned BIParentI = PN->getBasicBlockIndex(BIParent); 1471 Value *BB1V = PN->getIncomingValue(BB1I); 1472 Value *BIParentV = PN->getIncomingValue(BIParentI); 1473 if (TrueV == BB1V && FalseV == BIParentV) { 1474 PN->setIncomingValue(BB1I, SI); 1475 PN->setIncomingValue(BIParentI, SI); 1476 } 1477 } 1478 } 1479 1480 ++NumSpeculations; 1481 return true; 1482} 1483 1484/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1485/// across this block. 1486static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1487 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1488 unsigned Size = 0; 1489 1490 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1491 if (isa<DbgInfoIntrinsic>(BBI)) 1492 continue; 1493 if (Size > 10) return false; // Don't clone large BB's. 1494 ++Size; 1495 1496 // We can only support instructions that do not define values that are 1497 // live outside of the current basic block. 1498 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1499 UI != E; ++UI) { 1500 Instruction *U = cast<Instruction>(*UI); 1501 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1502 } 1503 1504 // Looks ok, continue checking. 1505 } 1506 1507 return true; 1508} 1509 1510/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1511/// that is defined in the same block as the branch and if any PHI entries are 1512/// constants, thread edges corresponding to that entry to be branches to their 1513/// ultimate destination. 1514static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { 1515 BasicBlock *BB = BI->getParent(); 1516 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1517 // NOTE: we currently cannot transform this case if the PHI node is used 1518 // outside of the block. 1519 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1520 return false; 1521 1522 // Degenerate case of a single entry PHI. 1523 if (PN->getNumIncomingValues() == 1) { 1524 FoldSingleEntryPHINodes(PN->getParent()); 1525 return true; 1526 } 1527 1528 // Now we know that this block has multiple preds and two succs. 1529 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1530 1531 // Okay, this is a simple enough basic block. See if any phi values are 1532 // constants. 1533 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1534 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); 1535 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; 1536 1537 // Okay, we now know that all edges from PredBB should be revectored to 1538 // branch to RealDest. 1539 BasicBlock *PredBB = PN->getIncomingBlock(i); 1540 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1541 1542 if (RealDest == BB) continue; // Skip self loops. 1543 // Skip if the predecessor's terminator is an indirect branch. 1544 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; 1545 1546 // The dest block might have PHI nodes, other predecessors and other 1547 // difficult cases. Instead of being smart about this, just insert a new 1548 // block that jumps to the destination block, effectively splitting 1549 // the edge we are about to create. 1550 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(), 1551 RealDest->getName()+".critedge", 1552 RealDest->getParent(), RealDest); 1553 BranchInst::Create(RealDest, EdgeBB); 1554 1555 // Update PHI nodes. 1556 AddPredecessorToBlock(RealDest, EdgeBB, BB); 1557 1558 // BB may have instructions that are being threaded over. Clone these 1559 // instructions into EdgeBB. We know that there will be no uses of the 1560 // cloned instructions outside of EdgeBB. 1561 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1562 DenseMap<Value*, Value*> TranslateMap; // Track translated values. 1563 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1564 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1565 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1566 continue; 1567 } 1568 // Clone the instruction. 1569 Instruction *N = BBI->clone(); 1570 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1571 1572 // Update operands due to translation. 1573 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1574 i != e; ++i) { 1575 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i); 1576 if (PI != TranslateMap.end()) 1577 *i = PI->second; 1578 } 1579 1580 // Check for trivial simplification. 1581 if (Value *V = SimplifyInstruction(N, TD)) { 1582 TranslateMap[BBI] = V; 1583 delete N; // Instruction folded away, don't need actual inst 1584 } else { 1585 // Insert the new instruction into its new home. 1586 EdgeBB->getInstList().insert(InsertPt, N); 1587 if (!BBI->use_empty()) 1588 TranslateMap[BBI] = N; 1589 } 1590 } 1591 1592 // Loop over all of the edges from PredBB to BB, changing them to branch 1593 // to EdgeBB instead. 1594 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1595 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1596 if (PredBBTI->getSuccessor(i) == BB) { 1597 BB->removePredecessor(PredBB); 1598 PredBBTI->setSuccessor(i, EdgeBB); 1599 } 1600 1601 // Recurse, simplifying any other constants. 1602 return FoldCondBranchOnPHI(BI, TD) | true; 1603 } 1604 1605 return false; 1606} 1607 1608/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1609/// PHI node, see if we can eliminate it. 1610static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { 1611 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1612 // statement", which has a very simple dominance structure. Basically, we 1613 // are trying to find the condition that is being branched on, which 1614 // subsequently causes this merge to happen. We really want control 1615 // dependence information for this check, but simplifycfg can't keep it up 1616 // to date, and this catches most of the cases we care about anyway. 1617 BasicBlock *BB = PN->getParent(); 1618 BasicBlock *IfTrue, *IfFalse; 1619 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1620 if (!IfCond || 1621 // Don't bother if the branch will be constant folded trivially. 1622 isa<ConstantInt>(IfCond)) 1623 return false; 1624 1625 // Okay, we found that we can merge this two-entry phi node into a select. 1626 // Doing so would require us to fold *all* two entry phi nodes in this block. 1627 // At some point this becomes non-profitable (particularly if the target 1628 // doesn't support cmov's). Only do this transformation if there are two or 1629 // fewer PHI nodes in this block. 1630 unsigned NumPhis = 0; 1631 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1632 if (NumPhis > 2) 1633 return false; 1634 1635 // Loop over the PHI's seeing if we can promote them all to select 1636 // instructions. While we are at it, keep track of the instructions 1637 // that need to be moved to the dominating block. 1638 SmallPtrSet<Instruction*, 4> AggressiveInsts; 1639 unsigned MaxCostVal0 = PHINodeFoldingThreshold, 1640 MaxCostVal1 = PHINodeFoldingThreshold; 1641 1642 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { 1643 PHINode *PN = cast<PHINode>(II++); 1644 if (Value *V = SimplifyInstruction(PN, TD)) { 1645 PN->replaceAllUsesWith(V); 1646 PN->eraseFromParent(); 1647 continue; 1648 } 1649 1650 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, 1651 MaxCostVal0) || 1652 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, 1653 MaxCostVal1)) 1654 return false; 1655 } 1656 1657 // If we folded the first phi, PN dangles at this point. Refresh it. If 1658 // we ran out of PHIs then we simplified them all. 1659 PN = dyn_cast<PHINode>(BB->begin()); 1660 if (PN == 0) return true; 1661 1662 // Don't fold i1 branches on PHIs which contain binary operators. These can 1663 // often be turned into switches and other things. 1664 if (PN->getType()->isIntegerTy(1) && 1665 (isa<BinaryOperator>(PN->getIncomingValue(0)) || 1666 isa<BinaryOperator>(PN->getIncomingValue(1)) || 1667 isa<BinaryOperator>(IfCond))) 1668 return false; 1669 1670 // If we all PHI nodes are promotable, check to make sure that all 1671 // instructions in the predecessor blocks can be promoted as well. If 1672 // not, we won't be able to get rid of the control flow, so it's not 1673 // worth promoting to select instructions. 1674 BasicBlock *DomBlock = 0; 1675 BasicBlock *IfBlock1 = PN->getIncomingBlock(0); 1676 BasicBlock *IfBlock2 = PN->getIncomingBlock(1); 1677 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) { 1678 IfBlock1 = 0; 1679 } else { 1680 DomBlock = *pred_begin(IfBlock1); 1681 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I) 1682 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1683 // This is not an aggressive instruction that we can promote. 1684 // Because of this, we won't be able to get rid of the control 1685 // flow, so the xform is not worth it. 1686 return false; 1687 } 1688 } 1689 1690 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { 1691 IfBlock2 = 0; 1692 } else { 1693 DomBlock = *pred_begin(IfBlock2); 1694 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I) 1695 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1696 // This is not an aggressive instruction that we can promote. 1697 // Because of this, we won't be able to get rid of the control 1698 // flow, so the xform is not worth it. 1699 return false; 1700 } 1701 } 1702 1703 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1704 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1705 1706 // If we can still promote the PHI nodes after this gauntlet of tests, 1707 // do all of the PHI's now. 1708 Instruction *InsertPt = DomBlock->getTerminator(); 1709 IRBuilder<true, NoFolder> Builder(InsertPt); 1710 1711 // Move all 'aggressive' instructions, which are defined in the 1712 // conditional parts of the if's up to the dominating block. 1713 if (IfBlock1) 1714 DomBlock->getInstList().splice(InsertPt, 1715 IfBlock1->getInstList(), IfBlock1->begin(), 1716 IfBlock1->getTerminator()); 1717 if (IfBlock2) 1718 DomBlock->getInstList().splice(InsertPt, 1719 IfBlock2->getInstList(), IfBlock2->begin(), 1720 IfBlock2->getTerminator()); 1721 1722 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1723 // Change the PHI node into a select instruction. 1724 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1725 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1726 1727 SelectInst *NV = 1728 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); 1729 PN->replaceAllUsesWith(NV); 1730 NV->takeName(PN); 1731 PN->eraseFromParent(); 1732 } 1733 1734 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement 1735 // has been flattened. Change DomBlock to jump directly to our new block to 1736 // avoid other simplifycfg's kicking in on the diamond. 1737 TerminatorInst *OldTI = DomBlock->getTerminator(); 1738 Builder.SetInsertPoint(OldTI); 1739 Builder.CreateBr(BB); 1740 OldTI->eraseFromParent(); 1741 return true; 1742} 1743 1744/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1745/// to two returning blocks, try to merge them together into one return, 1746/// introducing a select if the return values disagree. 1747static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, 1748 IRBuilder<> &Builder) { 1749 assert(BI->isConditional() && "Must be a conditional branch"); 1750 BasicBlock *TrueSucc = BI->getSuccessor(0); 1751 BasicBlock *FalseSucc = BI->getSuccessor(1); 1752 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1753 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1754 1755 // Check to ensure both blocks are empty (just a return) or optionally empty 1756 // with PHI nodes. If there are other instructions, merging would cause extra 1757 // computation on one path or the other. 1758 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator()) 1759 return false; 1760 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator()) 1761 return false; 1762 1763 Builder.SetInsertPoint(BI); 1764 // Okay, we found a branch that is going to two return nodes. If 1765 // there is no return value for this function, just change the 1766 // branch into a return. 1767 if (FalseRet->getNumOperands() == 0) { 1768 TrueSucc->removePredecessor(BI->getParent()); 1769 FalseSucc->removePredecessor(BI->getParent()); 1770 Builder.CreateRetVoid(); 1771 EraseTerminatorInstAndDCECond(BI); 1772 return true; 1773 } 1774 1775 // Otherwise, figure out what the true and false return values are 1776 // so we can insert a new select instruction. 