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/Instructions.h" 18 #include "llvm/IntrinsicInst.h" 19 #include "llvm/Type.h" 20 #include "llvm/DerivedTypes.h" 21 #include "llvm/GlobalVariable.h" 22 #include "llvm/Support/CFG.h" 23 #include "llvm/Support/Debug.h" 24 #include "llvm/Analysis/ConstantFolding.h" 25 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 26 #include "llvm/ADT/SmallVector.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/Statistic.h" 29 #include <algorithm> 30 #include <functional> 31 #include <set> 32 #include <map> 33 using namespace llvm; 34 35 STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 36 37 /// SafeToMergeTerminators - Return true if it is safe to merge these two 38 /// terminator instructions together. 39 /// 40 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 41 if (SI1 == SI2) return false; // Can't merge with self! 42 43 // It is not safe to merge these two switch instructions if they have a common 44 // successor, and if that successor has a PHI node, and if *that* PHI node has 45 // conflicting incoming values from the two switch blocks. 46 BasicBlock *SI1BB = SI1->getParent(); 47 BasicBlock *SI2BB = SI2->getParent(); 48 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 49 50 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 51 if (SI1Succs.count(*I)) 52 for (BasicBlock::iterator BBI = (*I)->begin(); 53 isa<PHINode>(BBI); ++BBI) { 54 PHINode *PN = cast<PHINode>(BBI); 55 if (PN->getIncomingValueForBlock(SI1BB) != 56 PN->getIncomingValueForBlock(SI2BB)) 57 return false; 58 } 59 60 return true; 61 } 62 63 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 64 /// now be entries in it from the 'NewPred' block. The values that will be 65 /// flowing into the PHI nodes will be the same as those coming in from 66 /// ExistPred, an existing predecessor of Succ. 67 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 68 BasicBlock *ExistPred) { 69 assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != 70 succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); 71 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 72 73 PHINode *PN; 74 for (BasicBlock::iterator I = Succ->begin(); 75 (PN = dyn_cast<PHINode>(I)); ++I) 76 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 77 } 78 79 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 80 /// almost-empty BB ending in an unconditional branch to Succ, into succ. 81 /// 82 /// Assumption: Succ is the single successor for BB. 83 /// 84 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 85 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 86 87 DOUT << "Looking to fold " << BB->getNameStart() << " into " 88 << Succ->getNameStart() << "\n"; 89 // Shortcut, if there is only a single predecessor it must be BB and merging 90 // is always safe 91 if (Succ->getSinglePredecessor()) return true; 92 93 typedef SmallPtrSet<Instruction*, 16> InstrSet; 94 InstrSet BBPHIs; 95 96 // Make a list of all phi nodes in BB 97 BasicBlock::iterator BBI = BB->begin(); 98 while (isa<PHINode>(*BBI)) BBPHIs.insert(BBI++); 99 100 // Make a list of the predecessors of BB 101 typedef SmallPtrSet<BasicBlock*, 16> BlockSet; 102 BlockSet BBPreds(pred_begin(BB), pred_end(BB)); 103 104 // Use that list to make another list of common predecessors of BB and Succ 105 BlockSet CommonPreds; 106 for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); 107 PI != PE; ++PI) 108 if (BBPreds.count(*PI)) 109 CommonPreds.insert(*PI); 110 111 // Shortcut, if there are no common predecessors, merging is always safe 112 if (CommonPreds.empty()) 113 return true; 114 115 // Look at all the phi nodes in Succ, to see if they present a conflict when 116 // merging these blocks 117 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 118 PHINode *PN = cast<PHINode>(I); 119 120 // If the incoming value from BB is again a PHINode in 121 // BB which has the same incoming value for *PI as PN does, we can 122 // merge the phi nodes and then the blocks can still be merged 123 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 124 if (BBPN && BBPN->getParent() == BB) { 125 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 126 PI != PE; PI++) { 127 if (BBPN->getIncomingValueForBlock(*PI) 128 != PN->getIncomingValueForBlock(*PI)) { 129 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 130 << Succ->getNameStart() << " is conflicting with " 131 << BBPN->getNameStart() << " with regard to common predecessor " 132 << (*PI)->getNameStart() << "\n"; 133 return false; 134 } 135 } 136 // Remove this phinode from the list of phis in BB, since it has been 137 // handled. 138 BBPHIs.erase(BBPN); 139 } else { 140 Value* Val = PN->getIncomingValueForBlock(BB); 141 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 142 PI != PE; PI++) { 143 // See if the incoming value for the common predecessor is equal to the 144 // one for BB, in which case this phi node will not prevent the merging 145 // of the block. 146 if (Val != PN->getIncomingValueForBlock(*PI)) { 147 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 148 << Succ->getNameStart() << " is conflicting with regard to common " 149 << "predecessor " << (*PI)->getNameStart() << "\n"; 150 return false; 151 } 152 } 153 } 154 } 155 156 // If there are any other phi nodes in BB that don't have a phi node in Succ 157 // to merge with, they must be moved to Succ completely. However, for any 158 // predecessors of Succ, branches will be added to the phi node that just 159 // point to itself. So, for any common predecessors, this must not cause 160 // conflicts. 161 for (InstrSet::iterator I = BBPHIs.begin(), E = BBPHIs.end(); 162 I != E; I++) { 163 PHINode *PN = cast<PHINode>(*I); 164 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 165 PI != PE; PI++) 166 if (PN->getIncomingValueForBlock(*PI) != PN) { 167 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 168 << BB->getNameStart() << " is conflicting with regard to common " 169 << "predecessor " << (*PI)->getNameStart() << "\n"; 170 return false; 171 } 172 } 173 174 return true; 175 } 176 177 /// TryToSimplifyUncondBranchFromEmptyBlock - BB contains an unconditional 178 /// branch to Succ, and contains no instructions other than PHI nodes and the 179 /// branch. If possible, eliminate BB. 180 static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, 181 BasicBlock *Succ) { 182 // Check to see if merging these blocks would cause conflicts for any of the 183 // phi nodes in BB or Succ. If not, we can safely merge. 184 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 185 186 DOUT << "Killing Trivial BB: \n" << *BB; 187 188 if (isa<PHINode>(Succ->begin())) { 189 // If there is more than one pred of succ, and there are PHI nodes in 190 // the successor, then we need to add incoming edges for the PHI nodes 191 // 192 const SmallVector<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 193 194 // Loop over all of the PHI nodes in the successor of BB. 195 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 196 PHINode *PN = cast<PHINode>(I); 197 Value *OldVal = PN->removeIncomingValue(BB, false); 198 assert(OldVal && "No entry in PHI for Pred BB!"); 199 200 // If this incoming value is one of the PHI nodes in BB, the new entries 201 // in the PHI node are the entries from the old PHI. 202 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 203 PHINode *OldValPN = cast<PHINode>(OldVal); 204 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) 205 // Note that, since we are merging phi nodes and BB and Succ might 206 // have common predecessors, we could end up with a phi node with 207 // identical incoming branches. This will be cleaned up later (and 208 // will trigger asserts if we try to clean it up now, without also 209 // simplifying the corresponding conditional branch). 210 PN->addIncoming(OldValPN->getIncomingValue(i), 211 OldValPN->getIncomingBlock(i)); 212 } else { 213 // Add an incoming value for each of the new incoming values. 214 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) 215 PN->addIncoming(OldVal, BBPreds[i]); 216 } 217 } 218 } 219 220 if (isa<PHINode>(&BB->front())) { 221 SmallVector<BasicBlock*, 16> 222 OldSuccPreds(pred_begin(Succ), pred_end(Succ)); 223 224 // Move all PHI nodes in BB to Succ if they are alive, otherwise 225 // delete them. 226 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 227 if (PN->use_empty()) { 228 // Just remove the dead phi. This happens if Succ's PHIs were the only 229 // users of the PHI nodes. 230 PN->eraseFromParent(); 231 continue; 232 } 233 234 // The instruction is alive, so this means that BB must dominate all 235 // predecessors of Succ (Since all uses of the PN are after its 236 // definition, so in Succ or a block dominated by Succ. If a predecessor 237 // of Succ would not be dominated by BB, PN would violate the def before 238 // use SSA demand). Therefore, we can simply move the phi node to the 239 // next block. 240 Succ->getInstList().splice(Succ->begin(), 241 BB->getInstList(), BB->begin()); 242 243 // We need to add new entries for the PHI node to account for 244 // predecessors of Succ that the PHI node does not take into 245 // account. At this point, since we know that BB dominated succ and all 246 // of its predecessors, this means that we should any newly added 247 // incoming edges should use the PHI node itself as the value for these 248 // edges, because they are loop back edges. 249 for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i) 250 if (OldSuccPreds[i] != BB) 251 PN->addIncoming(PN, OldSuccPreds[i]); 252 } 253 } 254 255 // Everything that jumped to BB now goes to Succ. 256 BB->replaceAllUsesWith(Succ); 257 if (!Succ->hasName()) Succ->takeName(BB); 258 BB->eraseFromParent(); // Delete the old basic block. 259 return true; 260 } 261 262 /// GetIfCondition - Given a basic block (BB) with two predecessors (and 263 /// presumably PHI nodes in it), check to see if the merge at this block is due 264 /// to an "if condition". If so, return the boolean condition that determines 265 /// which entry into BB will be taken. Also, return by references the block 266 /// that will be entered from if the condition is true, and the block that will 267 /// be entered if the condition is false. 268 /// 269 /// 270 static Value *GetIfCondition(BasicBlock *BB, 271 BasicBlock *&IfTrue, BasicBlock *&IfFalse) { 272 assert(std::distance(pred_begin(BB), pred_end(BB)) == 2 && 273 "Function can only handle blocks with 2 predecessors!"); 274 BasicBlock *Pred1 = *pred_begin(BB); 275 BasicBlock *Pred2 = *++pred_begin(BB); 276 277 // We can only handle branches. Other control flow will be lowered to 278 // branches if possible anyway. 