1777 Value *TrueValue = TrueRet->getReturnValue(); 1778 Value *FalseValue = FalseRet->getReturnValue(); 1779 1780 // Unwrap any PHI nodes in the return blocks. 1781 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1782 if (TVPN->getParent() == TrueSucc) 1783 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1784 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1785 if (FVPN->getParent() == FalseSucc) 1786 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1787 1788 // In order for this transformation to be safe, we must be able to 1789 // unconditionally execute both operands to the return. This is 1790 // normally the case, but we could have a potentially-trapping 1791 // constant expression that prevents this transformation from being 1792 // safe. 1793 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1794 if (TCV->canTrap()) 1795 return false; 1796 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1797 if (FCV->canTrap()) 1798 return false; 1799 1800 // Okay, we collected all the mapped values and checked them for sanity, and 1801 // defined to really do this transformation. First, update the CFG. 1802 TrueSucc->removePredecessor(BI->getParent()); 1803 FalseSucc->removePredecessor(BI->getParent()); 1804 1805 // Insert select instructions where needed. 1806 Value *BrCond = BI->getCondition(); 1807 if (TrueValue) { 1808 // Insert a select if the results differ. 1809 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1810 } else if (isa<UndefValue>(TrueValue)) { 1811 TrueValue = FalseValue; 1812 } else { 1813 TrueValue = Builder.CreateSelect(BrCond, TrueValue, 1814 FalseValue, "retval"); 1815 } 1816 } 1817 1818 Value *RI = !TrueValue ? 1819 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); 1820 1821 (void) RI; 1822 1823 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1824 << "\n " << *BI << "NewRet = " << *RI 1825 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); 1826 1827 EraseTerminatorInstAndDCECond(BI); 1828 1829 return true; 1830} 1831 1832/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the 1833/// probabilities of the branch taking each edge. Fills in the two APInt 1834/// parameters and return true, or returns false if no or invalid metadata was 1835/// found. 1836static bool ExtractBranchMetadata(BranchInst *BI, 1837 uint64_t &ProbTrue, uint64_t &ProbFalse) { 1838 assert(BI->isConditional() && 1839 "Looking for probabilities on unconditional branch?"); 1840 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 1841 if (!ProfileData || ProfileData->getNumOperands() != 3) return false; 1842 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); 1843 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); 1844 if (!CITrue || !CIFalse) return false; 1845 ProbTrue = CITrue->getValue().getZExtValue(); 1846 ProbFalse = CIFalse->getValue().getZExtValue(); 1847 return true; 1848} 1849 1850/// checkCSEInPredecessor - Return true if the given instruction is available 1851/// in its predecessor block. If yes, the instruction will be removed. 1852/// 1853static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) { 1854 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst)) 1855 return false; 1856 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) { 1857 Instruction *PBI = &*I; 1858 // Check whether Inst and PBI generate the same value. 1859 if (Inst->isIdenticalTo(PBI)) { 1860 Inst->replaceAllUsesWith(PBI); 1861 Inst->eraseFromParent(); 1862 return true; 1863 } 1864 } 1865 return false; 1866} 1867 1868/// FoldBranchToCommonDest - If this basic block is simple enough, and if a 1869/// predecessor branches to us and one of our successors, fold the block into 1870/// the predecessor and use logical operations to pick the right destination. 1871bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1872 BasicBlock *BB = BI->getParent(); 1873 1874 Instruction *Cond = 0; 1875 if (BI->isConditional()) 1876 Cond = dyn_cast<Instruction>(BI->getCondition()); 1877 else { 1878 // For unconditional branch, check for a simple CFG pattern, where 1879 // BB has a single predecessor and BB's successor is also its predecessor's 1880 // successor. If such pattern exisits, check for CSE between BB and its 1881 // predecessor. 1882 if (BasicBlock *PB = BB->getSinglePredecessor()) 1883 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator())) 1884 if (PBI->isConditional() && 1885 (BI->getSuccessor(0) == PBI->getSuccessor(0) || 1886 BI->getSuccessor(0) == PBI->getSuccessor(1))) { 1887 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 1888 I != E; ) { 1889 Instruction *Curr = I++; 1890 if (isa<CmpInst>(Curr)) { 1891 Cond = Curr; 1892 break; 1893 } 1894 // Quit if we can't remove this instruction. 1895 if (!checkCSEInPredecessor(Curr, PB)) 1896 return false; 1897 } 1898 } 1899 1900 if (Cond == 0) 1901 return false; 1902 } 1903 1904 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1905 Cond->getParent() != BB || !Cond->hasOneUse()) 1906 return false; 1907 1908 // Only allow this if the condition is a simple instruction that can be 1909 // executed unconditionally. It must be in the same block as the branch, and 1910 // must be at the front of the block. 1911 BasicBlock::iterator FrontIt = BB->front(); 1912 1913 // Ignore dbg intrinsics. 1914 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1915 1916 // Allow a single instruction to be hoisted in addition to the compare 1917 // that feeds the branch. We later ensure that any values that _it_ uses 1918 // were also live in the predecessor, so that we don't unnecessarily create 1919 // register pressure or inhibit out-of-order execution. 1920 Instruction *BonusInst = 0; 1921 if (&*FrontIt != Cond && 1922 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond && 1923 isSafeToSpeculativelyExecute(FrontIt)) { 1924 BonusInst = &*FrontIt; 1925 ++FrontIt; 1926 1927 // Ignore dbg intrinsics. 1928 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; 1929 } 1930 1931 // Only a single bonus inst is allowed. 1932 if (&*FrontIt != Cond) 1933 return false; 1934 1935 // Make sure the instruction after the condition is the cond branch. 1936 BasicBlock::iterator CondIt = Cond; ++CondIt; 1937 1938 // Ingore dbg intrinsics. 1939 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; 1940 1941 if (&*CondIt != BI) 1942 return false; 1943 1944 // Cond is known to be a compare or binary operator. Check to make sure that 1945 // neither operand is a potentially-trapping constant expression. 1946 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1947 if (CE->canTrap()) 1948 return false; 1949 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1950 if (CE->canTrap()) 1951 return false; 1952 1953 // Finally, don't infinitely unroll conditional loops. 1954 BasicBlock *TrueDest = BI->getSuccessor(0); 1955 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; 1956 if (TrueDest == BB || FalseDest == BB) 1957 return false; 1958 1959 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1960 BasicBlock *PredBlock = *PI; 1961 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1962 1963 // Check that we have two conditional branches. If there is a PHI node in 1964 // the common successor, verify that the same value flows in from both 1965 // blocks. 1966 SmallVector<PHINode*, 4> PHIs; 1967 if (PBI == 0 || PBI->isUnconditional() || 1968 (BI->isConditional() && 1969 !SafeToMergeTerminators(BI, PBI)) || 1970 (!BI->isConditional() && 1971 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) 1972 continue; 1973 1974 // Determine if the two branches share a common destination. 1975 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; 1976 bool InvertPredCond = false; 1977 1978 if (BI->isConditional()) { 1979 if (PBI->getSuccessor(0) == TrueDest) 1980 Opc = Instruction::Or; 1981 else if (PBI->getSuccessor(1) == FalseDest) 1982 Opc = Instruction::And; 1983 else if (PBI->getSuccessor(0) == FalseDest) 1984 Opc = Instruction::And, InvertPredCond = true; 1985 else if (PBI->getSuccessor(1) == TrueDest) 1986 Opc = Instruction::Or, InvertPredCond = true; 1987 else 1988 continue; 1989 } else { 1990 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest) 1991 continue; 1992 } 1993 1994 // Ensure that any values used in the bonus instruction are also used 1995 // by the terminator of the predecessor. This means that those values 1996 // must already have been resolved, so we won't be inhibiting the 1997 // out-of-order core by speculating them earlier. 1998 if (BonusInst) { 1999 // Collect the values used by the bonus inst 2000 SmallPtrSet<Value*, 4> UsedValues; 2001 for (Instruction::op_iterator OI = BonusInst->op_begin(), 2002 OE = BonusInst->op_end(); OI != OE; ++OI) { 2003 Value *V = *OI; 2004 if (!isa<Constant>(V)) 2005 UsedValues.insert(V); 2006 } 2007 2008 SmallVector<std::pair<Value*, unsigned>, 4> Worklist; 2009 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); 2010 2011 // Walk up to four levels back up the use-def chain of the predecessor's 2012 // terminator to see if all those values were used. The choice of four 2013 // levels is arbitrary, to provide a compile-time-cost bound. 2014 while (!Worklist.empty()) { 2015 std::pair<Value*, unsigned> Pair = Worklist.back(); 2016 Worklist.pop_back(); 2017 2018 if (Pair.second >= 4) continue; 2019 UsedValues.erase(Pair.first); 2020 if (UsedValues.empty()) break; 2021 2022 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { 2023 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 2024 OI != OE; ++OI) 2025 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); 2026 } 2027 } 2028 2029 if (!UsedValues.empty()) return false; 2030 } 2031 2032 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); 2033 IRBuilder<> Builder(PBI); 2034 2035 // If we need to invert the condition in the pred block to match, do so now. 2036 if (InvertPredCond) { 2037 Value *NewCond = PBI->getCondition(); 2038 2039 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 2040 CmpInst *CI = cast<CmpInst>(NewCond); 2041 CI->setPredicate(CI->getInversePredicate()); 2042 } else { 2043 NewCond = Builder.CreateNot(NewCond, 2044 PBI->getCondition()->getName()+".not"); 2045 } 2046 2047 PBI->setCondition(NewCond); 2048 PBI->swapSuccessors(); 2049 } 2050 2051 // If we have a bonus inst, clone it into the predecessor block. 2052 Instruction *NewBonus = 0; 2053 if (BonusInst) { 2054 NewBonus = BonusInst->clone(); 2055 PredBlock->getInstList().insert(PBI, NewBonus); 2056 NewBonus->takeName(BonusInst); 2057 BonusInst->setName(BonusInst->getName()+".old"); 2058 } 2059 2060 // Clone Cond into the predecessor basic block, and or/and the 2061 // two conditions together. 2062 Instruction *New = Cond->clone(); 2063 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus); 2064 PredBlock->getInstList().insert(PBI, New); 2065 New->takeName(Cond); 2066 Cond->setName(New->getName()+".old"); 2067 2068 if (BI->isConditional()) { 2069 Instruction *NewCond = 2070 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), 2071 New, "or.cond")); 2072 PBI->setCondition(NewCond); 2073 2074 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2075 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2076 PredFalseWeight); 2077 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2078 SuccFalseWeight); 2079 SmallVector<uint64_t, 8> NewWeights; 2080 2081 if (PBI->getSuccessor(0) == BB) { 2082 if (PredHasWeights && SuccHasWeights) { 2083 // PBI: br i1 %x, BB, FalseDest 2084 // BI: br i1 %y, TrueDest, FalseDest 2085 //TrueWeight is TrueWeight for PBI * TrueWeight for BI. 2086 NewWeights.push_back(PredTrueWeight * SuccTrueWeight); 2087 //FalseWeight is FalseWeight for PBI * TotalWeight for BI + 2088 // TrueWeight for PBI * FalseWeight for BI. 2089 // We assume that total weights of a BranchInst can fit into 32 bits. 2090 // Therefore, we will not have overflow using 64-bit arithmetic. 2091 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + 2092 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); 2093 } 2094 AddPredecessorToBlock(TrueDest, PredBlock, BB); 2095 PBI->setSuccessor(0, TrueDest); 2096 } 2097 if (PBI->getSuccessor(1) == BB) { 2098 if (PredHasWeights && SuccHasWeights) { 2099 // PBI: br i1 %x, TrueDest, BB 2100 // BI: br i1 %y, TrueDest, FalseDest 2101 //TrueWeight is TrueWeight for PBI * TotalWeight for BI + 2102 // FalseWeight for PBI * TrueWeight for BI. 2103 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + 2104 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); 2105 //FalseWeight is FalseWeight for PBI * FalseWeight for BI. 2106 NewWeights.push_back(PredFalseWeight * SuccFalseWeight); 2107 } 2108 AddPredecessorToBlock(FalseDest, PredBlock, BB); 2109 PBI->setSuccessor(1, FalseDest); 2110 } 2111 if (NewWeights.size() == 2) { 2112 // Halve the weights if any of them cannot fit in an uint32_t 2113 FitWeights(NewWeights); 2114 2115 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); 2116 PBI->setMetadata(LLVMContext::MD_prof, 2117 MDBuilder(BI->getContext()). 2118 createBranchWeights(MDWeights)); 2119 } else 2120 PBI->setMetadata(LLVMContext::MD_prof, NULL); 2121 } else { 2122 // Update PHI nodes in the common successors. 2123 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { 2124 ConstantInt *PBI_C = cast<ConstantInt>( 2125 PHIs[i]->getIncomingValueForBlock(PBI->getParent())); 2126 assert(PBI_C->getType()->isIntegerTy(1)); 2127 Instruction *MergedCond = 0; 2128 if (PBI->getSuccessor(0) == TrueDest) { 2129 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value) 2130 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value) 2131 // is false: !PBI_Cond and BI_Value 2132 Instruction *NotCond = 2133 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2134 "not.cond")); 2135 MergedCond = 2136 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2137 NotCond, New, 2138 "and.cond")); 2139 if (PBI_C->isOne()) 2140 MergedCond = 2141 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2142 PBI->getCondition(), MergedCond, 2143 "or.cond")); 2144 } else { 2145 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) 2146 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) 2147 // is false: PBI_Cond and BI_Value 2148 MergedCond = 2149 cast<Instruction>(Builder.CreateBinOp(Instruction::And, 2150 PBI->getCondition(), New, 2151 "and.cond")); 2152 if (PBI_C->isOne()) { 2153 Instruction *NotCond = 2154 cast<Instruction>(Builder.CreateNot(PBI->getCondition(), 2155 "not.cond")); 2156 MergedCond = 2157 cast<Instruction>(Builder.CreateBinOp(Instruction::Or, 2158 NotCond, MergedCond, 2159 "or.cond")); 2160 } 2161 } 2162 // Update PHI Node. 2163 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()), 2164 MergedCond); 2165 } 2166 // Change PBI from Conditional to Unconditional. 2167 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI); 2168 EraseTerminatorInstAndDCECond(PBI); 2169 PBI = New_PBI; 2170 } 2171 2172 // TODO: If BB is reachable from all paths through PredBlock, then we 2173 // could replace PBI's branch probabilities with BI's. 2174 2175 // Copy any debug value intrinsics into the end of PredBlock. 2176 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 2177 if (isa<DbgInfoIntrinsic>(*I)) 2178 I->clone()->insertBefore(PBI); 2179 2180 return true; 2181 } 2182 return false; 2183} 2184 2185/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 2186/// predecessor of another block, this function tries to simplify it. We know 2187/// that PBI and BI are both conditional branches, and BI is in one of the 2188/// successor blocks of PBI - PBI branches to BI. 2189static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 2190 assert(PBI->isConditional() && BI->isConditional()); 2191 BasicBlock *BB = BI->getParent(); 2192 2193 // If this block ends with a branch instruction, and if there is a 2194 // predecessor that ends on a branch of the same condition, make 2195 // this conditional branch redundant. 2196 if (PBI->getCondition() == BI->getCondition() && 2197 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2198 // Okay, the outcome of this conditional branch is statically 2199 // knowable. If this block had a single pred, handle specially. 2200 if (BB->getSinglePredecessor()) { 2201 // Turn this into a branch on constant. 2202 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2203 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2204 CondIsTrue)); 2205 return true; // Nuke the branch on constant. 2206 } 2207 2208 // Otherwise, if there are multiple predecessors, insert a PHI that merges 2209 // in the constant and simplify the block result. Subsequent passes of 2210 // simplifycfg will thread the block. 2211 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 2212 pred_iterator PB = pred_begin(BB), PE = pred_end(BB); 2213 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()), 2214 std::distance(PB, PE), 2215 BI->getCondition()->getName() + ".pr", 2216 BB->begin()); 2217 // Okay, we're going to insert the PHI node. Since PBI is not the only 2218 // predecessor, compute the PHI'd conditional value for all of the preds. 2219 // Any predecessor where the condition is not computable we keep symbolic. 2220 for (pred_iterator PI = PB; PI != PE; ++PI) { 2221 BasicBlock *P = *PI; 2222 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && 2223 PBI != BI && PBI->isConditional() && 2224 PBI->getCondition() == BI->getCondition() && 2225 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 2226 bool CondIsTrue = PBI->getSuccessor(0) == BB; 2227 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 2228 CondIsTrue), P); 2229 } else { 2230 NewPN->addIncoming(BI->getCondition(), P); 2231 } 2232 } 2233 2234 BI->setCondition(NewPN); 2235 return true; 2236 } 2237 } 2238 2239 // If this is a conditional branch in an empty block, and if any 2240 // predecessors is a conditional branch to one of our destinations, 2241 // fold the conditions into logical ops and one cond br. 2242 BasicBlock::iterator BBI = BB->begin(); 2243 // Ignore dbg intrinsics. 2244 while (isa<DbgInfoIntrinsic>(BBI)) 2245 ++BBI; 2246 if (&*BBI != BI) 2247 return false; 2248 2249 2250 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 2251 if (CE->canTrap()) 2252 return false; 2253 2254 int PBIOp, BIOp; 2255 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 2256 PBIOp = BIOp = 0; 2257 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 2258 PBIOp = 0, BIOp = 1; 2259 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 2260 PBIOp = 1, BIOp = 0; 2261 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 2262 PBIOp = BIOp = 1; 2263 else 2264 return false; 2265 2266 // Check to make sure that the other destination of this branch 2267 // isn't BB itself. If so, this is an infinite loop that will 2268 // keep getting unwound. 2269 if (PBI->getSuccessor(PBIOp) == BB) 2270 return false; 2271 2272 // Do not perform this transformation if it would require 2273 // insertion of a large number of select instructions. For targets 2274 // without predication/cmovs, this is a big pessimization. 2275 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 2276 2277 unsigned NumPhis = 0; 2278 for (BasicBlock::iterator II = CommonDest->begin(); 2279 isa<PHINode>(II); ++II, ++NumPhis) 2280 if (NumPhis > 2) // Disable this xform. 2281 return false; 2282 2283 // Finally, if everything is ok, fold the branches to logical ops. 2284 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 2285 2286 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() 2287 << "AND: " << *BI->getParent()); 2288 2289 2290 // If OtherDest *is* BB, then BB is a basic block with a single conditional 2291 // branch in it, where one edge (OtherDest) goes back to itself but the other 2292 // exits. We don't *know* that the program avoids the infinite loop 2293 // (even though that seems likely). If we do this xform naively, we'll end up 2294 // recursively unpeeling the loop. Since we know that (after the xform is 2295 // done) that the block *is* infinite if reached, we just make it an obviously 2296 // infinite loop with no cond branch. 2297 if (OtherDest == BB) { 2298 // Insert it at the end of the function, because it's either code, 2299 // or it won't matter if it's hot. :) 2300 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(), 2301 "infloop", BB->getParent()); 2302 BranchInst::Create(InfLoopBlock, InfLoopBlock); 2303 OtherDest = InfLoopBlock; 2304 } 2305 2306 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2307 2308 // BI may have other predecessors. Because of this, we leave 2309 // it alone, but modify PBI. 2310 2311 // Make sure we get to CommonDest on True&True directions. 2312 Value *PBICond = PBI->getCondition(); 2313 IRBuilder<true, NoFolder> Builder(PBI); 2314 if (PBIOp) 2315 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not"); 2316 2317 Value *BICond = BI->getCondition(); 2318 if (BIOp) 2319 BICond = Builder.CreateNot(BICond, BICond->getName()+".not"); 2320 2321 // Merge the conditions. 2322 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); 2323 2324 // Modify PBI to branch on the new condition to the new dests. 2325 PBI->setCondition(Cond); 2326 PBI->setSuccessor(0, CommonDest); 2327 PBI->setSuccessor(1, OtherDest); 2328 2329 // Update branch weight for PBI. 2330 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; 2331 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, 2332 PredFalseWeight); 2333 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, 2334 SuccFalseWeight); 2335 if (PredHasWeights && SuccHasWeights) { 2336 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; 2337 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; 2338 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; 2339 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; 2340 // The weight to CommonDest should be PredCommon * SuccTotal + 2341 // PredOther * SuccCommon. 2342 // The weight to OtherDest should be PredOther * SuccOther. 2343 SmallVector<uint64_t, 2> NewWeights; 2344 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + 2345 PredOther * SuccCommon); 2346 NewWeights.push_back(PredOther * SuccOther); 2347 // Halve the weights if any of them cannot fit in an uint32_t 2348 FitWeights(NewWeights); 2349 2350 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); 2351 PBI->setMetadata(LLVMContext::MD_prof, 2352 MDBuilder(BI->getContext()). 2353 createBranchWeights(MDWeights)); 2354 } 2355 2356 // OtherDest may have phi nodes. If so, add an entry from PBI's 2357 // block that are identical to the entries for BI's block. 2358 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); 2359 2360 // We know that the CommonDest already had an edge from PBI to 2361 // it. If it has PHIs though, the PHIs may have different 2362 // entries for BB and PBI's BB. If so, insert a select to make 2363 // them agree. 2364 PHINode *PN; 2365 for (BasicBlock::iterator II = CommonDest->begin(); 2366 (PN = dyn_cast<PHINode>(II)); ++II) { 2367 Value *BIV = PN->getIncomingValueForBlock(BB); 2368 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 2369 Value *PBIV = PN->getIncomingValue(PBBIdx); 2370 if (BIV != PBIV) { 2371 // Insert a select in PBI to pick the right value. 2372 Value *NV = cast<SelectInst> 2373 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux")); 2374 PN->setIncomingValue(PBBIdx, NV); 2375 } 2376 } 2377 2378 DEBUG(dbgs() << "INTO: " << *PBI->getParent()); 2379 DEBUG(dbgs() << *PBI->getParent()->getParent()); 2380 2381 // This basic block is probably dead. We know it has at least 2382 // one fewer predecessor. 2383 return true; 2384} 2385 2386// SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a 2387// branch to TrueBB if Cond is true or to FalseBB if Cond is false. 2388// Takes care of updating the successors and removing the old terminator. 2389// Also makes sure not to introduce new successors by assuming that edges to 2390// non-successor TrueBBs and FalseBBs aren't reachable. 