279 if (!isa<BranchInst>(Pred1->getTerminator()) || 280 !isa<BranchInst>(Pred2->getTerminator())) 281 return 0; 282 BranchInst *Pred1Br = cast<BranchInst>(Pred1->getTerminator()); 283 BranchInst *Pred2Br = cast<BranchInst>(Pred2->getTerminator()); 284 285 // Eliminate code duplication by ensuring that Pred1Br is conditional if 286 // either are. 287 if (Pred2Br->isConditional()) { 288 // If both branches are conditional, we don't have an "if statement". In 289 // reality, we could transform this case, but since the condition will be 290 // required anyway, we stand no chance of eliminating it, so the xform is 291 // probably not profitable. 292 if (Pred1Br->isConditional()) 293 return 0; 294 295 std::swap(Pred1, Pred2); 296 std::swap(Pred1Br, Pred2Br); 297 } 298 299 if (Pred1Br->isConditional()) { 300 // If we found a conditional branch predecessor, make sure that it branches 301 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 302 if (Pred1Br->getSuccessor(0) == BB && 303 Pred1Br->getSuccessor(1) == Pred2) { 304 IfTrue = Pred1; 305 IfFalse = Pred2; 306 } else if (Pred1Br->getSuccessor(0) == Pred2 && 307 Pred1Br->getSuccessor(1) == BB) { 308 IfTrue = Pred2; 309 IfFalse = Pred1; 310 } else { 311 // We know that one arm of the conditional goes to BB, so the other must 312 // go somewhere unrelated, and this must not be an "if statement". 313 return 0; 314 } 315 316 // The only thing we have to watch out for here is to make sure that Pred2 317 // doesn't have incoming edges from other blocks. If it does, the condition 318 // doesn't dominate BB. 319 if (++pred_begin(Pred2) != pred_end(Pred2)) 320 return 0; 321 322 return Pred1Br->getCondition(); 323 } 324 325 // Ok, if we got here, both predecessors end with an unconditional branch to 326 // BB. Don't panic! If both blocks only have a single (identical) 327 // predecessor, and THAT is a conditional branch, then we're all ok! 328 if (pred_begin(Pred1) == pred_end(Pred1) || 329 ++pred_begin(Pred1) != pred_end(Pred1) || 330 pred_begin(Pred2) == pred_end(Pred2) || 331 ++pred_begin(Pred2) != pred_end(Pred2) || 332 *pred_begin(Pred1) != *pred_begin(Pred2)) 333 return 0; 334 335 // Otherwise, if this is a conditional branch, then we can use it! 336 BasicBlock *CommonPred = *pred_begin(Pred1); 337 if (BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator())) { 338 assert(BI->isConditional() && "Two successors but not conditional?"); 339 if (BI->getSuccessor(0) == Pred1) { 340 IfTrue = Pred1; 341 IfFalse = Pred2; 342 } else { 343 IfTrue = Pred2; 344 IfFalse = Pred1; 345 } 346 return BI->getCondition(); 347 } 348 return 0; 349 } 350 351 /// DominatesMergePoint - If we have a merge point of an "if condition" as 352 /// accepted above, return true if the specified value dominates the block. We 353 /// don't handle the true generality of domination here, just a special case 354 /// which works well enough for us. 355 /// 356 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 357 /// see if V (which must be an instruction) is cheap to compute and is 358 /// non-trapping. If both are true, the instruction is inserted into the set 359 /// and true is returned. 360 static bool DominatesMergePoint(Value *V, BasicBlock *BB, 361 std::set<Instruction*> *AggressiveInsts) { 362 Instruction *I = dyn_cast<Instruction>(V); 363 if (!I) { 364 // Non-instructions all dominate instructions, but not all constantexprs 365 // can be executed unconditionally. 366 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 367 if (C->canTrap()) 368 return false; 369 return true; 370 } 371 BasicBlock *PBB = I->getParent(); 372 373 // We don't want to allow weird loops that might have the "if condition" in 374 // the bottom of this block. 375 if (PBB == BB) return false; 376 377 // If this instruction is defined in a block that contains an unconditional 378 // branch to BB, then it must be in the 'conditional' part of the "if 379 // statement". 380 if (BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator())) 381 if (BI->isUnconditional() && BI->getSuccessor(0) == BB) { 382 if (!AggressiveInsts) return false; 383 // Okay, it looks like the instruction IS in the "condition". Check to 384 // see if its a cheap instruction to unconditionally compute, and if it 385 // only uses stuff defined outside of the condition. If so, hoist it out. 386 switch (I->getOpcode()) { 387 default: return false; // Cannot hoist this out safely. 388 case Instruction::Load: { 389 // We can hoist loads that are non-volatile and obviously cannot trap. 390 if (cast<LoadInst>(I)->isVolatile()) 391 return false; 392 // FIXME: A computation of a constant can trap! 393 if (!isa<AllocaInst>(I->getOperand(0)) && 394 !isa<Constant>(I->getOperand(0))) 395 return false; 396 // External weak globals may have address 0, so we can't load them. 397 Value *V2 = I->getOperand(0)->getUnderlyingObject(); 398 if (V2) { 399 GlobalVariable* GV = dyn_cast<GlobalVariable>(V2); 400 if (GV && GV->hasExternalWeakLinkage()) 401 return false; 402 } 403 // Finally, we have to check to make sure there are no instructions 404 // before the load in its basic block, as we are going to hoist the loop 405 // out to its predecessor. 406 BasicBlock::iterator IP = PBB->begin(); 407 while (isa<DbgInfoIntrinsic>(IP)) 408 IP++; 409 if (IP != BasicBlock::iterator(I)) 410 return false; 411 break; 412 } 413 case Instruction::Add: 414 case Instruction::Sub: 415 case Instruction::And: 416 case Instruction::Or: 417 case Instruction::Xor: 418 case Instruction::Shl: 419 case Instruction::LShr: 420 case Instruction::AShr: 421 case Instruction::ICmp: 422 break; // These are all cheap and non-trapping instructions. 423 } 424 425 // Okay, we can only really hoist these out if their operands are not 426 // defined in the conditional region. 427 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 428 if (!DominatesMergePoint(*i, BB, 0)) 429 return false; 430 // Okay, it's safe to do this! Remember this instruction. 431 AggressiveInsts->insert(I); 432 } 433 434 return true; 435 } 436 437 /// GatherConstantSetEQs - Given a potentially 'or'd together collection of 438 /// icmp_eq instructions that compare a value against a constant, return the 439 /// value being compared, and stick the constant into the Values vector. 440 static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){ 441 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 442 if (Inst->getOpcode() == Instruction::ICmp && 443 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { 444 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 445 Values.push_back(C); 446 return Inst->getOperand(0); 447 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 448 Values.push_back(C); 449 return Inst->getOperand(1); 450 } 451 } else if (Inst->getOpcode() == Instruction::Or) { 452 if (Value *LHS = GatherConstantSetEQs(Inst->getOperand(0), Values)) 453 if (Value *RHS = GatherConstantSetEQs(Inst->getOperand(1), Values)) 454 if (LHS == RHS) 455 return LHS; 456 } 457 } 458 return 0; 459 } 460 461 /// GatherConstantSetNEs - Given a potentially 'and'd together collection of 462 /// setne instructions that compare a value against a constant, return the value 463 /// being compared, and stick the constant into the Values vector. 464 static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){ 465 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 466 if (Inst->getOpcode() == Instruction::ICmp && 467 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_NE) { 468 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 469 Values.push_back(C); 470 return Inst->getOperand(0); 471 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 472 Values.push_back(C); 473 return Inst->getOperand(1); 474 } 475 } else if (Inst->getOpcode() == Instruction::And) { 476 if (Value *LHS = GatherConstantSetNEs(Inst->getOperand(0), Values)) 477 if (Value *RHS = GatherConstantSetNEs(Inst->getOperand(1), Values)) 478 if (LHS == RHS) 479 return LHS; 480 } 481 } 482 return 0; 483 } 484 485 /// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a 486 /// bunch of comparisons of one value against constants, return the value and 487 /// the constants being compared. 488 static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, 489 std::vector<ConstantInt*> &Values) { 490 if (Cond->getOpcode() == Instruction::Or) { 491 CompVal = GatherConstantSetEQs(Cond, Values); 492 493 // Return true to indicate that the condition is true if the CompVal is 494 // equal to one of the constants. 495 return true; 496 } else if (Cond->getOpcode() == Instruction::And) { 497 CompVal = GatherConstantSetNEs(Cond, Values); 498 499 // Return false to indicate that the condition is false if the CompVal is 500 // equal to one of the constants. 501 return false; 502 } 503 return false; 504 } 505 506 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 507 Instruction* Cond = 0; 508 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 509 Cond = dyn_cast<Instruction>(SI->getCondition()); 510 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 511 if (BI->isConditional()) 512 Cond = dyn_cast<Instruction>(BI->getCondition()); 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. 521 static Value *isValueEqualityComparison(TerminatorInst *TI) { 522 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 523 // Do not permit merging of large switch instructions into their 524 // predecessors unless there is only one predecessor. 525 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()), 526 pred_end(SI->getParent())) > 128) 527 return 0; 528 529 return SI->getCondition(); 530 } 531 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 532 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 533 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 534 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 535 ICI->getPredicate() == ICmpInst::ICMP_NE) && 536 isa<ConstantInt>(ICI->getOperand(1))) 537 return ICI->getOperand(0); 538 return 0; 539 } 540 541 /// GetValueEqualityComparisonCases - Given a value comparison instruction, 542 /// decode all of the 'cases' that it represents and return the 'default' block. 543 static BasicBlock * 544 GetValueEqualityComparisonCases(TerminatorInst *TI, 545 std::vector<std::pair<ConstantInt*, 546 BasicBlock*> > &Cases) { 547 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 548 Cases.reserve(SI->getNumCases()); 549 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 550 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i))); 551 return SI->getDefaultDest(); 552 } 553 554 BranchInst *BI = cast<BranchInst>(TI); 555 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 556 Cases.push_back(std::make_pair(cast<ConstantInt>(ICI->getOperand(1)), 557 BI->getSuccessor(ICI->getPredicate() == 558 ICmpInst::ICMP_NE))); 559 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 560 } 561 562 563 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 564 /// in the list that match the specified block. 565 static void EliminateBlockCases(BasicBlock *BB, 566 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 567 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 568 if (Cases[i].second == BB) { 569 Cases.erase(Cases.begin()+i); 570 --i; --e; 571 } 572 } 573 574 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 575 /// well. 576 static bool 577 ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 578 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 579 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 580 581 // Make V1 be smaller than V2. 582 if (V1->size() > V2->size()) 583 std::swap(V1, V2); 584 585 if (V1->size() == 0) return false; 586 if (V1->size() == 1) { 587 // Just scan V2. 588 ConstantInt *TheVal = (*V1)[0].first; 589 for (unsigned i = 0, e = V2->size(); i != e; ++i) 590 if (TheVal == (*V2)[i].first) 591 return true; 592 } 593 594 // Otherwise, just sort both lists and compare element by element. 595 std::sort(V1->begin(), V1->end()); 596 std::sort(V2->begin(), V2->end()); 597 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 598 while (i1 != e1 && i2 != e2) { 599 if ((*V1)[i1].first == (*V2)[i2].first) 600 return true; 601 if ((*V1)[i1].first < (*V2)[i2].first) 602 ++i1; 603 else 604 ++i2; 605 } 606 return false; 607 } 608 609 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 610 /// terminator instruction and its block is known to only have a single 611 /// predecessor block, check to see if that predecessor is also a value 612 /// comparison with the same value, and if that comparison determines the 613 /// outcome of this comparison. If so, simplify TI. This does a very limited 614 /// form of jump threading. 615 static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 616 BasicBlock *Pred) { 617 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 618 if (!PredVal) return false; // Not a value comparison in predecessor. 619 620 Value *ThisVal = isValueEqualityComparison(TI); 621 assert(ThisVal && "This isn't a value comparison!!"); 622 if (ThisVal != PredVal) return false; // Different predicates. 623 624 // Find out information about when control will move from Pred to TI's block. 625 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 626 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 627 PredCases); 628 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 629 630 // Find information about how control leaves this block. 631 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 632 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 633 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 634 635 // If TI's block is the default block from Pred's comparison, potentially 636 // simplify TI based on this knowledge. 637 if (PredDef == TI->getParent()) { 638 // If we are here, we know that the value is none of those cases listed in 639 // PredCases. If there are any cases in ThisCases that are in PredCases, we 640 // can simplify TI. 641 if (ValuesOverlap(PredCases, ThisCases)) { 642 if (isa<BranchInst>(TI)) { 643 // Okay, one of the successors of this condbr is dead. Convert it to a 644 // uncond br. 645 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 646 // Insert the new branch. 647 Instruction *NI = BranchInst::Create(ThisDef, TI); 648 649 // Remove PHI node entries for the dead edge. 650 ThisCases[0].second->removePredecessor(TI->getParent()); 651 652 DOUT << "Threading pred instr: " << *Pred->getTerminator() 653 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 654 655 EraseTerminatorInstAndDCECond(TI); 656 return true; 657 658 } else { 659 SwitchInst *SI = cast<SwitchInst>(TI); 660 // Okay, TI has cases that are statically dead, prune them away. 661 SmallPtrSet<Constant*, 16> DeadCases; 662 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 663 DeadCases.insert(PredCases[i].first); 664 665 DOUT << "Threading pred instr: " << *Pred->getTerminator() 666 << "Through successor TI: " << *TI; 667 668 for (unsigned i = SI->getNumCases()-1; i != 0; --i) 669 if (DeadCases.count(SI->getCaseValue(i))) { 670 SI->getSuccessor(i)->removePredecessor(TI->getParent()); 671 SI->removeCase(i); 672 } 673 674 DOUT << "Leaving: " << *TI << "\n"; 675 return true; 676 } 677 } 678 679 } else { 680 // Otherwise, TI's block must correspond to some matched value. Find out 681 // which value (or set of values) this is. 682 ConstantInt *TIV = 0; 683 BasicBlock *TIBB = TI->getParent(); 684 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 685 if (PredCases[i].second == TIBB) { 686 if (TIV == 0) 687 TIV = PredCases[i].first; 688 else 689 return false; // Cannot handle multiple values coming to this block. 690 } 691 assert(TIV && "No edge from pred to succ?"); 692 693 // Okay, we found the one constant that our value can be if we get into TI's 694 // BB. Find out which successor will unconditionally be branched to. 695 BasicBlock *TheRealDest = 0; 696 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 697 if (ThisCases[i].first == TIV) { 698 TheRealDest = ThisCases[i].second; 699 break; 700 } 701 702 // If not handled by any explicit cases, it is handled by the default case. 703 if (TheRealDest == 0) TheRealDest = ThisDef; 704 705 // Remove PHI node entries for dead edges. 706 BasicBlock *CheckEdge = TheRealDest; 707 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 708 if (*SI != CheckEdge) 709 (*SI)->removePredecessor(TIBB); 710 else 711 CheckEdge = 0; 712 713 // Insert the new branch. 714 Instruction *NI = BranchInst::Create(TheRealDest, TI); 715 716 DOUT << "Threading pred instr: " << *Pred->getTerminator() 717 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 718 719 EraseTerminatorInstAndDCECond(TI); 720 return true; 721 } 722 return false; 723 } 724 725 namespace { 726 /// ConstantIntOrdering - This class implements a stable ordering of constant 727 /// integers that does not depend on their address. This is important for 728 /// applications that sort ConstantInt's to ensure uniqueness. 729 struct ConstantIntOrdering { 730 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 731 return LHS->getValue().ult(RHS->getValue()); 732 } 733 }; 734 } 735 736 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value 737 /// equality comparison instruction (either a switch or a branch on "X == c"). 738 /// See if any of the predecessors of the terminator block are value comparisons 739 /// on the same value. If so, and if safe to do so, fold them together. 740 static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { 741 BasicBlock *BB = TI->getParent(); 742 Value *CV = isValueEqualityComparison(TI); // CondVal 743 assert(CV && "Not a comparison?"); 744 bool Changed = false; 745 746 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 747 while (!Preds.empty()) { 748 BasicBlock *Pred = Preds.pop_back_val(); 749 750 // See if the predecessor is a comparison with the same value. 751 TerminatorInst *PTI = Pred->getTerminator(); 752 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 753 754 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 755 // Figure out which 'cases' to copy from SI to PSI. 756 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 757 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 758 759 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 760 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 761 762 // Based on whether the default edge from PTI goes to BB or not, fill in 763 // PredCases and PredDefault with the new switch cases we would like to 764 // build. 765 SmallVector<BasicBlock*, 8> NewSuccessors; 766 767 if (PredDefault == BB) { 768 // If this is the default destination from PTI, only the edges in TI 769 // that don't occur in PTI, or that branch to BB will be activated. 770 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 771 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 772 if (PredCases[i].second != BB) 773 PTIHandled.insert(PredCases[i].first); 774 else { 775 // The default destination is BB, we don't need explicit targets. 776 std::swap(PredCases[i], PredCases.back()); 777 PredCases.pop_back(); 778 --i; --e; 779 } 780 781 // Reconstruct the new switch statement we will be building. 782 if (PredDefault != BBDefault) { 783 PredDefault->removePredecessor(Pred); 784 PredDefault = BBDefault; 785 NewSuccessors.push_back(BBDefault); 786 } 787 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 788 if (!PTIHandled.count(BBCases[i].first) && 789 BBCases[i].second != BBDefault) { 790 PredCases.push_back(BBCases[i]); 791 NewSuccessors.push_back(BBCases[i].second); 792 } 793 794 } else { 795 // If this is not the default destination from PSI, only the edges 796 // in SI that occur in PSI with a destination of BB will be 797 // activated. 798 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 799 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 800 if (PredCases[i].second == BB) { 801 PTIHandled.insert(PredCases[i].first); 802 std::swap(PredCases[i], PredCases.back()); 803 PredCases.pop_back(); 804 --i; --e; 805 } 806 807 // Okay, now we know which constants were sent to BB from the 808 // predecessor. Figure out where they will all go now. 809 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 810 if (PTIHandled.count(BBCases[i].first)) { 811 // If this is one we are capable of getting... 812 PredCases.push_back(BBCases[i]); 813 NewSuccessors.push_back(BBCases[i].second); 814 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 815 } 816 817 // If there are any constants vectored to BB that TI doesn't handle, 818 // they must go to the default destination of TI. 819 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 820 PTIHandled.begin(), 821 E = PTIHandled.end(); I != E; ++I) { 822 PredCases.push_back(std::make_pair(*I, BBDefault)); 823 NewSuccessors.push_back(BBDefault); 824 } 825 } 826 827 // Okay, at this point, we know which new successor Pred will get. Make 828 // sure we update the number of entries in the PHI nodes for these 829 // successors. 830 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 831 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 832 833 // Now that the successors are updated, create the new Switch instruction. 