2391static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, 2392 BasicBlock *TrueBB, BasicBlock *FalseBB, 2393 uint32_t TrueWeight, 2394 uint32_t FalseWeight){ 2395 // Remove any superfluous successor edges from the CFG. 2396 // First, figure out which successors to preserve. 2397 // If TrueBB and FalseBB are equal, only try to preserve one copy of that 2398 // successor. 2399 BasicBlock *KeepEdge1 = TrueBB; 2400 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0; 2401 2402 // Then remove the rest. 2403 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) { 2404 BasicBlock *Succ = OldTerm->getSuccessor(I); 2405 // Make sure only to keep exactly one copy of each edge. 2406 if (Succ == KeepEdge1) 2407 KeepEdge1 = 0; 2408 else if (Succ == KeepEdge2) 2409 KeepEdge2 = 0; 2410 else 2411 Succ->removePredecessor(OldTerm->getParent()); 2412 } 2413 2414 IRBuilder<> Builder(OldTerm); 2415 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); 2416 2417 // Insert an appropriate new terminator. 2418 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) { 2419 if (TrueBB == FalseBB) 2420 // We were only looking for one successor, and it was present. 2421 // Create an unconditional branch to it. 2422 Builder.CreateBr(TrueBB); 2423 else { 2424 // We found both of the successors we were looking for. 2425 // Create a conditional branch sharing the condition of the select. 2426 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); 2427 if (TrueWeight != FalseWeight) 2428 NewBI->setMetadata(LLVMContext::MD_prof, 2429 MDBuilder(OldTerm->getContext()). 2430 createBranchWeights(TrueWeight, FalseWeight)); 2431 } 2432 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { 2433 // Neither of the selected blocks were successors, so this 2434 // terminator must be unreachable. 2435 new UnreachableInst(OldTerm->getContext(), OldTerm); 2436 } else { 2437 // One of the selected values was a successor, but the other wasn't. 2438 // Insert an unconditional branch to the one that was found; 2439 // the edge to the one that wasn't must be unreachable. 2440 if (KeepEdge1 == 0) 2441 // Only TrueBB was found. 2442 Builder.CreateBr(TrueBB); 2443 else 2444 // Only FalseBB was found. 2445 Builder.CreateBr(FalseBB); 2446 } 2447 2448 EraseTerminatorInstAndDCECond(OldTerm); 2449 return true; 2450} 2451 2452// SimplifySwitchOnSelect - Replaces 2453// (switch (select cond, X, Y)) on constant X, Y 2454// with a branch - conditional if X and Y lead to distinct BBs, 2455// unconditional otherwise. 2456static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { 2457 // Check for constant integer values in the select. 2458 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue()); 2459 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue()); 2460 if (!TrueVal || !FalseVal) 2461 return false; 2462 2463 // Find the relevant condition and destinations. 2464 Value *Condition = Select->getCondition(); 2465 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); 2466 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); 2467 2468 // Get weight for TrueBB and FalseBB. 2469 uint32_t TrueWeight = 0, FalseWeight = 0; 2470 SmallVector<uint64_t, 8> Weights; 2471 bool HasWeights = HasBranchWeights(SI); 2472 if (HasWeights) { 2473 GetBranchWeights(SI, Weights); 2474 if (Weights.size() == 1 + SI->getNumCases()) { 2475 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). 2476 getSuccessorIndex()]; 2477 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). 2478 getSuccessorIndex()]; 2479 } 2480 } 2481 2482 // Perform the actual simplification. 2483 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, 2484 TrueWeight, FalseWeight); 2485} 2486 2487// SimplifyIndirectBrOnSelect - Replaces 2488// (indirectbr (select cond, blockaddress(@fn, BlockA), 2489// blockaddress(@fn, BlockB))) 2490// with 2491// (br cond, BlockA, BlockB). 2492static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { 2493 // Check that both operands of the select are block addresses. 2494 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue()); 2495 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue()); 2496 if (!TBA || !FBA) 2497 return false; 2498 2499 // Extract the actual blocks. 2500 BasicBlock *TrueBB = TBA->getBasicBlock(); 2501 BasicBlock *FalseBB = FBA->getBasicBlock(); 2502 2503 // Perform the actual simplification. 2504 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 2505 0, 0); 2506} 2507 2508/// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp 2509/// instruction (a seteq/setne with a constant) as the only instruction in a 2510/// block that ends with an uncond branch. We are looking for a very specific 2511/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In 2512/// this case, we merge the first two "or's of icmp" into a switch, but then the 2513/// default value goes to an uncond block with a seteq in it, we get something 2514/// like: 2515/// 2516/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] 2517/// DEFAULT: 2518/// %tmp = icmp eq i8 %A, 92 2519/// br label %end 2520/// end: 2521/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] 2522/// 2523/// We prefer to split the edge to 'end' so that there is a true/false entry to 2524/// the PHI, merging the third icmp into the switch. 2525static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, 2526 const DataLayout *TD, 2527 IRBuilder<> &Builder) { 2528 BasicBlock *BB = ICI->getParent(); 2529 2530 // If the block has any PHIs in it or the icmp has multiple uses, it is too 2531 // complex. 2532 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false; 2533 2534 Value *V = ICI->getOperand(0); 2535 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); 2536 2537 // The pattern we're looking for is where our only predecessor is a switch on 2538 // 'V' and this block is the default case for the switch. In this case we can 2539 // fold the compared value into the switch to simplify things. 2540 BasicBlock *Pred = BB->getSinglePredecessor(); 2541 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; 2542 2543 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); 2544 if (SI->getCondition() != V) 2545 return false; 2546 2547 // If BB is reachable on a non-default case, then we simply know the value of 2548 // V in this block. Substitute it and constant fold the icmp instruction 2549 // away. 2550 if (SI->getDefaultDest() != BB) { 2551 ConstantInt *VVal = SI->findCaseDest(BB); 2552 assert(VVal && "Should have a unique destination value"); 2553 ICI->setOperand(0, VVal); 2554 2555 if (Value *V = SimplifyInstruction(ICI, TD)) { 2556 ICI->replaceAllUsesWith(V); 2557 ICI->eraseFromParent(); 2558 } 2559 // BB is now empty, so it is likely to simplify away. 2560 return SimplifyCFG(BB) | true; 2561 } 2562 2563 // Ok, the block is reachable from the default dest. If the constant we're 2564 // comparing exists in one of the other edges, then we can constant fold ICI 2565 // and zap it. 2566 if (SI->findCaseValue(Cst) != SI->case_default()) { 2567 Value *V; 2568 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2569 V = ConstantInt::getFalse(BB->getContext()); 2570 else 2571 V = ConstantInt::getTrue(BB->getContext()); 2572 2573 ICI->replaceAllUsesWith(V); 2574 ICI->eraseFromParent(); 2575 // BB is now empty, so it is likely to simplify away. 2576 return SimplifyCFG(BB) | true; 2577 } 2578 2579 // The use of the icmp has to be in the 'end' block, by the only PHI node in 2580 // the block. 2581 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); 2582 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back()); 2583 if (PHIUse == 0 || PHIUse != &SuccBlock->front() || 2584 isa<PHINode>(++BasicBlock::iterator(PHIUse))) 2585 return false; 2586 2587 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets 2588 // true in the PHI. 2589 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext()); 2590 Constant *NewCst = ConstantInt::getFalse(BB->getContext()); 2591 2592 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 2593 std::swap(DefaultCst, NewCst); 2594 2595 // Replace ICI (which is used by the PHI for the default value) with true or 2596 // false depending on if it is EQ or NE. 2597 ICI->replaceAllUsesWith(DefaultCst); 2598 ICI->eraseFromParent(); 2599 2600 // Okay, the switch goes to this block on a default value. Add an edge from 2601 // the switch to the merge point on the compared value. 2602 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", 2603 BB->getParent(), BB); 2604 SmallVector<uint64_t, 8> Weights; 2605 bool HasWeights = HasBranchWeights(SI); 2606 if (HasWeights) { 2607 GetBranchWeights(SI, Weights); 2608 if (Weights.size() == 1 + SI->getNumCases()) { 2609 // Split weight for default case to case for "Cst". 2610 Weights[0] = (Weights[0]+1) >> 1; 2611 Weights.push_back(Weights[0]); 2612 2613 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 2614 SI->setMetadata(LLVMContext::MD_prof, 2615 MDBuilder(SI->getContext()). 2616 createBranchWeights(MDWeights)); 2617 } 2618 } 2619 SI->addCase(Cst, NewBB); 2620 2621 // NewBB branches to the phi block, add the uncond branch and the phi entry. 2622 Builder.SetInsertPoint(NewBB); 2623 Builder.SetCurrentDebugLocation(SI->getDebugLoc()); 2624 Builder.CreateBr(SuccBlock); 2625 PHIUse->addIncoming(NewCst, NewBB); 2626 return true; 2627} 2628 2629/// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. 2630/// Check to see if it is branching on an or/and chain of icmp instructions, and 2631/// fold it into a switch instruction if so. 2632static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, 2633 IRBuilder<> &Builder) { 2634 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 2635 if (Cond == 0) return false; 2636 2637 2638 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2639 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2640 // 'setne's and'ed together, collect them. 2641 Value *CompVal = 0; 2642 std::vector<ConstantInt*> Values; 2643 bool TrueWhenEqual = true; 2644 Value *ExtraCase = 0; 2645 unsigned UsedICmps = 0; 2646 2647 if (Cond->getOpcode() == Instruction::Or) { 2648 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, 2649 UsedICmps); 2650 } else if (Cond->getOpcode() == Instruction::And) { 2651 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false, 2652 UsedICmps); 2653 TrueWhenEqual = false; 2654 } 2655 2656 // If we didn't have a multiply compared value, fail. 2657 if (CompVal == 0) return false; 2658 2659 // Avoid turning single icmps into a switch. 2660 if (UsedICmps <= 1) 2661 return false; 2662 2663 // There might be duplicate constants in the list, which the switch 2664 // instruction can't handle, remove them now. 2665 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); 2666 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2667 2668 // If Extra was used, we require at least two switch values to do the 2669 // transformation. A switch with one value is just an cond branch. 2670 if (ExtraCase && Values.size() < 2) return false; 2671 2672 // TODO: Preserve branch weight metadata, similarly to how 2673 // FoldValueComparisonIntoPredecessors preserves it. 2674 2675 // Figure out which block is which destination. 2676 BasicBlock *DefaultBB = BI->getSuccessor(1); 2677 BasicBlock *EdgeBB = BI->getSuccessor(0); 2678 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2679 2680 BasicBlock *BB = BI->getParent(); 2681 2682 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() 2683 << " cases into SWITCH. BB is:\n" << *BB); 2684 2685 // If there are any extra values that couldn't be folded into the switch 2686 // then we evaluate them with an explicit branch first. Split the block 2687 // right before the condbr to handle it. 2688 if (ExtraCase) { 2689 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test"); 2690 // Remove the uncond branch added to the old block. 2691 TerminatorInst *OldTI = BB->getTerminator(); 2692 Builder.SetInsertPoint(OldTI); 2693 2694 if (TrueWhenEqual) 2695 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); 2696 else 2697 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); 2698 2699 OldTI->eraseFromParent(); 2700 2701 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them 2702 // for the edge we just added. 