834 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, 835 PredCases.size(), PTI); 836 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 837 NewSI->addCase(PredCases[i].first, PredCases[i].second); 838 839 EraseTerminatorInstAndDCECond(PTI); 840 841 // Okay, last check. If BB is still a successor of PSI, then we must 842 // have an infinite loop case. If so, add an infinitely looping block 843 // to handle the case to preserve the behavior of the code. 844 BasicBlock *InfLoopBlock = 0; 845 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 846 if (NewSI->getSuccessor(i) == BB) { 847 if (InfLoopBlock == 0) { 848 // Insert it at the end of the function, because it's either code, 849 // or it won't matter if it's hot. :) 850 InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 851 BranchInst::Create(InfLoopBlock, InfLoopBlock); 852 } 853 NewSI->setSuccessor(i, InfLoopBlock); 854 } 855 856 Changed = true; 857 } 858 } 859 return Changed; 860 } 861 862 // isSafeToHoistInvoke - If we would need to insert a select that uses the 863 // value of this invoke (comments in HoistThenElseCodeToIf explain why we 864 // would need to do this), we can't hoist the invoke, as there is nowhere 865 // to put the select in this case. 866 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 867 Instruction *I1, Instruction *I2) { 868 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 869 PHINode *PN; 870 for (BasicBlock::iterator BBI = SI->begin(); 871 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 872 Value *BB1V = PN->getIncomingValueForBlock(BB1); 873 Value *BB2V = PN->getIncomingValueForBlock(BB2); 874 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 875 return false; 876 } 877 } 878 } 879 return true; 880 } 881 882 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 883 /// BB2, hoist any common code in the two blocks up into the branch block. The 884 /// caller of this function guarantees that BI's block dominates BB1 and BB2. 885 static bool HoistThenElseCodeToIf(BranchInst *BI) { 886 // This does very trivial matching, with limited scanning, to find identical 887 // instructions in the two blocks. In particular, we don't want to get into 888 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 889 // such, we currently just scan for obviously identical instructions in an 890 // identical order. 891 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 892 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 893 894 BasicBlock::iterator BB1_Itr = BB1->begin(); 895 BasicBlock::iterator BB2_Itr = BB2->begin(); 896 897 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 898 while (isa<DbgInfoIntrinsic>(I1)) 899 I1 = BB1_Itr++; 900 while (isa<DbgInfoIntrinsic>(I2)) 901 I2 = BB2_Itr++; 902 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) || 903 !I1->isIdenticalTo(I2) || 904 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 905 return false; 906 907 // If we get here, we can hoist at least one instruction. 908 BasicBlock *BIParent = BI->getParent(); 909 910 do { 911 // If we are hoisting the terminator instruction, don't move one (making a 912 // broken BB), instead clone it, and remove BI. 913 if (isa<TerminatorInst>(I1)) 914 goto HoistTerminator; 915 916 // For a normal instruction, we just move one to right before the branch, 917 // then replace all uses of the other with the first. Finally, we remove 918 // the now redundant second instruction. 919 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 920 if (!I2->use_empty()) 921 I2->replaceAllUsesWith(I1); 922 BB2->getInstList().erase(I2); 923 924 I1 = BB1_Itr++; 925 while (isa<DbgInfoIntrinsic>(I1)) 926 I1 = BB1_Itr++; 927 I2 = BB2_Itr++; 928 while (isa<DbgInfoIntrinsic>(I2)) 929 I2 = BB2_Itr++; 930 } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2)); 931 932 return true; 933 934 HoistTerminator: 935 // It may not be possible to hoist an invoke. 936 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 937 return true; 938 939 // Okay, it is safe to hoist the terminator. 940 Instruction *NT = I1->clone(); 941 BIParent->getInstList().insert(BI, NT); 942 if (NT->getType() != Type::VoidTy) { 943 I1->replaceAllUsesWith(NT); 944 I2->replaceAllUsesWith(NT); 945 NT->takeName(I1); 946 } 947 948 // Hoisting one of the terminators from our successor is a great thing. 949 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 950 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 951 // nodes, so we insert select instruction to compute the final result. 952 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 953 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 954 PHINode *PN; 955 for (BasicBlock::iterator BBI = SI->begin(); 956 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 957 Value *BB1V = PN->getIncomingValueForBlock(BB1); 958 Value *BB2V = PN->getIncomingValueForBlock(BB2); 959 if (BB1V != BB2V) { 960 // These values do not agree. Insert a select instruction before NT 961 // that determines the right value. 962 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 963 if (SI == 0) 964 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 965 BB1V->getName()+"."+BB2V->getName(), NT); 966 // Make the PHI node use the select for all incoming values for BB1/BB2 967 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 968 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 969 PN->setIncomingValue(i, SI); 970 } 971 } 972 } 973 974 // Update any PHI nodes in our new successors. 975 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 976 AddPredecessorToBlock(*SI, BIParent, BB1); 977 978 EraseTerminatorInstAndDCECond(BI); 979 return true; 980 } 981 982 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 983 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code 984 /// (for now, restricted to a single instruction that's side effect free) from 985 /// the BB1 into the branch block to speculatively execute it. 986 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 987 // Only speculatively execution a single instruction (not counting the 988 // terminator) for now. 989 Instruction *HInst = NULL; 990 Instruction *Term = BB1->getTerminator(); 991 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 992 BBI != BBE; ++BBI) { 993 Instruction *I = BBI; 994 // Skip debug info. 995 if (isa<DbgInfoIntrinsic>(I)) continue; 996 if (I == Term) break; 997 998 if (!HInst) 999 HInst = I; 1000 else 1001 return false; 1002 } 1003 if (!HInst) 1004 return false; 1005 1006 // Be conservative for now. FP select instruction can often be expensive. 1007 Value *BrCond = BI->getCondition(); 1008 if (isa<Instruction>(BrCond) && 1009 cast<Instruction>(BrCond)->getOpcode() == Instruction::FCmp) 1010 return false; 1011 1012 // If BB1 is actually on the false edge of the conditional branch, remember 1013 // to swap the select operands later. 1014 bool Invert = false; 1015 if (BB1 != BI->getSuccessor(0)) { 1016 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1017 Invert = true; 1018 } 1019 1020 // Turn 1021 // BB: 1022 // %t1 = icmp 1023 // br i1 %t1, label %BB1, label %BB2 1024 // BB1: 1025 // %t3 = add %t2, c 1026 // br label BB2 1027 // BB2: 1028 // => 1029 // BB: 1030 // %t1 = icmp 1031 // %t4 = add %t2, c 1032 // %t3 = select i1 %t1, %t2, %t3 1033 switch (HInst->getOpcode()) { 1034 default: return false; // Not safe / profitable to hoist. 1035 case Instruction::Add: 1036 case Instruction::Sub: 1037 // Not worth doing for vector ops. 1038 if (isa<VectorType>(HInst->getType())) 1039 return false; 1040 break; 1041 case Instruction::And: 1042 case Instruction::Or: 1043 case Instruction::Xor: 1044 case Instruction::Shl: 1045 case Instruction::LShr: 1046 case Instruction::AShr: 1047 // Don't mess with vector operations. 1048 if (isa<VectorType>(HInst->getType())) 1049 return false; 1050 break; // These are all cheap and non-trapping instructions. 1051 } 1052 1053 // If the instruction is obviously dead, don't try to predicate it. 1054 if (HInst->use_empty()) { 1055 HInst->eraseFromParent(); 1056 return true; 1057 } 1058 1059 // Can we speculatively execute the instruction? And what is the value 1060 // if the condition is false? Consider the phi uses, if the incoming value 1061 // from the "if" block are all the same V, then V is the value of the 1062 // select if the condition is false. 1063 BasicBlock *BIParent = BI->getParent(); 1064 SmallVector<PHINode*, 4> PHIUses; 1065 Value *FalseV = NULL; 1066 1067 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1068 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end(); 1069 UI != E; ++UI) { 1070 // Ignore any user that is not a PHI node in BB2. These can only occur in 1071 // unreachable blocks, because they would not be dominated by the instr. 1072 PHINode *PN = dyn_cast<PHINode>(UI); 1073 if (!PN || PN->getParent() != BB2) 1074 return false; 1075 PHIUses.push_back(PN); 1076 1077 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 1078 if (!FalseV) 1079 FalseV = PHIV; 1080 else if (FalseV != PHIV) 1081 return false; // Inconsistent value when condition is false. 1082 } 1083 1084 assert(FalseV && "Must have at least one user, and it must be a PHI"); 1085 1086 // Do not hoist the instruction if any of its operands are defined but not 1087 // used in this BB. The transformation will prevent the operand from 1088 // being sunk into the use block. 1089 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1090 i != e; ++i) { 1091 Instruction *OpI = dyn_cast<Instruction>(*i); 1092 if (OpI && OpI->getParent() == BIParent && 1093 !OpI->isUsedInBasicBlock(BIParent)) 1094 return false; 1095 } 1096 1097 // If we get here, we can hoist the instruction. Try to place it 1098 // before the icmp instruction preceding the conditional branch. 1099 BasicBlock::iterator InsertPos = BI; 1100 if (InsertPos != BIParent->begin()) 1101 --InsertPos; 1102 // Skip debug info between condition and branch. 1103 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos)) 1104 --InsertPos; 1105 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) { 1106 SmallPtrSet<Instruction *, 4> BB1Insns; 1107 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 1108 BB1I != BB1E; ++BB1I) 1109 BB1Insns.insert(BB1I); 1110 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 1111 UI != UE; ++UI) { 1112 Instruction *Use = cast<Instruction>(*UI); 1113 if (BB1Insns.count(Use)) { 1114 // If BrCond uses the instruction that place it just before 1115 // branch instruction. 1116 InsertPos = BI; 1117 break; 1118 } 1119 } 1120 } else 1121 InsertPos = BI; 1122 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst); 1123 1124 // Create a select whose true value is the speculatively executed value and 1125 // false value is the previously determined FalseV. 1126 SelectInst *SI; 1127 if (Invert) 1128 SI = SelectInst::Create(BrCond, FalseV, HInst, 1129 FalseV->getName() + "." + HInst->getName(), BI); 1130 else 1131 SI = SelectInst::Create(BrCond, HInst, FalseV, 1132 HInst->getName() + "." + FalseV->getName(), BI); 1133 1134 // Make the PHI node use the select for all incoming values for "then" and 1135 // "if" blocks. 1136 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1137 PHINode *PN = PHIUses[i]; 1138 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1139 if (PN->getIncomingBlock(j) == BB1 || 1140 PN->getIncomingBlock(j) == BIParent) 1141 PN->setIncomingValue(j, SI); 1142 } 1143 1144 ++NumSpeculations; 1145 return true; 1146 } 1147 1148 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1149 /// across this block. 1150 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1151 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1152 unsigned Size = 0; 1153 1154 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1155 if (isa<DbgInfoIntrinsic>(BBI)) 1156 continue; 1157 if (Size > 10) return false; // Don't clone large BB's. 1158 ++Size; 1159 1160 // We can only support instructions that do not define values that are 1161 // live outside of the current basic block. 1162 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1163 UI != E; ++UI) { 1164 Instruction *U = cast<Instruction>(*UI); 1165 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1166 } 1167 1168 // Looks ok, continue checking. 1169 } 1170 1171 return true; 1172 } 1173 1174 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1175 /// that is defined in the same block as the branch and if any PHI entries are 1176 /// constants, thread edges corresponding to that entry to be branches to their 1177 /// ultimate destination. 1178 static bool FoldCondBranchOnPHI(BranchInst *BI) { 1179 BasicBlock *BB = BI->getParent(); 1180 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1181 // NOTE: we currently cannot transform this case if the PHI node is used 1182 // outside of the block. 1183 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1184 return false; 1185 1186 // Degenerate case of a single entry PHI. 1187 if (PN->getNumIncomingValues() == 1) { 1188 FoldSingleEntryPHINodes(PN->getParent()); 1189 return true; 1190 } 1191 1192 // Now we know that this block has multiple preds and two succs. 1193 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1194 1195 // Okay, this is a simple enough basic block. See if any phi values are 1196 // constants. 1197 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1198 ConstantInt *CB; 1199 if ((CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i))) && 1200 CB->getType() == Type::Int1Ty) { 1201 // Okay, we now know that all edges from PredBB should be revectored to 1202 // branch to RealDest. 1203 BasicBlock *PredBB = PN->getIncomingBlock(i); 1204 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1205 1206 if (RealDest == BB) continue; // Skip self loops. 1207 1208 // The dest block might have PHI nodes, other predecessors and other 1209 // difficult cases. Instead of being smart about this, just insert a new 1210 // block that jumps to the destination block, effectively splitting 1211 // the edge we are about to create. 1212 BasicBlock *EdgeBB = BasicBlock::Create(RealDest->getName()+".critedge", 1213 RealDest->getParent(), RealDest); 1214 BranchInst::Create(RealDest, EdgeBB); 1215 PHINode *PN; 1216 for (BasicBlock::iterator BBI = RealDest->begin(); 1217 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1218 Value *V = PN->getIncomingValueForBlock(BB); 1219 PN->addIncoming(V, EdgeBB); 1220 } 1221 1222 // BB may have instructions that are being threaded over. Clone these 1223 // instructions into EdgeBB. We know that there will be no uses of the 1224 // cloned instructions outside of EdgeBB. 1225 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1226 std::map<Value*, Value*> TranslateMap; // Track translated values. 1227 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1228 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1229 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1230 } else { 1231 // Clone the instruction. 1232 Instruction *N = BBI->clone(); 1233 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1234 1235 // Update operands due to translation. 1236 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1237 i != e; ++i) { 1238 std::map<Value*, Value*>::iterator PI = 1239 TranslateMap.find(*i); 1240 if (PI != TranslateMap.end()) 1241 *i = PI->second; 1242 } 1243 1244 // Check for trivial simplification. 1245 if (Constant *C = ConstantFoldInstruction(N)) { 1246 TranslateMap[BBI] = C; 1247 delete N; // Constant folded away, don't need actual inst 1248 } else { 1249 // Insert the new instruction into its new home. 1250 EdgeBB->getInstList().insert(InsertPt, N); 1251 if (!BBI->use_empty()) 1252 TranslateMap[BBI] = N; 1253 } 1254 } 1255 } 1256 1257 // Loop over all of the edges from PredBB to BB, changing them to branch 1258 // to EdgeBB instead. 1259 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1260 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1261 if (PredBBTI->getSuccessor(i) == BB) { 1262 BB->removePredecessor(PredBB); 1263 PredBBTI->setSuccessor(i, EdgeBB); 1264 } 1265 1266 // Recurse, simplifying any other constants. 1267 return FoldCondBranchOnPHI(BI) | true; 1268 } 1269 } 1270 1271 return false; 1272 } 1273 1274 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1275 /// PHI node, see if we can eliminate it. 1276 static bool FoldTwoEntryPHINode(PHINode *PN) { 1277 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1278 // statement", which has a very simple dominance structure. Basically, we 1279 // are trying to find the condition that is being branched on, which 1280 // subsequently causes this merge to happen. We really want control 1281 // dependence information for this check, but simplifycfg can't keep it up 1282 // to date, and this catches most of the cases we care about anyway. 1283 // 1284 BasicBlock *BB = PN->getParent(); 1285 BasicBlock *IfTrue, *IfFalse; 1286 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1287 if (!IfCond) return false; 1288 1289 // Okay, we found that we can merge this two-entry phi node into a select. 1290 // Doing so would require us to fold *all* two entry phi nodes in this block. 1291 // At some point this becomes non-profitable (particularly if the target 1292 // doesn't support cmov's). Only do this transformation if there are two or 1293 // fewer PHI nodes in this block. 1294 unsigned NumPhis = 0; 1295 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1296 if (NumPhis > 2) 1297 return false; 1298 1299 DOUT << "FOUND IF CONDITION! " << *IfCond << " T: " 1300 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"; 1301 1302 // Loop over the PHI's seeing if we can promote them all to select 1303 // instructions. While we are at it, keep track of the instructions 1304 // that need to be moved to the dominating block. 1305 std::set<Instruction*> AggressiveInsts; 1306 1307 BasicBlock::iterator AfterPHIIt = BB->begin(); 1308 while (isa<PHINode>(AfterPHIIt)) { 1309 PHINode *PN = cast<PHINode>(AfterPHIIt++); 1310 if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) { 1311 if (PN->getIncomingValue(0) != PN) 1312 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 1313 else 1314 PN->replaceAllUsesWith(UndefValue::get(PN->getType())); 1315 } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, 1316 &AggressiveInsts) || 1317 !DominatesMergePoint(PN->getIncomingValue(1), BB, 1318 &AggressiveInsts)) { 1319 return false; 1320 } 1321 } 1322 1323 // If we all PHI nodes are promotable, check to make sure that all 1324 // instructions in the predecessor blocks can be promoted as well. If 1325 // not, we won't be able to get rid of the control flow, so it's not 1326 // worth promoting to select instructions. 1327 BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; 1328 PN = cast<PHINode>(BB->begin()); 1329 BasicBlock *Pred = PN->getIncomingBlock(0); 1330 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1331 IfBlock1 = Pred; 1332 DomBlock = *pred_begin(Pred); 1333 for (BasicBlock::iterator I = Pred->begin(); 1334 !isa<TerminatorInst>(I); ++I) 1335 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1336 // This is not an aggressive instruction that we can promote. 1337 // Because of this, we won't be able to get rid of the control 1338 // flow, so the xform is not worth it. 1339 return false; 1340 } 1341 } 1342 1343 Pred = PN->getIncomingBlock(1); 1344 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1345 IfBlock2 = Pred; 1346 DomBlock = *pred_begin(Pred); 1347 for (BasicBlock::iterator I = Pred->begin(); 1348 !isa<TerminatorInst>(I); ++I) 1349 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1350 // This is not an aggressive instruction that we can promote. 1351 // Because of this, we won't be able to get rid of the control 1352 // flow, so the xform is not worth it. 1353 return false; 1354 } 1355 } 1356 1357 // If we can still promote the PHI nodes after this gauntlet of tests, 1358 // do all of the PHI's now. 1359 1360 // Move all 'aggressive' instructions, which are defined in the 1361 // conditional parts of the if's up to the dominating block. 1362 if (IfBlock1) { 1363 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1364 IfBlock1->getInstList(), 1365 IfBlock1->begin(), 1366 IfBlock1->getTerminator()); 1367 } 1368 if (IfBlock2) { 1369 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1370 IfBlock2->getInstList(), 1371 IfBlock2->begin(), 1372 IfBlock2->getTerminator()); 1373 } 1374 1375 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1376 // Change the PHI node into a select instruction. 1377 Value *TrueVal = 1378 PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1379 Value *FalseVal = 1380 PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1381 1382 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); 1383 PN->replaceAllUsesWith(NV); 1384 NV->takeName(PN); 1385 1386 BB->getInstList().erase(PN); 1387 } 1388 return true; 1389 } 1390 1391 /// isTerminatorFirstRelevantInsn - Return true if Term is very first 1392 /// instruction ignoring Phi nodes and dbg intrinsics. 1393 static bool isTerminatorFirstRelevantInsn(BasicBlock *BB, Instruction *Term) { 1394 BasicBlock::iterator BBI = Term; 1395 while (BBI != BB->begin()) { 1396 --BBI; 1397 if (!isa<DbgInfoIntrinsic>(BBI)) 1398 break; 1399 } 1400 1401 if (isa<PHINode>(BBI) || &*BBI == Term || isa<DbgInfoIntrinsic>(BBI)) 1402 return true; 1403 return false; 1404 } 1405 1406 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1407 /// to two returning blocks, try to merge them together into one return, 1408 /// introducing a select if the return values disagree. 