2703 AddPredecessorToBlock(EdgeBB, BB, NewBB); 2704 2705 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase 2706 << "\nEXTRABB = " << *BB); 2707 BB = NewBB; 2708 } 2709 2710 Builder.SetInsertPoint(BI); 2711 // Convert pointer to int before we switch. 2712 if (CompVal->getType()->isPointerTy()) { 2713 assert(TD && "Cannot switch on pointer without DataLayout"); 2714 CompVal = Builder.CreatePtrToInt(CompVal, 2715 TD->getIntPtrType(CompVal->getContext()), 2716 "magicptr"); 2717 } 2718 2719 // Create the new switch instruction now. 2720 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); 2721 2722 // Add all of the 'cases' to the switch instruction. 2723 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2724 New->addCase(Values[i], EdgeBB); 2725 2726 // We added edges from PI to the EdgeBB. As such, if there were any 2727 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2728 // the number of edges added. 2729 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2730 isa<PHINode>(BBI); ++BBI) { 2731 PHINode *PN = cast<PHINode>(BBI); 2732 Value *InVal = PN->getIncomingValueForBlock(BB); 2733 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2734 PN->addIncoming(InVal, BB); 2735 } 2736 2737 // Erase the old branch instruction. 2738 EraseTerminatorInstAndDCECond(BI); 2739 2740 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); 2741 return true; 2742} 2743 2744bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { 2745 // If this is a trivial landing pad that just continues unwinding the caught 2746 // exception then zap the landing pad, turning its invokes into calls. 2747 BasicBlock *BB = RI->getParent(); 2748 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI()); 2749 if (RI->getValue() != LPInst) 2750 // Not a landing pad, or the resume is not unwinding the exception that 2751 // caused control to branch here. 2752 return false; 2753 2754 // Check that there are no other instructions except for debug intrinsics. 2755 BasicBlock::iterator I = LPInst, E = RI; 2756 while (++I != E) 2757 if (!isa<DbgInfoIntrinsic>(I)) 2758 return false; 2759 2760 // Turn all invokes that unwind here into calls and delete the basic block. 2761 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) { 2762 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator()); 2763 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 2764 // Insert a call instruction before the invoke. 2765 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II); 2766 Call->takeName(II); 2767 Call->setCallingConv(II->getCallingConv()); 2768 Call->setAttributes(II->getAttributes()); 2769 Call->setDebugLoc(II->getDebugLoc()); 2770 2771 // Anything that used the value produced by the invoke instruction now uses 2772 // the value produced by the call instruction. Note that we do this even 2773 // for void functions and calls with no uses so that the callgraph edge is 2774 // updated. 2775 II->replaceAllUsesWith(Call); 2776 BB->removePredecessor(II->getParent()); 2777 2778 // Insert a branch to the normal destination right before the invoke. 2779 BranchInst::Create(II->getNormalDest(), II); 2780 2781 // Finally, delete the invoke instruction! 2782 II->eraseFromParent(); 2783 } 2784 2785 // The landingpad is now unreachable. Zap it. 2786 BB->eraseFromParent(); 2787 return true; 2788} 2789 2790bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { 2791 BasicBlock *BB = RI->getParent(); 2792 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; 2793 2794 // Find predecessors that end with branches. 2795 SmallVector<BasicBlock*, 8> UncondBranchPreds; 2796 SmallVector<BranchInst*, 8> CondBranchPreds; 2797 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 2798 BasicBlock *P = *PI; 2799 TerminatorInst *PTI = P->getTerminator(); 2800 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 2801 if (BI->isUnconditional()) 2802 UncondBranchPreds.push_back(P); 2803 else 2804 CondBranchPreds.push_back(BI); 2805 } 2806 } 2807 2808 // If we found some, do the transformation! 2809 if (!UncondBranchPreds.empty() && DupRet) { 2810 while (!UncondBranchPreds.empty()) { 2811 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 2812 DEBUG(dbgs() << "FOLDING: " << *BB 2813 << "INTO UNCOND BRANCH PRED: " << *Pred); 2814 (void)FoldReturnIntoUncondBranch(RI, BB, Pred); 2815 } 2816 2817 // If we eliminated all predecessors of the block, delete the block now. 2818 if (pred_begin(BB) == pred_end(BB)) 2819 // We know there are no successors, so just nuke the block. 2820 BB->eraseFromParent(); 2821 2822 return true; 2823 } 2824 2825 // Check out all of the conditional branches going to this return 2826 // instruction. If any of them just select between returns, change the 2827 // branch itself into a select/return pair. 2828 while (!CondBranchPreds.empty()) { 2829 BranchInst *BI = CondBranchPreds.pop_back_val(); 2830 2831 // Check to see if the non-BB successor is also a return block. 2832 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 2833 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 2834 SimplifyCondBranchToTwoReturns(BI, Builder)) 2835 return true; 2836 } 2837 return false; 2838} 2839 2840bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { 2841 BasicBlock *BB = UI->getParent(); 2842 2843 bool Changed = false; 2844 2845 // If there are any instructions immediately before the unreachable that can 2846 // be removed, do so. 2847 while (UI != BB->begin()) { 2848 BasicBlock::iterator BBI = UI; 2849 --BBI; 2850 // Do not delete instructions that can have side effects which might cause 2851 // the unreachable to not be reachable; specifically, calls and volatile 2852 // operations may have this effect. 2853 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2854 2855 if (BBI->mayHaveSideEffects()) { 2856 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 2857 if (SI->isVolatile()) 2858 break; 2859 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 2860 if (LI->isVolatile()) 2861 break; 2862 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) { 2863 if (RMWI->isVolatile()) 2864 break; 2865 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) { 2866 if (CXI->isVolatile()) 2867 break; 2868 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) && 2869 !isa<LandingPadInst>(BBI)) { 2870 break; 2871 } 2872 // Note that deleting LandingPad's here is in fact okay, although it 2873 // involves a bit of subtle reasoning. If this inst is a LandingPad, 2874 // all the predecessors of this block will be the unwind edges of Invokes, 2875 // and we can therefore guarantee this block will be erased. 2876 } 2877 2878 // Delete this instruction (any uses are guaranteed to be dead) 2879 if (!BBI->use_empty()) 2880 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 2881 BBI->eraseFromParent(); 2882 Changed = true; 2883 } 2884 2885 // If the unreachable instruction is the first in the block, take a gander 2886 // at all of the predecessors of this instruction, and simplify them. 2887 if (&BB->front() != UI) return Changed; 2888 2889 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2890 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2891 TerminatorInst *TI = Preds[i]->getTerminator(); 2892 IRBuilder<> Builder(TI); 2893 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2894 if (BI->isUnconditional()) { 2895 if (BI->getSuccessor(0) == BB) { 2896 new UnreachableInst(TI->getContext(), TI); 2897 TI->eraseFromParent(); 2898 Changed = true; 2899 } 2900 } else { 2901 if (BI->getSuccessor(0) == BB) { 2902 Builder.CreateBr(BI->getSuccessor(1)); 2903 EraseTerminatorInstAndDCECond(BI); 2904 } else if (BI->getSuccessor(1) == BB) { 2905 Builder.CreateBr(BI->getSuccessor(0)); 2906 EraseTerminatorInstAndDCECond(BI); 2907 Changed = true; 2908 } 2909 } 2910 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2911 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2912 i != e; ++i) 2913 if (i.getCaseSuccessor() == BB) { 2914 BB->removePredecessor(SI->getParent()); 2915 SI->removeCase(i); 2916 --i; --e; 2917 Changed = true; 2918 } 2919 // If the default value is unreachable, figure out the most popular 2920 // destination and make it the default. 2921 if (SI->getDefaultDest() == BB) { 2922 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity; 2923 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2924 i != e; ++i) { 2925 std::pair<unsigned, unsigned> &entry = 2926 Popularity[i.getCaseSuccessor()]; 2927 if (entry.first == 0) { 2928 entry.first = 1; 2929 entry.second = i.getCaseIndex(); 2930 } else { 2931 entry.first++; 2932 } 2933 } 2934 2935 // Find the most popular block. 2936 unsigned MaxPop = 0; 2937 unsigned MaxIndex = 0; 2938 BasicBlock *MaxBlock = 0; 2939 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator 2940 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2941 if (I->second.first > MaxPop || 2942 (I->second.first == MaxPop && MaxIndex > I->second.second)) { 2943 MaxPop = I->second.first; 2944 MaxIndex = I->second.second; 2945 MaxBlock = I->first; 2946 } 2947 } 2948 if (MaxBlock) { 2949 // Make this the new default, allowing us to delete any explicit 2950 // edges to it. 2951 SI->setDefaultDest(MaxBlock); 2952 Changed = true; 2953 2954 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2955 // it. 2956 if (isa<PHINode>(MaxBlock->begin())) 2957 for (unsigned i = 0; i != MaxPop-1; ++i) 2958 MaxBlock->removePredecessor(SI->getParent()); 2959 2960 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 2961 i != e; ++i) 2962 if (i.getCaseSuccessor() == MaxBlock) { 2963 SI->removeCase(i); 2964 --i; --e; 2965 } 2966 } 2967 } 2968 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2969 if (II->getUnwindDest() == BB) { 2970 // Convert the invoke to a call instruction. This would be a good 2971 // place to note that the call does not throw though. 2972 BranchInst *BI = Builder.CreateBr(II->getNormalDest()); 2973 II->removeFromParent(); // Take out of symbol table 2974 2975 // Insert the call now... 2976 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); 2977 Builder.SetInsertPoint(BI); 2978 CallInst *CI = Builder.CreateCall(II->getCalledValue(), 2979 Args, II->getName()); 2980 CI->setCallingConv(II->getCallingConv()); 2981 CI->setAttributes(II->getAttributes()); 2982 // If the invoke produced a value, the call does now instead. 2983 II->replaceAllUsesWith(CI); 2984 delete II; 2985 Changed = true; 2986 } 2987 } 2988 } 2989 2990 // If this block is now dead, remove it. 2991 if (pred_begin(BB) == pred_end(BB) && 2992 BB != &BB->getParent()->getEntryBlock()) { 2993 // We know there are no successors, so just nuke the block. 2994 BB->eraseFromParent(); 2995 return true; 2996 } 2997 2998 return Changed; 2999} 3000 3001/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a 3002/// integer range comparison into a sub, an icmp and a branch. 3003static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { 3004 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3005 3006 // Make sure all cases point to the same destination and gather the values. 3007 SmallVector<ConstantInt *, 16> Cases; 3008 SwitchInst::CaseIt I = SI->case_begin(); 3009 Cases.push_back(I.getCaseValue()); 3010 SwitchInst::CaseIt PrevI = I++; 3011 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) { 3012 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor()) 3013 return false; 3014 Cases.push_back(I.getCaseValue()); 3015 } 3016 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered"); 3017 3018 // Sort the case values, then check if they form a range we can transform. 3019 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate); 3020 for (unsigned I = 1, E = Cases.size(); I != E; ++I) { 3021 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1) 3022 return false; 3023 } 3024 3025 Constant *Offset = ConstantExpr::getNeg(Cases.back()); 3026 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()); 3027 3028 Value *Sub = SI->getCondition(); 3029 if (!Offset->isNullValue()) 3030 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); 3031 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); 3032 BranchInst *NewBI = Builder.CreateCondBr( 3033 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); 3034 3035 // Update weight for the newly-created conditional branch. 3036 SmallVector<uint64_t, 8> Weights; 3037 bool HasWeights = HasBranchWeights(SI); 3038 if (HasWeights) { 3039 GetBranchWeights(SI, Weights); 3040 if (Weights.size() == 1 + SI->getNumCases()) { 3041 // Combine all weights for the cases to be the true weight of NewBI. 3042 // We assume that the sum of all weights for a Terminator can fit into 32 3043 // bits. 3044 uint32_t NewTrueWeight = 0; 3045 for (unsigned I = 1, E = Weights.size(); I != E; ++I) 3046 NewTrueWeight += (uint32_t)Weights[I]; 3047 NewBI->setMetadata(LLVMContext::MD_prof, 3048 MDBuilder(SI->getContext()). 3049 createBranchWeights(NewTrueWeight, 3050 (uint32_t)Weights[0])); 3051 } 3052 } 3053 3054 // Prune obsolete incoming values off the successor's PHI nodes. 3055 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); 3056 isa<PHINode>(BBI); ++BBI) { 3057 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I) 3058 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent()); 3059 } 3060 SI->eraseFromParent(); 3061 3062 return true; 3063} 3064 3065/// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch 3066/// and use it to remove dead cases. 3067static bool EliminateDeadSwitchCases(SwitchInst *SI) { 3068 Value *Cond = SI->getCondition(); 3069 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth(); 3070 APInt KnownZero(Bits, 0), KnownOne(Bits, 0); 3071 ComputeMaskedBits(Cond, KnownZero, KnownOne); 3072 3073 // Gather dead cases. 3074 SmallVector<ConstantInt*, 8> DeadCases; 3075 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3076 if ((I.getCaseValue()->getValue() & KnownZero) != 0 || 3077 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) { 3078 DeadCases.push_back(I.getCaseValue()); 3079 DEBUG(dbgs() << "SimplifyCFG: switch case '" 3080 << I.getCaseValue() << "' is dead.\n"); 3081 } 3082 } 3083 3084 SmallVector<uint64_t, 8> Weights; 3085 bool HasWeight = HasBranchWeights(SI); 3086 if (HasWeight) { 3087 GetBranchWeights(SI, Weights); 3088 HasWeight = (Weights.size() == 1 + SI->getNumCases()); 3089 } 3090 3091 // Remove dead cases from the switch. 3092 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { 3093 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); 3094 assert(Case != SI->case_default() && 3095 "Case was not found. Probably mistake in DeadCases forming."); 3096 if (HasWeight) { 3097 std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); 3098 Weights.pop_back(); 3099 } 3100 3101 // Prune unused values from PHI nodes. 3102 Case.getCaseSuccessor()->removePredecessor(SI->getParent()); 3103 SI->removeCase(Case); 3104 } 3105 if (HasWeight) { 3106 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); 3107 SI->setMetadata(LLVMContext::MD_prof, 3108 MDBuilder(SI->getParent()->getContext()). 3109 createBranchWeights(MDWeights)); 3110 } 3111 3112 return !DeadCases.empty(); 3113} 3114 3115/// FindPHIForConditionForwarding - If BB would be eligible for simplification 3116/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated 3117/// by an unconditional branch), look at the phi node for BB in the successor 3118/// block and see if the incoming value is equal to CaseValue. If so, return 3119/// the phi node, and set PhiIndex to BB's index in the phi node. 3120static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, 3121 BasicBlock *BB, 3122 int *PhiIndex) { 3123 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) 3124 return NULL; // BB must be empty to be a candidate for simplification. 3125 if (!BB->getSinglePredecessor()) 3126 return NULL; // BB must be dominated by the switch. 3127 3128 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator()); 3129 if (!Branch || !Branch->isUnconditional()) 3130 return NULL; // Terminator must be unconditional branch. 3131 3132 BasicBlock *Succ = Branch->getSuccessor(0); 3133 3134 BasicBlock::iterator I = Succ->begin(); 3135 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3136 int Idx = PHI->getBasicBlockIndex(BB); 3137 assert(Idx >= 0 && "PHI has no entry for predecessor?"); 3138 3139 Value *InValue = PHI->getIncomingValue(Idx); 3140 if (InValue != CaseValue) continue; 3141 3142 *PhiIndex = Idx; 3143 return PHI; 3144 } 3145 3146 return NULL; 3147} 3148 3149/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch 3150/// instruction to a phi node dominated by the switch, if that would mean that 3151/// some of the destination blocks of the switch can be folded away. 3152/// Returns true if a change is made. 3153static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { 3154 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap; 3155 ForwardingNodesMap ForwardingNodes; 3156 3157 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) { 3158 ConstantInt *CaseValue = I.getCaseValue(); 3159 BasicBlock *CaseDest = I.getCaseSuccessor(); 3160 3161 int PhiIndex; 3162 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest, 3163 &PhiIndex); 3164 if (!PHI) continue; 3165 3166 ForwardingNodes[PHI].push_back(PhiIndex); 3167 } 3168 3169 bool Changed = false; 3170 3171 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), 3172 E = ForwardingNodes.end(); I != E; ++I) { 3173 PHINode *Phi = I->first; 3174 SmallVector<int,4> &Indexes = I->second; 3175 3176 if (Indexes.size() < 2) continue; 3177 3178 for (size_t I = 0, E = Indexes.size(); I != E; ++I) 3179 Phi->setIncomingValue(Indexes[I], SI->getCondition()); 3180 Changed = true; 3181 } 3182 3183 return Changed; 3184} 3185 3186/// ValidLookupTableConstant - Return true if the backend will be able to handle 3187/// initializing an array of constants like C. 3188static bool ValidLookupTableConstant(Constant *C) { 3189 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 3190 return CE->isGEPWithNoNotionalOverIndexing(); 3191 3192 return isa<ConstantFP>(C) || 3193 isa<ConstantInt>(C) || 3194 isa<ConstantPointerNull>(C) || 3195 isa<GlobalValue>(C) || 3196 isa<UndefValue>(C); 3197} 3198 3199/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up 3200/// its constant value in ConstantPool, returning 0 if it's not there. 3201static Constant *LookupConstant(Value *V, 3202 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3203 if (Constant *C = dyn_cast<Constant>(V)) 3204 return C; 3205 return ConstantPool.lookup(V); 3206} 3207 3208/// ConstantFold - Try to fold instruction I into a constant. This works for 3209/// simple instructions such as binary operations where both operands are 3210/// constant or can be replaced by constants from the ConstantPool. Returns the 3211/// resulting constant on success, 0 otherwise. 3212static Constant *ConstantFold(Instruction *I, 3213 const SmallDenseMap<Value*, Constant*>& ConstantPool) { 3214 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3215 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); 3216 if (!A) 3217 return 0; 3218 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); 3219 if (!B) 3220 return 0; 3221 return ConstantExpr::get(BO->getOpcode(), A, B); 3222 } 3223 3224 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 3225 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3226 if (!A) 3227 return 0; 3228 Constant *B = LookupConstant(I->getOperand(1), ConstantPool); 3229 if (!B) 3230 return 0; 3231 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); 3232 } 3233 3234 if (SelectInst *Select = dyn_cast<SelectInst>(I)) { 3235 Constant *A = LookupConstant(Select->getCondition(), ConstantPool); 3236 if (!A) 3237 return 0; 3238 if (A->isAllOnesValue()) 3239 return LookupConstant(Select->getTrueValue(), ConstantPool); 3240 if (A->isNullValue()) 3241 return LookupConstant(Select->getFalseValue(), ConstantPool); 3242 return 0; 3243 } 3244 3245 if (CastInst *Cast = dyn_cast<CastInst>(I)) { 3246 Constant *A = LookupConstant(I->getOperand(0), ConstantPool); 3247 if (!A) 3248 return 0; 3249 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); 3250 } 3251 3252 return 0; 3253} 3254 3255/// GetCaseResults - Try to determine the resulting constant values in phi nodes 3256/// at the common destination basic block, *CommonDest, for one of the case 3257/// destionations CaseDest corresponding to value CaseVal (0 for the default 3258/// case), of a switch instruction SI. 3259static bool GetCaseResults(SwitchInst *SI, 3260 ConstantInt *CaseVal, 3261 BasicBlock *CaseDest, 3262 BasicBlock **CommonDest, 3263 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { 3264 // The block from which we enter the common destination. 3265 BasicBlock *Pred = SI->getParent(); 3266 3267 // If CaseDest is empty except for some side-effect free instructions through 3268 // which we can constant-propagate the CaseVal, continue to its successor. 3269 SmallDenseMap<Value*, Constant*> ConstantPool; 3270 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); 3271 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; 3272 ++I) { 3273 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { 3274 // If the terminator is a simple branch, continue to the next block. 3275 if (T->getNumSuccessors() != 1) 3276 return false; 3277 Pred = CaseDest; 3278 CaseDest = T->getSuccessor(0); 3279 } else if (isa<DbgInfoIntrinsic>(I)) { 3280 // Skip debug intrinsic. 3281 continue; 3282 } else if (Constant *C = ConstantFold(I, ConstantPool)) { 3283 // Instruction is side-effect free and constant. 3284 ConstantPool.insert(std::make_pair(I, C)); 3285 } else { 3286 break; 3287 } 3288 } 3289 3290 // If we did not have a CommonDest before, use the current one. 3291 if (!*CommonDest) 3292 *CommonDest = CaseDest; 3293 // If the destination isn't the common one, abort. 3294 if (CaseDest != *CommonDest) 3295 return false; 3296 3297 // Get the values for this case from phi nodes in the destination block. 3298 BasicBlock::iterator I = (*CommonDest)->begin(); 3299 while (PHINode *PHI = dyn_cast<PHINode>(I++)) { 3300 int Idx = PHI->getBasicBlockIndex(Pred); 3301 if (Idx == -1) 3302 continue; 3303 3304 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), 3305 ConstantPool); 3306 if (!ConstVal) 3307 return false; 3308 3309 // Note: If the constant comes from constant-propagating the case value 3310 // through the CaseDest basic block, it will be safe to remove the 3311 // instructions in that block. They cannot be used (except in the phi nodes 3312 // we visit) outside CaseDest, because that block does not dominate its 3313 // successor. If it did, we would not be in this phi node. 3314 3315 // Be conservative about which kinds of constants we support. 3316 if (!ValidLookupTableConstant(ConstVal)) 3317 return false; 3318 3319 Res.push_back(std::make_pair(PHI, ConstVal)); 3320 } 3321 3322 return true; 3323} 3324 3325namespace { 3326 /// SwitchLookupTable - This class represents a lookup table that can be used 3327 /// to replace a switch. 3328 class SwitchLookupTable { 3329 public: 3330 /// SwitchLookupTable - Create a lookup table to use as a switch replacement 3331 /// with the contents of Values, using DefaultValue to fill any holes in the 3332 /// table. 3333 SwitchLookupTable(Module &M, 3334 uint64_t TableSize, 3335 ConstantInt *Offset, 3336 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3337 Constant *DefaultValue, 3338 const DataLayout *TD); 3339 3340 /// BuildLookup - Build instructions with Builder to retrieve the value at 3341 /// the position given by Index in the lookup table. 3342 Value *BuildLookup(Value *Index, IRBuilder<> &Builder); 3343 3344 /// WouldFitInRegister - Return true if a table with TableSize elements of 3345 /// type ElementType would fit in a target-legal register. 3346 static bool WouldFitInRegister(const DataLayout *TD, 3347 uint64_t TableSize, 3348 const Type *ElementType); 3349 3350 private: 3351 // Depending on the contents of the table, it can be represented in 3352 // different ways. 3353 enum { 3354 // For tables where each element contains the same value, we just have to 3355 // store that single value and return it for each lookup. 