1409 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1410 assert(BI->isConditional() && "Must be a conditional branch"); 1411 BasicBlock *TrueSucc = BI->getSuccessor(0); 1412 BasicBlock *FalseSucc = BI->getSuccessor(1); 1413 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1414 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1415 1416 // Check to ensure both blocks are empty (just a return) or optionally empty 1417 // with PHI nodes. If there are other instructions, merging would cause extra 1418 // computation on one path or the other. 1419 if (!isTerminatorFirstRelevantInsn(TrueSucc, TrueRet)) 1420 return false; 1421 if (!isTerminatorFirstRelevantInsn(FalseSucc, FalseRet)) 1422 return false; 1423 1424 // Okay, we found a branch that is going to two return nodes. If 1425 // there is no return value for this function, just change the 1426 // branch into a return. 1427 if (FalseRet->getNumOperands() == 0) { 1428 TrueSucc->removePredecessor(BI->getParent()); 1429 FalseSucc->removePredecessor(BI->getParent()); 1430 ReturnInst::Create(0, BI); 1431 EraseTerminatorInstAndDCECond(BI); 1432 return true; 1433 } 1434 1435 // Otherwise, figure out what the true and false return values are 1436 // so we can insert a new select instruction. 1437 Value *TrueValue = TrueRet->getReturnValue(); 1438 Value *FalseValue = FalseRet->getReturnValue(); 1439 1440 // Unwrap any PHI nodes in the return blocks. 1441 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1442 if (TVPN->getParent() == TrueSucc) 1443 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1444 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1445 if (FVPN->getParent() == FalseSucc) 1446 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1447 1448 // In order for this transformation to be safe, we must be able to 1449 // unconditionally execute both operands to the return. This is 1450 // normally the case, but we could have a potentially-trapping 1451 // constant expression that prevents this transformation from being 1452 // safe. 1453 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1454 if (TCV->canTrap()) 1455 return false; 1456 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1457 if (FCV->canTrap()) 1458 return false; 1459 1460 // Okay, we collected all the mapped values and checked them for sanity, and 1461 // defined to really do this transformation. First, update the CFG. 1462 TrueSucc->removePredecessor(BI->getParent()); 1463 FalseSucc->removePredecessor(BI->getParent()); 1464 1465 // Insert select instructions where needed. 1466 Value *BrCond = BI->getCondition(); 1467 if (TrueValue) { 1468 // Insert a select if the results differ. 1469 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1470 } else if (isa<UndefValue>(TrueValue)) { 1471 TrueValue = FalseValue; 1472 } else { 1473 TrueValue = SelectInst::Create(BrCond, TrueValue, 1474 FalseValue, "retval", BI); 1475 } 1476 } 1477 1478 Value *RI = !TrueValue ? 1479 ReturnInst::Create(BI) : 1480 ReturnInst::Create(TrueValue, BI); 1481 1482 DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1483 << "\n " << *BI << "NewRet = " << *RI 1484 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; 1485 1486 EraseTerminatorInstAndDCECond(BI); 1487 1488 return true; 1489 } 1490 1491 /// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, 1492 /// and if a predecessor branches to us and one of our successors, fold the 1493 /// setcc into the predecessor and use logical operations to pick the right 1494 /// destination. 1495 bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1496 BasicBlock *BB = BI->getParent(); 1497 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1498 if (Cond == 0) return false; 1499 1500 1501 // Only allow this if the condition is a simple instruction that can be 1502 // executed unconditionally. It must be in the same block as the branch, and 1503 // must be at the front of the block. 1504 BasicBlock::iterator FrontIt = BB->front(); 1505 // Ignore dbg intrinsics. 1506 while(isa<DbgInfoIntrinsic>(FrontIt)) 1507 ++FrontIt; 1508 if ((!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1509 Cond->getParent() != BB || &*FrontIt != Cond || !Cond->hasOneUse()) { 1510 return false; 1511 } 1512 1513 // Make sure the instruction after the condition is the cond branch. 1514 BasicBlock::iterator CondIt = Cond; ++CondIt; 1515 // Ingore dbg intrinsics. 1516 while(isa<DbgInfoIntrinsic>(CondIt)) 1517 ++CondIt; 1518 if (&*CondIt != BI) { 1519 assert (!isa<DbgInfoIntrinsic>(CondIt) && "Hey do not forget debug info!"); 1520 return false; 1521 } 1522 1523 // Cond is known to be a compare or binary operator. Check to make sure that 1524 // neither operand is a potentially-trapping constant expression. 1525 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1526 if (CE->canTrap()) 1527 return false; 1528 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1529 if (CE->canTrap()) 1530 return false; 1531 1532 1533 // Finally, don't infinitely unroll conditional loops. 1534 BasicBlock *TrueDest = BI->getSuccessor(0); 1535 BasicBlock *FalseDest = BI->getSuccessor(1); 1536 if (TrueDest == BB || FalseDest == BB) 1537 return false; 1538 1539 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1540 BasicBlock *PredBlock = *PI; 1541 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1542 1543 // Check that we have two conditional branches. If there is a PHI node in 1544 // the common successor, verify that the same value flows in from both 1545 // blocks. 1546 if (PBI == 0 || PBI->isUnconditional() || 1547 !SafeToMergeTerminators(BI, PBI)) 1548 continue; 1549 1550 Instruction::BinaryOps Opc; 1551 bool InvertPredCond = false; 1552 1553 if (PBI->getSuccessor(0) == TrueDest) 1554 Opc = Instruction::Or; 1555 else if (PBI->getSuccessor(1) == FalseDest) 1556 Opc = Instruction::And; 1557 else if (PBI->getSuccessor(0) == FalseDest) 1558 Opc = Instruction::And, InvertPredCond = true; 1559 else if (PBI->getSuccessor(1) == TrueDest) 1560 Opc = Instruction::Or, InvertPredCond = true; 1561 else 1562 continue; 1563 1564 DOUT << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB; 1565 1566 // If we need to invert the condition in the pred block to match, do so now. 1567 if (InvertPredCond) { 1568 Value *NewCond = 1569 BinaryOperator::CreateNot(PBI->getCondition(), 1570 PBI->getCondition()->getName()+".not", PBI); 1571 PBI->setCondition(NewCond); 1572 BasicBlock *OldTrue = PBI->getSuccessor(0); 1573 BasicBlock *OldFalse = PBI->getSuccessor(1); 1574 PBI->setSuccessor(0, OldFalse); 1575 PBI->setSuccessor(1, OldTrue); 1576 } 1577 1578 // Clone Cond into the predecessor basic block, and or/and the 1579 // two conditions together. 1580 Instruction *New = Cond->clone(); 1581 PredBlock->getInstList().insert(PBI, New); 1582 New->takeName(Cond); 1583 Cond->setName(New->getName()+".old"); 1584 1585 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1586 New, "or.cond", PBI); 1587 PBI->setCondition(NewCond); 1588 if (PBI->getSuccessor(0) == BB) { 1589 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1590 PBI->setSuccessor(0, TrueDest); 1591 } 1592 if (PBI->getSuccessor(1) == BB) { 1593 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1594 PBI->setSuccessor(1, FalseDest); 1595 } 1596 return true; 1597 } 1598 return false; 1599 } 1600 1601 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1602 /// predecessor of another block, this function tries to simplify it. We know 1603 /// that PBI and BI are both conditional branches, and BI is in one of the 1604 /// successor blocks of PBI - PBI branches to BI. 1605 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1606 assert(PBI->isConditional() && BI->isConditional()); 1607 BasicBlock *BB = BI->getParent(); 1608 1609 // If this block ends with a branch instruction, and if there is a 1610 // predecessor that ends on a branch of the same condition, make 1611 // this conditional branch redundant. 1612 if (PBI->getCondition() == BI->getCondition() && 1613 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1614 // Okay, the outcome of this conditional branch is statically 1615 // knowable. If this block had a single pred, handle specially. 1616 if (BB->getSinglePredecessor()) { 1617 // Turn this into a branch on constant. 1618 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1619 BI->setCondition(ConstantInt::get(Type::Int1Ty, CondIsTrue)); 1620 return true; // Nuke the branch on constant. 1621 } 1622 1623 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1624 // in the constant and simplify the block result. Subsequent passes of 1625 // simplifycfg will thread the block. 1626 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1627 PHINode *NewPN = PHINode::Create(Type::Int1Ty, 1628 BI->getCondition()->getName() + ".pr", 1629 BB->begin()); 1630 // Okay, we're going to insert the PHI node. Since PBI is not the only 1631 // predecessor, compute the PHI'd conditional value for all of the preds. 1632 // Any predecessor where the condition is not computable we keep symbolic. 1633 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 1634 if ((PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) && 1635 PBI != BI && PBI->isConditional() && 1636 PBI->getCondition() == BI->getCondition() && 1637 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1638 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1639 NewPN->addIncoming(ConstantInt::get(Type::Int1Ty, 1640 CondIsTrue), *PI); 1641 } else { 1642 NewPN->addIncoming(BI->getCondition(), *PI); 1643 } 1644 1645 BI->setCondition(NewPN); 1646 return true; 1647 } 1648 } 1649 1650 // If this is a conditional branch in an empty block, and if any 1651 // predecessors is a conditional branch to one of our destinations, 1652 // fold the conditions into logical ops and one cond br. 1653 BasicBlock::iterator BBI = BB->begin(); 1654 // Ignore dbg intrinsics. 1655 while (isa<DbgInfoIntrinsic>(BBI)) 1656 ++BBI; 1657 if (&*BBI != BI) 1658 return false; 1659 1660 1661 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1662 if (CE->canTrap()) 1663 return false; 1664 1665 int PBIOp, BIOp; 1666 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1667 PBIOp = BIOp = 0; 1668 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1669 PBIOp = 0, BIOp = 1; 1670 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1671 PBIOp = 1, BIOp = 0; 1672 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1673 PBIOp = BIOp = 1; 1674 else 1675 return false; 1676 1677 // Check to make sure that the other destination of this branch 1678 // isn't BB itself. If so, this is an infinite loop that will 1679 // keep getting unwound. 1680 if (PBI->getSuccessor(PBIOp) == BB) 1681 return false; 1682 1683 // Do not perform this transformation if it would require 1684 // insertion of a large number of select instructions. For targets 1685 // without predication/cmovs, this is a big pessimization. 1686 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1687 1688 unsigned NumPhis = 0; 1689 for (BasicBlock::iterator II = CommonDest->begin(); 1690 isa<PHINode>(II); ++II, ++NumPhis) 1691 if (NumPhis > 2) // Disable this xform. 1692 return false; 1693 1694 // Finally, if everything is ok, fold the branches to logical ops. 1695 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1696 1697 DOUT << "FOLDING BRs:" << *PBI->getParent() 1698 << "AND: " << *BI->getParent(); 1699 1700 1701 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1702 // branch in it, where one edge (OtherDest) goes back to itself but the other 1703 // exits. We don't *know* that the program avoids the infinite loop 1704 // (even though that seems likely). If we do this xform naively, we'll end up 1705 // recursively unpeeling the loop. Since we know that (after the xform is 1706 // done) that the block *is* infinite if reached, we just make it an obviously 1707 // infinite loop with no cond branch. 1708 if (OtherDest == BB) { 1709 // Insert it at the end of the function, because it's either code, 1710 // or it won't matter if it's hot. :) 1711 BasicBlock *InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 1712 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1713 OtherDest = InfLoopBlock; 1714 } 1715 1716 DOUT << *PBI->getParent()->getParent(); 1717 1718 // BI may have other predecessors. Because of this, we leave 1719 // it alone, but modify PBI. 1720 1721 // Make sure we get to CommonDest on True&True directions. 1722 Value *PBICond = PBI->getCondition(); 1723 if (PBIOp) 1724 PBICond = BinaryOperator::CreateNot(PBICond, 1725 PBICond->getName()+".not", 1726 PBI); 1727 Value *BICond = BI->getCondition(); 1728 if (BIOp) 1729 BICond = BinaryOperator::CreateNot(BICond, 1730 BICond->getName()+".not", 1731 PBI); 1732 // Merge the conditions. 1733 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1734 1735 // Modify PBI to branch on the new condition to the new dests. 1736 PBI->setCondition(Cond); 1737 PBI->setSuccessor(0, CommonDest); 1738 PBI->setSuccessor(1, OtherDest); 1739 1740 // OtherDest may have phi nodes. If so, add an entry from PBI's 1741 // block that are identical to the entries for BI's block. 1742 PHINode *PN; 1743 for (BasicBlock::iterator II = OtherDest->begin(); 1744 (PN = dyn_cast<PHINode>(II)); ++II) { 1745 Value *V = PN->getIncomingValueForBlock(BB); 1746 PN->addIncoming(V, PBI->getParent()); 1747 } 1748 1749 // We know that the CommonDest already had an edge from PBI to 1750 // it. If it has PHIs though, the PHIs may have different 1751 // entries for BB and PBI's BB. If so, insert a select to make 1752 // them agree. 1753 for (BasicBlock::iterator II = CommonDest->begin(); 1754 (PN = dyn_cast<PHINode>(II)); ++II) { 1755 Value *BIV = PN->getIncomingValueForBlock(BB); 1756 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1757 Value *PBIV = PN->getIncomingValue(PBBIdx); 1758 if (BIV != PBIV) { 1759 // Insert a select in PBI to pick the right value. 1760 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1761 PBIV->getName()+".mux", PBI); 1762 PN->setIncomingValue(PBBIdx, NV); 1763 } 1764 } 1765 1766 DOUT << "INTO: " << *PBI->getParent(); 1767 1768 DOUT << *PBI->getParent()->getParent(); 1769 1770 // This basic block is probably dead. We know it has at least 1771 // one fewer predecessor. 1772 return true; 1773 } 1774 1775 1776 /// SimplifyCFG - This function is used to do simplification of a CFG. For 1777 /// example, it adjusts branches to branches to eliminate the extra hop, it 1778 /// eliminates unreachable basic blocks, and does other "peephole" optimization 1779 /// of the CFG. It returns true if a modification was made. 1780 /// 1781 /// WARNING: The entry node of a function may not be simplified. 1782 /// 1783 bool llvm::SimplifyCFG(BasicBlock *BB) { 1784 bool Changed = false; 1785 Function *M = BB->getParent(); 1786 1787 assert(BB && BB->getParent() && "Block not embedded in function!"); 1788 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 1789 assert(&BB->getParent()->getEntryBlock() != BB && 1790 "Can't Simplify entry block!"); 1791 1792 // Remove basic blocks that have no predecessors... or that just have themself 1793 // as a predecessor. These are unreachable. 1794 if (pred_begin(BB) == pred_end(BB) || BB->getSinglePredecessor() == BB) { 1795 DOUT << "Removing BB: \n" << *BB; 1796 DeleteDeadBlock(BB); 1797 return true; 1798 } 1799 1800 // Check to see if we can constant propagate this terminator instruction 1801 // away... 1802 Changed |= ConstantFoldTerminator(BB); 1803 1804 // If there is a trivial two-entry PHI node in this basic block, and we can 1805 // eliminate it, do so now. 1806 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 1807 if (PN->getNumIncomingValues() == 2) 1808 Changed |= FoldTwoEntryPHINode(PN); 1809 1810 // If this is a returning block with only PHI nodes in it, fold the return 1811 // instruction into any unconditional branch predecessors. 1812 // 1813 // If any predecessor is a conditional branch that just selects among 1814 // different return values, fold the replace the branch/return with a select 1815 // and return. 1816 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1817 if (isTerminatorFirstRelevantInsn(BB, BB->getTerminator())) { 1818 // Find predecessors that end with branches. 1819 SmallVector<BasicBlock*, 8> UncondBranchPreds; 1820 SmallVector<BranchInst*, 8> CondBranchPreds; 1821 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1822 TerminatorInst *PTI = (*PI)->getTerminator(); 1823 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 1824 if (BI->isUnconditional()) 1825 UncondBranchPreds.push_back(*PI); 1826 else 1827 CondBranchPreds.push_back(BI); 1828 } 1829 } 1830 1831 // If we found some, do the transformation! 1832 if (!UncondBranchPreds.empty()) { 1833 while (!UncondBranchPreds.empty()) { 1834 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 1835 DOUT << "FOLDING: " << *BB 1836 << "INTO UNCOND BRANCH PRED: " << *Pred; 1837 Instruction *UncondBranch = Pred->getTerminator(); 1838 // Clone the return and add it to the end of the predecessor. 1839 Instruction *NewRet = RI->clone(); 1840 Pred->getInstList().push_back(NewRet); 1841 1842 BasicBlock::iterator BBI = RI; 1843 if (BBI != BB->begin()) { 1844 // Move region end info into the predecessor. 1845 if (DbgRegionEndInst *DREI = dyn_cast<DbgRegionEndInst>(--BBI)) 1846 DREI->moveBefore(NewRet); 1847 } 1848 1849 // If the return instruction returns a value, and if the value was a 1850 // PHI node in "BB", propagate the right value into the return. 1851 for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end(); 1852 i != e; ++i) 1853 if (PHINode *PN = dyn_cast<PHINode>(*i)) 1854 if (PN->getParent() == BB) 1855 *i = PN->getIncomingValueForBlock(Pred); 1856 1857 // Update any PHI nodes in the returning block to realize that we no 1858 // longer branch to them. 1859 BB->removePredecessor(Pred); 1860 Pred->getInstList().erase(UncondBranch); 1861 } 1862 1863 // If we eliminated all predecessors of the block, delete the block now. 1864 if (pred_begin(BB) == pred_end(BB)) 1865 // We know there are no successors, so just nuke the block. 1866 M->getBasicBlockList().erase(BB); 1867 1868 return true; 1869 } 1870 1871 // Check out all of the conditional branches going to this return 1872 // instruction. If any of them just select between returns, change the 1873 // branch itself into a select/return pair. 1874 while (!CondBranchPreds.empty()) { 1875 BranchInst *BI = CondBranchPreds.pop_back_val(); 1876 1877 // Check to see if the non-BB successor is also a return block. 1878 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 1879 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 1880 SimplifyCondBranchToTwoReturns(BI)) 1881 return true; 1882 } 1883 } 1884 } else if (isa<UnwindInst>(BB->begin())) { 1885 // Check to see if the first instruction in this block is just an unwind. 1886 // If so, replace any invoke instructions which use this as an exception 1887 // destination with call instructions, and any unconditional branch 1888 // predecessor with an unwind. 1889 // 1890 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 1891 while (!Preds.empty()) { 1892 BasicBlock *Pred = Preds.back(); 1893 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) { 1894 if (BI->isUnconditional()) { 1895 Pred->getInstList().pop_back(); // nuke uncond branch 1896 new UnwindInst(Pred); // Use unwind. 1897 Changed = true; 1898 } 1899 } else if (InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator())) 1900 if (II->getUnwindDest() == BB) { 1901 // Insert a new branch instruction before the invoke, because this 1902 // is now a fall through... 1903 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 1904 Pred->getInstList().remove(II); // Take out of symbol table 1905 1906 // Insert the call now... 1907 SmallVector<Value*,8> Args(II->op_begin()+3, II->op_end()); 1908 CallInst *CI = CallInst::Create(II->getCalledValue(), 1909 Args.begin(), Args.end(), 1910 II->getName(), BI); 1911 CI->setCallingConv(II->getCallingConv()); 1912 CI->setAttributes(II->getAttributes()); 1913 // If the invoke produced a value, the Call now does instead 1914 II->replaceAllUsesWith(CI); 1915 delete II; 1916 Changed = true; 1917 } 1918 1919 Preds.pop_back(); 1920 } 1921 1922 // If this block is now dead, remove it. 1923 if (pred_begin(BB) == pred_end(BB)) { 1924 // We know there are no successors, so just nuke the block. 1925 M->getBasicBlockList().erase(BB); 1926 return true; 1927 } 1928 1929 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1930 if (isValueEqualityComparison(SI)) { 1931 // If we only have one predecessor, and if it is a branch on this value, 1932 // see if that predecessor totally determines the outcome of this switch. 1933 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1934 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 1935 return SimplifyCFG(BB) || 1; 1936 1937 // If the block only contains the switch, see if we can fold the block 1938 // away into any preds. 1939 BasicBlock::iterator BBI = BB->begin(); 1940 // Ignore dbg intrinsics. 