3356 SingleValueKind, 3357 3358 // For small tables with integer elements, we can pack them into a bitmap 3359 // that fits into a target-legal register. Values are retrieved by 3360 // shift and mask operations. 3361 BitMapKind, 3362 3363 // The table is stored as an array of values. Values are retrieved by load 3364 // instructions from the table. 3365 ArrayKind 3366 } Kind; 3367 3368 // For SingleValueKind, this is the single value. 3369 Constant *SingleValue; 3370 3371 // For BitMapKind, this is the bitmap. 3372 ConstantInt *BitMap; 3373 IntegerType *BitMapElementTy; 3374 3375 // For ArrayKind, this is the array. 3376 GlobalVariable *Array; 3377 }; 3378} 3379 3380SwitchLookupTable::SwitchLookupTable(Module &M, 3381 uint64_t TableSize, 3382 ConstantInt *Offset, 3383 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, 3384 Constant *DefaultValue, 3385 const DataLayout *TD) { 3386 assert(Values.size() && "Can't build lookup table without values!"); 3387 assert(TableSize >= Values.size() && "Can't fit values in table!"); 3388 3389 // If all values in the table are equal, this is that value. 3390 SingleValue = Values.begin()->second; 3391 3392 // Build up the table contents. 3393 SmallVector<Constant*, 64> TableContents(TableSize); 3394 for (size_t I = 0, E = Values.size(); I != E; ++I) { 3395 ConstantInt *CaseVal = Values[I].first; 3396 Constant *CaseRes = Values[I].second; 3397 assert(CaseRes->getType() == DefaultValue->getType()); 3398 3399 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) 3400 .getLimitedValue(); 3401 TableContents[Idx] = CaseRes; 3402 3403 if (CaseRes != SingleValue) 3404 SingleValue = 0; 3405 } 3406 3407 // Fill in any holes in the table with the default result. 3408 if (Values.size() < TableSize) { 3409 for (uint64_t I = 0; I < TableSize; ++I) { 3410 if (!TableContents[I]) 3411 TableContents[I] = DefaultValue; 3412 } 3413 3414 if (DefaultValue != SingleValue) 3415 SingleValue = 0; 3416 } 3417 3418 // If each element in the table contains the same value, we only need to store 3419 // that single value. 3420 if (SingleValue) { 3421 Kind = SingleValueKind; 3422 return; 3423 } 3424 3425 // If the type is integer and the table fits in a register, build a bitmap. 3426 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { 3427 IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); 3428 APInt TableInt(TableSize * IT->getBitWidth(), 0); 3429 for (uint64_t I = TableSize; I > 0; --I) { 3430 TableInt <<= IT->getBitWidth(); 3431 // Insert values into the bitmap. Undef values are set to zero. 3432 if (!isa<UndefValue>(TableContents[I - 1])) { 3433 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); 3434 TableInt |= Val->getValue().zext(TableInt.getBitWidth()); 3435 } 3436 } 3437 BitMap = ConstantInt::get(M.getContext(), TableInt); 3438 BitMapElementTy = IT; 3439 Kind = BitMapKind; 3440 ++NumBitMaps; 3441 return; 3442 } 3443 3444 // Store the table in an array. 3445 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); 3446 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); 3447 3448 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, 3449 GlobalVariable::PrivateLinkage, 3450 Initializer, 3451 "switch.table"); 3452 Array->setUnnamedAddr(true); 3453 Kind = ArrayKind; 3454} 3455 3456Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { 3457 switch (Kind) { 3458 case SingleValueKind: 3459 return SingleValue; 3460 case BitMapKind: { 3461 // Type of the bitmap (e.g. i59). 3462 IntegerType *MapTy = BitMap->getType(); 3463 3464 // Cast Index to the same type as the bitmap. 3465 // Note: The Index is <= the number of elements in the table, so 3466 // truncating it to the width of the bitmask is safe. 3467 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); 3468 3469 // Multiply the shift amount by the element width. 3470 ShiftAmt = Builder.CreateMul(ShiftAmt, 3471 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), 3472 "switch.shiftamt"); 3473 3474 // Shift down. 3475 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, 3476 "switch.downshift"); 3477 // Mask off. 3478 return Builder.CreateTrunc(DownShifted, BitMapElementTy, 3479 "switch.masked"); 3480 } 3481 case ArrayKind: { 3482 Value *GEPIndices[] = { Builder.getInt32(0), Index }; 3483 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, 3484 "switch.gep"); 3485 return Builder.CreateLoad(GEP, "switch.load"); 3486 } 3487 } 3488 llvm_unreachable("Unknown lookup table kind!"); 3489} 3490 3491bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, 3492 uint64_t TableSize, 3493 const Type *ElementType) { 3494 if (!TD) 3495 return false; 3496 const IntegerType *IT = dyn_cast<IntegerType>(ElementType); 3497 if (!IT) 3498 return false; 3499 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values 3500 // are <= 15, we could try to narrow the type. 3501 3502 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. 3503 if (TableSize >= UINT_MAX/IT->getBitWidth()) 3504 return false; 3505 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); 3506} 3507 3508/// ShouldBuildLookupTable - Determine whether a lookup table should be built 3509/// for this switch, based on the number of caes, size of the table and the 3510/// types of the results. 3511static bool ShouldBuildLookupTable(SwitchInst *SI, 3512 uint64_t TableSize, 3513 const DataLayout *TD, 3514 const SmallDenseMap<PHINode*, Type*>& ResultTypes) { 3515 // The table density should be at least 40%. This is the same criterion as for 3516 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. 3517 // FIXME: Find the best cut-off. 3518 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) 3519 return false; // TableSize overflowed, or mul below might overflow. 3520 if (SI->getNumCases() * 10 >= TableSize * 4) 3521 return true; 3522 3523 // If each table would fit in a register, we should build it anyway. 3524 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), 3525 E = ResultTypes.end(); I != E; ++I) { 3526 if (!SwitchLookupTable::WouldFitInRegister(TD, TableSize, I->second)) 3527 return false; 3528 } 3529 return true; 3530} 3531 3532/// SwitchToLookupTable - If the switch is only used to initialize one or more 3533/// phi nodes in a common successor block with different constant values, 3534/// replace the switch with lookup tables. 3535static bool SwitchToLookupTable(SwitchInst *SI, 3536 IRBuilder<> &Builder, 3537 const DataLayout* TD, 3538 const TargetTransformInfo *TTI) { 3539 assert(SI->getNumCases() > 1 && "Degenerate switch?"); 3540 3541 // Only build lookup table when we have a target that supports it. 3542 if (!TTI || !TTI->getScalarTargetTransformInfo()->shouldBuildLookupTables()) 3543 return false; 3544 3545 // FIXME: If the switch is too sparse for a lookup table, perhaps we could 3546 // split off a dense part and build a lookup table for that. 3547 3548 // FIXME: This creates arrays of GEPs to constant strings, which means each 3549 // GEP needs a runtime relocation in PIC code. We should just build one big 3550 // string and lookup indices into that. 3551 3552 // Ignore the switch if the number of cases is too small. 3553 // This is similar to the check when building jump tables in 3554 // SelectionDAGBuilder::handleJTSwitchCase. 3555 // FIXME: Determine the best cut-off. 3556 if (SI->getNumCases() < 4) 3557 return false; 3558 3559 // Figure out the corresponding result for each case value and phi node in the 3560 // common destination, as well as the the min and max case values. 3561 assert(SI->case_begin() != SI->case_end()); 3562 SwitchInst::CaseIt CI = SI->case_begin(); 3563 ConstantInt *MinCaseVal = CI.getCaseValue(); 3564 ConstantInt *MaxCaseVal = CI.getCaseValue(); 3565 3566 BasicBlock *CommonDest = 0; 3567 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; 3568 SmallDenseMap<PHINode*, ResultListTy> ResultLists; 3569 SmallDenseMap<PHINode*, Constant*> DefaultResults; 3570 SmallDenseMap<PHINode*, Type*> ResultTypes; 3571 SmallVector<PHINode*, 4> PHIs; 3572 3573 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { 3574 ConstantInt *CaseVal = CI.getCaseValue(); 3575 if (CaseVal->getValue().slt(MinCaseVal->getValue())) 3576 MinCaseVal = CaseVal; 3577 if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) 3578 MaxCaseVal = CaseVal; 3579 3580 // Resulting value at phi nodes for this case value. 3581 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; 3582 ResultsTy Results; 3583 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, 3584 Results)) 3585 return false; 3586 3587 // Append the result from this case to the list for each phi. 3588 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { 3589 if (!ResultLists.count(I->first)) 3590 PHIs.push_back(I->first); 3591 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); 3592 } 3593 } 3594 3595 // Get the resulting values for the default case. 3596 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; 3597 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, 3598 DefaultResultsList)) 3599 return false; 3600 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { 3601 PHINode *PHI = DefaultResultsList[I].first; 3602 Constant *Result = DefaultResultsList[I].second; 3603 DefaultResults[PHI] = Result; 3604 ResultTypes[PHI] = Result->getType(); 3605 } 3606 3607 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); 3608 uint64_t TableSize = RangeSpread.getLimitedValue() + 1; 3609 if (!ShouldBuildLookupTable(SI, TableSize, TD, ResultTypes)) 3610 return false; 3611 3612 // Create the BB that does the lookups. 3613 Module &Mod = *CommonDest->getParent()->getParent(); 3614 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), 3615 "switch.lookup", 3616 CommonDest->getParent(), 3617 CommonDest); 3618 3619 // Check whether the condition value is within the case range, and branch to 3620 // the new BB. 3621 Builder.SetInsertPoint(SI); 3622 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, 3623 "switch.tableidx"); 3624 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( 3625 MinCaseVal->getType(), TableSize)); 3626 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); 3627 3628 // Populate the BB that does the lookups. 3629 Builder.SetInsertPoint(LookupBB); 3630 bool ReturnedEarly = false; 3631 for (size_t I = 0, E = PHIs.size(); I != E; ++I) { 3632 PHINode *PHI = PHIs[I]; 3633 3634 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], 3635 DefaultResults[PHI], TD); 3636 3637 Value *Result = Table.BuildLookup(TableIndex, Builder); 3638 3639 // If the result is used to return immediately from the function, we want to 3640 // do that right here. 3641 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && 3642 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { 3643 Builder.CreateRet(Result); 3644 ReturnedEarly = true; 3645 break; 3646 } 3647 3648 PHI->addIncoming(Result, LookupBB); 3649 } 3650 3651 if (!ReturnedEarly) 3652 Builder.CreateBr(CommonDest); 3653 3654 // Remove the switch. 3655 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { 3656 BasicBlock *Succ = SI->getSuccessor(i); 3657 if (Succ == SI->getDefaultDest()) continue; 3658 Succ->removePredecessor(SI->getParent()); 3659 } 3660 SI->eraseFromParent(); 3661 3662 ++NumLookupTables; 3663 return true; 3664} 3665 3666bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { 3667 BasicBlock *BB = SI->getParent(); 3668 3669 if (isValueEqualityComparison(SI)) { 3670 // If we only have one predecessor, and if it is a branch on this value, 3671 // see if that predecessor totally determines the outcome of this switch. 3672 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3673 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) 3674 return SimplifyCFG(BB) | true; 3675 3676 Value *Cond = SI->getCondition(); 3677 if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) 3678 if (SimplifySwitchOnSelect(SI, Select)) 3679 return SimplifyCFG(BB) | true; 3680 3681 // If the block only contains the switch, see if we can fold the block 3682 // away into any preds. 3683 BasicBlock::iterator BBI = BB->begin(); 3684 // Ignore dbg intrinsics. 3685 while (isa<DbgInfoIntrinsic>(BBI)) 3686 ++BBI; 3687 if (SI == &*BBI) 3688 if (FoldValueComparisonIntoPredecessors(SI, Builder)) 3689 return SimplifyCFG(BB) | true; 3690 } 3691 3692 // Try to transform the switch into an icmp and a branch. 3693 if (TurnSwitchRangeIntoICmp(SI, Builder)) 3694 return SimplifyCFG(BB) | true; 3695 3696 // Remove unreachable cases. 3697 if (EliminateDeadSwitchCases(SI)) 3698 return SimplifyCFG(BB) | true; 3699 3700 if (ForwardSwitchConditionToPHI(SI)) 3701 return SimplifyCFG(BB) | true; 3702 3703 if (SwitchToLookupTable(SI, Builder, TD, TTI)) 3704 return SimplifyCFG(BB) | true; 3705 3706 return false; 3707} 3708 3709bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { 3710 BasicBlock *BB = IBI->getParent(); 3711 bool Changed = false; 3712 3713 // Eliminate redundant destinations. 3714 SmallPtrSet<Value *, 8> Succs; 3715 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 3716 BasicBlock *Dest = IBI->getDestination(i); 3717 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) { 3718 Dest->removePredecessor(BB); 3719 IBI->removeDestination(i); 3720 --i; --e; 3721 Changed = true; 3722 } 3723 } 3724 3725 if (IBI->getNumDestinations() == 0) { 3726 // If the indirectbr has no successors, change it to unreachable. 3727 new UnreachableInst(IBI->getContext(), IBI); 3728 EraseTerminatorInstAndDCECond(IBI); 3729 return true; 3730 } 3731 3732 if (IBI->getNumDestinations() == 1) { 3733 // If the indirectbr has one successor, change it to a direct branch. 3734 BranchInst::Create(IBI->getDestination(0), IBI); 3735 EraseTerminatorInstAndDCECond(IBI); 3736 return true; 3737 } 3738 3739 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { 3740 if (SimplifyIndirectBrOnSelect(IBI, SI)) 3741 return SimplifyCFG(BB) | true; 3742 } 3743 return Changed; 3744} 3745 3746bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ 3747 BasicBlock *BB = BI->getParent(); 3748 3749 if (SinkCommon && SinkThenElseCodeToEnd(BI)) 3750 return true; 3751 3752 // If the Terminator is the only non-phi instruction, simplify the block. 3753 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); 3754 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && 3755 TryToSimplifyUncondBranchFromEmptyBlock(BB)) 3756 return true; 3757 3758 // If the only instruction in the block is a seteq/setne comparison 3759 // against a constant, try to simplify the block. 3760 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) 3761 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) { 3762 for (++I; isa<DbgInfoIntrinsic>(I); ++I) 3763 ; 3764 if (I->isTerminator() && 3765 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) 3766 return true; 3767 } 3768 3769 // If this basic block is ONLY a compare and a branch, and if a predecessor 3770 // branches to us and our successor, fold the comparison into the 3771 // predecessor and use logical operations to update the incoming value 3772 // for PHI nodes in common successor. 3773 if (FoldBranchToCommonDest(BI)) 3774 return SimplifyCFG(BB) | true; 3775 return false; 3776} 3777 3778 3779bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { 3780 BasicBlock *BB = BI->getParent(); 3781 3782 // Conditional branch 3783 if (isValueEqualityComparison(BI)) { 3784 // If we only have one predecessor, and if it is a branch on this value, 3785 // see if that predecessor totally determines the outcome of this 3786 // switch. 3787 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 3788 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) 3789 return SimplifyCFG(BB) | true; 3790 3791 // This block must be empty, except for the setcond inst, if it exists. 3792 // Ignore dbg intrinsics. 3793 BasicBlock::iterator I = BB->begin(); 3794 // Ignore dbg intrinsics. 3795 while (isa<DbgInfoIntrinsic>(I)) 3796 ++I; 3797 if (&*I == BI) { 3798 if (FoldValueComparisonIntoPredecessors(BI, Builder)) 3799 return SimplifyCFG(BB) | true; 3800 } else if (&*I == cast<Instruction>(BI->getCondition())){ 3801 ++I; 3802 // Ignore dbg intrinsics. 3803 while (isa<DbgInfoIntrinsic>(I)) 3804 ++I; 3805 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder)) 3806 return SimplifyCFG(BB) | true; 3807 } 3808 } 3809 3810 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. 3811 if (SimplifyBranchOnICmpChain(BI, TD, Builder)) 3812 return true; 3813 3814 // If this basic block is ONLY a compare and a branch, and if a predecessor 3815 // branches to us and one of our successors, fold the comparison into the 3816 // predecessor and use logical operations to pick the right destination. 3817 if (FoldBranchToCommonDest(BI)) 3818 return SimplifyCFG(BB) | true; 3819 3820 // We have a conditional branch to two blocks that are only reachable 3821 // from BI. We know that the condbr dominates the two blocks, so see if 3822 // there is any identical code in the "then" and "else" blocks. If so, we 3823 // can hoist it up to the branching block. 3824 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) { 3825 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3826 if (HoistThenElseCodeToIf(BI)) 3827 return SimplifyCFG(BB) | true; 3828 } else { 3829 // If Successor #1 has multiple preds, we may be able to conditionally 3830 // execute Successor #0 if it branches to successor #1. 3831 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator(); 3832 if (Succ0TI->getNumSuccessors() == 1 && 3833 Succ0TI->getSuccessor(0) == BI->getSuccessor(1)) 3834 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0))) 3835 return SimplifyCFG(BB) | true; 3836 } 3837 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) { 3838 // If Successor #0 has multiple preds, we may be able to conditionally 3839 // execute Successor #1 if it branches to successor #0. 3840 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator(); 3841 if (Succ1TI->getNumSuccessors() == 1 && 3842 Succ1TI->getSuccessor(0) == BI->getSuccessor(0)) 3843 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) 3844 return SimplifyCFG(BB) | true; 3845 } 3846 3847 // If this is a branch on a phi node in the current block, thread control 3848 // through this block if any PHI node entries are constants. 3849 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 3850 if (PN->getParent() == BI->getParent()) 3851 if (FoldCondBranchOnPHI(BI, TD)) 3852 return SimplifyCFG(BB) | true; 3853 3854 // Scan predecessor blocks for conditional branches. 3855 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 3856 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 3857 if (PBI != BI && PBI->isConditional()) 3858 if (SimplifyCondBranchToCondBranch(PBI, BI)) 3859 return SimplifyCFG(BB) | true; 3860 3861 return false; 3862} 3863 3864/// Check if passing a value to an instruction will cause undefined behavior. 3865static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { 3866 Constant *C = dyn_cast<Constant>(V); 3867 if (!C) 3868 return false; 3869 3870 if (I->use_empty()) 3871 return false; 3872 3873 if (C->isNullValue()) { 3874 // Only look at the first use, avoid hurting compile time with long uselists 3875 User *Use = *I->use_begin(); 3876 3877 // Now make sure that there are no instructions in between that can alter 3878 // control flow (eg. calls) 3879 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i) 3880 if (i == I->getParent()->end() || i->mayHaveSideEffects()) 3881 return false; 3882 3883 // Look through GEPs. A load from a GEP derived from NULL is still undefined 3884 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use)) 3885 if (GEP->getPointerOperand() == I) 3886 return passingValueIsAlwaysUndefined(V, GEP); 3887 3888 // Look through bitcasts. 3889 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use)) 3890 return passingValueIsAlwaysUndefined(V, BC); 3891 3892 // Load from null is undefined. 3893 if (LoadInst *LI = dyn_cast<LoadInst>(Use)) 3894 return LI->getPointerAddressSpace() == 0; 3895 3896 // Store to null is undefined. 3897 if (StoreInst *SI = dyn_cast<StoreInst>(Use)) 3898 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I; 3899 } 3900 return false; 3901} 3902 3903/// If BB has an incoming value that will always trigger undefined behavior 3904/// (eg. null pointer dereference), remove the branch leading here. 3905static bool removeUndefIntroducingPredecessor(BasicBlock *BB) { 3906 for (BasicBlock::iterator i = BB->begin(); 3907 PHINode *PHI = dyn_cast<PHINode>(i); ++i) 3908 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) 3909 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) { 3910 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator(); 3911 IRBuilder<> Builder(T); 3912 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 3913 BB->removePredecessor(PHI->getIncomingBlock(i)); 3914 // Turn uncoditional branches into unreachables and remove the dead 3915 // destination from conditional branches. 3916 if (BI->isUnconditional()) 3917 Builder.CreateUnreachable(); 3918 else 3919 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) : 3920 BI->getSuccessor(0)); 3921 BI->eraseFromParent(); 3922 return true; 3923 } 3924 // TODO: SwitchInst. 3925 } 3926 3927 return false; 3928} 3929 3930bool SimplifyCFGOpt::run(BasicBlock *BB) { 3931 bool Changed = false; 3932 3933 assert(BB && BB->getParent() && "Block not embedded in function!"); 3934 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 3935 3936 // Remove basic blocks that have no predecessors (except the entry block)... 3937 // or that just have themself as a predecessor. These are unreachable. 3938 if ((pred_begin(BB) == pred_end(BB) && 3939 BB != &BB->getParent()->getEntryBlock()) || 3940 BB->getSinglePredecessor() == BB) { 3941 DEBUG(dbgs() << "Removing BB: \n" << *BB); 3942 DeleteDeadBlock(BB); 3943 return true; 3944 } 3945 3946 // Check to see if we can constant propagate this terminator instruction 3947 // away... 3948 Changed |= ConstantFoldTerminator(BB, true); 3949 3950 // Check for and eliminate duplicate PHI nodes in this block. 3951 Changed |= EliminateDuplicatePHINodes(BB); 3952 3953 // Check for and remove branches that will always cause undefined behavior. 3954 Changed |= removeUndefIntroducingPredecessor(BB); 3955 3956 // Merge basic blocks into their predecessor if there is only one distinct 3957 // pred, and if there is only one distinct successor of the predecessor, and 3958 // if there are no PHI nodes. 3959 // 3960 if (MergeBlockIntoPredecessor(BB)) 3961 return true; 3962 3963 IRBuilder<> Builder(BB); 3964 3965 // If there is a trivial two-entry PHI node in this basic block, and we can 3966 // eliminate it, do so now. 3967 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 3968 if (PN->getNumIncomingValues() == 2) 3969 Changed |= FoldTwoEntryPHINode(PN, TD); 3970 3971 Builder.SetInsertPoint(BB->getTerminator()); 3972 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 3973 if (BI->isUnconditional()) { 3974 if (SimplifyUncondBranch(BI, Builder)) return true; 3975 } else { 3976 if (SimplifyCondBranch(BI, Builder)) return true; 3977 } 3978 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 3979 if (SimplifyReturn(RI, Builder)) return true; 3980 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 3981 if (SimplifyResume(RI, Builder)) return true; 3982 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 3983 if (SimplifySwitch(SI, Builder)) return true; 3984 } else if (UnreachableInst *UI = 3985 dyn_cast<UnreachableInst>(BB->getTerminator())) { 3986 if (SimplifyUnreachable(UI)) return true; 3987 } else if (IndirectBrInst *IBI = 3988 dyn_cast<IndirectBrInst>(BB->getTerminator())) { 3989 if (SimplifyIndirectBr(IBI)) return true; 3990 } 3991 3992 return Changed; 3993} 3994 3995/// SimplifyCFG - This function is used to do simplification of a CFG. For 3996/// example, it adjusts branches to branches to eliminate the extra hop, it 3997/// eliminates unreachable basic blocks, and does other "peephole" optimization 3998/// of the CFG. It returns true if a modification was made. 3999/// 4000bool llvm::SimplifyCFG(BasicBlock *BB, const DataLayout *TD, 4001 const TargetTransformInfo *TTI) { 4002 return SimplifyCFGOpt(TD, TTI).run(BB); 4003} 4004