1941 while (isa<DbgInfoIntrinsic>(BBI)) 1942 ++BBI; 1943 if (SI == &*BBI) 1944 if (FoldValueComparisonIntoPredecessors(SI)) 1945 return SimplifyCFG(BB) || 1; 1946 } 1947 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 1948 if (BI->isUnconditional()) { 1949 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 1950 1951 BasicBlock *Succ = BI->getSuccessor(0); 1952 // Ignore dbg intrinsics. 1953 while (isa<DbgInfoIntrinsic>(BBI)) 1954 ++BBI; 1955 if (BBI->isTerminator() && // Terminator is the only non-phi instruction! 1956 Succ != BB) // Don't hurt infinite loops! 1957 if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ)) 1958 return true; 1959 1960 } else { // Conditional branch 1961 if (isValueEqualityComparison(BI)) { 1962 // If we only have one predecessor, and if it is a branch on this value, 1963 // see if that predecessor totally determines the outcome of this 1964 // switch. 1965 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1966 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 1967 return SimplifyCFG(BB) || 1; 1968 1969 // This block must be empty, except for the setcond inst, if it exists. 1970 // Ignore dbg intrinsics. 1971 BasicBlock::iterator I = BB->begin(); 1972 // Ignore dbg intrinsics. 1973 while (isa<DbgInfoIntrinsic>(I)) 1974 ++I; 1975 if (&*I == BI) { 1976 if (FoldValueComparisonIntoPredecessors(BI)) 1977 return SimplifyCFG(BB) | true; 1978 } else if (&*I == cast<Instruction>(BI->getCondition())){ 1979 ++I; 1980 // Ignore dbg intrinsics. 1981 while (isa<DbgInfoIntrinsic>(I)) 1982 ++I; 1983 if(&*I == BI) { 1984 if (FoldValueComparisonIntoPredecessors(BI)) 1985 return SimplifyCFG(BB) | true; 1986 } 1987 } 1988 } 1989 1990 // If this is a branch on a phi node in the current block, thread control 1991 // through this block if any PHI node entries are constants. 1992 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 1993 if (PN->getParent() == BI->getParent()) 1994 if (FoldCondBranchOnPHI(BI)) 1995 return SimplifyCFG(BB) | true; 1996 1997 // If this basic block is ONLY a setcc and a branch, and if a predecessor 1998 // branches to us and one of our successors, fold the setcc into the 1999 // predecessor and use logical operations to pick the right destination. 2000 if (FoldBranchToCommonDest(BI)) 2001 return SimplifyCFG(BB) | 1; 2002 2003 2004 // Scan predecessor blocks for conditional branches. 2005 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2006 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2007 if (PBI != BI && PBI->isConditional()) 2008 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2009 return SimplifyCFG(BB) | true; 2010 } 2011 } else if (isa<UnreachableInst>(BB->getTerminator())) { 2012 // If there are any instructions immediately before the unreachable that can 2013 // be removed, do so. 2014 Instruction *Unreachable = BB->getTerminator(); 2015 while (Unreachable != BB->begin()) { 2016 BasicBlock::iterator BBI = Unreachable; 2017 --BBI; 2018 // Do not delete instructions that can have side effects, like calls 2019 // (which may never return) and volatile loads and stores. 2020 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2021 2022 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) 2023 if (SI->isVolatile()) 2024 break; 2025 2026 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) 2027 if (LI->isVolatile()) 2028 break; 2029 2030 // Delete this instruction 2031 BB->getInstList().erase(BBI); 2032 Changed = true; 2033 } 2034 2035 // If the unreachable instruction is the first in the block, take a gander 2036 // at all of the predecessors of this instruction, and simplify them. 2037 if (&BB->front() == Unreachable) { 2038 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2039 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2040 TerminatorInst *TI = Preds[i]->getTerminator(); 2041 2042 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2043 if (BI->isUnconditional()) { 2044 if (BI->getSuccessor(0) == BB) { 2045 new UnreachableInst(TI); 2046 TI->eraseFromParent(); 2047 Changed = true; 2048 } 2049 } else { 2050 if (BI->getSuccessor(0) == BB) { 2051 BranchInst::Create(BI->getSuccessor(1), BI); 2052 EraseTerminatorInstAndDCECond(BI); 2053 } else if (BI->getSuccessor(1) == BB) { 2054 BranchInst::Create(BI->getSuccessor(0), BI); 2055 EraseTerminatorInstAndDCECond(BI); 2056 Changed = true; 2057 } 2058 } 2059 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2060 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2061 if (SI->getSuccessor(i) == BB) { 2062 BB->removePredecessor(SI->getParent()); 2063 SI->removeCase(i); 2064 --i; --e; 2065 Changed = true; 2066 } 2067 // If the default value is unreachable, figure out the most popular 2068 // destination and make it the default. 2069 if (SI->getSuccessor(0) == BB) { 2070 std::map<BasicBlock*, unsigned> Popularity; 2071 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2072 Popularity[SI->getSuccessor(i)]++; 2073 2074 // Find the most popular block. 2075 unsigned MaxPop = 0; 2076 BasicBlock *MaxBlock = 0; 2077 for (std::map<BasicBlock*, unsigned>::iterator 2078 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2079 if (I->second > MaxPop) { 2080 MaxPop = I->second; 2081 MaxBlock = I->first; 2082 } 2083 } 2084 if (MaxBlock) { 2085 // Make this the new default, allowing us to delete any explicit 2086 // edges to it. 2087 SI->setSuccessor(0, MaxBlock); 2088 Changed = true; 2089 2090 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2091 // it. 2092 if (isa<PHINode>(MaxBlock->begin())) 2093 for (unsigned i = 0; i != MaxPop-1; ++i) 2094 MaxBlock->removePredecessor(SI->getParent()); 2095 2096 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2097 if (SI->getSuccessor(i) == MaxBlock) { 2098 SI->removeCase(i); 2099 --i; --e; 2100 } 2101 } 2102 } 2103 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2104 if (II->getUnwindDest() == BB) { 2105 // Convert the invoke to a call instruction. This would be a good 2106 // place to note that the call does not throw though. 2107 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2108 II->removeFromParent(); // Take out of symbol table 2109 2110 // Insert the call now... 2111 SmallVector<Value*, 8> Args(II->op_begin()+3, II->op_end()); 2112 CallInst *CI = CallInst::Create(II->getCalledValue(), 2113 Args.begin(), Args.end(), 2114 II->getName(), BI); 2115 CI->setCallingConv(II->getCallingConv()); 2116 CI->setAttributes(II->getAttributes()); 2117 // If the invoke produced a value, the Call does now instead. 2118 II->replaceAllUsesWith(CI); 2119 delete II; 2120 Changed = true; 2121 } 2122 } 2123 } 2124 2125 // If this block is now dead, remove it. 2126 if (pred_begin(BB) == pred_end(BB)) { 2127 // We know there are no successors, so just nuke the block. 2128 M->getBasicBlockList().erase(BB); 2129 return true; 2130 } 2131 } 2132 } 2133 2134 // Merge basic blocks into their predecessor if there is only one distinct 2135 // pred, and if there is only one distinct successor of the predecessor, and 2136 // if there are no PHI nodes. 2137 // 2138 if (MergeBlockIntoPredecessor(BB)) 2139 return true; 2140 2141 // Otherwise, if this block only has a single predecessor, and if that block 2142 // is a conditional branch, see if we can hoist any code from this block up 2143 // into our predecessor. 2144 pred_iterator PI(pred_begin(BB)), PE(pred_end(BB)); 2145 BasicBlock *OnlyPred = *PI++; 2146 for (; PI != PE; ++PI) // Search all predecessors, see if they are all same 2147 if (*PI != OnlyPred) { 2148 OnlyPred = 0; // There are multiple different predecessors... 2149 break; 2150 } 2151 2152 if (OnlyPred) 2153 if (BranchInst *BI = dyn_cast<BranchInst>(OnlyPred->getTerminator())) 2154 if (BI->isConditional()) { 2155 // Get the other block. 2156 BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB); 2157 PI = pred_begin(OtherBB); 2158 ++PI; 2159 2160 if (PI == pred_end(OtherBB)) { 2161 // We have a conditional branch to two blocks that are only reachable 2162 // from the condbr. We know that the condbr dominates the two blocks, 2163 // so see if there is any identical code in the "then" and "else" 2164 // blocks. If so, we can hoist it up to the branching block. 2165 Changed |= HoistThenElseCodeToIf(BI); 2166 } else { 2167 BasicBlock* OnlySucc = NULL; 2168 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 2169 SI != SE; ++SI) { 2170 if (!OnlySucc) 2171 OnlySucc = *SI; 2172 else if (*SI != OnlySucc) { 2173 OnlySucc = 0; // There are multiple distinct successors! 2174 break; 2175 } 2176 } 2177 2178 if (OnlySucc == OtherBB) { 2179 // If BB's only successor is the other successor of the predecessor, 2180 // i.e. a triangle, see if we can hoist any code from this block up 2181 // to the "if" block. 2182 Changed |= SpeculativelyExecuteBB(BI, BB); 2183 } 2184 } 2185 } 2186 2187 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2188 if (BranchInst *BI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2189 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2190 if (BI->isConditional() && isa<Instruction>(BI->getCondition())) { 2191 Instruction *Cond = cast<Instruction>(BI->getCondition()); 2192 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2193 // 'setne's and'ed together, collect them. 2194 Value *CompVal = 0; 2195 std::vector<ConstantInt*> Values; 2196 bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values); 2197 if (CompVal && CompVal->getType()->isInteger()) { 2198 // There might be duplicate constants in the list, which the switch 2199 // instruction can't handle, remove them now. 2200 std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); 2201 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2202 2203 // Figure out which block is which destination. 2204 BasicBlock *DefaultBB = BI->getSuccessor(1); 2205 BasicBlock *EdgeBB = BI->getSuccessor(0); 2206 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2207 2208 // Create the new switch instruction now. 2209 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, 2210 Values.size(), BI); 2211 2212 // Add all of the 'cases' to the switch instruction. 2213 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2214 New->addCase(Values[i], EdgeBB); 2215 2216 // We added edges from PI to the EdgeBB. As such, if there were any 2217 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2218 // the number of edges added. 2219 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2220 isa<PHINode>(BBI); ++BBI) { 2221 PHINode *PN = cast<PHINode>(BBI); 2222 Value *InVal = PN->getIncomingValueForBlock(*PI); 2223 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2224 PN->addIncoming(InVal, *PI); 2225 } 2226 2227 // Erase the old branch instruction. 2228 EraseTerminatorInstAndDCECond(BI); 2229 return true; 2230 } 2231 } 2232 2233 return Changed; 2234 } 2235